TCRCan: Pan-cancer Prediction via Fine-Tuned ESM-2 Using Peripheral Blood TCRβ Sequences
Xin Yang, Xing Wei, Kaiyu Fu, Zhengwei Wan, Huaichao Luo, Jian HuangIntroduction:
Malignant tumors are a leading cause of mortality worldwide, underscoring the need for accurate and non-invasive diagnostic methods. T cell Receptors (TCRs) are promising biomarkers due to their antigen specificity and alterations in response to tumors.
Methods:
Here, we introduced TCRCan, a deep learning framework for pan-cancer prediction using TCRβ sequences from peripheral blood T cells. TCRCan was constructed by fine-tuning ESM-2, a protein language model, and integrates a Cancer Risk Index (CRI) to quantify individual cancer risk based on cancer-associated TCR frequencies.
Results:
Using 1,294 samples spanning nine solid tumors, two leukemia subtypes, non-cancer diseases, and healthy donors, TCRCan demonstrated robust performance and strong generalization, accurately identifying melanoma and two pediatric leukemias excluded from training. Non-cancer conditions did not significantly affect TCRCan-derived CRI. Additionally, TCRCan showed excellent performance in detecting early-stage non-small cell lung cancer.
Discussion:
By harnessing immune repertoire signals, TCRCan offers a minimally invasive alternative to conventional imaging and tissue-based diagnostics, though further validation in larger cohorts is warranted.
Conclusion:
TCRCan’s non-invasive nature supports large-scale screening applications. These findings establish TCRCan as a promising tool for cancer detection.