The Deep generative decoder, a general-purpose model for oncology

Medicine is inherently multimodal, requiring the analysis of diverse data types such as genomics, clinical metrics, and patient history. When a new patient arrives, clinicians must navigate multiple tasks, including diagnosis, condition assessment, and treatment planning. Yet, most biomedical AI models remain unimodal and single-task, limited to specific objectives like detecting breast cancer from mammograms.

Foundational models present a new opportunity to transform medical AI. These models, trained on large datasets without specific objectives, adapt easily to diverse tasks and contexts, making them ideal for complex medical needs. During my talk, I will introduce a foundational model for oncology, trained on The Cancer Genome Atlas (TCGA) dataset. I will demonstrate how this model learns detailed patient characteristics and adapts to various oncology tasks. Specifically, I will showcase its ability to detect tumor origins, classify the origins of metastases, predict patient survival, and assess immune infiltration, highlighting its potential to assist medical decisión making in complex scenarios.