Protein Discovery with Discrete Walk-Jump Sampling

Explore protein discovery techniques through discrete walk-jump sampling in this 55-minute conference talk by Nathan Frey from Valence Labs. Delve into the challenges of training and sampling from discrete generative models, and learn about a novel approach that combines energy-based and score-based modeling. Discover how this method simplifies training and sampling while improving sample quality. Examine the application of this technique to antibody protein generation, with impressive results in laboratory experiments. Follow along as the speaker discusses the learning of smoothed energy functions, Langevin Markov chain Monte Carlo sampling, and the projection back to the true data manifold. Gain insights into the evaluation of protein generative models using the distributional conformity score. Witness the first demonstration of long-run fast-mixing MCMC chains visiting diverse antibody protein classes. The talk covers background information, learning scores on a smooth manifold, discrete walk-jump sampling, high fitness molecule production, and concludes with a discussion on the implications and potential of this innovative approach in AI-driven drug discovery.

- Intro + Background
- Learning Scores on a Smooth Manifold
- Discrete Walk-Jump Sampling
- WJS Produces High Fitness Molecules
- Discussion
- Conclusion