Abstract

Deep learning is rapidly advancing the best computational tools in computational biology. In this talk, I will discuss recent advancements in using generative models and sampling algorithms to generate proteins and optimize their fitness. First, I will go over FrameDiff in which we lay the mathematical foundation of SE(3) diffusion and introduce a practical algorithm for training a frame-based generative model over protein backbones. SE(3) diffusion is then utilized in a state-of-the-art protein design method, RFdiffusion, that is pre-trained on protein structure prediction. To conclude, I will go over latest work in sequence-based protein fitness optimization using Gibbs with Gradients. We argue for the importance of model regularization for handling protein fitness sparsity and ensuring a smooth optimization landscape.

https://cl-cam-ac-uk.zoom.us/j/94094286505?pwd=T2V3MFVSQ1ZKTFFyYlFueHlzTE83Zz09

Bio: Jason Yim is a second year EECS PhD student at MIT advised by Tommi Jaakkola and Regina Barzilay. Previously, he graduated from Johns Hopkins University with a bachelors in computer science then worked at DeepMind at a research engineer on medical imaging and AlphaFold-multimer. His research interest is to develop machine learning methods with applications to scientific problems. His latest works focus on inverse problems in protein design.