Cancer is caused by germline and somatic mutations, which can share biological features such as amino acid change. However, integrated germline and somatic analysis remains uncommon. We present a framework that uses machine learning to learn features of recurrent somatic mutations to (1) predict somatic variants from tumor-only samples and (2) identify somatic-like germline variants for integrated analysis of tumor-normal DNA. Using data from 1769 patients from seven cancer types (bladder, glioblastoma, low-grade glioma, lung, melanoma, stomach, and pediatric glioma), we show that "somatic-like" germline variants are enriched for autosomal-dominant cancer-predisposition genes (p < 4.35 × 10-15), including TP53. Our framework identifies germline and somatic nonsense variants in BRCA2 and other Fanconi anemia genes in 11% (11/100) of bladder cancer cases, suggesting a potential genetic predisposition in these patients. The bladder carcinoma patients with Fanconi anemia nonsense variants display a BRCA-deficiency somatic mutation signature, suggesting treatment targeted to DNA repair.