FrustraMPNN: An ultra-fast deep learning tool for proteome-scale analysis of deep mutational single-residue local energetic frustration in proteins

Abstract

Energetic frustration, characterized by conflicting local interactions, is a key determinant of protein dynamics, allostery, protein-protein interactions, enzyme catalysis, and overall protein function. Although the principle of minimal frustration describes an evolutionary bias towards a reduction of energetic conflicts for efficient protein folding, local violations are selectively embedded to encode the complex energy landscapes necessary for protein functions. However, the computational cost of traditional single-residue frustration analysis has prevented the calculation of these energetic conflicts at the proteome scale, creating a significant bottleneck in structural biology. Here, we introduce FrustraMPNN, a message-passing neural network retrained via transfer learning that predicts a complete per-residue frustration-index mutation profile (full mutational scanning per position for a whole protein) orders-of-magnitude faster than existing methods while maintaining high accuracy. This is demonstrated by calculating the single-residue frustration of the E. coli proteome, reducing the calculation time from years to 12 hours. We validate FrustraMPNN on a diverse external set of over 3,400 human protein structures, achieving a Spearman correlation of up to 0.80 across all frustration categories, demonstrating robust generalizability. By conducting a thorough dataset ablation study, we find that the models performance improves when trained on a broader range of protein sizes, highlighting an inherent limitation of datasets like Megascale. We provide FrustraMPNN as an open-source tool, expecting it will enable exploration of new hypotheses in fields such as personalized structural biology and allow analysis of local energetic frustration patterns involved in protein function at an unprecedented proteomic scale.