TransTCR: Integrating TCRs and Transcriptomes through Optimal Transport for Antigen Specificity Prediction

TransTCR is an unsupervised multimodal representation learning framework that integrates single-cell TCR sequencing (scTCR-seq) and single-cell RNA sequencing (scRNA-seq) data through Optimal Transport (OT) for antigen specificity prediction and cell clustering.

TransTCR

The official implementation for "TransTCR".

Table of Contents

• Datasets
• Installation
• Usage

Datasets

The raw data can be downloaded here:

Dataset	Download
D1	Link
D2	Link
D3	Link
D4	Link

We provide easy access to the processed datasets in the Zenodo.

Installation

To reproduce TransTCR, we suggest first creating a conda environment by:

conda create -n TransTCR python=3.9.21
conda activate TransTCR

and then install the required packages below:

• scanpy=1.9.1
• scib=1.1.7
• scipy=1.13.2
• torch=2.6.0
• pot=0.9.5

Usage

Data Preprocessing

To reproduce TransTCR, paired scTCR-seq and scRNA-seq data in h5ad and csv formats must be processed.

• Process scTCR-seq Data

We use the pre-trained TCR-BERT to encode CDR3 sequences from both TCR chains (TCR-BERT must be downloaded separately):

cd Process/TCR
bash get_emb.sh

• Process scRNA-seq Data

We employ CellFM, a recently published foundational model for single-cell data, to process scRNA-seq data (CellFM must be downloaded separately).

Train and Evaluate

• Train and evaluate on intra-dataset classification and clustering:

bash run_Intra.sh

• Train and evaluate on inter-dataset classification and clustering:

bash run_Inter.sh

Citation

If you find our codes useful, please consider citing our work:

@article{TransTCR,
  title={TransTCR: Integrating TCRs and Transcriptomes through Optimal Transport for Antigen Specificity Prediction},
  author={Yuansong Zeng, Wenbing Li, Ruipeng Huang, Yuanze Chen, Jinyun Niu, Ningyuan Shangguan, Siyuan He, Yuedong Yang},
  journal={},
  year={2025},
}