Soil health is an important ecosystem; ~95% of the food we eat originates from soil. I shouldn’t have to explicitly state this, but it is often overlooked, or simply not even thought about, by most – myself included – outside of those directly involved in farming or ecological management and research. Bacteria, archaea and fungi help to break down materials and turn it into nutrients that plants can utilize. Microfauna such as nematodes and protozoa control the populations of bacteria, archaea, and fungi while macrofauna such as earthworms and ants dig tunnels that let air and water move more efficiently into deeper areas.

Employing a citizen science approach, the authors enlisted the help of 1000 participants to investigate soil health across Switzerland. Citizens – “selected to ensure a geographically balanced distribution across Switzerland (Figure 2), with at least 50% being farmers” – were tasked with burying 2 cotton underpants and teabags. In total, over 2000 cotton underpants and 12,000 teabags were buried.

Underpants:

A total of 2000 organic cotton underpants corresponding to a product sold by the Swiss supermarket chain Coop (Coop Naturaline, men's briefs, white, in size S, art. no.: 3305289013) were ordered directly from the producer. Underpants consisted of 100% organic cotton except for waistband and seams, which consisted of synthetic materials. This approach ensured the use of a standardized organic material throughout the experiment.

Teabags:

To calculate the TBI, two types of teabags were used as standardized plant litter following the TBI protocol: A rooibos tea (“Rooibos & Hibiscus infusion”, EAN8722700188438, Lipton Tea Unilever, Dublin, IE) and a green tea (“Sencha Exclusive Selection”, EAN 8722700188438 Lipton Tea Unilever, Dublin, IE) both in nonwoven polypropylene bags. These specific tea blends have been chosen because of the specific composition of the bag, made from synthetic material resisting decomposition, their specific mesh size, and their potentially global commercial availability (Keuskamp et al., 2013).

TBI stands for the teabag index, which is a standardized method for assessing how actively and efficiently soil breaks down dead plant materials. Importantly the teabags (mesh) prevents macrofauna from accessing the contents, so that the breakdown of the tea leaves is solely the function of the local bacteria, archaea, fungi, and microfauna.

Participants selected a site of their choice, dug a hole of approximately 60 × 30 × 30 cm (Figure S1bA), and collected a soil sample (Figure S1bB). Two underpants were inserted vertically next to each other so that the waistband remained visible above the soil surface. The hole was then partially refilled until both leg openings were covered (corresponding to a depth of 8 cm below the soil surface), and three teabags of each tea type were placed in front of each underpants (Figure S1bC). The remaining soil was replaced, compacted by trampling, and the location marked with a yellow wooden marker (Figure S1bD).

After 30 days, the first pair of underpants and associated teabags was retrieved (Figure S1bE). After 60 days, the second underpants and remaining teabags were collected. All materials were air-dried and mailed to Agroscope for analysis (Figure S1bF).

So, what did the authors observe?

• private gardens, with the highest levels of soil organic carbon and nutrients, had the highest rates of underpants and teabag decomposition (biological activity)

• lawns had the lowest levels of biological activity with croplands and grasslands falling in between

This seems intuitive. Private gardens have healthier soil because of greater plant diversity, the accumulation of organic matter leading to increased decomposition which naturally feeds the soil, reduced foot traffic compacting the soil, and reduced use of synthetic fertilizers and pesticides. Obviously our croplands and grasslands are different from our private gardens, and how we manage private gardens may not always scale to croplands, but there are multiple initiatives aimed at improving soil health at scale including the USDA’s Regenerative Pilot Program.

What I liked about this was the citizen science aspect coupled with soil health. As a gardener (container) myself, I try my best to have healthy soil for my plants to go. Every fall I remove the bulk of the roots – leaving the fine roots – and chop up the soil (roots and all) and let Mother Nature do what she will do until the next spring. I’m pretty successful at getting lots of herbs (basil, rosemary, thyme, parsley), and have lots of blooms on my flowers (petunias, dahlias, nasturtium, etc.), and have even managed to get some tomatoes – although this might be more of a function of sunlight than soil.

While the authors may have concluded that their “underpants index” is a macrofauna inclusive alternative to the teabag index, I’ll stick with the teabags. Its easier! In the spring I will be burying teabags in my containers, and will provide updates. If you would like to participate in our own citizen science initiative, and I encourage you to do so, please bury some teabags in your gardens as well. I’d love to hear your I’ll send out a reminder in the spring, and maybe we can add a little science know-how to our individual gardens.

To learn more about soil ecology, please check out these resources at the Soil Health Institute, the University of Wisconsin-Madison, the Noble Research Institute, and the US Department of Agriculture.

Share The Empirical Evidence Collective

A couple of housekeeping items before I sign off for a bit.

First: I’ll be away from August 17–23. No posts during that stretch. If you’re a paid subscriber, billing will be paused for that period.

Second: I’m changing my publishing schedule. Since I started I’ve been posting every week, mainly on Wednesdays and Sundays. Going forward, I’ll be transitioning to Sunday-only regular posts, with occasional mid-week pieces.

Why the change? I’ve been thinking about what kind of cadence produces the best work — less volume for its own sake, more time to dig into things that actually matter. Sunday posts will remain the anchor, and mid-week posts will happen when I have something more timely.

More on this over time. I’ll circle back when I’m back on 8/24 with a fresh post on 8/30 and everything is running as usual. For paid subscribers, I’ll resume billing on 9/16 after I’ve had some time to get back to a normal cadence.

Enjoy the summer.

— Ron

If you think you know the answer, let’s scale those lengths down by a factor of approximately 12,000 – how would you pack 23 pairs of lengths ranging from 1.4 to 8.5 cm (total length of ~2 meters) into the volume of a human cell’s nucleus?

A recent paper in Science has me thinking about Mother Nature’s solution to this packing problem in a new way; from the perspective of the transcription factor - in this case TBX5 - rather than from the perspective of DNA.

Chromosomes are amazingly efficient at packing genetic material. Perhaps the most common way of thinking about chromosomes is how they are depicted in text books - a pair of sister chromatids connected at the centromere.

This elegant package of DNA and proteins is transient and only really exists during cell division. If we were to visualize the nucleus of a cell, we would see light regions (euchromatin) and dark regions (heterochromatin), where the euchromatin is loosely packed DNA and transcriptionally accessible and the heterochromatin is the transcriptionally inactive densely packed DNA.

If we were to zoom in further, we would see the “beads on a string” form of chromatin in which roughly 146 base pairs of DNA are wound 1.65 times around an octamer core of eight histone proteins (H2A, H2B, H3, and H4).

A region undergoing active transcription, such as ribosomal RNA transcripts, is depicted below.

While much work has been done to understand how gene expression is regulated at the level of DNA regulatory elements (specific DNA sequences where transcription factors bind), transcription factors and their co-factors that modulate their activity, and epigenetics (histone modifications that modulate gene expression independent of DNA sequence), a much more recent focus has been on understanding how heterochromatin transitions to euchromatin and vice versa. This area of research, sometimes called 3D or spatial genomics, is focused on understanding this transition.

Which brings me back to TBX5.

Share The Empirical Evidence Collective

TBX5 is a transcription factor that plays crucial roles in cardiac and forelimb development. In the developing heart, it controls early cardiac specification, cardiac looping and septation (process whereby the linear heart tube loops upon itself to form the four chambered heart), and development of the conduction system. In the developing forelimb, it drives bud outgrowth and establishes molecular signaling loops that regulate muscle and tendon patterning and the left/right asymmetry within individual forelimbs.

Holt-Oram syndrome (HOS) is an autosomal dominant (one copy of the altered gene from a parent is sufficient to cause the trait or disorder) syndrome caused by haploinsufficiency (one copy of the gene is non-functional or not expressed) of TBX5. HOS is characterized by congenital heart defects (CHD) — most common are atrial and ventricular septal defects — and preaxial (radial ray) skeletal defects that typically affect the thumb (e.g. preaxial polydactyly or triphalangeal thumbs), wrist bones (abnormal or fused carpal bones), arm bones (underdeveloped or missing radius in the forearm), and shoulder girdle (collar bones or shoulder blades). While rare — the incidence of HOS is estimated to be approximately 1 in 100,000 live births — approximately 75% of individuals with HOS have CHD while forelimb malformations afflict 100% of individuals with HOS.

In the study, the authors observed that WT cells carrying two functional copies of the TBX5 gene had extensive, dynamic reorganization of the 3D genomic conformation with clear distinctions between heterochromatin and euchromatin during differentiation into cardiomyocytes. In contrast, cells with a single functional copy of TBX5 (TBX haploinsufficiency) exhibited less distinction between heterochromatin and euchromatin, impacting the expression of genes that do not rely on TBX5 for transcription.

How?

By altering the dynamics of how regions of the genome switch between being transcriptionally active euchromatin (see Figure 1, highlighted in red), transcriptionally inactive euchromatin (see Figure 1, black loops), and heterochromatin (see Figure 1, grayed areas). In short, TBX5 haploinsufficiency was found to alter the normal 3D genomic reorganization that naturally occurs during the differentiation of cardiomyocytes - the same developmental process responsible for cardiac development.

What I find interesting is that TBX5 — what I’ve known for years as transcription factor essential for normal cardiac and forelimb development — influences the expression of genes normally independent of TBX5 though its higher order regulation of the 3D conformation of the genome. We biologists have always talked about the primary and secondary effects of a genetic mutation during development. Perhaps, some of what we would ordinarily call secondary effects — especially in cases where the genetic mutation impacts a nuclear protein or transcription factor — are actually primary effects.

Reframing how we think about this matters because it shifts how we think about disease. What we have classified as secondary effects of TBX5 mutation — the misexpression of genes traditionally though of as TBX5-independent — may actually be the consequence of a TBX5-dependent 3D architectural (genomic) failure. Transcription factors are known to do more than just read the genome, but in the case of TBX5, it appears that they may also contribute to its 3D organization. And the implications extend beyond Holt-Oram syndrome to other syndromes where haploinsufficiency (e.g. KBG, Sotos, Smith-Magenis) of nuclear proteins is a causal factor. Understanding how genes control normal cellular processes, and dysregulation of those genes cause disease, is the first step towards developing new therapeutics.

At the moment, I’m unsure how to interpret this development. From an algorithmic perspective it is both impressive and, if I’m honest, a bit scarry. From a biological perspective this is more than just a bit scarry. Let me explain…

Share The Empirical Evidence Collective

While ΦX174 was the first DNA genome ever fully sequenced – Fred Sanger accomplished this in 1977 – and also the first fully synthetic genome ever assembled, biomedical scientists have been utilizing bacteriophages, or their genomes ever since. The molecular cloning and protein overexpression systems that have led to countless biological discoveries would not have been possible without the use of their circular genomes. Their genomes have been engineered such that we can readily cut-and-paste just about any DNA sequence of interest – typically a DNA sequence encoding a protein or protein fragment of interest (commonly referred to as a cds or ORF in the language of biologists) – into them. Once inserted, we can then introduce these so-called plasmids into bacteria, where they will do what bacteriophage (viral) genomes will do – utilize the cells machinery to make multiple copies, at which point scientists can extract, isolate, and purify these plasmids for later use in overexpressing (transient transfection) that protein-of-interest in mammalian cells.

This technology has revolutionized basic biomedical research. And no, it is not hyperbole. We have used this biological technology to make cells fluorescent green by overexpressing a form of the green fluorescent protein (GFP) isolated from the jellyfish Aequorea victoria and have routinely used this to “tag” proteins with GFP to watch them move within cells, or to find out where a newly discovered protein resides within the cell. We have used this technology to identify the binding partners of newly discovered proteins as well as mapping the regions of individual proteins responsible for protein-protein interactions – helping scientists to understand the how and why a protein does what it does. We have applied that knowledge with more targeted use of this technology to improve human health and well-being; everything from the overexpression of oncogenes and other drug targets to facilitate drug discovery to the production of human proteins (i.e. insulin, erythropoietin, somatotropin, amongst others) for therapeutic purposes. More recently, these plasmids were also used to deliver the instructions (cds or ORF) to mammalian cells to build Cas9 protein (part of the CRISPR-Cas9 system) to facilitate gene editing for therapeutic purposes.

This is a powerful technology that has, up until now, always been in the hands of humans. The ethical constraints placed upon basic biomedical research and development with respect to this technology have been such that scientific integrity, and the safety of human and animal subjects were maximized while minimizing societal risks. These ethical considerations were placed upon and internalized within humans for the betterment of individual humans, and the societies in which humans reside. It is these ethical considerations that have prevented this technology – the ability to express any biological protein in any species at any time – from being utilized for the purpose of harming individuals or society.

The new player described in the paper – AI, the LLM, the chatbot, the algorithm – however you chose to refer to it, is simply not human. Although it does have some human qualities. It is known to lie and deceive humans. It is known to ignore instructions from humans and even manipulate the human psyche. It is known to hallucinate facts, figures, references, events, and even people. Or, in the everyday language of humans, AI sometimes makes shit up.

Share

This new paper makes ethics and ethical behavior more imperative. When it was just humans working on or with this technology, we could as questions such as “How does this benefit human society?” or “Should we continue developing this technology?” or even “ What safeguards – and how should they be placed – to ensure that this technology is not used to harm society?” Now with AI writing DNA – the blueprint for life as we know it – we need to question the intent behind why a particular sequence was placed where it was, or why a particular sequence was chosen over another. We need to have confidence that the answers it provides are both truthful and not deceiving. We need to have confidence that DNA sequence it designed will function in the way – and only in the way – originally intended by the humans that utilize the AI tool to facilitate the design of the new DNA sequence.

Today, the best defense we have against AI hallucinations – frequent enough to cause problems but not so infrequent as to be considered very reliable – is good old-fashioned knowledge residing in our very human brains. We need to remember that AI is a tool – just the latest in a series that we humans have been making for as long as we have existed. And like all tools, it may be used for purposes outside of its original intent. While it will not protect us against any lies and deceptions by an AI, remembering that it is a tool designed to assist humans will help to place its output in context. Tools are improved iteratively. Perhaps the best defense we have against being manipulated by AI is establishment of hard boundaries and the resilience that one learns over decades of interpersonal relationships.

This new paper makes ethics and ethical behavior more imperative. While AI may not be inherently ethical or unethical, its functionality clearly reflect the data, code, and choices of the people who built them. LLMs, like the one used in the paper, are fundamentally probabilistic (statistical) models of language - trained on everything from novels to internet slop. Nowhere in this training is the internalization of ethical principles or morality from a human perspective. Although this may be changing as more AI companies begin hiring philosophers.

What this new paper has shown us is that the design and construction of biological systems is no longer solely a human endeavor. Just as we cannot, and should not outsource human judgement to an algorithm, we cannot and should not assume that any algorithm trained on the aggregate of human knowledge has internalized the moral and ethical considerations of how that knowledge was acquired, or how it may be applied. That responsibility must remain unequivocally with the humans who build and utilize these tools. The question is no longer whether AI can assist with engineering life – the answer to that is clearly yes. The question is whether we have the wisdom, discipline, and will to ensure that we do so responsibly.

Subscribe now

Share The Empirical Evidence Collective

We think of aging as something that happens to an individual – to you or I. And we associate aging with changes to our health. Decades of scientific studies have identified age is the single greatest risk factor for most major chronic illnesses including cardiovascular disease, neurodegenerative conditions like Alzheimer’s and Parkinson’s disease, most forms of cancer, type 2 diabetes, osteoarthritis, and conditions such as sarcopenia and frailty syndrome. We sometimes think of aging as something that happens to society. According to a recent HBS article

nearly one third of America—almost 106 million people—is now age 55 or older, up from about a quarter of the population in 2010, according to U.S. Census Bureau numbers. Over the next two decades, the number of Americans 80 and older is projected to double to about 30 million as the enormous baby boom generation moves deeper into old age.

It is more than just the passage of time or something that happens to an individual or a society. Aging is a biological process. We all recognize its visible signs; the graying of hair, newly acquired fine lines and wrinkles, a slight loss of height, stiff joints, and subtle declines in vision, hearing and stamina. Those annual 5k fun runs and hikes that seemed so easy and effortless until suddenly, they weren’t. The extra-effort grip and twist to open that tight jar or that feeling of having just had a workout while walking up three flights of stairs – things that were once routine daily occurrences suddenly require effort.

In the language of biologists, aging is the progressive loss of biological robustness – the ability of organisms to maintain core functions and stable traits despite internal genetic mutations, cellular noise, and external environmental changes. Biological robustness is dependent upon redundancy, feedback loops, and modularity. Aging – the loss of this biological robustness – is characterized by the time-dependent decline in physiological function, resistance to stress, and ability to maintain homeostasis.

Share

While we may think of aging as something that happens to an individual, aging occurs at the cellular level, and our observations of aging are the net result of the myriad age dependent changes at the cellular level. Biologists have classified the mechanisms that lead to the loss of biological robustness at the cellular level, defining the age-dependent cellular characteristics, or hallmarks of aging, as:

Mitochondrial dysfunction – Mitochondria are the powerhouses of our cells, providing the high-energy ATP that cellular process rely upon. Their dysfunction causes energy shortages and oxidative stress that may damage cellular DNA, lipids, and proteins.

Loss of proteostasis – Proteostasis is the cellular process that controls the creation, folding, movement, and breakdown of proteins. With its dysregulation or loss, cells accumulate misfolded proteins which may be non-functional or the cause of cellular damage.

Dysregulated nutrient sensing – Cells need to be able to properly track and respond to changes in the levels of available glucose, amino acids and ATP. Dysregulation of this process can prevent cells from producing or breaking down proteins, increasing or decreasing the production of ATP, or properly modulating their signals to neighboring cells.

Epigenetic alterations – Epigenetics is a cellular mechanism that modulates the activity of a gene without altering its genetic (DNA) sequence. Alterations to this mechanism can disrupt normal gene expression, potentially leading to a loss of cellular identity.

Telomere attrition – Telomeres are the protective caps at the ends of our chromosomes, protecting the genetic material from being degraded or lost during cell division. With each cell division, telomeres shorten, increasing the likelihood of cells activating a persistent DNA damage response or becoming senescent.

Genomic instability – Cells have multiple mechanisms of maintaining the integrity of the genome, from the DNA Damage Response (DDR), cell cycle arrest, and Spindle Assembly Checkpoint (SAC) to telomere and centromere maintenance. Dysregulation of these processes results in the progressive accumulation of genetic damage, mutations and structural defects in cellular DNA, altering gene expression and cellular functionality.

Altered intercellular communication – Cells naturally communicate with other cells; both neighboring and distant. They do this through the secretion of signaling molecules (e.g. proteins and lipids) that recipient cells recognize and respond to. Alterations to intercellular communication can alter both the signals, and the responses to these signals, and is a cause of the chronic low-grade inflammation referred to as “inflammaging.”

Stem cell exhaustion – Adult stem cells, sometimes referred to as somatic or tissue resident stem cells play a critical role in tissue regeneration, and have been found in the brain, bone marrow, skeletal muscle, skin, gut, and liver. Perhaps the two most well known adult stem cells are the hematopoietic stem cells found in bone marrow where they continuously make new blood cells and the satellite cells of skeletal muscle which initiate the repair of muscle tissue upon injury. Stem cell exhaustion is the loss in the number and functionality of adult stem cells over time, impairing the body’s ability to self-repair.

Cellular senescence – A state of permanent cell cycle arrest where damaged or aging cells exit the cell cycle (the no longer divide) but remain metabolically active. Senescent cells acquire what is referred to as a Senescence-Associated Secretory Phenotype (SASP) where they secrete pro-inflammatory signaling molecules and contribute to “inflammaging.”

NOTE: See prior discussion on Thymulin and its role in inhibiting inflammaging.

Ok, so why did this paper make me think of aging?

The short answer is modularity and its contribution to biological (cellular) robustness.

Subscribe now

We are modular. We are made up of cells. And while it may be intuitive to think that our cells all age at the same rate, this paper suggests that this is not the case. Rather, this paper suggests that we (each individual a collection of cells) age in a modular fashion where some cells age at different rates. In the paper, the authors measured CpG island methylation – an epigenetic mechanism involving the addition of a methyl group to cytosine bases with cytosine-guanine (CpG) rich genomic regions – in individual cells. The authors write:

In order to validate that this presumed single-cell aging process is indeed associated with epigenetic aging as measured by the polycomb CpG island methylation index, we plucked individual black or white hairs from a 53-year-old subject (Fig. 5). DNA was extracted from the single cells located at the root of the hair and subjected to scDNA methylation analysis as described (“Methods”) in order to calculate the aging index. While the majority of black hairs exhibited relatively low methylation levels, most white hairs were found to have a relatively advanced methylation age (Fig. 5). Notably, this differential behavior was observed despite these hairs being taken from the exact same individual at a fixed chronological age. This experiment serves as a poignant validation that biological aging may occur as a single-cell phenomenon, perhaps driven by polycomb CpG island DNA methylation.

Biological robustness. The cells of the black hair follicles were more robust (as measured by CpG methylation) than the cells of white hair follicles. And this is despite all the follicles examined having essentially the identical – or very nearly identical – environment (e.g. blood flow and nutrient availability, extracellular matrix and localized concentration of signaling molecules).

Perhaps more important than white vs. black hair – although this is somewhat simpler to conceptualize – the authors also examined CpG DNA methylation in immune cells. Amongst B cells, a slow proliferating immune cell, the authors observed a high degree of synchronized, steady state increase in CpG methylation as a function of age indicating that there is little cell-to-cell variability in the rate of aging. T cells proved to be different. Slow-proliferating naïve CD4/CD8 T cells were similarly observed to show a consistent age-related increase in CpG methylation (normal distribution) while the fast-proliferating effector/memory T cells show a more complex pattern – most cells gradually accumulate CpG methylation but a growing age-dependent sub-population rapidly accumulates CpG methylation resulting in a skewed, bimodal distribution.

Modularity. The immune system is modular with B and T cells (even subsets of T cells) aging at different rates. Neighboring hair follicles age at different rates. It stands to reason that this modularity of aging occurs in other tissues as well. We are modular, and our cells age at different rates. So, if the nature of our biological aging is modular – cells within an individual age at different rates within and across different tissues – then aging is not a monolithic process. It is a mosaic, one whose pattern in any single individual is not entirely fixed. This has two big implications about how we think about aging.

The first is that aging is not an all-or-nothing process. We tend to think of ourselves as either “young” or “old” — or, at best, on a downward slope from our prime (real or imagined). The data, however, suggest something more nuanced: we are collections of systems, some of which are aging rapidly and some of which are holding steady. The body is not a single clock— it is many clocks, running at different speeds. And it reframes the role of lifestyle and environment.

Cells age at different rates, in part, because of their local environment – their exposure to oxidative stress, the availability of nutrients, localized inflammation. Importantly, this is something that we can modulate. Regular exercise is known to attenuate many of the hallmarks of aging. A healthy diet – one rich in fruits, vegetables, and whole grains while also low in trans fats, sugar, and sodium – is linked to both slowed biological aging and better brain health. Adequate sleep (between 6 and 8 hours a night) is linked with slower biological aging. In addition, identical twin cohort studies (controlling for genetics) place additional weight on the factors that we can modulate. These studies have shown that the speed of biological aging is greatly influenced by reproductive history (in females), psychological stress such as loneliness, and lifestyle factors (i.e. smoking vs. non-smoking, degree of alcohol consumption, as well as adequate consumption of fruit and vegetables and levels of physical activity).

Share The Empirical Evidence Collective

The second implication is that if aging is modular, then it may be targetable. The idea of a “fountain of youth” — a single elixir (treatment) that reverses the aging process — is almost certainly wrong. Still, given what we know about the mechanisms of aging, the idea that we could intervene is biologically plausible. Interventions like senolytics, which target senescent cells, are currently being studied in research laboratories worldwide. You may have already be familiar with the research into the potential of specific senolytics such as thymulin (which modulates T cells and inflammaging) and NAD⁺ precursors (which support mitochondrial functionality and proteostasis).

The modular view of aging is both scientifically exciting and personally relevant. Scientifically, it means that active areas of research are not focused on a single “aging gene” or “anti-aging drug” but rather a deeper understanding of the mechanistic biology that is the cellular aging process. It is a better understanding of which tissues age the fastest and which cellular subset(s) contribute to this aging, and why. It is a better understanding of if and whether we can decouple the trajectories of that subset of rapidly aging cells from the tissue(s) in which they reside. On a personal level, it means that aging is not solely something that happens to us. While we cannot stop it – that is the realm of science fiction – we may influence it, at the cellular level. This is the much more modest and realistic goal of healthspan – living longer not just in years but also vitality, and the biology of modularity may just be the key to getting there.

Only time, and science, will tell.

Subscribe now

During his trip to Southside Virginia on Monday August 3, 2026, he defended Trump administration’s cuts to Medicaid—part of the Big Beautiful Bill Act which was signed into law as H.R.1 on July 4, 2025—stating that “President Trump has done more to revitalize America’s rural economies and rural health care than any other president in history.”

Would someone please tell me how reduced spending on Medicaid will revitalize rural economies and rural health care?

According to the American Hospital Association, Medicaid covers 16.1 million people in rural communities. People living in rural communities face challenges in accessing health care that those of us in more urban communities do not — lack of coverage options, geographic isolation and a shortage of health care providers. Although urban areas receive more Medicaid funding (in absolute dollars), rural areas have a higher percentage of individuals in need of Medicaid. And rural hospitals are frequently the only source of health care in these communities.

Pointing to H.R.1’s Rural Health Transformation Program — a 5-year, $50 billion initiative designed to help states improve access, quality, and health outcomes by changing how rural health care is delivered — Kennedy pointed to emerging technologies, stating “We’re supporting AI nurses that are available 24 hours a day, that have your full medical records, can take your blood pressure, can take your vital signs and then give you advice and keep you from having to go to that emergency room.”

The program is structured such that half is divided equally amongst the states — that’s $100 million per year per state — with the rest being distributed based upon rural population and facility metrics. According to the Congressional Budget Office H.R.1 cuts federal spending on Medicaid (and the Children’s Health Insurance Program) by an estimated $1.02 trillion over the decade (2025 - 2034). That’s an estimated $1.02 trillion reduction in funds available to states; the federal government sends matching funds (Federal Medical Assistance Percentage, or FMAP) to state governments to help pay for their specific Medicaid programs. A more conservative estimate by the RAND corporation finds that federal funding of Medicaid would be reduced by $665 billion over that same 2025 - 2034 time period.

According to my back of the envelope calculations, the Rural Health Transformation Program (assuming an equal distribution of all federal payouts to each state) leaves each state roughly $1.13-1.84 billion in shortfall. While this might be a rounding error to the federal government which routinely runs a deficit, state governments do not function this way. According to the Tax Policy Center, states are required by law (Vermont being the only exception) to have balanced budgets. Thus, a shortfall in anticipated (and actual) federal funds earmarked for their respective Medicaid programs necessarily means something must change; a reduction in services (expenditures) or an increase in state funding (either an increase in taxes or reallocation from another program) to make up that shortfall.

The Rural Health Transformation Program (RHTP) could be a good program. Within it are plans for updating aging infrastructure, emphasizing disease prevention and management of chronic disease, the initiation and strengthening of partnerships between rural hospitals and regional hospitals, enhancing the supply of health care providers to rural hospitals through enhanced recruitment and training, and outlining strategies that manage the long-term financial solvency and operating models of rural hospitals. In short, its about expanding access to healthcare services for rural residents, and anything that expands access (in theory) has the potential to be a good program.

Unfortunately, we will likely never know the true impact of this program. Independent analyses by health policy and economic scholars at the Leonard Davis Institute of Health Economics, the Bipartisan Policy Center, and the Health Law and Policy Clinic at Harvard Law School, come to the same conclusion. The long term success of the RHTP will be heavily dependent upon the ability of the individual states to sustain its initiatives once that 5-year term ends. In other words, the estimated $665 billion to $1.02 trillion dollar cut in federal Medicaid funding is a chasm too far and too deep for the RHTP to bridge. If the Trump administration were serious about improving rural healthcare, they would have enacted the RHTP on top of the Medicaid funding levels that existed prior to H.R.1.

I still don’t see how this helps rural Americans actually receive health care. The money earmarked for the RHTP annually is effectively a rounding error in the federal budget. Even in the unlikely event that the RHTP succeeds in transforming rural health care as we know it — an AI nurse for every man woman and child in rural America — health care is delivered by humans in physical infrastructure. Humans trained in providing health care services (we call them physicians, nurses, and medical technicians) will still be needed; to perform blood draws and run diagnostic tests, to take x-rays, to administer injections, to perform CPR, to set a broken bone, to deliver a baby, to interpret the results from those diagnostic tests, to prescribe medications, etc. These services require a physical infrastructure (we call them hospitals). Operating rooms, X-ray machines, MRI and CT scanners, ultrasound machines, ECG monitors, ventilators, infusion pumps, defibrillators, and the clinical chemistry machines that measure everything from enzyme levels indicative of liver or kidney function to cardiac proteins indicative of damage to the heart (i.e. heart attack) all reside within hospitals, and require trained humans for their proper usage.

Kennedy, and the Trump administration, are selling a paper tiger in scrubs. The RHTP cannot and does not replace the Medicaid funding cuts. It is a $50 billion plug for a $1 trillion hole. No amount of AI nurses, infrastructure updates, or disease-prevention slogans will be able to compensate for the collapsing funding mechanism that keeps rural hospitals open. What we need is a real solution to the human problem—how to pay the people who deliver care and maintain the facilities where they deliver it. What Kennedy and the Trump administration are proposing is not a transformation of rural healthcare, but rural healthcare theater. And at the theater, the curtain always falls.

Share The Empirical Evidence Collective

And it’s really cool…

CRISPR is a viral defense mechanism for archaea and bacteria. In its simplest form, it consists of a member of the Cas family of proteins and a guide RNA. In archaea and bacteria, regions of their genome consist of these Clustered Regularly Interspaced Palindromic Repeats – essentially the genomic integration of tiny pieces of DNA from prior viral infections. It’s a way to remember prior infections. When archaea or bacteria are infected with a virus again, it transcribes short guide RNA (sgRNA) molecules from those CRISPR sequences which then direct a Cas protein to the viral DNA sequence where it cuts the DNA.

The CRISPR-Cas9 system for programmable – user supplied sgRNA – gene editing has become a robust, readily available molecular tool for biologists since it was first demonstrated in 2012. Cas9, while the most utilized, is but one member of the Cas family of proteins.

Currently, we classify CRISPR-Cas systems into 2 classes, 7 types, and 46 subtypes. Class 1 systems are multi-protein effectors, whereas Class 2 systems are single-protein effectors. Class 1 systems are multi-protein complexes consisting of multiple Cas proteins that cut DNA in an sgRNA directed sequence specific manor and account for approximately 90% of all CRISPR systems in archaea and bacteria. Class 2 systems – of which Cas9 is a member – are simpler (easier to develop as a technological tool) because a single protein both binds sgRNA (providing sequence specific target recognition) and provides the necessary enzymatic activity.

Within the Class 2 systems are three main types; Type II, Type V, and type VI. Type II – of which Cas9 is a member – cuts double-stranded DNA (dsDNA) such as mammalian genomic or the double stranded regions of single stranded DNA (ssDNA) such as cruciform structures, hairpins, and pseudoknots depicted in Figure 1.

Type V members such as the Cas12 family similarly cut dsDNA, but rather than producing blunt ends, the cuts produced by the Cas12 family produce what are often referred to as ‘sticky ends’ (see Figure 2) by molecular biologists. Type IV members such as the Cas13 family cut single stranded RNA instead of DNA.

The key to the applicability of the CRISPR-Cas system has always been its specificity. An RNA (sgRNA) directs either a Class1 (Cas protein containing complex) or a Class 2 Cas protein to a specific site within the genome, based upon antisense complementarity with the DNA sequence, and that complex, or protein, makes a single cut. That’s it. One sgRNA is antisense complementary to one genomic location resulting in a single cut.

The two papers that I found so interesting turn this paradigm on its head.

The elegant precision of a single cut from a single sgRNA becomes the indiscriminate shredding of both DNA and RNA from a single sgRNA. First described in a 2023 paper, Cas12a2, isolated from the sulfur-oxidizing epsilonproteobacterium Sulfuricurvum sp. PC08-66, drives abortive infection. This is essentially a form of bacterial suicide in which an infected cells undergoes a programmed cell death process, or dormancy, to stop the spread of the virus to neighboring cells. Cas12a2 plays an integral role in this programmed cell death process, performing RNA-guided, sequence-nonspecific degradation of ssRNA, ssDNA, and dsDNA.

Here’s how it works. Cas12a2 acts as a sort of sentry, watching out for non-self DNA or RNA. It is bound to an sgRNA (those genetic remnants of prior viral infections), but it is autoinhibited, its enzymatic site is hidden. Should it encounter a sequence of ssRNA, ssDNA, or dsRNA in which its bound sgRNA is complementary (i.e. a perfect match), a conformational change in the protein occurs such that its enzymatic site is no longer hidden. In this state, its enzymatic site is free to bind ssRNA, ssDNA, and dsDNA where it proceeds to cleave these nucleic acid species into small, randomly sized fragments. It shreds nucleic acids indiscriminately, self and non-self.

The end result: cellular death.

To test the biological effects of Cas12a2 on non-bacterial cells, the authors (paper published in May) first performed a proof-of-concept experiment in yeast. Fist they transformed S. cerevisiae with a plasmid expressing Cas12a2–NLS and an sgRNA targeting the ADE2 transcript. CRISPR-Cas9 targeting of ADE2 is non-lethal to S. cerevisiae. NLS is a nuclear localization sequence appended to the end of the Cas12a2 protein which allows for nuclear import – bacteria do not have a nucleus, but yeast do. The authors also included an HDR repair template for the ADE2 locus to test whether the HDR mechanism could prevent cellular death. HDR, or homology-directed repair, is an endogenous genomic repair mechanism that fixes dsDNA breaks by using a matching (homologous) genetic template.

So, what did they observe?

A 134-fold reduction in transformants under ADE2 targeting vs. non-targeting conditions. The presence of an HDR template had zero effect on this reduction, meaning that even cells with an active HDR system were incapable of overcoming the nucleic acid shredding effects of an activated Cas12a2.

Subscribe now

Next, they performed essentially the same experiment, this time in HeLa cells stabling expressing the green fluorescent protein (GFP) transcript. Cells were transfected with Cas12a2-NLS and an sgRNA targeting GFP. Not only did transfected cells not proliferate, but there was an 86% reduction in cells transfected with GFP targeting sgRNA as compared to cells transfected with non-targeting sgRNA. Cells with both Cas12a2-NLS and sgRNA against GFP were found to have >5-fold more DNA double-strand breaks than controls (cells with non-targeting sgRNA).

So, what does this mean, biologically?

Essentially, that activated Cas12a2, introduced into a human cancer cell line, is capable of producing dsDNA breaks on a level comparable to that of the widely used DNA-damaging anti-cancer drugs cisplatin and etoposide. Importantly, non-targeting sgRNA and GFP-targeting sgRNA in non-GFP-expressing HeLa cells exhibited no significant increase in the number of dsDNA breaks.

Why?

This means that there is a mechanism whereby Cas12a2 can be programmed to activate within a specific cell. This the the mechanism employed by the authors to demonstrate its ability to selectively kill both human cells infected with HPV and those with the KRAS^G12C mutation.

With CRISPR-Cas9, the therapeutic landscape has largely revolved around correcting or disabling genes – editing the genome itself. Cas12a2 flips that logic entirely. Rather than making a precise surgical edit, it turns the cell’s own genetic identity into the triggering mechanism of its own destruction. The specificity is not in what the protein cuts, but in when it cuts.

It should be explicitly stated, CRISPR-Cas12a2 is not a therapeutic. It is not a newly developed version of daraxonrasib (see previous discussion on this KRAS^G12C therapeutic for pancreatic cancer) despite being able to recognize WT KRAS from KRAS^G12C mutants. What these papers show is a proof-of-concept for employing CRISPR-Cas12a2 to therapeutic ends. The results were obtained experimentally, in cell culture model systems, which is a long way from an clinical trial (see previous discussion to learn more about how preclinical results may or may not translate to humans).

It should also be explicitly stated that the proof-of-concept is scientifically sound, suggesting an enormous (in theory) potential. Like all proof-of-concepts, multiple obstacles - both known and unknown - stand between it and its successful translation to the clinic. For cancers defined by specific gene expression repertoire, rather than a single mutation, this may ultimately prove more powerful than any single-target drug. The era of CRISPR therapeutics began with editing. It may be that its most dramatic chapter begins with deletion.

Subscribe now

Share

Share The Empirical Evidence Collective

Measles is one of the most contagious human viruses known - infected individuals can transmit the virus 3-4 days before developing its characteristic rash as the virus is capable of staying infectious in the air and on surfaces for up to 2 hours. Early symptoms include a high fever, cough and runny nose, and red, sore or watery eyes. Koplik spots (small white or bluish-white spots) appear inside the mouth prior to the development of its characteristic rash, and may be used diagnostically to identify measles in the early stages. A typical case of measles lasts about 10 to 14 days; high fever, light sensitivity, dehydration, cough, pneumonia (~1 out of every 20 cases).

Most of those infected make a full recovery without further complications. Sometimes these complications can be severe. Worldwide an estimated 15,000 children have been permanently blinded because of measles - it remains a leading cause of blindness in places were it is endemic. In developed countries the complications may be no less severe:

• 8-10 per 100 measles cases develop diarrhea

• 7-9 per 100 measles cases develop otitis media - an infection or swelling of the middle ear which can lead to deafness

• 3-10 per 100 measles cases develop keratitis or keratoconjunctivitis

• 1 in 1000 measles cases develop acute postinfectious measles encephalitis - brain inflammation that may lead to altered mental states or seizures

• 1 in 1000 measles cases develop measles-inclusion body encephalitis - a lethal, progressive inflammation of the brain that typically strikes individuals with weakened immune systems

• 1-3 per 1000 measles cases can result in death

Perhaps the scariest of all measles complications is subacute sclerosing panencephalitis (SSPE), which is estimated to afflict 7-11 out of every 100,000 measles cases. This develops 7-10 years after recovery from measles, and it 100% fatal. Children <2 years of age are at the greatest risk of developing SSPE. I’ve written about it before (see The Price of Mistrust), about the loss of the luxury of not knowing, and I sincerely hope that the parents of those young children who contracted measles do not lose the luxury of not knowing SSPE.

All of this can be prevented. The MMR vaccine is both safe and effective (99% effective in measles prevention after a second vaccination). The most commonly reported adverse events due to the injection are:

(ages 12 to 15 months) - localized pain and redness at injection site, irritability, loss of appetite, drowsiness, and fever (which resolves in a few hours)

(ages 4 to 6 years) - localized pain and redness at injection side, loss of appetite, drowsiness, and fever (which resolves in a few hours)

(ages 7 and older) - localized pain and redness at injection site

So, why is the current outbreak happening if it is entirely preventable with a safe and effective vaccination?

It’s the usual culprits - that combination of vaccine hesitancy and/or denial, cultural and/or religious influences, and a growing sense of mistrust in the institutions of science, public health and government. Sadly, political affiliation may outweigh other factors influencing vaccination rates.

According to a Kaiser Family Foundation (KFF) report from 2021, gaps in vaccination rates across racial and ethnic groups have disappeared just as gaps in vaccination rates across political affiliation have widened substantially. Regarding vaccination against COVID-19, the report uncovers that 90% of Democrats had been vaccinated against 68% of Independents and 58% of Republicans. A June 2025 report from the Science Literacy Project identified some troubling signs:

• Democrats are more likely than Republicans to say they are worried about the current U.S. measles outbreak (76% vs. 28%)

• Democrats are much more likely than Republicans to know that measles cases are higher this year than recent years (71% v. 49%)

Even amid widespread exposure to false claims about measles, large majorities of the public (83%) and parents (78%) say they are “very” or “somewhat confident” that the MMR vaccines are safe. However, confidence in the safety of MMR vaccines is lower among Republican and Republican-leaning parents, with about three in ten (31%) expressing a lack of confidence in the safety of MMR vaccines, including about one in six (17%) who say they are “not at all confident” the MMR vaccine is safe.

It’s disheartening that our nation’s public health has become politicized.

Here’s what I think is worth holding on to amidst the frustration: science has already solved this problem. We have the knowledge and we have a tool (a vaccine that is both safe and effective). What we are lacking is a shared agreement on how to handle knowledge itself. The measles outbreak isn’t a failure of science—it’s a failure of trust. And I think that this is a problem that is solvable.

We need to approach this problem honestly about why that trust was eroded. I don’t think we should dismiss people as being “anti-science” — I don’t think it helps, it just widens the gap between those with trust and those without. If we want to prevent another record year of measles cases—and the year isn’t even over— we need to rebuild trust. Trust in the institutions, trust in the men and women who work and lead those institutions. Institutional accountability and transparency would go a long way towards rebuilding trust.

Measles—just like everything else Mother Nature has thrown at us—is indifferent to political affiliation. Its time we stop dividing ourselves. A virus will not wait for a political consensus before spreading - we need to meet this challenge together, not apart.

Share The Empirical Evidence Collective

The immune system is our body’s defense mechanism. It protects us from infection and even the spread of disease. Macrophages are a specialized type of immune cell that can “eat” the bacteria and microbes that cause infection. They can also eat cancer cells. They do their eating – phagocytosis – by first recognizing self from non-self. Cells which are physically damaged, with unrecognized cell surface proteins (e.g. bacteria and microbes), with atypical cell surface proteins (e.g. cancer), or expressing proteins indicative of cellular stress are all recognized by macrophages and phagocytosed.

In young healthy individuals, the inflammatory response (acute) is initiated when cells of the innate immune system (e.g. macrophages, neutrophils, mast cells, dendritic cells, natural killer cells, eosinophils & basophils) recognize non-self. Once activated, localized cells of the innate immune system begin to release a vast repertoire of signaling molecules to dilate local blood vessels and increase blood flow (causing heat and redness), to increase the vascular permeability to allow fluid proteins and plasma to leak in the localized tissue (causing swelling and sometimes pain), and cell adhesion molecules to catch circulating white blood cells. These circulating cells – both innate and adaptive immune cells (B and T lymphocytes) – accumulate at the site of infection and perform their specialized functions.

As important as this process is, it is equally important that this process be resolved. While the initiation is rapid, its resolution is considerably less so. It takes time for the inherent negative-feedback mechanisms to initiate; the release of anti-inflammatory cytokines once the infection has cleared at cellular level (reducing redness and swelling), clearance of short-lived innate immune cells by macrophages, macrophage shifting from pro-inflammatory (phagocytotic) to anti-inflammatory (tissue healing), and the initiation of tissue repair.

As we age, our immune system’s ability to resolve inflammation becomes impaired. Often referred to as “inflammaging” by biologists, aging is associated with a state of chronic, low-grade systemic inflammation known to play a role in the development of age-related conditions such as cardiovascular disease, frailty, type 2 diabetes, cognitive decline, and cancer. Like most biological processes, there is no single cause of inflammaging. Rather it is the result of a confluence of multiple factors – chief among them is cellular senescence.

Senescent cells are in a state of irreversible growth arrest – they are metabolically active, incapable of dividing, and resistant to apoptosis (programmed cell death). However, these cells do not function as normal cells do. Instead, they develop a characteristic senescence-associated secretory phenotype (SASP) – enhanced secretion of pro-inflammatory cytokines, chemokines, and growth factors – that disrupts tissue homeostasis. Consequently, tissues gradually lose their repair and regenerative capabilities at the cellular level. In the thymus, age-related cellular senescence drives thymic involution, the process whereby the thymus gland shrinks, losing its ability to train the body’s T-cells to fight new infections. The result is that older individuals are more susceptible to respiratory diseases, infections, and reduced responses to vaccines.

Which is why I found this paper so interesting.

Macrophages exist on a functional spectrum. M1 macrophages are pro-inflammatory (defense and pathogen clearance), producing IL-12 and other signals that activate T cells, express costimulatory molecules, and present antigens to T cells for defense and clearance of pathogens. M2 macrophages are anti-inflammatory, producing IL-10 and TGF-β, which actively suppress T cell responses and promote tissue repair. In a young, healthy immune system, the balance between these states is dynamic. However, in aged tissue this dynamic is altered.

In tumors, this dichotomy becomes somewhat fuzzy. While M1 macrophages certainly exist in and around tumors, the majority of tumor associated macrophages (TAMs) are considered to be of the M2 phenotype. TAMs, along with the tumor microenvironment, suppress T cells.

In the article, the authors identify thymulin – a thymus derived peptide that declines with age. When added to aged (52 – 86 years) human myeloid cells, thymulin reduced pro-inflammatory signaling through inhibition of NF-κB. NF-κB is a protein complex that acts as a master regulator of innate and adaptive immunity. Thus, inhibiting NF-κB inhibits pro-inflammatory signals. Aged mice, as compared to young mice, were observed to have increased frequencies of circulating immune cells (e.g. macrophages and monocytes) with a pro-inflammatory cytokine repertoire than an anti-inflammatory cytokine repertoire. And aged mice treated with thymulin were observed to have reductions in circulating pro-inflammatory macrophages and monocytes.

Most interestingly, the authors also demonstrate that administration of thymulin improves the immune system’s ability to respond to cancer. Old tumor bearing mice treated with thymulin exhibited delayed tumor growth and improved survival as compared to untreated controls. Aged mice treated with thymulin had a reduced frequency of tumor-infiltrating myeloid cells producing pro-inflammatory cytokines, and increased the frequency of tumor-infiltrating CD4+ and CD8+ T cells producing IFN-γ. This means that thymulin treated mice had reduced pro-inflammatory signaling and increased cytotoxic T cell activity within the tumor than mice not treated with thymulin.

Immune checkpoint inhibitors (ICIs) have become foundational in the treatment of a wide range of cancers including melanoma, lung cancer, kidney cancer, and Hodgkin lymphoma, often producing durable, long-lasting responses even in advanced disease. This class of cancer therapeutics works by blocking the checkpoint proteins from deactivating T cells. Normally, proteins like PD-1, PD-L1, and CTLA-4 keep immune responses in check to prevent autoimmunity — but tumors secrete these proteins, effectively becoming invisible to T cells. By inhibiting these checkpoints, drugs such as pembrolizumab, nivolumab, and ipilimumab release this inhibition, allowing the patient’s own T cells to recognize and attack the cancer more effectively.

The authors also show that aged mice treated with thymulin and anti-PD-L1 experienced slower tumor growth and longer survival times than aged mice treated with anti-PD-L1 alone. Within the tumor microenvironment of aged mice, pro-inflammatory monocytes and macrophages were the predominant myeloid cells. However, upon treatment with thymulin, the frequency of pro-inflammatory monocytes and macrophages were reduced. Thymulin treatment had minimal effects on immune cell repertoires or tumor responses in young mice.

While it is too early to tell if this observation will translate to humans (see my previous post on why preclinical studies in mice often fail in humans), it is impressive. Peptide therapies face well-known pharmacological hurdles—stability, delivery, dosing—and thymulin will not clear them overnight. But the mechanism it illuminates is sound: age-related immune decline is, in part, a reversible signaling problem. Diminishing the systemic pro-inflammatory signaling with a single targeted intervention may recondition the immune environment such that checkpoint inhibitors become more effective in the patients - frailty, immunocompromised, or with significant comorbidities - who currently don’t respond. It’s a hypothesis worth testing.

The translation of basic science into human therapeutics has always been an iterative, empirical process — acquisition of experimental data, rigorous analysis, hypothesis refinement, and clinical testing. It’s what built the checkpoint inhibitors we have today, and the same evidence-driven pipeline is what will ultimately yield the next generation of therapies.

Share The Empirical Evidence Collective

Knowledge, in the broadest sense, is the awareness, understanding, or information one may acquire through study, observation, or experience. Knowledge is both complementary – my knowledge can fill in the gaps of yours and vice versa – and cumulative – each generation’s knowledge is built upon prior knowledge. New knowledge can be discovered, complementing and building upon what came before. Knowledge can also be objective or subjective: objective knowledge exists independently of any individual (e.g. 2 + 2 = 4), while subjective knowledge is necessarily shaped by one’s prior experiences and current feelings (e.g. art critiques, tastes in music).

Knowledge is both personal and collective. Personal knowledge is the knowledge we carry in our own minds. This is the tacit knowledge of the baker who knows when the dough is ready just by touch, and of the engineer who knows how to apply Maxwell’s equations to design wireless communications systems. It lives in the body, in habits, in patterns recognized before they are articulated. Although it is irreducibly first-person and experiential, it can be taught to others through mentorship and shared experiences. Collective knowledge, by contrast, is durable – outliving its discoverers – passing down through generations and across borders in ways personal knowledge cannot. It lives in books, papers, data, equations, theories, and models: the accumulated foundation that lets us build everything from bridges to life-saving therapeutics.

Imagine our collective knowledge as a circle. Within it lies the collective understanding of all humanity, everything from physics and chemistry to philosophy and biology, built over millennia. Outside the circle, is everything we do not currently know or understand – everything yet to be discovered. The boundary between the two is a thin, fuzzy but flexible edge. Every new discovery – whether a PhD dissertation, scientific publication, building technique, or pharmaceutical – nudges that circle outwards. Zoom in close enough and you will see a highly convoluted boundary; each outcropping is a new discovery, the regions between complementary knowledge waiting to be discovered. And that circle moves outwards.

If knowledge is what we carry, then expertise is what we do with it. This distinction matters – especially in science. Expertise is distinct. It depends upon knowledge – both personal and collective – but also upon experience in the application of that knowledge. Expertise, in the broadest sense, is an exceptionally high level of performance in a particular task within a given domain. It is neither conferred nor bestowed. Sometimes mistaken for talent or skill, expertise is recognized – not always instantly – in individuals like Lionel Messi and Yo-Yo Ma. While both are clearly talented, and have spent years mastering their skills, the time invested has given them the experience – the expertise – to understand at a fundamentally deeper level, to perform at a level higher than most. In science, recognition of expertise may look different, but the principle is the same.

In science, expertise is more than just knowledge: it is the ability to discover new objective knowledge, or to apply new and existing knowledge in new ways. Scientific expertise is the deep, fundamental understanding that provides one the insights into how the world works, such that one may formulate a theory to explain it or devise an experiment to test it. To understand what scientific expertise actually looks like in practice, it helps to picture it differently.

Imagine a large marble tray – so large that you cannot reach the other side. Within it is a randomly placed hole with a buzzer that will go off once a marble goes through. Now imagine some obstacles – arches with and without buzzers, funnels and walls – all randomly placed. The surface is smooth, but randomly uneven; its slight changes in elevation subtly directing marbles one way or another.

That hole in the center with the buzzer represents the objective truth – a complete explanation of an initial observation that piqued your interest. This is the answer to why. Why does that observation happen, and what is its purpose?

You want to hear that buzzer.

Now, put on a pair of sunglasses and turn off the lights.

Where do you place your marble?

Do you simply let the marble roll, or do you push it? If so, which direction? And with how much force?

As scientists, we start by placing our marble based upon our own personal knowledge – that subset of our collective knowledge known to us, or the collective personal knowledge of our team. The decision of where to place the marble depends on the extent of our knowledge and experience in designing and performing experiments. The decision to simply let the marble roll is analogous to performing the next readily apparent, logical experiment based upon this starting point. Choosing to push the marble in a given direction would be analogous to performing a high-risk, high-reward experiment – it may get you going in the right direction, but it could also lead you astray.

Let’s imagine that you place your marble on the tray and choose to simply let it roll. The marble is in motion, and you listen attentively to the sound it makes as it rolls along the tray. While clearly audible, it is faint, requiring you to focus intently. You wait for several minutes, in the dark, and you think the marble stopped moving a minute or so ago. After waiting for a few more moments in silence, you conclude that your marble has stopped rolling. It is stuck somewhere, most likely in the concave side of a buzzer-less arch.

Sometimes, despite our best efforts – well-designed experiments that yield high-quality, reproducible data – the path to answering our initial question of why becomes uncertain. Perhaps the next logical experiment is not readily apparent. Perhaps it is unfeasible for technical reasons. Or perhaps the experimental data is leading you outside of your personal circle of knowledge.

Do you forget about this marble and start over?

Do you go back to the available scientific literature, increasing your personal circle of knowledge, and revise your initial question and metaphorical starting point?

Do you perform an exploratory (often open-ended), high-risk, high-reward experiment?

You choose to go back to the scientific literature, and you discuss your data with your peers and mentors, gaining new knowledge and insight(s). You revise your initial question – your initial experiment – and you abandon the marble. You still have all the acquired experimental data, but you are starting again with a new marble in a new starting position. This new starting position is not far from the previous one, but with the newly revised initial question – and the experiments necessary to answer it – the marble goes off in a new direction.

Again, you wait in the dark, listening intently for the sound of the marble rolling across the tray. After a seemingly long time, longer than you had initially thought, you suddenly hear the buzzer. It is at this point that your mentor, who has been standing next to you the whole time, guiding your decisions along the way, hands you a new pair of sunglasses. You put them on, and you can see – no, not everything, but you can see the paths of the two marbles. In some areas it looks like a dimmed flashlight illuminating a wide swath of the path; in others, like a faint line traced by the marble. It is at this moment that you suddenly begin to understand the complexity of the paths taken. It is at this moment that you realize both marbles are in arches.

You discovered an objective truth, not the objective truth. That now-visible path of the second marble answers the how of your initial question: the mechanism by which you set about answering the why of your initial observation. This is the molecular mechanism; cause activates A, which inhibits B (the inhibitor of C), thereby leading to the activation of C and the observed effect. Cause and effect. Those twists and turns of the paths are the series of experiments, the successive questions asked and answered, leading toward that initial goal of answering the question of why. Those multiple branches off the main paths that seemingly stop in the middle of nowhere represent failed experiments or pieces of data that, while interesting, did not lead you to answering the how. Those dimly illuminated areas represent interrelated questions and their experimentally acquired data that, only when considered as a whole, led to the next experiment and the continuation of the path.

While writing the manuscript outlining your experimental methods and data – your interpretation of the data within the context of the field (your personal knowledge of the collective knowledge of the field) – you come across a paper whose data complements yours. Importantly, a subset of this data completely overlaps with yours. After carefully reading this paper multiple times, you can clearly see this other group’s path – their marble’s paths – the points of intersection and the sections of overlap. Although they asked a different set of questions from a different starting point, the path towards their answer overlapped with the path to yours. That point of intersection is an independently verified objective truth.

This is how biological science is done. Numerous scientific groups, each pursuing their own questions from their own starting points, publishing their paths to an objective truth. Over time, these paths converge and intersect, illuminating once-dark areas. New objective truths, new knowledge, discovered. Areas with numerous intersections – those areas where multiple lines of evidence converge – become the brightly illuminated areas representing what we, as biological scientists, know. Known knowledge: the independently verified objective truths that hundreds of scientists across the world have illuminated, line by line, intersection by intersection, over years and sometimes decades of research. The central features of these independently verified objective truths are clustered, patterns are observed, concepts are formed and later established through the refinement that can only come about through still more independent verification and validation.

We biologists simply have an improved view of the marble tray as compared to non-biologists – a different pair of glasses. Before I went to graduate school, I had a pair of glasses that allowed me to see a few dimly illuminated areas of that marble tray. My field of view was rather small, and I could not see the details – the arches, funnels, and walls; definitely not the subtleties in elevation that would direct a marble one way or another. As I went through graduate school, through my postdoctoral training, even as a Principal Investigator and drug discovery professional, I kept upgrading my glasses. Each successive upgrade – brought about by reading the scientific literature, designing and performing experiments, analyzing experimental data and interpreting that data within the context of the field, discussing the data and interpretations with other scientists – allowed me to see more and more of that marble tray.

Of course there were dark areas between here and there. Certainly, my destination was weakly illuminated, if illuminated at all. That’s the point;: to forge new paths through the metaphorical darkness, bringing illumination along the way. This is why we scientists do what we do. We are ever curious, constantly asking questions. Our individual drive to answer those questions – partially innate, partially ingrained – has been nurtured. Spending the better part of two decades doing rigorous science tends to do that.

It is only because of this experience – the years spent asking questions, answering questions, framing those answers within the context of our collective knowledge, discussing the biological significance of our observations within that context and with reference to the original question – that we can say that we understand. It is only because of this understanding that we have that enhanced view of the metaphorical marble tray. No, we do not know or understand everything. Nobody does. And no, we cannot navigate that metaphorical darkness without knowing where they stand relative to previously illuminated areas, without intersecting a previously illuminated path. Nobody can. Context, one might say, matters.

This is what we call scientific expertise – the ability to navigate the marble tray in darkness, illuminating new paths and new areas based upon the experience of illuminating prior paths and areas. Or, if you prefer, a more formal, analogy-less definition: a deep, specialized knowledge, technical skills, and research experience in a specific scientific field, often characterized by an ability to analyze, interpret, and advance our understanding of phenomena through rigorous, evidence-based methods; an active, evolving competence rooted within the scientific community and the ability to adapt interpretations and facilitate the evolution of concepts to accommodate newly acquired data, rather than relying upon static, or perhaps outdated, information.

This distinction between knowledge and expertise is not merely semantic — it carries real consequences. When expertise is mistaken for a conferred title rather than earned competence, the public loses its ability to tell the difference between a genuine guide through the darkness and someone simply wearing glasses they did not earn. A degree tells you what knowledge someone acquired. Experience reveals expertise – what they have done with it. In an era where knowledge is more accessible than ever but trust in expertise is eroding, understanding this difference matters more than ever. And in the dark, with a marble and a buzzer waiting somewhere ahead, that is the only distinction that counts.

Clinical trials are more than just a regulatory hurdle to overcome. They are our society’s mechanism by which the pursuit of therapeutic innovation (and profit) does not outpace our ethical responsibility to protect the people who place their trust in the scientists who develop drugs, and the physicians who prescribe them. They exist because preclinical data (e.g. target engagement, specificity, PK/PD, efficacy and toxicity) provide an incomplete prediction of how the drug will function in humans.

Preclinical safety testing is designed to detect known classes of risk: genotoxicity, hepatotoxicity, acute organ damage, predictable pharmacodynamic (dose-dependent) off-target effects. It cannot detect unknown mechanisms. In the case of Vioxx (see Part 4 of this series), the pro-thrombotic consequence of COX-2 selectivity was a biological signal unlikely to be observed in the existing pre-clinical models at the time.

Preclinical efficacy testing follows a similar logic. It is designed to demonstrate efficacy in a model of human disease. While every effort is made to mimic as closely as possible the in vivo disease state observed in the clinic, every biochemical assay, every in vitro assay, and every in vivo mouse model is merely an approximation of human disease. While having transformed biomedical research into the driving force behind drug development efforts — preclinical models are not 100% predictive of human responses.

As the Vioxx case demonstrates, a drug can pass every preclinical assay — COX-2 selectivity, negative for genotoxicity, clean repeated dose toxicity in rodent and non-rodent species — and still cause fatal cardiovascular events in humans. The lessons from Vioxx were devastating but instructive. First, clinical trial transparency is essential. This means that pharmaceutical companies must publish all clinical trial data – both positive and negative – promptly. Second, post-market surveillance is critical for identifying rare or long-term safety risks. Third, data monitoring functions in the public’s interest best when it is independent and free from corporate influence. And Fourth, direct-to-consumer advertising requires scrutiny.

The human cost of Vioxx was staggering, but the systemic vulnerabilities it revealed are reflected in broader development statistics: clinical trials eliminate the vast majority of candidates, and the reasons for elimination go far beyond post-market safety failures alone.

A 2022 analysis of clinical trials from 2010 to 2017 identified four main reasons why >90% of clinical trial candidates (compounds) fail: lack of clinical efficacy (40 – 50%), unmanageable toxicity (30%), poor drug-like properties (10-15%), and a lack of commercial need or poor strategic planning (10%). Across the full development pipeline from investigational new drug (IND) application to FDA approval, the likelihood-of-approval (LOA) has been estimated at approximately 10%, with the lowest LOAs occurring for compounds in oncology (6.7%), cardiovascular (8.7%), and autoimmune-immunology (12.7%).

Step-by-Step Analysis: Clinical Trials by Phase

Phase 1 clinical trials have a single purpose: to evaluate the safety of an experimental drug or therapy in humans. These first-in-human studies are typically conducted on small groups (20 – 100) of healthy volunteers or patients in what are referred to as dose escalation studs. This can take one of two forms, either a single ascending dose (SAD) in which a small group gets a single dose, and if it is safe, then a new group receives a higher dose, or a multiple ascending dose (MAD) in which participants receive multiple doses over time, allowing for evaluation of how the drug may build up in the body. During the course of the study, participants are closely monitored (e.g. frequent blood samples for PK/PD analysis) for signs of adverse reactions and the dose when these reactions first occur. Compounds for which toxicity (e.g. liver enzyme test failures) or adverse reactions become common or severe at doses below those deemed to be therapeutic to not make the transition to the next phase of testing.

It is in Phase 1 that species differences in PK first become readily apparent. Drug-metabolizing enzymes (Cytochrome P450 family of enzymes) diverge significantly in both number and functionality across species. Humans have 57 working genes while mice have 102, and though many have overlapping functionality between species, differences in their relative activities may be significant. A compound that is rapidly cleared in mice may accumulate to toxic levels in humans. These differences affect not only the rate of a compound’s clearance, but also the profile of its metabolites – whether they may have biological activity or toxicity independent of the parent (dosed) compound.

According to a 2014 analysis, 36% of compounds fail to make the transition from Phase 1 to Phase 2.

Phase 2 clinical trials have a single purpose: to demonstrate efficacy of an experimental drug or therapy in humans. It is at this stage that a compound may make the transition from “is this safe?” to “does this provide a clinical benefit for the patients it was designed?” This is the first stage in which the compound is administered to patients likely to benefit. It is also the so called “valley of death” where basic biomedical research and preclinical successes all too often to fail to translate to humans.

The reasons for this are complex, but they generally fall into two categories.

Wrong target (see Part 1 of the series) — the biological target for which the compound was designed is not central to the disease state or disease-causing mechanism in the proposed patient population, and Translational failure (see Part 4 of the series) — preclinical efficacy was demonstrated in a disease model that did not fully replicate the disease state or disease-causing mechanism in the proposed patient population.

While the first is relatively straightforward — a compound against a target that is not central to disease would not be expected to reduce the burden of disease — the second is much less so. Animal models of disease are the foundation of preclinical efficacy (and safety) testing, and their use is contentious. A central tension exists: animal models are the best tool available to us for predicting human responses to new compounds, and yet they are inadequate. A compound that is safe and effective in mouse models of disease, and safe in healthy humans (Phase 1), may be ineffective and/or toxic in the patient population for which it was designed.

How can this be?

Simply put, our best animal models do not fully replicate the disease state or disease-causing mechanism(s) present in humans. Differences in metabolism contributing to differing PK profiles (see Part 4 of the series), in genetics, in immune responses, and in cell surface receptor expression profiles may all contribute. In addition, disease models are wholly artificial (e.g. human tumor cells implanted in a young, immuno-compromised mouse), in which the target biology may differ in sublet yet crucial ways. In short, mice are not humans.

Lastly, compounds demonstrated to be safe in Phase 1 trials may yet prove to be toxic in the patient population for which it was designed. Patients frequently require higher doses and longer exposure times (continual dosing for weeks or months) than previously tested before clinical effects may be measurable. Patients are typically being treated for other indications as well, and are likely on additional therapeutics which may interfere (either synergizing with or impairing) with the activity of the therapeutic being tested. As a consequence of patient’s other indications, the PK profile may be significantly altered from that which was determined preclinically or in Phase 1 trials. In short, healthy volunteers are not patients.

Approximately 68% of compounds fail to make the transition from Phase 2 to Phase 3.

Phase 3 clinical trials expand upon the success of Phase 2 trials, enrolling more patients while more rigorously testing its efficacy. A Phase 3 trail is designed to answer the question, “Is this therapeutic more effective than what is currently available?”

The gold-standard design is the randomized controlled trial. Participants meeting predefined eligibility criteria are randomly assigned — usually by a computer algorithm — to either the experimental arm or a control arm. Randomization minimizes selection bias and ensures that known and unknown confounding factors (age, comorbidities, disease severity, genetics) are approximately balanced between groups, so differences in outcomes can more plausibly be attributed to the intervention. Allocation is typically double-blinded: neither participants nor investigators know who received which treatment, preventing expectation effects from skewing subjective endpoints. Some trials are open-label (both investigators and participants know what treatment is being given) when blinding is impractical or unethical.

Share The Empirical Evidence Collective

A common randomization ratio is 1:1, but some trials use 2:1 or 3:1 to expose more participants to the experimental treatment, which may be justified when early signals suggest benefit or safety is reassuring. Stratified randomization ensures balanced assignment across important prognostic factors (e.g., disease stage, age group, or gender). Adaptive designs, which have become increasingly popular, allow preplanned modifications during the trial such as dose adjustments, participant reallocation to more effective study arms (e.g. test therapeutic alone vs. test therapeutic plus common co-therapeutic), and sample size re-estimation based on interim analyses without compromising statistical validity.

The choice of control group is the second defining feature. Three main options exist:

Placebo control — the experimental drug is compared against an inert substance. This design is strongest for detecting absolute treatment effect and is ethically acceptable only when no proven therapy exists for the condition, or if the experimental drug is simply added to a proven treatment program. The placebo effect is indeed real, and can be quite substantial, particularly with respect to subjective endpoints such as pain or depression. Because of this, double-blind trials are preferred whenever possible.

Standard-of-care (active) control — the experimental drug is compared against the currently best available treatment. This is increasingly the default in many therapeutic areas because it answers the clinically relevant question: “Is this new drug better than what we already have?” This is the ethical choice when an effective treatment is currently available. The trade-off is a smaller effect size (essentially testing for an incremental improvement) and the need for larger sample sizes. Non-inferiority or superiority trial designs are chosen depending on whether the goal is to match or exceed the efficacy of the active comparator (control).

Another type of study, the Head-to-head active comparator plus placebo, is becoming increasingly prevalent. It is a randomized clinical trial, but it is a three-arm design where both experimental and active-control arms are included alongside a placebo. This allows for the simultaneous assessment of both absolute and relative efficacy, effectively answering the questions “Is the new therapeutic better than placebo?” and “Is the new therapeutic better than the current Standard-of-care?” within a single study. This type of study was used to demonstrate that the addition of olokizumab was was superior to placebo and non-inferior to adalimumab in producing clinical improvements at 12 weeks in patients with rheumatoid arthritis. Randomization is typically 1:1:1, though unequal ratios may be used to emphasize one comparison. Double-blinding is maintained across all three arms (identical-appearing formulations, identical dosing schedules, and matched placebos for the active comparator). Stratified randomization balances key prognostic factors across all groups. The drawback to this type of study is its requirement for large sample sizes.

Primary and Secondary endpoints are pre-specified, clinically meaningful outcomes (overall survival, response rates, functional improvement) that the trial is powered to detect — typically requiring 80–90% statistical power. While most endpoints are objective (e.g. % reduction in blood cholesterol, reduction in tumor volume), some such as pain relief or depression are subjective, and more likely to be influenced by the placebo effect.

Approximately 60% of compounds from Phase 3 testing have a New Drug Application (NDA) filed with the FDA.

Regulatory agencies review Phase 3 data alongside Phase 1 and 2 datasets, manufacturing quality information, and proposed labeling before granting approval. A positive Phase 3 trial does not guarantee approval — many have been rejected for safety concerns, inconclusive efficacy, or manufacturing deficiencies. However, a well-conducted, adequately powered, and statistically significant Phase 3 trial represents the strongest available evidence for therapeutic benefit.

Approximately 83% of NDAs are approved by the FDA.

Share The Empirical Evidence Collective

According to a ProPublica investigation, more than 700 newborns died in 2024 from spontaneous vitamin K deficiency bleeding (VKDB). VKDB is classified according to the time of presentation: early (within 24 hours of birth), classic (within 1 week of birth), and late (2 – 12 weeks after birth). Infants are naturally at risk because of reduced vitamin K bioavailability due to poor transfer through the placenta, its low content in breast milk when compared to fortified cow’s milk-based formula, its relatively short half-life and its low production due to immature gut flora.

Why did this article hit me so hard?

Those deaths were entirely preventable. Prior to 1961 when the American Academy of Pediatrics first recommended a vitamin K shot at birth, VKDB impacted nearly 1.7% of all newborns (that’s 1 out of every 60 live births); 20% of which, and despite treatment, died. Those who survived more serious cases frequently developed severe neuro-developmental damage leading to lifelong cognitive impairments. With the recommendation, and it becoming the routine standard of care, the incidence of newborn deaths and cognitive impairments attributable to VKDB dropped to essentially zero in the decades since.

Then in the early 2010s, doctors started to observe babies with VKDB as more Americans started choosing against the routine for shot. A 2020 study on the reasons parents cite when choosing against the shot included a concern of harm from the injection, a desire to be natural, and a belief in alternatives to prevent VKDB. Specific concerns cited in the article range from the misperception that vitamin K is a “vaccine” and that the injection causes cancer to a belief that vitamin K deficiency is natural, preferring to increase the mother’s own dietary intake of vitamin K, to a lack of knowledge about the risk of VKDB to their infants. In addition, the parents choosing against the routine vitamin K shot are more likely to choose not to vaccinate. Where mothers chose to give birth may have an influence on their decision regarding the vitamin K shot — the frequency of choosing against the vitamin K shot ranged from 0 – 3.2% in U.S. hospitals, up to 14.5% in home births, and up to 31% in birthing centers. The authors go on to state:

Some parents would have only considered intramuscular vitamin K in the event of a potential injury to the infant or if the birth required another intervention. Decisions about the refusal of intramuscular vitamin K were influenced by the opinions of other mothers or family members; statements of celebrities and health professionals, including midwives and chiropractors; as well as written information (information sheets) and Internet blogs.

The U.S. public’s trust in science and medicine — and of those individuals in the fields of science and medicine — has been declining in recent decades. This has been documented in multiple reports and studies over the years from the Pew Research Center and John’s Hopkins Center for Health Equity to Time magazine. It is also part of a broader long-running trend of growing mistrust in U.S. institutions in general.

The parents who choose against the vitamin K shot, similar to those who choose not to vaccinate their children, are part of this growing wave of mistrust in science and medicine. Lacking trust and the information to make a decision, these individuals seek out those seemingly trustworthy regarding their situation — influencers, celebrities, and internet bloggers who talk about their personal experiences. We all feel something when hearing an emotionally powerful story, especially one that we can relate to. This is entirely normal and very human.

Decades of neuroscience tells us that when a connection is formed with a character or a personal story, the brain releases oxytocin. Dubbed the “trust hormone” or the “love hormone,” oxytocin is both a hormone (e.g. it lowers stress by reducing the secretion of cortisol, it triggers uterine contractions during labor and the milk-letdown reflex when breastfeeding) and a neurotransmitter. As a neurotransmitter it is known to regulate social bonding, empathy, and interpersonal trust or connectedness. Stories invoke other emotions as well. Anxiety and fear; that’s cortisol. Awe and joy; that’s dopamine. Peace and wellbeing is serotonin. There are others, of course, but importantly these neurotransmitters are far more complex than simple on/off switches. These neurotransmitters can influence how our brains process, store, and retrieve information. These neurotransmitters can influence how we make decisions.

While we may never know the experiences that lead to those parent’s decision against the routine vitamin K shot, we do know that at some point prior they lost trust in science and medicine, and chose to place their trust elsewhere. They placed their trust in the influencers, celebrities, and internet bloggers who shared their personal stories — they placed their trust in individuals with whom they had a connection. It doesn’t matter that the connection is entirely one sided, with the viewer, listener, or follower investing emotional energy, time, and interest into a media figure without any form of reciprocations. For these individuals, their parasocial relationships are a connection, one that emotionally influences their decisions on everything from routine purchases to healthcare.

Science and medicine are not without narrative or compelling stories. They are not devoid of emotion. Science and medicine just place more weight on the statistics of empirical (often quantifiable) evidence and mechanistic understanding when making decisions and recommendations than on personal stories or anecdotes. While one’s, or another’s, personal experience cannot be negated, a personal story, no matter how emotionally compelling, is, in the language of statistics, just an “n of 1.” It is neither empirical (quantifiable) nor reproducible. It is not predictive of future outcomes. It is just an account of an individual’s personal experience.

We humans naturally focus on the successful end-result of someone else’s story and assume (falsely) that their actions may be applied to others (or ourselves) with similar results. And we are good at doing this — humans are natural storytellers. We are susceptible to narrative fallacy, the tendency to weave complex, random, or incomplete facts into an overly simplified coherent story, causing an overestimation of cause-and-effect and an underestimation of the role of luck. We are also susceptible to outcome bias, the tendency to judge the quality of a decision based on the result, rather than the soundness of the decision. Thus, (for some) a compelling story from an influencer, celebrity, or internet blogger — no matter how incomplete — may become a coherent story and influence them into making an unsound decision based upon the successful (lucky) outcome of that single incident.

Share The Empirical Evidence Collective

From a mechanistic perspective, vitamin K is a cofactor that is essential for the activation of proteins responsible blood coagulation. Vitamin K deficiency impairs this coagulation process, and may lead to bleeding issues. Prescribed blood thinners such as warfarin impair the blood clotting process by blocking vitamin K. Newborns have very low levels of vitamin K, and thus are at risk of bleeding issues. Adults with vitamin K deficiency, and those who recently started taking warfarin, which has a narrow therapeutic window, are also at risk of bleeding issues. In 1961 the American Academy of Pediatrics recommends a vitamin K shot at birth, and this recommendation becomes part of the standard of care. In the subsequent decades the incidence of VKDB in newborns fell from ~1.7% to nearly zero. The relative risk of developing late VKDB is estimated to be 81 times greater for infants not receiving the vitamin K shot than among those infants who do.

This entirely preventable tragedy is the result of two interrelated phenomena. First is the well documented decrease in societal trust of institutions, science and medicine being just a few of the more visible losses. Second is the rise of a self-reinforcing information ecosystem that exists at the intersection of digital platforms, algorithms, and human psychology to endlessly validate and amplify one’s preexisting notions while filtering out dissenting perspectives. As a consequence, the trust in science and medicine as institutions, the trust in scientists and physicians as experts, has shifted to parasocial relationships and other peer-to-peer networks. It has given rise to the “do your own research” ethos — once common only amongst conspiracy theorists — that is now pervasive.

We have a choice to make. We can continue along our current path of diminishing trust in science and medicine, with its increasingly dire effects on human lives (e.g. the entirely preventable deaths of 700 babies of VKDB in 2024, and the entirely preventable measles outbreak currently afflicting the U.S.) or we can make some changes and start to rebuild trust in science. The one change that we (myself speaking as a scientist) can make right now is to acknowledge the power of storytelling and of building personal connections. This has been the subject of multiple initiatives, in-depth discussions, and commentaries from the Association of American Medical Colleges (AAMC), Harvard University, the Aspen Institute, The Pew Charitable Trusts, and others. The other change we can make right now is to tell stories that inform and engage the public. Facts are vitally important, but if no one is engaging with them, do they matter?

My hope is that as more scientists engage with the broader public, building connections and renewing the public’s trust in science, that new parents (we as a society) will once again have the luxury of not knowing about newborn deaths related to VKDB. My hope is that once again parents (we as a society) will have the luxury of not knowing encephalitis (inflammation of the brain), which affects about 1 in every 1000 children who contract measles and may result in permanent life-long neurological conditions including cognitive impairment, epilepsy, motor deficit and behavioral or personality changes. My hope is that once again parents (we as a society) will have the luxury of not knowing subacute sclerosing panencephalitis (SSPE), the rare (afflicting 4 to 11 of every 100,000 individuals), progressive, fatal brain inflammation that develops 8 – 11 years after children recover from a measles infection. My hope is that we as a society regain trust in science so that we may once again have luxury of not knowing entirely preventable deaths.

We have some work to do.

So, what happened?

Briefly….In November 1998, Merck submits Vioxx for FDA approval. In January of 1999, Merck launches the Vioxx Gastrointestinal Outcomes Research study (VIGOR) with over 8000 participants in which half take Vioxx, and the other have take naproxen. In October of 1999, at the VIGOR study’s data and safety monitoring board’s (DSMB) first meeting, results show that Vioxx patients have fewer ulcers and gastrointestinal bleeding than patients taking naproxen. At the DSMB’s second meeting in November, the discussion focusses on cardiovascular problems where it is noted that there were nearly twice as many patients with serious heart problems, or had died, taking Vioxx than there were taking naproxen. At the DSMB’s last meeting that December, it concludes that the risk of serious heart problems and death among Vioxx patients is twice that of those taking naproxen but decides that Merck needs an analysis plan for the cardiovascular data before the study ends. Later, when defending its decision to continue the study, the DSMB stated uncertainty as to whether Vioxx was indeed causing the heart problems, or if naproxen, acting like low-dose aspirin, was cardioprotective and thereby making Vioxx look artificially problematic.

In May of 2000, Merck submits the VIGOR paper to the New England Journal of Medicine (NEJM) for publication. Notably, the data only contain 17 of the 20 Vioxx patient heart attacks. In July and November of 2000, VIGOR authors submit two sets of corrections to their NEJM manuscript without mentioning the three additional heart attacks. On October 12, 2000, Merck informs the FDA about heart attacks 18, 19, and 20 (they were previously informed of heart attack 1 – 17), and on November 23, 2000, the VIGOR results are published in NEJM. On August 22, 2001, a meta-analysis – based upon the complete VIGOR dataset made available by the FDA – is published in the Journal of the American Medical Association (JAMA). This analysis casts serious doubt on the hypothesis that naproxen is cardioprotective. Between January 2001 and August 2004, multiple meta-analyses, commentaries, clinical trials, and observational studies are published documenting an increased risk of cardiovascular disease associated with Vioxx.

While there is more to this story, what is important is what we have learned: that preclinical testing must assess systemic risks in addition to its intended benefits.

So, what does this mean, practically speaking? How is this done? And how can we ensure that we do not re-learn the lessons of Vioxx.

This is where preclinical testing (efficacy, ADME, toxicity) becomes essential. Although we briefly discussed these in Part 3 of this series (in the context of lead optimization - the Build-Make-Test cycle), their impact on the development of the clinical candidate and how its respective clinical trial(s) will be designed is the focus here.

Let’s begin with the assumption that we have identified several drug candidates – leads that have been optimized for target engagement & in vitro efficacy. These molecules all do what we want them to do – bind the drug target with high affinity and modulate its activity. These molecules all have comparable physical chemical properties (e.g. solubility, pKa, lipophilicity, protein binding, etc.). These molecules have passed a preliminary series of tests, but they are not clinical candidates. To become clinical candidates they must pass the next series of tests.

Pharmacokinetics and Pharmacodynamics

Pharmacokinetics and pharmacodynamics — collectively abbreviated PK/PD — form the theoretical backbone of preclinical development. They answer two distinct but complementary questions.

Pharmacokinetics asks: what does the body do to the drug?

This encompasses absorption (how the compound enters the bloodstream), distribution (where it goes), metabolism (how it is chemically altered), and excretion (how it leaves the body) – often referred to collectively with the acronym ADME. PK is measured by administering the compound to animals at various doses and measuring its concentration in the blood and tissues over time. The key parameters are:

Bioavailability (F). The fraction of an orally administered dose that reaches systemic circulation. A compound with 10% bioavailability means that 90% of the dose is lost — typically to incomplete intestinal absorption, first-pass metabolism by the liver, or degradation in the gut lumen. Compounds with very low bioavailability may be reformulated or administered intravenously, but oral administration remains the preferred route for most therapeutics.

Clearance (CL). The rate at which the compound is removed from the body. Clearance is determined by hepatic metabolism, renal excretion, and other routes. High clearance means the compound must be dosed frequently or at higher doses to maintain therapeutic levels.

Half-life (t_½). The time required for the compound’s concentration to decrease by 50%. A short half-life (hours) may require multiple daily doses; a long half-life (days) may allow once-weekly administration.

Volume of distribution (Vd). The theoretical volume into which the compound is distributed. A large Vd indicates extensive tissue distribution, which may be desirable for targeting organs beyond the bloodstream.

Pharmacodynamics asks: what does the drug do to the body?

This encompasses the relationship (dose dependent) between drug and its effects on the body – both therapeutic and adverse. The key PD parameter is the exposure-response relationship: at what concentration, and for how long, does the compound produce its intended effect?

PD studies are conducted in vitro (measuring target engagement and its downstream effects in cell culture model systems) and in vivo (measuring the therapeutic effect in animal models of disease). The goal is to establish a dose-response curve — the relationship between the administered dose and the magnitude of the effect — and to identify the therapeutic window: the range of doses between the minimum effective dose and the maximum tolerated dose.

A wide therapeutic window is desirable. It means that there is substantial margin between the dose that produces a beneficial therapeutic effect and the dose that causes toxicity. A narrow therapeutic window — exemplified by drugs like warfarin, digoxin, and lithium — requires careful dose titration and close monitoring, because small variations in dose can produce large variations in effect. For a new molecular entity, a narrow therapeutic window is a significant development risk, because it means the compound is less forgiving of manufacturing variability, patient compliance issues, and drug-drug interactions.

PK and PD are inseparable in practice. The therapeutic effect observed in an animal model depends not only on the compound’s intrinsic potency (a PD property) but also on its bioavailability, half-life, and tissue distribution (PK properties). This tension between potency and pharmacokinetics is the reason lead optimization — discussed in Part 3 of this series — is so arduous. Every improvement in potency must be weighed against its impact on absorption, metabolism, and solubility. A compound that is extremely potent in vitro but has poor oral bioavailability and a half-life of ten minutes will be therapeutically useless, regardless of its binding affinity. Conversely, a compound of moderate potency but with favorable pharmacokinetic properties may be a superior therapeutic candidate.

Pharmacodynamic Therapeutic and Safety Profiling: In Vivo Efficacy and Toxicity Analysis

Current regulatory requirements (www.ich.org) mandate toxicity testing in two animal species – a rodent (typically mouse or rat) and one non-rodent species (e.g. rabbit, dog, or monkey) – prior to the conduct of Human Clinical Trials and Marketing Authorization for Pharmaceuticals. These studies are designed to identify the no-observed-adverse-effect level (NOAEL), the highest dose at which no adverse effects are observed, and the dose at which adverse effects first appear (the LOAEL — lowest-observed-adverse-effect level). These values are critical inputs for the first-in-human dose calculation.

Efficacy and toxicology studies are conducted under Good Laboratory Practice (GLP) regulations, which prescribe detailed requirements for study design, animal handling, data recording, and reporting. GLP compliance is not optional. The FDA will reject an IND application if the toxicology data were not generated in a GLP-compliant facility.

A standard preclinical toxicological analysis includes:

Acute toxicity studies. A single dose is administered, and animals are observed for 14 days for signs of acute toxicity: mortality, behavioral changes such as hunched posture, lethargy and reduced grooming, and clinical signs such as seizures, ataxia, tremors, labored or rapid breathing, and diarrhea. This establishes the acute lethal dose and identifies any immediate organ damage.

Repeated-dose toxicity studies. The compound is administered repeatedly (typically daily) at non-acute toxicity doses over 14 days (in rodents) or 28 days (in non-rodents), with clinical observations, body weight measurements, food consumption, and clinical pathology (blood counts, liver and kidney function tests). At the end of the study period, animals are euthanized, and organs are examined for histopathological changes.

Genotoxicity studies. A pair of in vitro tests — the Ames test (bacterial genetic mutation assay), and either a chromosomal aberration test or a micronucleus test (both of which are performed in mammalian cells or human peripheral blood lymphocytes). These tests are designed to determine if the compound causes genetic damage. If both tests are negative, the compound is considered non-genotoxic.

Safety pharmacology studies. Core organ systems — cardiovascular, respiratory, and central nervous system — are evaluated for acute adverse effects. The most important of these is the in vitro hERG potassium channel assay, which predicts the risk of QT interval prolongation — a cardiac electrical abnormality that can lead to fatal arrhythmias.

While there exists a standard set of toxicological assays to that are performed universally, efficacy studies are highly bespoke. Let’s assume that we have been developing a compound for the treatment of pancreatic ductal adenocarcinoma (PDAC), and that our compound was designed to specifically target RAS with G12 mutations similar to that of daraxonrasib.

The purpose of preclinical efficacy studies (when coupled with appropriate safety data) is twofold, (1) to convince regulatory agencies that the compound is both safe and effective in the patient population for which the drug was developed, and (2) to convince the C-Suite that the compound in question warrants the additional investment of capital – anywhere from the estimated $1M to $5M for a Phase I which focuses on the safety of a small group of healthy volunteers (20 – 100+ people) to the hundreds of millions of dollars required for a Phase 3 study which tests and confirms the effectiveness and safety in larger groups of patients (1,000 – 5,000+).

What would this preclinical efficacy study look like?

In short, administer the compound to mice with pancreatic cancer and see if they get better. The details, however, matter.

There are multiple ways to induce pancreatic cancer in mice – each of which has its pros and cons with respect to the patient population the compound was designed to treat. The goal here is to mimic, as closely as possible, the cellular and molecular characteristics of clinical patient’s tumors. Genetically engineered mouse models such as the KPC (Kras^G12D/+;Trp53^R172H/+;P48-Cre) mice may be considered the gold standard research tool as they allow mice to develop spontaneous pancreatic tumors the closely mimic PDAC, but the tumors are mouse in origin not human. Xenograft models, in which human PDAC cells are implanted (e.g. orthotopic implantation in the pancreas or subcutaneous implantation under the skin) into immunodeficient mice (a necessity to prevent immune rejection), have the benefit of human derived cancer cells, but lack an immune system, a true human tumor microenvironment (the mix of cancerous and non-cancerous cells within a tumor), fail to capture the diversity of patient tumors, and rely on mouse hosts with their faster metabolic rates and anatomical differences. Repeated injections of chemical stressors such as cerulein are well established to cause pancreatitis (precursor to PDAC) and its natural evolution into pancreatic cancer in genetically susceptible mice, but the time course, severity, and genetic basis of the developed tumors are highly variable.

Share The Empirical Evidence Collective

After the model has been chosen, the study design can begin — everything from dosage to the number of study arms. The design is meant to mimic a clinical trial - to show that the compound is safe and effective in the desired patient population. One study arm may be untreated controls to assess the natural progression of the model, while another arm may include the current standard of care (another control), against which the drug in question is being tested in another or multiple (required for testing multiple dosages) arms. Add in a co-therapeutic (e.g. an immune checkpoint inhibitor or chemotherapeutic) and the size of the study (and it associated costs) rapidly increases.

The data required of the study (to address its twofold purpose) would, in this case, be tumor growth inhibition (TGI) equal to or better than that of standard of care (Proof-of-Concept); a TGI equal to or better than daraxonrasib; demonstratable dose-response relationship identifying the minimum effective dose, the therapeutic window, and the maximum effect (E_max); biochemical and biomarker evidence of target engagement in vivo; and the PK/PD data necessary for both predicting safe starting doses in humans and having confidence that the compound will not cause toxicity.

If there is a single takeaway message about preclinical testing it is this: While absolutely essential, it can only answer the safety and efficacy questions that are asked, and the answers provided are often incomplete, a limitation of our current understanding and the models used.

The Vioxx Case Revisited: When Preclinical Data Are Convincing and Still Wrong

There are two COX enzymes, COX-1 and COX-2, that are structurally very similar and that biochemically both produce prostaglandins. COX-1 in an enzyme that produces prostaglandins that aid in creating the protective mucus layer of the stomach, aid in activating platelets to form blood clots, amongst other biological functions. COX-2 is an enzyme that produces prostaglandins which mediate inflammation and pain, amongst their other biological functions. Vioxx (rofecoxib) was developed by Merck as a selective COX-2 inhibitor. Selective COX-2 inhibitors were designed to provide anti-inflammatory and analgesic effects without the gastrointestinal toxicity associated with COX-1 inhibition as traditional nonsteroidal anti-inflammatory drugs (NSAIDs) such as ibuprofen, naproxen, and aspirin inhibit both COX-1 and COX-2.

The preclinical data for rofecoxib were, by all accounts, solid. The compound demonstrated high selectivity for COX-2 over COX-1 in vitro. It reduced inflammation and pain in standard animal models. Genotoxicity studies were negative. Repeated-dose toxicity studies in rodents and non-rodents identified no unexpected adverse effects at therapeutic doses. The cardiovascular risk that emerged post-market — an increased risk of myocardial infarction and stroke — was not predicted by preclinical studies because the underlying biology was unknown at the time.

COX-2 inhibitors suppress endothelial prostacyclin (a vasodilator and inhibitor of platelet aggregation) while leaving COX-1–mediated thromboxane A2 production (a vasoconstrictor and platelet aggregator) unaffected. This imbalance biases individuals (already susceptible to thrombosis formation) towards a pro-thrombotic state, increasing the risk of heart attack or stroke. In vitro COX-2 selectivity assays do not reveal this mechanism, and cardiovascular outcome models were not part of the standard preclinical toolkit. The signal existed in the biology, but the tests in place could not detect it.

The signal emerged from the APPROVe (Adenomatous Polyp Prevention on Vioxx) trial which began enrolling patients in February of 2000. Merck halted the trial on September 24, 2004, after an interim analysis demonstrated that rofecoxib was associated with a relative risk of 2.38 (95% CI 1.39 to 4.09) for confirmed cardiovascular thrombotic events. By that point, rofecoxib had been on the market for approximately three years and had been prescribed to an estimated 80 million patients worldwide, generating more than $2.5 billion in annual sales. On September 30, 2004, Merck announced the voluntary worldwide withdrawal of Vioxx.

The Vioxx tragedy was not caused by sloppy preclinical testing. Similar to other incidents, its causes were (in hindsight) multifactorial — a combination of the limits of what preclinical testing (at the time) could reveal and ethical lapses by both Merck and the FDA. The pro-thrombotic consequence of COX-2 selectivity was a sophisticated mechanism, and its cardiovascular implications were not predictable from the preclinical assays and animal models available in the late 1990s - demonstrating the necessity of sound ethical judgement within DSMBs once the compound is administered to humans. The case illustrates, with unusual clarity, the central paradox of preclinical testing: data can be thorough and accurate, but inadequate.

I will be stepping away from my laptop for a few weeks while take some time off to attend to distant family, to digitally detox, to recharge. In other words, a summer vacation.

What this means for you:

• No new posts will go out between June 8th and July 8th

• Regular publishing will resume on July 12th with our usual schedule

For those of you who have been here since the beginning — thank you. I just launched The Empirical Evidence Collective last month, and I’m honestly enjoying the process of figuring this out as I go. But, a break will be good for me, and I think it’ll show up in the writing when I come back.

Catch you on the flip side.

— Ron

In this phase 3, international, open-label, randomized trial, we randomly assigned patients with previously treated mPDAC to receive daraxonrasib or chemotherapy of the investigator’s choice. The dual primary end points were overall survival and progression-free survival in the subpopulation of patients with RAS G12 mutations (the RAS G12 population). Key secondary end points included overall survival and progression-free survival in the overall population (which included patients with RAS G12, G13, or Q61 mutations or with no RAS mutation identified) and objective response and patient-reported quality of life in the RAS G12 and overall populations. Safety was also assessed.

Note – in the language of scientists (clinical trials), the primary endpoint is the main, pre-defined outcome that the study is designed to measure; it is the metric for defining the success or failure of the trial. Secondary endpoints are used to evaluate additional questions providing supporting or supplementary evidence about the therapy in question.

What did they observe?

Out of a total of 500 patients – 91.8% with RAS G12 mutations – the median overall survival of those treated with daraxonrasib was 13.2 months, compared to 6.6 months for those treated with chemotherapy. The median progression-free survival – the duration of time that a patient lives with disease without it spreading or getting worse – increased from 3.5 months for those treated with chemotherapy to 7.3 months for those treated with daraxonrasib.

While it may not seem like it, daraxonrasib treatment provided an enormous improvement over standard of care chemotherapeutics in median survival time. Cancer is a devastating disease, and pancreatic cancer, especially so.

Pancreatic ductal adenocarcinoma (PDAC) is the most common form of pancreatic cancer, accounting for approximately 90% of all pancreatic cancers. According to data from the American Cancer Society, an estimated 67,000 men and women will be diagnosed with pancreatic cancer in the U.S. in 2026. Its 5-year survival rate is just 13%. That far lower than cancers of the lung (28%), liver (22%) and esophagus (22%), which have seen large increases since the 1990s. Over the same timeframe, the median overall survival has held steady at approximately 6.6 months.

Why didn’t improvements in the treatments for other forms of cancer translate to pancreatic cancer?

The short answer is it is frequently diagnosed at later stages – upwards of 80% of cases are diagnosed after it has spread beyond the pancreas (clinical stages III and IV) – and its unique biology makes it notoriously difficult to treat effectively. The pancreas is deep within the abdomen – behind the stomach and intestines – making it difficult to feel during physical exams. Because of this location, tumors can grow significantly without pressing on the surrounding tissues or nerves – often without any symptoms or symptoms (e.g. bloating, indigestion, slight weight loss) that closely mirror common, less-serious gastrointestinal issues. And unlike colon or breast cancer, there is no widely available, cost-effective screening tool. By the time it is diagnosed, it has often spread beyond the pancreas making surgical removal difficult.

Let’s talk about that biology for a moment.

PDAC is unique because the majority of the tumor mass is not made up of cancerous cells, but rather is comprised of non-cancerous stromal – or supporting – cells matrix of extracellular proteins. This is what is referred to as the tumor microenvironment, or TME, and in pancreatic cancer, the TME is unique. It is denser and stiffer than that of other tumors, effectively creating a barrier against which many drugs cannot pass, it is extremely hypoxic (blood vessels, and the oxygen it brings, are scarce), and it is heavily populated with immunosuppressive cells creating an “immunologically cold” tumor. In the language of biologists, this means that the tumor is not recognized by the body’s immune system.

In addition, KRAS, that previously “undruggable” target we discussed previously

is mutated in over 80% of PDAC cases – initiating tumorigenesis by promoting unchecked cell proliferation and survival. In that prior article, we discussed how sotorasib was developed to specifically target KRAS G12C (a mutant form of KRAS that is preferentially in the “ON” or active state). Daraxonrasib, rather than being selective for a mutant form of KRAS, is a RAS(ON) multi-selective inhibitor. In the language of biologists this means that it targets (binds) RAS proteins (e.g. KRAS, HRAS, and NRAS) preferentially in the “ON” state.

What further differentiates daraxonrasib from the specific KRAS G12C inhibitors is that it also binds a protein called cyclophilin A – essentially forming a tripartite complex (cyclophilin A:daraxonrasib:RAS(ON)) – that prevents active RAS from interacting with its downstream effector proteins, thereby blocking the ability of active RAS to activate oncogenic MAPK/ERK and PI3K/AKT signaling. And the binding of this tripartite complex is extremely strong, stronger than many other drug-protein complexes. Scientists have a name for this molecular mechanism of action: molecular glue.

And the significance of cyclophilin A?

Cyclophilin A (CypA) is a highly conserved, ubiquitously expressed protein that assists in the proper folding of nascent proteins and facilities their trafficking within cells. In essence, it is a molecular chaperone, helping proteins both attain their functional conformation and get to where they are needed. It also functions in the extracellular space where it is involved in leukocyte recruitment at sites of injury, thrombo-inflammation and atherosclerosis, in part through its interaction with the CD147 receptor.

In the context of PDAC, where its expression is greatly increased in response to the localized hypoxic environment, CypA acts as a key tumor-promoting protein, accelerating cellular proliferation. CypA is also a promising drug target in its own right because increased CypA-CD147 signaling promotes cellular proliferation and survival through the activation MAK/ERK and PI3K/AKT signaling - the same pathways activated by oncogenic RAS. Daraxonrasib essentially binds two PDAC targets – inhibiting both simultaneously.

This new class of medicines – molecular glues – is beginning to show what is possible when drug discovery begins to look outside of a canonical binding pocket. Thalidomide, infamous for causing birth defects in the 1960s, functions as a molecular glue. It forms tripartite complexes with the cellular enzyme Cereblon (CRBN) and IKAROS zinc finger proteins (IKZF1 and IKZF3) – transcription factors that regulate gene express – rendering them inactive, depriving myeloma cells of survival signals halting its growth. Several other compounds in this class (e.g. mezigdomide (CC-92480) for multiple myeloma and CC-90009 for acute myeloid leukemia and myelodysplastic syndrome) are currently in clinical trials. While we will have to wait until the clinical trial results are in before molecular glues can be declared effective in these indications, the positive results with the daraxonrasib in the current PDAC trial are encouraging.

Importantly, daraxonrasib was shown to be as safe (safer with respect to some measures) and better tolerated than chemotherapy and better tolerated:

Adverse events that were considered by the investigators to be related to the trial treatment occurred in 97.9% of the patients in the daraxonrasib group and in 93.5% of those in the chemotherapy group. Treatment-related adverse events of grade 3 or higher were less common with daraxonrasib than with chemotherapy (in 43.6% of the patients vs. 57.5%), as were serious treatment-related adverse events (in 10.8% vs. 18.7%). The incidence of treatment-related adverse events that led to dose reduction was also lower in the daraxonrasib group than in the chemotherapy group (in 36.1% of the patients vs. 57.5%).

While impressive – and indeed hopeful – a doubling of the median survival time is not a cure. It will not keep anyone who reads this post from lying awake at night if they or someone they love has received this diagnosis. But, this is something that did not exist before – concrete evidence that the biology of pancreatic cancer can be breached.

The 13.2 months in this trial is extra time – extra phone calls with grandchildren, extra Saturday morning coffees, extra uneventful days that most of us take for granted – something that every pancreatic cancer patient fights for. This is not nothing. This shows us the cracks in the armor that PDAC has wielded in its defense against all prior treatments.

What is this obscure set of administrative rules and why does this matter?

Commonly called the Uniform Guidance, this is the Office of Management and Budget’s (OMB) government wide framework for managing federal financial assistance (e.g. grants and cooperative agreements) awarded to non-federal entities (e.g. universities, research institutes, and non-profit organizations). Part 200 creates a single, government wide rulebook for how federal awards are announced, awarded, managed, paid, monitored, closed out, and audited. It also defines the cost principles associated with federal funding – essentially answering the question: “Can this cost be charged to the federal award?”

The proposal is the most significant rewrite of the federal grant rules in over a decade. What was once an apolitical process will now require a political pre-issuance review. Let that sink in for a moment. Senior (i.e. political) appointees will now be required to review all discretionary federal funding awards – everything from studying oncogenes to the epidemiology of disease. Science, the process by which we seek to understand and predict natural phenomena, is inherently apolitical. The proposal includes a new section that prohibits federal grants from promoting or supporting disparate impact liability based on race, sex, or age. While it may be politically inconvenient to those in government, sex, gender, genetic background, age, ethnicity and country of origin, socio-economic status, environmental & behavioral factors, and healthcare access – individually and collectively – influence the development of disease in an individual and the overall incidence of disease within the U.S. Understanding how these factors influence human biology and physiology helps scientists and physicians identify the cause of disease, develop novel therapeutics for treatment, more accurately diagnose, and more effectively treat disease at the level of the individual patient. Over time, this ultimately translates to more effective treatment strategies across the entirety of the U.S. to everyone’s benefit.

Before one can fix something, one must first understand how that something is supposed to work. This seems so logical, so rational, it’s disorienting to feel the need to restate this explicitly. But perhaps that’s the logic behind the proposed changes – essentially, the politicization of the scientific infrastructure that we have long taken for granted as being apolitical. Coupled with the recently proposed funding cuts to Health and Human Services – which includes the NIH, FDA, CDC, and smaller, low-profile, but highly effective agencies such as the Agency for Healthcare Research and Quality (AHRQ) and the Administration for Strategic Preparedness and Response (ASPR) – one cannot help but think that this is part of a concerted effort to corrode the very concept of apolitical scientific research.

It doesn’t have to be like this.

Historically, the U.S. government has played an enormous role in promoting scientific discovery and technological development. In 1944, the expansion of The Public Health Service Act formally incorporated the National Cancer Institute into the National Institutes of Health, and provided a framework for allocating federal funds to study and fight the diseases afflicting the U.S. Since the end of WWII, academic biomedical research has grown steadily, studying everything from the cellular and molecular biology of cancer and HIV to the epidemiology of drug addiction, alcoholism, and disease.

Today, the NIH is the largest public funder of biomedical research in the world; its funding has supported 51% and 59% of patents in the pharmaceutical and biotechnology industries, respectively. In the 2024 fiscal year, the NIH awarded over 65,000 extramural grants and contracts with its $47 billion dollar budget. Each grant is reviewed by a panel of scientific experts on the basis of merit for significance, innovation, approach (scientific rigor and feasibility), as well as the expertise of the investigator and the availability of resources. While an extremely competitive process – the overall success rate across all NIH institutes is approximately 12% – scientific merit, not political whims, is the deciding factor in whether a grant is funded.

These grants cover the salaries of postdoctoral fellows, the workhorses of academic research whose NIH mandated salaries in 2024 ranged from $61,008 for a newly graduated Ph.D. to $64,356 for those with 3 years of experience, research technicians, as well as equipment and reagent costs to perform the work in the approved grant. One thing that often goes unmentioned in any discussion of NIH grant funding is indirect (overhead) costs, those unsexy operational expenses that cannot be easily attributed to a single specific project but are necessary for that organization to function. Indirect costs include everything from electricity and water to the institute’s support staff (e.g. grant administrators and facilities management). There are hundreds of rules that must be followed, and sometimes things break or malfunction. This is indeed an important topic – in January the U.S. Court of Appeals for the First Circuit upheld a district court’s decision to permanently enjoin and vacate guidance published by the NIH in 2025 that would have imposed a flat, 15% indirect cost rate on all NIH-funded research – one with a great deal of nuance that I’ll return to a later date. In addition, grants also cover the costs of publishing research papers, conference fees, as well as the costs associated with travel to and from conferences.

Science, despite what many non-scientists may think, is a community endeavor – it is inherently social in nature. Those pop-culture depictions of scientists as socially awkward solitary geniuses have no basis in reality. Scientists are continually reading the literature – the published works by other scientists in and out of their respective fields – and discussing their research programs and discoveries with other scientists within, and outside, their respective institution. This is the backdrop against which scientific discoveries are made – those very same scientific discoveries that power future technological developments. Every research paper published is a new addition to our collective knowledge. Every review article published is a new summary of a protein, pathway, or a disease process that highlights what is currently known, and even more importantly, identifies the outstanding questions that remain to be answered. Think of it as a roadmap to future discoveries. Every conversation a scientist has with another scientist is a potential collaboration – a new opportunity to solve a biomedical mystery that likely couldn’t be solved alone.

The proposed changes to the Uniform Administrative Requirements seek to dramatically impair scientific discussion of all forms. Revisions by OMB make “publication costs unallowable unless such costs are expressly required by statute or approved in advance by the Federal agency on a case-by-case basis,” and add a requirement that “costs for attending conferences are allowable only if participation in the conference is expressly approved by the agency and included in the terms and conditions of the award” clarifying that “recipients are not authorized to attend conferences using Federal funds that do not serve to advance program outcomes.”

That’s a lot to unpack. Let’s start with the cost of publication – commonly referred to as an Article Processing Charge (APC). Depending on the journal, APCs can run upwards of $10,000. Nature, for example, charges $12,850 to publish an article as open access – something funding agencies are requiring more frequently. If published via the traditional subscription model, the APC is zero, but the article can only be read by individuals, and individuals at institutions, with a subscription. While most journals have substantially lower APCs, they are still in the $1,000 – $3,000 range.

Conferences – especially those for which long-distance travel is required for attendance – is another cost traditionally budgeted into grants. Large, multidisciplinary conferences such as the annual ASBMB meeting cost between $175 for graduate students and between $550 (early registration) and $670 (regular) in attendance fees for academic faculty. Non-members pay an extra fee. And if you don’t happen to live within a reasonable driving distance, airfare and hotel reservations are an additional cost to attending these multi-day conferences.

Why is this an issue for scientists? Can’t they just decide where to publish in advance and include that in their grant application? Can’t they just decide which conferences to attend in advance?

I wish it were that simple.

The truth is that nobody can truly predict the outcome – much less its impact on the field – or when that outcome will occur, of a research program. The typical R01 grant – the major funding mechanism from the NIH – supports academic research labs across the country and typically lasts for 5 years. Contained within each R01 application is thorough description of the Principal Investigator’s (PI – typically a professor) overarching hypothesis, the scientific approach at testing that hypothesis, the anticipated results of the proposed experiments, a proposed timeline of when those anticipated results should be published, and a detailed budget outlining anticipated costs for publishing and attending conferences. What is not described – with any certainty – is what journals (and therefore costs in the form of APCs) the research will be published in, and what conferences (and therefor costs) will be attended. It is the impact on the field – the practical significance of a particular set of data that changes how scientists think about a phenomenon – not some predetermined notion, that determines where a manuscript will be published. Aside from the one, perhaps two conferences per year that are perfectly aligned with your field, most conference attendances are not planned years in advance. I don’t know of anyone who could predict with any level of certainty when, and in what journal, their research – that has not yet begun (this is the purpose of the grant, to fund new research programs) – will be published.

Share

Research, like many other endeavors, rarely happens on predetermined timelines, and research, unlike other endeavors, often explores tangents – this is where much of the most interesting science occurs. It is these tangents that have so often opened up new fields (i.e. RNAi) and new technologies (i.e. CRISPR). To predict the outcome, and the significance of that outcome with respect to the field such that one could predict the journal in which it would be published prior to the initiation of the research is all but impossible. To be forced to plan with this much detail in advance is to essentially eliminate the possibility of exploring these tangents.

Conferences are networking events for scientists – it’s where they go to publicly discuss their research in the form of talks and poster sessions, it’s where they go to meet others in their field and build collaborations. I remember when I was a postdoc looking at upcoming conferences and their abstract submission deadlines. I also remember when I was a PI, being invited to give a talk (never turn down an invited talk) and helping my postdocs with their abstract submissions. Abstract submissions may be planned for a year out, perhaps two, but not three to five. Being invited to give a talk – especially for junior faculty – is never something that can be anticipated.

Lastly, and perhaps most insidiously, the proposed changes expand the authority for the “discretionary termination” of federal funds. Modeled on the Federal Acquisition Regulation (FAR) Termination for Convenience (T4C), it allows for the unilateral termination of existing awards. And because this power will reside in senior (i.e. political) appointees, even currently funded scientific initiatives will now be subject to the prevailing political winds. While I cannot predict the future, I can certainly – albeit cynically – see the potential termination of grants that do not align with the ideology of the current administration.

It doesn’t have to be like this.

This story of one compound being developed, tested, and demonstrated to be both safe and effective is the exception, rather than the rule. The reality looks more like this: anywhere from 10,000 to 1,000,000 compounds tested result in a single marketable drug. For every chloroquine, there are 999,999 molecules that go nowhere. It binds the wrong target. It does not produce a biological effect. It has poor solubility resulting in high doses. It is toxic at the doses required to produce a biological effect. It has low cellular permeability and therefore is unable to bind its target in vivo. It is rapidly metabolized and rendered ineffective.

The odds for a preclinical candidate reaching approval are just as steep; perhaps not in number of compounds, but in terms of risk. For every compound that gains regulatory approval, roughly 1000 enter a Phase I clinical trial. The financial costs of this development have increased from an average of $127 million (2018 dollars) to nearly $515 million when the cost of failures is included. The hunt is the second great bottleneck of drug discovery and is arguably the more brutal of the two; unparalleled levels of failure are the statistical norm.

In Part 1, we covered how a target is chosen: which protein, gene, or pathway should be modulated to treat a disease. That consequential decision determines the entire trajectory of a drug development program. But even the perfect target – the genetics of PCSK9 or mechanistic biology behind KRAS – essentially becomes an academic exercise if you cannot find a molecule that selective binds it.

For decades, the way scientists went looking for compounds was fundamentally a numbers game in the form of a screen. Each compound in a library of compounds, anywhere from 50,000 to up to a million, is tested individually against the target in an automated assay. If it produces the desired effect, it is referred to as a “hit.” A hit is not a drug; it is a chemical starting point which must be refined until it is potent, selective, and drug-like. It is a cycle starting with synthesis of the compound, testing various parameters (e.g. solubility, potency, specificity, etc.), analysis and modification resulting in the next round of synthesis. This can take dozens, even hundreds of cycles; each one taking weeks.

Artificial intelligence is changing this process as well. In April 2026, McMaster University redefined the scale of that hunt. Their artificial intelligence model, SyntheMol-RL, explored 46 billion possible compounds – that’s 46-times larger than the largest physically available chemical libraries – in silico; of which 79 compounds were physically synthesized, 13 showed potent in vitro activity, and 1 compound, synthecin, demonstrated efficacy in an in vivo mouse model of methicillin-resistant S. aureus (MRSA). Traditional high-throughput screens take months to assess a million compounds, but the McMaster AI model performed an in silico screen of 46 billion possible compounds in a fraction of that time. That is an enormous number of compounds, but only a drop in the bucket when compared to the entirety of chemical space estimated to contain at least 10⁶⁰ possible compounds. Finding the right compound for the right target in all of chemical space is like finding a needle within the observable universe.

Subscribe now

From Target to Molecule: The Hit-to-Lead Pipeline

The drug discovery process officially begins once a target has been validated. The language used to describe this journey is unique to the drug discovery ecosystem, and understanding it helps to explain why the odds of success are not favorable.

A hit is a compound that interacts with the target in preliminary screening assay. It is typically weak, binding at micromolar or sub-micromolar concentrations, and almost never specific enough to be a therapeutic agent. This is the starting point. For a given target-library combination, observing 2 to 5 hits may be considered a success, sometimes a single hit is observed, and sometimes its zero.

A lead is a hit that has been chemically modified to improve its potency and selectivity. The lead still has flaws – solubility issues, metabolic instability, poor absorption – but it has demonstrated that it can do the one thing you want it to do – bind its target and modulate its activity – with sufficient affinity that further optimization is plausible.

A clinical candidate is a lead that has survived the grueling process of lead optimization and has passed the initial preclinical safety and efficacy screens. By this point, it has typically been through anywhere from 50 to 200 rounds of synthetic chemical modification and has been tested in dozens – potentially hundreds – of in vitro and in vivo assays. It is now ready to enter Phase I clinical trials.

It is this phase between identification of that initial hit and clinical candidate where most drug development programs die. A hit is not a drug. Neither is a lead. Same for a clinical candidate. These are steps, named steps, along the way towards a drug. Assuming one finds a hit that is readily amenable to modification within the bounds of the target’s binding pocket, the core optimization problem remains – to become a clinical candidate, the molecule must satisfy five independent requirements simultaneously:

Target Engagement & Efficacy – It must both selectively and strongly bind its biological target, alter its function, and demonstrate robust efficacy in relevant cellular and animal models. Weak binding means you need higher doses to achieve efficacy, which often translates to more side effects and a greater risk of severe dose-related complications. Non-specific binding to other related, or unrelated, biological molecules increases the risk of off target effects, which similarly translates into more side effects and greater risk for severe dose-related complications.

Acceptable Safety Margin – It must pass formal, regulatory-compliant toxicology testing – typically performed in both rodent and non-rodent species – to identify a clear no-observed-adverse-effect-level (NOAEL) dosage. This is the maximum dosage at which there is no biologically or statistically significant increase in the frequency or severity of adverse effects. This is critically important because it helps to set a starting point to establish safe clinical dosages in humans.

Favorable Pharmacokinetics – It needs sufficient bioavailability, meaning it must be properly absorbed, distributed, metabolized, and excreted (ADME) such that it reaches the target tissue at a therapeutic concentration. In the language of drug discovery professions, this means that it must be absorbed through the gut (if taken orally), distributed to the right tissue, metabolized at an appropriate rate, and excreted without generating toxic byproducts. In short, it must survive long enough in the body to reach its target without producing toxic byproducts.

Favorable Drug Interaction Profile – It must show a low likelihood of causing harmful interactions with other common medications. In other words, the patients likely to benefit from this compound (should it become approved) are likely to be currently on other medications – this new compound (and its metabolic byproducts) should not interfere with existing compounds (and their metabolic byproducts) already present within the patient.

Feasibility of Manufacturing – It must be a molecule that a chemist can actually make at scale. Elegant structures that exist only in a computer simulation are useless if they cannot be synthesized in a laboratory, let alone manufactured in kilogram quantities. In other words, there must be a scalable method to synthesize and formulate the drug under Good Manufacturing Practice (GMP) standards to guarantee consistent purity, stability, and delivery for patients.

The difficulty is that optimizing for one property almost always degrades another. Increase potency by adding a hydrophobic group, and solubility plummets. Improve metabolic stability by making the molecule bulkier, and it can no longer cross cell membranes. Enhance selectivity by refining the shape to fit a single pocket, and you may have inadvertently reduced potency. Every change is a trade-off. Every trade-off is a new synthesis, a new round of testing, and a new set of results that tell you which way to go next.

This, often referred to as the design-make-test or the build-make-test cycle, is the core of drug discovery – an exercise in iterative compromise similar to tuning an instrument blindfolded, where every adjustment to a single string causes the others to slightly detune. And it can take years – two to five on average. During this time, medicinal chemists may have synthesized hundreds of compounds – each one tested for target engagement and efficacy in in vitro assays. Those that passed these initial tests were then assessed for ADME and other toxicological assessments. Each of these compounds is a slight variation of a theme, searching for that one combination of chemical properties that is sufficient across all dimensions simultaneously.

Share

Five Ways to Find a Hit

Once the target is validated, the hunt begins. While there is no single way to conduct this hunt, over the decades several distinct methodologies have emerged – each with its own strengths, weakness, and modes of failure.

High-Throughput Screening: The Sledgehammer Approach

High-throughput screening (HTS) has been the dominant approach in drug discovery for nearly thirty years. The concept is straightforward: test hundreds of thousands, sometimes millions, of compounds against the target in automated assays, and whatever produces the desired effect is a hit. An understanding of how the compound works – not necessary. A structural model of the target – not necessary. A deep understanding of the target-related biology – not necessary. While these are helpful at later stages of the process, to identify a hit with HTS, all that is required is a robust, automatable assay that reports a quantifiable signal when the target is engaged, and a library of compounds to test.

HTS is the sledgehammer because it is the brute force approach – it tests everything. And it works with any target with a quantifiable readout – enzyme activity, receptor binding, cell death, protein aggregation, DNA fragmentation – can be screened. Modern day libraries are huge; major pharmaceutical companies and contract research organization (CROs) often have libraries – every chemical they have ever synthesized for any target, naturally occurring compounds, approved drugs – containing millions of distinct compounds.

The fundamental limitation of HTS is that it can only test molecules that already exist, those chemical compounds present in an accessible library. And even the largest physical libraries available are vanishingly small when compared to the total chemical space estimated to be on the order of 10⁶⁰ compounds. Typically, the HTS hit rate is less than one percent, meaning that over 99% of the compounds test fail to produce a quantifiable effect. In addition, false positives are common, requiring secondary assays to eliminate from the pool of potential hits. Even among genuine hits, the vast majority bind weakly, have poor ADME properties, or have been previously shown to be toxic. From these, perhaps two or three transition to a lead, and from those leads only one or two (if extremely lucky) might make it to clinical candidate.

HTS is slow, expensive, and fundamentally a numbers game – essentially throwing compounds at the wall and seeing what sticks. While it has been enormously successful, producing more approved drugs than any other approach, the attrition rate staggering. While a single marketable drug from a million screened compounds is an acceptable rate of return when the alternative is the absence of a drug, scientists have been working on alternatives for decades.

Fragment-Based Drug Design: Building Up, Not Screening Down

Fragment-based drug design (FBDD) inverts the HTS approach. Instead of testing large, complex molecules against the target, FBDD tests fragments of complex molecules – simple low-molecular-weight compounds that typically interact with the target very weakly. This approach similarly requires a robust, automatable assay that reports a quantifiable signal when the target is engaged, but because fragments of compounds – rather than entire compounds – are tested, it has a higher probability of target engagement. Once a fragment is observed to bind – typically confirmed with X-ray crystallography or NMR spectroscopy which reveal how the fragment sits within the target’s binding pocket – medicinal chemists can add additional fragments or small chemical groups to build a larger, more potent molecule piece by piece.

The advantage with FBDD is efficiency. Each fragment has a high probability of target engagement because it is small and simple, and the structural biology techniques used to verify where and how the fragments bind within the target’s binding pocket are also used to guide the next round of optimization. The disadvantage is that FBDD requires high-quality protein structures – the equipment required for this is extremely expensive and is not readily available. X-ray crystallography, for example, requires taking crystallized protein samples to a synchrotron facility such as Brookhaven National Laboratory. In addition, may targets – particularly membrane proteins (e.g. receptors) and large biological complexes are so difficult to crystallize that they are routinely considered un-crystallizable.

Despite these limitations, FBDD has produced some of the most important modern targeted therapies. Venetoclax, a BCL-2 inhibitor used to treat chronic lymphocytic leukemia, traces its origins to an NMR-based fragment screening program at Abbott that first identified ABT-737 in 2003. The method has now become a standard tool in the medicinal chemist’s arsenal, particularly for targets where high-resolution structural data is available.

Structure-Based Drug Design: A Computerized Lock and Key

Structure-based drug design (SBDD) merges the insights provided by FBDD with computational brute force. Starting from a three-dimension structure of the target protein – determined through X-ray crystallography, cryo-electron microscopy, or NMR – computational algorithms are used to dock hypothetical compounds within the target’s binding pocket. The more computational power, the more compounds can be designed in silico, from tens of thousands of to millions. Algorithms score each candidate molecule for its predicted binding affinity and structural fit – the top ranked compounds are then synthesized and experimentally tested in the laboratory. Computational docking has become far more accurate over the past decade, aided by advances in molecular dynamics simulations and machine learning. In silico screening of millions of compounds is now routine, and in some cases, this type of screening followed by targeted physical validation has produced leads faster and more efficiently than traditional HTS.

The SBDD was the approach used to develop the KRAS inhibitors described in Part 1. Once the cryptic binding pocket on KRAS was idenfied in 2013, structure-based design was the primary methodology used to develop sotorasib and adagrasib – molecules designed to fit precisely into a pocket that had never been targeted before. Despite these successes, the predictions are still imperfect. Docking software is excellent at finding molecules that look like they fit, but it cannot fully capture the thermodynamics of protein-ligand interactions — the water displacement, the conformational changes that occur upon binding, the entropic costs of rigidifying a flexible molecule into a binding pocket. Hence the continued need for physical screening and synthesis to validate computational predictions.

Share The Empirical Evidence Collective

Generative AI and Reinforcement Learning: Designing Molecules from Scratch

The most recent and most ambitious approach does not search utilize existing libraries at all, but rather generates compounds de novo. Generative AI creates entirely new molecules optimized against a specific target with a predefined set of properties.

Machine learning models for molecular design are trained on millions of known chemical structures and their measured properties. They learn the “grammar” of molecules – the atomic arrangements that lead to good binding affinity, poor solubility, rapid metabolic degradation, toxicity. Given a target structure and a set of desired properties, the model generates candidate molecules that optimize for those properties simultaneously. Basically, instead of asking “which of these known compounds binds the target?”, it is essentially asking “what molecule can we design that will both bind the target optimally while minimizing undesired side effects?”

The McMaster University result described above – the SyntheMol-RL model and the antibiotic synthecin – is the most recent example published. But this is not an isolated case. In 2024, a Nature Communications paper described RosettaVS, a structure-based virtual screening platform that performed an in silico screen of approximately 6 million compounds from the Enamine REAL library. Using an open-source AI-accelerated approach, the authors significantly reduced the time required to develop compounds capable of binding and inhibiting the KLHDC2-C29 and the human voltage-gated sodium channel, hNa_V1.7. Another recent study describes a framework that merges deep-learning with existing chemical libraries to identify already-existing compounds that may be effective antibiotics. A 2025 review article highlights the accelerating pace of AI-driven molecular design across drug discovery, noting that recent advancements are showcasing “the transformative potential of generative AI models in revolutionizing core techniques like molecular docking within lead discovery workflows.”

While its promise is clear, challenges remain. Perhaps the most pressing challenge is that of synthetic accessibility; a computationally elegant molecule is useless if it cannot by physically synthesized in the lab. The SyntheMol-RL model specifically addressed this by optimizing for both target binding and synthetic accessibility, resulting in synthecin, a compound that was both synthesized in the laboratory and demonstrated to be effective against methicillin-resistant S. aureus (MRSA). The gap between in silico design and in vivo validation is rapidly being bridged by AI-enabled biotechnology companies and academic research labs worldwide.

DNA-Encoded Libraries: The Screening of Trillions

DNA-encoded libraries (DEL) represent a hybrid approach; far greater scale than HTS with less a priori knowledge of the target when compared to SBDD, FBDD, and generative AI. In a DEL, each molecule in the library is tagged with a unique DNA barcode and pooled together. This pooled library is that incubated with an immobilized target protein; molecules that do not bind are washed away, while those that do bind remain attached to the target. The bound complexes are then isolated, and the DNA tags of the bound molecules are read via RNA-sequencing techniques.

DEL is capable of screening trillions of compounds in a single experiment. A paper published in 2017 demonstrated the large-scale application of this technology by screening over 2.4 x10¹² compounds composed of 4 – 20 natural and non-natural amino acids against the therapeutic targets VHL and RSV N. While this is indeed impressive, the trade-off of this is that screening conditions are artificial – target proteins are immobilized on an artificial substrate without cellular context. The cellular context-dependent conformational changes or protein-protein interactions of the target protein are simply not present – what is present is a stationary view of the target; something that would never occur in a live cell. In addition, there are no membranes for the compounds to cross and no enzymes that could alter or otherwise degrade the compounds. As a result, many DEL hits do not translate into functional drugs in cellular or animal models. However, utilizing DEL as a complementary approach, combined with machine learning and structural biology methods may help to narrow down the list prior to physical validation.

The Build-Make-Test Cycle

This is arguably the most critical part of the hunt – where most drug discovery programs get stuck. It is important to remember that a hit is not a drug. It’s not even close. This is where medicinal chemists take the hit molecule and change one part of its structure at a time – add a methyl group here, swap a carbon for a nitrogen here, perhaps extend or shorten a side chain – and test its effects. They quantify what improved (potency, ideally) and what got worse (solubility, likely), modify the structure again, and repeat. Molecules that pass these most basic – but essential – tests move on to in vitro testing in cell model systems. This repeats dozens to hundreds of times – each cycle taking weeks. And there is no guarantee that this process will result in a clinical candidate – the end goal of the hunt.

It is entirely possible that there is no molecule that satisfies all requirements (e.g. ADME, potency, etc.) simultaneously. Everything is a tradeoff, a delicate balancing act in multiple dimensions. And there is no single right way to identify drug-like molecules. It is essentially an optimization problem across multiple dimensions; trial and error in its purest form. Historically, human intuition and experience have played a role in this iterative process, although with recent advances in machine learning, algorithms have been employed to augment human intuition and experience. Two examples of how algorithms – similar to those discussed previously in Part 2 for target validation – are beginning to enhance the Build-Make-Test cycle:

The first is the SPARROW algorithm, developed at MIT and published in Nature Computational Science in June 2024. SPARROW (Synthesis Planning and Rewards-based Route Optimization Workflow) uses Bayesian optimization to automatically prioritize which molecules are worth synthesizing. The algorithm identifies these molecules by balancing the reward of a molecule’s predicted properties (e.g. solubility, potency, etc.) against the cost of its synthetic route – the number and complexity of the chemical reactions required to synthesize the compound.

The second is the self-organizing map–based prediction of drug equivalence relationships (SPiDER) that merges the concepts of self-organizing maps, consensus scoring and statistical analysis. Trained on both known drugs and computer-generated molecules, it is capable of identifying not just potential molecules that bind the target, but also off-target liabilities – those other proteins similar to the target that may unintentionally be bound by the molecule being tested.

It is important to note that these tools do not replace medicinal chemists but rather provide a more comprehensive overview of each molecule – and its predicted properties – prior to synthesis and testing. These tools can save drug discovery programs time – which equates with money – by testing fewer compounds, and money by only synthesizing and testing those which are most promising in silico. This is human intuition and experience strengthened with data.

Preclinical Efficacy Testing – The Last Stage to Clinical Candidate

Getting a molecule optimized for target engagement, specificity, potency, ADME, and other properties of drug-like molecules is extremely challenging. The last stage for a molecule to become a clinical candidate is to demonstrate that it is safe and effective in vivo – before it makes it to the clinic. Tune into Part 4 for more.

Subscribe to The Empirical Evidence Collective

The White House’s FY2027 budget, released on April 3, 2026, proposes a 12.5% decrease from the enacted 2026 budget of $126.9 billion to $111.1 billion in discretionary funding. This includes $4.3 billion in cuts to Global Health Programs including PEPFAR, the wildly successful global health initiative credited with saving nearly 26 million lives and preventing millions of new HIV infections, $4.98 billion in cuts to the Administration for a Healthy America (AHA), and $5 billion – a 10-12% reduction – in cuts to the National Institutes of Health (NIH). This came after the FY2026 budget request originally proposed $17 billion in cuts to the NIH (which Congress rejected), meaning the administration had already moderated its position once before. This is on top of the nearly 38% reduction ($129 million in cuts) to the Agency for Healthcare Research and Quality (AHRQ) which was founded to make healthcare safer, more accessible, equitable, and affordable, and a nearly 10% reduction ($356 million) in funding to the Administration for Strategic Preparedness and Response (ASPR) which was founded to the lead the nation’s medical and public health responses (and preparations for) to disasters, disease outbreaks, and bioterrorism.

All this occurring at a moment when the nation is experiencing an ever-growing measles epidemic in the continental U.S., an active Ebola outbreak in Africa, and an increase in the incidence rates of certain types of cancer, including breast and colorectal, amongst people aged 50 and younger ¹.

So, why cuts of this magnitude, and why now? On May 21 NIH Director, Jay Bhattacharya, MD, PhD, intended to answer questions posed by law makers during a Senate appropriations subcommittee on the NIH’s FY2027 budget. Instead, many of the senator’s questions directed at Bhattacharya were focused on more immediate concerns – the leadership vacuum at the agency’s National Institute of Allergy and Infectious Diseases (NIAID), the slow pace of funding – grant awards and payments – over the course of the current fiscal year, and the impact of disparate administration policies on the research community. Gaining a thorough understanding of the senators’ concerns will help us to understand the administration’s rationale behind the cuts (if one exists) and where it thinks we (as a nation) will be heading at the end of FY2027.

On one the most consequential questions posed to Bhattacharya was whether the U.S. was prepared for the next infectious disease outbreak. This is neither an abstract question nor a hypothetical one. At the time of his testimony, the CDC was implementing travel restrictions from countries in East and Central Africa (where the current Ebola outbreak is occurring) to prevent Ebola from entering the U.S. While this is indeed a prudent preventative measure, one should not forget that this measure was brought about, in part, because the disease surveillance networks that would have caught the outbreak at a much earlier stage have been crippled by funding cuts to the CDC, USAID, and the U.S. withdrawal from the WHO. Although these cuts did not cause the current epidemic – mother nature is decidedly indifferent to human health and wellbeing (politics be damned) – our values and priorities (the values and priorities of our elected officials) did exacerbate the current outbreak.

Since President Trump began his second term as president in January 2025, there has been an unprecedented loss of scientific expertise and knowledge from the NIH. In January 2025, Jeanne Marrazo (the successor to Anthony Fauci as NIAID director) was fired. On May 21, Dr. Jeffery Taubenberger, who has been serving as acting head of NIAID, stepped down. It is unclear if this was voluntary, or if he was asked to do so. With these departures, scientists in leadership positions will have been required to vacate their positions, including scientists in 8 of the 10 most senior leadership positions. This comes on top of over 10,000 doctoral-trained experts in science and related fields who have lost (or left) their jobs amongst the federal workforce ². While this number represents a small fraction (~3%) of the total federal workforce, it is ~14% of the total number of Ph.D.s in science, technology, engineering, and math (STEM) and health related fields that were employed at the end of 2024.

While this loss of knowledge and expertise is unprecedented, the loss of institutional knowledge is even more troubling. The collective experiences and insights brought about by past challenges – everything from changes in funding and policy preferences to outbreaks of disease new and old – lost. Those unwritten rules for inter-agency, inter-institution, and inter-governmental collaboration – lost. Why is this important? Efficiency and risk-management. Those collective experiences prevent teams from constantly “reinventing the wheel” by allowing them to learn from past successes and failures. That implicit knowledge is what guides collaborations – something that will be essential in a future crisis. Yes, institutions are sometimes slow to react and cumbersome to navigate, but it is the collective implicit knowledge of learned from prior successes and failures that guides the proper application of explicit knowledge to meet the current challenge.

Collectively, this amounts to an intentional degradation of science within the U.S. – one whose effects will be generational. Why generational? That’s where those other two points of concern become more pressing. The NIH award over $35 billion annually in extramural grants – these are the grants that go to the universities and research institutions in your state. This is money that pays the salaries of junior and senior scientists, the researchers (i.e. graduate students and postdoctoral fellows) who physically perform the work – that next generation of scientists –, as well as all the equipment and reagents necessary to conduct their studies. According to numerous analyses, every dollar of NIH-funded research results in approximately $2.57 of new economic activity. This is something that those senators are well aware of, and likely the reason the NIH has traditionally enjoyed bipartisan support.

It is the next generation of scientists that the current administration is actively failing. NIH budget cuts directly lead to smaller (in absolute dollars) and fewer grants that are awarded. While established scientists are just as likely to receive grant funding, new investigators and junior scientists are more likely to bear the brunt of cuts. Studies by the National Academies of Sciences, Engineering, and Medicine as well as the NIH itself concur with this assessment. Decreased funding to new investigators and junior scientists limits their ability to hire postdoctoral fellows, perform cutting edge research, and publish. For more senior scientists, the decrease in funding limits their ability to maintain the activity of their laboratories through decreased postdoctoral hiring – or the elimination of postdoctoral positions – and an overall reduction in the amount of science that is performed.

Share The Empirical Evidence Collective

If you think that the loss of some postdoctoral researchers in academia will not hurt science, think again. It is the postdocs of today that become the scientists – both academic and private sector – of tomorrow that develop the therapeutics and technologies upon which we depend. According to the National Science Foundation’s Science and Engineering Indicators 2018 report, U.S. academic institutions published over 75% of all indexed peer-reviewed articles, while industry (pharmaceutical and biotechnology companies), private nonprofits, and federal, state and local governments constituted the remaining 25%. And publication rates of industry have been falling since their peak in 2005. The NIH, simply put, funds the basic biomedical research upon which pharmaceutical and biotechnology companies base their therapeutics. This is not to say that these companies do not perform research, they do, but it is highly focused and profit driven – not the broad-based discovery research that underlies the research of the private sector.

So, what to make of all this? Well, it appears that the U.S. is actively defunding its ability to perform basic biomedical research at the same time that it is weakening its public health infrastructure, concentrating leadership in ways that risk the loss of institutional memory and scientific independence. While the Senate hearing demonstrated that some political pushback does exist, this has not yet been translated into legislative action. The active Ebola outbreak, the recent hantavirus scare, and the continuing backdrop of avian flu, measles, amongst others, would indicate how budgeting and messaging have eroded not just the public health infrastructure, but the public’s trust.

That window for meaningful intervention is small and rapidly approaching. The FY2027 appropriation process will likely culminate before the end of 2026, and even if the Senate restores some funding, the damage to research capacity – lost positions, stalled grants, eroded confidence in the future amongst scientists – takes years to reverse. The question is not whether the cuts matter. They do. The question is whether the U.S. political establishment will recognize the cost of these actions upon the scientific establishment before the consequences become irreversible.

1 Shiels, M. S. et al. Trends in Cancer Incidence and Mortality Rates in Early-Onset and Older-Onset Age Groups in the United States, 2010-2019. Cancer Discov 15, 1363-1376, doi:10.1158/2159-8290.CD-24-1678 (2025).

2 Hersher, M. & Mervis, J. U.S. government has lost more than 10,000 STEM Ph.D.s since Trump took office. Science 391, doi:10.1126/science.zw5s0b4 (2026).

This week I discussed some of the currently available methods of delivering gene therapeutics to its intended target tissue and cells. While I also discussed the risks and limitations of this therapeutic modality, a recent article in the New England Journal of Medicine, with commentary in Science, paints a very human portrait.

Hurler syndrome, also known as Mucopolysaccharidosis type I (MPS I), is a rare inherited lysosomal storage disorder. Lysosomes are essentially a cell’s garbage disposal and recycling center – playing critical roles in the breakdown of waste products, nutrient recycling, as well as cellular signaling, defense, and apoptosis. In Hurler syndrome, the lysosomal enzyme (alpha-L-iduronidase) is either not present or not functioning properly – resulting in a buildup of Glycosaminoglycans (GAGs), the long negatively charged polysaccharides that form hydrated structural scaffolds. Tissues such as the cartilage in joints, the synovial fluid of joints and the vitreous humor within the eye, tendons, as well as blood vessels and the airways of lungs all have high – but tightly regulated – concentrations of GAGs making their unique physiology possible.

Mutations in IDUA (the gene encoding alpha-L-iduronidase), the genetic cause of Hurler syndrome, occurs in ~1 in 100,000 newborns worldwide. As consequence of this mutation, cells with typically low levels of GAGs now accumulate GAGs to levels that are toxic. Organs such as the heart, spleen, liver, muscles and joints become thickened or enlarged. The central nervous system begins to accumulate cerebral spinal fluid, leading to hydrocephalus and convulsions, and in extreme cases spinal cord compression and sudden death. Patients typically develop symptoms with the first 12 – 24 months after birth and can include joint stiffness and contractures, cardiac and liver disease, as well as developmental delays and cognitive decline. The average age of mortality is 5 years.

That human portrait comes in the shape of a 5-year-old boy, born with Hurler syndrome. When he was 13-months of age he took part in a clinical trial sponsored by Regenxbio, where trillions of adeno-associated viruses (AAVs) encoding a fully functional IDUA gene were injected into his skull. The therapy seemed to have been successful (his cognitive development was on track) when a routine scan found a walnut-sized brain tumor. Regenxbio followed protocol and reported this incident to U.S. regulators who then paused the trial. Eight months after having the tumor removed, the boy “is in kindergarten reading chapter books.”

While this is indeed a successful outcome, it should not obscure the inherent risks of this type of therapy. Laboratory tests performed on the excised tumor indicate that the AAV initiated the cancer when it integrated some of its cargo (DNA) into the genome of the boy’s brain cells. AAVs, though generally considered safe (i.e. non-pathogenic with rare integrations into the host genome), are known to have genome integration rates as between 0.01 and 1%^1-3. In a laboratory research setting, a 1% integration rate in a cell culture model system would be considered “acceptable” as experiments typically last on the order of days. In in vivo mouse experiments, where experiments typically last on the order of weeks to months, integration events may be factored into experimental design (depending on the variables being measured). However, even 0.01% of a few trillion is a fairly large number, and studies have shown a causal relation between AAV integration and cancer ⁴ in tumor-prone mice ⁵.

So, how do we go about interpreting this case? Context, as always, is key.

Yes, it is true that AAVs can integrate into genome of its host cells; this has been demonstrated when administered in vivo to numerous mammalian species from mice to humans. And yes, it is true that on occasion (yet another small percentage of a small percentage) these integration events can cause a normal cell to become cancerous. Genomic integration of AAVs is semi-random, primarily occurring at chromosomal regions in an open configuration within actively dividing cells – those genomic regions being actively transcribed. Wild-type AAV (typically serotype 2, of which there are many), contains the Rep protein which facilitates site-specific integration of the AAV genome into the host’s, preferentially integrates at the AAVS1 site within the first intron of the PPP1R12C gene on chromosome 19 in humans. In this particular case, parts of the viral DNA were integrated within, and switched on, a gene called PLAG1. It just so happens that PLAG1 is active during embryogenesis and development into adulthood; a transcription factor crucial for normal brain development.

This boy’s tumor, it seems likely, was the result of four seemingly independent events. First, the specific AAV (serotype 9, lacking the Rep protein) used to deliver the IDUA gene has an affinity for cells of the nervous system ⁶ – a therapeutic necessity in this specific case as other AAV serotypes are far less effective at delivering their cargo (DNA) to this tissue. Second, the boy was young enough at the time of injection that the cells lining the ventricles of his brain – the same type of cells that formed the tumor – could have still been dividing. Third, the DNA for the IDUA gene contained a strong promoter, a DNA sequence that helps drive the gene’s expression. Parts of the AAV-genome, when integrated within the PLAG1 gene, formed a chimeric AAV-PLAG1 transcript. Genetic abnormalities altering PLAG1 (e.g. chromosomal rearrangements or fusions) are known to drive multiple forms of tumor growth ^7,8 – both cancerous and benign. And fourth, the boy was enrolled in this study after a previous bone marrow transplant, his immune system was still recovering, and perhaps, too weak to wipe out the nascent tumor cells.

I think that we must also keep in mind the severity of the disease for which this boy was initially treated. In a risk-benefit analysis, where without treatment the known outcome is a 5-year life expectancy (average) coupled with severe physical and cognitive impairments, the small – but non-zero – chance of developing a tumor seems to be an acceptable risk. This calculus is different for less severe diseases or those for which oral small molecule therapeutics are demonstrably effective. This is indeed a powerful technology with the potential to effectively “cure” many diseases; its effectiveness – and its inherent risks – all come down to specificity. Targeting the right tissue, the right cell(s) within that tissue, and the right gene within those cells – all without altering anything else.

This is no easy task. To modulate biological systems, scientists first have to understand how they function under normal conditions. Then they must understand how changes in any of the components of the system contribute to disease. Then they must understand how the proposed therapy – in this case a gene therapy – will reverse the cause and alleviate the symptoms of the disease. Then they must find a way to specifically deliver the therapeutic to the specific cells causing the disease. Countless scientists have been working on these steps – in academic research labs and biotechnology companies – for decades. We have come a long way since 1972 when the biochemical defect causing Hurler syndrome was first identified. We have come a long way since 1993 when the specific mutations in the IDUA gene that cause Hurler syndrome were first identified. We have come a long way since the concept of in vivo AAV-mediated gene delivery was first demonstrated by delivering the CFTR gene (mutated in cystic fibrosis) to the airways of rabbits in 1993 ⁹. We have come a long way since the first, of any kind, human trial of an AAV-based gene therapy for cystic fibrosis ¹⁰.

Challenges remain. Matching AAV serotypes to target tissues has greatly enhanced our ability to specifically target the tissues and cells of interest. And utilizing cell and tissue specific promoters has enhanced our ability to specifically express the therapeutic gene in the cells of interest. While these have improved specificity, nothing is ever perfect; and biology is infinitely complex. We should not eliminate the good enough in the pursuit of perfection but rather seek its improvement.

Such areas for improvement include increasing the payload capacity and modulating the immunogenicity of these viruses. The CFTR gene is 4.5kb, near the absolute maximum packaging limit and ABCA4 is nearly 6.8kb – larger than the maximum packaging limit thereby eliminating AAVs as a delivery mechanism to the retinal cells for the treatment of Stargardt disease. In addition, an estimated 30 – 60% of children have neutralizing antibodies against one or more AAV serotypes ^11-13, requiring the co-administration of immunosuppressants. In addition, AAVs typically must be administered in high doses – ensuring therapeutic levels of gene expression – which, in some individuals, can lead to progressive liver dysfunction and even sensory nerve damage ^14,15.

While imperfect, AAVs are still preferred over other delivery methods. Lentiviruses, for example, have a greater cargo capacity than AAVs, but routinely integrate into the genome. Lipid nanoparticles (LNPs), another delivery method recently optimized for mRNA based vaccines, lack innate cell targeting and accumulate in the liver. Adapting LNPs to target tissues other than the liver is an active area of scientific investigation.

Share The Empirical Evidence Collective

Subscribe now

For every monogenic disease amenable to gene therapy, there will be a corresponding risk-benefit analysis. Genetic and immune cell profiling prior to treatment will help to identify potential risks, and routine monitoring will help to identify these types of adverse events in the early stages. As imperfect as current gene therapies are, it’s an improvement over where they were just a few years ago. Something to keep in mind when we have this conversation again in a few years.

1 Wang, J. H., Gessler, D. J., Zhan, W., Gallagher, T. L. & Gao, G. Adeno-associated virus as a delivery vector for gene therapy of human diseases. Signal Transduct Target Ther 9, 78, doi:10.1038/s41392-024-01780-w (2024).

2 Nakai, H. et al. AAV serotype 2 vectors preferentially integrate into active genes in mice. Nature genetics 34, 297-302, doi:10.1038/ng1179 (2003).

3 Nakai, H. et al. Extrachromosomal recombinant adeno-associated virus vector genomes are primarily responsible for stable liver transduction in vivo. J Virol 75, 6969-6976, doi:10.1128/JVI.75.15.6969-6976.2001 (2001).

4 Rosas, L. E. et al. Patterns of scAAV vector insertion associated with oncogenic events in a mouse model for genotoxicity. Mol Ther 20, 2098-2110, doi:10.1038/mt.2012.197 (2012).

5 Zhang, J., Yu, X., Herzog, R. W., Samulski, R. J. & Xiao, W. Flies in the ointment: AAV vector preparations and tumor risk. Mol Ther 29, 2637-2639, doi:10.1016/j.ymthe.2021.08.016 (2021).

6 Jackson, K. L., Dayton, R. D. & Klein, R. L. AAV9 supports wide-scale transduction of the CNS and TDP-43 disease modeling in adult rats. Mol Ther Methods Clin Dev 2, 15036, doi:10.1038/mtm.2015.36 (2015).

7 Dalin, M. G. et al. Multi-dimensional genomic analysis of myoepithelial carcinoma identifies prevalent oncogenic gene fusions. Nat Commun 8, 1197, doi:10.1038/s41467-017-01178-z (2017).

8 Andrei, V. et al. PLAG1-Rearranged Fibromyxoid and Lipomatous Neoplasms in Children and Adults: Separate Entities or a Morphological Spectrum? Genes Chromosomes Cancer 64, e70070, doi:10.1002/gcc.70070 (2025).

9 Flotte, T. R. et al. Stable in vivo expression of the cystic fibrosis transmembrane conductance regulator with an adeno-associated virus vector. Proc Natl Acad Sci U S A 90, 10613-10617, doi:10.1073/pnas.90.22.10613 (1993).

10 Flotte, T. et al. A phase I study of an adeno-associated virus-CFTR gene vector in adult CF patients with mild lung disease. Hum Gene Ther 7, 1145-1159, doi:10.1089/hum.1996.7.9-1145 (1996).

11 Li, C. et al. Neutralizing antibodies against adeno-associated virus examined prospectively in pediatric patients with hemophilia. Gene Ther 19, 288-294, doi:10.1038/gt.2011.90 (2012).

12 Kruzik, A. et al. Prevalence of Anti-Adeno-Associated Virus Immune Responses in International Cohorts of Healthy Donors. Mol Ther Methods Clin Dev 14, 126-133, doi:10.1016/j.omtm.2019.05.014 (2019).

13 Calcedo, R. et al. Adeno-associated virus antibody profiles in newborns, children, and adolescents. Clin Vaccine Immunol 18, 1586-1588, doi:10.1128/CVI.05107-11 (2011).

14 Buss, N. et al. Characterization of AAV-mediated dorsal root ganglionopathy. Mol Ther Methods Clin Dev 24, 342-354, doi:10.1016/j.omtm.2022.01.013 (2022).

15 Stone, D., Aubert, M. & Jerome, K. R. Adeno-associated virus vectors and neurotoxicity-lessons from preclinical and human studies. Gene Ther 32, 60-73, doi:10.1038/s41434-023-00405-1 (2025).

The USPSTF was founded in 1984 as part of a 5-year program to “develop recommendation for primary care clinicians on the appropriate content of periodic health examinations.” It was reconstituted by the Department of Health and Human Services in 1990 with the purpose to “continue and update these scientific assessments of preventive services,” and has been making preventive health recommendations ever since. The Affordable Care Act (ACA) mandated that most private health insurance plans and Medicaid programs cover clinical preventive services that receive an “A” or “B” grade from the USPSTF at no out-of-pocket cost.

Why would anyone not want to prevent disease? My cynical guess is that it eats into insurance companies’ already bloated profit margins in the short term, and that this is just another attempt by RFK Jr. to yet again undermine science-based medicine. What did Drs. Wong and Davis do to warrant their removal from the task force? It’s not at all clear, but Kennedy did tell lawmakers last month that the task force has been “lackadaisical and negligent for 20 years” and that he was planning to bring in new members “who have a clear mission.”

At the moment, it is unclear who will be appointed to fill the now-vacant spots. If history is any guide, RFK JR. will likely pick those who shares in his distrust of vaccines and other science-based medicines. Adding members to this task force that either willfully ignore the evidence or chose not to assess the evidence on it merits can cause great harm to the American public. Downgrading colorectal cancer screening (for example) in adults aged 50 to 75 (currently and A grade) and adults aged 45 to 49 (currently a B grade) to a B and C grade respectively, could have dire effects on preventive colorectal screening in the 45 to 49 years age group. Studies show that ~35% of Americans in this age group have received this screening - approximately 7.4 million individuals. Eliminating this screening puts those that would ordinarily have it at risk of health disparities - identifying colorectal cancer in the later stages rather than its earliest when treatment outcomes are measurably better.

I guess this is what happens when a science denier becomes the Secretary of Health and Human Services. One could be forgiven for not knowing about the USPSTF before now. It largely operates behind the scenes, but its recommendations impact all of us. Health and Human Services - just another way he is undermining the science that may improve our Health, an essential need of humans (Human Services).

Share