Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about amortized sampling techniques for molecular conformations through this conference talk that introduces Prose, a 280 million parameter all-atom transferable normalizing flow designed to overcome the computational limitations of classical sampling methods in computational chemistry. Discover how deep learning enables the creation of scalable and transferable samplers that can generate zero-shot uncorrelated proposal samples for arbitrary peptide systems, achieving transferability across sequence lengths while maintaining efficient likelihood evaluation. Explore the extensive empirical evaluation demonstrating Prose's efficacy as a proposal mechanism for various sampling algorithms, including a simple importance sampling-based finetuning procedure that outperforms established methods like sequential Monte Carlo on unseen tetrapeptides. Understand the theoretical foundations proving that deep learning can design effective sampling algorithms, and examine the practical applications of this approach trained on a corpus of peptide molecular dynamics trajectories up to 8 residues in length. Access insights into the open-source implementation, model weights, and training dataset that advance research into amortized sampling methods and finetuning objectives for drug discovery applications.
