Project Description The project consists in optimizing and expanding TorchProteinLibrary, our library of differentiable primitives for deep neural network models of protein structure (see our preprint “TorchProteinLibrary: A computationally efficient, differentiable representation of protein structure” https://arxiv.org/abs/1812.01108). The library implements the functionalities needed to perform end-to-end learning of protein structure prediction.

The recent success of DeepMind’s AlphaFold2 has shown the need for an open software platform for the development of machine learning (ML) models focused on molecular structure and function, so that a diverse ecosystem of new methods can grow on top of a stable software base. Research interest on the topic is rapidly moving beyond protein structure prediction, towards integrating greater molecular diversity (DNA, RNA, small ligands, etc.) and towards understanding the interactions, dynamics, and context-specific details responsible for a given function or activity.

TorchProteinLibrary (TPL), our own open-source initiative, is ideally positioned to become a central tool for future development of deep learning models of biomolecular structure. The library is meant to develop along a number of different axes: 1) performance optimization (CUDA implementation and refactoring), 2) implementation of new differentiable layers (protein sequence/structure conversion and transformation), and 3) extension to nucleic acids (DNA and RNA).