Abstract 5040: Transformer-based foundation model for clinically relevant DNA alterations: Filling gene status gaps and enhancing classification of rare cancer types

Differences in target panels, data collection strategies, and in data processing among clinical genomic profiling platforms, including MSK-IMPACT and FoundationOne, results in incomplete and incompatible data. For example, other GENIE panels lack the comprehensive genomic features provided by IMPACT. To address this challenge, we developed a Transformer-based foundation model, called CanBART, that generates “plausible patients” to harmonize patient profiles collected on different platforms. To illustrate the value of these generated patients, we adapted the pre-trained CanBART model for a tumor classification task to predict cancer types across 39 tumor types. The results demonstrated strong performance, showing that these plausible patients significantly enhance the robustness of classification, particularly for underrepresented cancer types. CanBART leverages a BART-based transformer architecture to predict missing genomic data. Patient profiles were represented in a “sentence-like” format, for which they were tokenized using a word-level approach and sorted by chromosomal location, capturing clinically relevant alterations, including mutations, hotspots, copy number variations, and structural variants. We used genomic data from the MSK-IMPACT and GENIE databases, encompassing a total of 144,000 patient profiles. Missing gene statuses were addressed by generating plausible patients, producing a range of possible alterations grouped in profiles that could be sorted by their probability of presence in a particular patient. To validate these profiles, we fine-tuned the model for a classification problem to predict cancer types across the dataset. CanBART was applied to harmonize genomic profiles across patient genomic profiling platforms, producing biologically plausible patient profiles that augmented the dataset for classification tasks. Adding these patients to the training dataset improved tumor-type classification, with a notable 20% enhancement in accuracy in rare cancer predictions compared to using real patient data alone. The enriched dataset demonstrated the benefits of the proposed data representation and transformer-based architecture in addressing the challenges of cancer classification when data is scarce. CanBART is a Transformer-based foundation model that provides an innovative solution for harmonizing patient genomic profiling platforms and generating plausible patients to train clinical prediction tools. CanBART not only addresses the challenge of incomplete data but also offers a robust method for enhancing tumor classification accuracy, particularly for rare cancers. This approach demonstrates potential for transformative advancements in genomic profiling and precision oncology, facilitating research and clinical decision-making where data scarcity has been a barrier. Citation Format: Dmitrii Chebanov, Quaid Morris. Transformer-based foundation model for clinically relevant DNA alterations: Filling gene status gaps and enhancing classification of rare cancer types [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 5040.