Learning Hierarchical Interaction for Accurate Molecular Property Prediction

TL;DR

HimNet introduces a Hierarchical Interaction Message Passing framework that models atoms, motifs, and global molecular context to predict molecular properties with high accuracy and interpretability. By combining a dual-pathway message passing mechanism, cross-scale attention, and consensus fingerprint fusion, the method captures cooperative effects among substructures that traditional GNNs often miss. Thorough evaluation on MoleculeNet benchmarks plus challenging real-world datasets demonstrates state-of-the-art or near-state-of-the-art performance, while case studies and attention visualizations provide chemically meaningful explanations. The approach offers a scalable, interpretable solution for ADCETM property prediction in early drug discovery and motivates future incorporation of 3D structure and multimodal data.

Abstract

Discovering molecules with desirable molecular properties, including ADMET profiles, is of great importance in drug discovery. Existing approaches typically employ deep learning models, such as Graph Neural Networks (GNNs) and Transformers, to predict these molecular properties by learning from diverse chemical information. However, these models often fail to efficiently capture and utilize the hierarchical nature of molecular structures, and often lack mechanisms for effective interaction among multi-level features. To address these limitations, we propose a Hierarchical Interaction Message Passing Mechanism, which serves as the foundation of our novel model, the Hierarchical Interaction Message Net (HimNet). Our method enables interaction-aware representation learning across atomic, motif, and molecular levels via hierarchical attention-guided message passing. This design allows HimNet to effectively balance global and local information, ensuring rich and task-relevant feature extraction for downstream property prediction tasks, such as Blood-Brain Barrier Permeability (BBBP). We systematically evaluate HimNet on eleven datasets, including eight widely-used MoleculeNet benchmarks and three challenging, high-value datasets for metabolic stability, malaria activity, and liver microsomal clearance, covering a broad range of pharmacologically relevant properties. Extensive experiments demonstrate that HimNet achieves the best or near-best performance in most molecular property prediction tasks. Furthermore, our method exhibits promising hierarchical interpretability, aligning well with chemical intuition on representative molecules. We believe that HimNet offers an accurate and efficient solution for molecular activity and ADMET property prediction, contributing to advanced decision-making in the early stages of drug discovery.

Figures (4)

• Figure 1: Overview of the proposed HimNet framework. (a) Hierarchical Interaction Message Passing Module (HIMPM). (i) Message Passing: Combines a dual-channel architecture consisting of directed message passing (D-MPNN) and a hierarchical interactive attention mechanism. (ii) Interaction Module: Models multi-scale interactions including atom–atom, motif–motif, and atom–motif relationships using cross-hierarchical attention mechanisms. (iii) Global-Local Attention Module: Aggregates information from atom-level and motif-level nodes into a unified graph-level representation through hierarchical attention fusion. (b) HimNet Architecture. (i) Molecule Preprocessing: Converts SMILES strings into 2D molecular structures and constructs hierarchical molecular graphs comprising atom, motif, and global nodes. (ii) Molecular Fingerprint Representation Learning: Encodes multiple types of molecular fingerprints and extracts common features via similarity-guided fusion strategies. (iii) HIMPM: Applies the hierarchical message passing and interaction mechanisms described in (a) to the hierarchical graph representations. (iv) Attention Fusion: Integrates multi-source features from fingerprints and hierarchical graphs using multi-head self-attention to produce expressive molecular embeddings for downstream tasks.
• Figure 2: Hierarchical attention visualization for molecule CC1COc2c(N3CCN(C)CC3)c(F)cc3c(=O)c(C(=O)O)cn1c23. (a) Atomic-level attention: Highlights atom-wise contributions to BBB permeability, with red indicating positive and blue negative effects, consistent with chemical intuition that hydrophobic atoms promote and polar atoms inhibit penetration. (b) Motif-level attention: Aggregates attention to chemically meaningful substructures, showing enhanced positive attention on aromatic and fluorinated motifs, and negative attention on polar or acidic fragments. (c) Hierarchical interaction attention: Integrates atomic, motif, and global graph levels, revealing context-dependent interactions—such as competitive and cooperative effects—between motifs and their surrounding atomic environments, thereby supporting chemically interpretable property prediction.
• Figure 3: Hierarchical attention visualization of CCC(=O)C(CC(C)N(C)C)(c1ccccc1)(c1ccccc1), illustrating chemically interpretable multi-level feature interactions for BBB permeability prediction. (a) Atomic-level attention: The model assigns strong positive attention to the carbon atoms of both phenyl rings and the adjacent aliphatic carbons, reflecting their role in enhancing hydrophobicity and promoting BBB permeability. In contrast, negative attention is focused on the carbonyl oxygen and the nitrogen atom of the tertiary amine side chain, indicating these polar centers hinder BBB penetration. (b) Motif-level attention: The two phenyl rings, though structurally similar, are distinguished by the model: one is assigned strong positive attention, while the other receives negative attention, reflecting their distinct chemical environments and contextual influence from neighboring groups. The carbonyl and amine motifs both receive negative attention, consistent with their polar and hydrogen-bonding character. (c) Hierarchical interaction attention: The integrated view demonstrates that the model captures competitive and complementary relationships among motifs. The strongly positive aromatic motif (M2) competes with the negatively weighted aromatic motif (M3) and the polar carbonyl and amine motifs (M0, M1) in determining the global property. Notably, the context-dependent differentiation of the two phenyl motifs highlights the model's ability to capture non-additive and position-sensitive effects, aligning with chemical intuition regarding the interplay between hydrophobic and polar regions in BBB permeability.
• Figure 4: Performance comparison of HimNet and its ablated variants on eight benchmark datasets. Each bar shows the mean performance metric of the full model (HimNet) or a variant with a specific module removed (see legend for abbreviations). Error bars represent the standard deviation across repeated runs. Removing any module leads to a consistent performance drop, confirming the necessity of each component.