DeepLION2 - Deep multi-instance contrastive learning framework enhancing the prediction of cancer-associated T cell receptors

Sep 21, 2026

Uploaded by Finn

DeepLION2 presents a deep multi-instance contrastive learning framework for predicting cancer-associated T cell receptors from TCR repertoire sequencing data. The method models inter-TCR correlations using content-based sparse self-attention that focuses on the top k related TCRs for each TCR. It further applies contrastive learning to bootstrap attention-matrix updates, reducing overfitting to non-cancer-associated signals. Extensive experiments across more than ten cancer types show consistent gains in accuracy, sensitivity, specificity, MCC, and AUC, while also identifying cancer-associated TCRs and their key motifs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages14

Reading Time35 min

AccessFree to download

Introduction

Methods

Results

Conclusion

Keywords

1 Introduction

TYPE Original Research PUBLISHED 07 March 2024 DOI 10.3389/fimmu.2024.1345586 OPEN ACCESS EDITED BY Yi Shi, Shanghai Jiao Tong University, China REVIEWED BY Scott Christley, University of Texas Southwestern Medical Center, United States Chong Chu, Harvard Medical School, United States *CORRESPONDENCE Jiayin Wang [wangjiayin@mail.xjtu.edu.cn](wangjiayin@mail.xjtu.edu.cn)[ ](wangjiayin@mail.xjtu.edu.cn)Tao Wang [tddocwangt@163.com](tddocwangt@163.com) †These authors have contributed equally to this work RECEIVED 28 November 2023 ACCEPTED 19 February 2024 PUBLISHED 07 March 2024 CITATION Qian X, Yang G, Li F, Zhang X, Zhu X, Lai X, Xiao X, Wang T and Wang J (2024) DeepLION2: deep multi-instance contrastive learning framework enhancing the prediction of cancer-associated T cell receptors by attention strategy on motifs. Front. Immunol. 15:1345586 . doi: 10.3389/fimmu.2024.1345586 COPYRIGHT © 2024 Qian, Yang, Li, Zhang, Zhu, Lai, Xiao, Wang and Wang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. DeepLION2: deep multi-instance contrastive learning framework enhancing the prediction of cancer-associated T cell receptors by attention strategy on motifs Xinyang Qian 1,2†, Guang Yang 3†, Fan Li 1,2, Xuanping Zhang 1,2, Xiaoyan Zhu 1,2, Xin Lai 1,2, Xiao Xiao 4, Tao Wang 5* and Jiayin Wang 1,2* 1School of Computer Science and Technology, Xi’an Jiaotong University, Xi’an, China, 2Shaanxi Engineering Research Center of Medical and Health Big Data, Xi’an Jiaotong University, Xi’an, China, 3 Department of Clinical Oncology, The Second Afﬁliated Hospital of Air Force Medical University, Xi’an, China, 4Genomics Institute, Geneplus-Shenzhen, Shenzhen, China, 5 Department of Thoracic Surgery, The Second Afﬁliated Hospital of Air Force Medical University, Xi ’an, China Introduction: T cell receptor (TCR) repertoires provide valuable insights into complex human diseases, including cancers. Recent advancements in immune sequencing technology have signiﬁcantly improved our understanding of TCR repertoire. Some computational methods have been devised to identify cancerassociated TCRs and enable cancer detection using TCR sequencing data. However, the existing methods are often limited by their inadequate consideration of the correlations among TCRs within a repertoire, hindering the identiﬁcation of crucial TCRs. Additionally, the sparsity of cancer-associated TCR distribution presents a challenge in accurate prediction. Methods: To address these issues, we presented DeepLION2, an innovative deep multi-instance contrastive learning framework speciﬁcally designed to enhance cancer-associated TCR prediction. DeepLION2 leveraged content-based sparse self-attention, focusing on the top k related TCRs for each TCR, to effectively model inter-TCR correlations. Furthermore, it adopted a contrastive learning strategy for bootstrapping parameter updates of the attention matrix, preventing the model from ﬁxating on non-cancer-associated TCRs. Results: Extensive experimentation on diverse patient cohorts, encompassing over ten cancer types, demonstrated that DeepLION2 signiﬁcantly outperformed current state-of-the-art methods in terms of accuracy, sensitivity, speciﬁcity, Matthews correlation coefﬁcient, and area under the curve (AUC) . Notably, DeepLION2 achieved impressive AUC values of 0.933, 0.880, and 0.763 on thyroid, lung, and gastrointestinal cancer cohorts, respectively. Furthermore, it effectively identiﬁed cancer-associated TCRs along with their key motifs, highlighting the amino acids that

FAQs

What problem does DeepLION2 address in cancer-associated TCR prediction?

It targets limitations of existing methods that inadequately model correlations among TCRs in a repertoire and that face sparsity in the distribution of cancer-associated TCRs.

How does DeepLION2 model relationships between TCRs?

It uses content-based sparse self-attention, focusing on the top k related TCRs for each TCR to capture inter-TCR correlations.

What is the purpose of the contrastive learning strategy in DeepLION2?

It bootstraps parameter updates of the attention matrix, helping prevent the model from fixating on TCRs that are not associated with cancer.

Sep 21, 2026

Uploaded by Finn