From Fragments to Systems: How can we assemble biological mechanisms in the age of AI?
Abstract: Modern cancer research routinely generates petabytes of data across vast spatial and molecular scales. Our primary focus is to build interpretable and steerable AI models of disease that integrates with existing biological and clinical knowledge. I will present two recent tools developed by our teams towards this goal: (i) UniVI is a scalable framework designed to unify multimodal single-cell data. UniVI learns distinct but aligned embedding spaces for each modality. This approach enables consistent cross-modal integration, denoising, and label transfer without the need for pre-annotated reference atlases or curated feature graphs. (ii) MiroSCOPE is an AI-driven digital pathology platform for the systematic annotation of functional tissue units (FTUs). While cell-level modeling provides high resolution, it often overlooks the structural organization of FTUs—the repeating units (such as glands or stromal compartments) essential to tissue function and pathological grading. MiroSCOPE provides an end-to-end, "human-in-the-loop" workflow that integrates multiclass segmentation models with curation-specific features. Using this system, we have annotated over 100.000 FTUs across 200 prostate cancer samples, creating a high-quality resource for the community and facilitating the use of complex structural information in clinical assessments.
Dr. Emek Demir is an Associate Professor at Oregon Health & Science University (OHSU) and a leader in the Knight Cancer Institute, where his lab turns high-dimensional molecular and imaging data into mechanistic biological insight. Trained at Bilkent University (M.S. Molecular Biology & Genetics; Ph.D. Computer Science), he went on to Memorial Sloan Kettering’s Computational Biology Center, where he helped build foundational community resources including the BioPAX standard and Pathway Commons. At OHSU, his work advances tumor heterogeneity modeling and AI–human causal inference, including DARPA-supported “human-in-the-loop” tools and an NVIDIA collaboration scaling multicellular segmentation and annotation of functional tissue units across prostate, breast, and pancreatic tumors.