🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI

What It Covers

Evan Feinberg (CEO) and Sergey Edunov (CTO, former Llama 2/3 pretraining lead) explain how Genesis Molecular AI built Pearl, a diffusion-based protein-ligand co-folding model achieving sub-one-angstrom pose prediction accuracy. They cover synthetic data generation, inference-time scaling, reinforcement learning loops with wet lab partners, and why the most innovative diffusion research now happens in 3D structural biology rather than image or language domains.

Key Questions Answered

• •Sub-angstrom accuracy threshold: Protein-ligand pose prediction must reach below one angstrom RMSD to be useful for downstream drug discovery tasks. The standard two-angstrom benchmark inherited from pre-AI docking studies allows entire aromatic rings to flip undetected, producing structurally plausible but chemically wrong outputs. Hydrogen bonds require donor-to-acceptor distances of 2.7–3.3 angstroms, a 0.6-angstrom window, meaning two-angstrom model errors routinely destroy the physical validity of predicted interactions.
• •Diffusion as the correct primitive for molecular structure: GANs, the dominant generative framework circa 2017–2018, failed for protein-ligand systems due to mode collapse and training instability. Diffusion models solved these problems and now represent the frontier of generative AI research in 3D structural biology. Counterintuitively, the most novel diffusion research is no longer in image or video generation but in co-folding tasks, where iterative denoising maps naturally onto conformational refinement of molecular complexes.
• •Inference-time scaling via physics-guided diffusion: Pearl applies inference-time compute scaling analogous to chain-of-thought reasoning in LLMs, but instead of language tokens, the model iterates over intermediate crystal structure representations. Physics-based force fields steer the diffusion trajectory during each refinement step, functioning like a verifier in a loop. This approach substantially improves pose accuracy without retraining and mirrors the three-stage LLM scaling roadmap: pretraining, post-training, then inference-time scaling.
• •Synthetic data from molecular physics simulations: The RCSB Protein Data Bank contains roughly 200,000 crystal structures, an insufficient and slowly expanding dataset for training generalizable models. Small molecules, unlike large proteins, are computationally tractable for physics-based simulation, enabling Genesis to generate large volumes of synthetic conformational data for pretraining. This physics-derived synthetic data pipeline is the primary mechanism for scaling beyond public experimental data, directly paralleling how LLM pretraining uses synthetic text corpora.
• •Reinforcement learning with wet lab feedback loops: Genesis runs design-make-test-analyze cycles with Insilico Medicine as a lab partner, synthesizing model-proposed compounds, measuring binding and ADMET properties experimentally, and feeding results back into model fine-tuning. This closes the RL loop at the wet lab level rather than relying solely on GPU-based self-play. The bottleneck is synthesis complexity: novel chemistries required for Pareto-optimal multi-parameter molecules cannot yet be reliably automated, limiting cycle speed for frontier compounds.
• •Multi-property optimization as the real drug discovery bottleneck: Binding affinity, solubility, oral bioavailability, CYP3A4 inhibition, hERG channel activity, and roughly 30 additional ADMET endpoints must all fall within acceptable ranges simultaneously. These properties frequently anti-correlate: hydrophobicity improves binding but reduces solubility; adding polarity to fix solubility impairs membrane permeability. Pearl's 3D structural outputs feed directly into ADMET prediction models, enabling multi-parameter optimization from a shared structural representation rather than treating each property as an independent flat-feature regression.
• •Agentic drug discovery requires model quality thresholds first: Genesis is building an agentic platform codenamed Sapphire that orchestrates Pearl and ADMET models for continuous 24/7 drug design. The prerequisite was achieving pose prediction quality where model outputs would not be rejected by medicinal chemists. Below the one-angstrom threshold, agents amplify errors rather than productivity, analogous to mid-2024 coding agents that compounded subtle LLM bugs. The intended human role shifts to strategic direction-setting while agents handle tool orchestration, parameter selection, and iterative hypothesis generation.