Explore a comprehensive lecture on machine learning techniques for generating new molecules, covering fundamental concepts from basic autoencoders to advanced applications in drug discovery. Delve into the theoretical foundations of variational autoencoders (VAEs), examining their probabilistic perspective, information theory concepts, and the mathematics behind Bayes' Law and evidence lower bounds. Master the intricacies of VAE architecture, including the challenges in variational computing and the structure of latent spaces, before advancing to Generative Adversarial Networks (GANs). Learn how junction tree VAEs are specifically applied to small molecule generation, culminating in practical applications of AI for identifying candidate antibiotics. Access complementary materials including detailed notes, slides, and a dedicated chapter to enhance understanding of these cutting-edge molecular generation techniques.
Generating New Molecules Using Machine Learning and AI - Lecture 15
Overview
Syllabus
Basic autoencoders
Variational autoencoders VAEs
VAEs from a probabilistic perspective
Information, entropy, and the KL divergence
Rewriting Bayes’ Law
Evidence lower bound
The opposing forces on a VAE
Issues with computing in a variational model
Structure of the latent space and b-VAEs
Generative Adversarial Networks GANs
A Junction tree VAE for small molecules
Using AI to identify candidate antibiotics
Taught by
Manolis Kellis
