Deep Learning and Generative Modeling for High-Throug hput Biological Discovery

Abstract

This presentation provides an overview of how deep learning and generative models are reshaping high-throughput biology and early-stage drug discovery. I will begin by introducing representation learning approaches that eliminate the need for conventional, hand-engineered image analysis pipelines in phenotypic screening, enabling more efficient and scalable hit detection. I will then show how these learned biological
representations can be extended to guide the preselection of compound libraries using Cell Painting data and multimodal learning strategies.

Finally, I will illustrate how generative modeling can be used to propose new molecular structures directly from cellular phenotypes, uncover subtle phenotypic changes, and reconstruct cellular, developmental or disease trajectories that may be experimentally inaccessible.