Equivariant Shape-Conditioned Generation of 3D Molecules for Ligand-Based Drug Design

Authors: Keir Adams, Connor W. Coley

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our 3D generative model in tasks relevant to drug design including shape-conditioned generation of chemically diverse molecular structures and shape-constrained molecular property optimization, demonstrating its utility over virtual screening of enumerated libraries.
Researcher Affiliation Academia 1Department of Chemical Engineering, MIT 2Department of Electrical Engineering and Computer Science, MIT
Pseudocode Yes Algorithm 1 Genetic algorithm for shape-constrained optimization with SQUID
Open Source Code Yes Our source code can be found at https://github.com/keiradams/SQUID.
Open Datasets Yes We train SQUID with drug-like molecules (up to nh = 27) from MOSES (Polykovskiy et al., 2020) using their train/test sets.
Dataset Splits Yes The final dataset contains 1.3M 3D molecules, partitioned into 80/20 train/validation splits.
Hardware Specification No The authors acknowledge the MIT Super Cloud and Lincoln Laboratory Supercomputing Center for providing HPC resources that have contributed to the research results reported within this poster. Training the graph generator on 1 GPU takes 5-7 days.
Software Dependencies No We define graph (chemical) similarity sim G [0, 1] between two molecules as the Tanimoto similarity computed by RDKit with default settings (2048-bit fingerprints).
Experiment Setup Yes We use the Adam optimizer with default parameters. We use an initial learning rate of 2.5 10 4, which we exponentially decay by a factor of 0.9 every 50K iterations to a minimum of 5 10 6. ... max batch size (generation sequences) 400