Unified modeling of mutation-induced protein stability responses with OmegaTherm

Abstract

Accurate prediction of mutation-induced protein stability responses, including changes in folding free energy and melting temperature, is critical for protein engineering. Existing models represent these related but non-equivalent measurements separately and often focus on single mutations. Here we present OmegaTherm, a sequence-based framework that jointly learns both stability responses for single and multiple mutations. OmegaTherm introduces a shared protein language model to capture transferable mutation context, measurement-specific encoders to preserve measurement identity, and paired-measurement-informed distillation to learn complementary knowledge from heterogeneous data. An antisymmetric prediction architecture further enforces physical consistency between forward and reverse mutations. Across independent benchmarks, OmegaTherm achieves state-of-the-art performance in quantitative prediction, mutation classification and protein-level ranking. Beyond benchmark prediction, OmegaTherm-derived stability landscapes correlate strongly with fitness landscapes from deep mutational scanning, enrich favorable variants and reveal function-associated mutation-sensitive regions. These results establish unified representation learning as an effective strategy for integrating related but non-equivalent biochemical measurements.