Mutational signatures are increasingly used to understand the mechanisms causing cancer, predict prognosis and stratify patients for therapy. However, inference of mutational signatures can be error-prone, particularly in the case of featureless, low-sparsity signatures, which often get confounded. One of them is the homologous recombination deficiency-associated signature SBS3, relevant because of its association with prognosis in ovarian and breast cancer and because of its potential use as a biomarker for synthetic lethality therapies. Here, we present the multimodal method for mutational signature extraction, operating on single-base substitutions (SBS) and indels jointly, and highlight its accuracy signature identification and patient survival prediction. Across four different cohorts of whole-genome sequenced ovarian cancers, the multimodal SBS/indel approach correctly distinguished the commonly confused signatures SBS3, SBS8, SBS39, SBS40 and SBS5. Moreover, we identified two different multimodal m-SBS3 signatures, m-SBS3a and m-SBS3b, with distinct patterns in the indel spectrum. Specifically, the m-SBS3b signature was strongly predictive of better survival in high-grade serous ovarian cancer patients, replicating across the four cohorts, with effect sizes greatly exceeding other genetic markers of survival. m-SBS3 further predicted survival in platinum-treated patients with various cancer types, supporting a general utility of the multimodal mutational signatures for generating biologically and clinically meaningful readouts.