Dual-Attention Multimodal Framework for Molecular Property Prediction

Abstract

Molecular property prediction plays a vital role in drug discovery and chemical research by facilitating the efficient selection and optimization of candidate compounds. Traditional Quantitative Structure–Activity Relationship models and early machine learning methods often rely on handcrafted molecular descriptors, which limit their generalizability and predictive performance. Recent advancements have shown that the powerful representation capabilities of Large Language Models on textual molecular data, combined with the relational learning strengths of Attention-based Graph Neural Networks, offer a new paradigm for capturing both the semantic and structural information about molecules. In this study, we introduce a novel Dual-Attention Multimodal framework for Graphs and Sequence-based representations, so-called DAM-GS. We evaluated our approach on multiple benchmark data sets covering diverse molecular prediction tasks. Experimental results showed that DAM-GS had outperformed the compared graph-based methods in most dataset–split settings. By leveraging the complementary advantages of graph-based and language-based modeling, our framework provides a promising solution for molecular property prediction with broad applications in drug discovery and computational molecular science.