Statistical Bioinformatics Seminar: Dr Guillaume Jaume (Harvard Medical School)

Title: Towards Generalist AI Models in Pathology: The Unique Role of Molecular Data'

Abstract: How can we develop generalist AI models for pathology? How can we leverage these models for better diagnosis, prognosis, response-to-treatment prediction, and biomarker discovery? Foundation models have taken the field of computational pathology by storm—bringing a whole new perspective on AI model development, training, and evaluation. Whole-slide image classification now largely relies on pretrained “patch encoders”, such as UNI, and increasingly relies on “slide encoders”, such as Threads. Multimodal learning, in particular based on morphomolecular data, emerges as a critical component for training and evaluating these models. In this talk, I will present our recent works in this direction: (1) HEST (NeurIPS’24) for joint analysis of spatial transcriptomics and histology, and (2) Tangle (CVPR’24), Madeleine (ECCV’24) and Threads (in review) for molecular-guided slide representation learning. I will close by sharing my perspective on the potential future direction of the field.

About the speaker: Guillaume is a 3rd-year postdoctoral researcher at Harvard Medical School and Brigham & Women’s Boston Hospital in the group of Prof. Faisal Mahmood. He obtained his Ph.D. in Electrical and Electronic Engineering from EPFL in collaboration with IBM Research and ETH Zurich in 2022. Guillaume’s research focuses on computational pathology to integrate AI tools into the clinical and research facets of pathology. His research involves two main objectives: first, enhancing the representation learning of tissue by developing general-purpose foundation models for pathology and oncology; and second, integrating AI tools in drug development to improve drug safety assessment, detect toxicity, and discover safety biomarkers.