Projects Research + code
TCR Pathology Classifier
A protein language model that ranks T-cell receptors by the disease they likely recognize, so the lab screens a short list instead of a whole repertoire.
- When
- Sep 2025 – Mar 2026
- My role
- Built it end to end: data, modeling, deployment, and validation
- Built with
- PyTorch, ESM2, Hugging Face, Flask, pandas, scikit-learn
- 35% validated hit rate among model-prioritized clones (11 of 31)
- ~10× smaller search space, from ~7,000 clones to 572
- 8 wk → 10 d from raw repertoire to a testable shortlist
Project introduction
The problem
Disease-reactive T-cell receptors are rare — some of the clones we care about sit at 0.0002% of a repertoire. Finding them the classic way takes about eight weeks of antigen-driven enrichment before anything reaches a functional assay.
Approach
I asked whether sequence alone could tell us where to look first.
- Data. Paired α/β TCRs with pathology labels from the curated McPAS-TCR database, cleaned and split 70/15/15, with attention to class imbalance.
- Model. Paired TRA and TRB CDR3 sequences are embedded with ESM2 (650M parameters), fine-tuned end to end, combined through cross-attention, and fused with learned V-gene embeddings before a classifier head.
- Comparison. I benchmarked three designs — frozen embeddings with an MLP, a binary reactive/non-reactive model, and the fine-tuned model — and kept the fine-tuned one as the best performer.
- Delivery. A Flask web app for point-and-click inference, with model weights on Hugging Face.
Validation at the bench
In a proof of concept, a 6,737-clone repertoire from a healthy donor was stimulated with the MART-1 melanoma antigen. The model narrowed it to 572 melanoma-associated candidates. We advanced 31 to functional testing and 11 were confirmed MART-1-reactive — a 35% hit rate, including clones at 0.0002% frequency — and the time from raw repertoire to shortlist fell from about eight weeks to ten days.
What’s next
Extending the classifier toward open-set TCR–peptide–MHC ranking with bidirectional cross-attention and a contrastive objective, so the model can score antigens it has never seen.