Jun 30, 2026 · 1h 48m · latent-space
🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation"
Evan Feinberg · 59m spoken Sergei Yudinov · 17m spoken Brandon Anderson · 12m spoken RJ Haneke · 10m spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Latent Space AI for Science podcast, Genesis Molecular AI leaders Evan Feinberg and Sergei Yudinov discuss how frontier 3D diffusion models, synthetic physics data, and sub-angstrom precision are transforming small-molecule drug discovery and multi-parameter optimization.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 6.3 Guest teaching 4.4 Guest disagreement 1.8 The hosts pushing back 2.2
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:18:00 Pushback on automated robotic chemistry hype
Evan forcefully rejects the industry narrative that automated cloud labs have solved wet-lab bottlenecks, pointing out abysmal correlation rates and synthetic purification challenges.
Hardest push from the hosts ▶ 44:21 Challenging the physical reality of static binding posesBrandon directly challenges whether a static 3D pose is an artificial human abstraction rather than a thermodynamic probability ensemble.
Biggest teaching moment ▶ 50:20 The 0.6-angstrom physics of hydrogen bondingEvan gives a detailed physics breakdown of hydrogen bonding donor-acceptor distances, demonstrating why predictions above 1 Angstrom produce unusable slop.
The host holds their own ▶ 18:18 Host drilling into diffusion steering and verification loopsRJ demonstrates strong technical fluency by dissecting how inference-time reasoning maps to diffusion verifiers and loop transformers.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The Evolution of AI in Molecular Biology and Drug Discovery | 6 | 3 | 1 | 1 | Brandon frames the decade-long arc of machine learning in molecular biology and zeroes in on the historically resistant problem of protein-small molecule interactions. Evan enthusiastically agrees and elaborates on the iterative compounding nature of AI in drug discovery. | |
| Fundamentals of Drug Discovery & the 3D Structure Hypothesis | 6 | 4 | 1 | 2 | Brandon asks about historical failure modes and computational expense in 3D structure prediction. Evan explains lock-and-key biology and how testing the 3D coordinate hypothesis historically required cost-prohibitive crystallographic experiments. | |
| Introducing Pearl: 3D Co-Folding and Synthetic Physics Data | 6 | 5 | 1 | 2 | Hosts and guests discuss co-folding models like Pearl, AlphaFold 3, and OpenFold. Sergei and Evan explain why small molecules present a 10^60 search space and how physics-based synthetic data augments the sparse PDB database. | |
| Scaling Laws and Inference-Time Reasoning for Structure Prediction | 7 | 4 | 1 | 3 | Sergei connects LLM inference-time reasoning to crystal structure diffusion heads, while RJ drills into the iterative steering loop, mechanistic interpretability, and loop transformers. | |
| Incorporating Physical Priors and Focusing on Small Molecules | 6 | 5 | 2 | 1 | Evan discusses representation learning, physical priors, and why AlphaFold-predicted structures historically failed in docking force fields due to low pocket resolution. | |
| High-Leverage Intervention Points in the Drug Development Pipeline | 6 | 4 | 2 | 2 | RJ prompts a walkthrough of the drug development lifecycle, and Evan argues that hit-to-lead and target drug design are the highest-leverage intervention points for AI. | |
| First-in-Class vs. Best-in-Class Targets and Industry Partnerships | 7 | 4 | 2 | 3 | Brandon challenges whether known-biology targets have already been picked over. Evan counters that biology validation is orthogonal to druggability and cites ALK inhibitor generational progress. | |
| Beyond Static Structures: ADMET Properties and Multi-Parameter Optimization | 6 | 4 | 2 | 2 | Sergei and Brandon debunk the popular notion that Nobel-winning structure prediction solved drug discovery, stressing ADMET and off-target selectivity. Evan shares progress on partner programs. | |
| The Sub-Angstrom Resolution Imperative and Metric Evals | 7 | 5 | 2 | 4 | Brandon challenges whether static poses are an artificial abstraction rather than probability distributions. Sergei and Evan defend rigid poses as necessary for docking verification and explain sub-angstrom thresholds. | |
| Agentic Chemistry (Sapphire) and Sub-Angstrom Precision | 6 | 6 | 3 | 2 | Evan details how poor 1.9 Angstrom RMSD predictions yield agentic slop and explains hydrogen bonding geometry tolerances (0.6 Angstrom margin), introducing the Sapphire agentic platform. | |
| Achieving 1-Angstrom Precision & The Eval Evolution | 7 | 5 | 2 | 3 | Sergei breaks down the eval-centric engineering approach. Evan compares misleading 2-Angstrom metrics to SWE-bench discrepancies and cites PoseBusters as a necessary eval evolution. | |
| The Evolution of Machine Learning for ADME Prediction | 7 | 4 | 1 | 2 | Brandon asks about the history of graph neural networks and ADMET prediction. Evan discusses PotentialNet, MoleculeNet, and predicting multi-parameter endpoints like hERG channel inhibition. | |
| The Convergence of Diffusion Models in Structural Biology | 6 | 4 | 1 | 1 | Evan reviews the transition from GAN mode collapse to 3D diffusion models, noting that the most innovative diffusion research is currently happening in molecular co-folding. | |
| Wet-Lab Integration, Automation Realities, and RL Feedback Loops | 6 | 6 | 4 | 3 | Evan pushes back hard against overhyped claims of fully robotic automated labs, explaining chemical purification, NMR verification, low high-throughput screening R-squared, and anti-correlating molecular properties. | |
| Company Strategy: Genesis Molecular AI and Human-Agent Symbiosis | 7 | 4 | 2 | 3 | RJ and Brandon explore the company's rebrand and strategic shift toward agents. Sergei explains why LLM agents allow med chemists to orchestrate complex tools without manual hyperparameter tuning. | |
| Outperforming Benchmarks on the OpenBind EV-A71 Protease Challenge | 6 | 4 | 1 | 2 | Sergei and Evan review Pearl's performance on the OpenBind EV-A71 protease challenge, specifically highlighting its ability to model dynamic flexible loop movements and induced-fit binding. | |
| Industry Bottlenecks, Compute Scarcity, and the Future of AI for Science | 6 | 4 | 3 | 2 | Sergei calls out LLM architectures as comparatively boring transformer layers from 2017 and flags GPU compute bottlenecks, while Evan suggests that alpha in generic LLMs is peaking relative to life sciences. |
Statements from this episode (24)
Feinberg: Solving experimental 3D protein structures takes months and tens of thousands
“Like you might as well solve the three D crystal structure or cryo-EM structure is what you're saying, which can cost tens of thousands of dollars. It can be months or years. Right. Entire PhD theses, postdocs can sometimes work literally 24 seven trying to so…”
Edunov: There are 10^60 drug-like small molecules in the universe
“There are 10 to the 60 drug-like small molecules in the universe.”
Edunov: The Protein Data Bank contains around 200,000 crystal structures
“So a lot of training data that people are used is so-called PDB which is a public database of like all of the historical Crystal structures, and it's not that big. It's like 200,000 crystal structures, and it's very hard to expand.”
Edunov: Small molecules can be modeled with physics to generate training data
“In small molecular space, you can actually model your small molecules with physics. You can model their behavior, and that allows you to create more data that you can train model on. Something which is not necessarily possible in protein to protein.”
Edunov: Physics-guided inference-time scaling significantly boosts molecular model performance
“We are doing very similar thing with our models where a model is forced to think, except it's not thinking in language tokens. It's thinking in terms of crystal structures, not fully materialized crystal structures, but some sort of a crystal structure represe…”
Feinberg: Genesis CTO Sergey Edunov led Meta's Llama 2 research team
“Sergei led the LLAMA II research team at Meta when he was still there.”
Feinberg: Small molecules make up 65% of FDA-approved drugs
“Small molecules are still 65% of FDA approved drugs, so we're talking about the biggest part of the pie.”
Feinberg: Studies show AlphaFold structures provided no value for drug docking
“There is this, ah, a few papers that came out, one was in Cell, I think last year, which showed that for all of the claims about AlphaFold-solving drug discovery, people try to take AlphaFold-produced protein structures, use them for traditional docking, and f…”
Feinberg: Drug discovery and design is AI's highest-leverage application in healthcare
“Our contention is that the highest leverage application of artificial intelligence
Is the drug discovery and drug design process.”
Feinberg: Biological target validation is orthogonal to druggability
“Known biology is orthogonal to the ease with which one can drug that target. Sometimes, unfortunately, it seems they anti-correlate in that often it seems that the most appealing targets from a validation perspective seem to be really hard to drug.”
Edunov: Predicting ADMET properties is as critical as protein structure prediction
“Predicting all of those properties is also just as important as, or maybe even more important than predicting the crystal structure itself.”
Edunov: Predicting molecular poses is necessary to validate drug discovery models
“Yes, it's an abstraction, but it's a very useful abstraction. It helps us to build up confidence that a particular model output is actually valid. Whether you did not just straight up hallucinate at something, because yes, ultimately what matters is binding af…”
Feinberg: Genesis is developing autonomous drug discovery agent platform 'Sapphire'
“We're working on an agentic platform for 24 seven drug discovery. If you can imagine just fleets of hundreds of med chemists and CAD scientists Working nights and weekends all the time for your drug targets. The code name for that gem is Sapphire.”
Feinberg: Potent ligands have well-defined 3D positions down to 0.5 angstroms
“The reality is that for a highly potent ligand, Almost certainly, there is a large portion of that molecule with a very well-defined three-D position, down to even half angstrom.”
Yudinov: Only data, infrastructure, and evals matter in AI development
“Where three things matter in AI it's data, infrastructure, and evals. Right. So you can only improve what you measure. And once you are very careful about measuring, What matters? And you have really talented people on the team. We're going to figure out how t…”
Feinberg: Gemini is obviously worse at coding despite benchmark wins
“So Gemini does pretty well on Sweebench. Sometimes Gemini publishes models that win on some of those software benchmarks. Raise your hand if you're using Gemini to write code right now instead of, you know, the obvious other name competitors. No one. Like, why…”
Feinberg: Drug candidates must pass over 30 distinct assay properties
“I would say there's over 30 or so assays, each of which you could imagine if you're a neural net person, like a multitask neural network or multi-head, it's got to predict over 30, you know, three dozen-ish properties, each of which, if it's in the wrong range…”
Feinberg: GANs failed for protein modeling due to mode collapse before diffusion emerged
“For all the same reasons that those models were really tricky to train for images. Mode collapse was the most famous sort of problem. 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 c…”
Feinberg: The most innovative diffusion research is happening in 3D structure prediction
“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 happening in our field, is happening in three D structure prediction right now.”
Feinberg: High-throughput screening exhibits a shockingly low R-squared to actual synthesis
“The reality is that the translation of high throughput screens, whether it's Dell, DNA encoded libraries, or more traditional screens, the R-squared of those predictions to, like, the actual business of resynthesizing a molecule de novo and doing a Low through…”
Feinberg: Robotic lab automation trades off molecule novelty and quality for speed
“The kinds of chemistries that today can be automated are fairly constrained. And so your ability to search chemical space broadly for those really top top Pareto optimal compounds is, is actually very limited. So the benefits you get in speed have a very harsh…”
Yudinov: Crystal structures can become a native modality for LLMs
“Or you can even train a model which is able to natively understand crystal structure representation, and like image modality is a modality for LLMs these days. Crystal structure can be a modality for LLMs as well.”
Feinberg: Chipmakers will invest in life sciences as LLM alpha shrinks
“I do think that chip makers, including Nvidia, are going to want to get a lot more invested in life sciences because it will always be high in demand. And The amount of alpha left in pure LLM space is just getting a little questionable.”
Edunov: LLM architectures are boring and fundamentally unchanged since 2017
“And honestly, LLM architectures are relatively boring. I don't know, probably alienate half of your audience. But it's like, it's a transforming layer in the end, like paper was published in 2017, and you go to any LLM lab today, you will see very, very simila…”