Student: Ms. Ziru HUANG
Department: Department of Chemistry, HKUST
Supervisor: Prof. Haibin SU
Abstract
Exploring the vast chemical space—estimated between 1023
and 1060
candidate molecules—presents a fundamental challenge for traditional drug and molecular discovery. While conventional virtual screening methods are often computationally expensive and limited to existing libraries, generative machine learning models offer a paradigm shift by directly navigating and sampling novel molecular structures. Among these, Variational Autoencoders (VAEs) have proven particularly effective by mapping discrete chemical structures into continuous probabilistic latent spaces, enabling smooth interpolation and targeted property optimization. This seminar provides an overview of VAE architectures in molecular design, tracing their evolution across structural dimensions. We first introduce VAE fundamentals, contrasting them with traditional autoencoders to explain how probabilistic latent space regularizations ensure continuous molecular navigation. We then review foundational milestones, moving from 1D SMILES-based sequence models (Chemical VAE) to 2D graph-based substructure trees (JT-VAE) that guarantee 100% chemical validity. Finally, we highlight modern applications bridging advanced chemical requirements, including 3D stereochemistry modeling for complex natural products (NP-VAE) and latent space sampling for biological macromolecules and dynamic protein conformational ensembles. Overall, VAEs serve as a powerful and highly versatile tool that significantly accelerates generative AI-driven molecule discovery.