Contextual Molecule Representation Learning from Chemical Reaction Knowledge
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: Self-supervised Pre-training, Chemical Reaction, Molecular Representation Learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Self-supervised learning has emerged as a powerful tool for harnessing large amounts of unlabelled data to obtain meaningful representations. However, prevailing techniques such as reconstructing masked sub-units are inapplicable to Molecular Representation Learning (MRL) due to the high degree of freedom in possible combinations of atoms in molecules. In this work, we propose a self-supervised learning framework, \textit{REMO}, which pre-trains graph/Transformer encoders on 1.7 million chemical reactions by taking advantage of well-defined rules of atom combinations in common chemical reactions. Specifically, two pre-training objectives are proposed, including masked reaction centre reconstruction and reaction centre identification. \textit{REMO} offers a novel solution to MRL by leveraging the unique characteristics of chemical reactions as knowledge context for pre-training, which effectively supports diverse downstream molecular tasks with minimum finetuning. Experimental results show that \textit{REMO} outperforms masked modeling of single molecule in various downstream tasks.
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: 5944
Loading