AutoTFCNNY: A multi-instance neural network for enhanced early cancer detection using TCR data

Abstract

For most cancers, early diagnosis and intervention can significantly improve cure rates and patient survival. Consequently, achieving early and accurate cancer detection has always been a central focus in both medical practice and scientific research. Recently, studies based on peripheral blood T-cell receptors (TCRs) have attracted considerable attention due to their noninvasiveness and potential for high sensitivity. It has been reported that cancer-associated TCRs (caTCRs) exist in the peripheral blood of cancer patients, suggesting that discerning whether a TCR repertoire is associated with cancer provides a viable strategy for early cancer prediction. However, extracting crucial cancer-related information from a large and heterogeneous TCR repertoire remains a major challenge.To address this issue, we propose AutoTFCNNY, a multi-instance deep neural network model that combines a Transformer and a convolutional neural network (CNN). Built upon a multi-instance learning (MIL) framework, AutoTFCNNY leverages the Transformer’s global dependency modeling alongside the CNN’s local feature enhancement to effectively extract TCR sequence features, thereby significantly improving early cancer detection accuracy. Experimental results demonstrate that AutoTFCNNY performs well in detecting 22 different cancer types, achieving an average area under the ROC curve (AUC) exceeding 0.94. Notably, in 18 of these types—including brain cancer and non-small cell lung cancer et al.—the average AUC surpasses 0.99. These findings indicate that AutoTFCNNY possesses high accuracy, stability, and favorable generalization ability, suggesting its potential as a non-invasive tool for early cancer detection based on peripheral blood TCR repertoires.