Abstract 2491: Conditional generation of mutational profiles in cancer

Abstract The accumulation of DNA damage on the cancer genome, often summarized as mutational profiles, is influenced by a complex interplay of environmental and cellular mutagenic conditions. Despite cancer genomics having some of the largest biological datasets in existence collected by pan-cancer consortiums, a significant portion of the diverse spectrum of cancer etiologies is still underrepresented and understudied. We leveraged the power of generative artificial intelligence (Gen AI) to develop MutaGen, a conditional variational autoencoder and generative adversarial network hybrid model (CVAE-GAN), to generate realistic mutational profiles conditioned on various clinical and genetic features. Trained on the PCAWG cohort, MutaGen can generate pan-cancer mutational profiles across four mutation types: single base substitutions (SBS96), small insertions and deletions (ID83), structural variations (SV32), and copy number alterations (CN48). MutaGen's cancer-conditioned generations show high similarities to the real PCAWG samples, with median cosine similarities of 0.98, 0.97, 0.92, and 0.96 for SBS96, ID83, SV32, and CN48 respectively. When conditioned on homologous recombination deficiency (HRD) status, MutaGen can generate zero-shot (i.e. cases not seen in training set) HRD lung cancers with high similarity to a real BRCA2-null lung cancer patient in both SBS96 and ID83 (0.93 and 0.78 cosine similarities respectively). The generation is significantly more similar than non-HRD lung cancer or non-lung HRD cancer in the PCAWG training set. Inspection of the generated profiles revealed the generations harbouring both lung cancer-associated and HRD-associated mutational patterns, showing that MutaGen can generate novel and biologically plausible profiles. By providing a versatile tool for generating synthetic mutational data, MutaGen can facilitate the exploration of understudied cancer subtypes and enable the training and benchmarking of predictive models. Citation Format: Andy Wu, Hannan Wong, Aakansha Narain, Vedant Sandhu, Jason Pitt. Conditional generation of mutational profiles in cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2491.