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

Genesis Molecular AI is applying advanced generative AI to the complex challenge of small molecule drug discovery. By utilizing diffusion models, the company has developed a system called PEARL that predicts how drug candidates interact with proteins. Unlike traditional methods that treat proteins as static structures, PEARL models protein flexibility and induced fit, allowing it to accurately simulate how molecules bind to their targets.

A major hurdle in this field has been the reliance on insufficient academic benchmarks. Current standards often accept a 2-angstrom margin of error, which researchers at Genesis argue is inadequate for capturing critical chemical interactions like hydrogen bonds. By pushing for higher precision, the team has demonstrated that their models can outperform existing methods, as evidenced by PEARL’s success on the OpenBind benchmark, where it accurately predicted binding poses for hundreds of previously unseen complexes without needing fine-tuning.

These technical improvements have enabled the creation of agentic workflows, such as the company’s internal system, SAPPHIRE. These agents can reason about molecular structures, consult literature, and iterate on drug candidates autonomously. By integrating these AI models with automated laboratory testing, the company aims to accelerate the discovery process. This shift represents a transition from early, unreliable AI tools to sophisticated systems capable of performing complex, multi-parameter optimization, potentially bringing the industry closer to solving the challenge of protein-ligand binding.