Artificial intelligence-based integration of genomic and histopathologic data to improve precision cancer medicine - Project Abstract Microscopic images of tumors profile disease biology in a way that complements disaggregated genomic sequencing. Digital pathology images and tumor genomics are increasingly available together in clinical settings, and there is thus an unmet need for artificial intelligence (AI) methods to integrate them toward clinical and biological insights. This proposal will develop multimodal, interpretable AI models that fuse spatial and molecular tumor profiles with prior biological knowledge to sharpen diagnosis and improve selection of systemic therapy, focusing on breast cancer, non-small cell lung cancer, and muscle-invasive urothelial bladder cancer. In the K99 phase, I will study the added value of multimodal self-supervised learning to fuse pathology and genomics data for diagnosis and prediction of response to neoadjuvant therapy, with a research focus on using graphical methods to incorporate structured biological knowledge into high-performing black-box models. I will go beyond estimating prognosis with causal-inference analyses to estimate treatment effect by regimen. In the R00 phase, I will deepen the integration of structured biological knowledge into the models, especially the causal inference methods, and extend them into a treatment recommender system. Also during the independent phase, I will work to translate these models toward deployment via federated multi-institution validation, testing on newly accrued cases, and human-factors studies with pathologists/oncologists to assess interpretability, calibration, and decision impact. Building on recent successes in AI, the work will develop optimized, interpretable architectures for oncology to integrate complementary tumor profiles with pre-existing knowledge for actionable insights in cancer data science. Concretely, resultant models could eventually be developed toward de-escalating ineffective therapy, expediting effective treatment for cancers of unknown primary, and informing trial design for biomarker-guided neoadjuvant care. My career development plan builds on my prior training in computational pathology and clinical oncology to develop skills in multimodal self-supervised learning, causal inference, graph theory, and cancer biology, along with leadership and communication for team science. I will take targeted coursework, workshops, and grant-writing seminars; present at major meetings; and lead collaborative manuscripts. During the K99 phase, my training will be well supported by Dr. Sohrab Shah (Chief, Computational Oncology) and Dr. Nikolaus Schultz (Director, Cancer Data Science Initiative) and a panel of experts including pathologists, molecular biologists, and physician-scientists, with access to already-curated, world-class cohorts and the resources that I need to transition to independence.
