🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI
Evan Feinberg and Sergey Edunov from Genesis Molecular AI discuss how diffusion models have become the superior primitive for protein-ligand structure prediction compared to GANs, enabling sub-angstrom accuracy predictions that are finally useful for drug discovery. They explain their focus on small molecule drug discovery, the importance of synthetic training data from physics simulations, and their shift toward agentic systems while maintaining human-in-the-loop validation.
Summary
The conversation explores the evolution of AI in drug discovery over the past decade, with particular focus on how diffusion models have replaced GANs as the generative primitive of choice. Evan Feinberg, founder and CEO of Genesis Molecular AI, and Sergey Edunov, former LLAMA 2/3 pre-training lead at Meta, discuss their company's work on PEARL, a structure prediction model for protein-ligand complexes.
A central theme is the importance of achieving sub-angstrom accuracy in 3D pose prediction. The speakers explain that two-angstrom accuracy—previously considered acceptable—is insufficient because aromatic rings can be flipped and other critical details can be wrong without being detected. Sub-angstrom accuracy is essential because hydrogen bonds require precise geometric relationships (2.7-3.3 angstroms between donor and acceptor), and mistakes compound downstream in potency and ADMET predictions.
Genesis's approach differs from competitors through several key innovations: (1) focus on small molecules rather than protein-protein interactions, which presents unique challenges due to the vast search space of 10^60 drug-like molecules; (2) extensive use of synthetic training data generated through physics-based molecular dynamics simulations rather than relying solely on the ~200,000 crystal structures in the PDB; (3) incorporation of physical priors throughout the model architecture and inference process; and (4) inference-time scaling using diffusion steps steered by physics-based guidance, analogous to how LLMs use thinking tokens.
The speakers discuss their partnership model with companies like Exscientia, where they collaborate on both first-in-class targets (discovering entirely new chemical matter) and best-in-class improvements to existing drugs. They emphasize that while AlphaFold3 and similar models have made important contributions, protein structure prediction alone doesn't solve drug discovery—ADMET properties, dynamics, conformational sampling, and potency prediction remain critical unsolved problems.
A significant portion of the discussion addresses the broader drug discovery pipeline and where AI provides the highest leverage. The speakers argue that the discovery and design phase—finding molecules that bind selectively to disease targets with appropriate pharmacokinetics and safety profiles—is where AI can have the greatest impact, rather than in target identification, clinical trials, or regulatory approval phases.
The company recently announced a name change from Genesis Therapeutics to Genesis Molecular AI, reflecting their identity as an AI-first company while maintaining proprietary drug discovery programs to demonstrate real-world utility of their models. They work with pharma partners while conducting their own research to ensure models are battle-tested in practice.
Regarding future directions, the speakers discuss their development of agentic systems powered by large language models that can orchestrate multiple tools and make decisions about molecular design without requiring deep expertise from individual scientists. They position this as an evolution similar to how coding assistants have progressed from autocomplete to autonomous code generation, emphasizing that agents require underlying models of sufficient quality to avoid amplifying errors.
The transcript also covers practical challenges in drug discovery that pure ML approaches must address: the inability of highly automated synthesis platforms to efficiently explore novel chemical space, the large number of anti-correlated properties that must be optimized simultaneously, and the high false positive rate of high-throughput screens compared to careful resynthesis and validation.
Finally, the speakers identify GPU availability as the primary bottleneck limiting their research velocity and discuss hiring priorities for researchers interested in working on novel architectures in the biology space as opposed to incremental transformer improvements in LLMs.
About this episode
<p>This episode has a fun personal twist: There’s a counterfactual world where I was employee #1 at <a href="https://www.genesis.ml/" target="_blank">Genesis Molecular AI</a>, the company behind today’s episode. A certain introduction happened a few weeks too late and I had already happily signed at Atomwise, another ML-for-drug-discovery startup. Same problem, different company. I was certain ML was going to transform small molecule drug discovery. Early results were underwhelming. Useful at times, but nowhere near revolutionary. In the last year I’ve seen signs that ML is finally ready to deliver on my convictions from a decade ago. Genesis is one of the places that might have finally cracked this problem. I was super excited to come full circle and catch up with co-founder <a href="https://www.linkedin.com/in/evanfeinberg" target="_blank">Evan Feinberg</a> and CTO <a href="https://www.linkedin.com/in/edunov/" target="_blank">Sergey Edunov</a>.</p><p>If you are at all interested in small molecule drug discovery, we think you will find this fascinating!</p><p>In our nearly two hour chat we cover:</p><p>* What is small molecule drug discovery, and why is it hard</p><p>* Structure prediction as a hotbed of innovation in AI algorithms</p><p>* How advances in AI elsewhere have enabled stepwise improvements in predictive power</p><p>* How the community benchmarks are essentially calling AI slop good enough</p><p>* The Genesis flagship model (PEARL) can routinely hit a threshold that is necessary for real-world applications</p><p>* New agentic workflows enabled by these highly accurate models</p><p>Read on for more, and also some personal thoughts on the future at the end.</p><p>The coolest diffusion research is happening at Genesis</p><p>Sergey Edunov came to Genesis from Meta where he led Llama 2 training and Llama 3 pretraining. Sergey was a former physicist who thought he was done with physics after many years of training LLMs. Then, he discovered Genesis, and was blown away with all the novel architecture work they’ve been developing.</p><p>It probably surprises no one that modern LLM research has not resulted in fundamentally novel or exciting updates in architectures since almost the advent of the transformer — the entire field is using variants on the same idea that came out in the original “Attention is all you need” paper. Sure, some were quite useful (mixture-of-experts in particular allowed for the massive model paradigm we’re at today), but there was very little conceptually exciting.</p><p>“We sort of had to wait for the right primitive to get created, and that turned out to be diffusion… Actually, some of the most innovative diffusion research that’s happening in our field is happening in 3D structure prediction right now.” — Evan Feinberg</p><p>The field of 3D structure prediction on the other hand has been a hotbed of research. Genesis’ recent model <a href="https://www.genesis.ml/news/introducing-pearl" target="_blank">PEARL</a> (Place Every Atom at the Right Location) is able to understand protein flexibility, and model not just where the ligand goes, but also make small adjustments of the protein so that the two fit better than either alone. The field knew this was missing for a long time, but it was really hard to model until now.</p><p>Agentic Discovery</p><p>What makes this problem so hard? As Sergey points out, there are 10^60 possible drug-like small molecules. You’ll never be able to search them all, and trying to find the good ones is something like finding a needle in a haystack — except everything except your needle is dangerous.</p><p>“There are 10 to the 60 drug-like small molecules in the universe… it’s like finding a needle in a haystack, where everything except your needle is very, very dangerous.” — Sergey Edunov</p><p>“Or finding hay in a needle stack might be a more apt analogy.” — Evan Feinberg</p><p>Trying to solve the multi-parameter optimization problem is even worse. What makes a strong binder and a molecule with good “ADMET Properties” are oftentimes at tension with each other. For example, a good binder is likely greasy, but a greasy molecule is likely insoluble so it won’t enter the bloodstream and get to where it needs to go!</p><p>Genesis’ advances in generative AI have now pushed them beyond the threshold where they believe agentic drug discovery loops are finally possible. We all remember the early days of LLMs. They were great chatbots but terrible agents, as small errors compounded rapidly into uselessness. As LLMs got better, the usefulness of agents rapidly improved. Evan and Sergey argue that their models at Genesis recently passed a similar threshold. Their internal agentic drug-discovery system (code named SAPPHIRE) can now iterate like a chemist: look at and reason about poses, form hypotheses, read literature, use internal tools, create candidates for the next iteration. Combining this with automated lab partnerships like the one Genesis has with <a href="https://incyte.com/" target="_blank">Incyte</a>, we’re rapidly approaching a time of drug discovery agents running 24/7 making/testing new molecules. Exciting times!</p><p>Benchmark crisis: Everyone’s favorite benchmark is slop</p><p>One surprising point that isn’t talked enough about: the academic field of “co-folding” has settled on a benchmark value of “2 Angstrom RMSD” as a metric for a “good pose”. Evan does not mince words: this threshold is just bad. Perhaps even deceptively bad. For many strong binders, there’s a very clear pose, one that you can even directly resolve in the PDB electron density! And yet, with a 2Å RMSD threshold, you can get the pose quite wrong in ways that might even mislead a medicinal chemist. For example, flip around an aromatic ring, and everything looks reasonable, but you’re no longer modeling the right interactions.</p><p>Evan makes the strong claim that 1Å RMSD is really the threshold necessary to ensure the core of the molecule is sitting where it needs to be, and models all interactions.</p><p>“If your model is sitting at 1.8, 1.9 Angstrom RMSD, that’s slop, most likely.” — Evan Feinberg</p><p>As a simple example, he points out hydrogen bonds which are responsible for many of the most important interactions in protein-ligand systems. Hydrogen bonds only have a 0.6Å range to be valid! Clearly if you’re accurately resolving all H-bonds, you generally have to be doing much better than the 2Å threshold.</p><p>This is clearly a hard-fought lesson for Evan and Genesis. In their opinion, the community is stuck on these benchmarks because academics developing methods were not users. Evan does see signs of life, with the use of new metrics such as lDDT for co-folding. Hopefully soon the community can agree that “1.8Å RMSD is slop”, and start hill climbing on this much harder task.</p><p>For a more thorough exploration of the weaknesses in conventional benchmarks, see the <a href="https://arxiv.org/abs/2510.24670" target="_blank">PEARL technical report</a>.</p><p>PEARL tops OpenBind</p><p>Which makes what happened next all the more striking. Near the end of the podcast, we talked about a recent “proof-is-in-the-pudding” moment for Genesis — evaluating their <a href="https://www.genesis.ml/news/zero-shot-pearl-system-surpasses-all-cofolding-models-on-openbind" target="_blank">PEARL model</a> on a recently released OpenBind benchmark. This benchmark featured 802 never before seen co-complexes on a target protein EV-A71. This target seems almost custom-chosen to give most classical docking methods a problem. When a ligand binds to the main binding site, the protein moves around to close off the path the ligand used to enter the binding pocket. This process, known as “induced fit” is notoriously hard for traditional methods to model. The tradeoff is easy to understand: treating the protein as a static structure, it becomes difficult to place a ligand in a binding pocket. Treat the protein as dynamic, and now you have to simulate complicated processes that take a long time to resolve.</p><p>PEARL was able to model the induced fit of the ligand without running long MD simulations. Across the different evaluation metrics, PEARL came out not just ahead, but oftentimes well ahead of any public model. A truly impressive result.</p><p>“Where PEARL was exceptionally good is figuring out how to move this loop. We are basically correct for every single pose.” — Sergey Edunov</p><p>Even more exciting, this was done without any fine-tuning, or using any data on the target or homologous targets — the template PDB was released after PEARL’s training cutoff.</p><p>Where does co-folding go now?</p><p>As someone who has followed or participated in ML techniques for protein-ligand interactions for almost a decade, I was genuinely impressed with the results that Genesis has released recently. This has been many years in development, and I’m sure Evan and the team had many sleepless nights trying to get to this point. I also think other teams are making similar progress — both Isomorphic and Deep Origin have released results that seem spiritually similar and combine computation, wetlab data, ML, to achieve genuine predictive power that seemed impossible a decade ago. Sadly, all of the above are closed source so there’s no way to honestly compare them. Looking at the results I think there might be a time in the not so distant future where we can consider protein-ligand binding “solved”.</p><p>I sincerely hope that the academic community can take inspiration from these developments. Once you know something can be done, it’s much easier to execute. Still, I believe that the key enabler in all of the above was the tight integration of ML, large-scale computation, and real-world drug discovery applications. Sadly academia is just not structured in a way that makes such a development easy.</p><p>With those parting thoughts, we hope you give the podcast a listen!</p> <br /><br />This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit <a href="https://www.latent.space/subscribe?utm_medium=podcast&utm_campaign=CTA_2">www.latent.space/subscribe</a>
Key Insights
- Diffusion models proved to be a superior generative primitive compared to GANs for protein-ligand systems, enabling the creation of accurate 3D structure predictions where GANs struggled with mode collapse and other training instabilities.
- Sub-angstrom accuracy (less than 1 angstrom RMSD) is necessary for useful drug discovery predictions because hydrogen bonds require precise geometric positioning within a 0.6 angstrom window (2.7-3.3 angstroms), and even small errors like aromatic ring flips create invalid predictions that appear correct to downstream analyses.
- Traditional benchmarks using 2-angstrom RMSD accuracy originated from academic docking studies focused on publication rather than drug discovery, and the field has been shifting toward more stringent metrics like LDDT and PoseBuster validity checks.
- Genesis creates synthetic training data through physics-based molecular dynamics simulations of small molecules because the public PDB contains only ~200,000 crystal structures expanding at a glacial pace, whereas small molecules can be modeled with physics at much lower computational cost than large proteins.
- The search space for small molecules is vastly larger than protein-protein interactions (10^60 drug-like molecules) making small molecule prediction harder despite appearances, because the model must distinguish a binding molecule from overwhelming noise of non-binding conformations.
- Inference-time scaling in their models uses diffusion steps guided by physics-based verification that iteratively refines predicted structures, analogous to how LLMs use thinking tokens, rather than providing a single pass prediction.
- The highest leverage AI application in drug discovery is the molecular design phase (finding selective binders with good pharmacokinetics), not target identification or clinical trial design, because other phases require different expertise and approaches.
- Agentic systems for drug discovery require underlying model quality to be sufficiently good that they don't amplify errors into useless predictions, similar to how early LLM coding agents produced buggy code that reduced overall productivity.
- High-throughput screening methods like DNA-encoded libraries have surprisingly low correlation to actual resynthesis validation results, creating false positives that waste downstream effort, making precise computational predictions potentially more reliable than massive screens.
- Critical drug discovery properties often anti-correlate: increased binding potency typically requires greater hydrophobicity while solubility requires polarity, and increased cellular penetration requires lipophilicity that limits water solubility, forcing multi-parameter optimization.
- Genesis's partnership with Exscientia combines AI model development with rapid experimental validation through in-house synthesis and screening, enabling true design-make-test-analyze loops that continuously improve models with fresh data.
- The company shifted its name from Genesis Therapeutics to Genesis Molecular AI to reflect its identity as an AI-first company while maintaining internal drug discovery programs to demonstrate real-world utility and validity of its models rather than just publishing benchmarks.
Topics
Transcript
I remember very clearly in 2017, 2018, talking about GANs and how generative adversarial networks and how they're clearly the future of image generation, obviously. They didn't work very well for proteins or protein ligand systems. And we sort of had to wait for the right primitive to get created. And that turned out to be diffusion, which turns out to be a much more useful primitive for the right primitive to get created. And that turned out to be diffusion, which turns to be a much more useful primitive for the space. What's kind of cool is right now for people that are interested in really core fundamental AI research, actually some of the most innovative diffusion research is…
Full transcript available for MurmurCast members
Sign Up to Access