P-TEAM is a Random Forest-based prediction model to estimate the effect of mutations in epitopes on the TCR activation. Depending on the provided label, the model can classify TCR activation binary, or predict a continuous binding score. While we observed that a total of 25% (ca. n=38) of randomly sampled mutations are sufficient for high-performance prediction, the number of samples can further be reduced to 24 mutations by employing iterative experimental design and our active learning framework
To recreate the prediction results of the paper:
git clone https://github.com/SchubertLab/TcrPrediction_MutatedAPLs.git
cd TCRPrediction_MutatedAPLs
conda env create -f requirements.yml
sh ./activation-prediction/run_all.sh
Alternatively, you can work with the essential packages numpy==1.20.2, pandas==1.2.5, scikit-learn=0.24.2, scipy==1.6.2, seaborn==0.11.1.
For the active learning results, run activation-prediction/active_learning/al_tcr_specific.ipynb for both epitopes.
The positional distances were calculated within the pymol software via the script Modelling3D/PositonalDistances.py on the structural models provided in the Supplementary Data 6-7 of the paper.
To recreate the baseline results, refer to the notebook baseline/SOTA_Comparisson.ipynb and the GitHub repositories of ERGO-II link and ImRex link.
Tutorials are provided for:
- novel predictions for a specific TCR and active learning (
tutorials/01_within_tcr.ipynb) - a full mutation profile of a novel TCR (
tutorials/02_across_tcr.ipynb)
If P-TEAM is helpful in your research, please consider citing the following paper:
@article{dorigatti2023predicting,
title={Predicting T Cell Receptor Functionality against Mutant Epitopes},
author={Dorigatti, Emilio and Drost, Felix and Straub, Adrian and Hilgendorf, Philipp and Wagner, Karolin Isabel and Bischl, Bernd and Busch, Dirk and Schober, Kilian and Schubert, Benjamin},
journal={bioRxiv},
pages={2023--05},
year={2023},
publisher={Cold Spring Harbor Laboratory}
}