Pre-trained Ensembles for Bayesian Optimization of Protein Sequences

Explore a cutting-edge approach to protein sequence design through Bayesian optimization using pre-trained models in this 49-minute talk by Ziyue Yang from Valence Labs. Delve into the transformative potential of pre-trained sequence models in enabling accurate predictions with minimal labeled data. Learn how this method can significantly reduce the number of experiments needed for various sequence design tasks, including creating novel peptide inhibitors with AlphaFold. Discover the advantages of using pre-trained models for good predictive accuracy at low data points and how Bayesian optimization guides the selection of sequences to test. Gain insights into key concepts such as latent space optimization, the Gumbel Softmax trick, and calibration at low data. The presentation concludes with a Q&A session, offering a comprehensive overview of this innovative approach to protein sequence optimization.

- Intro
- Sequences & Pre-trained Models
- Bayesian Approach
- Latent Space for Optimization
- Gumbel Softmax Trick
- Calibration at Low Data
- Conclusions
- Q+A