ThermoESM: Bridging Experimental Tm and OGT Data with Protein Language Model for Reliable Protein Thermal Stability Classification

Protein thermal stability is fundamental to protein function, evolutionary adaptation and biotechnological applications. However, current predictors are limited by the noisy nature of optimal growth temperature (OGT) labels and the scarcity of experimentally determined melting temperatures (Tm), particularly for highly thermostable proteins. To address these challenges, we curated a unified dataset of [Formula: see text]90,000 proteins by integrating Tm values from high-throughput Thermal Proteome Profiling (TPP) measurements with OGT-derived high-stability sequences. This strategy increases thermal diversity while avoiding the introduction of OGT-associated noise into the low-stability class. Leveraging this dataset, we developed three feature-specific deep learning classifiers that separately combine global self-attention encoders with local convolutional residual modules, using sequence–evolutionary descriptors, ESM-2 contextual embeddings, or ProteinMPNN structural embeddings, respectively. Among these models, the ESM-2–based classifier — termed ThermoESM — achieved the strongest overall performance and exhibited excellent generalization on independent Limited Proteolysis–Mass Spectrometry (LiP–MS) datasets. Benchmarking against existing thermostability predictors further demonstrated ThermoESM’s superior accuracy, along with balanced sensitivity and precision. Collectively, these results establish ThermoESM as a robust and widely applicable framework for large-scale thermal stability annotation and thermostability-guided protein engineering.