Abstract 1295: Deep-learning profiling of regulatory somatic variants across clonal cancer evolution.

Abstract Somatic noncoding mutations are abundant in cancer genomes, yet their regulatory functional consequences remain poorly understood. To address this gap, we generated a pan-cancer atlas of variant regulatory effects by profiling >35M SNVs within +/-20 kb of transcription start sites across more than 8,000 TCGA samples. Chromatin impact was predicted with the Sei deep-learning model (>20,000 features across 40 sequence classes), and clonal architecture was inferred with PyClone-VI, enabling genome-wide, evolution aware annotation of proximal regulatory alterations. We then aggregated these variant-level functional predictions into gene-level and patient-level representations of regulatory dysregulation. Despite comparable mutation counts near COSMIC Tier1/2 and non-COSMIC genes, the atlas revealed substantially higher predicted regulatory impact in promoter-proximal mutations, uncovering a noncoding functional dimension not detectable from burden alone. Aggregating these predicted effects into a personalized somatic dysregulation representation across COSMIC Hallmark genes produced a patient-level promoter-proximal regulatory effect score. In a multivariable Cox model of progression-free interval stratified by cancer type and adjusted for log1p(FGA), HRD score, non-silent mutations per Mb, and age at diagnosis, this regulatory score was a strong and independent predictor of adverse progression (HR = 1.10, 95% CI 1.04-1.15, P < 0.005), whereas noncoding mutation count contributed no additional information. Partitioning regulatory burden by evolutionary timing revealed distinct processes: clonal promoter-proximal disruption consistently predicted worse pan-cancer outcomes and showed stronger within-cancer associations, whereas subclonal burden was prognostic only in specific cancers. For example, subclonal promoter-proximal disruption within Hallmark genes was significantly associated with progression in Uterine Corpus Endometrial Carcinoma (HR = 1.36, 95% CI 1.10-1.70, P = 0.01), possibly indicating late-arising regulatory programs contributing to tumor progression. Finally, deep-learning regulatory sequence classes mapped to tissue-specific vulnerabilities: lower-grade glioma showed disproportionate impact from the E3 brain/melanocyte enhancer class, whereas sarcoma was most strongly associated with PC4 polycomb/bivalent stem-cell-linked disruption. Together, these results show that functional promoter-proximal disruption, not mutational load, captures clinically relevant noncoding dysregulation, shaped by context-specific and evolution-linked regulatory programs. This atlas provides a scalable foundation for integrating noncoding regulatory functions into prognostic modeling and will be made available to the research community. Citation Format: Ksenia Sokolova, Vessela N. Kristensen, Olga G. Troyanskaya. Deep-learning profiling of regulatory somatic variants across clonal cancer evolution [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1295.