🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI · Latent Space: The AI Engineer Podcast summary

Key takeaways

Significant AI advances, especially diffusion models and high-precision structure prediction, are transforming drug discovery.
Current benchmarks like RMSD < 2 Å are inadequate; stricter metrics such as 1 Å RMSD are necessary to model critical biological interactions reliably.
Genesis’s PEARL model demonstrates exceptional promise by accurately modeling protein flexibility and induced fit, often surpassing traditional methods.
Agentic workflows—automated, iterative AI-driven drug discovery loops—are becoming feasible, supported by partnerships with automated labs.
Progress in physics-based modeling, data synthesis, and real-world application integration is key to reaching a point where protein-ligand interactions are effectively 'solved'.

Summary

Advancements in AI for Molecular and Structural Biology

The podcast highlights how revolutionary progress in AI, especially diffusion models and structure prediction, has transformed drug discovery. Recent innovations like Genesis’s PEARL model demonstrate the ability to model protein-ligand interactions with unprecedented accuracy, surpassing traditional benchmarks like 2 Å RMSD. The discussion emphasizes how these advances enable understanding of protein flexibility and induced fit, crucial for real-world applications in drug design.

The Role of Benchmarks and Metrics in AI Development

The hosts critique existing benchmarks such as RMSD < 2 Å for protein-ligand pose prediction, arguing they are insufficient and sometimes misleading for practical drug discovery. Evan Feinberg advocates for stricter metrics like 1 Å RMSD to ensure core interactions, especially hydrogen bonds, are accurately modeled. The importance of evolving benchmarks (e.g., lDDT) to better reflect realistic goals is discussed, with Genesis’s models achieving surpassing results on new challenging benchmarks like OpenBind.

GeneSIS’s Innovative Models and Drug Discovery Philosophy

Genesis’s PEARL model embodies the latest in diffusion and structure prediction, capable of modeling dynamic conformational changes and induced fit. The company focuses on building models that provide high-resolution, physically meaningful predictions, enabling 'agentic' discovery workflows—automated, iterative cycles of hypothesis generation, testing, and synthesis, supported by partnerships like Insight. The goal is to reach a threshold of precision (around 1 Å RMSD) that makes these models practically useful for real-world drug development.

Industry Impact and Future Directions in Drug Discovery AI

The discussion covers the shift toward AI tools being core components in pharma, facilitating workflows from target identification to clinical trials. Genesis emphasizes their full-stack approach, including in-house wet lab capacity, and their focus on solving hard targets with novel chemistries. Sergey Edunov discusses upcoming developments like reinforcement learning in labs and agent-based systems for continuous, automated drug discovery. The importance of better physics-based modeling and more accurate benchmarks to truly 'solve' protein-ligand binding are also highlighted.

Show notes

This episode has a fun personal twist: There’s a counterfactual world where I was employee #1 at Genesis Molecular AI, the company behind today’s episode. A certain introduction happened a few weeks too late and I had already happily signed at Atomwise...

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