Improving Molecular Property Prediction via Topology-Enhanced Chemical Language Model

Anonymous

Abstract: Pre-trained chemical language models (CLMs) excel in the field of molecular property predictions, utilizing string-based molecular descriptors such as SMILES for learning universal representations. However, the one-dimensional format of SMILES can impede the effectiveness of the model because it lacks the topological information necessary for accurate property predictions. In this work, we introduce HINT, a novel framework to enhance the understanding of molecular structures within CLMs with topological fingerprints. HINT enhances molecular representations of CLMs through a molecular substructure prediction task and fingerprint-based contrastive learning. Experimental results on various tasks verify that HINT significantly improves the molecular property prediction performance of CLMs.

Paper Type: short

Research Area: NLP Applications

Contribution Types: NLP engineering experiment

Languages Studied: English