MSigHdp: hierarchical Dirichlet process mixture modeling for mutational signature discovery. Issue 1 (23rd January 2023)

• Record Type: Journal Article
• Authors: Liu, Mo; Wu, Yang; Jiang, Nanhai; Boot, Arnoud; Rozen, Steven G
• Is Part Of: NAR genomics and bioinformatics. Volume 5:Issue 1(2023)
• Publication Date: 2023-01-23
• Subjects: Genomics -- Periodicals; Bioinformatics -- Periodicals; 572.8
• DOI: 10.1093/nargab/lqad005
• Languages: English
• ISSNs: 2631-9268

Abstract

Abstract: Mutational signatures are characteristic patterns of mutations caused by endogenous or exogenous mutational processes. These signatures can be discovered by analyzing mutations in large sets of samples—usually somatic mutations in tumor samples. Most programs for discovering mutational signatures are based on non-negative matrix factorization (NMF). Alternatively, signatures can be discovered using hierarchical Dirichlet process (HDP) mixture models, an approach that has been less explored. These models assign mutations to clusters and view each cluster as being generated from the signature of a particular mutational process. Here, we describe mSigHdp, an improved approach to using HDP mixture models to discover mutational signatures. We benchmarked mSigHdp and state-of-the-art NMF-based approaches on four realistic synthetic data sets. These data sets encompassed 18 cancer types. In total, they contained 3.5 × 10 7 single-base-substitution mutations representing 32 signatures and 6.1 × 10 6 small insertion and deletion mutations representing 13 signatures. For three of the four data sets, mSigHdp had the best positive predictive value for discovering mutational signatures, and for all four data sets, it had the best true positive rate. Its CPU usage was similar to that of the NMF-based approaches. Thus, mSigHdp is an important and practical addition to the set of tools available for discovering mutational signatures.