Decoding whole-genome mutational signatures in 37 human pan-cancers by denoising sparse autoencoder neural network

• Published in: Oncogene 2020-07, Vol.39 (27), p.5031-5041
• Main Authors: Pei, Guangsheng, Hu, Ruifeng, Dai, Yulin, Zhao, Zhongming, Jia, Peilin
• Format: Article
• Language: English
• Subjects: 45/23; 45/70; 631/1647/514; 631/67; 692/53/2423; 96/47; Apoptosis; Base excision repair; Cancer; Carcinogenesis; Cell Biology; Cisplatin; Deep Learning; DNA - genetics; Gene expression; Gene mutations; Genetic aspects; Genome, Human - genetics; Genomes; Health aspects; Human Genetics; Humans; INDEL Mutation - genetics; Internal Medicine; Medicine; Medicine & Public Health; Mutagenesis; Mutation; Neoplasms - genetics; Neural networks; Oncology; Oncology, Experimental; Phenotypes; Point Mutation - genetics; Whole genome sequencing
• ISSN: 0950-9232; 1476-5594
• DOI: 10.1038/s41388-020-1343-z

Millions of somatic mutations have recently been discovered in cancer genomes. These mutations in cancer genomes occur due to internal and external mutagenesis forces. Decoding the mutational processes by examining their unique patterns has successfully revealed many known and novel signatures from whole exome data, but many still remain undiscovered. Here, we developed a deep learning approach, DeepMS, to decompose mutational signatures using 52,671,908 somatic mutations from 2780 highly curated cancer genomes with whole genome sequencing (WGS) in 37 cancer types/subtypes. With rigorous model training and comparison, we characterized 54 signatures for single base substitutions (SBSs), 11 for doublet base substitutions (DBSs) and 16 for small insertions and deletions (Indels). Compared to the previous methods, DeepMS could discover 37 SBS, 5 DBS, and 9 Indel new signatures, many of which represent associations with DNA mismatch or base excision repair and cisplatin therapy mechanisms. We further developed a regression-based model to estimate the correlation between signatures and clinical and demographical phenotypes. The first deep learning model DeepMS on WGS somatic mutational profiles enable us identify more comprehensive context-based mutational signatures than traditional NMF approaches. Our work substantially expands the landscape of the naturally occurring mutational signatures in cancer genomes, and provides new insights into cancer biology.