?
High-throughput computational design of protein binders for complex targets using deep learning models
Computational protein design methods has transformed structural bioinformatics by overcom- ing many experimental limitations. Previously, experimental methods such as directed evo- lution were utilized to create protein binders. Many advancements in computational protein design have made it possible to generate de novo binders solely based on target structure and sequence information. However, despite recent progress, designing de novo protein binders still poses difficulties, as the mean success rate of experimental testing remains relatively low (1).
Deep learning approaches has shown promise in addressing this challenge, especially after the success of AlphaFold model in the task of protein structure prediction (2). This study aims to combine many different approaches of geometrical and generative neural networks into a single semi-automatic pipeline for protein binder design. The proposed pipeline includes methods of structural analysis and binding interface prediction, binder backbone and sequence generation, and AlphaFold 2 model as the main tool for validation. Many studies have applied similar techniques to generate binders for well-known protein targets, some of which may have limited geometric complexity. However, in this particular case, the pipeline is applied to the more challenging landscapes of large protein complexes. We generate several hundred designs, depict the pros and cons of different binder generation approaches and evaluate their performance and computational resource consumption. The developed approach can serve as a base for high- throughput in silico binder design as well as the benchmark test for similar protein design tools.