Protein design updates its engine
The field is shifting from physical models to deep learning
Computational protein design is shifting its center of gravity from mechanistic calculations based on physical forces to more powerful machine learning approaches trained on reams of data, according to David Baker, one of the field’s pioneers. The gains could extend to biologics, vaccines, cell therapies and diagnostics.
Papers published over the past year by Baker’s lab at University of Washington’s Institute for Protein Design (IPD) introduce new methods based on deep learning — a subset of machine learning that uses multiple layers of artificial neural networks — that are helping drive that shift, including a strategy dubbed “protein hallucination” that the team used to design therapeutically relevant proteins including an RSV vaccine antigen...