Lee-CBG / ActiveTCR

Official Tensorflow implementation for ActiveTCR: Active Learning Framwork for Cost-Effective TCR-Epitope Binding Affinity Prediction

ActiveTCR: An Active Learning Framework for Cost-Effective TCR-Epitope Binding Affinity Prediction

ActiveTCR is a unified framework designed to minimize the annotation cost and maximize the predictive performance of T-cell receptor (TCR) and epitope binding affinity prediction models. It incorporates active learning techniques to iteratively search for the most informative unlabeled TCR-epitope pairs, reducing annotation costs and redundancy. By leveraging four query strategies and comparing them to a random sampling baseline, ActiveTCR demonstrates significant cost reduction and improved performance in TCR-epitope binding affinity prediction. ActiveTCR is the first systematic investigation of data optimization in the context of TCR-epitope binding affinity prediction.

Publication

An Active Learning Framework for Cost-Effective TCR-Epitope Binding Affinity Prediction
Pengfei Zhang^1,2, Seojin Bang³, Heewook Lee^{1,2, *
1} School of Computing and Augmented Intelligence, Arizona State University, ² Biodesign Institute, Arizona State University, ³ Google DeepMind
Accepted for publication: IEEE BIBM 2023

Paper | Code | Poster | Slides | Presentation (YouTube)

Major results of ActiveTCR

1. Use case a: reducing more than 40% annotation cost for unlabel TCR-epitope pools.

2. Use case b: minimizing more than 40% redundancy among already annotated TCR-epitope pairs.

Dependencies

• Linux
• Python 3.6.13
• Keras 2.6.0
• TensorFlow 2.6.0

Steps to train a Binding Affinity Prediction model for TCR-epitope pairs.

1. Clone the repository

git clone https://github.com/Lee-CBG/ActiveTCR
cd ActiveTCR/
conda create --name bap python=3.6.13
pip install -r requirements.txt
source activate bap

2. Prepare TCR-epitope pairs for training and testing

• Download training and testing data from datasets folder.
• Obtain embeddings for TCR and epitopes following instructions of catELMo. Or directly download embeddings from Dropbox.

3. Train and test models

An example for use case a of ActiveTCR: reducing annotation cost for unlabeled TCR-epitope pools.

python -W ignore main.py \
                --split epi \
                --active_learning True \
                --query_strategy entropy_sampling \
                --train_strategy retrain \
                --query_balanced unbalanced \
                --gpu 0 \
                --run 0

An example for use case b of ActiveTCR: minimizing redundancy among labeled TCR-epitope pairs.

python -W ignore main.py \
                --split epi \
                --query_strategy entropy_sampling \
                --train_strategy retrain \
                --query_balanced unbalanced \
                --gpu 1 \
                --run 0

Citation

If you use this code or use our catELMo for your research, please cite our paper:

License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.