Abstract 7429: Illuminating the dark genome in cancer using artificial intelligence

Abstract Understanding the role of noncoding regions of the human genome in cancer requires innovative, high-resolution approaches. Artificial intelligence (AI) is transforming cancer research by providing advanced tools to identify, analyze, and interpret cancer genomes. This study aims to 1) discuss how noncoding mutations in key regulatory elements and other regions can contribute to cancer development, and 2) evaluate the leading AI models used for analyzing noncoding variants, focusing on their evolution and application in cancer research. We present examples of mutations in promoters, e.g. TERT [Huang et al., Science 2013] APC [Snow et al., Clin Gen 2015]), enhancers (a SNP in gliomas [Yanchus et al., Science 2022]), and super-enhancer hypermutation in B-cell lymphomas [Bal et al., Nature 2022; Leeman-Neil et al., Nat Gen 2023]. These cases demonstrate the functional impact of noncoding variants in tumorigenesis. However, for many noncoding regions and mutations, our understanding of the underlying mechanisms, binding motifs, and their roles in gene regulation remains limited. Here, AI can play a crucial role by analyzing large genomic datasets to uncover hidden patterns and interactions. This study compares AI models developed over the past decade, including DeepSEA[ Zhou et al., Nat Meth 2015], DeepBind [Alipanahi et al., Com Bio 2015], DeltaSVM [Lee et al., Nat Gen 2015], Basset [Kelley et al., Gen Res 2015], CADD [Rentzsch et al., Nuc Ac Res 2018], Expecto [Zhou et al., Nat Gen 2018], DeepC [Schwessinger et al., Nat Meth 2020], Enformer [Avsec et al., Nat Meth 2021], and GET [Fu et al., Nature 2024 (in press)]. We assess and scored key factors such as model complexity, data integration, cell-type specificity, interpretability, and usability within cancer research. Older models (2015-2018) score lower in complexity than recent ones (2018-2024), reflecting simpler architectures due to earlier computational limitations. In contrast, newer models leverage advanced computational power and deep learning, enhancing their ability to capture complex biological patterns and improve data integration and cell-type specificity. However, usability in practical research does not correlate with publication year, as user-friendly interfaces and comprehensive documentation play crucial roles in application, regardless of model complexity. In conclusion, AI has proven instrumental in elucidating the role of noncoding mutations in cancer, revealing intricate regulatory patterns and previously unknown mechanisms. Over 11, 000 publications cite the models discussed, underscoring their scientific impact. Such efforts will support predictive models that can generalize across diverse contexts, advancing genomic research and its implications for cancer biology. Our work highlights the potential of these AI models to assist researchers, particularly those less familiar with computational techniques, in advancing cancer genomics. Citation Format: Maria del Mar Alvarez-Torres, Xi Fu, Raul Rabadan. Illuminating the dark genome in cancer using artificial intelligence [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 7429.