Robust Multi-Mutant Protein Stability Prediction from a Fine-Tuned Evolutionary Scale Model

Abstract

Recently, high-throughput experimental techniques have propelled improvements in deep learning-based prediction of mutation effects on protein stability. However, leading stability predictors still struggle to predict the combined effect of multiple mutations and prefer mutations that negatively impact other properties, including expressibility. To mitigate these limitations, we apply Low-Rank Adaptation (LoRA) to specialize ESM3 for stability prediction by fine-tuning on the Megascale protease susceptibility dataset, developing a novel dual-perspective inference mechanism to provide explicit mutant context information. ESM-Mutant Stability Ranker (ESM-MSR) significantly exceeds all contemporary methods tested on the prioritization of stabilizing mutations ({Delta}NDCG@96 [≥] +0.12), double mutant ranking ({Delta}{rho}avg [≥]+0.068) and direct epistasis ranking ({Delta}{rho}avg [≥] +0.164) within the Megascale test set. Further, it generalizes effectively to heterogeneous thermostability benchmarks, consistently matching or exceeding current approaches across our comprehensive suite. Finally, a single parameter{sigma} enables tunable control of the models compromise between stability and more general sequence fitness, leading to state-of-the-art performance in the Human Domainome 1 benchmark ({rho}avg = 0.573) at{sigma} = 0.5, demonstrating the broad applicability of ESM-MSR as a protein engineering tool.