FINE-T

Predict T-cell response for peptide sequences with FINE-T, the fine-tuned transformer model of the paper "Transfer learning for T-cell response prediction" by Josua Stadelmaier, Brandon Malone, and Ralf Eggeling, published in BMC Bioinformatics.

FINE-T is fine-tuned per peptide source and MHC class, and this package provides the models for human peptides of MHC class I and MHC class II. The models only take the peptide sequence as input.

For training the model and for reproducing the domain-aware evaluation of the paper, use the scripts of the repository.

Installation

pip install fine-t-cell

This provides the fine-t-cell command and the finetcell Python package.

For a faster installation, use PyTorch without CUDA support since the FINE-T models are small enough to run on a CPU:

pip install --index-url https://download.pytorch.org/whl/cpu torch numpy
pip install fine-t-cell

Usage

fine-t-cell --peptides peptides.csv --mhc I --source human

The file with the peptides is a CSV file with a peptide column (Epitope Description or peptide), a CSV file with a single column, or a text file with one peptide per line. The predictions are written next to it, or to the path given by --predictions_file.

--source currently only accepts human, since the models that are fine-tuned on other peptide sources are not part of this package.

--mhc is either I or II and selects the model trained specifically for the respective MHC class. So use as input the peptides of one MHC class at a time.

From Python:

import finetcell

predictions = finetcell.predict(['GILGFVFTL', 'NLVPMVATV'], mhc='I', source='human')

finetcell.predict returns a DataFrame with the peptides and their predictions.

Peptides need to consist of the 20 standard amino acids and they can be at most 28 amino acids long, which is the longest peptide that the positional encodings of the model support. Peptides that contain other characters, such as X, and peptides that are longer are printed with the reason for skipping them, and they are left out of the predictions.

Tests

The tests check that the package reproduces the predictions of the repository scripts. They are part of the source code, so run them from a clone of the repository:

pip install -e .[test]
pytest tests

License

FINE-T is licensed under the GNU Affero General Public License v3.0, see LICENSE.

Citation

Stadelmaier, J., Malone, B. & Eggeling, R. Transfer learning for T-cell response prediction. BMC Bioinformatics 27, 124 (2026). https://link.springer.com/article/10.1186/s12859-026-06492-2