Training and evaluation datasets for the TmProt tool

These datasets were used for training and evaluating TmProt 1.0, a protein melting temperature (Tm) predictor based on LoRA-adapted ESM-2 embeddings. The source code is available at the GitHub repository.

ProMelt: Training Dataset

ProMelt (train_promelt_seq.csv, val_promelt_seq.csv, test_promelt_seq.csv) comprises 45,441 proteins assembled from two proteomics-based sources: Meltome Atlas (Jarzab A et al., 2020) and ProThermDB (Nikam R et al., 2021). Sequences were filtered to lengths of 20–2,000 amino acids and deduplicated by UniProt ID. Proteins overlapping with the independent validation sets were removed before training. The dataset was split at 25% maximum sequence identity using USEARCH into training (77.1%, n = 35,054), validation (8.6%, n = 3,895), and test (14.3%, n = 6,492) subsets to prevent data leakage.

Independent Biophysics-Based Validation Datasets

Five datasets assembled from low-throughput biophysical experiments, used exclusively for benchmarking:

• BRENDA (BRENDA.csv, n = 321): Wild-type enzymes with Tm values manually curated from the BRENDA enzyme database (Chang A et al., 2021), accessed January 2025. Filtered to neutral pH, wild-type entries only, and measurements free of exogenous ligands, ions, or apoenzyme annotations.
• FireProt (FIREPROT.csv, n = 94): Wild-type protein Tm values extracted from FireProtDB (Stourac J et al., 2021), accessed June 2023. A manually curated thermostability database of literature-derived and experimentally validated entries.
• ERED_WT / ERED_ASR (ERED_WT.csv, n = 39; ERED_ASR.csv, n = 44): Ene reductases (390–417 residues) comprising wild-type proteins and variants designed via ancestral sequence reconstruction (ASR) for enhanced thermostability (Livada J et al., 2023). Tm measured by differential scanning fluorimetry (DSF); range 37–67°C.
• CAS (CAS.csv, n = 29): CRISPR-Cas Class II effector proteins (1,049–1,535 residues) (Pudžiuvelytė I et al., 2024). Tm values span 33–67°C, measured by nano differential scanning fluorimetry (nanoDSF).
• HLD (HLD.csv, n = 24): Haloalkane dehalogenases (271–336 residues) obtained through genome mining (Vasina M et al., 2022). Tm values range 35.2–58.7°C, measured by DSF.

For more details about the tool, please get in touch with us.