Navigating high-order protein mutational landscapes via deep learning on directed evolution trajectories

Accurately predicting the fitness effects of high-order mutations is a grand challenge in understanding and engineering proteins. Existing models, including pre-trained protein language models, struggle to capture the multi-residue interactions that govern these effects. Here, we introduce DENet, a deep learning framework that harnesses the rich co-mutation information within directed evolution (DE) trajectories to reconstruct high-resolution fitness landscapes for deciphering and engineering of complex protein variants. Applied to the cancer target KRAS, DENet-guided screening systematically identified high-order mutants with potent activities and uncovered hidden allosteric mechanisms. For MEK1, DENet discovered complex variants with >1,000-fold increased drug resistance, revealed synergistic tail mutations, and retrospectively identified over 75% of known clinical mutations, largely outperforming existing models. To broaden the framework's applicability, we developed an in silico strategy that simulates directed evolution to generate crucial co-mutation information from widely available single-mutant datasets. DENet provides a quantitative framework for navigating complex fitness landscapes, uniting the rational engineering of multi-mutation proteins with the elucidation of their allosteric and clinical implications. Overall design: High-throughput sequencing data were collected from the most highly mutated regions during the directed evolution campaigns of KRAS and MEK1 to identify and quantify mutant variants generated and selected at multiple time points.