Protein melting temperature (Tm) prediction accelerates the discovery of thermostable enzymes crucial for industrial biotechnology, where proteins must endure harsh reaction conditions. Experimental determination of Tm remains labour-intensive and varies across techniques, motivating the development of in silico predictors. Recent advances in high-throughput mass spectrometry have led to large-scale proteomics-based Tm datasets enabling effective training of machine learning models. However, the generalisability of such tools across diverse proteomics- and biophysics-based datasets remains an open question.
This study aims to (i) assemble comprehensive Tm datasets including high-quality biophysics sources for independent evaluation, (ii) evaluate generalisability of state-of-the-art approaches based on deep learning, ESM-2 sequence embeddings, and parameter-efficient low-rank adaptation (LoRA), and (iii) explore ESM-3 structural embeddings as a counterpart to sequence-based approaches.
We assembled the ProMelt dataset (45,441 proteins) from Meltome Atlas and ProThermDB and trained baseline and advanced embedding-based predictors. Models were evaluated on five independent biophysics-based datasets derived from BRENDA, FireProtDB, and literature.
Our analysis revealed substantial inconsistencies in reported Tm values between proteomics- and biophysics-based measurements, highlighting the need for generalizable predictors. Fine-tuned embedding-based models showed competitive performance compared to DeepSTABp, TemBERTure, and SaProt and achieved superior performance in binary classification of thermostable proteins (Tm ≥ 60 °C): ESM2-LoRA achieved AUC = 0.75, while ESM3-MLP achieved AUC = 0.77.
The best-performing model is deployed as TmProt, a user-friendly web server on Hugging Face.
Together, these results show that LoRA-adapted ESM-2 and ESM-3 structural embeddings improve thermostability prediction and emphasize the importance of rigorous evaluation across diverse experimental methodologies.

Co-authors: None