Structured Variational Autoencoders for Prediction and Optimization; Diffusion for Molecule Generation - Primer

This conference talk from the Broad Institute's Models, Inference and Algorithms series features two presentations on advanced machine learning techniques. The first part is a primer by Alexandru Dumitrescu and Dani Korpela from Aalto University exploring diffusion models for molecule generation in drug discovery. Learn how deep generative models can rapidly create drug-like molecules, with detailed explanations of point cloud representations, E(3) invariant neural network parametrizations, and Field-based Molecule Generation (FMG). The speakers discuss the challenges of maintaining rotational invariance while accounting for molecular chirality, and outline future integration possibilities for drug discovery pipelines. The main presentation by Harri Lähdesmäki focuses on Structured Variational Autoencoders (VAEs), extending traditional VAEs by incorporating probabilistic graphical models to handle dependencies in latent variable priors. Discover how these models apply to temporal and spatio-temporal data modeling and high-dimensional Bayesian optimization through Gaussian process priors and latent neural ODEs/PDEs. Applications include longitudinal modeling of electronic health records, dynamical modeling of physical systems, and single-cell data analysis.