Protein melting temperature prediction from sequence, pairing frozen ESM2 embeddings with growth-temperature and thermophilicity features.
No providers recorded yet. Browse all providers
ESMStabP predicts a protein's melting temperature (Tm) directly from its amino acid sequence. It was developed by Marcus Ramos, Robert L. Jernigan, and Mesih Kilinc at Iowa State University and posted to bioRxiv in February 2025.
Thermostability determines whether a protein survives the temperatures encountered during cultivation, processing, and storage, which makes it a gating property for industrial enzymes, biologics formulation, and food science applications. Measuring it experimentally is slow and costly: differential scanning calorimetry and circular dichroism spectroscopy are labor-intensive per protein, and the resulting scarcity of labeled data has been the field's binding constraint as much as any modeling limitation.
ESMStabP builds directly on DeepSTABp, an earlier regression model for the same task, and makes two changes. First, it replaces the sequence representation with embeddings from a frozen ESM-2 protein language model, and identifies empirically which layer of that model carries the relevant signal. Second, it enlarges the training data by consolidating and cleaning the datasets used by several prior thermostability models into a single deduplicated corpus, which addresses the data scarcity problem rather than working around it.
Sequences are encoded with the esm2_t33_650M_UR50 checkpoint and truncated to a maximum of 1,022 residues to fit the model's context. Layer-33 output logits are pooled into a fixed-dimension embedding, concatenated with the three auxiliary features, and passed to a random forest regressor. Labels come from high-throughput mass-spectrometry thermal proteome profiling assays; the thermophilic classification follows the convention used by TemBERTure, with growth temperature above 60°C marking thermophilic and below 30°C non-thermophilic. The corpus was split 80/20 for training and testing, with 5-fold cross-validation.
In a head-to-head comparison where all models were retrained on the identical balanced dataset, ESMStabP reached R² 0.94 and Pearson correlation 0.92, against DeepSTABp at 0.81 and 0.88 and ProTstab2 at 0.51 and 0.68. On those balanced-set error metrics ESMStabP's MAE of 3.42 was better than DeepSTABp's 3.62, though its MSE and RMSE were slightly worse. Trained on the original unbalanced corpus it improved across all five metrics, reaching MSE 13.71, RMSE 3.70, and MAE 2.79. The paper's headline figures are R² 0.95 and PCC 0.97. Ablations showed that dropping either optimal growth temperature or the thermophilic class reduced PCC to 0.87 and 0.90 respectively, so both contribute despite being closely related quantities.
The practical use is triage in protein engineering: ranking candidate enzyme variants by predicted stability so that only the promising ones consume bench time, which matters most in directed evolution campaigns and industrial biocatalyst development where variant libraries are large. Proteome-scale stability surveys are another fit, since scoring is cheap enough to run across an organism's full complement of proteins. Formulation and process scientists can use predicted Tm to anticipate which proteins are at risk during heating steps. The hosted endpoint means groups without local GPU capacity can still obtain predictions.
ESMStabP's most transferable finding is not the model but the layer-isolation result: performance varied considerably across ESM2 layers, and the common practice of taking the final layer by default may leave signal unused in other property predictors built the same way. The architecture itself is deliberately modest — a random forest on frozen embeddings — which keeps it cheap to train and easy to interpret relative to end-to-end deep models. Several caveats apply. The work remains a preprint, not peer-reviewed. The code is public but unlicensed: the authors' repository holds the dataset-assembly pipeline, its three base corpora, the training script, and LoRA fine-tuning experiments, and a companion Flask repository ships the four fitted regressors — but neither carries a licence, so the artifacts are inspectable rather than legally reusable. The reported R² and PCC figures differ between the abstract and the comparison table, and the auxiliary features it depends on — optimal growth temperature in particular — are not always available for an arbitrary sequence of interest. Finally, training labels derive from thermal proteome profiling, whose melting temperatures reflect cellular context and do not always match purified-protein measurements.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.