Recent years have witnessed increased interest in applying generative models to molecular libraries for the purposes of representation learning, exploring nearby chemical space, and accelerating directed search via a learned continuous representation space. A large body of literature exists for generating molecules using both string-based representations (e.g., SMILES or SELFIES) or on graph-based representations. In both regimes, generation proceeds auto-regressively, sampling one token, node, or edge per decoder step. This limits their utility for large scale search applications and reinforcement learning. Standard message passing schemes are provably limited in the types of molecules they can distinguish, especially in the presence of highly symmetric molecules or chiral centers.<br/><br/>In this work, we develop expressive graph encoders and computationally efficient graph decoders in a permutation-invariant graph autoencoder framework which we apply to large molecular libraries. While increased expressivity of the encoders is achieved at the expense of increased computational complexity, the decoders we investigate have at most quadratic time complexity in the size of the graph. In part, we achieve quadratic scaling by borrowing the iterative updating scheme inspired by AlphaFold2. We benchmark our network and find competitive performance on standard benchmarks in the field (e.g., MOSES, GuacaMol) as well as internal benchmarks for lead optimization (LO) and virtual-high-throughput screening (vHTS) applications. Our architecture therefore drastically improves the scalability of graph generative models for search and optimization applications.

Towards computationally efficient graph generative models for molecular search

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.