ImmunoMatch

ImmunoMatch is a machine learning framework for deciphering the molecular rules governing the pairing of antibody chains. Fine-tuned on an antibody-specific language model (AntiBERTA2), ImmunoMatch learns from paired H and L sequences from single human B cells to distinguish cognate H-L pairs and randomly paired sequences.

A total of three variants of ImmunoMatch, trained on different subsets of the data, are made available on huggingface:

Checkpoint name	Trained on
ImmunoMatch	A mixture of antibodies with both κ and λ light chains
ImmunoMatch-κ	Antibodies with κ light chains
ImmunoMatch-λ	Antibodies with λ light chains

Please note that the ImmunoMatch models are provided under a CC-BY-NC-4.0 license.

Try it out on Google Colab

Run_ImmunoMatch.ipynb contains example code on how to apply any ImmunoMatch model to obtain H-L pairing scores for a given VH-VL sequence pair, or to annotate sequences in batch upon supplying a CSV. You can also try it out on Google Collaboratory:

Requirements

There are no specific prerequisites to use ImmunoMatch beyond standard installation of Huggingface libraries on Python. On a clean virtual environment on Google Colab, the installation of these libraries took around 1 minute.

Figure reproducibility

Folder figure_code contains all Python and R code used to generate figure panels in the manuscript.

Cite

If you have used any of the ImmunoMatch models in your research please cite:

@article {Guo2025.02.11.637677,
	author = {Guo, Dongjun and Dunn-Walters, Deborah K and Fraternali, Franca and Ng, Joseph CF},
	title = {ImmunoMatch learns and predicts cognate pairing of heavy and light immunoglobulin chains},
	elocation-id = {2025.02.11.637677},
	year = {2025},
	doi = {10.1101/2025.02.11.637677},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/02/15/2025.02.11.637677},
	eprint = {https://www.biorxiv.org/content/early/2025/02/15/2025.02.11.637677.full.pdf},
	journal = {bioRxiv}
}