Reproduce “BetterBodies: Reinforcement Learning guided Diffusion for Antibody Sequence Design” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yannick Vogt, Mehdi Naouar, Maria Kalweit, Christoph Cornelius Miething, Justus Duyster, Joschka Boedecker, Gabriel Kalweit
BetterBodies is a computational method for designing antibody Complementarity Determining Region H3 (CDRH3) sequences with improved binding affinity, using a combination of Variational Autoencoders and Reinforcement Learning-guided latent diffusion. The approach demonstrated superior mean binding affinity, achieving lower free energy values compared to basic diffusion and GFlowNets across various datasets.
Key insights
- Integrating offline RL guidance into a latent diffusion model effectively shifts the generative policy towards regions of higher predicted binding affinity, enabling property-guided sequence design.
- Embedding biophysical properties into the VAE's latent space via contrastive learning creates a structured representation that correlates with higher Q-values, aiding in the generation of functional sequences.
- Framing sequence generation as a stepwise process guided by Q-functions allows for potentially "stitching together" improved sequences, offering a robust way to navigate the vast design space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.