Molecule Generation by Heterophilious Triple Flows
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Molecule generation, graph neural networks, heterophily, generative models
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Interacting flows for improved molecule generation
Abstract: Generating molecules with desirable properties is key to domains like material design and drug discovery. The predominant approach is to encode molecular graphs using graph neural networks or their continuous-depth analogues. However, these methods often implicitly assume strong homophily (i.e., affinity) between neighbours, overlooking repulsions between dissimilar atoms and making them vulnerable to oversmoothing. To address this, we introduce HTFlows. It uses multiple interactive flows to capture heterophily patterns in the molecular space and harnesses these (dis-)similarities in generation, consistently showing good performance on chemoinformatics benchmarks.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5318
Loading