Could Chemical LLMs benefit from Message Passing

TL;DR

This paper explores whether integrating message passing neural networks with chemical language models can improve molecular property prediction, especially for smaller molecular graphs, by proposing and testing two information fusion strategies.

Contribution

It introduces two novel strategies for integrating MPNNs with language models and evaluates their effectiveness on molecular graph datasets.

Findings

Integration improves performance on small molecular graphs.

No significant gains on large molecular graphs.

Contrast learning and fusion strategies are effective for smaller graphs.

Abstract

Pretrained language models (LMs) showcase significant capabilities in processing molecular text, while concurrently, message passing neural networks (MPNNs) demonstrate resilience and versatility in the domain of molecular science. Despite these advancements, we find there are limited studies investigating the bidirectional interactions between molecular structures and their corresponding textual representations. Therefore, in this paper, we propose two strategies to evaluate whether an information integration can enhance the performance: contrast learning, which involves utilizing an MPNN to supervise the training of the LM, and fusion, which exploits information from both models. Our empirical analysis reveals that the integration approaches exhibit superior performance compared to baselines when applied to smaller molecular graphs, while these integration approaches do not yield performance enhancements on large scale graphs.