Research Rundown

Prompt Like a Scientist

Santhosh Eapen — Sat, 22 Aug 2026 02:00:40 GMT

Here is a number that should reassure anyone who feels behind on AI. In a cohort of biomedical PhD students, 81% were already using AI tools for their science, yet only 5% could write a competent prompt. After a single training module, that rose to 48%. Using AI well turns out to be teachable, and it needs no code.

The finding comes from James M. Dewar, a biochemist at Vanderbilt University, writing in Nature‘s Career column on 3 August 2026. His argument seems obvious only once said aloud: researchers are not bad at AI because they lack a computing background. They are bad at it because they treat a scientific instrument like a vending machine - put in a question, take the answer, walk away. His fix is to do instead what we were all trained to do. Treat every prompt like an experiment.

After four decades in crop protection, biotechnology, and bioinformatics, I find this reframing persuasive precisely because it asks nothing new of us. A prompt is an experimental design; the output is a result; and no result is trustworthy until it has been checked, repeated, and stress-tested against the chance that you fooled yourself. You do not need to know how a language model works internally, any more than you needed semiconductor physics to run a qPCR machine. You need the habit of asking whether a result is robust or a one-off. That reflex is second nature to a working scientist.

Set Up the Run

A good experiment starts before the first measurement, and so does a good prompt. Dewar’s setup tips are concrete. Give the model its instruments - connect web search and literature access (PubMed, Scite, Elicit) before prompting, so it reasons from the current record rather than blind from training data. Structure your inputs with Markdown, XML, or JSON, separating context from instruction from data; it is the AI equivalent of labelling your tubes. And do not write from scratch if you need not: start from validated prompts others have tested, or use meta-prompting - asking the model to draft an optimised prompt for your task before you run it.

Control the Conditions

Any experimentalist knows conditions drift over a long run, and language models are no different: reliability falls as a conversation grows. Keep conversations short. When a line of work is done, condense the useful output into one clean prompt and start fresh, because a new conversation is an uncontaminated run. And when errors creep in mid-thread, do not argue the model out of them - branch back to the last point where it was working and go from there. This is version control applied to a dialogue, and it heads off the slow accumulation of nonsense that has quietly soured many researchers on these tools.

Reproduce Before You Trust

If one tip deserves taping above every desk, it is this. Run the same prompt several times, in separate conversations, and compare. What appears every time is robust; what appears once is an artefact - the AI equivalent of a contaminated well. We do not trust a Western blot we ran once, and we should not trust a literature summary or a drafted paragraph generated once either. Reproducibility is not a checkbox we impose on our science; it is how we know anything at all. Extending it to AI costs a few minutes and buys a great deal of protection against the failure mode everyone fears: confident, fluent, and wrong.

Dewar backs this with resources rather than exhortation. His open guide on Zenodo illustrates worked prompts for everyday research tasks - manuscript summarization and on-demand literature reviews among them - the point being that a validated, shareable prompt is itself a reproducible method, no different in spirit from a published protocol.

Why This Matters for Our Field

None of this is confined to biomedicine. Those of us in spices, plantation crops, and plant pathology face exactly the multi-source, easy-to-get-wrong questions where experimental prompting earns its keep. Ask an AI for the evidence on biocontrol of Phytophthora in black pepper, or for candidate resistance loci in a Colletotrichum–host interaction, and one unverified answer is worth little. Run it three times across fresh conversations, literature search connected and inputs cleanly structured, and the results that survive all three are the ones worth taking to the bench.

The deeper point I find genuinely encouraging. The scientific method we spent our careers internalising - controls, replicates, scepticism toward an answer that arrives too easily - is not obsolete in the age of AI. It is the exact competence these tools demand. We are not starting from zero. We are the people best equipped to use them well.

The Experimental Prompting Checklist

Connect the instruments: enable web and literature search before you prompt, not after.
Structure the input: separate context, instruction, and data with Markdown, XML, or JSON.
Borrow validated protocols : start from a tested prompt, or have the model draft one.
Keep runs clean: short conversations; condense and restart rather than let a thread sprawl.
Branch back on failure: return to the last working version instead of arguing with a drifting model.
Replicate before trusting: run the prompt several times in separate conversations; keep only what reproduces.

The AI has not changed what makes a claim trustworthy. It has only changed how fast we can generate claims - which makes verification more important, not less. Prompt like a scientist, because you already are one.

Ten Thousand Tongues, One Narrowing World

Santhosh Eapen — Wed, 12 Aug 2026 02:41:42 GMT

Two thousand years ago, humanity may have spoken close to ten times as many languages as it does today. You might expect a smaller number in the older world. But it was not so! In fact, it was ten times more.

A landmark study published on 23 July 2026 in Science by Damián Blasi, Marcus Hamilton, Russell Gray and Claire Bowern, completely reorders a historical narrative most of us carry without examining it. We tend to assume that human diversity has been draining away steadily since some deep and crowded past. This new work suggests the opposite: language diversity did not decline in a straight line. Instead, it rose for thousands of years, crested around 2,000 years ago at perhaps tens of thousands of distinct tongues, and only then began the long fall that has left us with the roughly 7,500 languages counted worldwide today.

And the engine behind both the spectacular rise and the subsequent fall appears to be the exact same historical catalyst: farming.

For anyone who works on domesticated plants, that arc will feel deeply familiar. It is the identical arc of the crop itself. I want to take you through what the study actually found, how much weight its numbers can carry, and why the pattern it describes is one every agricultural scientist has already met in the field.

How do you count a language that left no trace?

Let us start with the honest problem. We know that spoken language leaves almost nothing behind in the archaeological record. Joe Salmons, a historical linguist at the University of Wisconsin–Madison who was not involved in the work, put it plainly: researchers know “surprisingly little” about how many languages people once spoke. Most previous estimates, as co-author Claire Bowern notes, were merely “back-of-the-envelope figures”.

Because of this complete lack of physical traces, the research team built a sophisticated statistical model instead of digging up an archaeological answer. What they did to estimate the number of languages spoken 12,000 years ago near the end of the last glacial period was ingenuous. They divided the era’s approximate global population by the typical size of a hunter-gatherer group that shared one language. The size of this baseline group was set at around 800 people, drawing from twentieth-century empirical studies of Alaska Native communities. That calculation returned a range of roughly 3,300 to 7,800 languages for the end of the Ice Age. They then used statistical models to project how these numbers changed as societies transitioned from mobile bands to settled farming communities and, eventually, to highly centralized states.

You need to be absolutely clear-eyed about this approach. Every single step of this model rests on a core assumption that could significantly shift the final count: one language per group, a group size borrowed entirely from one specific modern region, and a smooth demographic model laid over twelve millennia of intense human upheaval. The output is a mere projection, not a census. Heidi Harley, a linguist at the University of Arizona, described the sheer scale as “mind-blowing,” admitting that the projected numbers initially made her think “what? That’s crazy.” Both reactions belong in the same sentence: the result is striking, and it is a model. It must be treated as a well-argued hypothesis about the shape of our past, not a settled tally.

Why settling down multiplied our tongues

The reason for this is simple. Ancestral groups moved from a nomadic life to settled farming communities. Over generations, their languages naturally began to drift apart. Small, stationary groups that marry within their own community and trade locally create their own unique vocabulary and sounds. Over time, these local dialects grew into distinct languages in each isolated settlement.

The study points to strong living evidence for this. Take New Guinea and Vanuatu, for example. Small farming settlements there speak languages that differ sharply from their neighbors. This divergence is much more pronounced than among nearby hunter-gatherer groups. It turns out that staying in one place is a powerful way to diversify speech. Constant movement helps a common language spread across large distances. Staying put, however, allows that language to splinter into local variations.

We often view agriculture as a force for homogenization and the origin of monoculture. Yet, at its beginning, it did the exact opposite. This is the start of a paradox that is rarely told. Farming anchored people to specific locations. This caused linguistic diversity to grow on a massive scale. This growth surpassed anything seen during the mobile Palaeolithic era.

And then the same revolution took them back

The decline followed the systems farming created. Surplus grain supported specialists, and those specialists built states. These states needed armies, bureaucracies, and markets, which all required a common, dominant language. Trade, conquest, and migration also played a role. They mixed populations together, and this mixing reduced differences in speech.

The genetic data supports this. A 2016 study by David Reich, Iosif Lazaridis, and their colleagues found that early Near Eastern farmers were as genetically distinct from each other as modern Europeans are from East Asians. Reich noted that this diversity disappeared as these groups intermarried and spread. More recently, in April 2026, Ali Akbari and his team analyzed DNA from over 16,000 ancient individuals in West Eurasia. They found that natural selection sped up after farming began. It reshaped genes related to diet, disease, and skin color.

These pieces of evidence tell a consistent story. Farming first created a patchwork of languages, lineages, and traits. Then, it created the states and migration patterns that wore that diversity down. The rise and fall of this diversity were not accidental. They were two sides of the same change.

The lens from the plantation sector: crops did exactly this

I have spent four decades working with spices and plantation crops. Linguistics feels like a practical, everyday problem to me. It is not just distant anthropology.

Domestication is simply crop-scale settling down. It follows the same arc. Early farmers saved and sowed seeds. They did not limit their options right away. They multiplied them over thousands of years. Every valley and soil type developed its own unique landrace. Heirloom varieties like Navara rice, aromatic Wayanad pepper, and cardamom thrived in specific regions. What grew in one district was unique. It did not look or taste like the crop from the next valley over. This is the botanical version of those ten thousand languages. Small communities tended their own lines in isolation.

Modern agriculture changed this. It did to crops what empires did to speech. Markets demanded yield and uniformity. Elite cultivars spread across the globe. They pushed out the old heirloom varieties. We narrowed a gene pool that took millennia to build in just a few generations. Coffee is a prime example. It is a massive global commodity, but it sits on one of the thinnest genetic bases in agriculture. Coffee is not an exception. It is the language study sitting in your cup.

What this means for us

This finding is more than just a historical fact. It changes how we think about diversity. We often act like it is a fixed inheritance that slowly leaks away. But this study shows that the same forces can both build and destroy it. If we understand these forces, we can do more than just mourn what is lost. We can act. Here are a few things to keep in mind:

Watch the whole curve, not just the end. Don’t just look at a declining number. Ask where it sits on the arc of rise and fall. Managing a system that has already peaked is different from protecting one that is still growing.

Check the assumptions before you repeat the stat. The “ten thousand languages” figure is a strong estimate. But it relies on a specific group-size model. When you see a big number, find out what is doing the heavy lifting before you quote it.

Variety needs active work, not just a seed bank. Saving seeds in an attic isn’t enough. We need to keep things like Navara rice or Wayanad pepper growing in the ground. Conservation is a choice we make every day to resist the push toward a single, uniform monoculture.

Look for the double-edged sword in your field. Farming created thousands of tongues and then erased them. Most tools work both ways. Figure out if the systems you study are building diversity or grinding it down. That is more useful than calling them “good” or “bad.”

The languages we lost at the dawn of farming will not come back, and neither will most of the landraces. But the mechanism the Science team has sketched is still turning, in our crops and our speech alike. Knowing its shape is the first step toward deciding, this time, how much of the mosaic we intend to keep.

AI-Powered Graphical Abstracts: Building Your Paper's Best Advertisement

Santhosh Eapen — Mon, 03 Aug 2026 02:45:47 GMT

I have never met a researcher who enjoyed making a graphical abstract. We are trained to write. Nobody trained us to draw.

And yet the figure is the part that travels. Your methods section will be read closely by perhaps a dozen people. The graphical abstract gets screenshotted, dropped into someone else’s slide deck, and shown at a conference you were not invited to. The evidence supports this, though less cleanly than the enthusiasts suggest. Papers carrying a graphical abstract reliably attract more online attention and more downloads. On citations the literature is split. A study in gastroenterology and hepatology journals found a lift; a larger propensity-matched analysis across fields found none. So I would not promise you citations. Visibility, yes.

My career began nearly forty years ago, a time when the graphical abstract was unheard of. Their widespread adoption is a recent phenomenon, gathered mostly over the last decade. For the better part of my professional life, the visibility they offer was locked behind software I simply could not use. We were a generation raised on Word and PowerPoint; the Adobe suite was a luxury we could rarely justify. The choice was binary: spend months mastering a vector editor, or hire a professional who already had. Most of us chose a third path and just sent the manuscript without a figure.

That barrier has now gone, and I think this is one of the genuinely good things AI has done for working researchers. Ananya Thakur made the case in Nature recently and called it a democratizing force. She is right. But she also wrote the sentence that ought to sit above every one of these tools: the graphic carries your name, not the AI’s.

What follows is the method I would use. It is not the only one.

Before anything else, read the journal’s rules

This is dull and it is also the step that decides what you are permitted to submit, so do it first.

The policies are not consistent with one another. Nature allows AI-assisted figures but wants the tool disclosed, usually in the Methods, and separates AI-assisted data visualisation (broadly acceptable) from AI-generated conceptual illustration (more scrutiny, and never standing in for real data). Science is stricter and treats undisclosed AI content as misconduct. Cell Press wants the figure labelled in the legend and bars AI images that purport to show primary experimental data.

The rule that keeps you safe under all three: AI builds the schematic, never the result. The arrows, the cartoon cell, the pathway diagram. Not the data. Keep a record of the tool and the prompts, the way you keep a lab notebook. You will need it for the disclosure statement, and it takes thirty seconds at the time and an afternoon to reconstruct later.

Decide the one thing you are saying

Most bad graphical abstracts are not ugly. They are overloaded.

A figure that faithfully maps every interaction in your paper communicates nothing at all. The reader’s eye has nowhere to go. A figure carrying one idea gets remembered for years. This is the hardest part of the whole exercise and it happens on paper, before you open anything.

Three questions get you there. What shape is the story: a mechanism, a process, a set-up, a workflow, a comparison, a timeline? Naming it narrows everything downstream.

Who is looking at it? This one is underrated. A plant pathologist wants the infection pathway. An extension officer wants to know what to spray and when. Same finding, two different figures, and trying to serve both produces a figure that serves neither.

And then the sentence. Force the paper down to one claim you could say out loud. Mine, for a worked example: Phytophthora moves through the black pepper vine’s vascular system, and a soil-applied biocontrol stops it at the root collar. One pathway. One intervention point. Everything else has to argue for its place against that sentence, and most of it loses.

Our recent Gene paper linked two sibling species - Phytophthora capsici and P. tropicalis - to black pepper foot rot disease. The core takeaway is that this disease is caused by a species complex sharing a convergent carbohydrate-degrading and effector arsenal, which management strategies must target. The graphical abstract should have highlighted this headline finding, omitting supporting details like scaffold counts, assembly sizes, and exact gene numbers.

A chat model is useful here, and this is where I find it most useful of all. Paste in your abstract and ask it to strip the paper to a single message, name the audiences, and propose ways of drawing an abstract concept like induced resistance or systems-level interaction. Treat what comes back as a first sketch from a well-read colleague who has not done the work. Sometimes it sees the framing you had missed. Often it does not.

“Here is the abstract of my paper. In one sentence, what is the single most important finding a graphical abstract should communicate? Then suggest three simple visual metaphors for [your concept], and tell me which audience each would suit.”

Match the tool to the job

There is no best tool. There is a best tool for each part of the work, and the commonest mistake is asking one tool to do all of it.

Broadly: a conversational model to think with, a science-figure platform to build the accurate schematic, a vector editor to finish. Generative image models belong at the sketching stage and nowhere near the final file. They will draw you a chloroplast with the wrong number of membranes and do it beautifully.

One thing worth insisting on regardless of what you pick: the final file must be editable vector. SVG, EPS or PDF. A journal will ask you to change one label at proof stage and you do not want to rebuild the figure to do it.

Briefing the tool

Output quality tracks brief quality almost exactly. Three decisions are worth making deliberately before you type a prompt.

Colour, and how little of it to use. Warm colours pull the eye and read as danger or activity, which is why they suit the pathogen or the key interaction. Cool colours read as healthy, normal, baseline. That much is standard. The part people get wrong is quantity. Two or three colours, high contrast, used consistently, will beat a rainbow every single time. And roughly one man in twelve has some colour-vision deficiency, so colour must never be the only thing separating two elements. Shape or label has to carry it too.

“I want the infected vascular pathway in warm tones and healthy tissue in cool tones, on white. Suggest a limited, high-contrast, colour-blind-safe palette with hex codes.”

Composition, which is really about where the eye goes. Left to right for a process with a start and a finish. A loop for a cycle. A fork for a comparison or a decision. Decide before you generate, because retrofitting a flow onto a finished figure rarely works.

“Here is my draft graphical abstract. Suggest two alternative compositions and better label placements so the eye moves from infection to intervention.”

Opacity, occasionally. You will not need this often. But fading a background structure, the plant or the soil or the cell, lets the central reaction dominate while the reader still knows where they are standing.

“Which elements should stay visually dominant and which can fade back? Do I need a context background at all, or would plain white serve the message better?”

Then check it properly

The first output is a draft. Treat it as one.

This next part is the bit that cannot be delegated, and it is where the scientist rather than the software does the work. My list:

Variation should mean something. A small set of text sizes, line weights, colours. Change an arrowhead from open to solid and a careful reader will spend twenty seconds hunting for the significance of it. There is none. You have cost them twenty seconds.
White space is not wasted. A clean background usually beats a busy one. Clutter is what makes a figure forgettable.
Print it in greyscale. This single test catches most accessibility failures. If two colours collapse into the same grey, you have lost the colour-blind reader and the person reading a photocopy at the same time.
Give it to someone outside your subfield. Ten seconds. If they cannot tell you the message, it is not finished. This is uncomfortable and it is the most useful thing on the list.
Proofread every character in the image. AI models are careless with text inside graphics in a way that is genuinely startling. Then check legend, scale bars, units, abbreviations, gene and protein formatting, and the journal’s own rules.
Confirm the science. Every arrow. Does the schematic imply a relationship your data do not support? A handsome figure that overstates the finding is worse than no figure, because it will be believed and shared by people who have never read the paper
.

What has changed and what has not

The barrier is down. Someone with no design training can now produce a clear, publishable graphical abstract in an afternoon, and I would encourage every author to take advantage of that.

What has not changed is who answers for it. The tool did the labour. It does not carry the judgement. That figure goes out under your name, makes a claim about your science to people who will never read the paper, and stands as the most widely seen thing you publish that year.

While it is futile to complain about the subpar performance of Indian academic journals, it is high time they embrace graphical abstracts. This single addition would significantly enhance their impact and global reach. Given the abundance of modern tools designed to elevate our skills, researchers should actively leverage them to deliver clear, compelling presentations of their findings.

Use the tools. Then check the figure the way you would check a result, because that is what it is.

The short version

Read the journal’s AI policy before you start, and keep a record of tools and prompts.
Write the one sentence. Do not open a tool until you can.
Chat model to think, science-figure platform to build, vector editor to finish. Export editable vector.
Two or three colours, one clear flow, plenty of white space.
Greyscale test, proofread the text, verify every arrow, then show a colleague.

Which "AI Scientist" Should You Actually Use? A Field Guide for the Perplexed Researcher

Santhosh Eapen — Tue, 21 Jul 2026 02:30:15 GMT

Earlier this month, Euan Ashley, a geneticist and cardiologist at Stanford, did something that would have been science fiction a decade ago. In 2010, he led the first clinical analysis of a human genome, a project that took his team of 31 scientists nine months. This July, while unpacking after a holiday, he handed his own genome to Anthropic’s Claude and asked it to do the same job. Thirty minutes later, the tool had flagged an Alzheimer’s risk allele and gene variants affecting how he metabolises drugs. “There is no world in which this is not utterly remarkable,” Ashley wrote.

I felt that story in my bones. For the better part of a decade, my team worked to decipher the genome of the Phytophthora species that devastates black pepper - the very crop that once drew the world to India’s shores. We tried outsourcing. Then we trained ourselves. Then we built our own computing facilities. It still took more than ten years. So reading that a holidaying scientist had matched a nine-month, 31-person effort in half an hour was, for me, equal parts marvel and quiet ache.

If you have spent any time in a lab, that story lands somewhere between exhilarating and unsettling. A new class of software - variously called “AI scientists,” “research agents,” or “co-scientists” - now promises to compress hours of literature review, data wrangling, and analysis into minutes. But a fresh wave of these tools launched in just the past few weeks, each pitching itself as the one your lab needs. Claude Science. OpenAI’s Prism. Google DeepMind’s Co-Scientist. The open-source Biomni. And, researchers will tell you, many more.

So which one suits your work? Drawing on a recent Nature explainer by Ewen Callaway and the launch details of each platform, here is a practical guide to sorting the genuinely useful from the merely impressive - without losing your scientific judgement in the process.

What these tools actually are

Start with what they are not. An “AI scientist” is not the same as a specialized model like AlphaFold, which does one narrow thing - predict protein structures - extraordinarily well. These new tools are generalists. They are built on the large language models that power ordinary chatbots, but wrapped in an agentic layer: give one a request, and it breaks the job into steps, then recruits external software, databases, and sometimes specialised models to carry each step out.

In practice that means help with the unglamorous middle of research - literature reviews, data analysis, figure generation, protocol drafting, manuscript preparation. Gabriele Corso, who runs the protein-design firm Boltz, set a Claude agent loose on designing an antibody that recognised two therapeutic targets, using his company’s own folding tools. The AI’s output matched what his experienced designers would have intuited.

“Work that usually takes me hours now takes minutes, I can really spend my time on the science that needs a human.” - Yuanhao Qu of the start-up Phylo

That last phrase is the whole game. These tools are not replacing the scientist. They are clearing the underbrush so the scientist can get to the clearing faster.

The main players, and what each is good at

The field is crowded, but the leading tools sort into recognisable roles. (I wrote about Claude Science in this newsletter not long ago; it is only one of a fast-growing crowd, so let me start there.)

Claude Science (Anthropic, launched 30 June 2026) is a workbench rather than a new model - it runs existing Claude models inside an environment purpose-built for biology. A coordinating agent draws on more than 60 curated skills and connectors for genomics, single-cell, proteomics, structural biology, and cheminformatics, and can spin up specialist sub-agents. It renders 3D protein structures, genome browser tracks, and chemical structures natively, and - importantly - a reviewer agent checks citations and calculations and flags errors. It is strong for hands-on analytical tasks late in a project.

Co-Scientist (Google DeepMind) sits at the opposite end of the workflow. It is a hypothesis machine: a multi-agent system that generates ideas, then debates and ranks them through a kind of internal tournament before evolving the best. Clare Bryant, a Cambridge immunologist and early adopter, fed it a grant application and her data. Some suggestions were unworkable; others were squarely in her wheelhouse. Her lab is now testing one - an experiment she reckons might otherwise have taken two years to arrive at. “You feel like you’re talking to an oracle,” says Stanford’s Gary Peltz, who used it to find existing drugs for a liver-fibrosis model.

Biomni (Stanford, published in Science in 2026) is the open-source option, already used by more than 15,000 scientists and now spun out into the start-up Phylo. It handles casual-language prompts - “Why are these patients responding differently to the drug?” - and was tested across more than 400 research tasks. It approaches human-level performance on database querying, sequence analysis, and molecular cloning, but the team is candid that it still struggles where nuanced clinical judgement or deep biological synthesis is needed.

Prism (OpenAI) is narrower in scope but useful: a free, LaTeX-native writing and collaboration workspace powered by GPT-5.2, aimed at drafting, revising, and preparing manuscripts for publication rather than running analyses.

How to actually choose

Here is the uncomfortable truth from the people building and studying these tools: fewer than 20% of labs have genuinely embedded AI scientists into their research. Most exposure is still the shallow kind - generating slides, drafting emails. “It’s really important that people actually try these things out,” says Ashu Singhal of the platform Benchling, “rather than simply trusting what gets shared in headlines.”

Three principles cut through the noise.

Match the tool to the stage of your project. Hypothesis-generating systems like Co-Scientist earn their keep at the very start, when you are still deciding what is worth doing. Analytical workbenches like Claude Science and Biomni pay off later, on defined tasks such as genomic data analysis. A writing tool like Prism belongs at the end. No single tool owns the whole pipeline, so expect to trial several.

Start small and verifiable. Corso’s advice is the most practical thing in the entire discussion: begin with tasks whose output you can check easily. “Worst case, you have to do them over.” A small, checkable task teaches you where a tool is reliable and where it quietly invents things - at a stage where a wrong answer costs you an afternoon, not a paper.

Weigh openness against convenience. An open-source tool like Biomni gives you transparency and no licence fee, at the cost of setup effort. A commercial workbench gives you polish and support, at the cost of dependence on a vendor. For an academic lab used to assembling its own pipelines, that trade-off may already feel familiar.

Can you trust an AI scientist?

The honest answer is: sometimes, and only if you check. The Science paper on Biomni found it matched experts on certain tasks - rare-disease diagnosis, cell-sequencing data analysis - often in a fraction of the time. On tasks demanding deep biological reasoning, it fell short.

Anthony Gitter, a computational biologist at the Morgridge Institute for Research in Wisconsin, put several general-purpose tools through a tougher test: a 2025 community challenge to predict preclinical drug properties such as toxicity, judged against unreleased experimental data from a biotech company. The tools did reasonably - but not as well as human experts, including those at large pharmaceutical firms. “They were doing generally reasonable things,” Gitter says, “but it was not pushing the boundaries.” His hands-on experience with Biomni was mixed: “Some things I wanted to try with it worked well, and some crashed pretty badly.”

That unreliability is survivable - a tool does not have to be right to be useful - but only if it shows its working. The feature researchers prize most is not the answer itself but the reasoning trail behind it. “When using any of these tools, it’s my responsibility to know what claims are being made and how those claims are being supported,” says Gitter. Some models do this far better than others, and that transparency should weigh heavily in which one you pick.

The catch nobody advertises

For Ashley, the productivity gain has been “unspeakably huge.” But the scientists Nature spoke to share a quieter worry worth taking seriously: what these tools do to the people still learning the craft. “You only learn to see certain types of mistakes by having made them yourself many times in the past,” Gitter notes - and an agent that clears away the grunt work also clears away part of the apprenticeship. The same risk shadows experienced scientists working outside their own fields, where the instinct for “that can’t be right” simply isn’t there.

Clare Bryant encourages her trainees to use Co-Scientist to form ideas, but she leans on decades of experience to catch its errors. “I have the experience to know when something’s not true,” she says. “The biggest question I have is, how do I train students.” That tension - between the acceleration these tools offer and the hard-won judgement they can quietly short-circuit - is the one every lab will have to manage for itself.

A note from where I sit

In India, where I have spent four decades in crop-protection research, AI tools are still treated by many as something close to taboo - a shortcut to be suspected rather than a scalpel to be mastered. I understand the caution; I lived the years of doing everything the hard way, and I know what it costs to trust a result you did not build by hand. But these tools are evolving at a pace that caution alone cannot outlast. Scientists across the globe now lean on them daily, and the refrain is remarkably consistent: work that used to take hours takes minutes, freeing them to spend their time on the science that genuinely needs a human. That is not a threat to good science. Used with eyes open, it may be the most time our discipline has ever been handed back.

What to do this week

If you are curious but unconvinced, four concrete steps will tell you more than any launch announcement:

Pick one real, small task from your current project - a literature scan, a tidy dataset, a figure - whose correct answer you already know or can quickly check.

Match a tool to it. Need ideas? Try a hypothesis generator. Need analysis? Try a workbench. Need a cleaner manuscript? Try a writing tool.

Run it, then audit the trace. Do not just read the answer - follow how it got there. That is where you learn whether to trust it on a task you can’t check.

Compare two tools on the same task. The differences in how they reason, cite, and fail will teach you more than any review, including this one.

The genome that took 31 people nine months now takes one person half an hour. That shift is real, and it is not slowing down. The researchers who benefit first will not be the ones who trust the headlines, nor the ones who dismiss the whole thing as hype - but the ones who roll up their sleeves, run a small experiment on the tools themselves, and keep their scientific scepticism switched firmly on.

Additional Reading

Anthropic. (2026, June 30). Claude Science, an AI workbench for scientists. Anthropic. https://www.anthropic.com/news/claude-science-ai-workbench
Callaway, E. (2026, July 10). Which ‘AI scientist’ suits your lab? A guide for the perplexed. Nature. https://doi.org/10.1038/d41586-026-02091-6
Huang, K., et al. (2026). Autonomous biomedical research with an artificial intelligence agent. Science. https://doi.org/10.1126/science.adz4351
Google DeepMind. (2026). Accelerating scientific discovery with Co-Scientist. Nature. https://doi.org/10.1038/s41586-026-10644-y
OpenAI. (2026). Introducing Prism. OpenAI. https://openai.com/index/introducing-prism/
Santhosh Eapen (2026, July 07). Claude Science: One workbench, not a dozen tabs. https://open.substack.com/pub/sanjoe

Subscribe now

Leave a comment

Seven Beans, Two Centuries, One Narrow Gene Pool

Santhosh Eapen — Mon, 13 Jul 2026 06:20:30 GMT

Every Indian coffee story begins with the same legend. The Sufi saint Baba Budan, returning from Mecca in the seventeenth century, carried seven coffee seeds strapped to his body and planted them in the hills of Chandragiri, in what is now Chikkamagaluru. From those seeds, the story goes, grew an industry that today employs around two million people and earns the country a record 1.8 billion dollars a year in exports.

It is a good legend. It is also, unintentionally, a precise description of a genetic problem.

Seven seeds is a founder population. A founder population is a bottleneck. And Coffea arabica was already bottlenecked before Baba Budan touched it, having arisen from a chance cross between C. canephora and C. eugenioides in a small corner of Ethiopia and Yemen, then propagating largely by self-pollination. India inherited a narrow crop and narrowed it further.

We have spent four centuries building a globally admired coffee sector on that inheritance. Climate change is now asking what else the inheritance contains. The honest answer is that we do not have much genetic room left, and the reserves that could give us more are sitting in a research collection in Karnataka and in forests we are steadily clearing.

Research Rundown reads global science through an Indian researcher's lens, one story a week. No noise, no filler. If that is your kind of reading, come aboard.

Subscribe now

What India actually grows, and what it has quietly stopped growing

Look past the legend to the ledger. India is the world’s seventh-largest coffee producer. Karnataka accounts for roughly 70 percent of output, Kerala about 20 percent, and Tamil Nadu around 5 percent. Some 250,000 growers work this crop, and 98 percent of them are smallholders.

Now look at the composition of that output, because this is where the story hides. Robusta makes up close to three-quarters of Indian production. Arabica, the species with the higher market value and the finer cup, has been in a long, quiet retreat. The Coffee Board’s post-blossom estimate for 2025 to 2026 put arabica at 118,125 tonnes against robusta’s 284,875 tonnes. Yields tell the same story more brutally. Arabica returns roughly 384 kilograms per hectare against robusta’s 1,276.

Growers are not abandoning arabica out of fashion. They are abandoning it because it keeps losing. It loses to white stem borer. It loses to leaf rust. It loses to a monsoon that arrives when it should not and vanishes when it should not. Robusta is the rational choice, and every hectare that shifts is a hectare of genetic and ecological simplification.

That shift is not a Karnataka curiosity. It is the same shift happening across tropical Asia, and it is what a crop looks like when it runs out of adaptive options.

The pests that drew the map

India’s coffee geography was drawn by two organisms, and neither has gone away.

White stem borer, Xylotrechus quadripes, began devastating plantations in the 1860s. Coffee leaf rust, Hemileia vastatrix, arrived from Sri Lanka in 1869 and rewrote the industry across South Asia. Indian planters responded with an agronomic answer that has since become the sector’s signature: they grew coffee under dense, diverse native shade. Indian coffee is shade-grown not primarily as a marketing story but as a survival strategy.

That strategy has held for a century and is now under pressure from two directions at once. Kodagu, the single largest coffee-growing district, lost about 77,000 hectares of natural forest, roughly 30 percent, between 1977 and 1997, even as coffee area doubled. Growers under financial stress increasingly replace native shade with exotic timber species, and research from the Western Ghats shows the substitution has real costs: multi-species native canopies buffer the effects of extreme rainfall better than monospecific Grevillea stands, and shade-tree diversity itself lifts both coffee yield and quality.

Meanwhile the climate has stopped cooperating. Analysis reported in early 2026 found that climate change has added roughly 30 extra extreme-heat days a year to India’s coffee farms. Scenario modelling for the cardamom and coffee hotspots of the Western Ghats projects temperature increases spanning 0.5 to 8.5 degrees Celsius and enormous swings in precipitation under RCP 4.5 and 8.5. The 2025 monsoon delivered the consequence directly: excessive rain through berry formation triggered black rot in arabica and fruit rot in robusta, which is why planter associations in Karnataka and Kerala publicly doubted the Coffee Board’s optimistic 403,000-tonne forecast.

A crop with thin genetics, a thinning canopy, and a destabilising monsoon is a crop with very little margin.

What Kodagu taught me

I did not come to this subject from the literature. I came to it from the field.

The early part of my career (1987-93) was spent in Cardamom Research Centre, Appangala, Kodagu, in the coffee tracts of Coorg, and those years shaped how I read every plantation-crop crisis since. Today it is a Regional Station of ICAR-Indian Institute of Spices Research, Kozhikode. What I remember is not the statistics. It is the sound of a planter tapping an arabica stem and knowing, from the hollowness, that the white stem borer had already won. It is the ridge lines where the old native shade still stood, and the neighbouring blocks where it did not, and how differently the two behaved in a bad year.

Kodagu is not an abstraction to me. It is the district that grows a third of India’s coffee, and it is where I learned that a pest problem is never only a pest problem. Stem borer pressure tracked the canopy. Rust tracked the variety. The estates that had kept their diversity, in trees and in planting material, absorbed shocks that flattened the estates that had not.

Those lessons were unfashionable then. Yields were rising, replanting with a narrow set of high-performing selections looked like progress, and the argument for keeping messy genetic and ecological variety around sounded like sentimentality. It was not sentimentality. It was insurance, and the premiums were being quietly cancelled.

Four decades later I am watching the same argument play out with better data and less time.

We have done this before, and we have forgotten

Here is the part of the story Indian researchers should be telling louder.

India has already been rescued once by genes from outside cultivated arabica. The most widely planted arabica selections in the country today, Chandragiri (Sln.13) and Selection 9, are Sarchimor and related types. Their rust resistance traces back to Hybrido de Timor, a spontaneous hybrid between C. arabica and C. canephora. Cauvery, Sln.12, comes from the same lineage. Our commercial robusta hybrid CxR is itself an interspecific cross.

Read that again. The varieties holding up Indian arabica cultivation right now carry introgressed genes from another species. Interspecific germplasm is not an exotic future proposal for Indian coffee. It is the load-bearing wall of the present crop, installed decades ago by breeders who understood exactly what they were doing.

So when the international literature urges breeders to look at wild Coffea relatives such as the heat-tolerant C. stenophylla, rediscovered in West Africa after being written off, or the drought-adapted Madagascan species, Indian scientists should not hear a novel idea. They should hear a familiar one, arriving late, with higher stakes.

The collection we are sitting on

India is better positioned than the pessimism suggests, and worse positioned than the complacency admits.

The Central Coffee Research Institute at Balehonnur, which marked its centenary in 2025, maintains a germplasm collection of about 300 arabica accessions, 73 robusta, and 17 other Coffea species. Over its history CCRI has released 13 arabica and 3 robusta varieties, and a white stem borer tolerant arabica is now in multi-location trials.

That is a serious national asset, built patiently over a century by people who will never be famous for it. It is also modest against the scale of the problem. Seventeen non-cultivated species is a start, not a strategy, when a global assessment led by Kew concluded that around 60 percent of wild Coffea species face extinction. The insurance policy for the world’s coffee is burning while we are still deciding whether to read it.

And breeding a perennial is punishingly slow. A new coffee variety takes years for the plant to mature and many generations of backcrossing to stabilise. Marker-assisted selection, now genuinely feasible since the arabica genome was resolved in detail, can compress that timeline. It cannot compress it into a monsoon cycle. Every year we delay pre-breeding is a year subtracted from a response we will need in the 2040s.

The smallholder is the whole game

None of this reaches a farm by itself.

Ninety-eight percent of India’s coffee growers are smallholders. A climate-resilient variety that exists only in a trial plot at Balehonnur protects nobody in Wayanad. The bottleneck between a released variety and a planted one is nursery capacity, clean planting material, extension reach, and the simple economics of a grower who cannot afford a failed replanting.

This is where India’s coffee story converges with every other plantation crop I have worked on. The pattern I first saw on those Kodagu estates repeated itself later in black pepper and in cardamom. The elite lines get narrower, the landraces get neglected, the wild relatives go unconserved, and then a new pathogen or an unfamiliar season exposes exactly how little variation remains in the cultivated stock. The germplasm that could have answered the question sits underfunded, or uncollected, in a forest that is no longer there.

India holds Geographical Indication tags for Coorg Arabica, Wayanad Robusta, Chikmagalur Arabica, Araku Valley Arabica, Bababudangiris Arabica, and Monsooned Malabar. We have learned to market our coffee’s distinctiveness with real sophistication. We have not learned to conserve the genetic base that distinctiveness rests on with anything like the same energy.

What India should do now

The Coffee Board has set a target of 9 lakh tonnes by 2047. That number is achievable only if the genetic and ecological foundations hold. Four priorities follow directly from the evidence.

Fund pre-breeding at CCRI as national infrastructure. Expand the wild and interspecific holdings well beyond 17 species, characterise them genomically, and support the slow work of moving traits into usable Indian breeding lines. This is unglamorous, decade-scale science that no market will pay for.
Conserve the Western Ghats canopy as a coffee asset, not a competing land use. Native multi-species shade demonstrably buffers climate extremes and improves yield and quality. Incentivise growers to keep native trees instead of penalising them for it.
Solve the last mile. Resilient varieties and clean planting material must reach smallholders at a price and scale that matter. Otherwise our breeding programme is an academic exercise with a good annual report.
Treat arabica’s retreat as a warning, not a market signal. Every hectare converted to robusta narrows our options. Some conversion is rational. All of it is a loss of adaptive capacity we are not accounting for anywhere.

Baba Budan’s seven seeds were a spectacular piece of luck. They were also, genetically speaking, a very thin margin, and we have been living on it for four hundred years. The margin is now closing. India has the institutions, the collection, and the scientific tradition to widen it again, and roughly one breeding generation in which to start.

The breeders at Balehonnur are ready. The planters of Kodagu, where I learned all of this the slow way, have been telling us the same thing for years. The question, as always, is whether we will fund the century of work before we need it, or after.

Pass it on · Share with a colleague · Forward to someone in the field

Davis, A. P., Chadburn, H., Moat, J., O’Sullivan, R., Hargreaves, S., & Nic Lughadha, E. (2019). High extinction risk for wild coffee species and implications for coffee sector sustainability. Science Advances, 5(1), eaav3473. https://doi.org/10.1126/sciadv.aav3473
Davis, A. P., Mieulet, D., Moat, J., Sarmu, D., & Haggar, J. (2021). Arabica-like flavour in a heat-tolerant wild coffee species. Nature Plants, 7(4), 413–418. https://doi.org/10.1038/s41477-021-00891-4
Murugan, M., Shetty, P. K., Anandhi, A., Ravi, R., Alappan, S., Vasudevan, M., & Dhanya, M. K. (2022). Climate change scenarios, their impacts and implications on Indian cardamom-coffee hot spots. Frontiers in Sustainable Food Systems, 6, 1057617. https://doi.org/10.3389/fsufs.2022.1057617
Nesper, M., Kueffer, C., Krishnan, S., Kushalappa, C. G., & Ghazoul, J. (2017). Shade tree diversity enhances coffee production and quality in agroforestry systems in the Western Ghats. Agriculture, Ecosystems & Environment, 247, 172–181. https://doi.org/10.1016/j.agee.2017.06.024
Salojärvi, J., Rambani, A., Yu, Z., Guyot, R., … Albert, V. A. (2024). The genome and population genomics of allopolyploid Coffea arabica reveal the diversification history of modern coffee cultivars. Nature Genetics, 56(4), 721–731. https://doi.org/10.1038/s41588-024-01695-w

If this piece gave you something to think about, that is the whole point of Research Rundown. New pieces land weekly, always refracted through what they mean for researchers here. Subscribe free and I will see you in the next one.

Subscribe now

Add your take · What is your take? · Tell me what I missed

Leave a comment

Claude Science: One Workbench, Not a Dozen Tabs

Santhosh Eapen — Tue, 07 Jul 2026 01:30:35 GMT

Four decades in crop protection, biotechnology, and bioinformatics have taught me one unglamorous truth: research is rarely slowed down by hard questions. It is slowed down by fragmentation. A day’s work means logging into PubMed for one thing, opening a Jupyter notebook for another, switching to R for a plot, then SSH-ing into a cluster to submit a job and waiting to see if it succeeded. Each tool is fine on its own. Stitched together, they cost hours that should have gone to science.

Anthropic’s new Claude Science - released in beta on June 30, 2026 - is built squarely around that problem. It is not a chatbot that talks about biology. It is a desktop workbench that runs your analyses, queries the databases behind them, and keeps a full, checkable record of how every result was produced. Let’s go through what it does, how to install it, and how researchers are actually using it.

What It Is

Claude Science is a desktop application, available in beta for macOS and Linux, that pairs Claude with a real analysis environment on your own computer. You describe a task in plain language - “cluster these single-cell samples and annotate the marker genes” - and Claude writes and runs Python, R, or shell code inside a sandbox, reads whichever folders you grant it, pulls data from more than 60 scientific databases through built-in connectors, and saves the results as versioned artifacts with a complete provenance record. A background reviewer agent independently checks Claude’s claims against what actually ran, flagging anything that doesn’t hold up.

Your files stay on your machine. Code runs in a sandbox that is deny-by-default on the network; you approve each new folder, host, or remote job before Claude touches it. Anthropic is direct about its limits: the reviewer reduces errors, it doesn’t eliminate them, and Claude Science is explicitly a research tool, not intended for clinical or diagnostic use.

Key Features

Reproducible artifacts - every figure ships with its code, execution log, and environment.
Native renderers - proteins, alignments, genomic tracks, chemical structures, no extra viewers.
The reviewer - flags unsupported claims, bad citations, mismatched DOIs.
Scales compute - persistent kernels locally; SSH to your HPC/SLURM cluster or Modal for GPUs on demand.
Domain-ready connectors - Ensembl, UniProt, PDB, AlphaFold, PubChem, ClinVar, and more, on by default.
Skills - AlphaFold2, ESMFold2, ProteinMPNN, Evo 2, scGPT, and others; you can add your own.

Installation

macOS: download installer from claude.com/product/claude-science, double-click.

Linux:

curl -fsSL https://claude.ai/install-claude-science.sh | bash

claude-science serve

Windows: no native build yet - run the Linux binary under WSL 2 (Ubuntu 24.04+).

Requires Pro, Max, Team, or Enterprise. Sign in with your Claude account - no API key needed.

Workflows

Pre-configured for single-cell RNA-seq, phylogenetic/evolutionary analysis, protein structure work, and cheminformatics. Open a project, point Claude at a folder, approve its permission requests, and results land as artifacts in the Files panel. For heavier runs, connect your lab’s HPC cluster or Modal account under Settings > Compute, and Claude drafts, submits, and retrieves the job.

In beta use: Manifold Bio for drug-target nomination, the Allen Institute for a 20-skill literature-review pipeline, and UCSF for germline variant studies - including catching a virus contaminant in RNA-seq data that had stumped the team for a year.

Why It Matters Here

For crop protection and plant bioinformatics work, the same fragmentation applies - just with different databases. A single environment that traces every figure back to its code is exactly the reproducibility standard grant reviewers and journal editors expect.

Anthropic is also running an AI for Science grants program - up to $30,000 in credits, applications open through July 15, 2026.

Getting Started Today

Claude Science is in beta on Pro, Max, Team, and Enterprise plans. Documentation covering installation, connectors, and admin setup lives at claude.com/docs/claude-science. Download it, sign in with your existing Claude account, and open the bundled Example project to get a feel for how permission cards and artifacts work before pointing it at your own data.

The most interesting thing about the tool, in my early impression, isn’t any single feature - it’s that provenance is not an afterthought bolted onto the output. It’s structural. That is exactly the standard research has always needed and rarely gotten from software built for speed alone.

When the Cell Ignores Its Own Genetic Code

Santhosh Eapen — Fri, 03 Jul 2026 14:30:59 GMT

We are taught the central dogma as a faithful copying process: DNA is transcribed into RNA, RNA is translated into protein, and the genetic code - the 64-codon dictionary that maps triplets of RNA letters to amino acids - is read with near-perfect fidelity. Deviations from that code have long been dismissed as rare errors, molecular noise to be cleaned up or ignored.

A new paper in Nature, published on 24 June 2026 by Shira Tsour, Nikolai Slavov and colleagues at the Parallel Squared Technology Institute, argues that this picture is incomplete in a consequential way. By mining deep proteomic, transcriptomic and genomic data from more than 1,000 human samples - spanning 6 cancer types and 26 healthy tissues - the team shows that cells deviate from the genetic code far more systematically than assumed. And the surprise is not just that it happens. It is that the “wrong” proteins are frequently the better ones: more stable, and in hundreds of cases more abundant than the versions the gene actually encodes.

The authors call the underlying process alternate RNA decoding. This edition of Research Rundown unpacks what they found, why the numbers are hard to wave away, and how the work reframes a process most textbooks still treat as a footnote.

Subscribe now

What “alternate decoding” actually means

Start with the vocabulary, because the distinction matters. A classic mutation changes the DNA itself - alter the gene, and every protein copied from it carries the change. Alternate RNA decoding is different. The gene is untouched. Instead, during translation, the ribosome inserts an amino acid that the codon does not specify - a substitution introduced at the protein level, not written into the genome.

This is sometimes called mistranslation, but that word smuggles in a verdict. The team’s central claim is that many of these events are not random slip-ups. When the genome is intact and the polymorphism databases cannot account for the change, yet the substituted protein keeps appearing across samples and tissues in a reproducible pattern, “error” becomes an unsatisfying explanation. Something is selecting for these alternate proteoforms.

To find them at scale, the researchers searched mass-spectrometry fragmentation spectra for peptides whose sequence departed from the genome-encoded reference, then cross-checked each candidate against matched DNA and RNA data to rule out ordinary genetic variants. The result is the first genuinely proteome-wide, multi-tissue map of where the cell goes off-script.

The scale of the off-script proteome

The headline numbers are what give the paper its weight. Across the dataset, the team identified 60,803 fragmentation spectra corresponding to 8,746 unique amino acid substitutions, in proteins derived from 1,767 genes - including 1,955 confidently localized sites, where the exact position of the swap could be pinned down.

This is not a handful of curiosities. It is a recurring feature of human proteomes, captured in tissue after tissue. Some substitutions showed up broadly, shared across many samples. Others were strikingly specific - switched on in particular tissue types or particular cancers, which is exactly the signature you would expect from a regulated process rather than uniform background noise.

The most provocative finding sits one layer deeper. For hundreds of proteins, the alternately translated version was more abundant than the canonical one encoded by the gene. That reversal is the paper’s intellectual centre of gravity. It is difficult to square with the idea of rare error and points instead toward what the authors term sense-codon recoding - the cell systematically reading certain codons as a different amino acid, and accumulating that product as the dominant form.

Why the “wrong” protein wins

If alternate proteoforms can out-number the originals, the obvious question is how. The team traces the abundance of a substituted protein to several converging mechanisms rather than a single cause.

The first is protein stability. A substituted residue can yield a proteoform that resists degradation better than its canonical counterpart, so even if it is made less often, it lingers longer and builds up - a turnover effect, not just a synthesis effect. The second is codon frequency: how often a given codon appears, and how its decoding interacts with the available tRNA pool. The third is mechanical - codon–anticodon mismatches, the near-miss pairings (a long-studied “two out of three” reading) that let a tRNA deliver the wrong amino acid. And the fourth is RNA modifications: chemical marks on the mRNA itself that can change how a codon is read.

The team also mapped where on a protein these swaps tend to land. Substitution ratios were positively associated with intrinsically disordered regions - the floppy, structurally permissive stretches of proteins that tolerate sequence changes without collapsing. In other words, the cell’s deviations cluster precisely where they can be absorbed, another hint that this is shaped, not chaotic.

Not noise, not mutation — and conserved across species

Two further results harden the case that alternate decoding is biology rather than artifact.

First, the genomics check. The substitution ratios were positively associated with genetic polymorphisms catalogued in the gnomAD database - but, crucially, those polymorphisms could not account for the substitutions themselves. The DNA-level variation correlates with where decoding flexes, yet the protein changes are not simply hidden mutations. This is the firewall against the most obvious skeptical reading: these are not just unannotated genetic variants masquerading as something new.

Second, and perhaps most persuasive, is conservation. The sequence, the relative abundance, and even the tissue specificity of alternately translated proteins were conserved between humans and mice. Random error does not reproduce itself faithfully across a roughly 90-million-year evolutionary gap. A process preserved between two mammals is, almost by definition, doing something the organism has reason to keep.

The functional roster makes the stakes concrete. Recoded proteins included transcription factors, proteases, signalling proteins, and proteins associated with neurodegeneration - not obscure bystanders, but the kinds of molecules whose precise sequence governs cell fate, regulation, and disease.

Where this sits in the wider field

The finding lands in a literature that has been circling this idea from several directions for decades, without assembling it at this scale.

The notion that mistranslation is mechanistically constrained, and even functionally relevant, has a long pedigree. Drummond and Wilke argued in Cell that mistranslation-induced misfolding is a dominant force shaping how coding sequences evolve, meaning the cell’s tolerance for translational error has left fingerprints all over our genomes. Netzer and colleagues showed in Nature that oxidative and immune stress can deliberately modify translational fidelity, making “errors” a tunable response rather than a fixed leak. And Mordret and colleagues, in Molecular Cell, built an early systematic proteomic method to detect amino acid substitutions and dissect the ribosome’s error modes, a direct methodological ancestor of the new work.

There is also a parallel story in RNA editing, where cells chemically rewrite their own transcripts. A Nature Communications survey mapped the landscape of adenosine-to-inosine recoding across human tissues, and the classic case of AMPA-receptor GluR2 editing in the brain (Wright and Vissel) shows how a single recoding event can be physiologically essential. Tsour and colleagues extend this spirit from a few known edits to a proteome-wide phenomenon detected directly at the protein level.

What ties it to the present moment is the technology. Predicting which amino acid changes matter has become a marquee problem - DeepMind’s AlphaMissense, in Science, scored the functional impact of missense variants proteome-wide. But every such tool assumes the protein sequence follows from the gene. This paper’s quiet provocation is that the protein sequence sometimes does not, and that the difference is measurable, reproducible, and disease-relevant.

The honest caveats

This is a discovery-scale, observational study built on existing public datasets, and it should be read as a map rather than a mechanism nailed shut.

Detecting single amino acid substitutions in mass-spectrometry data is technically demanding, and false positives are the field’s perennial worry; the authors lean on confident localization (the 1,955 pinned-down sites) and orthogonal genomic checks precisely because the analytical bar is high. The work establishes that alternate proteoforms exist, recur, and correlate with stability, codons, modifications and disordered regions - but the precise molecular machinery that decides when a codon is recoded, gene by gene, remains to be worked out. And “associated with neurodegeneration” is a flag for investigation, not yet a demonstrated cause. Whether any specific recoded protein drives a specific disease is the next experiment, not this one’s conclusion.

None of this blunts the core contribution. The central barriers that mattered - detecting substitutions proteome-wide, separating them from genetic variants, showing they are abundant, tissue-specific, and conserved - have been cleared with an unusually large and well-controlled dataset.

Why it matters

Step back and the implication reaches the foundations. For seventy years, the genetic code has been treated as a near-deterministic lookup table: gene in, predictable protein out. This work suggests the cell uses a looser, partly regulated dictionary - that the proteome is more diverse than the genome dictates, and that some of that diversity is conserved, tissue-tuned, and pointed at exactly the proteins that run regulation and disease.

The practical consequences are large. Every reference proteome, every genotype-to-phenotype prediction, every variant-effect model that assumes one gene yields one protein sequence is, on this account, missing a real layer of biology. If a drug target exists predominantly as a recoded proteoform in the tissue you care about, the “canonical” sequence you designed against may be the wrong molecule. And in cancer, where the team found tissue- and tumour-specific substitutions, alternate proteoforms could represent a previously invisible source of antigens, or vulnerabilities.

It reframes a deep assumption: the genome is the blueprint, but the cell is not an entirely literal builder.

What to watch next

Mechanism, gene by gene. The single most important follow-up is causal: pinning down which tRNAs, modifications, and codon contexts trigger recoding at specific sites. Watch for targeted experiments that move from correlation to control.
A disease test case. The neurodegeneration association is the headline-grabber. Look for the first study that shows a recoded proteoform actively contributing to a disease phenotype, not merely co-occurring with it.
Reference proteomes catch up. Whether databases and variant-effect predictors begin to incorporate alternate decoding will signal how seriously the field takes the claim. This is the infrastructure tell.
Cancer immunology angle. Tumour-specific substitutions raise the question of novel antigens. Keep an eye on whether immuno-oncology groups start mining recoded proteoforms as neoantigen candidates.

The Worm That Makes Medicine

Santhosh Eapen — Sat, 13 Jun 2026 12:31:57 GMT

For most of medical history, the hookworm has had exactly one reputation: unwanted guest. It hooks into the lining of the gut, sips blood, and-at scale-drives anaemia in nearly half a billion people worldwide. So it takes a certain scientific nerve to look at this animal and ask: what if we put it to work?

That is precisely what a team led by researchers at Washington University School of Medicine and James Cook University has now done. In a paper published in Nature Communications in June 2026, Singh and colleagues report something that sounds like science fiction but reads like careful molecular biology: they genetically engineered a human hookworm to manufacture and secrete a human antibody inside its host, continuously, from a single dose of infective larvae.

The antibody they chose, s16-HuScFv, neutralizes tetrodotoxin, the lethal pufferfish poison for which no commercial antidote exists. The toxin was a proof-of-concept target, not the point. The point is the platform. The researchers have built what is, in effect, a self-replicating, self-installing, slow-release biologic factory that lives in the gut for years. If biologics are the most powerful drugs we have but the hardest to deliver, this is a genuinely new answer to the delivery problem.

This edition of Research Rundown unpacks how they did it, why the numbers matter, and how the work sits alongside a fast-moving field of “living medicines” emerging in other journals.

Why biologics needed a new delivery idea

Protein- and peptide-based drugs - monoclonal antibodies, enzyme replacements, engineered binders - now treat everything from cancer to autoimmune disease. But they share a stubborn weakness. Swallow them and the gut digests them; so they are injected, often repeatedly, at high cost and with the compliance problems that dog any chronic therapy requiring frequent needles.

The field has been hunting for platforms that can produce biologics in situ, inside the patient, on a sustained schedule. The hookworm is an unlikely but oddly logical candidate. It has spent 350 million years co-evolving with vertebrate hosts, and it already runs an industrial-scale secretion operation: adult worms release more than 800 excretory/secretory proteins to feed, migrate, and quiet the host immune system. The audacious idea here is to hijack that secretory machinery and add one more protein to the export list, a human therapeutic.

Crucially, the groundwork already existed. Controlled human hookworm infections have been shown to be safe and well tolerated in clinical trials, and the worms are long-lived and cleared easily with a cheap oral dewormer when treatment ends. The biology was waiting; the missing piece was the genetic engineering.

The engineering: finding a safe place in the genome

The central technical problem was where to put the transgene. Parasitic nematodes have been notoriously resistant to genetic modification - their complex life cycles and tough cuticles defeat the microinjection tricks that work in the lab roundworm C. elegans. And dropping foreign DNA into a random spot risks either silencing the transgene or disrupting a gene the worm needs.

The team’s solution was to first map “genome safe harbours” (GSH) - loci where an insert sits among consistently, highly expressed genes in open, accessible chromatin, so the transgene stays switched on without collateral damage. Using multi-omics data across the Ancylostoma ceylanicum life cycle, they prioritized two such sites and settled on GSH2.

Into that harbour they used CRISPR/Cas9 with two overlapping guide RNAs to knock in a roughly 3.2-kb cassette: a CMV promoter, a signal peptide borrowed from the worm’s own abundant ASP-1 protein to route the antibody into the secretory pathway, the antibody coding sequence, a histidine tag, and a termination signal - all flanked by ~600-bp homology arms. Delivery into fragile hookworm eggs was the other hurdle; electroporation cleanly outperformed lipofection, and Oxford Nanopore sequencing confirmed the insert landed exactly where intended.

Proof it actually works - and that it breeds true

Two results elevate this from a clever construct to a credible platform.

First, heritability. The transgene was confirmed not only in the first-generation (F0) eggs and adult worms recovered from infected hamsters, but in the F1 generation of eggs as well. The engineered trait passed through the full, complicated parasitic life cycle and into the offspring - stable vertical transmission, something achieved before only in the more C. elegans-like Strongyloides worms.

Second, it didn’t break the worm. This is the quietly impressive part. Bulk RNA-sequencing comparing transgenic and wild-type worms found gene expression correlated at 0.998 on a natural scale, with only 6 differentially expressed genes out of 16,066 detected. The neighbours of the insertion site were untouched. The safe-harbour strategy worked: the worm carries its new payload without measurable disruption to its own biology or fitness.

Then the functional test. Hamsters infected with the transgenic larvae had circulating s16-HuScFv in their blood, and serum from those animals neutralized about 16.3% of tetrodotoxin in a cell-based assay (P = 9.1 × 10⁻⁵), while serum from wild-type-infected animals neutralized none. The antibody concentration was actually below the limit of the ELISA - yet the functional neutralization assay still detected real, statistically robust activity. Reassuringly, the worm’s own cocktail of secreted proteases did not chew up or inactivate the antibody in vitro, an essential check for any protein meant to survive in that environment.

Sixteen percent is partial, and the authors are candid about it. But for a first-generation, unoptimized system delivering a functional human antibody into the bloodstream from a parasite, it is a remarkable floor, not a ceiling.

Where this sits in the wider “living medicine” wave

What makes the hookworm result resonate is that it is not an isolated curiosity - it is the helminth entry in a broader movement toward organisms that manufacture drugs inside us.

The most direct parallels come from synthetic biology. In Cell Host & Microbe, Lynch and colleagues engineered commensal E. coli (a system they call PROTEcT) to secrete therapeutic nanobodies directly in the gut; a single dose of a strain secreting an anti-TNF nanobody prevented inflammation in a mouse model of colitis. More recently in Science Advances, Xie and colleagues turned engineered bacteria into an in-situ “antibody factory” that releases checkpoint-blocking antibodies inside tumours. The hookworm shares the same ambition as these designer microbes - sustained, local production - but trades a microbe that can be hard to retain in the gut for a long-lived animal that naturally colonises for years.

The helminth side of the story is older and clinically richer than many realize. Years of work established that worms are not merely pathogens but powerful immune modulators. Randomized trials of experimental hookworm infection in coeliac disease - Daveson and colleagues in PLoS One, and the larger 54-person trial by Croese and colleagues in Clinical and Translational Gastroenterology - tested whether the worms could restore gluten tolerance, with the larger study finding fewer gluten-related adverse events even though it did not fully restore tolerance. And in Science Translational Medicine, Navarro and colleagues isolated a single hookworm protein, AIP-2, that suppressed asthma in mice by inducing regulatory T cells. The new Nature Communications paper is the logical next step: from harvesting what worms naturally secrete to programming what they secrete.

The transgenesis itself also builds on a foundation. Shao, Li, and Lok showed heritable genetic transformation of Strongyloides stercoralis by injecting the male germline, an advance that helped pave the way for CRISPR in parasitic nematodes. And a companion review by Walther and colleagues in Advances in Parasitology lays out the broader vision: genetically modified helminths and schistosomes as pharmaceutical biofactories for inflammatory bowel disease, type 2 diabetes, coeliac disease, and arthritis.

The honest caveats

This is a first report, and the authors treat it as one. Neutralization was partial. Serum antibody levels sat below standard detection. The system needs optimization - stronger or inducible promoters, better signal peptides, codon tuning, longer antibody half-life - before the doses become clinically meaningful. Multi-generation stability will need long-term monitoring. And deliberately releasing a genetically modified human parasite raises real biocontainment questions; the team explicitly flags engineering suicide genes or inducible controls to prevent the modified organism spreading beyond the patient.

None of this dents the central achievement. The barriers that mattered - a safe insertion site, heritable transgenesis, secretion into host circulation, and proof the payload survives and works - have been cleared.

Why it matters

Step back and the implication is striking. A single oral or cutaneous dose of engineered larvae could, in principle, install a biologic factory that runs for years, calibrated by how many larvae you give, and switched off with a standard deworming pill. Because the worms live in the gut, they are ideally positioned for gastrointestinal and metabolic disease. Because the approach is disease-agnostic, the antibody payload is a swappable module - neutralize a toxin today, deliver an anti-inflammatory or an enzyme tomorrow.

It is a genuinely different way to think about chronic-disease therapy: not a drug you take, but a drug you grow.

What to watch next

Potency optimization. Track whether the group can push neutralization well beyond 16% with stronger promoters and tuned secretion, the single most important near-term metric.
A therapeutic, not a toxin. The proof-of-concept used an anti-toxin antibody. Watch for the first transgenic helminth secreting an anti-inflammatory biologic in an IBD or coeliac model, where the host biology already favours worms.
Containment engineering. Inducible “off-switches” and suicide genes will be the gate to any human trial; progress here signals seriousness about the clinic.
The synthetic-biology race. Keep one eye on engineered-bacteria platforms (PROTEcT and successors). Whoever solves durable, controllable in-situ production first - microbe or worm - defines the category.

A New Era in Dementia: What the Latest Science Means for You and the People You Love

Santhosh Eapen — Sat, 23 May 2026 23:35:59 GMT

Walk the corridors of any senior living community and you will sense it - the quiet grief of a family member who notices their mother struggling with names she once knew by heart, or the gentle patience of a caregiver guiding a resident through a morning routine he used to complete without a thought. Dementia does not merely erase memories. It slowly dismantles the scaffolding of a person’s independence, identity, and daily life.

But something is shifting in the science. A remarkable conversation published in JAMA on May 20, 2026 - a wide-ranging dialogue between JAMA Senior Editor Dr. Derek Angus and Dr. Gil Rabinovici, neurologist and Director of the UCSF Alzheimer’s Disease Research Center - lays out what is arguably the most hopeful picture of dementia research in decades. For those of us living alongside people at different stages of cognitive decline, this is not merely academic. It is directly relevant, quietly urgent, and in parts genuinely encouraging.

Research Rundown translates peer-reviewed science into plain language - for researchers, curious readers, and anyone navigating a world shaped by health decisions. If this kind of writing is useful to you, subscribe below.

Subscribe now

Nearly Half of All Dementia Cases May Be Preventable

That sentence deserves a pause.

Dr. Rabinovici points to the landmark 2024 Lancet Commission report, which identified a cluster of modifiable risk factors that together account for close to half of all dementia cases worldwide. These are not obscure molecular pathways - they are conditions and habits many of us already know to watch:

Poorly controlled high blood pressure and diabetes
Obesity and physical inactivity
Smoking and excessive alcohol use
Social isolation and depression
Chronic stress, lower educational attainment, and long-term exposure to air pollution

Some risk factors, of course, cannot be changed - age foremost among them, along with genetic inheritance. But the message from the science is clear: biology is not destiny. The choices we make, and the conditions in which we live, shape the brain’s trajectory over decades.

For anyone living in or caring for people in a senior community, this list carries both a caution and an invitation. Loneliness, in particular, is no small concern in residential settings. Social isolation is not just a quality-of-life issue - it is a measurable risk factor for cognitive decline.

If someone you know should read this - a family member, a caregiver, a colleague - the share button below does the work.

Lifestyle Interventions: Structured Support Makes a Difference

If the risk factors are partly behavioural, the interventions are too - and recent clinical evidence is encouraging. Dr. Rabinovici highlights the US POINTER trial, a large randomised clinical trial published in JAMA in 2025, which tested structured, multidomain lifestyle programmes against self-guided approaches in older adults at elevated risk for dementia.

The structured programme combined:

A heart-healthy diet (think Mediterranean or MIND-style eating patterns)
Regular physical exercise
Cognitive engagement - puzzles, learning, mental challenge
Social activity and meaningful connection
Careful management of vascular risk factors, particularly blood pressure

The result? The structured programme improved global cognitive function in participants. Not marginally, enough to be clinically meaningful.

This matters enormously in the context of senior living. Many communities already offer components of this cocktail - group fitness classes, hobby programmes, communal dining, social activities. The science now tells us these are not merely amenities. They are, if implemented with structure and consistency, potentially brain-protective.

The takeaway is not that individuals need to transform their lives overnight. It is that small, consistent, supported changes - especially when made in community - accumulate into real protection over time.

If you find this kind of evidence-to-practice translation useful, Research Rundown publishes regularly on science that matters beyond the lab.

A Blood Test That Can Detect Alzheimer’s Disease? It’s Here.

One of the most striking breakthroughs Dr. Rabinovici describes is the emergence of highly accurate blood-based biomarkers for Alzheimer’s disease.

Traditionally, confirming the presence of Alzheimer’s pathology in the brain required either a PET scan - expensive, specialist, and not universally available - or a lumbar puncture (spinal tap), invasive and uncomfortable. These tests detected the abnormal protein accumulations, amyloid plaques and tau tangles, that are the biological hallmark of Alzheimer’s disease.

Now, a new generation of blood tests can detect these same proteins circulating in the bloodstream. A 2026 meta-analysis in JAMA Neurology demonstrated that plasma phosphorylated tau 217 (p-tau217), one of the leading blood biomarkers, can accurately identify amyloid burden even in cognitively unimpaired older adults.

What this means in practice:

Earlier diagnosis - potentially years before symptoms become severe
A simpler, cheaper, less invasive pathway to understanding what is happening in the brain
The possibility of screening at-risk individuals and intervening sooner

We are not yet at the point where these tests are routine in every clinic or community health setting, but the trajectory is clear. The biology of dementia is becoming readable from a simple blood draw. That changes everything about when and how we can act.

New Drugs That Slow the Disease — Not Just the Symptoms

Perhaps the most groundbreaking development is the arrival of disease-modifying therapies for Alzheimer’s disease. For decades, available drugs could only blunt symptoms, offering modest help with memory and daily functioning without touching the underlying disease.

That has changed.

Two monoclonal antibody therapies - lecanemab and donanemab - now target the root biology of Alzheimer’s disease. They work by clearing amyloid protein plaques from the brain. The pivotal TRAILBLAZER-ALZ 2 trial, published in JAMA in 2023, demonstrated that donanemab slowed the rate of cognitive decline by approximately 30% in people with early symptomatic Alzheimer’s disease.

Thirty percent may not sound dramatic. But in a disease that robs people of their independence progressively and relentlessly, slowing that progression by nearly a third means more months - sometimes years - of meaningful independent living, recognising loved ones, making decisions, and being present in one’s own life.

These drugs are not cures. They carry risks, including a side effect called ARIA (amyloid-related imaging abnormalities), which requires monitoring. They work best - and perhaps only meaningfully - when given in the early stages of the disease, which is precisely why the new blood tests matter so much. Early detection enables early treatment.

Dr. Rabinovici is cautiously but genuinely optimistic: lecanemab and donanemab are, he believes, the opening chapter of a new frontier. More drugs, potentially with larger effects, are in development.

A Note of Measured Hope

Dr. Rabinovici’s conversation with Dr. Angus does not promise a cure. It does not erase the grief that dementia carries. But it maps a landscape that looks meaningfully different from where we stood even five years ago: new tools for early detection, new evidence for prevention, and the first drugs in history to alter the course of the disease itself.

For those of us living in communities where dementia is not an abstraction but a daily reality - a neighbour’s confusion, a friend’s fading recall, a fellow resident’s quiet withdrawal - this science arrives as something worth holding onto.

The era of watching helplessly is not entirely over. But it is ending.

The science in this piece is moving fast, and most people haven't heard it yet. If you found it useful, sharing it is the most direct way to change that.

Share 'A New Era in Dementia'

Baker LD, Espeland MA, Whitmer RA, et al. 2025. Structured vs self-guided multidomain lifestyle interventions for global cognitive function: the US POINTER randomized clinical trial. JAMA. 334(8):681–691. doi:10.1001/jama.2025.12923
Kim B, Blam K, Elser H, et al. 2025. Ambient air pollution and the severity of Alzheimer disease neuropathology. JAMA Neurol. 82(11):1153–1161. doi:10.1001/jamaneurol.2025.3316
Livingston G, Huntley J, Liu KY, et al. 2024. Dementia prevention, intervention, and care: 2024 report of the Lancet standing commission. Lancet. 404(10452):572–628. doi:10.1016/S0140-6736(24)01296-0
Malek-Ahmadi M, Sharma S, Stipho F, et al. 2026. Plasma phosphorylated tau 217 and amyloid burden in older adults without cognitive impairment: a meta-analysis. JAMA Neurol. 83(1):13–19. doi:10.1001/jamaneurol.2025.4721
Reuben DB, Kremen S, Maust DT. 2024. Dementia prevention and treatment: a narrative review. JAMA Intern Med. 184(5):563–572. doi:10.1001/jamainternmed.2023.8522

If you live or work close to this - as a resident, a caregiver, a clinician, or a family member - I'd genuinely like to hear what you make of the science. The comment section is open.

Leave a comment

Dark Proteome, Revealed: How Microproteins and Peptideins Are Rewriting the Human Protein Catalogue

Santhosh Eapen — Sat, 09 May 2026 04:30:31 GMT

A landmark multi-consortium study published in Nature (May 6, 2026) has mapped the protein-level evidence for 7,264 non-canonical open reading frames (ncORFs) in the human genome - the so-called “dark proteome.” The team identified 484 microproteins with high-confidence detection, coined a new biological entity called the “peptidein”, developed a novel evolutionary constraint tool called ORBL, and pinpointed a pan-essential microprotein in a long non-coding RNA transcript with direct implications for cancer biology. This is not an incremental update to protein databases. It is a structural rethink of how we define protein-coding genes.

Why This Matters - Even If You Work on Plants

If you have spent any time in proteomics, transcriptomics or functional genomics in the last decade, you will have heard the claim that the human genome encodes approximately 19,500 protein-coding genes. That number has been treated as essentially settled. But evidence has been quietly accumulating that thousands of small open reading frames - sitting within long non-coding RNAs, upstream of known genes, embedded in untranslated regions, or nestled in intergenic space - are actually translated into real protein molecules. They have been called microproteins, small ORF-encoded peptides (SEPs), or smORF-encoded polypeptides.

What we did not have, until now, was a systematic, annotation-quality framework to evaluate which of these are genuinely protein-coding, which are biologically relevant but of uncertain functional status, and which remain speculative. The TransCODE Consortium - a collaboration across more than 30 institutions spanning Europe, North America, Asia and Australia - has delivered precisely that framework, and the results are both impressive and appropriately cautious.

This development is highly relevant to crop biology and agricultural bioinformatics, as the resulting annotation standards will inform tools applied to complex plant genomes like Piper nigrum, or Coffea arabica.

The Scale of the Effort

To appreciate this study, start with the numbers. The consortium analysed:

7,264 ncORFs supported by GENCODE’s Ribo-seq data
95,520 proteomics experiments drawn from the PeptideAtlas repository
3.5 billion mass spectrometry (MS/MS) spectra from non-HLA samples
240 million MS spectra from HLA immunopeptidomics datasets
Data from PeptideAtlas, GENCODE, HUPO-HPP and the HIPP immunopeptidomics project

The analytical framework set a stringent false discovery rate (FDR) of <0.1% at the protein level, requiring two distinct, uniquely mapping peptides of at least 9 amino acids and a minimum coverage of 18 residues, the HUPO-HPP standard for canonical proteins. This is not a permissive search. These are the same verification standards applied to known human proteins.

From the 7,264 ncORFs queried, 484 ncORFs (~2.5%) passed detection thresholds mapping to 183 microproteins in non-HLA data. In the HLA immunopeptidome, a far richer hunting ground for short peptides, 3,116 peptides were detected mapping to 1,785 out of 7,264 ncORFs (24.6%), with 94.3% found exclusively on HLA class-I molecules.

The Annotation Framework: Tiers, Not Binary Calls

One of the most intellectually honest contributions of this paper is the decision not to force a binary classification. Rather than asking “is this a protein or not?”, the consortium built a tiered annotation system that mirrors the graded certainty with which evidence accumulates in any annotation project.

Through this pipeline, the team identified 20 Tier 1A candidates, which after rigorous manual inspection were reduced to 15 candidates for potential GENCODE annotation as protein-coding genes. Of these, three have already been annotated by GENCODE: c12norep105 in CYP27B1, c21norep46 in ERVH48-1, and c11riboseqorf4 in PIDD1 — the latter being a 171-amino-acid uORF peptide with evidence of physiological expression in non-malignant tissues, cancer samples and cell lines.

A fourth candidate, c2riboseqorf47 (a uORF within the GMCL1 gene), was so well-supported - a high ORBL evolutionary constraint score, positive PhyloCSF score, HLA-I and HLA-II peptide evidence, and a loss-of-function CRISPR phenotype - that GENCODE has now annotated it as a protein-coding gene: ENSG00000310604. This is annotation in action: evidence accumulates, classification follows.

The New Entity: What Exactly Is a “Peptidein”?

The paper introduces “peptidein”—a portmanteau of peptide and protein—as a new biological category. A peptidein is defined as an ORF with confirmed RNA translation and protein synthesis, but currently lacks sufficient data for conventional protein-coding gene status. This acts as a probationary state, acknowledging its existence without claiming function, thus creating a clear research goal for “graduating” it to a protein-coding gene.

The paper identifies three subtypes of peptideins:

Tier 1B - high-confidence HLA evidence confirming protein synthesis at these sites
Tier 2B - HLA evidence, but candidates too short to generate multiple tryptic peptides
Tier 2A - one tryptic MS peptide supporting the microprotein; may capture candidates that short-circuit standard detection

In total, 12 peptideins were formally annotated in this study, with 34 further detected ncORFs and the full annotation guidelines promised as a forthcoming resource.

Whether peptidein proves as durable a term as “pseudogene” or “enhancer RNA” remains to be seen. But the need for a category between “protein” and “translated-but-of-unknown-consequence” is real, and the community has been circling this problem for years.

Cracking Evolutionary Constraint: The ORBL Approach

A persistent challenge in evaluating ncORFs is that conventional evolutionary constraint metrics, comparing amino acid sequence conservation across species, do not work well for very short proteins. If a microprotein is only 50 amino acids long, its evolutionary signal drowns in noise.

The consortium developed ORBL: ORF Relative Branch Length, a method that quantifies the evolutionary conservation of the ORF structure itself (start, stop codons, and reading frame openness) across species, regardless of amino acid changes. The score (ORBLv) is the fraction of total branch length where the ORF structure is conserved. To adjust for chance, they use a comparative score (ORBLq): the quantile among untranslated ORFs of similar length.

Results show 30.4% of 7,264 ncORFs (2,211) exhibit evolutionary constraint (ORBLq > 0.9), significantly higher than the 10% expected under neutral evolution (P < 2.3 × 10⁻³⁰). This constraint is particularly high in uORFs and doORFs, suggesting regulatory roles. Crucially, ncORFs detected by HLA-I immunopeptidomics have significantly higher ORBLq scores, confirming ORBL as a meaningful predictor of protein synthesis.

ORBL is now publicly available for annotating non-canonical transcripts across various genomes.

A Pan-Essential Peptidein: The OLMALINC Story

Among the 51 ncORFs exhibiting a pan-essential knockout signature from CRISPR screening across 8 human cell lines, one attracted particular attention: c10riboseqorf92, a 123-amino-acid sequence encoded within the OLMALINC long non-coding RNA transcript (LINCO0263).

This ncORF scored as pan-essential across 415 of 485 cancer cell lines in the Dependency Map (DepMap) dataset. Its knockout phenotype was confirmed to be ORF-specific (not a consequence of disrupting the adjacent OLMALINC transcript), and transcriptome profiling after OLMALINC knockdown versus c10riboseqorf92 coding sequence expression revealed 513 genes of increased abundance and 456 of decreased abundance - with differentially regulated genes pointing toward hypoxia, glycolysis, angiogenesis, DNA damage response, UV light exposure and TNF signalling pathways.

Single-cell RNA-seq after c10riboseqorf92 knockout confirmed induction of mitotic and chromosome-related processes and downregulation of translation and metabolism pathways across 12 cell lines.

This is a peptidein - not yet classified as a protein-coding gene, because its evidence remains restricted to transformed and cancer cell lines - but one whose biology is already rich enough to warrant serious investigation. The paper’s cautious annotation reflects genuine scientific rigour, not lack of interest.

Research Agenda: Seven Questions the Field Must Address

The paper’s final section, produced through consensus across the multi-consortium group, lays out seven research challenges the ncORF field must engage:

Are HUPO-HPP guidelines (requiring ≥2 peptides of ≥9 amino acids covering ≥18 residues) appropriate for ncORF-encoded microproteins that are frequently below 50 amino acids in length? 28.3% of the 7,264 ncORFs studied are under 25 amino acids.
Should HLA immunopeptidomics - where 1,785 ncORFs showed peptide-level evidence - be used as protein-coding gene evidence? The immune system presents what cells translate; that is a form of functional evidence.
Should peptides detected only in cancer samples or immortalised cell lines support annotation as protein-coding genes?
What role should evolutionary inference play, given the limitations of amino acid conservation for short sequences?
Which alternative experimental approaches (beyond tryptic MS) could support protein-coding gene annotation?
How should microproteins for which function cannot be demonstrated or inferred be classified?
Should deep learning approaches inform gene or protein annotation in high-throughput settings?

These are not rhetorical questions. They will require formal community decisions from GENCODE, UniProtKB, RefSeq and HUPO-HPP. The answers will reshape how genomic databases handle tens of thousands of currently unannotated translation products - in humans, and eventually in all other organisms.

Key Takeaways for Researchers

Update your database assumptions. The human proteome is not ~19,500 proteins. The real number is larger, and the TransCODE Consortium has now begun defining what the expanded catalogue looks like and how to evaluate evidence for its members.
The peptidein concept is a tool, not a final answer. Use it as a prompt to ask: what experiment would demonstrate function? HLA presentation is evidence of existence, not function. CRISPR essentiality is closer to function but requires context. Annotation requires both.
ORBL is worth incorporating into your ORF annotation workflows. Whether you are working on plant transcriptomes, microbial genomes, or fungal proteomes, the concept of ORFness conservation as an evolutionary signal - distinct from amino acid conservation - is portable and powerful.
The immunopeptidomics window is real. 24.6% of 7,264 ncORFs were detectable in HLA data - far higher than in conventional proteomics. For any organism or tissue type where immunopeptidomics is feasible, this represents an underutilised discovery modality.
Pan-essential microproteins exist. At least 51 ncORFs show essentiality signatures in human cancer cell lines. Drug target space is not exhausted.

Additional Reading

Chothani, S., et al. (2024). An expanded reference catalog of translated open reading frames for biomedical research. Nucleic Acids Res., 54, gkag234 (2026).

Deutsch, E. W. (2026). The 2025 Report on the Human Proteome from the HUPO Human Proteome Project. J. Proteome Res., 25, 539–555.

Deutsch, E. W., Kok, L. W., Mudge, J. M., Valls, C. F., Jungreis, I., & van Heesch, S. (2026). Expanding the human proteome with microproteins and peptideins. Nature. https://doi.org/10.1038/s41586-026-10459-x

Nature News. (2026, May). Revealed: the mysterious ‘dark’ proteins that might play a big role in biology. Nature. https://doi.org/10.1038/d41586-026-01492-x

Whited, A. M., et al. (2024). Biophysical characterization of high-confidence, small human proteins. Biophys. Rep., 4, 100167.

Research Rundown is curated for researchers who want the signal without the noise. If this issue was useful, share it to a colleague — and consider what your own datasets might be hiding in the dark.

16,000 Ancient Genomes: The Data That Proved Human Evolution Never Took a Break

Santhosh Eapen — Sat, 25 Apr 2026 05:15:05 GMT

For the past decade, the consensus held that modern humans had largely stopped evolving, with culture replacing biology as the main adaptive force, as suggested by a 2011 analysis.

However, a new Nature study, Ancient DNA reveals pervasive directional selection across West Eurasia, by Akbari, Reich and colleagues (covered in Science 15 April 2026), challenges this. The team sequenced and re-analysed 15,836 ancient human genomes, the largest such time-series analysis.

Their finding: 479 genetic variants changed significantly in the 10,000 years since farming began in West Eurasia. A second surge, focused on immunity, occurred during the Bronze Age (~5,000 years ago) as people lived in closer proximity to domesticated animals.

Reich states, “the genome is alive with signal.” The signal is real, detectable due to the dataset’s scale. While the press will simplify this as “humans are still evolving,” the paper is more nuanced.

What the Paper Actually Demonstrates

Earlier claims that human evolution stopped failed due to a lack of ancient DNA (aDNA). Examining only living genomes confuses selection with random drift. aDNA, sourced from dated remains, provides a necessary time series to track allele frequency changes over millennia, moving beyond a single snapshot.

Initially, aDNA samples were too small to distinguish genuine selection from large-scale ancestry replacements, like the migration of Anatolian farmers or Yamnaya steppe herders.

However, Reich’s lab overcame this by analyzing nearly 16,000 genomes, a scale ten times larger than prior studies. This massive dataset allowed them to detect hundreds of selected variants, not just a dozen. As Alexander Young notes, this scale is now powerful enough “to powerfully interrogate selection over the last 10,000 years.”

Fig. 1 | Two waves, 10,000 years apart. The Neolithic farming transition (~10,000 BCE) drove shifts in metabolism and body composition; the Bronze Age (~3,000 BCE) triggered the sharpest surge — immunity genes under pressure from dense populations and domesticated animals. The genome was writing both times.

The findings strongly correlate with the archaeological record. For example, variants for higher body fat/mass decreased after farming began, consistent with a shift from food-scarce to grain-surplus life. Tuberculosis-resistance variants rose, then fell, reflecting changes in pathogen pressure. The clearest signal is the Bronze Age immunity-gene spike, driven by higher population density, livestock contact, and new pathogens.

Fig. 2 | When livestock changed everything. As Bronze Age settlements packed humans and animals together, novel pathogens crossed species barriers — and the human immune genome responded. Five thousand years later, that signal is still readable in ancient bone.

Multiple sclerosis risk variants are now seen within this coherent pattern. These are strong, data-driven claims, consistent across independent variants, validating the paper’s importance.

The Over-Extensions the Field Must Not Let Stand

The paper reports measurable shifts in polygenic scores for modern traits like walking pace, income, and years of schooling over the last 5,000 years, a finding the authors present cautiously but which will likely be sensationalized.

A polygenic score for “years of schooling” in 2026 is a modern proxy capturing a complex phenotype shaped by institutions. Projecting it onto the Neolithic period doesn’t mean Bronze Age farmers were selected for “education,” but rather that an upstream, unknown selected trait correlated with today’s outcome has shifted frequency.

The core issue is reporting selection signals with demographic rigor. As Arbel Harpak notes, the study offers “amazing data and provocative hypotheses” requiring scrutiny, especially since population shifts (like the Steppe replacement in Europe) also move allele frequencies, making disentangling true selection from demographic reshuffling difficult. The authors acknowledge they are only “getting closer.”

The scientific community must emphasize these caveats as strongly as the findings, given the potential misuse of claims like “selection on income-associated variants.”

The Gap in the Sample No One Wants to Lead With

Fig. 3 | A map of what we know — and what we're missing. Fifteen thousand ancient genomes, almost entirely from West Eurasia. South Asia, Southeast Asia, and sub-Saharan Africa remain genomically dark. A study of "human evolution" that excludes most of humanity is a starting point, not a conclusion.

The Nature paper on “pervasive directional selection across West Eurasia,” despite its 15,836 ancient genomes, is geographically limited to West Eurasia (Europe and adjacent West Asia) and relies heavily on UK Biobank data. Its inferences on farming, pathogens, and selection are thus confined to the West Eurasian gene pool.

This focus is a critical gap, as agricultural origins and human coevolutionary histories outside West Eurasia are excluded, particularly in regions like South and Southeast Asia, where crops like rice, pepper, and others were domesticated.

For a country like India, with ancient agricultural traditions and immense genetic diversity, this is a serious oversight. Although India has skeletal collections and growing ancient-DNA capacity (CCMB, BHU), it lacks a systematic, large-scale ancient-DNA program to address questions about its own population history comparable to the research done in Europe.

Fig. 4 | The archive exists. The program does not. India has the bones, the biodiversity, and the sequencing infrastructure. What is missing is the institutional will to ask the questions. The tools are no longer the barrier.

This is not a critique of the paper itself, but a challenge to Indian institutions. The tools and low-cost sequencing exist; the question is whether Indian scientific instiutions will integrate with ancient DNA studies or remain marginalized.

What This Means for Crop Science

Fig. 5 | The same question, a different kingdom. Reich’s methods — tracking allele frequency shifts across time — work equally well for domesticated plants. Pepper, cardamom, turmeric, and coconut are waiting for someone to apply them. Their wild relatives are still alive. That advantage won’t last indefinitely.

For those familiar with plant genomics, the Reich paper reiterates two known points. First, scale is transformative: questions insoluble with small samples become tractable with large ones, making subtle signals undeniable. This shift has already revolutionized crop domestication genetics and is now evident in human ancient DNA studies. Second, the analytical methods used to detect human selection across 10,000 years are directly applicable to crops. While major staples have ancient-DNA coverage, spice and plantation crops like cardamom, pepper, turmeric, and coconut are under-characterized. Reich’s computational frameworks will translate directly to plant populations, with the advantage that we can still study many wild relatives alongside ancient DNA. This presents a clear opportunity for Indian agricultural and plantation-crop research institutes, if they choose to pursue it.

Where Does This Leave Ancient-DNA Research?

Reich admits the current study is limited. His real goal is to study the period between 1,800,000 and 300,000 years ago, when hominin brains tripled, but the necessary data (better DNA preservation, sequencing, and uncollected fossils) is currently unavailable.

The paper is a credible, scale-driven demonstration that natural selection continued after agriculture began, framed with appropriate caution. It’s the starting point for the conversation on human evolution, not the end.

The true test is not whether the 479 variants replicate, which they likely will. Instead, it is whether the community can accept that the genome is dynamic while refusing to allow polygenic scores (like for “years of schooling”) to be mistakenly equated with Neolithic education.

Key Takeaways from the West Eurasian Ancient Genomes Study

Evolution is Ongoing: Using 16,000 ancient genomes proved that human evolution is still happening, a finding only made possible by the scale of the data.
Bronze Age Immunity Surge: The clearest finding is the sharp increase in immunity genes during the Bronze Age, driven by greater population density and contact with new pathogens from domesticated animals.
Caution on Behavioral Traits: Claims about selection for behavioral traits must be treated with caution. The polygenic scores used are based on modern metrics (like income/education), and the ancient selective pressures for these gene variants are unknown.
Migration vs. Selection: Distinguishing natural selection from ancestry replacement (like the two major migrations in Europe) remains the biggest challenge. The study progresses, but doesn’t fully resolve, this confound.
Regional Focus Only: The study is limited to West Eurasia. Broad statements about “human evolution” are premature and incomplete until similar-scale sequencing is achieved for other populations (e.g., India, Africa, East Asia).
Framework for Research: The analytical methods (tracking allele frequencies, separating demography from selection) are immediately applicable and valuable for plant/crop domestication research.

The best time to begin was a decade ago. The second best time is today.

If this was useful, share it with a colleague who should be reading it.

If someone forwarded this to you and you’d like to receive Research Rundown regularly,

Subscribe now

Have a question or a pushback on something I wrote? Leave a comment — I read every one.

Leave a comment

The Body Decoded: Mapping Life, Cell by Cell

Santhosh Eapen — Fri, 10 Apr 2026 23:30:22 GMT

Introduction

Imagine holding a map so precise that it tells you not just which organ you are looking at, but exactly which type of cell occupies every square micrometre of that organ - and what each of those cells is saying at the molecular level, at the very moment the tissue was sampled. That map now exists. And it covers an entire mammalian body.

Published in Cell this week (Clevenger et al., 2026), the study titled “Whole-body molecular and cellular mapping of the laboratory mouse” achieves what the field of spatial transcriptomics has been inching toward for a decade: a comprehensive, genome-wide molecular and cellular atlas of the adult laboratory mouse, with every major organ system captured in its native anatomical context. The work drew immediate recognition from the broader scientific community, earning a dedicated perspective commentary in Science (doi: 10.1126/science.zgr6u94) - a rare honour that signals just how transformative this resource is considered to be.

For the AAAS, the journal that publishes Science, this was a milestone moment in biology. For researchers in life sciences - whether in biomedicine, crop protection, or bioinformatics - the tools, AI pipelines, and conceptual frameworks this study unleashes deserve close attention. What was done to a mouse today will be done to plants, insects, and pathogens tomorrow.

The Problem This Study Was Built to Solve

The laboratory mouse (Mus musculus) has anchored biomedical discovery for over a century. Yet, for all our reliance on it, a fundamental gap persisted: we had no way to simultaneously measure gene expression across every tissue and cell type of the whole adult animal body in its native anatomical context.

Previous spatial transcriptomic platforms — Visium, Slide-seq, MERFISH — worked beautifully on small tissue sections, typically a centimetre or less in area. Scaling to a whole-body cross-section of an adult mouse (roughly 2 cm × 6 cm) introduced three compounding problems that no single team had cracked:

The surface area for spatial profiling far exceeded what existing commercial slides could handle.
Preserving RNA integrity while making ultra-thin histological sections of an intact frozen adult animal requires specialised equipment and optimised protocols not previously described for this purpose.
No computational tools existed to automatically annotate the hundreds of distinct organs, tissue subregions, and cell types that appear simultaneously on a body-wide section.

The Science commentary explicitly highlights these barriers and frames the Cell paper’s success in overcoming all three as the reason the work constitutes such a genuine step-change — not an incremental advance, but a new category of experiment.

The Experimental Platform: Array-seq on Whole-Mouse Sections

The team, led by Nicolas Chevrier and Feng Bao at the University of Chicago and Fudan University respectively, used Array-seq - a spatial transcriptomics platform previously developed in the Chevrier lab - and adapted it for whole-body profiling.

Six-week-old C57BL/6J mice - the most widely used laboratory strain - were carefully sectioned, stained with haematoxylin and eosin (H&E), and profiled across two biological replicates. The resulting slides captured between 588,299 and 610,689 spots per section (each spot sampling a local cluster of cells), detecting a mean of 662 genes and 1,241 unique molecular identifiers per spot - impressive sensitivity for such large-scale spatial data.

Workflow Infographic The Array-seq whole-body pipeline in four steps. Cryo-sectioning → H&E staining → Array-seq capture (~600,000 spots) → computational cell-type assignment against a 59-million-cell reference, yielding a body-wide molecular map of 379 cell types. Source: Clevenger et al., Cell 189, 2026.

The sections accurately captured the tissue architectures of 16 key murine organ systems, including:

Brain — with subregional resolution to individual cortical and cerebellar layers, including the molecular layer, granular layer, and fiber tracts
Liver — periportal vs. pericentral hepatocyte zones resolved by expression of Glu1, Sds, and Cyp2a5
Small intestine — proximal epithelium vs. muscularis mucosa distinguished by Krt20, Oat, and Defa30
Kidney, spleen, thymus, lung, pancreas, skin, bone marrow, heart, and additional structures

Spatial expression of canonical marker genes confirmed the anatomical fidelity throughout. The Purkinje cell layer of the cerebellum, the chief cell-rich and parietal cell-rich regions of stomach gastric glands, α/β/δ islet cells in the pancreas - all resolved with expected marker gene patterns. Two biological replicates showed strong correlation (Pearson’s coefficient 0.968 ± 0.013), affirming the platform’s reproducibility.

A Reference Universe of 59 Million Cells

To assign cell types to each spatial spot, the team built what is arguably the most comprehensive mouse single-cell reference dataset assembled to date: 59 million cells drawn from 466 published single-cell RNA sequencing studies, covering all 16 major organ systems. From this, they defined 495 granular cell types organised into 79 broad and 30 core categories.

Using this reference with a scalable computational cell-type assignment framework, they spatially identified 379 cell types across the ~5 million nuclei captured in their whole-mouse sections. Key findings from the body-wide cell-type map include:

Skeletal muscle cells were the most abundant, accounting for over 32% of all spots across both replicates.
Hepatocytes (8.8%), gut epithelial cells (7.8%), and bone marrow haematopoietic stem cells (5.5%) were all faithfully localised.
Rare populations - osteocytes in the skull and spine, Purkinje cells in the cerebellum, α, β, and δ cells in the pancreatic islets - were accurately detected and assigned to their expected anatomical niches.
All 27 T cell subsets present in their reference were detected across the two sections, alongside 48 total broad cell types spanning 16 organs.
The method detected 67–69% of all reference granular cell types - remarkable coverage for a single two-dimensional tissue section.

The concordance was benchmarked against three independent published methods for spatial cell-type deconvolution: robust cell type decomposition (RCTD), cell2location, and CARD. All three yielded concordant results, with mean concordance scores of 93.7–94.1% for RCTD and CARD respectively, validating the approach against the state of the art.

LABEL: AI That Reads H&E Images Like a Trained Pathologist

Perhaps the most immediately translatable innovation in this paper is the LABEL algorithm - a machine learning pipeline designed to annotate organs, tissue regions, and cell types directly from standard H&E-stained histology images, without the need for sequencing.

LABEL works in two stages:

It uses UNI, a pre-trained pathology foundation model, for unsupervised feature extraction from H&E image patches across the whole-mouse section.
It applies a spatially-aware K-nearest neighbours (KNN) classifier - trained on the Array-seq spatial transcriptomic annotations - to transfer molecular cell-type labels onto the image features.

The performance data are notable. Using an “in-distribution” training approach (80% of data for training, 20% for testing), LABEL achieved:

Mean organ-level accuracy of 0.91 across 16 organ systems
Above 0.95 for brain, liver, and kidney
0.73 mean accuracy for tissue subregion annotation across 33 unique subregions

Even under the far more stringent “leave-one-out” approach - where an entire mouse section is excluded from training and used for testing - LABEL achieved 0.71 mean organ accuracy, with brain, liver, small intestine, thymus, skin, and muscle all exceeding 0.80.

LABEL AI Algorithm Infographic How LABEL converts an H&E slide into a molecularly annotated map. H&E image patches are processed through the UNI pathology foundation model, then classified by a spatially-aware KNN algorithm to assign organ and cell-type identities — no sequencing required. Mean accuracy: 0.91. Source: Clevenger et al., Cell 189, 2026.

At the cell-type level, LABEL correctly annotated hepatocytes (0.98), renal epithelial cells (0.94), and skeletal muscle cells (0.90) with high accuracy. The broader implication is transformative: a cheap, widely available H&E slide becomes a molecularly annotated spatial map. Institutions without access to expensive spatial transcriptomics infrastructure can, in principle, use LABEL to infer cellular composition from histology images alone, a direct democratisation of molecular pathology.

The Science commentary specifically highlights LABEL’s design - the combination of a pre-trained vision foundation model with spatially-aware classification - as an exemplary use of AI to solve a classical problem in histopathology at a scale not previously possible.

The Disease Application: Mapping Systemic Inflammation Body-Wide

The team demonstrated the platform’s disease-profiling power through a model of endotoxemia, systemic inflammation induced by injection of lipopolysaccharide (LPS), mimicking bacterial sepsis. Mice received a sublethal LPS dose; whole-body sections were collected 12 hours later alongside untreated controls. The whole-body sections showed a massive response across 37 tissue subregions/16 organs:

5,143 differentially expressed genes.
Canonical inflammation biomarkers (Saa1, Saa2, Saa3) exhibited expected, precise, tissue-specific patterns (e.g., Saa1 dominant in liver), confirming and integrating decades of isolated-tissue studies.
Chemokine expression was strikingly organ-specific (e.g., Ccl12 peaked in brain/thymus).
The interferon-stimulated gene Irgm1 was selectively elevated in kidney, liver, lung, and heart, potentially explaining organ-specific cytoprotection.
Ligand-receptor signalling was systematically disrupted across organs, reshaping vascular adhesion, chemokine, cytokine, and matrix remodelling pathways, a systems-level view impossible via isolated sampling.
LPS increased macrophage proportions (lung, liver, spleen, kidney) and triggered epithelial/endothelial remodelling.
Concurrently, anti-inflammatory ligand-receptor pathways were upregulated, showing a simultaneous immunoregulatory counter-response.

Body-Wide Inflammation Heatmap LPS endotoxemia triggers a body-wide transcriptomic response. Compared to controls (left), LPS-treated mice (right) show differential expression of 5,143 genes across 37 tissue subregions and 16 organs, with the liver, kidney, and lung as primary inflammatory hubs. Source: Clevenger et al., Cell 189, 2026; Commentary: Science, doi: 10.1126/science.zgr6u94.

This demonstrates that a single disease leaves a body-wide molecular fingerprint, which the platform’s whole-body map - the most detailed picture of inflammatory gene expression across a mammal to date - can fully decode, proving its utility in disease biology.

What This Paper Means for Researchers Beyond Biomedicine

The Cell study, centered on the lab mouse and biomedical disease, establishes frameworks with broad applicability.

For bioinformaticians, the study’s computational pipelines, especially the 59M-cell database, LABEL foundation model, and deconvolution benchmarking, set new standards for spatial data integration. The UNI pre-trained vision model with KNN transfer offers an elegant, replicable solution for general annotation.

For agricultural and plant scientists, the parallels are direct. Spatial transcriptomics in plant tissues faces the same challenge of mapping genes across complex architectures while retaining spatial context. The study’s whole-organism sectioning, reference-based cell-type assignment, and AI-assisted histology methods are all applicable to plant and crop biology workflows, especially with platforms like Stereo-seq already generating plant spatial atlases.

For those studying host-pathogen interactions, the endotoxemia application is a direct proof of concept. Whole-body spatial transcriptomics under pathogen challenge offers a new way to understand systemic resistance, susceptibility, and the tissue geography of immune signaling, with the methodology easily extending from mouse endotoxemia to crop disease response (e.g., pepper plant’s response to Phytophthora capsici).

Limitations and What Comes Next

The authors are candid about current constraints. Brown fat and bone marrow showed lower LABEL classification accuracy, attributed to insufficient representation in training data. Some tissue subregions - lung bronchioles, splenic white pulp, brain ventricles - remain harder to classify at fine resolution. Two-dimensional sections, however detailed, do not capture the full three-dimensional architecture of the body; stacking serial sections for volumetric reconstruction is the logical next step, and the platform is explicitly positioned as a foundation for exactly that.

The dataset, models, and code have been made publicly available, inviting the community to extend and apply them. The mouse body atlas, like the human genome before it, is not an endpoint but a reference infrastructure - one that will be interrogated, expanded, and redeployed across many biological systems in the years ahead.

Key Takeaways

Five things worth carrying from this week’s reading:

Whole-body spatial transcriptomics is now technically feasible for adult mammals, opening a new tier of biological observation that bridges single-cell resolution with organism-level systems understanding.
LABEL’s use of a pathology foundation model for H&E annotation is a practical AI innovation with immediate translational value - any lab with H&E slides and computational infrastructure can build on this approach.
Systemic disease leaves a precise, body-wide molecular signature that only whole-organism profiling can fully decode; the endotoxemia map is the most detailed inflammatory atlas of an intact mammal ever produced.
The 59M-cell reference dataset is a community resource as valuable as the spatial atlas itself, a new baseline for mouse cell-type annotation in any tissue context.

The frameworks are discipline-agnostic. Researchers in plant biology, entomology, and crop protection should be watching these methodological developments closely: the pipelines are coming to your system within the decade, and engaging with them now pays dividends later.

Additional Reading

Chen, A., Luo, Z., Xia, Y., & the BGI Spatial Transcriptomics Consortium. (2022). Spatiotemporal transcriptomic atlas of mouse organogenesis using DNA nanoball-patterned arrays. Cell, 189(7), 1777–1792. https://doi.org/10.1016/j.cell.2022.04.003
Chen, H., Shao, X., Wei, R., Liu, H., Fan, X., & Wang, Y. (2024). Spatial transcriptomics: A new frontier in plant science. Plant Communications, 5(4), 100834. https://doi.org/10.1016/j.xplc.2024.100834
Clevenger, M. H., Cipurko, D., Patil, A., Li, B., Takahama, M., Mei, L., Plaster, M., Richey, G., Kawamoto, T., Bao, F., & Chevrier, N. (2026). Whole-body molecular and cellular mapping of the laboratory mouse. Cell, 189, 1–18. https://doi.org/10.1016/j.cell.2026.03.006
Moses, L., & Pachter, L. (2022). Museum of spatial transcriptomics. Nature Methods, 19, 534–546. https://doi.org/10.1038/s41592-022-01409-2
Science Commentary. (2026, April 2). [Perspective on Clevenger et al., 2026 — milestone whole-body spatial transcriptomics]. Science. https://doi.org/10.1126/science.zgr6u94

Leave a comment

Molecular Ghosts in the Mud: How Ancient Soil DNA Is Resurrecting the Earth’s Lost Worlds

Santhosh Eapen — Sat, 28 Mar 2026 08:26:22 GMT

Imagine reconstructing the heat of a single July afternoon from 30,000 years ago. The witnesses to that day - the wildflowers, the sub-arctic grasses, the larch trees - have long since vanished. For decades, palaeoclimatologists relied on fossilised pollen and the mouthparts of midges to guess at the past. These gave us blurred, regional sketches at best. A new kind of molecular forensics has changed this entirely. By extracting Sedimentary Ancient DNA (sedaDNA) from the mud of ancient lakebeds and ice-age cave floors, we are no longer guessing. We are listening to the genetic whispers of vanished worlds.

A Nature feature published on 24 March 2026 documents how this field has matured from a ridiculed doctoral hypothesis - Eske Willerslev, watching a dog relieve itself on a Copenhagen street in 2000, wondering whether DNA persists in soil - into one of the most transformative tools in the study of the past. His professors thought the idea was stupid. His 2003 paper in Science recovered DNA from 400,000-year-old Siberian permafrost. His 2022 paper pushed that record to two million years, from Greenland permafrost. The field that did not exist 25 years ago now routinely reads the genetic record of entire ecosystems from a soil core.

But the popular narrative has focused too narrowly on the human story - Neanderthals, Denisovans, ancient migrations. SedaDNA is simultaneously rewriting palaeobotany, palaeozoology, palaeoclimatology, and microbial ecology. The real story is that every organism that ever shed a cell leaves a molecular ghost in the mud. This Research Rundown covers all of them.

🌿 Plants: The Wind’s Deception, Finally Corrected

Pollen analysis has a fundamental flaw: pollen is designed to travel. A lake sediment sample captures pollen from forests hundreds of kilometres away, creating what scientists call a “regional blur” with a built-in warm bias - wind-pollinated boreal species consistently inflate temperature estimates in what were actually tundra environments. SedaDNA from the same sediment is a homebody. It originates from the immediate lake catchment - the leaves, stems, and cells that wash directly from the surrounding slopes - giving a high-definition local signal that pollen cannot provide.

The resolution difference is decisive. Pollen routinely stops at family or genus level. Plant sedaDNA, using the trnL-gh chloroplast marker, resolves 77.2% of taxa to genus or species level. That taxonomic precision transforms climate reconstruction from approximate to quantitative: species within the same genus can have radically different temperature tolerances, and those differences are exactly what a climate signal requires.

Calibrated against 203 Siberian lake samples, the plantDNA_PDF/SDM framework - which uses machine learning-based species distribution models from GBIF data to build temperature probability functions for 180 plant taxa - achieves a median temperature bias of 0.5–1.1°C. Pollen and chironomid methods produce errors of 1.5–2.5°C in the same settings. SedaDNA is now the most precise biological thermometer for the deep past.

There is a profound irony here: we are using 21st-century machine learning and global biodiversity databases to decode a world that existed 32,000 years ago. Our most futuristic technology has become the ultimate tool for our most ancient history.

🐘 Animals: The Census the Bone Record Could Never Take

The bone record is systematically biased. It preserves large animals that were hunted, butchered, and left in quantities large enough to survive millennia. Small animals - birds, rodents, reptiles - are largely absent. Animals that passed through a site without dying there leave nothing. SedaDNA bypasses every one of these biases.

The study of El Mirón Cave in Cantabria, Spain (Nature Communications, 2025) illustrates the point. From cave sediments spanning the Neanderthal period through the Magdalenian, researchers identified 28 animal taxa. Fifteen of these had never been identified from the bone record at that site - including cave hyena, Iberian lynx, falcon, woolly rhinoceros, shrew, and owl. These animals were there. The bones simply did not survive or were never recovered. Sediment DNA provides what is effectively a non-selective faunal census: a record of every organism that shed DNA into the water or soil, regardless of body size, hunting value, or preservation potential.

At Denisova Cave, the sedaDNA timeline documents multiple turnovers in the broader mammalian community - bears, hyenas, and other fauna - correlated with climate transitions, alongside the hominin population shifts. The sediment held a complete ecological record that no bone assemblage could replicate.

🦠 Microbes: The Invisible Majority With a Lot to Say

In any sedaDNA sample, microbial sequences - bacteria, archaea, fungi, protists - constitute the overwhelming majority of recoverable DNA. This is why detecting human or animal sequences requires targeted molecular probes. But this microbial signal is not noise. It is arguably the richest information channel in the sample.

Shifts in fungal assemblages mark wet-to-dry transitions. Changes in bacterial communities reflect altered nutrient inputs, fire events, and vegetation turnover. The presence of specific microbial taxa indicates past flooding or intensive land use. And coprolites - fossilised faeces, one of the “archaeological composites” that sedaDNA methodology encompasses - are preserved gut microbiome snapshots: a precise molecular portrait of the defecator’s identity, diet, and intestinal microbial community at a specific moment in time.

Ancient gut microbiome studies from European and American sites are already documenting dramatic losses in microbial diversity between hunter-gatherer populations and modern post-industrial guts. Ancient pathogen DNA survives in sediment as well: Yersinia pestis has been recovered from European archaeological sediments, documenting plague spread through time. The same approach could, in principle, trace the history of rice blast or black pepper root rot in South Asian archaeological contexts - giving crop protection research a deep evolutionary baseline it currently lacks entirely.

🧑 Humans: What Dirt Revealed That Bones Could Not

The 2017 watershed moment came when Max Planck Institute researchers used molecular probes to pull human mitochondrial DNA from ice-age cave soils - the first demonstration that human genetic history could be read from sediment without physical remains. The findings came quickly after that. At Denisova Cave, a sediment sample from a deep layer placed Neanderthal occupation 170,000 years ago - 30,000 years before the oldest fossil had suggested. Early modern humans were placed at the site from ~45,000 years ago with not a single human bone recovered from those levels. DNA also linked specific hominin groups to specific stone tool assemblages - answering who made which tool, one of archaeology’s most persistent debates, through the soil rather than the specimen.

At the Baishiya Karst Cave on the Tibetan Plateau, sedaDNA from cave soil in 2020 provided the first incontrovertible evidence that Denisovans lived outside Siberia - confirming what protein analysis of a jawbone had suggested but could not definitively prove. The frontier has since moved to nuclear DNA. Benjamin Vernot’s 1.6 million-probe system recovered sequences at ~10,000 genomic positions from the best sediment samples - sparse, but sufficient to distinguish two separate Neanderthal populations at a Spanish cave site. Targeted probe-based capture outperforms shotgun sequencing by 32-fold at informative genomic positions. The bottleneck now is not chemistry. It is bioinformatics.

⚖️ One Non-Negotiable: Ethics Is Not Optional

Sediment is ubiquitous, and some researchers have treated it as a convenient bypass around the ethical and legal requirements that govern work with skeletal remains. This is a mistake. Even sediment devoid of visible remains can contain ancestral human DNA. The cautionary case is Pueblo Bonito in Chaco Canyon, where tribal groups were not consulted, data descriptions were culturally insensitive, and researchers did not consider how results might affect descendants. Technically compliant. Ethically indefensible.

“The sampling and analysis of human DNA from composites cannot be seen as a workaround for collaboration and communication with descendants.” — Johnson et al.

The emerging consensus is clear: engage descendant communities as equal partners from the planning stage, not as an afterthought. Apply the Belmont Report principles - Respect for Persons, Beneficence, Justice - and treat data sovereignty as a non-negotiable. Genetic datasets should not be deposited in public repositories without explicit community consent. Only proceed with destructive sedaDNA analysis when the question genuinely cannot be answered by less invasive means. Leave a voucher specimen. Always.

🇮🇳 What This Means for Indian Researchers

This methodology is not an abstraction. It has specific, concrete applications here.

The Western Ghats are an origin centre for black pepper, cardamom, and turmeric. Lake sediments in Kerala and Tamil Nadu hold molecular records of these crops’ transition from wild to cultivated - the domestication timeline that archaeobotany alone cannot resolve.
South Asia’s population genetic history is among the most complex in the world. Cave sedaDNA from India’s extensive cave archaeological record could address unresolved questions about early human dispersal into the subcontinent.
Ancient plant pathogen DNA from agricultural archaeological sites could provide evolutionary baselines for crop diseases that are still economically significant - precisely the kind of deep history that resistance breeding programmes need.
India’s wildlife conservation challenges are inseparable from questions about historical range and population structure. SedaDNA from lake sediments can document faunal histories that the bone record cannot.

The infrastructure is within reach: CSIR, ICAR, and IISc facilities have sequencing capacity. GBIF data and ancient DNA analysis tools are open-access. What is needed is the disciplinary bridge between archaeology, ecology, genomics, and bioinformatics - and the institutional will to build it.

The soil beneath our feet is a living archive. From the frigid peaks of the Last Glacial Maximum to the forest floors of the Western Ghats, it holds a record of every breath, leaf-fall, and footstep that ever graced a landscape. SedaDNA is giving us the tools to read that record - for the plants that shaped climates, the animals that shared ecosystems, the microbes that colonised guts, and the humans who walked before us. If we are careful with our science and respectful of our ethics, the mud will prove to be our most honest historical record.

What is your take?

Is your institution positioned to engage with sedaDNA or environmental genomics? I’d like to hear from researchers across India working in archaeogenomics, palaeobotany, conservation genetics, or crop disease history.

Key Takeaways

SedaDNA captures plants, animals, microbes, and humans from a single sample - a non-selective biological census no other method can replicate.
Plant sedaDNA resolves 77.2% of taxa to genus/species level; climate reconstruction precision reaches 0.5°C - outperforming pollen and chironomid methods.
Animal sedaDNA from El Mirón Cave found 28 taxa, 15 absent from the entire bone record at that site. The non-selective faunal census is its defining advantage.
Microbial sedaDNA (including coprolites) preserves ancient gut microbiomes and pathogen histories. Ancient crop disease evolution is an untapped application.
Human sedaDNA placed Neanderthals 30,000 years earlier than fossil evidence; Denisovans confirmed outside Siberia from cave sediment alone.
Ethics is non-negotiable: descendant community partnership, data sovereignty, and minimal-destruction protocols are methodological requirements, not formalities.
India has concrete opportunities in spice crop domestication, South Asian population history, faunal range reconstruction, and ancient pathogen evolution.

Additional Reading

Aldeias, V., & Stahlschmidt, M. C. (2024). Sediment DNA can revolutionize archaeology—if it is used the right way. Proceedings of the National Academy of Sciences, 121, e2317042121. https://doi.org/10.1073/pnas.2317042121
Herzschuh, U., Böhmer, T., Jia, W., & Lisovski, S. (2025). Quantitative climate reconstruction from sedimentary ancient DNA: Framework, validation and application. EGUsphere [Preprint]. https://doi.org/10.5194/egusphere-2025-2678
Johnson, T., Thakar, H. B., Watkins, J., & Linderholm, A. (2024). Working ethically with ancient DNA from composites in the United States. Advances in Archaeological Practice, 12(2), 86–97. https://doi.org/10.1017/aap.2023.32
Kjær, K. H., Winther Pedersen, M., et al. (2022). A 2-million-year-old ecosystem in Greenland uncovered by environmental DNA. Nature (London), 612(7939). https://doi.org/10.1038/s41586-022-05453-y
Lewis, D. (2026, March 24). How DNA in dirt is shaking up the study of human origins. Nature, 651, 868–870. https://doi.org/10.1038/d41586-026-00879-0

Leave a comment

AlphaGenome: The AI That Learned to Read the Language of Life

Santhosh Eapen — Fri, 20 Mar 2026 01:15:15 GMT

INTRODUCTION

The Genome’s Dark Matter — and Why It Matters

Here is a number worth pausing over: 98. That is the percentage of the human genome that does not encode proteins. For decades it was dismissed as ‘junk DNA’ — a biological warehouse of evolutionary debris with little practical relevance. We now know this was spectacularly wrong.

That 98% is where the real instructions live. It is the genome’s control panel — the vast network of regulatory sequences, enhancers, silencers, splice signals, and chromatin architects that decide when a gene switches on, in which tissue, at what intensity, and for how long. Variation in this regulatory dark matter underpins susceptibility to cancer, rare genetic diseases, agricultural stress responses, and pathogen resistance. Yet interpreting it has remained one of biology’s most obstinate unsolved problems.

Until now. On January 28, 2026, the Google DeepMind team that gifted science the Nobel Prize-winning AlphaFold published AlphaGenome in Nature — a unified deep-learning model that can ingest one million base pairs of DNA and simultaneously predict thousands of molecular properties characterising that sequence’s regulatory behaviour. In doing so, it may have handed researchers the most powerful genomic interpretation tool ever built.

A horizontal timeline tracing Google DeepMind’s Alpha series of life-sciences AI tools across six years. AlphaFold (2020) solved the protein structure prediction problem; AlphaFold 2 (2021) achieved near-experimental accuracy, earning the 2024 Nobel Prize in Chemistry; AlphaMissense (2023) extended predictions to the functional impact of coding variants; AlphaGenome (preprint June 2025; Nature January 2026) pushes further still — from protein-coding sequence into the vast regulatory genome. A rising diagonal arrow illustrates the expanding biological scope, from individual protein structures to the full regulatory language of DNA.

🔬 At a Glance

AlphaGenome is published in Nature (January 2026) by the same DeepMind team behind AlphaFold. It predicts regulatory function from raw DNA sequence across 11 molecular modalities, at single-base resolution, over 1 megabase windows — outperforming all previous specialist models on 25 of 26 benchmarks.

THE CHALLENGE

Why Interpreting Non-Coding DNA Has Been So Hard

To appreciate what AlphaGenome achieves, it helps to understand what researchers were working with before it arrived. The field had produced a proliferation of specialist computational models — each designed to interpret one type of regulatory signal. One model predicted splicing; another modelled chromatin accessibility; a third estimated transcription factor binding. Using them together required stitching outputs from multiple incompatible systems, each with different input requirements, resolutions, and underlying assumptions.

Worse, these models faced a fundamental trade-off: accuracy versus context length. Models that could process long DNA sequences (hundreds of kilobases) sacrificed resolution, producing predictions at 128 or 32 base-pair bins rather than at individual nucleotides. Those with single-base resolution could only process very short windows, missing the influence of long-range regulatory elements that can sit hundreds of thousands of bases from the genes they control.

The consequence was significant: non-coding variants — the majority of hits from genome-wide association studies (GWAS) — were notoriously difficult to functionally interpret. Researchers could identify associations between genetic variants and disease, but pinpointing the molecular mechanism behind those associations required expensive, time-consuming wet-lab experiments.

HOW IT WORKS

A Swiss Army Knife for the Regulatory Genome

AlphaGenome resolves the resolution-versus-context trade-off that constrained its predecessors by combining two complementary deep-learning architectures: convolutional layers that capture fine-scale sequence motifs, and transformer layers that model long-range dependencies across the megabase input window. The result is a model that sees both the detail and the landscape simultaneously.

A single API call to AlphaGenome returns predictions across 11 molecular modalities — simultaneously and without requiring the user to manage multiple specialist tools. These include:

Gene expression levels across hundreds of cell types and tissues
Transcription initiation sites and their cell-type specificity
RNA splicing — splice site usage, junction coordinates, and expression strength
Chromatin accessibility (ATAC-seq tracks)
Histone modification patterns (H3K27ac, H3K4me3, and others)
Transcription factor binding sites
3D chromatin contact maps (Hi-C proximity)

Variant scoring is equally elegant. To assess the functional impact of a genetic variant, AlphaGenome simply compares predictions between the mutated and unmutated sequences — producing a multimodal readout of effect in under one second. This computational efficiency transforms what was previously a multi-day, multi-tool analysis into an instantaneous query.

A left-to-right pipeline diagram showing AlphaGenome’s five-stage architecture. A raw DNA sequence of up to one million base pairs enters the model (Stage 1), passes through convolutional layers that detect local sequence motifs (Stage 2), then transformer layers that capture long-range regulatory interactions across the full megabase window (Stage 3). The model simultaneously outputs predictions across 11 molecular modalities — gene expression, splicing, chromatin state, histone marks, transcription factor binding, and 3D contacts — across hundreds of cell types (Stage 4). Finally, variant scoring compares reference and mutated sequences to produce a multimodal effect readout in under one second (Stage 5).

A Landmark Achievement: Modelling RNA Splicing

Perhaps the most clinically significant advance is AlphaGenome’s explicit modelling of RNA splicing. Splicing — the editing of pre-mRNA to remove introns and join exons — is exquisitely sensitive to nearby DNA sequence, and errors in splicing cause a wide range of rare genetic diseases. Previous models could predict splice site strength but not the full landscape of splice junctions at the expression level. AlphaGenome predicts both, directly from sequence, opening new avenues for interpreting variants in intronic regions that were previously opaque to computational analysis.

🧬 Case Study: Cancer Variant Interpretation

Researchers used AlphaGenome to investigate mutations observed in T-cell acute lymphoblastic leukaemia (T-ALL). The model predicted that the mutations create a MYB transcription factor binding motif, which activates the TAL1 oncogene located 7.5 kb away — precisely replicating the known disease mechanism. This demonstration illustrates the model’s capacity to connect non-coding variants to downstream gene activation through specific molecular intermediaries.

PERFORMANCE & RECEPTION

Benchmarks, Expert Reactions, and Honest Limitations

AlphaGenome has been subjected to rigorous benchmarking against 26 variant effect prediction tasks — and it matches or exceeds all challenger models on 25 of them. Notably, many of the models it surpassed were purpose-built specialists optimised for exactly the tasks where AlphaGenome now leads. Its predecessor in context-length processing, Borzoi, worked on 500,000 base-pair windows and predicted properties at 32 base-pair resolution. AlphaGenome doubles the context and achieves single-nucleotide precision — an improvement that one computational biologist at Stanford described as ‘not just a bigger model’ but ‘quite a leap forward in overall utility.’

For gene-activity changes across specific cell types, AlphaGenome outperformed Borzoi by 14.7% — a meaningful gain when studying tissue-specific regulatory variants. Expert reception from the scientific community has been largely enthusiastic:

💬 Expert Voice
“For the first time, we have a single model that unifies long-range context, base-level precision and state-of-the-art performance across a whole spectrum of genomic tasks.” — Caleb Lareau, PhD, Memorial Sloan Kettering Cancer Center
💬 Expert Voice
“AlphaGenome represents a major advance in computational genomics... As seen previously with AlphaFold, the full impact of such approaches is likely to unfold over time.” — Prof. Kristian Helin, CEO, The Institute of Cancer Research, London

Acknowledged Limitations

The DeepMind team is admirably candid about the model’s current boundaries. Distant regulatory elements — those more than 100,000 base pairs from their target genes — remain poorly captured. The model is trained on bulk tissue datasets, limiting reliability for rare cell types or specific developmental stages. And critically for the research community reading these pages: the model currently covers only human and mouse genomes.

This species restriction is not a permanent ceiling. Fergal Martin of EMBL-EBI explicitly noted that the architecture ‘is expected to expand to the DNA interpretation of more species such as plants, animals, and microorganisms in the future.’ The framework is sound; it awaits training data for new species.

IMPLICATIONS FOR CROP & PLANT SCIENCE

Why Plant and Crop Researchers Should Watch AlphaGenome Closely

For those of us whose research lives are rooted in the regulatory biology of spices, plantation crops, and their pathogens, AlphaGenome’s current human-mouse limitation is a temporary frustration rather than a fundamental barrier. The underlying scientific problems it addresses are deeply familiar to plant biologists — and the solutions it demonstrates translate directly.

Non-Coding Variation in Crop Genomes

Plant genomes — particularly the complex, polyploid genomes of pepper, cardamom, black pepper, vanilla, and cinnamon — are replete with regulatory variation. Traits like resistance to Phytophthora, tolerance to drought, expression of volatile biosynthesis pathways, and adaptation to elevation gradients are all mediated substantially by regulatory sequences rather than coding changes. GWAS studies in crops face exactly the same interpretive bottleneck as human genetics: thousands of associated loci, predominantly non-coding, with no molecular mechanism to explain them. An AlphaGenome equivalent trained on spice crop genomes would be transformative.

Splice Variants and Pathogen Interaction

AlphaGenome’s RNA splicing modelling is particularly relevant for crop protection researchers. Alternative splicing plays documented roles in plant immune responses — including the NLR (nucleotide-binding leucine-rich repeat) immune receptor genes that are central to resistance breeding. The ability to predict how a variant in a promoter or intronic region affects splice junction usage — without a full wet-lab transcriptome experiment — would accelerate the characterisation of disease resistance loci enormously.

A split two-panel infographic connecting AlphaGenome’s core capabilities (top, dark teal) to their most significant applications in plant and crop research (bottom, forest green with botanical spice plant illustrations). Four gold arrows bridge the panels: regulatory variant prediction maps to GWAS follow-up in spice crop genomes; splice junction modelling applies to NLR immune receptor genes central to disease resistance breeding; synthetic DNA design enables in silico promoter engineering for secondary metabolite pathways; and planned multi-species expansion charts the path toward plant-trained models. Botanical motifs — rice, wheat and potato — ground the scientific framework in the crops that stand to benefit most.

Synthetic Biology and Promoter Design

Beyond variant interpretation, AlphaGenome enables the design of synthetic regulatory sequences. A researcher could specify a desired expression pattern — tissue-specific, inducible by pathogen attack, or constitutive — and use AlphaGenome to in silico evaluate synthetic promoter constructs before any laboratory work. This application, explicitly highlighted by DeepMind, has profound implications for crop biotechnology and the engineering of metabolite pathways in secondary metabolite-rich crops.

🌿 Looking Ahead for Plant Science

The key enabling step for plant genomics is training data. Comprehensive regulatory genomics datasets for crop species — the equivalent of ENCODE and GTEx for human cells — would allow the AlphaGenome architecture to be fine-tuned for plant genomes. Existing resources like PlantCARE, ENCODE-plant extensions, and emerging functional genomics datasets for Arabidopsis and major crops provide a foundation. The community has both the data and the motivation. DeepMind has provided the architecture.

PRACTICAL ACCESS

How to Access AlphaGenome — A Researcher’s Guide

Since its preprint in June 2025, nearly 3,000 scientists across 160 countries have registered to use AlphaGenome. Following Nature publication in January 2026, DeepMind made the model’s source code and weights publicly available for non-commercial academic use. Here is how to get started:

Web API (no coding required): Access the model at deepmind.google.com/science/alphagenome — input a DNA sequence and receive multimodal predictions via a browser interface or the accompanying Python SDK.
GitHub Repository: Full source code, model weights, and example scripts available at github.com/google-deepmind/alphagenome_research — enables customisation, fine-tuning for new species, and integration into existing bioinformatics pipelines.
Genome Interpretation Suite: A bundled toolkit provides streamlined variant scoring with quantile calibration and contribution score mapping to identify which specific bases drive model predictions.
Python SDK: Allows programmatic batch querying — essential for GWAS follow-up analyses where hundreds of thousands of variants need scoring simultaneously.

For researchers in plant and crop sciences, the immediate practical applications lie in (a) validating AlphaGenome predictions against existing plant regulatory datasets as a benchmarking exercise, and (b) contributing to community efforts to build ENCODE-equivalent functional genomics resources for economically important crops. Both activities will accelerate the arrival of a plant-trained equivalent.

CONCLUSION

Reading the Recipe — and Finally Understanding What It Says

DeepMind’s vice-president for science, Pushmeet Kohli, has called AlphaGenome ‘like AlphaFold 1 — a big first step.’ That framing is instructive. AlphaFold 1 was already revolutionary; AlphaFold 2 transformed structural biology. If AlphaGenome follows the same developmental arc — absorbing new training data, expanding species coverage, and improving its handling of long-range regulatory interactions — its eventual impact on functional genomics may be equally profound.

The immediate contribution is clear and valuable: a single, openly accessible model that can interpret any DNA variant across the full spectrum of regulatory molecular biology, at single-nucleotide resolution, in seconds. For researchers grappling with the avalanche of GWAS data from crop species, for plant biotechnologists designing precision regulatory constructs, and for crop protection scientists seeking the molecular basis of disease resistance, this represents a fundamentally new capability — even before plant-specific training arrives.

The regulatory code of life is being deciphered. The language is not yet fully understood, but for the first time, we have a unified grammar book.

KEY TAKEAWAYS FOR RESEARCHERS

AlphaGenome is free to access for non-commercial academic research via a web API and GitHub.
It is currently trained on human and mouse genomes, but the architecture is explicitly designed for extension to plants, microorganisms, and other species.
Its unified, multimodal approach eliminates the need to manage multiple specialist genomics models.
RNA splicing, chromatin state, and transcription factor binding are now simultaneously predictable from raw DNA sequence at single-base resolution.
Plant and crop scientists should engage now: benchmarking the model on available plant datasets and contributing to community functional genomics resources will accelerate plant-specific training.
For crop biotechnology, the synthetic DNA design capabilities offer immediate utility for in silico promoter and regulatory element engineering.

REFERENCES

Avsec, Ž., Latysheva, N., Cheng, J. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature, 649, 1206–1218. https://doi.org/10.1038/s41586-025-10014-0
García González, J. & Ghoussaini, M. (2026). AlphaGenome: a Swiss-army knife for exploring non-coding DNA. Trends in Genetics, 42(1), 4–6. https://doi.org/10.1016/j.tig.2025.11.007
Google DeepMind. (2026). AlphaGenome: AI for better understanding the genome [Blog post]. https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/
Raza, A. et al. (2026). From the genome to super-pangenome: a new paradigm for accelerated crop improvement. npj Science of Plants, 2, Article 4. https://doi.org/10.1038/s44383-025-00019-z
Science Media Centre. (2026, January 28). Expert reaction to paper on Google DeepMind’s AlphaGenome. https://www.sciencemediacentre.org
Thanks for reading Research Rundown! Subscribe for free to receive new posts and support my work.
Leave a comment
Share

The Microbiome Our Crops Lost — and the Science of Getting It Back

Santhosh Eapen — Fri, 13 Mar 2026 18:07:54 GMT

A paper that crossed my desk recently - submitted in February 2026 by Mike Bos under the supervision of Prof. George Kowalchuk at Utrecht University - raises a question that the plant science community needs to answer more honestly than it has so far: when we talk about rewilding crop microbiomes, are we talking about a genuine breakthrough for sustainable agriculture, or are we papering over serious unresolved trade-offs with premature optimism?

The paper is titled Rewilding Crop Microbiomes: Payoff or Trade-Off? The title alone is a signal. In a domain where the language of boundless potential dominates, the author asks the harder question. That deserves attention.

What Domestication Took Away from Our Crops

About 13,000 years ago, humans began domesticating plants. The logic was straightforward: select for yield, select for ease of harvest, select for the traits that made farming manageable. The Green Revolution of 1944 accelerated this further, introducing high-yielding varieties, synthetic fertilizers, and pesticides that reshaped agriculture globally.

What did we lose? More than most people realise.

The numbers tell the story plainly. Domesticated soybean (Glycine max) has lost about 49% of its genetic diversity compared to its wild relative Glycine soja. Domesticated tomato (Solanum lycopersicum) has lost about 77%, rice (Oryza sativa) about 30%, maize (Zea mays) 27%, bread wheat (Triticum aestivum) 69%, and durum wheat (Triticum turgidum ssp. durum) a staggering 84%. These are not abstract genomic statistics. They represent the erosion of adaptive potential — the biological capital that plants draw on when conditions become difficult.

But what often goes unmentioned is what this genetic bottleneck did to the microbial partnerships these plants depend on.

Plants are not solitary organisms. They live in intimate alliance with a consortium of bacteria, fungi, archaea, viruses, and protozoa — collectively termed the microbiome. These microbes inhabit the rhizosphere (the soil zone surrounding roots), the phyllosphere (leaf surfaces), and the endosphere (internal tissues). Their contributions are not minor: they fix atmospheric nitrogen, mobilize phosphorus and zinc, produce antibiotics, prime the plant’s immune defenses through induced systemic resistance (ISR), and help plants survive drought, heat, salinity, and flooding.

Crops rely on the rhizosphere microbiome for functions like managing nitrogen, phosphorus, stress, and pathogens. Synthetic inputs only partially replace these essential services, which domestication eroded by diminishing the plant’s ability to recruit the necessary organisms.

When we bred crops for yield under fertilizer-rich, low-stress conditions, we inadvertently selected away from the genes governing microbial recruitment. Modern wheat cultivars show far lower dependence on arbuscular mycorrhizal fungi (AMF) than their wild ancestors. Modern maize is colonized by AMF to a lesser extent than wild teosinte, and the diversity of AMF associated with it is reduced. Domesticated rice has lost the ability to recruit nitrogen-fixing Frankiaceae bacteria that are abundant in the rhizosphere of wild rice. Domesticated legumes appear to have weakened their ability to sanction ineffective nitrogen-fixing rhizobia — a quality-control mechanism wild legumes retain.

In wheat, researchers found a decrease in nitrogen transformation and phosphorus mineralization in domesticated cultivars compared to their wild counterparts. In rice, domestication increased nitrous oxide emissions — a potent greenhouse gas — while reducing nitrogen fixation in the soil.

The paradox is this: we bred crops to need less from their microbiome, then supplied the difference artificially. Fertilizer replaced nitrogen fixation. Pesticide replaced microbial disease suppression. The plants delivered yields. The ecological support structure quietly collapsed.

This is a serious gap. Not a minor concern to be managed around the edges. A serious gap.

The Rewilding Idea: What It Proposes

In a landmark 2022 paper in Science, researchers Jos Raaijmakers and E. Toby Kiers proposed “microbiome rewilding” — restoring the ancestral microbial interactions that crops have lost. Their four-step research roadmap has become the field’s organizing framework.

The roadmap starts with identifying microbes associated specifically with wild relatives of domesticated crops. The second step validates which ancestral microbes actually benefit crop phenotype. The third identifies the plant genetic loci responsible for recruiting and maintaining these microbes. The fourth step identifies the root exudates and signaling molecules that activate the beneficial interaction.

The Raaijmakers–Kiers four-step roadmap proposes a research framework to recover ancestral plant-microbe interactions: identify microbes in wild relatives, validate benefits, map plant genetic loci for recruitment, and reconnect signaling chemistry. The framework’s logic is sound, but field validation is still needed.

The ambition is clear: reintroduce the biological machinery that thousands of years of domestication switched off.

Several early results are genuinely encouraging. In cannabis, seed priming with Bacillus sp. and Pseudomonas putida isolated from low-domesticated varieties significantly improved germination rates and root formation in more domesticated lines. Bacteria isolated from the wild grass Celtica gigantea improved grain yield and carotenoid content in the cereal crop tritordeum under field conditions. Synthetic microbial communities (SynComs) constructed from wild rice microbiomes improved root elongation in domesticated rice. In soybean, SynComs including Rhodococcus, Lysobacter, and Arthrobacter promoted growth and nitrogen acquisition without relying on conventional rhizobial nodulation.

In barley, a genomic locus named QRMC-3HS — linked to the recruitment of specific bacterial taxa in the rhizosphere — was identified and found to sit close to a yield QTL. This suggests that yield-directed selection directly altered microbiome composition through genetic linkage. It is not a coincidence. It is a consequence.

These results matter. They demonstrate that the lost interactions are, at least in principle, recoverable.

But this is precisely where the honest accounting must begin.

The Trade-Offs That Need to Be Stated Plainly

The Bos thesis makes a point that most rewilding literature conspicuously avoids: the trade-offs are real, they are under-researched, and they may determine whether rewilding succeeds at scale or remains a greenhouse phenomenon.

Three categories of cost are relevant.

Genetic costs arise when the introduction of beneficial traits from wild relatives drags in linked undesirable alleles — a phenomenon called linkage drag. Introducing the scab resistance gene Rvi6 in apple (Malus domestica) improved resistance to the fungal pathogen Venturia inaequalis, but led to poor fruit quality, rendering the variety commercially unviable. QTLs for Phytophthora resistance in tomato and pepper were linked to negative effects on yield and fruit size. This is not an exception to the rule. It is the rule.

Allocation costs are perhaps the most structurally important trade-off. Microbial associations are metabolically expensive. Nitrogen-fixing microbes in root nodules consume 20–25% of the plant’s photosynthate. AMF take 10–20%. Maintaining induced systemic resistance reduces growth rate in the absence of pathogens. Inoculation of common bean with bacteria previously ascribed growth-promoting ability actually downregulated expression of genes involved in defense responses and detoxification. The interaction with nitrogen-fixing Rhizobium suppresses the plant’s innate immunity, increasing susceptibility to other pathogens.

Microbial symbiosis is costly. Nitrogen fixers and mycorrhizal fungi consume 30-45% of a plant’s photosynthate, directly competing with growth and pathogen defense. This structural trade-off must be accounted for in rewilding strategies to ensure replicable field results.

The trade-off is not theoretical. It is a basic biological constraint: a plant cannot simultaneously maximize growth, maximize defense, and maintain a full complement of energy-costly microbial symbionts.

The Bos paper highlights a study (Favela et al., 2026, Science Advances) that claimed to introduce reduced nitrification rates in maize lines through teosinte introgressions “without a performance penalty.” The author’s scrutiny reveals a different picture: both the NIL047 and NIL021 hybrid lines showed reduced grain yield, albeit non-significant in a small sample size (n=16). The larger study would likely reveal the trade-off the authors chose to downplay. This raises uncomfortable questions. It is not about whether rewilding shows promise. It is about whether the field is prepared to report trade-offs as rigorously as it reports benefits.

Field challenges add another layer of difficulty. The rhizosphere is a place of intense microbial competition. Native microbes routinely outcompete introduced ones. A SynCom that improved sorghum growth in greenhouse conditions lost many of its strains in field conditions within weeks. Bacterial strains improving sugar beet yield in California proved ineffective in Idaho. Strains effective in Idaho failed in California. Rhizobium leguminosarum improved rice in greenhouse experiments but showed no significant effect on grain yield in the field. The context dependency of these results is not a small problem. It is the central problem.

Greenhouse magic fails in the field. SynComs, effective in controlled settings, collapse in real-world soil. Location matters—what works in California won’t in Idaho. Deployment demands rigorous, multi-year, multi-environment field trials to measure the full trade-offs.

The Health and Ecological Risks No One Wants to Lead With

There is a dimension of microbial inoculation research that receives considerably less attention in promotional discussions: some beneficial microbes are simultaneously opportunistic human pathogens.

Burkholderia cepacia promotes growth in pepper and increases resistance to F. oxysporum, P. capsici, and R. solani - all significant plant pathogens. The same organism is an opportunistic human pathogen, particularly dangerous for immunocompromised patients. Our studies showed a strain of Pseudomonas aeruginosa to protect black pepper (Piper nigrum) against Phytophthora capsici and Radopholus similis - yet the same strain showed virulence similar to clinical P. aeruginosa strains in murine lungs.

Horizontal gene transfer (HGT) adds further risk. Rhodococcus spp. can transition from plant-beneficial bacterium to plant pathogen through acquisition of a virulence plasmid. Many commercial plant growth-promoting bacteria carry antibiotic resistance genes that can transfer to native soil microbes through HGT. The rhizosphere is already a reservoir for opportunistic human pathogens, and introducing non-native microbial species into it may have consequences we cannot fully model in advance.

The disruption of native microbial ecosystems is real. Introduction of Pseudomonas aeruginosa significantly reduced abundance of native bacteria. Introduction of S. meliloti and Azospirillum lipoferum reduced the abundance of soil nitrification and denitrifying bacteria under low carbon conditions.

These are not reasons to abandon rewilding. They are reasons to conduct long-term field trials before commercial deployment and to assess human safety as rigorously as agronomic efficacy. The Bos paper makes this point clearly, and it deserves to be made more loudly in public discussion.

Where Does This Leave Agriculture?

The Bos thesis concludes with a nuanced verdict: rewilding is a promising framework, but its success is highly context-dependent. In well-managed, fertilizer-rich, low-stress conditions, modern crop cultivars will likely outperform rewilded crops - precisely because they are optimized for those conditions and rewilded crops bear the metabolic cost of maintaining microbial partnerships. In stressed conditions - drought, pathogen pressure, nutrient-poor soils - crops with restored microbiomes may outperform, because their biological support system activates where chemical inputs cannot reach.

This is the key insight. Rewilding is not a universal replacement for conventional agriculture. It is a targeted strategy, most valuable where external inputs are unavailable or unaffordable, and where stress conditions are the norm rather than the exception. In subsistence farming, smallholder agriculture, organic systems, and climate-stressed environments, it may represent the most viable path to resilience.

The field needs to be honest about this. Rewilding as a concept should not be marketed as a wholesale alternative to modern agricultural practice. It should be positioned where the evidence actually places it - as a precision strategy for specific contexts, with identified benefits and acknowledged costs.

There is also a question that Indian agricultural researchers and policymakers should be asking directly: crops grown predominantly in smallholder, low-input systems in ecologically complex environments - what ancestral microbiome resources remain in wild relatives? What has been lost through centuries of selective cultivation? What can SynComs or targeted genetic approaches restore? These questions have not received the systematic attention they deserve.

Key Takeaways

Crop domestication has imposed a severe microbial cost. Genetic diversity losses of 30–84% across major crops have been accompanied by reduced ability to recruit nitrogen-fixing, phosphorus-mobilizing, and pathogen-suppressing microbes. These losses are measurable, documented, and consequential.

Rewilding is scientifically grounded, not speculative. The Raaijmakers–Kiers four-step roadmap provides a testable framework. Early results from SynCom studies and QTL mapping studies demonstrate that lost microbial interactions are recoverable in principle.

The trade-offs are real and under-reported. Genetic costs from linkage drag, allocation costs from microbial symbiosis, and the context-dependency of field performance are not minor caveats. They are central to whether rewilding scales from the greenhouse to the farm.

Field validation is non-negotiable. Greenhouse results and field results diverge dramatically and consistently. No rewilding intervention should move toward commercial application without multi-environment, multi-year field trials that specifically test for yield penalties alongside reported benefits.

Health and ecological safety assessments must be standard, not optional. The dual nature of some plant-beneficial microbes as human pathogens is documented. HGT risks are documented. Long-term ecological monitoring must accompany field deployment.

The real test of rewilding will not lie in the publication of promising greenhouse results. It will lie in whether crops with restored microbiomes outperform conventional cultivars under the stressed, resource-limited, unpredictable conditions that define farming in the places where food security is most precarious.

The science is ready to be taken seriously. The field needs to take its own trade-offs equally seriously.

Additional Reading

Raaijmakers, J. M., & Kiers, E. T. (2022). Rewilding plant microbiomes: Microbiota of crop ancestors may offer a way to enhance sustainable food production. Science, 378(6620), 599–600. https://doi.org/10.1126/science.abn6350
Ramirez-Villacis, D. X., Leon-Reyes, A., Pieterse, C. M. J., & Raaijmakers, J. M. (2025). Born to rewild: Reconnecting beneficial plant-microbiome alliances for resilient future crops. Cell Host & Microbe, 33(8), 1241–1255. https://doi.org/10.1016/j.chom.2025.06.017
Favela, A., Kent, A. D., Sible, C. N., et al. (2026). Lost and found: Rediscovering microbiome-associated phenotypes that reshape agricultural sustainability. Science Advances, 12(1), eaed3360. https://doi.org/10.1126/sciadv.aed3360
Chang, J., Costa, O. Y. A., Sun, Y., et al. (2025). Domesticated rice alters the rhizosphere microbiome, reducing nitrogen fixation and increasing nitrous oxide emissions. Nature Communications, 16(1), 2038. https://doi.org/10.1038/s41467-025-57213-x
Smith, M. E., Kavamura, V. N., Hughes, D., et al. (2025). Uncovering functional deterioration in the rhizosphere microbiome associated with post-green revolution wheat cultivars. Environmental Microbiome, 20(1), 64. https://doi.org/10.1186/s40793-025-00723-4

If this was useful, share it with a colleague who should be reading it.

If someone forwarded this to you and you’d like to receive Research Rundown regularly,

Subscribe now

Have a question or a pushback on something I wrote? Leave a comment — I read every one.

Leave a comment

Writing Life from Scratch

Santhosh Eapen — Fri, 06 Mar 2026 07:24:20 GMT

On 4 March 2026, a paper appeared in Nature that made many in the genomics community stop and read twice. A team led by Brian Hie at the Arc Institute and bioengineer Patrick Hsu published results from a model called Evo2 — the largest AI model trained on biological sequences to date. It had, among other things, designed a genome inspired by the bacterium Mycoplasma genitalium, generated human mitochondrial DNA sequences, and produced a yeast chromosome. The headlines arrived quickly. “AI can now write genomes.”

The question embedded in the Nature news piece accompanying the paper — “how long until it creates synthetic life?” — was equally revealing about the mood in the field. A reasonable question. But not the most important one.

The more important question is this: what does Evo2 actually demonstrate, what does it fall short of, and what does the gap between those two things require of us as researchers, policymakers, and institutions?

From Chemical Synthesis to AI Design: The 18-Year Road

To understand what Evo2 represents, you need to go back to 2008. Gibson and colleagues published the first-ever chemical synthesis of a living organism’s genome — the 580,000-nucleotide sequence of Mycoplasma genitalium, a bacterium with one of the smallest known genomes. Two years later, the same group showed they could “reboot” a synthetic genome inside a cell, producing what was claimed to be the first example of synthetic life.

It was a milestone. But it was also, at its core, copying. The sequence was already known. The researchers replicated it chemically and assembled it. What they demonstrated is that you can make life-functional DNA if you already know exactly what that DNA is supposed to say.

What followed over the next decade was incremental. The E. coli genome was recoded at nearly 20,000 locations, dispensing with three redundant codons — remarkable bioengineering, but still working from an existing blueprint. The Sc2.0 project has been rewriting the Saccharomyces cerevisiae genome segment by segment and is nearly complete. As genome engineer Patrick Yizhi Cai at the University of Manchester described it, these earlier efforts “amount to tinkering — the equivalent of editing a chapter of an existing book or removing all the commas.”

Evo2 represents a different kind of ambition. Not editing. Writing.

What Evo2 Is, and What It Was Built On

Think of Evo2 as something functionally analogous to the large language models that generate text — except its training corpus is not human language but the molecular language encoded in DNA. It was built collaboratively by the Arc Institute, Stanford University, UC Berkeley, UCSF, and NVIDIA, and trained on the OpenGenome2 dataset: 128,000 whole genomes and 9.3 trillion nucleotide pairs spanning all domains of life — bacteria, archaea, and eukaryotes, including complex organisms.

The model has 40 billion parameters and uses a novel architecture called StripedHyena 2, which the developers report trained nearly three times faster than comparable transformer-based models. It can process up to one million DNA letters in a single pass — a context window that matters biologically, since gene regulatory relationships often play out across long stretches of a genome.

Training required over 2,000 NVIDIA H100 GPUs running for several months. The model, its training data, and weights have been made publicly available, making Evo2 the largest fully open-source AI model in biology to date.

Evo2 can not only predict the effects of genetic mutations — including clinically significant variants in BRCA1 — but also generate novel DNA sequences it has never encountered. That capability is what makes the current paper worth taking seriously.

What Evo2 Actually Did

Before Evo2, Hie’s team had already used an earlier Evo model to design 285 genomes for bacteriophages — viruses that infect bacteria. When inserted into E. coli cells, 16 of those designs produced functional viruses capable of killing bacteria, including some antibiotic-resistant strains. The phage genomes were short — a few thousand nucleotides, encoding a handful of genes — but 16 out of 285 is not a figure to dismiss. It is proof that AI can write biological sequences that work.

The new Evo2 paper goes further. The model designed a genome inspired by M. genitalium, sequences modelled on human mitochondrial DNA, and a yeast chromosome. Computational predictions suggested that approximately 70% of the genes in the M. genitalium-inspired sequences looked realistic.

Cai’s assessment of the moment is worth recording:

“These AI models are the ‘ChatGPT moment’ for synthetic genomics. You can start writing things that never existed in nature.”

Here is where the honest accounting begins.

The 70% Problem — and Why It Matters

Seventy percent sounds impressive until you understand what a missing 30% means in a living system.

Nico Claassens, synthetic biologist at Wageningen University, stated it plainly: “You cannot design life 70%. You can do that on a computer, but it will not be functional.”

If even one essential gene is absent or poorly modelled, the genome fails inside a cell. Biology does not grade on a curve.

There is more. Even when all essential genes are present, how they are arranged within the genome matters. Maciej Wiatrak at the University of Cambridge, who developed the Bacformer model for bacterial genome generation, observed: “Evaluating whether your genome looks correct and works correctly are two different beasts.” A sequence that passes computational plausibility tests may fail the biological test entirely.

A preprint posted the same week from University of Texas, Austin added a further concern: Evo2’s genome designs are organised differently from natural genomes and lack some key structural features. The researchers were careful to say this does not definitively mean they are non-functional. But it means we should not conflate computational resemblance with biological viability.

And then there is the experimental bottleneck. Testing whether a designed genome works requires physically synthesising the DNA — at genome scale, a costly and technically demanding process. Hie himself acknowledged:

“Experiments are quickly becoming the bottleneck. At this scale, we bump up against the cost of DNA synthesis and the cost of DNA construction.”

The model can generate sequences far faster than we can test them. This is a serious gap.

What AI Genomics Could Actually Be Used For

The conversation about AI and synthetic life tends to converge quickly on the dramatic endpoint — will we create a synthetic organism from scratch? That question may be premature. But it should not distract from what AI genomics can offer in the nearer term, where applications are already taking shape.

The Broad Institute developed a model called DNA-Diffusion, based on image-diffusion architecture, to design synthetic cis-regulatory elements (CREs) — short DNA sequences that govern when and where specific genes are switched on. In early tests, the model successfully reactivated a protective gene in leukemia cells. The prospect this opens is significant: gene therapies targeted to specific cell types, reducing off-target effects that have long been a barrier to clinical use.

For agricultural biotechnology — where applications will eventually reach the laboratories and fields many in this audience work in — the implications are equally concrete. Designing operons, sets of inter-working bacterial genes that carry out defined tasks, is a far more tractable goal than building a full genome from scratch. Wiatrak’s group is pursuing exactly this: using models like Bacformer and Evo2 to design functional operon clusters for applications including biofuel production. Analogous approaches could yield microbial systems for biopesticide production, nitrogen fixation enhancement, or biosynthesis of secondary metabolites relevant to spice and plantation crop protection.

Industry investment is following this trajectory. Ginkgo Bioworks entered a five-year partnership with Google Cloud to train large language models on Ginkgo’s proprietary database of over two billion protein sequences, targeting protein design, drug discovery, and agriculture. This is not research for its own sake. These are commercial bets on AI-designed biological function delivering results.

The Ethical Ledger — and a Risk That Deserves Equal Attention

Every technology that accelerates biological design also accelerates its potential misuse. The ability to engineer genomes at speed raises the risk of designing novel pathogens or enhancing the virulence of known ones. Introducing synthetically engineered organisms into ecosystems poses unpredictable threats to species diversity and ecological balance.

These risks are not hypothetical, and they are not being ignored. But a quieter risk deserves equal attention: the misuse of AI not in a wet laboratory, but in the literature. Researchers at the University of Wisconsin–Madison have warned that over-relying on AI to fill data gaps in genome-wide association studies (GWAS) can produce false positives — spurious correlations between genetic variants and complex diseases like diabetes or Alzheimer’s that do not exist in the underlying biology. Advanced machine learning models can manufacture apparent signals where only noise exists.

This is not about AI being untrustworthy. It is about researchers failing to apply the same rigour to AI-assisted analysis they would demand of any other method. The Wisconsin–Madison group has introduced statistical corrections to restore validity. But the warning should travel further than it has.

Experts have emphasised the need for strict bio-governance, international safety protocols, data-breach protections, and rigorous statistical validation in AI-assisted genomic studies. The regulatory architecture does not yet match the pace of the science. That is a gap that should make any thoughtful researcher uncomfortable.

The Question We Actually Need to Ask

After reading the Evo2 paper, the accompanying Nature commentary, the UT Austin preprint, and the broader research landscape, a conclusion takes shape.

It is not about whether AI can write a genome. It has begun to do so. The more important question is whether the infrastructure for testing, validation, regulatory oversight, and biosafety can match the speed at which these models improve.

Hsu outlined a vision: autonomous laboratories pairing AI with robotics to iteratively design, test, and refine small genome segments, later assembled into working wholes. That vision is interesting. Whether equivalent investment is being directed toward validation infrastructure, statistical rigour, and governance frameworks is far less certain.

The real test of Evo2 will not lie in what the Nature paper reports. It will lie in how many of these designed sequences are physically synthesised, inserted into cells, and rigorously evaluated. Not 16 out of 285. At scale. With independent replication. With biosafety review.

That is what separates a promising model from a verified tool. Biology has never accepted press releases as evidence.

KEY TAKEAWAYS

Follow the experimental results, not the headline. A genome that looks plausible in computational predictions is not a functional organism. The 70% realistic-gene figure in the M. genitalium-inspired design is a milestone, not a finish line. Track what gets synthesised, tested in cells, and independently replicated.
Watch the nearer-term applications more closely than the synthetic-life headline. Operon design, cis-regulatory element synthesis, targeted gene therapy elements, and phage-based antimicrobials are where AI genomics will deliver verifiable results first. These deserve at least as much attention as the creation-of-life question.
Treat AI-assisted GWAS results with the same critical eye you apply to any other method. The Wisconsin–Madison warning about false positives in AI-augmented genomic datasets is a direct signal to researchers. Methodological rigour must keep pace with computational ambition.
Ask for specifics in policy discussions. “Bio-governance” and “international safety protocols” are easy to assert. Ask which bodies, what mandates, what enforcement mechanisms, and what timelines. Ceremony without accountability produces nothing.
Recognise the testing bottleneck as the rate-limiting step. The gap between what AI can design and what we can physically synthesise and test is widening. Investment in DNA synthesis infrastructure and high-throughput genome assembly is not peripheral to this field. It is central.

Claassens put it well: pairing AI with other approaches, including human insight, offers the best chance of success. That is not a limitation on the technology. It is a description of how science has always worked. The models can write. The harder question is whether we have built the systems to read what they have written — and act on what they say.

The best time to start building those systems is now. The second-best time is tomorrow.

ADDITIONAL READING

Brixi, G., et al. (2026). Genome modelling and design across all domains of life with Evo 2. Nature. https://doi.org/10.1038/s41586-026-10176-5
Gibson, D. G., et al. (2010). Creation of a bacterial cell controlled by a chemically synthesized genome. Science, 329, 52–56.
Goold, H. D., et al. (2025). Synthetic genome progress in yeast. Nature Communications, 16, 841.
King, S. H., et al. (2025). AI-designed phage genomes. Preprint at bioRxiv. https://doi.org/10.1101/2025.09.12.675911
Tzanakakis, A., Mouratidis, I., & Georgakopoulos-Soares, I. (2026). Structural analysis of Evo2 genome designs. Preprint at bioRxiv. https://doi.org/10.64898/2026.01.17.700093
Wiatrak, M., et al. (2025). Bacformer: bacterial genome generation model. Preprint at bioRxiv. https://doi.org/10.1101/2025.07.20.665723

Leave a comment

Your Brain Is Not Fixed: The Science of How Exercise Rewires Your Mind

Santhosh Eapen — Wed, 18 Feb 2026 08:40:51 GMT

A recent Nature article titled “Exercise rewires the brain — boosting the body’s endurance” reveals a profound discovery. University of Pennsylvania researchers found that regular treadmill exercise in mice doubled excitatory synapses in a key brain region. This increase in neural connections, which correlated with the amount of exercise, was fundamental to improved endurance. Without this brain rewiring, the mice failed to improve, demonstrating the brain’s coordinating role in endurance.

This finding underscores a critical, often unappreciated truth: the human brain is dynamic and adaptable, constantly restructuring itself based on daily actions. Exercise stands out as one of the most powerful, scientifically verified triggers for this change.

This shifts the focus from exercise as merely a strategy for weight loss or cardiovascular health to viewing it as a direct biological tool for reshaping core cognitive functions: memory, mood, focus, creativity, and resilience.

We are living through a mental health crisis. The World Health Organization identifies depression as the primary cause of disability worldwide, affecting over 264 million people (WHO, 2023). Anxiety disorders are among the fastest-growing health concerns globally. And yet, one of the most rigorously studied interventions for both conditions sits largely unused — not in a pharmacy, but on a footpath outside your door.

This article is not a fitness motivational piece. It is an evidence-based walkthrough of what exercise actually does to your brain, drawn from the latest neuroscience research, and why understanding this could change not just how you exercise, but how you think about your own cognitive potential.

The Brain Is Designed to Change — Here’s Why That Matters

The long-held 20th-century belief that the adult brain was permanent and irreversible is definitively wrong.

Neuroplasticity—the brain’s lifelong ability to reorganize itself by forming new neural connections—has overturned this old model. Your brain is constantly shaped by experiences, habits, and thoughts. This process is driven by three mechanisms:

Synaptic plasticity: Strengthening or weakening connections between neurons based on usage (the basis of learning and habit).
Structural plasticity: Physical changes, including the birth of new neurons (neurogenesis), new connections (synaptogenesis), and regional growth.
Functional plasticity: Redistributing cognitive tasks to healthy areas or reorganizing processing as skills develop.

Crucially, neuroplasticity is not passive; it is triggered by specific behaviours. Research confirms that exercise is one of the most potent known triggers.

BDNF — The Molecule That Makes It Happen

Exercise “rewires” the brain by boosting Brain-Derived Neurotrophic Factor (BDNF), a protein acting as “biological fertiliser” primarily in the hippocampus (memory/learning) and prefrontal cortex (executive function). BDNF promotes neuron survival, synaptic growth, and adult neurogenesis (new neuron birth).

Research confirms aerobic exercise elevates BDNF, enhancing synaptic plasticity, neurogenesis, memory, attention, and emotional regulation, while alleviating anxiety/depression.

BDNF: The Molecule That Makes It Happen Every step you take triggers a molecular cascade. BDNF is the signal your brain has been waiting for....

The optimal “dose” is moderate-intensity aerobic exercise (60–70% max heart rate) for 30–40 minutes, three to four times per week (brisk walk/slow jog). Maintaining healthy BDNF levels across the lifespan through consistent exercise also provides long-term neuroprotection, as declining levels are linked to neurodegenerative diseases like Alzheimer’s and Parkinson’s.

What Exercise Does to Your Brain, Region by Region

The effects of exercise on the brain are not vague or generalised. They are specific, measurable, and increasingly well-mapped. Here is what the research shows, broken down by brain region.

The Hippocampus (Memory/Learning): Highly sensitive to exercise, regular aerobic activity increases hippocampal volume. This reverses age-related shrinkage and improves spatial memory, directly resisting deterioration linked to Alzheimer’s disease.
The Prefrontal Cortex (Focus/Decision-Making/Emotional Regulation): Regular aerobic exercise measurably benefits this “executive command centre.” Older adults show improved executive function, including attention and problem-solving. This is key for managing demanding cognitive workloads.
The Default Mode Network (Creativity/Insight): Moderate exercise activates and strengthens the DMN, the network for mind-wandering and divergent thinking. This connection explains why novel ideas often arise during or after a walk.
Cerebral Blood Flow (Nutrition for Thought): Exercise increases the flow of oxygenated blood to the brain. This enhanced circulation delivers nutrients and removes waste (like beta-amyloid), leading to immediate mental sharpness.

Exercise and Mental Health — Beyond “Mood Lifting”

The mental health benefits of exercise go beyond mere stress relief, functioning as a significant clinical intervention. A landmark 2024 BMJ review (over 14,000 participants) found exercise, including walking, jogging, yoga, and strength training, as effective as antidepressants and psychotherapy for major depressive disorder.

The mechanisms are complex: exercise releases endorphins (reducing tension), increases serotonin, regulates cortisol (the stress hormone that inhibits neuroplasticity), and promotes BDNF (supporting hippocampal neurogenesis and having antidepressant effects).

Further evidence shows that both aerobic and resistance exercise significantly reduce depressive and anxiety symptoms, with mixed exercise being highly effective. Crucially, regular physical activity also prevents the incidence of depression and anxiety. A ten-year study even linked stronger handgrip strength (a proxy for fitness) to a reduced risk of developing these disorders, highlighting fitness as a long-term protective psychological factor.

While exercise is not a universal cure and is most effective for sedentary individuals, the evidence is unambiguous. For severe conditions, it works best as part of a comprehensive treatment plan including therapy and medication.

Starting, Sustaining, and Making It Work for Your Life

Understanding neuroscience is one thing. The harder question is what to actually do with it. Here is what the research recommends, translated into practical terms.

Starting, Sustaining, and Making It Work The post-exercise learning window isn't a productivity hack. It's biology. Use it deliberately....

Minimum Effective Dose: Just 20 minutes of daily physical activity can start positive brain changes. The optimal goal is 30–40 minutes of moderate aerobic activity (brisk walking, cycling, etc.) 3-4 times weekly. Consistency is more vital than intensity.

Post-Exercise Learning Window: BDNF is elevated for 1–2 hours after moderate aerobic exercise, creating a peak window for learning, memory, and cognitive tasks. Schedule demanding work for this time.

Mix Your Movement: Combine aerobic exercise (best for BDNF) with resistance training (elevates IGF-1 and myokines) and mind-body practices like yoga or Tai Chi (reduces cortisol, anxiety, and depression). This combination offers the broadest benefits.

The Role of Social Context: Exercising with others boosts mental health benefits through social connection, triggering positive neurochemicals like oxytocin and dopamine.

Age Is Not a Barrier: Neuroplasticity remains active well into old age. Consistency and challenge are the key variables, not age.

The Bigger Picture — Exercise as a Cognitive Investment

There is a framing problem with how most people think about exercise. We have been conditioned to see it as a physical maintenance activity — something you do to your body, for your body. Neuroscience tells a different story.

Exercise is a cognitive investment. Every session is deposited into a neurobiological account that pays dividends in memory, focus, emotional resilience, and creative capacity. Every week of consistent moderate activity is preserving and growing brain tissue. Consistent activity significantly lowers the risk of depression and cognitive decline.

The sedentary lifestyle of desk-bound, screen-heavy modern life is neurologically erosive. Chronically elevated cortisol from stress without physical release suppresses BDNF, starving brain tissue due to reduced cerebral blood flow. The absence of exercise does not simply mean the absence of benefits — it means the gradual accumulation of neurological cost.

This reframe matters because motivation to exercise consistently often falters when framed around aesthetics or cardiovascular health — both distant, abstract benefits. But understanding that a thirty-minute walk this evening will measurably improve your memory and mood is a more visceral and effective motivator than focusing on abstract aesthetics or long-term heart health.

Your brain is not fixed. It is waiting to be shaped. And the most powerful shaping tool available to you requires nothing more than movement.

Key Takeaways

Here are five things worth walking away with:

BDNF is the key mechanism. Exercise, particularly moderate aerobic activity, triggers BDNF production — the protein that drives the formation of new neurons and neural connections. Without it, neuroplasticity is significantly diminished.
The dose is achievable. 30–40 minutes of moderate aerobic exercise, three to four times per week, is the research-supported minimum for meaningful neuroplastic and cognitive benefits. You do not need high intensity or extreme duration.
The timing is leverage. The post-exercise window of one to two hours is your brain’s peak state for learning. Use it deliberately.
Exercise treats depression as effectively as medication. The 2024 BMJ meta-analysis confirmed this for major depressive disorder. For anxiety, the evidence is similarly strong. This is a serious clinical finding with serious lifestyle implications.
Start now, regardless of age. Neuroplasticity is active throughout the entire lifespan. The best time to begin is today. The second best time is tomorrow.

Additional Reading

Ben Ezzdine, I., et al. (2025). Physical activity and neuroplasticity in neurodegenerative disorders: A comprehensive review of exercise interventions, cognitive training, and AI applications. Frontiers in Neuroscience, 19, 1502417. https://doi.org/10.3389/fnins.2025.1502417
Kindel, M., et al. (2026). Exercise-induced synaptic plasticity in ventromedial hypothalamus SF1 neurons. Neuron. https://doi.org/10.1016/j.neuron.2025.12.033
Lenharo, M. (2026, February 12). Exercise rewires the brain — boosting the body’s endurance. Nature. https://doi.org/10.1038/d41586-026-00414-1
Noetel, M., et al. (2024). Effect of exercise for depression: Systematic review and network meta-analysis of randomised controlled trials. BMJ, 384, e075847. https://doi.org/10.1136/bmj-2023-075847
Romero Garavito, A., et al. (2025). Impact of physical exercise on the regulation of brain-derived neurotrophic factor in people with neurodegenerative diseases. Frontiers in Neurology. https://doi.org/10.3389/fneur.2024.1505879
Zhang, X.-Y., et al. (2024). Research status and trends of physical activity on depression or anxiety: A bibliometric analysis. Frontiers in Neuroscience, 18, 1337739. https://doi.org/10.3389/fnins.2024.1337739
Thanks for reading Research Rundown! Subscribe for free to receive new posts and support my work.
Share Research Rundown
Leave a comment

The End of Animal Testing? How iPSCs, AI, and "Clinical Trials in a Dish" Are Rewriting Drug Discovery

Santhosh Eapen — Wed, 11 Feb 2026 06:36:02 GMT

What if the biggest bottleneck in modern medicine isn’t a lack of ideas — but the wrong testing model?

Every year, pharmaceutical companies pour billions into developing drugs that will never reach a patient. The numbers are staggering: 92% of drugs that pass preclinical animal testing fail in human clinical trials. Cardiotoxicity alone kills 30% of promising candidates before they ever leave the lab. From 1953 to 2013, 462 drugs were withdrawn from the market — 63 of them due to cardiovascular complications that animal models simply couldn’t predict.

Now, a landmark review from Stanford University — published in the Annual Review of Pharmacology and Toxicology — lays out a roadmap for replacing this broken system. The tools? Human induced pluripotent stem cells (iPSCs), artificial intelligence, and a concept called “clinical trials in a dish.” Here’s what you need to know.

The Problem: Why Animal Models Keep Failing Us

Drug development is a decade-long, multi-billion-dollar process. In the United States alone, $28 billion is spent annually on preclinical research — the stage where compounds are tested in lab animals before they reach humans. Yet fewer than 10% of those compounds survive to approval.

The Species Gap

The root cause? Animal physiology simply doesn’t mirror human biology closely enough. Cardiac drug responses in mice, for instance, diverge sharply from human heart tissue. This “species gap” means that drugs deemed safe in animals can cause fatal arrhythmias, liver failure, or immune reactions in people. Phase II and III clinical trials — the most expensive stages — see failure rates of 22% and 35%, respectively, often because of toxicities invisible in earlier animal studies.

A Regulatory Turning Point

The FDA Modernization Act 2.0, signed into law in December 2022, formally acknowledges this problem. For the first time, the legislation permits pharmaceutical companies to use alternatives to animal testing — including iPSC-based assays, organoids, organ-on-a-chip technologies, and AI — to support drug safety applications. The Act doesn’t ban animal testing, but it removes the requirement that held back innovation for decades. And in December 2025, the Senate passed the FDA Modernization Act 3.0, further instructing the FDA to establish formal qualification processes for these new approach methodologies (NAMs).

> The Takeaway: The regulatory landscape has shifted decisively. For the first time, the law says: you don’t have to test on animals — if you can prove your alternative works.

iPSCs: Growing a Patient’s Heart in a Petri Dish

At the center of this revolution is a technology that won Shinya Yamanaka the Nobel Prize in 2012: induced pluripotent stem cells (iPSCs). These are ordinary adult cells — skin cells or blood cells — reprogrammed back into a stem-cell-like state using four transcription factors (OCT3/4, SOX2, KLF4, and c-MYC). Once reprogrammed, they can be coaxed into becoming virtually any cell type: cardiomyocytes (heart muscle cells), hepatocytes (liver cells), neurons, and more.

Why iPSCs Matter for Drug Testing

Unlike animal models, iPSC-derived cells retain the exact genetic profile of the donor patient. This means researchers can:

Model rare genetic diseases like hypertrophic cardiomyopathy (HCM) or dilated cardiomyopathy (DCM) using cells from actual patients carrying those mutations
Use CRISPR-Cas9 to introduce or correct specific mutations, creating isogenic controls for rigorous comparison
Build 3D organoids — miniature, self-organizing tissue structures that replicate the architecture and function of real organs far better than flat cell cultures

The Stanford team highlights cardiac organoids that mimic fetal heart development, including beating activity, metabolic responses, and drug-induced fibrosis. Researchers have already used these organoids to model myocardial infarction, test doxorubicin cardiotoxicity, and study heart defects caused by gestational diabetes.

From Single Cells to “Cell Villages”

One of the most innovative concepts is the “cell village” — iPSCs from dozens of genetically diverse donors, pooled together and exposed to the same drug simultaneously. This eliminates batch-to-batch variability and captures population-level diversity in drug response, mimicking what a real clinical trial would reveal, but in weeks instead of years.

> The Takeaway: iPSCs let researchers test drugs on human tissue — with the patient’s own genetics — without needing the patient, an animal, or a clinical trial.

AI Meets the Petri Dish: Smarter Screening, Faster Discovery

If iPSCs provide the biological substrate, AI provides the analytical engine. The review describes four domains where AI is already accelerating drug discovery:

1. In Silico Drug Screening

Before a compound ever touches a cell, AI algorithms predict its pharmacological potential. Graph neural networks (GNNs) like Chemprop model molecules as mathematical graphs — atoms as nodes, bonds as edges — to forecast properties like toxicity, bioavailability, and metabolic stability. Tools like ADMET-AI can evaluate thousands of compounds in hours, flagging dangerous molecules before they waste lab resources. Deep learning models like DeepDTA predict protein-ligand binding affinity, while DiffDock uses diffusion models to simulate how drugs dock to target proteins.

2. AI-Designed Molecules

Generative AI isn’t just screening existing compounds — it’s designing new ones. SyntheMol, developed at Stanford, navigates a vast chemical space of easily synthesizable molecules to propose viable antibiotic candidates. Multi-objective models simultaneously optimize for protein binding, drug-like properties, and synthetic accessibility. These AI-generated molecules are then validated on iPSC platforms.

3. Image-Based Drug Screening

Convolutional neural networks (CNNs) trained on iPSC-derived cardiomyocyte images can detect subtle morphological changes caused by cardiotoxic drugs. In one study, a CNN screened 1,280 bioactive compounds and flagged several FDA-approved drugs with documented cardiovascular concerns — demonstrating that AI can catch what traditional screening misses.

4. Measurement-Based Analysis

AI models analyze functional data — calcium transient signals in cardiomyocytes, cytokine responses in kidney cells, action potential waveforms — to predict toxicity at a resolution no human observer could match. Random forests, support vector machines, and CNNs have all been applied successfully to these datasets.

> The Takeaway: AI doesn’t replace biology — it supercharges it. By pre-screening compounds computationally and then validating winners on iPSC platforms, the combined approach collapses years of trial-and-error into months.

What This Means for India

India stands at a unique crossroads. The country is the world’s largest producer of generic drugs and has a rapidly expanding biotech sector — but its adoption of iPSC and organ-on-a-chip technologies lags behind the US, Japan, and Europe.

The Opportunity

India’s AI talent pool is among the deepest globally. The country’s IT infrastructure makes it well-positioned to adopt AI-driven drug discovery platforms at scale.
Genomic diversity is a major asset. India’s population represents a vast genetic landscape that is underrepresented in current iPSC biobanks, which are predominantly built from donors of European descent. Indian-origin iPSC lines would be invaluable for population-specific drug safety screening.
The India AI for Drug Discovery market is projected to grow significantly through 2035, transitioning from an experimental phase to a core component of pharmaceutical R&D strategy.

The Challenges

Infrastructure: Establishing iPSC biobanks, clean-room facilities, and organ-on-a-chip fabrication units requires significant capital investment.
Regulatory alignment: India’s CDSCO (Central Drugs Standard Control Organisation) has not yet adopted equivalent legislation to the FDA Modernization Act 2.0. Without regulatory incentives, pharma companies have little motivation to invest in NAMs.
Access: Many Indian research institutions lack access to high-throughput screening platforms and the computational infrastructure needed for AI-driven drug design.

The Path Forward

Indian researchers, biotech startups, and policymakers should watch these developments closely. Collaborative iPSC biobanks reflecting India’s genetic diversity, regulatory sandboxes for NAMs, and partnerships with global platforms like Evotec’s E.iPSC could position India not just as a generic drug manufacturer but as a leader in precision medicine.

What This Means

The Stanford review paints a picture of drug development that is fundamentally different from the one we’ve known for decades. The convergence of iPSC biology, organoid engineering, organ-on-a-chip systems, and AI-driven analytics is creating a pipeline where:

Drugs are tested on human tissue before they ever enter a human body
Patient-specific responses can be predicted from a single blood draw
Toxic compounds are eliminated computationally, saving billions in failed clinical trials
Diverse populations are represented through cell village models

But limitations remain. iPSC-derived cells still exhibit an immature, fetal-like phenotype rather than fully adult characteristics. Organ-on-a-chip systems are promising but not yet standardized. And no AI-discovered drug has achieved full FDA approval as of late 2025 — though several are in advanced clinical trials.

The question is no longer whether these technologies will replace animal testing. It’s how fast — and whether the regulatory, commercial, and research ecosystems can keep pace.

Your Spice Rack Contains 3,000 Bioactive Compounds - Here's What They Actually Do

Santhosh Eapen — Thu, 05 Feb 2026 10:28:45 GMT

Plants wage chemical warfare. Over millions of years, they evolved complex arsenals of bioactive compounds to defend against pests, pathogens, and environmental stress. When we eat these compounds in herbs and spices, we co-opt plant defense systems for our own biology.

The scale is staggering. A single herb or spice contains roughly 3,000 distinct phytochemicals. Yet according to FooDB, 85% of these compounds remain uncharacterized, untracked by national databases, unexplored through experimental research. We cook with molecular libraries we barely understand.

A 2025 study in npj Science of Food attempts to change that. Researchers used network analysis to map relationships between 1,094 herbs and spices, nearly 3,000 phytochemicals, and 1,597 health indications. Their approach: build bipartite networks connecting spices to their constituent chemicals and to their traditional medicinal uses, then infer which specific compounds drive which health associations.

The results validate some traditional knowledge, generate testable hypotheses, and reveal something unexpected about how cuisines evolve.

The Network Approach: From Correlation to Mechanism

Traditional ethnobotanical knowledge tells us garlic supports heart health and turmeric reduces inflammation. But these associations function as black boxes. They don’t explain which of the thousands of compounds in each spice produce which effects through what mechanisms.

The researchers addressed this by constructing three interconnected networks:

Spice → Phytochemical: Which compounds does each spice contain?
Spice → Indication: Which health conditions is each spice traditionally used for?
Phytochemical → Indication (inferred): Which specific compounds associate with which conditions?

They developed a “specificity score” to quantify how uniquely a given phytochemical associates with a particular disease. This filters out compounds that appear everywhere (and thus explain nothing) from those with targeted associations.

A simplified network diagram showing the interconnection of spices, their biochemicals and disease indications.

Validation was rigorous. Of their top 100 predicted phytochemical-disease relationships:

40% were confirmed by existing experimental literature
20% had been independently predicted through gene-chemical interaction databases (high confidence, not yet experimentally proven)
40% represent novel hypotheses for future testing

This hit rate suggests the method successfully distills signal from the noise of traditional knowledge.

Four Spices Under the Microscope

Garlic: The Hub

Garlic (Allium sativum) emerged as the most connected node in the spice-indication network, linking to an unusually broad range of health conditions.

The active chemistry centers on organosulfur compounds, particularly allicin (diallyl thiosulfinate). When garlic cells are crushed, the enzyme alliinase converts alliin to allicin. A 2024 review in Chinese Medicine summarized allicin’s cardiovascular mechanisms: it induces vasodilation via hydrogen sulfide signaling, reduces oxidative stress by upregulating glutathione, inhibits platelet aggregation, and modulates lipid metabolism.

Epidemiological data support these mechanisms. A systematic review found garlic consumption associated with reduced blood pressure, LDL cholesterol, and inflammatory markers. The effect sizes are modest (garlic isn’t a drug) but consistent across studies.

Turmeric: The Most-Studied Spice

Curcumin, the principal curcuminoid in turmeric (Curcuma longa), has been tested in over 400 clinical trials. A 2023 scoping review found that 75% of double-blind, randomized, placebo-controlled trials reported beneficial effects, primarily for metabolic disorders (29% of studies) and musculoskeletal conditions like osteoarthritis (17%).

The molecular targets are numerous: curcumin modulates NF-κB (the master inflammatory regulator), COX-2, STAT3, and multiple kinase pathways. This promiscuity is both its strength and its weakness. It affects many systems but achieves blockbuster efficacy in none.

The bioavailability problem is real. Curcumin absorbs poorly, metabolizes rapidly, and clears quickly. Plasma concentrations after oral dosing are often undetectable. Co-administration with piperine (from black pepper) increases bioavailability roughly 2,000% by inhibiting hepatic and intestinal glucuronidation. This is not folk wisdom but documented pharmacokinetics.

Traditional Indian cuisine pairs turmeric with black pepper routinely. Whether this represents accumulated empirical optimization or coincidence remains unknowable, but the biochemistry validates the combination.

An illustration depicting the major targets of four spices under the microscope.

Ginger: Multi-Pathway Anti-Inflammatory

Ginger (Zingiber officinale) contains gingerols, shogaols, and paradols, structurally related compounds with overlapping bioactivity. A 2024 review in Frontiers in Immunology detailed their mechanisms in immune cells:

Macrophages: Gingerols suppress LPS-induced cytokine production by inhibiting NF-κB and MAPK signaling
Neutrophils: Ginger compounds reduce reactive oxygen species generation and NETosis
Dendritic cells: 6-gingerol and 6-shogaol decrease TLR-triggered costimulatory molecule expression

Clinical studies suggest ginger’s analgesic effects rival NSAIDs for some conditions. The mechanism appears distinct from COX inhibition, instead involving desensitization of TRPV1-expressing nociceptors and modulation of substance P signaling.

Chilli: From Pain to Pain Relief

Capsaicin, the vanilloid responsible for chili heat, exemplifies a counterintuitive pharmacological principle: agonists can become functional antagonists through receptor desensitization.

Capsaicin activates TRPV1 channels on nociceptive neurons, causing initial burning pain. But sustained or repeated exposure produces “defunctionalization,” meaning loss of membrane potential, depletion of substance P, and eventual retraction of nerve terminals. A 2024 comprehensive review noted that high-concentration (8%) capsaicin patches carry FDA approval for neuropathic pain, with single applications providing relief for up to 12 weeks.

The network analysis found high specificity scores for capsaicin-family compounds (capsaicin, dihydrocapsaicin, capsorubin) with cluster headaches, diabetic neuropathy, and musculoskeletal pain, consistent with current clinical applications.

Regional Cuisines as Natural Experiments

Regional cuisines of India as natural experiments

The study’s second contribution analyzes 5,636 Indian recipes across 18 regional cuisines to understand how spice usage patterns relate to health indication coverage.

Geographic clustering emerged clearly. Principal component analysis revealed a North-South gradient: South Indian cuisines (Kerala, Tamil Nadu, Karnataka, Andhra) cluster together, characterized by curry leaves, coconut, and mustard seeds. Northern cuisines (Punjabi, Kashmiri, Mughlai) share saffron, cardamom, and clove.

Jain and Gujarati cuisines form a distinct cluster shaped by religious proscriptions against onion and garlic. These communities substituted asafoetida (Ferula assa-foetida), which contains diallyl disulfide, the same organosulfur compound responsible for garlic’s pungency. The substitution preserved flavor and potentially some bioactivity.

Health indication coverage varies by cuisine. Hyderabadi, Goan, Punjabi, and Mughlai cuisines showed the broadest coverage across disease categories. Notably, these cuisines historically absorbed influences from multiple cultures: Persian, Portuguese, Mughal.

The Fusion Hypothesis

Here’s the unexpected finding: randomly generated spice combinations outperformed traditional recipes in covering health indications with fewer ingredients.

The researchers used copy-mutate models to simulate culinary evolution, generating synthetic recipes and comparing the minimum number of spices needed to cover various disease categories. Random combinations required fewer spices than traditional regional recipes to achieve equivalent health indication coverage.

Why? Traditional cuisines optimize for multiple constraints simultaneously: flavor, ingredient availability, preservation, religious requirements, cultural identity. Health benefits, if considered at all, represent one factor among many. Random combinations face no such constraints.

This doesn’t mean traditional cuisines are poorly designed. It means they weren’t designed primarily for health optimization. The finding suggests that culinary globalization, the mixing of spices across traditions, may incidentally improve health coverage by filling gaps that no single cuisine addresses.

Bar chart comparing traditional vs. random spice combinations

Caveats Worth Noting

Cooking chemistry remains unaccounted for. Heat degrades some phytochemicals and enhances others. Cardamom and clove show increased antioxidant activity after cooking, possibly due to release of bound phenolics. The network analysis treats raw phytochemical content as a proxy for bioactivity, which introduces error.

Dose-response relationships matter. A 2018 Food and Chemical Toxicology review emphasized that phytochemicals exhibit dose-dependent effects: beneficial at dietary levels, potentially toxic at pharmacological doses. Capsaicin works as an analgesic topically but causes mucosal irritation orally. Curcumin is well-tolerated at grams per day but shows limited systemic bioavailability regardless.

Association does not equal causation. Network analysis identifies correlations and generates hypotheses. It does not prove that consuming a spice will produce a clinical effect. The 40% of predictions labeled “new hypotheses” are exactly that: testable guesses, not established facts.

Synergies and antagonisms remain unexplored. Phytochemicals don’t act in isolation. Piperine enhances curcumin absorption; other combinations may inhibit it. The network treats compounds independently, missing interaction effects.

What This Means in Practice

Diversify across traditions. No single cuisine optimizes health coverage. Incorporating spices from multiple culinary traditions broadens phytochemical exposure. If you cook Western food, add curry leaves. If you cook South Indian, try saffron.

Respect traditional pairings. Turmeric with black pepper isn’t arbitrary. Fat-soluble compounds (curcumin, capsaicin) absorb better with dietary lipids. Traditional cooking methods often incorporate these principles implicitly.

Prioritize whole foods over supplements. The health benefits of spices likely emerge from complex interactions among thousands of compounds. Isolating single molecules for supplementation may miss synergies and introduce imbalances. Curcumin supplements are also frequently adulterated.

Calibrate expectations. Spices are not drugs. Effect sizes in clinical trials tend toward modest. The value lies in cumulative, long-term dietary patterns, not acute therapeutic interventions.

A quick-reference cheat sheet summarizing actionable takeaways.

The Bigger Picture

This research represents a methodological advance: using network science to extract testable hypotheses from traditional knowledge systematically. The 40% validation rate against existing literature suggests the approach works. The 40% novel predictions offer a roadmap for future pharmacological investigation.

More broadly, the study illustrates how cuisines function as repositories of accumulated empirical knowledge. They are imperfect, constrained by history and culture, but they encode real information about bioactive compounds and their effects. As food systems globalize and traditions mix, we may inadvertently be running natural experiments in dietary optimization.

Your spice rack is a molecular library. We’re only beginning to catalog it.

References

Gao, Y., et al. (2024). Therapeutic potentials of allicin in cardiovascular disease. Chinese Medicine, 19, 93. doi.org/10.1186/s13020-024-00936-8
Guldiken, B., et al. (2018). Phytochemicals of herbs and spices: Health versus toxicological effects. Food Chem Toxicol, 119, 37–49. doi.org/10.1016/j.fct.2018.05.051
Kaur, R., et al. (2025). From phytochemicals to recipes: Health indications and culinary uses of herbs and spices. npj Science of Food, 9, 89. doi.org/10.1038/s41538-025-00458-z
Pázmándi, K., et al. (2024). The “root” causes behind the anti-inflammatory actions of ginger compounds in immune cells. Frontiers in Immunology, 15, 1400956. doi.org/10.3389/fimmu.2024.1400956
Perna, S., et al. (2023). Curcumin supplementation and human disease: A scoping review of clinical trials. Int J Mol Sci, 24(5), 4476. doi.org/10.3390/ijms24054476

Leave a comment

Black Pepper 2035

Santhosh Eapen — Fri, 23 Jan 2026 04:39:32 GMT

Pepper at the Crossroads

The esteemed spice, long celebrated for its distinctive flavor and myriad uses, now faces an increasingly precarious future. This pressing concern took center stage at the International Pepper Community’s Annual Meeting 2025 in Kochi, a pivotal event hosted by the Spices Board of India. Throughout the lively discussions and presentations, the shifting landscape of the pepper industry was thoroughly traced, revealing both formidable challenges and exciting opportunities.

Industry leaders at the forefront of agricultural innovation unveiled a series of bold possibilities, focusing on novel chemistry, functional molecules, and aroma intelligence, offering new possibilities for cultivation, processing, and product development. However, concerns were raised regarding increasing chemical dependency, the persistent issue of labor shortages, and fragmented value chains, which negatively impact the environment, health, and fair practices.

The conference concluded that pepper’s future relies on reinvention beyond mere yield. This means fundamental changes in practices, products, and processes, embracing innovation, exploring new markets, and addressing socio-economic and environmental issues for long-term sustainability.

The Ground Beneath the Vine

Chemical Dependency and Soil Fatigue: Heavy pesticide use over many years has led to the depletion of microbial life and reduced resilience in the environment. Regenerative practices are now focused on restoring biological health through the integration of biocontrol agents like Trichoderma, Pochonia chlamydosporia, organic amendments, and microbial consortia.

Climate Volatility: Erratic monsoons and warming trends are disrupting pepper’s rhythm. Climate-smart cultivation, including precision irrigation, AI-based disease forecasting, and resilient genotypes, must become standard practice.

Labour and Cost Pressures: Costly manual staking and harvesting can be improved through mechanization, women-led processing clusters, and digital farm management, which would reduce reliance on traditional methods and increase productivity.

Fragmented Post-Harvest Systems: India’s reputation in the export market is continually undermined by inconsistent drying and grading practices. Establishing cluster-level spice hubs equipped with shared dryers, sorters, and testing laboratories could provide the consistency that today’s markets require.

From Spice to Science

The recent deliberations held in Kochi have fundamentally altered the prevailing perception of pepper. Historically viewed primarily as a mere commodity, its significance has now been elevated. These discussions underscored a crucial shift in understanding, repositioning pepper as a vital and promising source of bio-innovation. This new perspective highlights its potential far beyond its traditional culinary and trade roles, opening avenues for research and development into its diverse biological properties and applications across various industries.

Functional Molecules: Piperine, piperyline, and caryophyllene are being utilized in nutraceuticals and cosmeceuticals due to their metabolic, anti-inflammatory, and antioxidant properties.
Green Extraction: Advancements in solvent-free and bio-transformation technologies are leading to the development of clean-label, highly efficacious ingredients.
Circular Economy: Utilizing pepper waste offers a dual advantage for plantations. It can be used for composting and biogas generation, fostering more sustainable, low-waste, and energy-efficient operations.

The message from innovators was clear—value lies in chemistry, not just commodity.

Sensing Tomorrow: Aroma Intelligence

“Aroma Intelligence,” a term that resonated throughout the recent meeting, perfectly encapsulates the transformative journey pepper is undergoing—a true sensory rebirth. This innovative approach moves beyond traditional classifications, delving deep into the very essence of pepper’s aromatic profile.

By meticulously decoding the volatile compounds that contribute to pepper’s distinct fragrance and taste, and subsequently linking this intricate chemistry with sophisticated consumer analytics, producers are now empowered to design bespoke flavors. This precision allows for the creation of pepper varieties and blends specifically tailored to the nuances of regional cuisines, catering to diverse palates and culinary traditions. Furthermore, it opens up unprecedented opportunities for expanding into lucrative export markets by delivering products perfectly aligned with international consumer preferences.

This paradigm shift represents pepper’s remarkable journey from its humble origins in the field to its sophisticated expression as a refined flavor. It is a testament to how cutting-edge data science, combined with a profound understanding of the human palate, is revolutionizing the world of spices. This is where data meets the palate, forging a future where every sprinkle of pepper is a carefully curated sensory experience.

Traceability and Trust

Consumers demand verifiable proof of quality, driving a global shift towards accountability in agriculture, influenced by ethical, environmental, and safety concerns. Advanced digital traceability systems, using QR codes, blockchain, and field-level data, meticulously track products. QR codes offer instant access to origin, blockchain ensures data integrity, and field-level data details cultivation and pesticide use. These systems provide unparalleled supply chain transparency, guarantee residue compliance, and bolster market credibility, giving businesses that prove product provenance a significant advantage.

Farmers benefit from recognition and fair returns, empowered by verifiable data that dismantles opaque pricing, linking quality and sustainable practices with financial rewards. Consumers gain profound confidence. For instance, pepper can be traced to its ethically sourced, pesticide-free farm. This transforms purchases into informed decisions, building trust and peace of mind. Digital traceability fosters transparency, trust, and fairness.

Convergence: The Critical Missing Link

Despite an optimistic outlook, Kochi highlighted the system’s weaknesses. The Spices Board successfully gathered international trade representatives, but the noticeable absence of crucial research institutions like ICAR–IISR and State Agricultural Universities exposed a fundamental structural disconnect. It is imperative that India’s pepper ecosystem overcomes its current fragmentation and works collaboratively.

Spices Board should evolve into a knowledge integrator, coupling regulation with real-time market and sustainability data.
ICAR–IISR and allied institutes must spearhead innovation in genomics, biocontrol, and precision agronomy.
Universities can translate science into training and entrepreneurship.
Private industry must drive investment in green processing and AI-based quality systems.
FPOs should anchor collective marketing and traceability at the grassroots.

Collaboration, not competition, is the true spice of the future.

A Shared Roadmap to 2035

The document outlines a comprehensive roadmap for the future of pepper, emphasizing a transformative journey rather than simply increased production - from a bulk spice grown in isolation to a science-driven, climate-smart, and high-value ingredient ecosystem. This strategic approach is crucial for the industry to adapt to evolving market demands, embrace innovation, and ensure long-term sustainability. By focusing on key areas, the roadmap aims to elevate the entire pepper value chain, from cultivation practices to market reach, ultimately redefining its global standing. The path ahead demands collaboration, innovation, and shared purpose across every tier of the sector.

A schematic diagram of the roadmap for black pepper

From Commodity to Conscious Ingredient

By 2035, black pepper has the potential to become a powerful symbol of India’s commitment to sustainable agriculture. This vision sees black pepper as a product that is not only traceable from farm to table but also deeply linked to health and wellness, thereby earning global respect and demand.

The realization of this renaissance, however, hinges on one critical factor: a shared vision. When scientists, dedicated growers, innovative processors, and forward-thinking policymakers align their efforts and move forward with a unified purpose, the world’s oldest traded spice can on

ce again rise to prominence. This collective synergy has the power to redefine the future of flavour, setting new global standards for quality, sustainability, and ethical production within the spice industry. This collaborative approach will ensure that every stage of black pepper production, from cultivation to consumption, adheres to the highest standards, solidifying India’s position as a leader in sustainable and responsible agricultural practices.

Five Strategic Shifts for the Pepper Industry in the Next Decade

[Insights from the 53rd Annual Meeting of International Pepper Community held at Kochi, India from 27-30 October 2025 and summarises my presentation titled ‘Next-Generation Crop Protection: Strengthening Regional Resilience and Sustainability in the Pepper Sector’. This article is published in the 2025 November issue of Spice India (Volume 38(09): 13-16)].

Leave a comment

Research Rundown

Prompt Like a Scientist

Set Up the Run

Control the Conditions

Reproduce Before You Trust

Why This Matters for Our Field

The Experimental Prompting Checklist

Further Reading

Ten Thousand Tongues, One Narrowing World

How do you count a language that left no trace?

Why settling down multiplied our tongues

And then the same revolution took them back

The lens from the plantation sector: crops did exactly this

What this means for us

Further Reading

AI-Powered Graphical Abstracts: Building Your Paper's Best Advertisement

Before anything else, read the journal’s rules

Decide the one thing you are saying

Match the tool to the job

Briefing the tool

Then check it properly

What has changed and what has not

The short version

Further reading

Which "AI Scientist" Should You Actually Use? A Field Guide for the Perplexed Researcher

What these tools actually are

The main players, and what each is good at

How to actually choose

Can you trust an AI scientist?

The catch nobody advertises

A note from where I sit

What to do this week

Additional Reading

Seven Beans, Two Centuries, One Narrow Gene Pool

What India actually grows, and what it has quietly stopped growing

The pests that drew the map

What Kodagu taught me

We have done this before, and we have forgotten

The collection we are sitting on

The smallholder is the whole game

What India should do now

Additional Reading

Claude Science: One Workbench, Not a Dozen Tabs

What It Is

Key Features

Installation

Workflows

Why It Matters Here

Getting Started Today

When the Cell Ignores Its Own Genetic Code

What “alternate decoding” actually means

The scale of the off-script proteome

Why the “wrong” protein wins

Not noise, not mutation — and conserved across species

Where this sits in the wider field

The honest caveats

Why it matters

What to watch next

Further Reading

The Worm That Makes Medicine

Why biologics needed a new delivery idea

The engineering: finding a safe place in the genome

Proof it actually works - and that it breeds true

Where this sits in the wider “living medicine” wave

The honest caveats

Why it matters

What to watch next

Further Reading

A New Era in Dementia: What the Latest Science Means for You and the People You Love

Nearly Half of All Dementia Cases May Be Preventable

Lifestyle Interventions: Structured Support Makes a Difference

A Blood Test That Can Detect Alzheimer’s Disease? It’s Here.

New Drugs That Slow the Disease — Not Just the Symptoms

A Note of Measured Hope

Additional Reading

Dark Proteome, Revealed: How Microproteins and Peptideins Are Rewriting the Human Protein Catalogue

Why This Matters - Even If You Work on Plants

The Scale of the Effort

The Annotation Framework: Tiers, Not Binary Calls

The New Entity: What Exactly Is a “Peptidein”?