TorchProteinLibrary: A computationally efficient, differentiable representation of protein structure

Predicting the structure of a protein from its sequence is a cornerstone task of molecular biology. Established methods in the field, such as homology modeling and fragment assembly, appeared to have reached their limit. However, this year saw the emergence of promising new approaches: end-to-end protein structure and dynamics models, as well as reinforcement learning applied to protein folding. For these approaches to be investigated on a larger scale, an efficient implementation of their key computational primitives is required. In this paper we present a library of differentiable mappings from two standard dihedral-angle representations of protein structure (full-atom representation "$φ,ψ,ω,χ$" and backbone-only representation "$φ,ψ,ω$") to atomic Cartesian coordinates. The source code and documentation can be found at https://github.com/lupoglaz/TorchProteinLibrary.