Modeling Proteins and Genes with Generative AI

Abstract

The intersection of biology and generative AI is rapidly reshaping how we understand and design molecules, proteins, and genes. This study blog is a curated learning journey through the foundational biological concepts—molecular structures, protein folding, and genetic sequences—paired with state-of-the-art generative models that simulate and engineer them. From graph-based diffusion models for molecule generation to large-scale protein language models like ESM-3, and the emergence of generative tools in synthetic biology like Evo2, we explore how AI is augmenting the design, prediction, and discovery of life’s building blocks. This blog aims to provide a structured and accessible path to the future of biological design with AI.

Study Plan

Part 1: Molecules

• PoseX: AI Defeats Physics-based Methods on Protein Ligand Cross-Docking (paper)
• Graph Diffusion Transformer for Multi-Conditional Molecular Generation (NeurIPS 2024 Oral) (paper)
• HOG-Diff: Higher-Order Guided Diffusion for Graph Generation

Part 2: Proteins

• RFdiffusion2 Youtube Lecture
• ESM3 Youtube Lecture
• Protein Language Model, ESM3 (ESM3 tutorial)
  1. Input tracks of ESMProtein
  2. How to embed a batch of sequences using ESM C and explore different layers
  3. Design a GFP Candidate with ESM3
  4. Prompting Examples to Generate with ESM3: Scaffolding a Motif, Helix Shortening, Exposing a Buried Helix
  5. Guided Generation with ESM3 using scoring_function

Part 3: Genes

• EVO2 ( (colab) hello_evo.ipynb, (colab) BioForge_Evo2_Synthetic_Biology_Tutorial.ipynb, Evo2 Github)

References

• Reference 1