Evan Feinberg and Sergey Edunov of Genesis Molecular AI discuss how diffusion-based co-folding models (Pearl) can predict protein–small molecule 3D structures with sub-1 angstrom accuracy, enabling precise binding affinity predictions. They highlight the use of physics simulations for data scaling and inference-time refinement, and describe their progress toward agentic platforms that can automate drug discovery workflows.
Summarized by Podsumo
Genesis Molecular AI's **Pearl** model achieves sub-1 angstrom accuracy on protein–ligand pose prediction, a critical threshold for downstream drug design.
Synthetic data from **physics-based simulations** (e.g., MD) scales training beyond the limited experimental PDB database, a key differentiator from protein–protein models.
The team applies **inference-time scaling** with a diffusion head guided by physics verifiers, analogous to "thinking tokens" in LLMs.
They are building **Sapphire**, an **agentic platform** where an LLM orchestrates drug discovery tools (pose prediction, ADMET, synthesis) to automate medicinal chemistry campaigns.
CEO and CTO both identify **GPU shortage** as the biggest bottleneck, urging that drug discovery AI should get more hardware priority over conversational LLMs.
"Some of the most innovative diffusion research is happening in our field. It's happening in 3D structure prediction right now."
"For small molecules, there are 10^60 drug-like molecules in the universe. It's finding a needle in a haystack."
"In the same way that for coding, agents are only as useful as the underlying models… we needed sub-angstrom accuracy before agents could be trusted."