[A benchmark dataset for evaluating antibody V(D)J gene reconstruction algorithms]

Accurate reconstruction of V(D)J recombination is fundamental to deciphering antibody diversity and critically underpins downstream analyses, including allele assignment and diversity estimation. Despite the development of multiple reconstruction algorithms, systematic performance comparison has been hampered by the lack of a standardized benchmark that integrates experimentally validated clonal sequences with diverse simulated data. In this study, we constructed a multi-source dataset comprising experimental clonal data, immuneSIM simulated data, and GenAIRR simulated data, with a total of over 110 000 sequences. The experimental component comprises B cell clones with validated rearrangements. The immuneSIM subset offers large-scale sequence data and includes complete D segment alignment sequences. The GenAIRR dataset, generated under two mutation models, carries comprehensive annotations for allele identity, somatic mutation frequency, and functional status. To evaluate the performance of the dataset, we evaluated five representative algorithms: IgBLAST, IMGT/HighV-QUEST, MiXCR, partis, and AlignAIR, against this dataset, assessing the performance at multiple dimensions: family-level and allele-level accuracy, full V(D)J segment assignment, and D-gene-stratified performance. The results showed that the dataset effectively discriminated algorithm performance and matched the key characteristics of real immune repertoires in terms of gene family distribution, CDR3 length, mutation rate, and D segment length. This benchmark dataset thus serves as a unified reference for algorithm selection, parameter optimization, and the development of new algorithms for antibody V(D)J recombination. Dataset profile.