Deep learning methods for protein structure prediction have gained considerable momentum over the past few years. Though structures for many proteins are available at adequate or better quality via these methods, the community is interested in improving their reliability, quality, and speed. In particular, the prediction accuracy of all-atom protein structures may not be high enough for direct use in drug discovery campaigns, especially concerning protein sidechain placement. Even the most advanced deep learning methods for protein structure prediction utilize downstream energy minimization steps to improve their predictions' quality and physical properties. In this work, we propose a framework (SidechainNet) that allows all-atom protein structure predictions to be directly optimized during training using a specialized loss function that coordinates information between molecular dynamics and machine learning software. We hypothesize that our method will enable more physically realistic protein structure predictions, eliminate potentially costly downstream minimization tasks, and incorporate valuable physical knowledge into the trained deep learning model.

SidechainNet: All-atom structure predictions enable training of deep learning models with molecular dynamics forcefields

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.