Abstract 1116: AI framework for imputing genome-wide somatic mutations in pan-cancer targeted sequencing data for biomarker discovery

Abstract Background: Tumor genotyping via high-throughput sequencing has revolutionized our understanding of various cancer types. The genotyping of tumor biopsies or liquid biopsies has now become an integral part of clinical routines to survey tumors and to identify potential markers for treatment management. This has led to the development of many assays that are either commercially or institutionally available for such testing, where they often target less than 0.5% of the entire genome, focusing on known and recurrent genes for cancer-specific panels, pan-cancer solid tumor panels, or heme panels targeting blood cancers. This enrichment of certain regions has allowed us to find mutations that occur at lower allele fractions. Nonetheless, this approach results in missing hundreds or even thousands of potential mutations, some of which could be potential biomarkers for prognostic or predictive models, or even potential targets for future drug developments. Here, we hypothesize that an artificial intelligence (AI) framework can aid us in unraveling the complex interactions between mutations and thereby impute the off-target space. Methods: Here, we introduce SPARK (Somatic mutation Prediction via Artificial intelligence for Rapid discovery and Knowledge translation), an algorithm for imputing missing genomic loci in existing genomic panels. SPARK relies on the assumption that there is a latent space in which mutations co-occur or are mutually exclusive across multiple malignancies. This AI-based framework utilizes other data modalities (e.g., gene expression) to train a unified model for solid tumors. During training, we use somatic mutations and RNA expression data, and then employ another network to transfer knowledge, enabling us to take only a small number of mutations from any sequencing panel and impute the rest. Results: We used genotyping and RNA expression data from TCGA across multiple carcinomas to train the model (n>2000 samples). This pre-trained model is then fine-tuned for each panel for further optimization. To evaluate model performance, we focused on the genes present on the MSK-IMPACT panels obtained from the GENIE Project v16. We found a number of genes that occur with over 1% frequency across multiple carcinomas in TCGA but were not covered by the MSK-IMPACT panels. We then aimed to impute these genes and achieved a precision ranging from 60% to 100% for a minority of genes (e.g., ETV4 and ETV5). The precision was much lower (less than 10%) for many genes, suggesting that those mutations are likely not functional and do not show any strong complex association with the genes in the panel. As the next step, we plan to apply this model to the entire set of samples from the GENIE Project and impute the genes that showed acceptable precision in our TCGA analysis. This larger imputed dataset will enable us to find potential biomarkers from a large number of samples in the GENIE Project. Citation Format: Amirsaman Ashtari, Mohammad Shahrokh Esfahani. AI framework for imputing genome-wide somatic mutations in pan-cancer targeted sequencing data for biomarker discovery [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 1116.