Improving protein side-chain accuracy using conditioned torsion angle loss function

Abstract

Determining the three-dimensional protein structure traditionally needs laborious and time-consuming experimental techniques. However, recent advancements in machine learning have led to the development of predictive methods, such as AlphaFold2 and RoseTTAFold, for protein structure prediction. AlphaFold2 has made significant strides in accurately predicting the backbone structure of proteins. Nevertheless, the prediction accuracy of AlphaFold2 for side-chains falls short of the backbone prediction accuracy. In this work, we propose a novel approach to enhance the side-chain prediction accuracy based on AlphaFold2 by introducing a conditional torsion angle loss function. The effectiveness of this approach is evaluated through experiments on diverse protein structures, demonstrating its potential to improve side-chain prediction accuracy and contribute to the field of protein structure determination.