Use TemStaPro Online
Commercially Available TemStaPro No-Code Web Server
TemStaPro: Protein Thermostability Prediction from Sequence Using Protein Language Models
TemStaPro (Temperatures of Stability for Proteins) is a tool for predicting protein thermostability from amino acid sequence using embeddings from large protein language models. It was developed by Ieva Pudziuvelyte, Kliment Olechnovic, Darius Kazlauskas and colleagues at Vilnius University and CasZyme, and published in Bioinformatics in 2024. The article is open access under a CC BY license, and the software is available on GitHub.
Knowing whether a protein stays folded at high temperature matters for enzyme engineering and industrial biotechnology. Experimental measurements are slow and costly, and earlier sequence-based predictors were trained on comparatively small datasets. The authors note that BertThermo used 2,803 sequences and ProLaTherm 7,409, whereas TemStaPro uses transfer learning from protein language models trained on hundreds of millions of sequences and is trained on over one million sequences from organisms with annotated growth temperatures.
Property | Detail |
|---|---|
Task | Sequence-based classification of protein thermostability at several temperature thresholds |
Features | Mean ProtT5-XL embeddings (vectors of length 1024) |
Model | Multi-layer perceptron with hidden layers of 256 and 128 units |
Training data | Proteomes of non-eukaryotic organisms with annotated optimal growth temperatures; 943,605 training sequences in the major dataset |
Thresholds | Nine binary classifiers from 40 to 80 C; default mode uses six (40 to 65 C) |
Code and data | GitHub repository and Zenodo archive |
How TemStaPro Predicts Thermostability
Embedding: each protein sequence is converted to a mean ProtT5-XL embedding. The authors chose ProtT5-XL over ESM-2 because it performed similarly while being faster to compute and shorter.
Binary classifiers: a neural network is trained for each temperature threshold, and each threshold in default mode is an ensemble of five networks.
Labels: proteome-level optimal growth temperatures serve as lower bounds on the thermostability of each protein. Sequences were clustered at 30% identity so training, validation and test sets share no clusters.
Output interpretation: left-hand and right-hand labels are derived by scanning the six binary predictions from each side; agreement between them indicates the highest temperature range at which the protein is predicted to be thermostable.
Prediction Modes
TemStaPro can produce global, per-residue or per-segment (window-averaged) predictions. An alternative mode adds 70, 75 and 80 C classifiers, which the authors regard as less reliable because positive examples are scarce.
TemStaPro Performance Versus Other Thermostability Predictors
Evaluation | Result |
|---|---|
Imbalanced test set, best threshold | Maximum MCC 0.638 at 50 C |
Balanced test subsets | Outperformed BertThermo and ProLaTherm at every threshold; maximum MCC 0.903 at 80 C |
SAPPHIRE benchmark (742 sequences) | TemStaPro-t70 reached MCC 0.919 vs. 0.884 for SAPPHIRE |
iThermo benchmark (562 sequences) | TemStaPro-t65 reached MCC 0.972 vs. 0.950 for BertThermo |
The authors also applied TemStaPro to 16,376 CRISPR-Cas Class II effector protein sequences. For example, Cas12a had no members predicted thermostable at 60 C or above, while more than half of Cas12b members were predicted to function at 45 C or above. Of 39 comparisons with experimental data, 35 (90%) were within 10 C, and nanoDSF measurements on Ghy2Cas9, CbaCas12b and SpyCas9 supported the predictions.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with TemStaPro on Tamarind Bio
Enzyme screening: rank candidate enzymes by predicted thermostability before expressing them.
Protein engineering: compare variants and identify regions with lower predicted stability using per-residue or per-segment output.
Genome and metagenome mining: screen large sequence sets, as the authors did for CRISPR-Cas effectors.
Hit prioritization: choose which proteins to characterize experimentally with methods such as nanoDSF.
How to Use TemStaPro on Tamarind Bio
Log in: sign in at tamarind.bio and open the TemStaPro tool.
Provide sequences: enter one or more protein sequences; TemStaPro takes amino acid sequences as input.
Choose the output type: select global, per-residue or per-segment predictions, if exposed.
Choose the threshold mode: use the default six thresholds (40 to 65 C) or the extended mode, if exposed.
Run the job: submit it and wait for predictions.
Download results: retrieve per-threshold predictions and any plots.
Interpret: read the temperature range at which each protein is predicted to be thermostable.
Follow up: shortlist proteins for experimental validation or engineering.
Things to Keep in Mind
Training labels are lower bounds, so the authors describe TemStaPro as more pessimistic than optimistic and likely to underestimate the highest stable temperature.
Each threshold classifier is independent, so outputs can conflict.
The 70, 75 and 80 C classifiers rely on few positive examples and are excluded from default mode.
In rare cases predictions were more than 10 C below experimental values.
Experimental validation covered only three enzymes.
Speed depends on GPU availability; the authors report about 10,000 sequences of average length 1,000 residues in under 2 hours on one GPU.
Source: Pudziuvelyte I, Olechnovic K, Godliauskaite E, Sermokas K, Urbaitis T, Gasiunas G, Kazlauskas D. TemStaPro: protein thermostability prediction using sequence representations from protein language models. Bioinformatics 40(4):btae157, 2024. DOI: 10.1093/bioinformatics/btae157. Code: github.com/ievapudz/TemStaPro. Data: 10.5281/zenodo.7743637.