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

2026-06-24 · Source: Latent.Space - Www.latent.space · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI in Drug Discovery · Depth: Expert, extended

Summary

Genesis Molecular AI, co-founded by Evan Feinberg and led by former Llama 2/3 lead Sergey Edunov, is advancing AI in drug discovery, particularly in 3D protein-ligand structure prediction. Their PERL model leverages diffusion networks and physics-based guidance, achieving sub-1 Angstrom accuracy, a crucial resolution for effective drug design. This capability, demonstrated by outperforming other models on challenging dynamic targets like the EVA721A protease in the OpenBind challenge, addresses historically "undruggable" targets and improves existing therapies. Genesis integrates synthetic data generation and physical priors to overcome limited public datasets. The company focuses on small and medium-sized molecule discovery, collaborating with pharmaceutical partners like Incyte, and is developing AI agents to enhance drug hunters' productivity by orchestrating complex tools.

Key takeaway

For AI Scientists and Machine Learning Engineers developing drug discovery models, you should prioritize sub-1 Angstrom accuracy in 3D protein-ligand prediction. This resolution is crucial for enabling effective downstream potency prediction and prospective design, moving beyond traditional 2 Angstrom benchmarks. Focus on integrating physical priors and synthetic data generation, as demonstrated by PERL, to achieve generalizable models that can tackle "undruggable" targets and accelerate therapeutic development.

Key insights

Diffusion models, enhanced with physical priors and synthetic data, are achieving critical resolution in 3D protein-ligand prediction, transforming drug discovery.

Principles

• Diffusion models are superior for 3D molecular generation than GANs.
• Sub-1 Angstrom accuracy is critical for effective drug design.
• AI models benefit from physical priors and synthetic data.

Method

PERL uses a diffusion-based head for iterative structure refinement, guided by physics-based feedback and synthetic data from molecular physics simulations, similar to LLM scaling.

In practice

• Use diffusion models for 3D molecular structure prediction.
• Integrate physics-based guidance for higher accuracy.
• Develop AI agents to orchestrate complex drug discovery tools.

Topics

• Diffusion Models
• Protein-Ligand Prediction
• Drug Discovery AI
• Molecular Co-folding
• AI Agents
• Synthetic Data Generation

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Two-dimensional geometric template diffusion for boosting single-sequence protein structure prediction — TDFold uses 2D geometric template diffusion for efficient, accurate single-sequence protein structure prediction.
• AI model used to generate complete models of proteins in motion — LD-FPG is the first AI framework to generate complete, all-atom dynamic protein models, capturing full motion.
• LEAP: A closed-loop framework for perovskite precursor additive discovery — Coupling a domain-specialized LLM with active learning and expert review accelerates perovskite additive discovery under low-data conditions.
• Author Correction: Mask-prior-guided denoising diffusion improves inverse protein folding — A published correction clarifies MapDiff's amino acid prediction accuracy in inverse protein folding.
• Escape Velocity Is the Wrong Dream — AlphaFold's solution to protein folding demonstrates that profound scientific breakthroughs are occurring in plain sight, not just in distant futures.
• Move over, AlphaFold: open-source model predicts shape of 1 billion proteins — ESMFold2, an open-source AI, predicts over a billion protein structures, outperforming AlphaFold3 and enabling novel protein design.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.