Abstract:Obtaining comprehensive and informative representations of molecules is a crucial prerequisite for efficient molecule property prediction in artificial intelligence-driven drug design and discovery. While recent advancements in graph neural networks (GNNs) and sequence modeling have achieved satisfactory results in molecular research, integrating molecular representations of different modalities effectively still faces significant challenges. Existing ensemble methods often rely on a simple concatenation of features and the valuable domain information may be overlooked. Therefore, we propose the novel EMPPNet ( E nhancing M olecular P roperty P rediction via Cross-modal Information Flow and Hierarchical Attention) method that leverages cross-modal information flow to capture the varied granularities of molecular characterization. The model incorporates a hierarchical attention mechanism to enable fully interactive multi-modal information encodings. Moreover, molecular reconstruction is used as a self-supervised task to enhance molecular representation. Extensive experimental results have demonstrated that the EMPPNet significantly outperforms the current state-of-the-art methods, including classification and regression tasks. Our model also highlights the significance of leveraging molecular information from different modalities, each of which contributes unique and diverse information essential for precise property prediction.

EMPPNet: Enhancing Molecular Property Prediction via Cross-modal Information Flow and Hierarchical Attention