Improving De Novo Protein Binder Design With Deep Learning

This article investigates the ability of deep learning methods to discriminate binders from non-binders in a set of 1 million experimentally characterized designs for 10 different targets. 15,000-100,000 designs were tested for each target, and the number of actual binders ranged from 1 to 584. As a baseline, the Rosetta energy of the monomer, normalized by chain length, was used. This metric provided little discriminatory power. In contrast, the deep learning-based accuracy prediction method DAN was able to partially discriminate binders from non-binders. AF2 structure predictions were also evaluated for the five minibinder structures from Cao et al. for which structures have been solved experimentally, and were found to predict the monomer structure with binder Cɑ accuracy between 0.2 Å-0.8 Å. An updated version of RoseTTAFold (RF2) was also found to predict all monomer structures with binder Cɑ accuracy between 0.2 Å-0.8 Å.