Read on LinkedIn

The promise of bioengineering is to design new biological tools: diagnostics that help us detect diseases, antibodies that can destroy cancers, new vaccines to protect against viruses. To build those tools, AI will ultimately need to interact with the physical world and not just with a static dataset.

The next phase of AI in biology is an agentic science loop: an AI system that can reason over scientific knowledge, use specialised tools to design thousands of molecules, send those designs into an automated wet-lab, and learn from the resulting data.

Three layers need to work together:

• Generalist models to interpret a scientific goal, plan a campaign, and orchestrate workflows.

• Specialist models and tools to perform tasks such as protein structure prediction and protein design models.

• Automated wet labs to turn digital designs into measurements from the physical world.

In this post, we showcase how Anthropic benchmarked Claude Mythos Preview and Opus 4.8 at designing novel protein binders that we tested in our automated wet lab. Overall, Claude models showed expert level skill at protein design, matching or exceeding human experts on many tasks. Together with the Anthropic team, we’re releasing the actual protein sequences Claude designed as well as the experimental data on Proteinbase, the open protein data platform.

How to test AI-designed proteins in the real world

At Adaptyv, we have built an automated, AI-native wet lab for turning protein designs into experimental data.

Instead of scientists in lab coats pipetting samples in tiny tubes one by one, we have built automated workcells that run the same experiments many times faster and cheaper.

Human protein designers and agents can access the lab via our web platform and API to send digital protein sequences for wet lab validation on different assays.

Our platform automatically processes the protein sequences (a string of amino acids) and converts it into a special DNA sequence that encodes the biological instructions for how to create the specific protein. This digital DNA sequence is then turned into an actual physical DNA molecule in the lab by assembling the DNA building blocks one by one. Next, cell-free protein synthesis takes the machinery a cell uses to read DNA and build proteins (ribosomes, enzymes, amino acids, an energy supply) and runs it with no cell around it. This is done using automated robots able to pipette incredibly small amounts of liquid really fast and at high-throughput.

At this point we have now made physical proteins in the lab from the digital AI-designed sequence. But we haven’t yet tested if the protein actually performs well at what it was designed to do. Proteins are the molecular machinery of all of life and can do many different things: digesting the food we eat, breaking down harmful molecules and making ones that we need to survive, generating energy in our cells, cutting and editing DNA, and a million other things.

Here, we’re testing binders: proteins that should stick to one target and nothing else. Most antibody cancer drugs are binders. So are the reagents in a pregnancy test and the capture molecules in most diagnostics. The “sticking strength” of the protein binder to its target can be measured with special instruments, giving us a so called K_D value. A lower K_D means a better binder, an often desirable property when developing new therapeutics. At Adaptyv, we run these specialized instruments with our own software to process all raw data to obtain clean K_D values against a wide range of target proteins that are relevant for therapeutic, diagnostic, and research applications.

Behind this simple input-output interface is a complex experimental process: many reagents, instruments, protocols and measurements must be coordinated and verified. At Adaptyv we package that complexity into an automated, quality-controlled workflow that answers a biological question reproducibly and generates a defined type of data at scale.

So how good is Claude at designing proteins?

The team at Anthropic chose 16 existing protein targets from our previous public competitions, hackathons, and BenchBB to benchmark different Claude versions for de novo binder design, all using Claude Science. They were all prompted similarly and had access to the same publicly available tools.

Then, designs against all these targets were sent to our wet lab, in an anonymized way. We did not have any information on which Claude model designed which protein. We ran our Affinity Characterization assay, to measure the binding strength on SPR using 5 target concentrations and with duplicate measurement, ensuring the results are robust, accurate, and matching our quality control standards.

95% of the designs expressed, which three years ago would have been an impressive headline, showing the rapid progress of AI tools for protein design in recent years. This number matched the best expression rates of our EGFR competition which had hundreds of expert protein designers, and surpassed other challenges such as the RBX1 one. Out of these, 354 of all designs (1,320) bound their target, an overall hit rate of 26.8%, and the per-target hit-rates vary quite widely. One target (MBP) yielded no binders. Overall, Claude designed proteins that bound 14 out of 15 targets, with 1 target (GDF-8 mature) being excluded due to low-quality measurements in our setup, likely due to the target aggregating, meaning it was binding to other molecules of itself.

When compared to our competitions, Claude surpasses all their hit rates, especially when looking at every single run for each target Anthropic submitted as in the plot above. For a fair comparison, we have subsetted each competition’s results to only include de novo minibinders. Claude achieved an 80% hit rate on TREM2, greatly improving over the 38.3% we reported in our competition, and even on trickier targets such as 15-PGDH, it has a success rate more than 3-fold higher than observed on Proteinbase.

The best Claude binders also show greater binding affinities compared to 5 out of 6 competition winners. It greatly surpassed the best 15-PGDH binder, going from 1.7 uM to 33.4 nM, and similarly for RBX1 (from 25.7 nM in the competition to 3.9 nM). Surprisingly, it was unable to beat the best binder on the Nipah Virus Competition. For GDF-8, RBX1, and Nipah, Claude seems to have explored a wide range of possible epitopes, while it converges on single well-defined epitopes in case of TREM2. Overall, Claude would have been a prolific protein designer if it were to participate in our Protein Design Competitions.

Looking at all these impressive results, Claude seems to have matched or even surpassed expert protein designers when it comes to orchestrating openly available design tools. We are definitely excited to see where this is heading towards and what the future of agentic science and protein design might look like. We’re sharing some of our thoughts down below.

Agentic science and curing actual diseases

So can AI now solve all diseases? Well no, not yet.

This case study shows that Claude is at least expert-level at orchestrating protein design tools. That’s great news, since protein design tools are hard to use. Before AI, even setting up a protein structure prediction tool could take hours of debugging opaque conda errors.

Now, imagine a researcher working on a new cancer therapeutic who needs an assay that can distinguish between two mutated receptor variants. They can give Claude the sequencing results from their cell samples and ask it to help design binders. Claude would sift through papers and research databases, find where in the genome the receptors are encoded, compare the two sequences, and map the differences. It would use protein folding models like Boltz to generate 3D structures of the receptor variants, then use protein design models like BindCraft to generate binders that attach to one variant but not the other. It would generate thousands of computational designs and score them to identify the most promising candidates to test. Then, thanks to our cloud lab API, Claude could submit the best candidates to the lab and have them tested experimentally in a couple of weeks. Once the results are in, Claude could analyze the data and find that the strongest binders are actually too promiscuous and bind both variants. Informed by the first round, it could then launch another design campaign to reduce that cross-reactivity and produce better proteins in the next round. Those can be linked to a fluorescent marker and then bind and mark the cancer cells to differentiate them from healthy cells.

Just a few years ago, this process alone would’ve required a year’s worth of a whole lab’s work. Today, anyone who is scientifically curious can run this campaign end-to-end from their laptop in a few weeks and for a couple thousand $ in AI tokens, cloud compute and wet lab credits.

The road to curing all disease needs to be built first

Of course, those proteins that we tested here are not real therapeutics. They completed only the first step of the process: demonstrating that they can function as binders. Still, this study shows a path towards making actual therapeutics with AI.

Imagine making a drug is like climbing a mountain. We have clearly been able to climb some mountains, as humanity has made many drugs already. But the way to the top is a dangerous narrow path and climbing it takes many years and costs billions of dollars (and the lives of many biotechs).

The goal of AI for drug discovery is turning this narrow mountain path into a highway, making it easier and cheaper to get to the top so that we can develop 100x more therapeutics than we have right now. Similarly, writing code was a more of a high-expertise craft before LLMs, now it’s mostly automated and it has made generating software accessible for anyone.

In practice, for drug discovery, that means:

• Automating molecular design and bioinformatics work to make better drug candidates

• Building good experimental readouts that answer relevant biological questions

• Automating those experimental workflows to increase throughput and lower costs at scale

On the wet-lab side, the road then looks roughly like this:

1. Expression: Can we reliably synthesise newly designed proteins?

2. Binder design: Can we make proteins that bind their intended target?

3. Therapeutic formats and developability: Can a candidate be made in the formats industry uses, remain stable and be manufactured reliably?

4. Cell-based function: Does binding produce the intended effect in a living-cell system?

5. Organoids and more realistic models: Does it work in a model that better captures human biology?

6. Translational and in-vivo evidence: Is it safe and effective in the settings that ultimately matter for patients?

At every step, the mountain gets steeper and more challenging, requiring better models (both intelligence/generalists and domain-specific tools), more compute and more wet lab data.

Closing the physical feedback loop

This Anthropic campaign was an open-loop experiment. Claude designed proteins, Adaptyv tested them experimentally, and here we have presented the results. The next step is to close that loop. An agent proposes a batch, receives experimental data, learns from the failures and successes, and chooses the next batch on the strength of what it just learned.

The long-term vision is to give AI a real path to help cure all disease. It will not happen in one leap. It happens by extending the physical feedback loop from binding to each harder biological question that follows, until AI can learn from the evidence needed to develop medicines.

Resources

• Read Anthropic’s blog post here: https://www.anthropic.com/research/Claude-accelerates-protein-design

• Check out the protein designs and experimental data here on Proteinbase: [COMING SOON]

• Benchmark your own protein designs: adaptyvbio.com

• Connect your agent to our API and MCP server

• Join Adaptyv: adaptyvbio.com/careers

• We hosted our third installment of the Protein Design Competition, this time challenging participants to design a binder against the Nipah virus glycoprotein.

• The response was incredible: Over 10’000 designs were submitted, over 5 times the number of designs in the past 2 competitions combined, with more than 650 total participants.

• We’re currently testing all 1’200 selected designs, but, in the meantime, we’re taking a brief detour into what happened during the competition. Which model was the fan favorite? What did the top-ranking designs look like? And how can we select a better ranking metric for future competitions?

• Plus, we opened a prediction market on Manifold for this competition so you can bet on hit rates, best performing models, and more

After a busy few weeks with over 680 protein designers submitting their best binder designs against the surface glycoprotein of the Nipah virus, the submissions and community vote doors have closed and we are waiting in anticipation for the experimental results to arrive. In the meantime, we want to highlight some interesting new insights and cool protein structures that we have gathered during the course of this competition.

Let’s start with a quick recap: Nipah is one of the deadliest viruses in the world and considered one of the top future pandemic risks. It’s a virus that’s found in fruit bats and was first discovered near the Nipah river in Malaysia during an outbreak in 1999. The virus is asymptomatic in bats but can infect livestock like pigs if they eat fruit contaminated by bat feces. From there it can infect humans if they are in contact with sick animals. In humans, it causes severe respiratory and neurological disease (Ganguly et al.), with mortality rates of up to 70%. In comparison, SARS-CoV-2 had mortality rates of ˜1-3%. Because of this high mortality rate, Nipah is considered a top-priority virus for vaccine development. No approved treatments or vaccines currently exist (Chan et al.). So the goal of this competition is as simple as it is difficult: Designing the best protein binders capable of neutralizing the Nipah virus.

Specifically, we asked participants to create binders that can disrupt the interaction between the viral glycoprotein G and its human receptor, ephrin-B2/B3 - an essential step the virus uses to enter host cells and initiate infection. Blocking this interaction has shown to reduce viral infection.

Out of the thousands of designs submitted we wanted to select the 1200 most promising designs for experimental validation:

• 600 designs were chosen based on the best Boltz-2 ipSAE score

• 400 designs were seleced by a panel of protein design experts

• 200 designs were selected by the community vote

So in total, more than 1 000 binders are being tested in our lab right now for binding to Niv-G. Stay tuned for the results of the competition, which we will release on January 16th on Proteinbase!

Four design models dominate the competition

Mirroring the trend we observed in our previous binder design competition, participants relied heavily on RFDiffusion and BindCraft to design their binders. However, the design toolkit is evolving and in this latest competition, the newly launched BoltzGen model (Stark et al.) emerged as the most popular design tool. We’ll explain the top design choices in more detail below.

BoltzGen is an all-atom generative diffusion model that unifies structure design and prediction into a single framework. By embedding structural reasoning directly into the generative process, the model achieves state-of-the-art accuracy in both design and folding. The model was validated across 26 targets in eight experimental campaigns, reaching a 66% success rate for designing binders with affinities in the nanomolar range. BoltzGen unifies the traditionally separate stages of design, inverse folding and filtering into one streamlined pipeline. This ease of use might also explain why the tool was very popular among designers.

The 2nd most popular design tool was design-a-protein.com, a platform we developed specifically for this competition to run a protein design workflow in less than 2 min! The goal here was not to run the most powerful model but allow non-experts and beginners to explore protein design in a playful way with a very fast design model. Users can generate Nipah virus binders by selecting structural hotspots on the viral protein, which serve as input to the design pipeline. Protpardelle-1c (Lu et al.) generates 3D binder backbones around these hotspots, and ProteinMPNN (Dauparas et al.) is used to designs sequences that fold into these structures. Additionally, users can control parameters such as chain length, number of designs, and temperature, and inspect predicted ipSAE scores in a dashboard to easily select high-scoring candidates.

The animation below shows the cumulative growth of submissions throughout the competition. Many participants submitted multiple entries over the course of several weeks, and momentum clearly built over time, with a sharp increase in submissions as the deadline approached.

Optimizing ipSAE score from Boltz-2

As in the previous competition, we selected designs for validation using an in silico metric, Boltz-2 ipSAE (more details on that metric later). This naturally fostered competition for the top spots on the leaderboard. The leading ipSAE score changed repeatedly over the course of the competition, with participants actively pushing the limits of this metric. This led to a steady upward trend in the highest-scoring ipSAE designs submitted over time. Competition intensity was high, with some participants even opting to experimentally test their top in silico candidates experimentally before submission, aiming to ensure that only the most promising designs were entered (here).

Let’s take a closer look at the top 100 binders

When comparing the molecule types between all submissions with the top 100 designs ranked by ipSAE score, miniproteins are approximately twice as prevalent in the top-ranked set, accounting for 55% of the top 100 submissions. Miniproteins are defined as designs with a molecular weight below 10.5 kDa and less than 35% loop content. Miniprotein-like designs meet the same loop-content criterion but have molecular weights between 10.5 and 15 kDa.

This enrichment is consistent with the tendency of Boltz-2, when used without a template multiple sequence alignment (MSA), to favor de novo miniprotein designs rather than antibody derivatives such as nanobodies and scFvs (Yin et al.).

Interestingly, the top 100 designs are enriched for proteins dominated by α-helical structures compared to the full set of submissions. Correspondingly, designs containing mixed α-helical and β-sheet architectures, as well as predominantly β-sheet structures, are underrepresented among the top-ranked entries. This shift may reflect the greater fold stability and lower structural ambiguity of α-helical proteins, which can lead to more confident structure predictions and higher ipSAE scores.

Choosing a good computational filtering metric

Previous competitions have shown that choosing an appropriate computational score to prioritize the most promising designs for experimental validation is in itself a challenge. Previously, we had used either the interface PAE (iPAE) metric from AlphaFold2 or a combination of iPAE, AlphaFold2’s interface pTM (iPTM) score and ESM2’s log-likelihood score (ESM-PLL). Both turned out to have their shortcomings. In the last competition, our use of unnormalized ESM log-likelihoods created a bias toward shorter designs.

For this competition, we opted to use the Boltz-2 ipSAE score for filtering and selected the 600 most promising designs based on this metric. You can find our reasoning for choosing it in the FAQ. Already during the course of the competition, some interesting questions and concerns were raised regarding the reliance of this metric.

Many participants evaluated their designs locally prior to submission and encountered reproducibility issues of the ipSAE score. These differences arise in part because the model can yield slightly different results depending on hardware configuration and random seeds. A certain degree of unpredictibility was intentional, as it makes direct optimization for the metric more difficult. However, as it turned out over the course of the competition, the ipSAE score variance was not uniform across designs and therefore problematic. Importantly, variability in ipSAE reproducibility may also have conferred an unintended advantage to designs with more stable scores, as these could be more easily optimized directly for the metric.

Given the difficulty of choosing a universally robust scoring function, we wanted to complement the computational filtering with community input and expert knowledge when deciding which designs would ultimately be selected for experimental testing. We’d like to sincerely thank the community for enthusiastically participating in the community vote, championing their favorite designs, and generally making this process far more interesting than a simple ranking table. We also owe a big thank-you to a group of protein design experts who volunteered to dive into the submissions and hand-pick proteins they found particularly promising.

This combination of computational scoring, community engagement, and expert review was intended to mitigate the limitations of any single selection strategy and to arrive at a more balanced and informed set of designs for experimental validation. Additionally, to mitigate the impact of ipSAE reproducibility issues, we added 200 additional expert-selected sequences to the pool of designs designated for experimental validation.

Prediction markets for protein design

While we’re waiting for the experimental results to come out of the lab, we thought it would be fun to create some prediction markets for people to bet on the hit rate for this competition, which model(s) will perform best and which designers will have binders.

Check out the Proteinbase questions on Manifold

Manifold uses a play money currency called Mana that you can bet with. You can just sign up for free and you’ll get 1000 mana to play with.

Nipah Gallery

Welcome to our mini “Nipah Gallery” — a highlight reel of protein structures we found particularly interesting from this competition. Each image links directly to its Proteinbase entry. Think of it like “Spotify Wrapped” for our Nipah design competition. The featured designs were selected based on how unique and compelling the structures were, the creativity of the design approaches used, and our collective intuition as the Adaptyv team about which candidates might express well and bind effectively.

Highlighting some cool approaches and community efforts

Some of the most interesting competition submissions are those that offer insight into the design process, including the decisions, motivations, and reasoning behind them. We’d like to thank everyone who took the time to talk about the strategies. We believe this contributes to the collective understanding of protein design tools quite a lot! In the following, we highlight several of these submissions, along with blog posts and social media discussions that reflect participants’ experiences during the competition.

• Andrés Torrubia learned firsthand that aggressively optimizing leaderboard metrics does not necessarily translate into experimental success: although his designs ranked #1 in silico in the EGFR competition, they ultimately failed in the wet lab. In this competition, he adopted a markedly different strategy, imposing constraints on compute and intentionally avoiding direct optimization of the ipSAE score in order to mimic a more realistic therapeutic design workflow. Check out his submission here.

• Several members from the Wells Wood lab took part in the competition and employed different design strategies and models. But they all assessed their designs using molecular dynamics simulations to evaluate the stability of the receptor–binder complexes. Final candidates were then ranked using a custom DE-STRESS–based score to identify the most promising designs. As a result, two team members finished in the top 100, including Leonardo Castorina who secured the top position on the leaderboard. He generated a large set of designs by combining BindCraft with TIMED sequence generation, large-scale sampling, and Boltz-2–based evaluation, followed by surrogate modelling and one round of in silico directed evolution.

• Yehlin Cho did not only rank 14th on the learderboard with her designs but also created an interactive site to demonstrate how the Boltz-2 ipSAE and ipTM scores compare to the corresponding metrics from AlphaFold3. This comparison further underscores how strongly such metrics can vary depending on the prediction tool used and the specific protein sequence being evaluated. Check out here super intuitive webpage that also allows you to look at your own data by simply uploading them as a CSV file.

• Ariax Bio submitted both nanobody and miniprotein designs to the competition and also created an excellent tutorial for getting started with protein design using BoltzGen. The tutorial walks through all the key steps required to run a design campaign and highlights the most important aspects to pay attention to along the way.

• Jacob DeRoo documented in detail his design decisions and released his code here. He tried different design modes using BoltzGen, from free design, to motif grafting and hotpot/binding site specification. He also experimented with RFDiffusion, ProteinHunter and BindCraft.

• In their submission, Silico Biosciences used their own peleke-1 suite of antibody language models to generate heavy and light chain Fv sequences conditioned on an epitope-annotated antigen. In their blogpost they highlight their issues with reproducibility of ipSAE scores and show that ipSAE scores fail to correlate with physics-based estimates of binding affinity. Highlighting again the need for more robust and reproducible scoring methods and good benchmark datasets for the field.

We’d like to thank all participants the Nipah competition, the people who voted for their favorite designs, our panel of experts, and everyone else involved who made this such a cool experience. We’re super excited to release all the experimental results and reveal the final rankings soon!

How to vote:

➡️ Head to proteinbase.com
🧐 Browse the submissions
⭐ Star your favorites

You can vote for as many as you’d like.

The top submissions (up to 200 proteins in total) based on community votes will be experimentally tested in the lab.

Voting closes December 10th 11:59 PM PST.

We’re hitting new records in terms of interest in protein design:

• Over 600 people participated
• Over 10’000 proteins were submitted
• Over 188 different protein design methods were used

When uploading their proteins, participants had the option to declare the design method used for each submission. Over half of all entries (5,648 proteins) included a self-declared method.

The most popular methods were:
• BoltzGen family: 1,247 proteins
• RFdiffusion family: 807 proteins
• Design-a-protein: 646 proteins
• BindCraft: 457 proteins

Overall, we saw impressive diversity and many new methods. We’re excited to see how they perform in the lab!

The Best ipSAE Award goes to Nick Boyd for his protein ‘noble-swan-clay’ with a whopping score of 0.92!

Check it out here

We received many questions about how ipSAE is calculated - we’ll share the exact methods and the reasoning behind in an email to participants soon

Want to dive deeper? Explore individual proteins, full design methods, and other stats in our Proteinbase collection:

https://proteinbase.com/collections/nipah-binder-competition-all-submissions

Please share this announcement on LinkedIn and X :)

➡️ http://design-a-protein.com

This is part of our protein design competition on Proteinbase, where we test completely novel protein sequences in a real lab.

The deadline to submit proteins is this Sunday EOD.

We’ve already received thousands of proteins from hundreds designers from all over the world.

And now you can join too!

The Nipah virus protein design competition on Proteinbase is heating up! We’re now averaging over 500 proteins submitted per day.

Want to take the top spot on the leaderboard? Modal now sponsors $500 in credits for you to run protein design models to participate in the competition!

Modal is the easiest way to run AI models and we use them extensively at Adaptyv for all our compute needs. Thanks a lot to the team there for supporting AI for biology research!

Find out how to redeem the credits here:

https://proteinbase.com/competitions/adaptyv-nipah-competition

As a reminder:

For wetlab testing, we’re selecting
- 600 sequences based on the ipSAE ranking
- 200 sequences based on an expert ranking
- 200 sequences based on a community vote

For expert ranking and community vote more details will follow soon!

Based on popular demand, we’re also extending the submission deadline by 1 week -> now open until Sunday November 30.

Check out the leaderboard and submit your sequences here:

https://proteinbase.com/competitions/adaptyv-nipah-competition

We tested 24 of their designs for a single property - binding affinity - with our affinity characterization platform. 23 out of 24 expressed and 6 bound to their target, showing that MOG-DFM is quite effective at designing functional peptide binders from just the sequence information!

Want to see the peptides they designed? All sequences + lab validation data are now available on Proteinbase!

Introducing designed peptides

You’ve probably heard about GLP-1 peptides: they’re the technology behind drugs like Ozempic and Wegovy that have transformed diabetes and obesity treatment. But what you might not know is that it took decades of work to make these peptides be clinically viable, meaning that they not only bind their targets effectively, but also survive in the body, avoid toxicity, dissolve well, and actually work as safe medicines. Most peptide candidates fail somewhere along this pipeline, not because they can’t bind, but because they don’t have a combination of good therapeutic properties. Now imagine if we could design GLP-1-like peptides for any disease and have them be therapeutically viable from the very beginning.

Recent advances in machine learning are making this vision possible. Instead of tweaking one molecule at a time, researchers now use generative models (basically algorithms that can explore vast sequence spaces) to imagine entirely new peptide candidates. The Programmable Biology Group has pushed the boundaries with models for motif-specific targeting (moPPIt), fusion-breakpoint detection (SOAPIA, FusOn-pLM), post-translational modification prediction (PTM-Mamba), the design of non-canonical or cyclized peptides (PepTune), and target-binding peptides for rare-disease treatment (with the recent Gumbel-Softmax flow matching framework).

However, the development of therapeutic peptides involves more than just designing a peptide that binds effectively to its target. Researchers must consider a variety of physical and biological properties, such as solubility, stability, affinity, hemolysis, and non-fouling behavior. The challenge lies in balancing these conflicting properties, as optimizing one may negatively affect another.

In this post, we’ll break down the group’s newest model, Multi-Objective-Guided Discrete Flow Matching (MOG-DFM), which addresses these challenges by leveraging a novel multi-objective optimization algorithm for discrete flow matching to guide the generation of peptide sequences that optimize multiple properties at the same time. We have validated their designs for two properties we can readily test: binding and expression, with some impressive results!

Why is peptide design such a challenging problem?

Designing peptides that satisfy multiple, often conflicting, functional and biophysical criteria is no simple task. The difficulty arises on two intertwined fronts. First, unlike single-objective problems where there is a clear optimum, multi-objective optimization (MOO) yields a Pareto front of trade-off solutions where improving one property typically degrades another. The second difficulty is biological: the properties themselves are not independent and the mapping from sequence to phenotype is nonlinear, due to epistatic interactions and context-dependent effects.

MOG-DFM for a more efficient multi-objective optimization

MOG-DFM offers a novel solution to this problem by integrating both generative sequence modeling with multi-objective optimization. At is core, MOG-DFM leverages discrete flow matching (DFM), a generative modeling approach tailored for biological sequence data like peptides. DFM learns how a sequence should evolve, step by step, from a random initialization toward a realistic, functional target. This involves the concept of a “velocity field”: for each position in the sequence, the model predicts the probability (or “velocity”) of switching the current token (such as an amino acid) to any other possible choice. Unlike methods that operate in continuous space, DFM works natively with discrete symbols, making it especially suitable for applications in biology.

To design a new peptide, MOG-DFM begins with a random sequence and sets a specific “trade-off direction” that represents the desired balance among properties (such as affinity vs. solubility) from a Das–Dennis simplex lattice . This direction is chosen so that different runs of the algorithm can explore different parts of the possible trade-off landscape.

After initialization, MOG-DFM will perform multiple sampling steps to gradually evolve the starting sequences to ones with desired properties. At each step, MOG-DFM selects one random position in the sequence to update. For that position, the algorithm evaluates all possible candidate tokens by calculating two scores: a rank score, reflecting how much each option improves the desired properties compared to the alternatives, and a directional score, which measures how well the change moves the sequence toward the chosen trade-off direction. These are combined into a guidance score, which adjusts the underlying DFM “velocity” for each possible transition. In practice, transitions with higher guidance scores are exponentially favoured, actively steering sequence evolution toward peptides that optimally balance all objectives.

To keep the generative process efficient, MOG-DFM uses a technique called adaptive hypercone filtering. Imagine each possible sequence change as an arrow pointing in a direction: some arrows point toward the desired trade-off (the optimal balance of properties), while others don’t. Hypercone filtering works by only allowing changes whose arrows fall within a certain angle (“cone”) of the target direction. If the algorithm finds itself with too few options, the cone automatically widens, encouraging more exploration! If it’s admitting too many, the cone narrows to maintain focus. This dynamic adjustment helps the algorithm avoid both getting stuck and wandering aimlessly. The sequence is then updated using the Euler method: the chosen position is switched to the selected token with a probability determined by its transition velocity to the selected token, otherwise the original token is retained. This approach maintains stochasticity in the sequence evolution, while ensuring that updates remain consistent with the desired trade-off direction.

Overall, discrete flow matching supplies the underlying generative backbone, while MOG-DFM injects multi-objective guidance through the rank-directional scores and maintains exploitation and exploration balance via the adaptive hypercone filtering mechanism.

In silico MOG-DFM benchmarks

MOG-DFM was benchmarked on a peptide binder design task guided simultaneously by five therapeutic properties:

• Hemolysis: A measure of toxicity, specifically the ability of a peptide to damage red blood cells. Lower values indicate safer, less toxic peptides.

• Non-fouling: Reflects the peptide’s tendency to avoid sticking to unintended surfaces, reducing unwanted interactions and side effects.

• Solubility: Determines how readily the peptide dissolves in biological fluids, a key factor for delivery and bioavailability.

• Half-life: Indicates the stability of the peptide in the body. Longer half-life means the peptide persists longer, allowing for lower dosing and improved efficacy.

• Binding affinity: Measures how tightly the peptide binds to its intended target, such as a disease-related protein or receptor.

Benchmarking was performed using a set of protein targets that included structured proteins with pre-existing binders, structured proteins without known binders, and intrinsically disordered proteins. Significantly, MOG-DFM-designed peptides consistently achieve low hemolysis (0.06–0.09), high non-fouling (>0.78) and solubility (>0.74), extended half-life (28–47 h), and good affinity scores (6.4–7.6).

MOG-DFM was also compared to four classical multi-objective optimization baselines: NSGA-III, SMS-EMOA, SPEA2, and MOPSO. Although MOG-DFM incurs longer runtimes, it consistently yields superior trade-offs: it reduces predicted hemolysis by over 10%, increases non-fouling and solubility by roughly 30–50%, and extends half-life by a factor of three to four relative to the next-best competitor, while maintaining comparable affinity. These results highlight MOG-DFM’s ability to navigate high-dimensional, conflicting property landscapes and produce peptide binders with well-balanced profiles that would be difficult to obtain via traditional optimizers.

Benchmarking MOG-DFM and 4 other multi-objective optimization baseline on all 5 key properties for therapeutic peptides. Table provided by the Programmable Biology Group.

Validating 24 MOG-DFM designs in our Adaptyv Foundry

We received 24 ten-residue de novo peptide binders targeting human FcRn, We then experimentally characterized all 24 designs by Biolayer Interferometry (BLI) using our Affinity Characterization workflow, with two replicates per peptide. We can see 6/24 sequences with clear KDs in the hundreds of nanomolar range (defined as binders by our assay, but still with some room for optimization).

What’s next for MOG-DFM

As noted before, the framework can become computationally intensive as sequence length or output dimensionality grows. Extending to longer proteins or other high-dimensional biological sequences will increase the number of candidate transitions per step and and the number of sampling iterations needed. Second, while MOG-DFM steers generation toward Pareto-efficient regions, it does not come with theoretical guarantees of Pareto optimality or coverage. The adaptive guidance and hypercone filtering induce positive expected improvement in the desired directions, but there is no formal assurance that the sampled set will fully represent or converge to the true Pareto front.

Thus, these limitation motivate two directions to improve upon:

1. Scale MOG-DFM to longer sequences, including those with non-canonical amino acids,

2. Strengthen Pareto convergence guarantees and better characterizing coverage, potentially via uncertainty-aware or feedback-driven extensions to the guidance mechanism.

Resources and links

• Try out MOG-DFM here and read the preprint here.

• The MOG-DFM experimental results are hosted on Proteinbase.

• Check out what the Programmable Biology Group is working on!

• We thank both Tong Chen and Prof. Pranam Chatterjee for actively working on this blog post!

• Have some novel proteins you want to test in the lab? Come talk to us — we’d like to run many more of those protein designer spotlights, so if you have a cool new hypothesis or model to test we’d love to hear from you!

Go to http://proteinbase.com to get started and submit your designs

Submission Rules:

- Up to 10 proteins per submission.

- Each sequence must be max 250 AA.

- You can submit once every 24 hours

- New submission overwrites old submission

- Designs need to be at minimum 10 AA away from any published sequence -> We have a similarity check for that

Submit your proteins

All you need is a laptop to participate: submit your computational protein designs, and Adaptyv will synthesize and experimentally test 1,000 of the most promising proteins, with all results released open-source on Proteinbase.

🦠 Why Nipah Virus?

Nipah virus is one of the deadliest known pathogens, with fatality rates up to 75% and no approved treatment or vaccine. It’s a WHO-listed pandemic threat due to its ability to spread from animals and between humans.

Our target is the Nipah virus Glycoprotein G (PDB: 2VSM), the viral surface protein that binds to human ephrin-B2 and ephrin-B3 receptors to enter host cells. Designing a protein that blocks this interaction could prevent the virus from attaching and stop infection before it begins.

⚖️ Selection

1,000 computational protein designs will be synthesized and tested in our lab, chosen through three paths:

• 600 by ipSAE score

• 200 by an expert panel

• 200 by community vote

🏆 Winners

All 1,000 selected proteins will be experimentally tested for expression and binding affinity against Nipah virus Glycoprotein G.

Winners will be recognized in two categories: De Novo Design and Lead Optimization, rewarding both novel binders and improved versions of existing leads.

🗓️ Timelines

Submissions open on Monday, October 27 and close on November 24.

Experimental validation begins on December 1, with results expected by January 6.

👀 Want to get started designing proteins?

You can find a list of popular AI models and design methods here: https://proteinbase.com/design-methods

We also recommend these 3 blog posts:

• A primer on AI in antibody engineering by Owlposting: owlposting.com/p/a-primer-on-ai-in-antibody-engineering

• AI antibody design 2025 by Brian Naughton: blog.booleanbiotech.com/ai-antibody-design-2025

• Takeaways from Adaptyv’s latest protein design competition by Tudor-Stefan Cotet: https://www.adaptyvbio.com/blog/po104/

All details about the competition are available on the Proteinbase competition page. You’ll also be able to submit your designs there once submissions open:

➡️ proteinbase.com/competitions/adaptyv-nipah-competition

Check out the X and LinkedIn post

Today we’re launching Proteinbase, a single hub for experimental protein design data. Over 1,000 novel proteins are already live, each with computational predictions, experimental validation, and the method used to design them. Everything comes from the Adaptyv lab under standardized protocols, which means the results are reproducible, comparable, and include negative data that usually never gets shared.

To encourage open science, we’re also offering a 20% discount on Adaptyv lab validation services if you open source your results on Proteinbase!

What are we trying to fix?

📊 Lack of open, high-quality protein experimental data (including negative data)

How we’re fixing it: We’ll periodically release thousands of experimental data points generated in the Adaptyv lab, from customers who choose to open source their result and from our internal benchmarking campaigns.

⚖️ Lack of real-world benchmarks for protein design pipelines

How we’re fixing it: Proteinbase links every protein to the design method that created it. As data accumulates, you can see how each model performs across different tasks. We’re also introducing standardized benchmarks like BenchBB to establish clear performance metrics.

🧑‍🔬 Lack of standardisation in experimental protocols which makes the data hard to compare

How we’re fixing it: All data on Proteinbase comes from the Adaptyv Lab, using standardized protocols. Every result can be traced back to its exact experimental conditions.

🧠 Lack of experimental validation opportunities, which makes it hard to see novel ideas emerge

How we’re fixing it: We will organize regular protein design competitions that are free to enter, with testing fully funded by Adaptyv or partner organizations.

How does Proteinbase work?

Proteins

The core elements of Proteinbase are proteins. For each protein published on proteinbase, we’ll show both computational predictions and experimental measurements. All submissions on Proteinbase go through our structure folding and annotation pipeline: they get assigned a folded structure via the recent Boltz-2 model, several metrics to characterize their designs, and structural domain annotations. Once the protein has been validated in the lab, all experimental information will be available on the protein’s page, from BLI curves to expression measurements or thermostability.

Collections

Collections group related proteins together, like a curated showroom or a playlist. These can be organized around a hypothesis, model launch, benchmark, or any theme that makes sense for your work. Collections can contain binders against a target, optimized enzyme variants, de novo proteins from a model like RFdiffusion, or any other designed protein.

Design Methods

Design Methods are the tools and approaches used to create proteins. These range from multi-step pipelines like BindCraft to single models like EvoDiff. Each protein links back to its design method whenever possible. The most successful methods (or state-of-the-art) can be easily retrieved within Proteinbase, each containing expression and hit-rates.

Targets

Targets are the molecules that designed proteins aim to bind. Each target has its own page showing general information and performance statistics across all tested proteins. We’re making it easy to find which methods performs best against any chosen target or which targets are the most popular.

Designers Profiles

All personal collections and designs can be showcased in a Designer Profile. We’ll also keep track of the designers’ favourite targets, success rates, and more, making it easy to quantify their progress. We’re expanding the Profiles even more in the future updates.

Contributing / Downloading data

All current data on Proteinbase is under ODC-BY license, meaning it’s free to explore, download, and use for anything the designers would want. This includes searching for leads to improve against a difficult target difficult or training some machine learning models to predict binding affinity. Any protein tested on Adaptyv can be published on Proteinbase. Learn more here: https://proteinbase.com/publish

What’s coming up next?

Collection drops: We’ll release new experimental data every week, be those from our own experiments, from community contributors, or from competitions. We’re open-sourcing the experimental measurements, so designers can use this data however they want and keep designing!

Designer profiles: We’re putting the designer in protein designer. We’ll enhance the profiles to include achievements, track metrics across proteins, and showcase rankings in the recent competitions. As people contribute and run more experiments, they’ll build a track record, and we’re making sure they’re rewarded for that.

Competitions: We’ll launch a new Protein Design Competition soon! But isolated competitions, albeit helpful, won’t help the field progress at the exponential rate we imagine it could. So we’ll launch multiple, constantly, gauging any shifts in the meta or any problems that appear to be “solved”. And, on top of that, we’re having ad-hoc research sprints and crowdsourced challenges - imagine a bio-hackathon every week. These are the fast design-build-test cycles we imagine, but open for everyone to contribute and engineer biology.

Why now

The protein design field is advancing at an incredible pace. New models, tools, and communities are emerging every month, pushing the boundaries of what can be designed and tested. However, the protein design landscape is still very fragmented.

We’ve seen this expansion along three major axes:

1. Open-source models and communities. Over the past year, open-source innovation has accelerated. Teams like Boltz, BindCraft, Mosaic, and Germinal have released powerful models with flexible licenses, allowing anyone to explore, modify, and build on them. This has created a thriving ecosystem of variants, forks, and experimental methods built from shared foundations. Proteinbase provides a common space for these open models to be tested, compared, and documented, turning scattered efforts into a cohesive public resource.

2. Industry-led research and proprietary models. At the same time, established companies are driving rapid progress with their own models and datasets. Chai Discovery, LatentLabs, DeepMind, Nabla Bio, Microsoft, and Cradle are releasing new design methods and publishing validation results at an unprecedented pace. These groups often set benchmarks and define new frontiers for what’s possible in protein design. Proteinbase complements this by offering a transparent reference point where results from both open and proprietary efforts can coexist and be compared.

3. Independent researchers and new entrants. The field is also attracting a wave of new participants, from academics and small startups to individual enthusiasts, who are eager to design proteins but often face fragmented tools and information. Many rely on incomplete information, online communities, and trial-and-error workflows. Proteinbase serves as an entry point for this growing group, helping them discover methods, learn from real experimental data, and connect their work to the wider ecosystem.

We’ve seen a massive momentum across all three axes, and we’re sure this is just the beginning! With Proteinbase, we’re building the home of protein design data, where protein designs, experimental results and design methods come together to be shared, compared and learned from.

Go check out Proteinbase now: proteinbase.com

When we started Adaptyv a few years ago, our core belief was: AI models for biology are only as good as the experimental data they're trained on and the hypotheses they can test in the real world.

Think about it this way - imagine your AI generates code for you but you could never actually run it. How would you know if the model is doing a good job? You can look at the code, check that there are no obvious errors, but you wouldn't really know if it works until you compile and run the code.

Now when you’re designing proteins, the equivalent to compiling and running your code means to make and test the proteins in the wet lab. Everyone can now run code thanks to AWS, Vercel, Modal and a million other cloud platform (or just on your local device) — but do you have a wet lab? Probably not.

So the choice here is either to build your own lab or work with so called Contract Research Organizations (CROs) that will set up experiments for you. If you go the lab route, you're looking at millions in equipment costs, hiring molecular biologists, and spending months developing protocols that might not even work. If you go the CRO route, you're sending emails back and forth comparing quotes, waiting 6-8 weeks for results, and when the data finally comes back, you're not entirely sure how they ran the experiments or why half your proteins didn't express. And then you need to iterate and do it all over again: More emails, more Teams meetings and less time to actually design proteins.

That's been the reality for protein designers. You design something incredible on your computer, and then you wait months to know if it actually works.

We saw that the ability to generated experimental data was the bottleneck to enabling a million people to design new proteins using AI. That's why we went after the hard and unsexy problem - building a fast, reliable, automated lab.

Now, after a year of working with many great partners, we’ve scaled our infrastructure to the point that we're now open to anyone who wants to use our platform!

Today we're dropping the "beta" tag from Adaptyv and launch our new website…

Where we are now

The main thing to understand about AI protein design is: The field is moving incredibly fast.

To give an idea: Last year’s Protein Design Competition brought together 150+ designers and we tested 600 proteins in our lab. Within just the two months between the two rounds, the hit rates went from 2.5% to 13%. Check out our series of blog posts or the community-written preprint about the lessons learned if you want to dive deeper!

And as the models are getting better, a massive design space is opening up, meaning even more proteins to be tested in the lab. Teams are designing de novo binders against every target they can think of. We’ve seen dozens of labs use BindCraft and test their designs in our lab — with success rates that people just 1 year ago were dreaming of. It is now possible to build novel biosensors like we demonstrated in our MBP sensor case study. Cradle is routinely optimizing antibodies to improve affinity, like they showed in our competition and their subsequent lab validation with Adaptyv. Microsoft Research validated their new model EvoDiff with us, Escalante just released Adaptyv data to demonstrate their new framework for multi-objective protein design and Chai Discovery recently released their zero-shot design model Chai-2, benchmarked using our lab.

Overall, this year, over 30 companies started using Adaptyv to validate their protein designs - from some of the biggest pharmas to frontier AI labs to many seed-stage techbio startups. We've run hundreds of experiments, tested well over 10,000 proteins this year and are generating the data that validates the best AI models currently in development. More than 10 preprints have been published in 2025 using Adaptyv data and more are currently being written.

Our team has more than doubled in the last 6 months and we just moved into our new lab and office in Lausanne to have more space to expand. There's a lot to build when you're trying to make biology work turn into

This is also the right time to announce that we raised an $8M seed round led by Ace Ventures earlier this year, with previous investors ByFounders and Founderful doubling down and LongGame and many great angels joining new. This lets us build faster, hire the right people, and scale our infrastructure to handle this massive demand. A big thanks to everyone that has been believing in us since the beginning!

What’s coming next

The real metric that matters to us is this: Every week, people are uploading new protein designs to our platform. New companies get started and test their proteins with us. People who otherwise wouldn't have been able to validate their proteins can now do it thanks to our platform. That's what we're building for.

So here’s what happening:

• Making the lab faster, cheaper, better - We're pushing hard on all three. Lower costs through automation and scale. More assay types for multi-modal AI models. And we're working toward bringing the turnaround time down even more.

• Alpha test our API - Biology should be as programmable as any other API. Connect your protein design agents directly to our lab. Run automated design-test-learn cycles and let your models order their own validation experiments.

• More open source data, more competitions - We're building something new for the protein design community to share designs, data, and insights. And there will be more competitions too. Stay tuned, we’ll release something cool soon!

• Join the team - We're hiring across bioengineering, software engineering, and lab automation. Also looking for our first non-engineering roles to help scale operations and onboard protein design teams. If you think biology should work more like software development, come work with us.

The protein design revolution is happening now. We're making sure everyone can participate.

• Julian for Team Adaptyv

TL;DR

• ProtRL - a framework for aligning (protein) language models with the distribution or score function you want using reinforcement learning (RL). You can use it for binders, enzymes, novel folds, and anything you can imagine.

• In this blog post, we’ll give you a brief review of reinforcement learning and how it has been applied to protein engineering - we hope this will get you up-to-speed with what ProtRL is doing!

• After, Filippo Stocco from the Ferruz Lab will be telling us all about ProtRL: why reinforcement learning is critical in protein design, how it works, how he’s currently using it to improve his initial EGFR binders with multiple rounds of experimental validation using Adaptyv’s binding characterization platform, and what’s next for ProtRL.

• We’ll finally speculate on some interesting reinforcement learning trivia: the infamous AlphaGo Move 37 - could this be possible in protein engineering? What would this look like? Are we there yet?

Protein language models are quite limited to their training data

ProtRL starts from a simple observation: protein language models (pLMs) capture the distribution of their training dataset. This condition is far to be ideal in protein engineering, where we aim to sample high fitness variants. For example, if that dataset is more biased towards highly stable, alpha-helical proteins (such as the PDB), then we should expect the same bias in the model when when we use a pLM for sampling new proteins (this has been a known problem for pre-training). But what if we want to sample something that is not well-represented in the training set, maybe a new enzyme?

This is where reinforcement learning comes into play: by biasing the model’s learned distribution in a variant of fine-tuning generally called “alignment”, we can increase the likelihood of sampling unlikely events.

Take for example the ZymCTRL model: a GPT-like pLM trained to generate new enzyme sequences starting from a simple enzyme commission number (EC) input in an autoregressive manner. When asked to sample 40,000 new carbonic anhydrases, the sequences captured the training dataset distribution. We can measure this by looking at metrics such as structural confidence (pLDDT) from ESMFold, the increased proportion of topologies like β (there are at least 5 families of carbonic anhydrases - α, β, γ, δ and ζ), and sequence lengths being closely matched. After aligning to a representative α carbonic anhydrase with the TM-score as an oracle (measuring structural similarity), ProtRL completely shifted ZymCTRL’s distribution: 95% of the generated sequences had the desired fold by the 6th round of reinforcement learning.

We see that ProtRL can be incredibly powerful for alignment!

But what is reinforcement learning?

In nature, goal-directed behaviours are shaped by rewards and punishments. RL in machine learning (ML) operates in a similar manner: an agent (model) interacts with its environment with a set of actions given its current state, refining its decisions (policy) over the set of actions based on the feedback it receives (reward). RL has successfully been applied to a wide range of challenges, from beating human players in Go and chess, to being world-class video game competitors.

For a deep dive into RL concepts, we recommend this OpenAI resource.

More recently, Reinforcement Learning has successfully been applied to Large Language Models (LLMs), following Human Feedback (RLHF). If you are familiar with Gemini, ChatGPT, Claude, or DeepSeek, you have experienced the power of RLHF first-hand! For instance, users can vote on ChatGPT responses, and sometimes the model presents them with two different options from which they can select the one that better aligns with their expectations. User feedback is then used to make the model more performant.

ProtRL is a framework where RL algorithms are implemented to be easily applied to any biological language model, now focusing on autoregressive pLMs like ZymCTRL. Currently, weighted directed preference optimization (wDPO) and group relative policy optimization (GRPO) are implemented, which we will explain in more detail later!

The landscape of protein RL

In the last years, RL innovations have also flourished in the protein engineering field. Broadly speaking, current efforts for protein RL can be abstracted into two main approaches:

1. Planning-based RL (seach-centric), which leverage a search algorithm (for example Monte Carlo Three Search) to explore the space of possible actions at each state. In this case, the possible actions are usually discrete. For instance, we can consider introducing a mutation in a sequence. The possible action space is defined as 20, the number of possible amino acids.

2. Policy-based RL (generative-centric), which learns the best way to explore the protein space by explicitly learning a policy (e.g. LLM) that maps from states → to optimal actions. This actions can be then queried directly to pick the best actions for each of the states, without passing through a search. This paradigm is the current hallmark for large scale modern generative AI.

We’ll briefly describe these and point you to some relevant papers.

Search-centric RL

In the search centric RL, the main objective is explore the space applying a search algorithm and applying a discrete set of possible actions in a very vast, yet constrained, space. For example, with the game of Pac-Man, we can decide between 4 different actions: up, down, left, and right. From this set of actions, we can explore the space applying a search algorithm and find the best combination of actions that will result in the higher outcome (e.g final score).

To better understand this, let's directly jump in one example of these approaches in protein design, that yet can generalize to other trends in RL. Monte Carlo Tree Search (MCTS) has been popularized by AlphaGo and AlphaZero to guide exploration (although strictly speaking they result in a kind of hybrid approach between the two categories here described). In this framework, the protein is modelled as a discrete sequence with each residue position as a potential mutation site. The policy network assigns probabilities for single-site changes across the sequence. Mutations are made and each of these "actions" spawns a new candidate sequence, and the search tree branches as we iteratively apply further mutations. Finally, each branch is evaluated to identify the mutation paths yielding the highest functional performance. This strategy was applied in EvoPlay, where the model "takes actions" generating several of these mutate-score trajectories to improve both its policy and value network (representing the mutation’s effect on fitness). As a result, EvoPlay generates luciferase mutants with 7.8x higher luminescence than the wild-type.

The agent can also operate on larger sections of a protein: for example, it can propose new secondary motifs (its domain being the protein backbone rather than the sequence) or monomers that should correctly form a nanocage. These still fit the search-centric RL category: the algorithm explores an exhaustive set of possible combinations, gets rewards for these, and can learn from the trajectories (self-play).

Generative-centric RL

Generative-centric RL is founded on a base generative model, which can be fine-tuned using reward or preference data to update model's parameters (θ).

Very recently, the field of research on aligning models with reinforcement learning algorithms has gained significant traction. This trend is particularly evident in the recent success of applying this technique to generative LLMs. Algorithms such as proximal policy optimization (PPO), direct preference optimization (DPO), and group relative policy optimization (GRPO) have emerged as powerful tools for aligning models to generate outputs that adhere to human preferences (e.g., generate correct answers to our questions, avoid toxic words), modelled as a preference probability. This field experienced remarkable success, particularly considering the widespread adoption of such technologies by the general public, emerging from research and academic circles, making it even more important to ensure an ethical alignment for these powerful tools.

PPO, DPO, and GRPO explained

One of the first policy gradient methods in RL is the REINFORCE algorithm. In REINFORCE, the agents interacts with the environments and the policy is updated directly by multiplying the reward. Intuitively, the higher the rewards of that specific action, the more likely that action will be in the updated model.

In the case of LLMs and PPO, a policy model generates a set of candidate sentences, which are then scored. The score itself represents a value that a human annotator would assign to a given output. However, directly using humans to score every model output is extremely costly and impractical. To address this, a reward model is first trained on a dataset of human-labeled examples. Once trained, this model can automatically evaluate and assign scores to new outputs, mimicking human judgment.

In addition to the reward model, a value model is trained, which has the role to predict the long term cumulative reward. Essentially, it estimates how good a partially completed output is, considering the future potential of the full sentence.For example, if we're training a model to talk about animals, the sentence "In the savanna" would be assigned a higher value than "In the pub". This is because "In the savanna" is more likely to lead to future content about animals, resulting in a higher expected cumulative reward.

This method helps the system optimize for long-term success, not just immediate reward, which reduces the risk of the model getting stuck in local optima. Finally, the reward and value outputs are combined to compute the advantage, which is used to update and improve the policy model (e.g LLM).

The need to train a reward model and a value model on top of the policy model makes PPO quite challenging to be applied to large scale alignments. To alleviate these issues, DPO allows the alignment of LLMs without explicitly learning and approximating a reward function, by learning directly from a preference dataset. In other words, compared to training a reward model where we need to connect the "inner knowledge of the model to a scalar value with limited data (compared to the training), with DPO we can directly tune the implicit reward", resulting thus in a potentially better approximation of the reward function (see this thread discussing the differences between DPO and PPO). DPO shifts the “preference burden” from the reward model to the training set, making the preferences itself align the model directly. This is a great resource if you want to understand the mathematical derivation from classic RLHF to DPO.

There is also some debate if DPO is RL or not but we leave that debate to the philosophers.

More recently, GRPO has gained great popularity. GRPO makes PPO more efficient by removing the value model and simply drawing multiple samples from its policy, then applying for each token the same advantage - computed by subtracting the sequence reward to the group mean reward and normalizing for the standard distribution of the group. Compared to DPO, GRPO provides greater flexibility making it potentially more adaptive to any type of distribution of the dataset - not just preference data.

Recent protein engineering applications of RL, including ProtRL, have been motivated by the success of PPO, DPO, and GRPO. For example, one of the earlier works - ProteinDPO - trained an inverse folding model (ESM-IF) with a DPO objective on pairs of proteins with high and low thermostability. This yielded improved correlations between the log-likelihoods and stability - thus shifting the generative model’s distribution to a higher stability one. Other works applied DPO for redesigning sequences with lower immunogenicity, better designability (as AlphaFold2 confidence), or more diverse peptides. PPO has been applied in the RLXF framework (reinforcement learning from experimental feedback) where experimental data and the PPO objective are used to align the ESM2 pLM to high fitness regions. It produced multiple variants of a fluorescent protein (CreiLOV), with one variant outperforming the best-known variant until then (1.7-fold improvement over wild-type versus 1.2). Other generative-centric RL models applied to protein engineering include: ProteinRL, LatProtRL, ProteinZero. Recent works have also been aligning discrete diffusion models instead of pLMs with an RL objective: RL-DIF for inverse folding, DRAKES, VIDD.

ProtRL : align your auto-regressive protein language mode

To explore the power of reinforcement learning in steering pLMs toward functional phenotypes, Filippo et al. developed ProtRL: a general framework that iteratively refines any decoder-only pLM against an arbitrary oracle. In each ProtRL iteration, the authors sample a batch of candidate sequences (e.g. 200), score them with their chosen oracle, and feed back the resulting phenotype–genotype pairs as reward signals to update the model.

In a progressively more complex tasks, they used ProtRL to align the model to generate a specific fold of carbonic anhydrase with higher probability than another one (beta to alpha). Over just six iterations, more than 95 % of generated sequences adopted the intended α-topology. Filippo et al. have also applied it to many other bounded and unbounded scores such as ESM1v, predicted enzymatic activity, and ProteinMPNN.

Yet, the most informative and real trustworthy oracle is the real-world itself. To demonstrate a full lab-in-the-loop campaign, Filippo et al. turned to binder design against the Epidermal Growth Factor Receptor (EGFR). They first fine-tuned ZymCTRL on 600 EGFR-related sequences (BLAST‐retrieved with wild-type EGF), generating 10,000 candidates and filtering them by perplexity and length (FT method). Six designs were tested experimentally in round 1 (coinciding with round 2 of the Protein Design Competition), yielding three binders with affinities of 328–819 nM. They then applied DPO (RL method), using measured Kd, sequence length, TM-score, and expression level as reward components. In round 2, tested again with the Adaptyv platform, four of nine tested sequences outperformed the best round 1 binder, including a top candidate with Kd = 27.4 nM. Remarkably, diversity naturally increases across rounds, with mean sequence identity dropping from 88.6% → 73.1%. Experimental results are summarized in the figure below.

ProtRL combines several powerful advantages in a single, seamless workflow: it can be driven entirely by synthetic data, needs only a handful of reinforcement learning iterations to deliver large fitness gains, and anything can serve as the feedback signal. Given these properties, ProtRL can be used to guide unconditional pLMs to explore vast, high-fitness regions of the sequence space while tailoring the sequences toward a property, producing entirely de novo proteins (such as BindCraft). In this scenario, the pLM is aligned to exhaustively sample vast, high fitness regions of the protein space, while meeting the reward function’s criteria. This paradigm has the potential to generate de novo binders or enzymes, all while avoiding intellectual property constraints.

Limitations

ProtRL and, in general, RL for protein design face several key challenges. First, reward hacking is pervasive: when the model over-optimizes a poorly specified reward, it can drift toward artifacts that score well but lack real biological function. Designing robust reward functions becomes therefore the most critical component in every RL campaign and requires many passes of hyper-parameter engineering. Here is an interesting blog post about this. Second, oracles (whether in silico predictors or wet-lab assays) are imperfect, and in the second case, often expensive to run. Training or integrating multiple orthogonal predictors (e.g. stability, expression, off-target effects) can increase computational and experimental cost, limiting throughput. Third, real biological systems are high-dimensional and context-dependent, far more complex than the closed environments of games. Generalizing across folds, functions, and host contexts remains an open problem. That said, we remain optimistic! There are many exciting opportunities to be explored in the field, that are advancing at breakneck pace!

Going further: Move 37 in protein design?

Move 37 is a famous move played by AlphaGo against the Go champion Lee Sedol in March 2016 (Netflix also made a documentary about this!). This move represented a play so non-intuitive that it left even one of the best players in the world speechless and, in the end, was a winning move. This move was described as an emergent phenomenon from RL training. AlphaGo was trained so that the model was able to self-play countless games and compute the score for each play, thus the set of decisions was made based on the outcomes of the match (winning or losing).

RLHF has revolutionized NLP by aligning language models to human-annotated data. Yet, the nuances and the underlying workings of the biological “grammar” are quite challenging to grasp. For this, relying on human annotation in pLM is not feasible. Additionally, aligning outputs with existing human knowledge risks confining generative models to familiar evolutionary motifs or research bias/trends rather than pioneering truly novel solutions.

In a recent substack, Prof. Levine describes LLMs as projections of human capabilities, distilled from the processing of the enormous amount of information of data present on the internet. Yet these models are still far from achieving the generality and flexibility of human learning processes, which can extrapolate and connect multiple subjects and concepts. In other words, LLMs have just learned to mimic our behaviour without implementing/exploiting the underlying learning processes we use. In a way, LLMs are constrained to a world of shadows, exactly how Plato, in his Allegory of the Cave, described humans able to see only the mere shadows of the "real world", chained in the cave of the ignorance. For this reason, according to Prof. Levine, AI systems will not acquire the flexibility and adaptability of human intelligence until they can actually learn like humans do, shining with their own light rather than observing a shadow from ours, and finally escape from the Plato’s Cave.

RL, especially in protein design, can play a different role compared to the NLP field, as each generation can be assessed and scored, things that we cannot easily do with language (how can we computationally rank and compare Virginia Woolf with Cervantes?). For this reason, as interestingly pointed out by Max Jaderberg in an interview, protein design problems may be seen more from an AlphaGo point-of-view rather than the ChatGPT one, where the model can autonomously learn the best policies / moves to achieve a determined objective, unconstrained by our knowledge.

But what would Move 37 look like in biology? Just as AlphaGo’s Move 37 emerged as a counterintuitive, unexpected strategy, a true biological Move 37 will be a mutation or fold no expert anticipated yet one that increases functional performance. For a protein engineering campaign, it could be a mutation that drastically reduces fitness for several rounds, until it yields a way better variant (e.g., the model anticipating a “fitness valley” that needs to be traversed to reach a peak). Translating this paradigm to proteins will require hybrid architectures integrating pLMs with reward functions that explicitly value novelty (there are some works to robustly assess protein novelty), and high-throughput testing or assemblies of oracles to navigate vast mutation landscapes, paving the way for real unconstrained exploration of the space and discovering the first Move 37 in protein design. Simply put, a Move 37 needs more search / exploration - AlphaGo achieved that through self-play in a constrained environment with a clear reward feedback. For biology, we either need better experimentally-aligned oracles (to create a constrained environment similar to AlphaGo’s), more experimental testing (to search and validate), or better methods to enforce diversity, novelty, and going “out-of-distribution”. Or all of the above!

Conclusion

ProtRL tries to lay the groundwork for iterative, oracle-driven protein design, but the ultimate frontier lies in discovering “blind spots” where neither nature nor existing models nor experiments have ventured. By combining synthetic datapoints, wet-lab feedback, and autonomous learning, we would like to get close to what Move 37 was for Go: a design leap so unexpected it rewrites our understanding of what proteins can do!

Resources and links

• Try out ProtRL here and read the preprint here.

• Check out what the Ferruz Lab is working on!

• Say hi to Filippo: X, LinkedIn.

• Have some novel proteins you want to test in the lab? Come talk to us — we’d like to run many more of these protein designer spotlights, so if you have a cool new hypothesis or model to test we’d love to hear from you!

Acknowledgments

We are grateful for the invaluable feedback and insightful discussions provided by Noelia Ferruz and Michele Garibbo.

TL;DR

• Michael Hla built Pro-1, a language-based reasoning model able not just to design proteins, but also reason about how it’s designing them.

• We offered him free binding affinity and thermostability testing in the Adaptyv Foundry for 19 of his FGF-1 (Fibroblast growth factor 1) sequences optimized by Pro-1.

• The results are interesting: Pro-1 to improve the melting temperature of 3 designs while maintaining the binding affinity at the same time.

• One variant (K116E) reached a similar melting temperature to the most optimized design from literature (Q40P, S47I, H93G, K112N)!

• We will run more benchmarks or test cool protein design hypotheses you have - reach out to us!

Pro-1: the protein reasoning model

In March, Michael Hla took the protein design community by storm when he released Pro-1 - the first protein reasoning language model. His proposal is simple: let a language model distil biochemical intuition so it can do all the protein optimization you want. What this means more precisely: use a complex training scheme, with recent innovations in large language model reasoning used by AI labs like OpenAI and DeepSeek, to have a model that both proposes mutations and explains why it chose them.

Michael makes several good points for why protein design should be delegated to language models in his blog post. Some of the more interesting ones we found are interpretability - the model argues for each mutation it makes, pointing to relevant (or hallucinated) biochemical motives and paper references; and flexibility — he mentions Pro-1 can be prompted with sequences, PDB structures, even experimental results.

How to train your protein thinker

We found his training scheme incredibly unique. Michael combines biochemical intuition with synthetic data generation from specialized protein language models, a training framework from reinforcement learning, and a physics-based representation of protein stability. We will briefly describe it, but you should check out Michael’s blog post and thread for more info!

1. Fine-tuning on synthetic reasoning traces

   Language models are often trained on large text corpora in two ways: either words are masked from the input and the model is tasked to predict them (masked language models), or the model has to correctly predict the next token (autoregressive language models). This is the pre-training stage.

   Pro-1 uses the pre-trained autoregressive Llama-3.1-8B Instruct and Llama-3.3-70B-instruct, adapting them to the protein design task in a process called fine-tuning. To make models reason about their design, Michael generated synthetic “reasoning traces”: initial proteins from a collection of enzyme sequences (BRENDA database) were “perturbed” with the ESM-3 protein language model. He then generated text explanations for how to get from the perturbed to original proteins with a different language model. This is incredibly unique! Michael points out that “this method needs to be tested more but has substantial implications if it scales well, especially since bio data is exceedingly scarce”.

2. Reinforcement learning with the Rosetta energy function

   Next, Pro-1 uses the group relative policy optimization (GRPO) - a reinforcement learning algorithm now well-known because of DeepSeek’s R1. In summary, Michael takes the Rosetta energy function, which accounts for several physical interactions and it well-correlated with protein stability, to score proteins designed by Pro-1 and then folded with ESMFold. The final value is integrated into GRPO and fed back into the model to improve it. Pro-1 should now output more stable proteins and “learn heuristics about the physical world and the effects of specific mutations”.

3. Creativity rewards

   Michael mentions the final model got “somewhat repetitive and bland, suggesting the same types of point mutations (polar aa -> nonpolar aa)”. He then included a judge model into the training scheme, which scores mutations based on how “creative” they were. It boosted the performance on his benchmark from 43% to 47%!

We were all impressed by Pro-1, so we wanted to put it to the ultimate test: lab validation. We gave Michael some free binding affinity and thermostability assays for any designs he wanted. He chose to optimize the fibroblast growth factor 1 (FGF-1).

Why FGF-1?

As a growth factor, FGF-1 is one of the most versatile proteins. It regulates the fate of bone marrow cells and may promote bone repair, the development of lung epithelial cells with a therapeutic effect on pulmonary fibrosis, and is highly expressed in inflammatory cells. Its role in type 2 diabetes is becoming better understood, with experiments showing FGF-1 injections reduced the levels of glucose and increased the sensitivity to insulin in mice.

It binds to plenty of targets, including the fibroblast growth factor receptor 1 (FGFR1) and FGFR2. FGFR1 aberrations occur in several types of cancer and there are already FGFR1-inhibiting drugs like Pemigatinib for bile duct cancer treatment. FGF-like binders to FGFR1, especially when conjugated with cytotoxic drugs, could be a potent cancer therapeutic.

Michael mentioned another interesting fact about FGF-1: it has a pretty low denaturation temperature. Maintaining its binding while also increasing the melting temperature is a worthwhile task for Pro-1.

How we are measuring melting temperatures and binding affinities

We ran our standard automated assay for affinity characterization and thermostability. Proteins were expressed with a cell-free system, followed by affinity characterization via bio-layer interferometry (BLI) with the FGFR-1 target, FGF-1 wild-type control, and the designs Michael uploaded on our Foundry Portal. BLI measures the binder association and dissociation kinetics via the interference pattern of light reflected from a sensor surface. With these measurements and our in-house post-processing and curve-fitting software, we can calculate the binding affinity ($K_D$) of a protein to its target.

To measure the melting temperature of Michael’s designs, we used our newly-developed thermostability assay. The melting temperature (or $T_m$) represents the temperature at which 50% of a protein is in its unfolded state. This is around 49 °C for the wild-type FGF-1. We are using an automated nanoDSF (nano differential scanning fluorimetry) protocol: proteins are heated up and we measure the fluorescence shift of the tryptophan and tyrosine amino acids as they get more exposed from the protein’s core. We normalize these values and quantify the melting temperatures in our post-processing pipeline.

Pro-1 yields more stable binders

Most of the variants tested were expressed. Out of these, only 6 maintained binding to their target in the same range as the wild-type. In the figure above, we have highlighted the 3 variants with a $T_m$ higher than the measured control - the wild-type FGF-1 with 50.8 °C - that also bind to FGFR1.

What is more impressive is that a single-point mutant (K116E) reached a melting temperature improvement of 24 °C over the wild-type, and that Pro-1 even suggested this variant. When we consider it was trained on synthetic “perturbed” data and reasoning traces and an in silico objective (the Rosetta energy function), these are spectacular results! Most other Rosetta-based thermostability optimization studies also reach an improvement of 20 °C, yet none of them have a model able to explain in writing why it chose those mutations.

Michael showed an example of Pro-1’s reasoning trace for v37 variant (the 7-mutant in our bar plot). We found it interesting how it knows FGF-1’s binding partners and some plausible biochemical interpretations (e.g., mutations that reduce flexibility should increase stability, targeting hydrophobic patches to reduce the chance of aggregation). However do not attempt to fact-check its references: we tried that and we could not find any “Wu et al., 2017” that suggested the K127E mutation could increase stability of the FGF2 heterodimer, nor any “Kim et al., 2015). But this should not diminish Pro-1’s success - who knows, maybe the Pro-2 will align its reasoning with verified references. If OpenAI’s Deep Research can do it, so could Pro-1.

Resources and links

• You can find all thermostability data here and the binding affinity data here

• Try out Pro-1 here

• Say hi to Michael Hla: Website, X, LinkedIn

• Have some novel proteins you want to test in the lab? Come talk to us — we’d like to run many more of those protein designer spotlights, so if you have a cool new hypothesis or model to test we’d love to hear from you!

TL;DR

We wrote a community paper about our Protein Design Competition, teaming up with your favourite protein designers from both rounds.

We aimed to explore the competition data in as many ways as we could, take stock of the state of the field, what SotAs have been established across the rounds, and which metric is most predictive of binding and expression.

The one thing we kept meeting throughout: the lack of a standardized benchmark set for protein binder design, leading to difficult comparisons and a lack of consistent, high quality data. This is why we close the paper by creating BenchBB, the Bench-tested Binder Benchmark — a curated set of 7 protein targets designed to capture diverse binder design challenges by remaining accessible enough for wide scale lab validation.

Want to read the community paper? Check it out on Biorxiv!

Want to test your protein-design model on BenchBB? Go to benchbb.bio and get started!

The community paper after the Protein Design Competition

As you might remember, we hosted the Protein Design Competition. Briefly, we called for protein designers to create novel binders for EGFR and received over 1857 total submissions, 600 which we then experimentally tested in our lab, validating a total of 60 novel binders! For details you should check our previous blog posts. However, in these posts we could barely scratch the surface of possible data exploration. There was a strong demand to do more analyses and collect all the learnings, including those that our participants had already been doing throughout the rounds.

The spirit of collaboration was strong throughout the competition so it did not seem fitting for us to act as the curators and gatekeepers for what would be included or said in such a writeup. Instead, we decided to form a consortium and launched an open call for collaboration. To our delight, several of the participants responded, and the excitement to contribute turned into a wide range of analyses, with people applying their expertise in different types of biomolecules (from antibodies to peptides), statistics, and general understanding of the current state of computational protein design, and vision for its future. The discussions in our Slack channel were some as delightful as they were interesting, and we finally distilled them into the preprint Crowdsourced Protein Design: Lessons From The Adaptyv EGFR Binder Competition. We will now summarize some of its key takeaways, but we strongly encourage you to read the full paper (honest clickbait: table 4 might surprise you!)

In summary: More data = better metrics & models = better proteins

We took another look at the data we had already analyzed in our blog posts, to see whether we can identify patterns and trends on the combined dataset of both rounds.

While there was a sizeable increase in expression success and binder hit rate (see below), there remained many questions about when certain protein design approaches might be better than others and how to predict whether a protein makes a good binder from just the sequence alone. During the competition, we received many more protein designs than we could test in our lab during the timeframe of the contest. We thus had to select which designs to experimentally validate and which not.

Our strategy was the following:

1. We ranked designs by a set of computational metrics (ipAE, iPTM and ESM2 PLL) and chose the top 100 designs according to this ranking.

2. We then additionally selected another 300 designs based on whether the protein designers had described a particularly interesting or novel design method.

As we already suspected that the computational metrics might not correlate well with the binding affinity, this approach also rewarded protein designers for proposing new design methods that not just aimed at maximizing the computational score.

In the paper, we run lots of statistical analyses to check the correlation of various surrogates (including iPTM, ipAE, and ESM 2/3/C) with binding strength and found that:

1. at least on our dataset, ipAE, iPTM and ESM2 PLL (normalized or not) only correlate weakly with KDKD​, despite some of them being part of the competition target metric.

2. the good news, as Nikhil Haas (BioLM) had already noted, ESM3 and ESMC, when length normalized do correlate with KDKD​, at least on our dataset.

We don’t recommend you to rush towards blindly maximising this metric though, because as we show later in the paper, specific antibody domains might require different metrics to inform binder design — huge shoutout to Nikhil and the team at BioLM for donating both the data and their valuable time for making this analysis possible.

Beyond this, there’s a lot more in the paper and the supplementary, e.g. a detailed study of the specific EGFR domains and their role in the successful binders, highlights of methods used throughout the competition (including Cradle’s winning entry, which they have been further evaluating and explaining in more detail in their own series of posts) and more context on the competition and the community response. We thank all the authors for the time and resources they to get the paper done this way.

However, the one thing that kept being said as we did all of these analyses and was confirmed with every insignificant p-value our analyses yielded is: we need more data.

Thus, in the discussion we tried to acknowledge the great advances that became obvious as we looked back on the data, but also stress the limitations and challenges the field still faces:

1. Computational metrics are getting better, but are not yet plug-and-play reliable, largely caused by the extremely non-standardized datasets they are derived from, often using different assays

2. This lack of standardization extends as far as the definition of a binding hit, making it very difficult to compare results from one report to another

3. Even if the assays and hit definitions were stable, everyone makes up their slightly tweaked sets of targets right now, with only a few if any targets being shared across studies.

Of course, it is too easy to only complain about things, so we also suggest a first step in fixing this situation.

Introducing BenchBB: the Bench-tested Binder Benchmark

BenchBB is a curated set of 7 protein targets designed to provide a rigorous, consistent, and practical benchmark for computational binder design methods. While recent papers (RFdiffusion, AlphaProteo, BindCraft) have started to partially use a few common targets, the field lacks a standardized minimal set to objectively compare computational approaches. BenchBB directly addresses this gap.

We selected targets by balancing multiple factors:

• Novel interfaces: the targets should not be heavily represented in standard ML training datasets.

• Challenging conformations: the targets have significant conformational changes and thus require varied binding mechanisms.

• Therapeutic relevance: target should have potential translational impact.

• Accessibility: ease of recombinant expression, primarily in E. coli, enabling broad lab validation.

To ensure consistent, comparable evaluation across studies, we propose the following standardized assay approach for binding measurements:

• Use label-free sensing methods such as Bio-Layer Interferometry (BLI) or Surface Plasmon Resonance (SPR) to accurately measure kinetic parameters (kon, koff) and compute affinity constants (KD) from them.

• Whenever feasible, share assay parameters and conditions (e.g., sensor type, immobilization method, analyte concentration range, and fitting methods) as well as the raw kinetic data.

• Define a binder or “hit” as having a clearly measurable interaction signal with KD ≤ 10 µM.

So let’s meet the 7 target proteins!

• This one needs no introduction - it was the catalyst that launched our entire competition. But, briefly, EGFR’s extracellular domain (~620 AA) binds EGF and TGF-α; it is frequently overexpressed or mutated in several cancers; and several therapeutic antibodies (e.g. Cetuximab) target it. PDB ID: 8HGO.

• We have accumulated a solid data set thanks to the participants in both competition rounds. Designers can further expand it or compare their tools or results to the data we have released - this is one of the main reasons to include EGFR.

• Cao et al. 2022 designed 50–65 aa miniproteins that bound EGFR’s Domain I and Domain III, successfully blocking EGF-induced signaling. They reported, however, an 0.01% hit-rate. We saw in the community paper how this was significantly improved upon: almost 3% in Round 1, then 13% in Round 2. And let’s not forget about the 8.2x binding affinity improvement over Cetuximab that Cradle achieved.

• IL7Ra is the alpha subunit of the IL-7 receptor (CD127), a 219 AA cytokine receptor critical for T-cell development. PDB ID: 3DI3.

• Other than its therapeutic relevance (blocking IL-7/IL-7R interaction could modulate immune responses), we chose it because its “ectodomain is easily produced in human cells and has been benchmarked in multiple prior studies”.

• RFdiffusion yielded multiple IL-7Rα binders where earlier Rosetta designs yielded almost none (original ~2.2% with AlphaFold selection to a reported ~34% for RFdiffusion). One designed binder showed nanomolar binding and inhibited IL-7 signaling in vitro. Cao et al. 2022 reported an 0.05% pre-AlphaFold hit-rate for de novo binders. AlphaProteo also generated strong IL-7Rα binders in one round. All these were de novo mini-proteins (~50–60 AA) that expressed well in E. coli and bound IL-7Rα with high affinity (comparable to or better than a natural IL-7:IL-7R interaction). They report a 24.5% success rate, greater than the remeasured RFdiffusion one (16.8% versus the original 34% published).

• PL-L1 is an immune checkpoint ligand (~290 AA) expressed on cancer cells and APCs. PD-L1 binds PD-1 on T cells, suppressing immune responses. PDB ID: 4Z18.

• Gainza et al. 2023 noted PD-L1’s surface “displays a flat interface considered to be ‘hard to drug’ by small molecules”, thus making it ideal for testing advanced design methods. We consider it a “de facto binder design benchmark target”

• RFdiffusion reported a 12.6% hit-rate. It was additionally benchmarked by AlphaProteo, MaSIF, BindCraft, and Yang et al., 2025.

• BBF-14 is a de novo designed 112-residue β-barrel protein (13.8 kDa) with an internal hydrophobic pore. PDB ID: 9HAG.

• Serves as a stress-test for binder design on a novel, non-natural target. With BBF-14, there are no evolved binders or known epitopes – designers must rely solely on the computed structure. Thus, it “can assess generalization beyond natural interfaces”.

• It was previously used as a target in the BindCraft paper, where one design (“binder4”) bound BBF-14 with KDKD​ of 20.9 nM (SPR). BindCraft achieved a 55% hit-rate (6/11) on BBF-14.

• BHRF1 is a viral anti-apoptotic protein from Epstein–Barr virus (EBV) that mimics Bcl-2, allowing infected cells to evade apoptosis. It is associated with EBV-linked cancers. PDB ID: 2WH6.

• Our main reasons for choosing were that it is “easily expressed in E. coli and commercially
  available with many antibody controls”. Additionally, it has a known hydrophobic hotspot - the BH3-binding cleft, which restricts the search space.

• Targeted initially by Procko et al. 2014 - their de novo 86 AA minibinder (“BINDI”) could bind BHRF1 with 220 pM affinity (PDB ID: 4OYD). More recently, AlphaProteo reported an 88% experimental hit-rate, far above prior methods, yielding multiple nanomolar binders without any optimization.

• MBP is a 42-kDa periplasmic binding protein in E. coli that binds maltose/maltodextrins as part of a sugar transport system. It is very stable and well-expressed; commonly used as an N-terminal fusion solubility tag to aid recombinant protein expression. PDB ID: 1PEB.

• MBP’s abundance and stability make it easy to produce and assay, thus it can be tested in any lab. Another reasons for choosing it is that it features “a well-characterized active site allowing straightforward binder screening via elution from amylose resin”.

• Zhou et al. 2025 employed de novo design and computational screening to create MBP binders: “6 candidate binders targeting MBP” were identified without any directed evolution. These hits were small folded proteins (≈80–100 aa) that bound MBP with low micromolar to nanomolar affinity.

• This RNA-guided DNA endonuclease needs to further introduction - it the key in CRISPR gene editing, widely used in gene editing and any genomic biotechnology application. PDB ID: 4OO8.

• We chose it for its “easy structural characterization via cryoEM; stable, easily expressed in E.
  coli, and with multiple known binding sites and conformations”. Cas9’s size and moving parts make binder design difficult, but a successful binder can act as an “off-switch” for genome editing (de novo binders regulating the enzyme function). The BindCraft authors note including “multi-domain nucleases, such as CRISPR-Cas9” as challenging targets.

• Used as a target for BindCraft - where a small designed protein binder (~100 AA) bound Cas9 and inhibited its genome editing activity. Surprisingly, it also yielded a 100% hit-rate, with the best binder measuring a KDKD​ of 267 nM (SPR).

Next steps

Want to read the full paper? Check it out on Biorxiv!

Want to test your protein-design model on BenchBB? Go to benchbb.bio and get started!

Got questions about the paper or BenchBB? Just email us at benchbb@adaptyvbio.com

Acknowledgements

Thanks to all the consortium authors for joining us throughout this journey, their contribution to the paper and in-depth discussions over the benchmark targets!

• Filippo Stocco, Noelia Ferruz from Centre for Genomic Regulation, Pompeu Fabra University

• Anthony Gitter from Department of Biostatistics and Medical Informatics, University of Wisconsin–Madison; Morgridge Institute for Research

• Yoichi Kurumida from School of Frontier Engineering, Kitasato University

• Lucas de Almeida Machado from Instituto Oswaldo Cruz, Fiocruz

• Francesco Paesani, Cianna N. Calia from Department of Chemistry and Biochemistry, University of California San Diego

• Chance A. Challacombe, Nikhil Haas, Ahmad Qamar from BioLM

• Bruno E. Correia, Martin Pacesa, Lennart Nickel from École Polytechnique Fédérale de Lausanne (EPFL)

• Maxwell J. Campbell from Hearth Industries

• Constance Ferragu, Patrick Kidger from Cradle Bio

• Logan Hallee from Synthyra; Center for Bioinformatics & Computational Biology, University of Delaware

• Christopher W. Wood, Michael J. Stam, Tadas Kluonis, Kartic Subr, Süleyman Mert Ünal, Leonardo Castorina from University of Edinburgh

• Elian Belot

• Alexander Naka from Science Corporation

At Adaptyv Bio, we make it easier for researchers to test their protein designs by providing fast and reliable experimental validation. In this case study, we highlight how Microsoft Research used our automated lab to validate proteins generated by EvoDiff, their novel sequence-first protein design model, in just a few weeks.

For a deeper dive into EvoDiff, including its methodology and broader applications, you can check out the EvoDiff paper.

Want to experimentally validate your own proteins? Start configuring your first experiment here!

What is EvoDiff?

EvoDiff is a generative model developed by Microsoft Research that operates directly in sequence space, bypassing the need to generate 3D structural models as part of the design process. This makes it fundamentally different from structure-first approaches like RFdiffusion, which rely on predicting and optimizing 3D protein structures before converting them into sequences.

By focusing on sequences, EvoDiff can explore regions of the design space that structure-based models may struggle to reach. This includes intrinsically disordered regions (IDRs), which lack a stable 3D structure but can still be highly functional, and other unconventional protein scaffolds.

For example, where RFdiffusion would need to design a specific structural fold and then identify a sequence to stabilize that fold, EvoDiff directly generates sequences optimized for the desired function, skipping the intermediate structural step. This gives it more flexibility, particularly for targets where structural constraints are less critical or poorly understood.

EvoDiff can be tailored to different design scenarios via conditioning:

• Filling Gaps: If you have part of a protein sequence, EvoDiff can fill in the missing regions while optimizing for specific properties, like binding affinity or solubility.

• Scaffolding Functional Motifs: The model can generate sequences that incorporate known binding motifs into diverse protein backbones.

• Mimicking Evolutionary Patterns: Using evolutionary alignments, EvoDiff can design sequences that resemble naturally occurring proteins, potentially improving stability and functionality.

This versatility makes EvoDiff a promising tool for both exploratory and application-driven protein design projects.

Protecting the ‘guardian of the genome’

MDM2 is a protein present in most human cells, where it plays a crucial role in regulating cell growth and survival. Its primary function is to control the activity of p53, another protein which often referred to as the “guardian of the genome.” p53 protects against cancer by halting cell division or triggering cell death when it detects DNA damage. MDM2 binds to p53 and signals for its degradation, effectively silencing this protective function when it is no longer needed.

In many cancers, MDM2 is overproduced, disrupting the balance and allowing cells to grow unchecked. This makes MDM2 an important target for cancer therapies. In this study, EvoDiff was used to design proteins that bind to MDM2. Competitive binding to MDM2 via a synthetic protein could help restore p53’s tumor-suppressing role.

However, designing molecules to bind MDM2 is challenging because you need to precisely target the p53-binding pocket to block the interaction.

EvoDiff employs a strategy known as “motif scaffolding,” where the binding motif, derived from p53, is preserved, but its stability and binding efficiency are enhanced by a scaffold designed by the model. This scaffold is optimized to hold the motif in the ideal conformation for binding to MDM2, improving the chances of effectively blocking the MDM2-p53 interaction. Importantly, with EvoDiff’s sequence-based design framework, this motif scaffolding occurs entirely in sequence space – without any information about the MDM2 target or the structure of the p53 binding peptide.

Validating the protein designs in the lab

We used our high-throughput Affinity Characterization workflow to validate EvoDiff’s designs. This streamlined process provides fast, high-quality data with minimal manual intervention.

1. DNA Synthesis & Protein Expression: We designed and optimized DNA constructs for the EvoDiff designs and produced the proteins in a high-efficiency protein expression system.

2. Binding Analysis: Using biolayer interferometry (BLI), we measured how well each protein bound to MDM2, calculating dissociation constants (KDs).

3. Data Analysis: We classified binders, analyzed binding strength, and generated detailed kinetic profiles to evaluate performance.

After the team from Microsoft Research uploaded their protein designs to our platform, the entire workflow from designing the DNA sequences to uploading the final results took just about 3 weeks thanks to our automated wet lab.

Results & Takeaways

Strong Binders

Out of 24 tested EvoDiff designs, 8 designs showed KD values below 50 nM for full-length MDM2, with the top performers reaching 25.9 nM and 40.2 nM, respectively.

Fragment vs. Full-Length Performance

When tested against the MDM2 fragment (containing just the p53-binding domain), most binders showed reduced affinity compared to the full-length protein, with KD values for the fragment often falling into the micromolar range. The full-length protein likely provides additional interactions that stabilize binding, making it critical to validate designs across multiple contexts.

Evolutionary Conditioning Advantage

Designs generated with evolutionary conditioning (using multiple sequence alignments, or MSAs) consistently outperformed unconditioned designs. These binders demonstrated stronger affinities and greater reliability across full-length and fragment tests. Conditioning with evolutionary data seems to enhance design quality by leveraging patterns that have been naturally optimized over millions of years.

Design your own experiment

Ready to test your own designs? Use our Experiment Configurator to validate your sequences and take your protein engineering projects to the next level!

We’d like to thank everyone that contributed to EvoDiff: Sarah Alamdari, Nitya Thakkar, Rianne van den Berg, Neil Tenenholtz, Robert Strome, Alan M. Moses, Alex X. Lu, Nicolò Fusi, Ava P. Amini, Kevin K. Yang. Check out their paper here!

TL;DR

The results of our second Protein Design Competition show dramatic progress in the field. Out of 400 tested proteins, 378 expressed successfully (95%) and 53 of these bound to their target EGFR (14% success rate). This represents a 5x improvement in binding success compared to just three months ago. The best designs even matched or exceeded the performance of Merck's Cetuximab, a clinically approved therapeutic antibody.

The competition demonstrated that protein design is becoming more reliable and accessible - 30 out of 130 participants managed to create at least one successful binder using various approaches, from optimizing existing antibodies to designing completely new proteins from scratch.

While we can now generate functional proteins more consistently, a key challenge remains: predicting which designs will work best before testing them in the lab.

We released the results of our second Protein Design Competition just about 2 weeks ago… and they were great!

A quick recap of what this was all about:

• We tasked designers to create new protein binders for EGFR, an important binding target for therapeutic proteins.

• We then ranked them based on a weighted score of AlphaFold2’s ipTM and iPAE metrics (scoring how well AlphaFold “likes” the interface) and ESM2 pseudolikelihood (ESM2 PLL), a measure of how “natural” a sequence looks like given ESM2 learned distribution.

• From these, we chose 400 sequences to be tested experimentally in our lab: the top 100 as determined by the rankings and then another 300 handpicked by us at Adaptyv, where we included at least one protein from each designer and then tried to maximize the diversity of interesting models and design strategies.

And the designers did not disappoint. In our previous analysis here, we went through how the meta shifted from the first round, in particular how close the race was in the last few days as some submitters tried to increase their chances of landing in the top 100 spots.

We had also given you some guidance regarding which hotspots to target on EGFR, which models to use to improve your binder’s expression chances, and which models to use if you’re truly interested in venturing into the de novo world. We believe these guidelines, and the community momentum that was bubbling up the scenes, with “protein design recipes”, blog posts, and general advice, were the causal factors that led to an increase in both hit rates and expression: from 2.5% for the former to 13%, and from 76% to 96%. These figures are something to be proud of, and we would like to thank all participants for joining, getting involved in the community, and especially those who were actively writing about their methods. We believe these nuggets of info will be valuable to move the field of binder design forward.

After all this preamble, let’s outline what this blog post will cover: We’ll first recognize the winners: the top 3 designers as ranked by binding affinity and a special prize for the de novo designs. The top 3 designers had the chance to present their workflows at the recent NeurIPS AIDrugX workshop. Then, we will go through a rapid-fire round of data analysis.

Spoiler: we’ll publish a paper that will go more in-depth into both rounds and what we’ve learned from them. If you want to join in on this effort (either by writing about the method you used, making some nice figures, or just debating with participants about statistical analysis and confounding factors), and if you took part in one of the rounds, please fill out this form here. Thus we’re keeping our analysis short and snappy here, focusing on only the second round, but there is more to investigate!

Lastly, you can download all binding data from the first and second rounds. Tag us and let us know if you have done any analysis on this and have some cool findings!

Meeting the winners

The number 1 spot was taken by Cradle and their optimized Cetuximab variant.

Designed in ”about 30 minutes and over some kombucha”, it highlights the strength of Cradle’s platform to optimize existing proteins for a set of biologically relevant properties, binding affinity in this case. As mentioned in their design process notes, they used a zero-shot approach without training their model on any labelled data. This yielded an improvement of Merck’s Cetuximab, by making exactly 10 stabilizing mutations in the framework regions (”FRs” as opposed to “CDRs”, the complementarity-determining regions, which are highly variable and form the typical antibody binding loops that determine their specificity).

Similar approaches to induce loop-stabilizing mutations in framework regions have been proven effective before. Moreover, recent advances in antibody engineering propose modulating the Fc regions as well, resulting in improved half-lives, as opposed to the current approaches of Fab changes (including FRs and CDRs) to modulate affinity and specificity. So there is a wider potential in changing or improving an antibody’s function than only CDR engineering, and Cradle demonstrated it in this competition to great success.

A optimized nanobody takes the number 2 spot. Designer Chris Xu mentions that his workflow involves CDR grafting, in which the FR regions are replaced with homologues human germline ones, followed by optimization of CDR stability and hydrophobicity, followed by developability filtering, with metrics for “humanness” and immunogenicity. Despite not ranking high in the leaderboard as his approach did not directly optimize for them, we are very excited that a lot of people approached our challenge from a similar perspective: not only designing a binder, but aiming to also optimize other therapeutic properties at the same time.

In terms of binding affinity, Chris’ nanobody is also on par with Cetuximab. Even more interestingly, his approach yielded two other binders, one with a binding affinity in the tens of nM range, and one in the hundreds. We’d love to see more experiments on those designs to see if their optimized developability properties hold up in the lab!

Our 3rd spot was claimed by Aurelia Bustos and her TGFα optimization approach. We’ve highlighted this method before, and you can find a more detailed explanation from her as well. Briefly, she started from TGFα, masked out the binding region, then used it to scaffold the non-binding backbone using RFdiffusion. She followed up with ProteinMPNN inverse folding and trimming the excess residues to maximize ESM2’s log-likelihood. Thus, her resulting sequence is 50 amino acids in length, likely due to the unnormalized likelihood score we have been using.

This achieved the best of both worlds in this competition: several of her designs ranked in the top 20 as they optimized the metrics and a lot were good binders. 8 of her submissions bound, in fact, which results in an 80% success rate, with some in the tens, hundreds of nM, or mM ranges. Almost all had better binding affinities than the EGF control. However, we did not have a TGFα control to compare and see if the affinities also improved compared to the natural ligand.

The best de novo designed protein comes from Lennart Nickel & Martin Pacesa using BindCraft.

If you paid attention, you will have noticed that the top 3 spots were dominated by redesigned binders, from known antibody therapeutics, to nanobodies and binders from nature. Yet, a lot of participants wanted to push their strategies to their limits by generating something completely brand new, a de novo binder. Here

For this, BindCraft seemed to be the favorite approach: The model generates a de novo binder by sampling a noisy initial state, and increase the chances for landing in the top 100 by optimizing (via gradient descent) for metrics like iPAE and ipTM. We noticed that it wasn’t as successful at landing in the top 100 spot compared to directed evolution, trimming excess amino acids, or other methods. We speculate it was due to other metrics like ESM2’s likelihood that were not directly accounted for in the model and that it was optimizing for multiple other properties (like compactness and number of interface contacts).

Lennart Nickel & Martin Pacesa wanted to make their design task even harder by asking BindCraft to create beta-sheet binder, introducing a helicity-penalizing loss. You can see that they achieved this and a good binding affinity of 91.5 nM. There is yet another impressive achievement for BindCraft in this round: it claimed 6 out of the 7 leaderboard spots won by *de novo* binders, with the remainder being designed with RFdiffusion and ProteinMPNN.

At least for this competition, BindCraft and hallucination seem to have dethroned diffusion models due to their increased hit rates, one of the greatest meta change we’ve observed from Round 1 just a few months ago to today. We also think there is a huge avenue for customizing BindCraft and making the design task more challenging (e.g., losses with fewer interface contacts, more tunable physicochemical properties of the interface, targeting more difficult, hydrophilic sites, scaffolding, or including pre-trained models to tune both specificity and affinity, or making the overall process less computationally demanding). We’re incredibly excited to see how this nifty design tool will evolve in the future!

Now, onto some data analysis!

How many binders and expressed designs did we see?

We’ve talked before about how popular the diversified and optimized binder choices were for landing the top 100 spots. Most of these were submitted within a couple of days from the deadline. And yet, all designers choosing these were rightfully doing so: we have a 28% hit rate for diversified binders and 18% for optimized. We’ve already seen how successful taking an existing binder and redesigning its non-interface regions was: Cradle got a better binder than Cetuximab, Chris Xu had a few strongly binding ones, and Aurelia Bustos had an 80% hit-rate. The message is clear: if you want to create a better binder, it helps to have an existing one to start from!

The expression rate for those diversified binders is quite surprising: 92% and thus lower than hallucination and on par with other de novo methods. We think some people departed too far from the natural sequence for this, trying to maximize the scores. Yet, this is something worthy of a deeper investigation.

For de novo, hallucination (BindCraft) got 6 out of the 7 binders, with diffusion (RFdiffusion) claiming the final one. The overall success rate of hallucination (9%, 6/65 tested, as seen in the bar plot above) is on par, if not slightly lower, than reported for other targets in the BindCraft paper. All of them expressed, showcasing what a difference the SolubleMPNN redesign step makes. Along with filtering for expression using tools like NetSolP, these pipelines and the reliable high expression rate across all algorithm categories make us say claim that expression has basically been solved.

We also wanted to see whether the new metrics had an impact on the number of resulting binders, or if a random selection is still better. Ultimately, we see a stark difference: 24% of the top 100 designs were binders, whereas only 10% were from our selection maximizing model diversity. The causal relationship is still blurry. Did any of these metrics (ipTM, iPAE, ESM2 PLL) play a actual role in predicting binders, or were the top 100 approaches that mostly diversified natural binders like TGFα and EGF bound to bind already, and landing in the top 100 was just a side-effect?

We could assume the metrics helped with expression, given the higher rate in the top 100. But the question remains the same: is there a correlation between metric optimization and expression rates, or between the metrics and binding? We’ll investigate these 2 questions next.

Did our selection have any influence on the binding affinity?

There is almost no difference when looking at the distribution of binding affinities per design categories. The median values are similar and, when performing an ANOVA test on the mean values, there is no significant difference. At most, we see some outliers - Cradle and Chris’s binders outcompeting Cetuximab’s binding affinity.

Some differences arise when comparing the K_D per selection status (distribution shifts, slight differences in medians), yet there is still not statistical significant when comparing the mean values with an ANOVA test. It is interesting that the pseudo-random selection has a higher affinity range. There are some aspects we’ve skipped in this analysis, mainly the normality assumption. We can see how the Adaptyv selection category forms an almost bimodal distribution, with the designs in the 1e-6 M likely originating from EGF/TGFα, and in subsequent analyses we can account for both the design origin (or de novo) and selection status. But, we can highlight our main conclusion: landing in the top 100 w.r.t in silico metrics played no detectable role in increasing the binding affinity.

Evaluating the in silico scores: we’re still at the beginning

Let’s look more directly into the correlation (or lack thereof) between our chosen scores and binding affinity. In all cases, we find no significant correlation for any of the metrics. For most, the trend is quite peculiar: we would expect and increase in K_D as iPAE decreases, as ipTM increases, or as ESM PLL increases, yet all trends are the opposite. We also need more data points for a conclusive statement, as we are likely undersampling our binders.

Thus, we cannot ascertain if any of these scores are meaningful when modulating the continuous affinities. But, are they predictive of binding or expression as binary labels?

From the figure above, the definite answer is “No!”. When plotting the ROC (receiver operating characteristic) curves and calculating the area-under-the-curve (AUC), most values range in the 0.5 region (random classifier, meaning these metrics cannot discern between binding and non-binding or expressing and non-expressing designs better than a coin-flip). Some metrics are better predictors. Leo Castorina and other people from the Wells Wood lab showed that scores like %identity to PDB and TMscore are better predictors, with an AUC above 0.6, while glutamic acid or lysine composition can discriminate highly expressing from weak or non-expressing designs. They computed even more scores that could correlate with expression or binding using their DE-STRESS tool, and you can find all of them here. We thank the people from the Wells Wood lab for their active involvement in this competition and manuscript preparation!

It’s interesting that the normalized ESM pseudolikelihood (divided by the sequence length) is one of the better binding predictors, with an AUC of 0.72. As mentioned in our consortium discussions, it is likely due to the several binders being modified versions of existing ones, towards which ESM2 would be biased.

All tested designs formed 3 main classes: EGF or TGFα-like, antibody-like, and de novo/others

Previously, we noticed a homology-based clustering for the designs starting from EGF and TGFα, yet other clusters were not as distinguishable. For only the tested designs, new clusters for antibody-like and de novo-like designs appear. The antibody-like cluster contains all strong binders. We cannot discern a separation by binding strength or expression level in the ESM2 embedding space, although it seems the antibody-like cluster had more non-expressing designs. For a more in-depth analysis, looking at the percentages of binders or expressed designs per starting point is key!

Similarly, in the structure-informed SaProt space, we find the same clusters. This time, there is a higher overlap between EGF and TGF-like designs, and a more continuous distribution for the other designs or de novo. We still cannot discern specific expression or binding clusters, although it seems the EGF-like region is more biased towards medium expression designs. Thus, it’s still an art to adequately shape-up a known binder, optimize it, and ensure it will still express or bind.

Domains I and III of EGFR were the preferred targets

We advised you to select a specific EGFR epitope (residues 11-13, 15-18, 356, 440-441), at the interaction between Domains I and III of EGFR. We believe this is reflected in the targeted site distribution per domain, especially for all submissions and all tested ones. Domain III is consistently the most targeted region, accounting for about 46-62% of binding sites across all categories. Interestingly, this preference is most pronounced in de novo binders (62%), but less so for the top 3 strong binders (46%), even though Cetuximab interacts strongly with Domain III, while EGF binds to the aforementioned hot spots. Moreso, some binders preferred Domains II and IV as well, albeit at insignificant percentages.

This poses even more questions we are unable to address in our exploratory analysis. Why do de novo binders prefer Domain III? Which sites are exactly targeted by starting point or de novo? Is there a large deviation between original and modified binder sites in the diversified or optimized cases? We invite you to join our consortium if you have answers for these questions.

We illustrate the same patterns with these pseudo 3D EGFR plots (thanks to ColabDesign for its implementation), this time separating binders as de novo or from an existing starting point via a homology search using MMseqs2. We see the same trend: Domains I and III are the preferred targets.

Glutamic acid is more prevalent in de novo designs’ interfaces

We next wanted to see if the amino acid composition of binding sites differs between non-binders, de novo binders, and designs starting from known ones. The main difference is for glutamic acid (E), a hydrophilic amino acids, which has a higher frequency in de novo binders (14%) compared to the other 2 classes (around 10%). We assume this is likely due to the SolubleMPNN redesigning step most de novo methods used to improve solubility and expression, which biases proteins towards more hydrophilic surfaces (thus binding sites as well). Glutamic acid composition is also an effective predictor of expression, as we mentioned above.

There are other amino acid preferences for de novo binders, such as methionine (M) and phenylalanine (F), both hydrophobic amino acids, yet these occur at lower frequencies.

The interface properties are more indicative of the true binders

We next investigated the interface physicochemical properties of binders versus non-binders. The main differences between binders and non-binders are for the interface ΔG, the percentage of hydrogen bonds, and the interface size, all statistically significant following an ANOVA test. Unsurprisingly, keeping an existing binding site yields the lowest interface ΔG, whereas de novo is in close range to the non-binder ones. De novo proteins, in turn, have more interface contacts compared to non-binders. We have not found any significant differences for the interface hydrophobicity, nor the shape complementarity between binding site and epitope.

We can now see why both the diversified and optimized binder approaches were so popular: keeping an existing binding region with adequate predicted binding free energy, then simply scaffolding it in a backbone that optimizes the set of metrics we’ve chosen, seemed like the most straight-forward approach to both generate binders and land in the top 100.

The top 5 de novo binders and existing ones have some interface property overlap

Finally, we investigated if the top 5 de novo designs (as ranked by KD) differ significantly from the top 5 designs starting from existing binders. The most striking difference appears in their interface sizes - most of the top-ranking de novo designs aim to increase the number of interface contacts. We assume this is due to BindCraft’s (used in 6/7 de novo binders) loss function that accounts for this. Both de novo and existing starting point methods achieve high interface ΔG values. De novo designs show consistency in their interface properties, especially size and energetics. Meanwhile, the existing binder designs display more variability, likely due to their diverse evolutionary origins.

Interestingly, no single design optimizes all interface properties simultaneously. Some excel at shape complementarity but show average hydrogen bonding, while others achieve great energetics with moderate hydrophobicity scores. This is evidence of the fact that (except for filtering based on predicted ΔG or maximizing the number of contacts), there is no secret recipe for binder designs and several of these methods optimize different properties.

Conclusion: Has protein binder design been solved?

Not yet — but the rate of progress in protein engineering is astounding. Within just a few months from Round 1 (September) to Round 2 (December), we witnessed dramatic improvements in expression and binding success rates, along with a surge in participation from the protein design community.

Generating soluble proteins – those that express and fold well – seems to be essentially solved, at least in our small-scale bacterial cell-free expression systems. The next challenge here will be to figure out how to translate expression between different organisms to ensure high production yields in larger reactions so that all those novel proteins can be put to good use outside of lab-scale testing. Luckily, people are already thinking about generating datasets for predictive models of expression like e.g. at Align to Innovate.

Optimization of existing binders is proving remarkably effective and can systematically yield better proteins than previous candidates. Cradle's success in this competition demonstrates how this approach is now being rapidly industrialized and we hope many applications can benefit from such tools.

In de novo design, BindCraft has brought about a significant increase in success rates and has immediately been adopted by the protein design community. It remains to be seen whether this represents a broader shift toward hallucination models outperforming diffusion approaches, or if the next generation of diffusion models will shift the meta back.

Both in this competition and across the field, we still struggle to predict binding success through computational metrics. This gap between the ability to generate candidates and predict their performance represents one of the field's most pressing challenges and we need to generate more data to build better models here.

Thanks again to all participants for making this competition an immense success through sharing your cool designs and interesting approaches. We have exciting plans for 2025 and can't wait to see what kinds of proteins you'll design next.

If you want to participate in a paper write-up in the coming weeks about the two rounds of the Protein Design Competition, you can fill out this form to join.

If you want to stay in the loop about future competitions and releases from us, leave your email here!

About Adaptyv Adaptyv is the cloud lab for protein designers. Our platform is the fastest way for you to experimentally validate your protein designs. Just select an assay, upload your sequences and get high-quality experimental results in under 3 weeks. Check out our pricing here or just email us at hi@adaptyvbio.com to set up your data generation campaign!

The long awaited results are here -- and they don’t disappoint!

📈 We doubled the number of proteins we tested in our lab from 200 to 400!

🧬 Out of those 400 proteins, 378 expressed (95% expression rate!)

🚀 Out of those 378 expressed proteins, 53 did successfully bind the target protein EGFR (that’s a 14% success rate, more than 5x of the success rate of the previous round just 2 months ago!)

💪 The best binders reached single-digit nanomolar affinities which is within the range of the commercially sold Cetuximab antibody by Merck

😎 30 out of 130 protein designers managed to design at least one binding protein

Check out the full results

Want to join the discussion about the results?

Check out the thread on X or on LinkedIn

Here’s the breakdown:

The team at Cradle takes the top spot in terms of binding affinity by using their protein language models to introduce 10 mutations into the framework regions of the Cetuximab antibody.

This way, they generated an scFv with a KD of 1.2 nM which is lower than the original Cetuximab scFv. Congrats!

---

Tihumab aka Chris Xu aka Xushaoyong (https://lnkd.in/e8Mj8eq3) lands two of his nanobodies on rank 2 & 3 using his antibody humanization technique.

The pipeline consist of CDR grafting, iterative optimization using a deep learning model, antigen-antibody complex prediction and developability index prediction.

---

Aurelia Bustos MD, PhD's approach of modifying the TGFa protein with diffusion models to yield an optimized binder for EGFR was already super successful in our computational ranking, with 8 out of her 10 designs landing in the top 25.

This turned out to be also successful in the lab, with all 8 of those binders working!

---

Also in the top 10:
- Jeff Vogt with his scFv optimization pipeline
- Lennart Nickel using BindCraft to generate fully de novo binders
- Brian Naughton and Alan Blakely who both had the idea of using a linker approach to combine binders from previous rounds
- Shosuke Suzuki's de novo binder
- Yoichi Kurumida’s RFdiffusion scaffolded binder
- Young Su Ko who’s using Raygun+Protrek to filter his sequences

Congrats to everyone that ended up designing one or more of those binders and thanks to everyone that participated!

We loved reading about all those interesting design approaches and tried to give as many shoutouts as possible. The energy in the protein design community is truly amazing and we can’t wait to see how many more cool proteins you'll all be designing in the future!

Our API allows you to programmatically create experimental campaigns, get experiment updates and query lab results.

By abstracting away the complexity of the wet lab we’re enabling thousands of protein design teams to get easier access to experimental data to validate their protein designs.

From the smallest techbio to the largest pharma company - we’re making experimental data generation easier than ever before to allow you to engineer better proteins!

Check it out

How 310 AI is using the Adaptyv API

310 is building a new kind of protein design platform, powered by their AI models. They integrated the Adaptyv API to their Copilot product, which allows you to design proteins using natural language.

By integrating the Adaptyv API, Copilot now allows to do true end-to-end protein design, going from in silico predictions to validated molecules in one seamless workflow.

Try it out here!

With so many people enthusiastic about designing novel proteins and so many interesting submissions, we decided to double the number of proteins that we will test in the lab!

So instead of just 200 proteins we will experimentally characterize 400 novel proteins in our lab and open-source all the results.

Here’s how we selected them:

• Top 100 designs according to the computational selection metric

• For the other 300, we picked between 1 and 5 designs per protein designer, depending on the quality and novelty of their design process descriptions

That means that every participant that wrote at least some description of their design process will have at least one of their proteins validated in the lab! 🎉

Find out which designs got selected here: https://foundry.adaptyvbio.com/egfr_design_competition_2

Want to see the full leaderboard? -> https://design.adaptyvbio.com/

Big thanks to Twist for giving us a discount for the DNA synthesis to make this happen! 🧬

And big thanks again to Modal for sponsoring the compute for the competition 💻

—> Join the discussion on Twitter & LinkedIn

TL;DR

• We analyzed the submissions for the first EGFR competition (results here) for you!

• Most people explored RFdiffusion for backbone generation, ProteinMPNN for obtaining sequences, with some other interesting approaches we’ve highlighted.

• Winning combination: AlphaFold2 hallucination and SolubleMPNN inverse-folding.

• We’ve provided actionable recommendations on design algorithms based on our previous literature review and the learnings from our competition. You can find out which is the right method for you with a protein engineer’s “personality test”.

• We also include some domain knowledge-derived tips that should help participants with a purely computational background to avoid some pitfalls when it comes to the in-silico → in-vitro domain shift

In our last blog post, we looked at the different ways you can optimize a protein for a given task using machine learning models.

We made the distinction between fixed and sequential model optimization and further split the latter into model-based sampling and greedy/heuristic methods. Model-based methods use a learned…well…model to explicitly sample novel candidates. This model tries to balance between exploitation and exploration using the knowledge that it incorporates from observations via training. Greedy/heuristic methods often use the fitness model to rank randomly proposed candidates, but generally don’t adapt to the observations as much as model-based methods. 

In the first half of this post, we will use this taxonomy to take a look at the methods employed by designers in the first round of our EGFR competition. Different approaches come with different theoretical trade-offs, and we will see whether we can see those in the submissions’ practical performance.

In the second half, we will give you some actionable advice on how to choose an approach for the second round, summarized with two handy decision charts.

Overview of Adaptyv’s EGFR binder design competition

Competition overview and winners.

In our competition, creative protein engineers were tasked with designing a new binder to the extracellular domain of EGFR, a cancer-associated drug target. 

We selected 200(+1) of the most promising sequences for screening. The top 100 were based on the AlphaFold2 interface pAE as a proxy for binding (Bennett et al., 2023), whereas the rest were selected across a wide range of design techniques. AlphaFold2 iPAE as a binding metric has its problems, which we will briefly discuss, but we chose it because it has been exhaustively used in the communities, including luminaries such as the Baker lab (Bennett et al., 2023; Watson et al., 2023; Zhang et al., 2024). Most successful designs (in terms of leaderboard placement) aimed to directly optimize this objective. For a brief overview of protein design competitions, check out this Nature article we have been featured in!

Design and model categories

We had requested competitors to include a brief description of their submissions’ strategy. Some wrote very brief stubs (which we still thank them for!), and some wrote detailed descriptions and blog posts (you guys are awesome!).

The figure below shows the diversity of techniques designers used, categorized into certain themes by us. 

You can browse through all the categorized (tested) designs here. 

Number of designs submitted and selected in the first round of the EGFR competition by design and model category.

The most popular options included:

• sampling from a generative model like RFdiffusion (Watson et al., 2023), sometimes conditioned on EGFR’s structure, followed by inverse-folding with ProteinMPNN (Dauparas et al., 2022), then using AlphaFold’s predicted metrics or others for filtering (de novo + filter). We called this playing the “de novo slot machine” in our previous post. In a typical experimental setting, these approaches can be prone to relatively low hit rates, although recent models have made great leaps to address this (e.g., Google DeepMind’s AlphaProteo, Zambaldi et al., 2024, which has a 24.5% hit rate for interleukin-7 receptor-𝛼 binders). 

• diversifying known binders from the literature with ProteinMPNN (Dauparas et al., 2022) and subsequent selection after re-folding with AlphaFold and computing selection metrics (diversified binder + filter). This category includes partial diffusion of a known binder with RFdiffusion. In our last post’s taxonomy, this would be a fixed model optimization or local search strategy. Other designers chose a rational approach (Korendovych, 2018) for diversifying: starting from known sequences (such as EGF), they mutated regions (often not in the binding site!) using methods such as BLOSUM62 substitutions (Eddy, 2004). Diversified binder + filter was the most common design strategy.

• gradient-based input space optimization via AlphaFold2 (aka hallucination) (Anishchenko et al., 2021; Goverde et al., 2023; Frank et al., 2023). For this, a random sequence is fed through AlphaFold2, then a custom loss is computed (often including terms like pLDDT, pAE, and other constraints) and its gradient with regards to the input is taken by backpropagating through the folding network. Several iterations of input optimization can achieve a structurally stable design. Feeding a random sequence and its desired binding target through AlphaFold-Multimer enables the design of binders, with losses that can take into account the interaction pLDDT and pAE, the number of interface residues, etc. 

• directly optimizing a starting sequence using unsupervised protein languge model directed evolution (fixed model optimization), and building custom active learning loops (sequential model optimization with greedy/heuristic sampling). We included both in the optimized binder category. This is very similar to input space optimization which hallucination is also a part of, yet we decided it should have its separate category to highlight its “resurgence” in protein design, as we will see from our winners. 

• strategies that mainly employed molecular dynamics, docking, or Rosetta design protocols were labelled as physics-based. Most of these still filtered the final candidates using AlphaFold2’s iPAE to ensure they made it on the leaderboard. 

Submissions with missing or unclear descriptions are labelled as “not mentioned” and excluded from this analysis (398 out of 726 total submissions, with 65 of these being selected for validation). 

Looking more in-depth at the models represented, we found that RFdiffusion for backbone design followed by ProteinMPNN diversification and filtering is by far the most common strategy (176 total submissions).

Here RFdiffusion is used either for target-conditioned generation or partial diffusion/scaffolding of known binder motifs. This is a completely valid choice given the iPAE objective and lack of fully characterized EGFR binder datasets: people opted to sample a large number of potential designs and simply select the top candidates. But, as Brian Naughton reports, it is often hard to get both RFdiffusion and ProteinMPNN to reliably sample high-confidence designs. For validation, this would entail screening massive designed libraries and it might not be the best choice if you’re working with a limited experimental budget. 

Expression rates

Only 146 sequences (excluding the Cetuximab control) expressed, leaving some people surprised and some disappointed. This is the reality of translating in silico designs into experimentally-valid candidates. A 73% (146/201) expression rate is about on par with Wicky and colleagues, who achieved a 74% (71/96) expression rate for their AlphaFold2 hallucinated + ProteinMPNN-designed symmetric assemblies. Goverde and colleagues compared the designs obtained from the standard ProteinMPNN for an AlphaFold2-hallucinated backbone (with the AF2Seq pipeline) with biasing ProteinMPNN’s sampling towards hydrophilic amino acids, and with their fine-tuned version on soluble proteins (SolubleMPNN): for a single redesigned protein, normal ProteinMPNN has a 0% expression rate (0/12), the biased version 75% (6/8), and SolubleMPNN 93.1% (27/29). Recently, a paper from the Baker lab (Glögl et al., 2024) achieved a 98% (94/96) expression rate for TNFR1 binders.

With this context, 73% could be considered on the lower side compared to the Baker lab. However, only half of the participants we could verify via their socials had protein design experience (defined as PhD/postdoc/professor or working in a therapeutics biotech). This combined with the wide range of methods employed and the fact that not many people optimized for expression, only binding/iPAE, makes us at Adaptyv relatively happy with the outcome. 

For the second round, we are including an expression proxy in the metrics. We also recommend using SolubleMPNN as a final check before submitting your sequences!

Hit rates

Out of these, 5 were considered strong binders, with a KDKD​ between 3e-8 M to 2.3e-5 M, with 2 labelled as weak binders (KDKD​ above 1e-5 M). This yields a 2.5% total hit rate (5/201), considerably higher than past EGFR-targeting design campaigns (0.01% previously reported using a Rosetta protocol, Cao et al., 2022). We should emphasize that hit rates are highly dependent on the design target, with highly accessible, hydrophobic (Cao et al., 2022; Pacesa et al., 2024), and reduced flexibility epitope regions (Kim, Choi & Kim, 2021) often yielding better results. To ensure you get a lot of binders, we recommend you target hydrophobic surface regions or the EGFR epitope region we provided, especially if you want to succeed in the EGF neutralization assay we are doing!

iPAE as a proxy for KDKD​ 

Seven binders (3 true binders, with 2 disqualified due to similarity to other therapeutics, and 2 weak) are too few data points to test the quality of iPAE as a competition proxy. Brian Naughton recently looked at 55 entries from PDBBind in his blog post, showing a Pearson’s correlation coefficient of -0.25 (p-value << 0.05). iPAE might not be an ideal score for binding affinity.

We recently asked for input from the protein design community regarding which metric should be used as a binding affinity proxy. This sparked quite an interesting conversation:

• Nikhil Haas conducted an extensive analysis of the competition sequences and other data ordered from Adaptyv, looking into expression and binding correlations. He showed that ESM2 log-likelihoods correlate well with expression. You can watch his video explanation here.

• Martin Pacesa says ipTM could be used for binary binding predictions.

• Sergey Ovchinnikov recommends using no filters for the ranking process or, at best, looking at iPAE from AlphaFold2’s predictions with initial guess and also selecting designs that did not pass the filters. 

We hope to see more filtering and optimization proxies suggested and experimented with in the second round!

The champion binders

First place 🥇- Bindcraft: Martin Pacesa and Lennart Nickel

Our undisputed winners are Martin Pacesa and Lennart Nickel, with a 4.91e-7 M KDKD​. This was achieved with their custom AlphaFold2 hallucination pipeline called BindCraft (Pacesa et al., 2024). As opposed to others, it hallucinates the binding interface instead of the binder’s structure only. It uses four stages of optimization: first gradient-based optimization on the continuous sequence logits, then on the softmax matrix, followed by one-hot encoding without and with randomly sampled mutations. It is also highly customizable via the combined loss function: some terms aim to optimize the binding interface’s prediction confidence, the number of residues, as well as a radius of gyration for the binder (to prevent “spaghetti”-like long binding stretches) and a “helicity” loss (as AlphaFold2 hallucination is biased towards helices, this promotes more non-helical designs). 

You can try it out here using Google Colab. As the authors pointed out, it is super straightforward to adapt it to your own design campaign and implement additional loss terms (see here!). For the second round, give BindCraft a try!

Second place 🥈 - Khondamir Rustamov

The second spot was taken by a binder designed using the AF2Seq protocol for backbone hallucination and SolubleMPNN for inverse-folding, as initially explored by Goverde and colleagues (see the section on expression rates). It had a KDKD​ of 4.77e-6 M and ranked 54th for iPAE. We can now see a common theme: hallucinated backbones with AlphaFold2, followed by SolubleMPNN inverse-folding!

The methods we have seens so far indirectly optimized multiple objectives and their designs are likely on the Pareto-optimal front between experimental binding, iPAE, and expression.

Third place 🥉- Adrian Tripp and Sigrid Kaltenbrunner

This binder had a KDKD​ of 2.29e-5 M, ranked 89th in the iPAE leaderboard, and had a high expression rate. Sigrid and Adrian used the ProtFlow pipeline (not to be confused with Adaptyv Bio’s ProteinFlow - which you should check out for processing protein structures!) to orchestrate their complex workflow. First, RFdiffusion tackled the binder backbone generation, targeting EGFR’s residues 18, 39, 41, 108 and 131, followed by initial filtering, then LigandMPNN inverse-folding on Rosetta-relaxed structures, and sequential folding and filtering with ESMFold and ColabFold. The ESMFold step accounted only for the binder structures and selected for pLDDT+TMscore, whereas the ColabFold step predicted the entire complex and filtered based on pLDDT, TMscore, iPAE, ipTM, and the number of hotspot contacts. The resulting designs were once again fed through the entire pipeline, for 3 cycles in total. We can see how they took the de novo design + filter approach to its limit!

Special mentions

Some approaches were not as successful, but were extensively documented and so deserve being highlighted as well.

In his blog post, Brian Naughton tried out a couple of methods, including ESM2 directed evolution with an iPAE oracle, RFdiffusion, ProteinMPNN, and Bayesian optimization. He provides the Modal commands for several of these tools here and you should definitely check this out. It is a great way to get started in protein design. And you can make use of your $30 free Modal credits as well!!

Anthony Gitter found that a de novo, language-instructed binder from ProTrek (Su et al., 2024) performed quite well, despite not ranking in the top 100. He tested the “non-biological/domain unadapted” language model Llama 3.1’s ability to design proteins, which still suggested antibody-like EGFR inhibitors. We are pretty hopeful for what language-instructed, chat-based protein design might look like in the future!

Alex Naka’s strategy was the only one that fits the definition of sequentially-optimized, model-based sampling we established in the last blog post. For his in silico oracle, he first opted for a simple AlphaFold2 implementation using Modal with PepMLM (Chen et al., 2023) as his EGFR-conditioned sequence generator. He trained a surrogate (an ensemble of 1D CNNs on one-hot encodings) on a starting dataset mined from this in-silico oracle, then continued training it during optimization. He used EvoProtGrad (Emami et al., 2023) as his surrogate-conditioned generator, iterating between scoring new candidates and then retraining. A complete active learning loop! Now we know why all his designs were at the top of the leaderboard (the ”custom active learning” category is entirely comprised of these). This is an interesting strategy you could use for the second competition round, but make sure you also co-optimize for expressibility!

The top 5 submitted designs ranked by AlphaFold2’s iPAE in the first design competition round.

We thank Brian Naughton, Anthony Gitter, and Alex Naka for their contributions, both to the competition, and especially for these highly detailed writeups. Make sure you read their posts! 

We thank all other design competition participants as well - it was a great experience seeing so many creative protein engineers’ solutions and, even beyond round two, we plan to organize more competitions in the future!

Recommendations for ML based protein optimization 

Main considerations for a protein engineering campaign. For simplicity, we grouped the starting data, lab budget, and risk-reward trade-off as business goals - willingness to spend on your campaign’s lab validation. ML power includes computational resources and ML experience - available expertise and resources to use ML models.

A) Questions to ask yourself before an optimization campaign

When selecting a model for your protein optimization project, ask yourself the following questions:

1. Starting data: How much labeled data do you have available to start with (low-N vs. large-N regime)? Do you plan to obtain more data given your model’s suggestions?

2. Lab budget: What is your total budget for screening? How many variants do you think you can reliably test out (consider replicates as well!)? Does your lab offer the necessary screening platforms and machines?  

3. Risk-reward trade-off: Are you trying to be more conservative and incrementally improve an existing lead? Or are you trying to discover the next moonshot therapeutic?

4. Computational resources: How many computational resources (GPUs, available time to implement or train) do you have? 

5. ML experience: What is your ML experience? Are you able to implement novel architectures, simply fine-tune a model, or are you familiar with tools like ColabFold, 310.ai, tamarind.bio?

B) The protein engineer’s personality test

The answers to these questions determine in which of a couple of profiles a protein engineer might fall into, and which ML models likely to achieve their aims during a binder optimization campaign. 

Decision flow chart for your next optimization campaign.

If you only have a single known binder or no starting data, you might be forced to de novo design using ProTrek or RFdiffusion, or mine some data from the literature and maybe replace some residues in the binding site using a BLOSUM62 scheme. If you have a small starting set, you might try low-N fine-tuning techniques and carefully plan to test what your model will suggest. If you’re lucky or proficient enough in the lab, you could have a large dataset, mapping both single point and combinatorial mutations to their binding affinities. In this case, custom architectures and from-scratch training become feasible for you. 

Next, when it comes to the lab resources for validation, most at-home protein engineers have a small to non-existent screening budget: they would validate at most their final best binder. This was the case for our EGFR competition, where completely in silico oracles are necessary. Alex Naka really leaned into this constraint, building a small dataset of low iPAE binders, training CNN ensembles for prediction, and finally creating a completely in silico active learning loop. 

If you can afford to validate about 200 designs, we recommend sequential optimization using either explicit model-based sampling or greedy/heuristics. We prefer Bayesian optimization/Active learning with simple, uncertainty-aware surrogates (Gaussian Processes or ensembles), as argued for in our first blog post! The ALDE (Yang et al., 2024) or EVOLVEpro (Jiang et al., 2024) frameworks are great starting points. If you want to do some input optimization via AlphaFold2 backpropagation, take a look at BindCraft (Pacesa et al., 2024) - use the default settings or even implement your own design campaign-tailored loss function!

If you are fortunate enough, your validation budget could be large to limitless: your best bet would be to take the largest protein foundation model you can find and sequentially fine-tuning it on new batches of data - you can retrain any fixed model. Be as greedy as you want: generate combinatorial libraries (Wu et al., 2019) and select the top-performing variants at each step!

Another axis of consideration is the risk-reward trade-off you are willing to make. Running more iterations of active learning, exploring more of the sequence space, or validating a larger batch at each step means higher investment, but also a higher chance of discovering moonshot binders, if you think those should exist. The alternative is to simply take an existing binder and do some local search that just gets you away from patent protections while improving enough to be worth it. These areas of consideration (starting sequences, lab resources, risk-reward trade-off) ultimately depend on your business goals.

Compute is rarely, if ever the bottleneck in protein design. Most protein engineers can make do with a single GPU on their local machine. Some use Modal or AWS, with a low resource consumption. With very limited resources it’s even possible to set up an active learning loop (as seen here) and only run it for a couple hours or days. We recommend using simple surrogates (1D CNN ensemble or Gaussian Processes, specifically). Simple one-hot encodings can also suffice, In fact, they perform about as well as embeddings from a protein language model like ESM2 for fitness prediction and optimization (Shanehsazzadeh et al., 2020; Greenman, Amini & Yang, 2023; Yang et al., 2024). 

Brief detour: cloud-GPUs

In the past, the way to do "budget" ML was to buy a used gaming PC and run stuff locally (getting you started on a few hundred USD with high FLOP/USD ROI). This is still the highest ROI if you will be running experiments 24/7 and don't pay for electricity (shoutout to Tim Dettmers), yet the hype around chatGPT created a very competitive cloud computing environment. 

We therefore recommend that you use Modal, Google Colab, or other cloud-GPU platforms to get started. 

Most of them now offer generous free tiers (in the case of Modal, 30 USD recurrent per month) and it means you can get started for as low as 0 USD (with credits) and for about 5 USD per designed protein, following Brian’s approach. 

Click here for a cost comparison between different cloud-computing platforms. And, lucky for you, we have partnered with Modal to offer free credits worth 500 USD on a first come first serve basis: you can get them from here. Make sure you mention which computing platform you used when you submit your binders. Good luck designing!!

That being said, more is in fact better. If you have a high computational resource budget - you can fine-tune any large language model (Li et al., 2023) or diffusion models (Bennett et al., 2024). While these larger models will also require more data, your limitless computational budget means you could collect data from in silico oracles as a proxy, or augment your sequence features for model training. For example, you could train on energy scores from Rosetta, perform molecular dynamics simulations, observe multiple blind-docking instances, run AlphaFold2 on your variants and extract structural features to train your surrogates on, and many more options.

Finally, you should consider your ML experience. This, combined with compute resources, describes your ML power. Lucky for you, you live in the best times to do computational protein design because the barrier of entry has been lowered so much.

Use platforms like 310, BioLM, Lab.Bio, Tamarind, Latch, or many more if you lack any ML experience. For example, 310’s Copilot lets you chat with your personal protein design assistant: you can ask it to find functionally annotated proteins, fold them, diversify them with ProteinMPNN, compare their structures, then export. 

Then, for someone with an intermediate level of experience in ML, there is the option of fine-tuning an existing language model from HuggingFace and performing in silico directed evolution. Implement your own architectures if you’re an experienced ML practitioner, or loss functions, optimization objectives, try out a bunch of acquisition functions, participate in design competitions, publish on arXiv, write Substack articles and post on X to get involved in the scene! You know what to do!

Platforms for protein design and optimization

We hope to see you in the second round of the EGFR competition!

References

[1] Anishchenko, I., Pellock, S.J., Chidyausiku, T.M., Ramelot, T.A., Ovchinnikov, S., Hao, J., Bafna, K., Norn, C., Kang, A., Bera, A.K., DiMaio, F., Carter, L., Chow, C.M., Montelione, G.T. & Baker, D. (2021) De novo protein design by deep network hallucination. Nature 2021 600:7889. 600 (7889), 547–552. doi:10.1038/s41586-021-04184-w.

[2] Bennett, N.R., Coventry, B., Goreshnik, I., Huang, B., Allen, A., Vafeados, D., Peng, Y.P., Dauparas, J., Baek, M., Stewart, L., DiMaio, F., De Munck, S., Savvides, S.N. & Baker, D. (2023) Improving de novo protein binder design with deep learning. Nature Communications 2023 14:1. 14 (1), 1–9. doi:10.1038/s41467-023-38328-5.

[3] Bennett, N.R., Watson, J.L., Ragotte, R.J., Borst, A.J., See, D.L., et al. (2024) Atomically accurate de novo design of single-domain antibodies. bioRxiv : the preprint server for biology. doi:10.1101/2024.03.14.585103.

[4] Cao, L., Coventry, B., Goreshnik, I., Huang, B., Sheffler, W., et al. (2022) Design of protein-binding proteins from the target structure alone. Nature 2022 605:7910. 605 (7910), 551–560. doi:10.1038/s41586-022-04654-9.

[5] Chen, T., Dumas, M., Watson, R., Vincoff, S., Peng, C., et al. (2023) PepMLM: Target Sequence-Conditioned Generation of Therapeutic Peptide Binders via Span Masked Language Modeling. https://arxiv.org/abs/2310.03842v3.

[6] Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R.J., et al. (2022) Robust deep learning–based protein sequence design using ProteinMPNN. Science. 378 (6615), 49–56. doi:10.1126/SCIENCE.ADD2187/SUPPL_FILE/SCIENCE.ADD2187_SM.PDF.

[7] Eddy, S.R. (2004) Where did the BLOSUM62 alignment score matrix come from? Nature Biotechnology 2004 22:8. 22 (8), 1035–1036. doi:10.1038/nbt0804-1035.

[8] Emami, P., Perreault, A., Law, J., Biagioni, D. & St. John, P. (2023) Plug & play directed evolution of proteins with gradient-based discrete MCMC. Machine Learning: Science and Technology. 4 (2), 025014. doi:10.1088/2632-2153/ACCACD.

[9] Frank, C., Khoshouei, A., Stigter, Y. de, Schiewitz, D., Feng, S., Ovchinnikov, S. & Dietz, H. (2023) Efficient and scalable de novo protein design using a relaxed sequence space. bioRxiv. 2023.02.24.529906. doi:10.1101/2023.02.24.529906.

[10] Glögl, M., Krishnakumar, A., Ragotte, R.J., Goreshnik, I., Coventry, B., Bera, A.K., Kang, A., Joyce, E., Ahn, G., Huang, B., Yang, W., Chen, W., Sanchez, M.G., Koepnick, B. & Baker, D. (2024) Target-conditioned diffusion generates potent TNFR superfamily antagonists and agonists. bioRxiv. 2024.09.13.612773. doi:10.1101/2024.09.13.612773.

[11] Goverde, C.A., Pacesa, M., Goldbach, N., Dornfeld, L.J., Balbi, P.E.M., Georgeon, S., Rosset, S., Kapoor, S., Choudhury, J., Dauparas, J., Schellhaas, C., Kozlov, S., Baker, D., Ovchinnikov, S., Vecchio, A.J. & Correia, B.E. (2024) Computational design of soluble and functional membrane protein analogues. Nature 2024 631:8020. 631 (8020), 449–458. doi:10.1038/s41586-024-07601-y.

[12] Goverde, C.A., Wolf, B., Khakzad, H., Rosset, S. & Correia, B.E. (2023) De novo protein design by inversion of the AlphaFold structure prediction network. Protein Science : A Publication of the Protein Society. 32 (6). doi:10.1002/PRO.4653.

[13] Greenman, K.P., Amini, A.P. & Yang, K.K. (2023) Benchmarking Uncertainty Quantification for Protein Engineering. bioRxiv. 2023.04.17.536962. doi:10.1101/2023.04.17.536962.

[14] Jiang, K., Yan, Z., Bernardo, M. Di, Sgrizzi, S.R., Villiger, L., Kayabolen, A., Kim, B., Carscadden, J.K., Hiraizumi, M., Nishimasu, H., Gootenberg, J.S. & Abudayyeh, O.O. (2024) Rapid protein evolution by few-shot learning with a protein language model. bioRxiv. 2024.07.17.604015. doi:10.1101/2024.07.17.604015.

[15] Kim, D.G., Choi, Y. & Kim, H.S. (2021) Epitopes of Protein Binders Are Related to the Structural Flexibility of a Target Protein Surface. Journal of Chemical Information and Modeling. 61 (4), 2099–2107. doi:10.1021/ACS.JCIM.0C01397/SUPPL_FILE/CI0C01397_SI_002.ZIP.

[16] Korendovych, I. V. (2018) Rational and Semirational Protein Design. Methods in molecular biology (Clifton, N.J.). 1685, 15. doi:10.1007/978-1-4939-7366-8_2.

[17] Li, L., Gupta, E., Spaeth, J., Shing, L., Jaimes, R., Engelhart, E., Lopez, R., Caceres, R.S., Bepler, T. & Walsh, M.E. (2023) Machine learning optimization of candidate antibody yields highly diverse sub-nanomolar affinity antibody libraries. Nature Communications 2023 14:1. 14 (1), 1–12. doi:10.1038/s41467-023-39022-2.

[18] Pacesa, M., Nickel, L., Schmidt, J., Pyatova, E., Schellhaas, C., et al. (2024) BindCraft: one-shot design of functional protein binders. bioRxiv. 2024.09.30.615802. doi:10.1101/2024.09.30.615802.

[19] Shanehsazzadeh, A., Belanger, D., Research, G. & Dohan, D. (2020) Is Transfer Learning Necessary for Protein Landscape Prediction? https://arxiv.org/abs/2011.03443v1.

[20] Su, J., Zhou, X., Zhang, X. & Yuan, F. (2024) ProTrek: Navigating the Protein Universe through Tri-Modal Contrastive Learning. bioRxiv. 2024.05.30.596740. doi:10.1101/2024.05.30.596740.

[21] Watson, J.L., Juergens, D., Bennett, N.R., Trippe, B.L., Yim, J., et al. (2023) De novo design of protein structure and function with RFdiffusion. Nature 2023 620:7976. 620 (7976), 1089–1100. doi:10.1038/s41586-023-06415-8.

[22] Wicky, B.I.M., Milles, L.F., Courbet, A., Ragotte, R.J., Dauparas, J., Kinfu, E., Tipps, S., Kibler, R.D., Baek, M., DiMaio, F., Li, X., Carter, L., Kang, A., Nguyen, H., Bera, A.K. & Baker, D. (2022) Hallucinating symmetric protein assemblies. Science. 378 (6615), 2024. doi:10.1126/SCIENCE.ADD1964/SUPPL_FILE/SCIENCE.ADD1964_SM.PDF.

[23] Wu, Z., Jennifer Kan, S.B., Lewis, R.D., Wittmann, B.J. & Arnold, F.H. (2019) Machine learning-assisted directed protein evolution with combinatorial libraries. Proceedings of the National Academy of Sciences of the United States of America. 116 (18), 8852–8858. doi:10.1073/PNAS.1901979116/SUPPL_FILE/PNAS.1901979116.SAPP.PDF.

[24] Xue, L.C., Rodrigues, J.P., Kastritis, P.L., Bonvin, A.M. & Vangone, A. (2016) PRODIGY: a web server for predicting the binding affinity of protein–protein complexes. Bioinformatics. 32 (23), 3676–3678. doi:10.1093/BIOINFORMATICS/BTW514.

[25] Yang, J., Lal, R.G., Bowden, J.C., Astudillo, R., Hameedi, M.A., Kaur, S., Hill, M., Yue, Y. & Arnold, F.H. (2024) Active Learning-Assisted Directed Evolution. bioRxiv. 2024.07.27.605457. doi:10.1101/2024.07.27.605457.

[26] Zambaldi, V., La, D., Chu, A.E., Patani, H., Danson, A.E., et al. (2024) De novo design of high-affinity protein binders with AlphaProteo. https://arxiv.org/abs/2409.08022v1.

[27] Zhang, J.Z., Li, X., Liu, C., Jiang, H., Wu, K. & Baker, D. (2024) De novo design of Ras isoform selective binders. bioRxiv. 2024.08.29.610300. doi:10.1101/2024.08.29.610300.