Benchmarking unsupervised methods for inferring TCR specificity

对用于推断TCR特异性的无监督方法进行基准测试

阅读:1

Abstract

Identifying T-cell receptor (TCR) specificity is crucial for advancing the understanding of adaptive immunity. Despite the development of computational methods to infer TCR specificity, their clustering behavior has not been thoroughly compared. We addressed this by curating a unified database of 190 670 human TCRs with known specificities for 2313 epitopes across 121 organisms, combining data from IEDB, McPAS-TCR, and VDJdb. We asked whether widely used TCR clustering methods produce comparable results on the same high-confidence dataset. We hypothesized that shared assumptions about conserved CDR3 motifs would yield similar patterns, with differences reflecting algorithmic design. Nine methods for clustering TCRs based on similarity were benchmarked against this dataset. DeepTCR demonstrated the best retention, while ClusTCR, TCRMatch, and GLIPH2 excelled in cluster purity but had lower retention. GLIPH2, Levenshtein distance, Hamming distance, and clusTCR generated large clusters in contrast to TCRMatch and DeepTCR. Smaller, antigen-specific clusters were produced by GIANA and iSMART. DeepTCR was the most sensitive in capturing antigen-specific TCRs. We confirmed these observations using a larger dataset from 10X Genomics containing antigen-specific labeled TCRs as well non-labeled cells. This study offers a unified TCR database and a benchmark of specificity inference methods, guiding researchers in selecting appropriate tools.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。