DETANGO: Disentangling mutation effects on protein stability and function

Understanding how evolutionary constraints shape protein sequences is fundamental to deciphering the molecular mechanisms underlying protein stability and function. Protein language models (pLMs) accurately predict mutation effects through evolutionary information, but their scores conflate stability- and function-driven selective pressures, obscuring the mechanisms underlying mutation effects. We introduce DETANGO, a deep learning framework that explicitly deconvolves the mutation effects on protein functions by removing components attributable to stability perturbations from the pLM-predicted mutation effects. Guided by computational or experimental stability measurements, DETANGO estimates a functional plausibility score for each single-point mutation that is the component of the mutation effect not accounted for by changes in stability. Single-point mutations with low functional plausibility are predicted to be stable-but-inactive (SBI) variants, whose compromised activities are caused by direct perturbations on functional mechanisms rather than structural stability. Residues enriched for such variants are inferred to be functionally critical, as indicated by the evolutionary pressures to maintain protein function. Across extensive benchmarks, DETANGO accurately identifies SBI variants and pinpoints functionally important sites across contexts, including ligand binding, catalysis, and allostery. These results establish DETANGO as a biologically grounded framework for understanding mutation mechanisms and informing rational protein engineering and therapeutic development.