Physics-Based Energy Functions for Computational Protein Design

Computational protein design (CPD) aims to conceive new proteins or modify existing ones to achieve a functional or structural goal, using numerical methods. Among the various branches of CPD, one consists in predicting sequences given a protein backbone. This is known as the inverse folding problem. It has been particularly fruitful over the last 40 years, has given rise to numerous methodological approaches, and has obtained experimental successes, such as the design of new folds and new enzymatic functions. One criterion for distinguishing between the methods proposed to tackle this problem is the scoring or energy function, which enables different possible sequences and conformations to be compared quantitatively. A traditional classification of scoring functions distinguishes between statistical, empirical, and physics-based approaches. Recent developments in CPD have brought to the fore new approaches based on deep learning. Improvements in prediction performance are undeniable. However, physics-based methods retain advantages due to their greater explanatory power and independence from a training dataset. We review here CPD works that have been using physics-based energy functions and discuss their interests and perspectives.