MolHFCNet: Enhancing Molecular Graph Representations with Hierarchical Feature Combining and Hybrid Pretraining

Duy-Long Nguyen, Duc-Luong Ho-Viet, Anh-Thu Ngo-Tran, Quang H. Nguyen, Binh P. NguyenIJCAI 2025 · DOI

TL;DR

This paper introduces MolHFCNet, a novel graph neural network architecture that enhances molecular representations by combining hierarchical feature extraction with a hybrid pretraining strategy. MolHFCNet achieves state-of-the-art performance in molecular property prediction tasks, demonstrating improved accuracy and efficiency compared to existing methods.

What Problem Does This Paper Address?

Predicting molecular properties is crucial for fields like drug discovery and materials science. While deep learning has shown promise, traditional graph neural networks (GNNs) face challenges:

  • Computational Complexity: Propagating full-dimensional features across many layers leads to high memory and processing demands, hindering scalability.
  • Limited Adaptability: Most GNNs are designed for either 2D or 3D molecular graphs, limiting their flexibility across different tasks and data availability.
  • Information Overlooking: 2D graphs miss crucial 3D geometric information, while incorporating 3D data can be computationally expensive.

MolHFCNet aims to address these limitations by offering a more efficient, adaptable, and comprehensive approach to learning molecular representations.

Methods

MolHFCNet introduces two main innovations:

  • n-Hierarchical Features Combining (n-HFC) Module: This core module aggregates information across multiple hierarchical feature spaces. It dynamically captures local and global graph structures by decomposing features and recursively processing them. Crucially, it achieves this with computational complexity comparable to a single full-dimensional graph layer, making it highly efficient. The n-HFC module can handle both 2D and 3D molecular graphs.
  • Hybrid Pretraining Strategy: This strategy combines two objectives:
  • Masked Attribute Prediction: The model learns to infer missing node and edge features (and 3D coordinates for 3D graphs) based on the remaining graph structure. This is applied to both 2D and 3D graphs.
  • Contrastive Learning: This objective encourages consistency between different "views" of the same molecule, generated through data augmentation techniques like subgraph removal and attribute masking. This helps the model learn robust embeddings that capture underlying molecular relationships.

The overall MolHFCNet architecture is modular and hierarchical, consisting of four stages, each with two block modules that integrate the n-HFC module, residual connections, and feed-forward networks.

Datasets

The paper utilizes the following datasets:

  • Pretraining: 10 million SMILES strings from the PubChem database.
  • Evaluation: Nine benchmark datasets from MoleculeNet for molecular property prediction: ESOL, FreeSolv, Lipophilicity, BACE, BBBP, HIV, ClinTox, SIDER, and Tox21.
  • Binding Affinity Prediction: The Atom3D dataset for Ligand Binding Affinity (LBA) and Ligand Efficacy Prediction (LEP) tasks.

Key Insights and Ideas

  • Hierarchical Feature Extraction for Efficiency: The n-HFC module's ability to achieve multi-scale feature learning with computational complexity similar to a single layer is a significant breakthrough. This suggests that hierarchical processing doesn't inherently need to be computationally prohibitive.
  • Hybrid Pretraining is Key: Combining masked attribute prediction with contrastive learning provides a powerful foundation for learning rich molecular embeddings. This dual approach captures both local structural details and global contextual information.
  • Adaptability to 2D/3D is Crucial: The model's flexibility to work with both 2D and 3D graph representations, or even mixed inputs, makes it highly practical for real-world scenarios where 3D data might not always be available or easily generated.
  • Pretraining Benefits Regression More: The paper observes that pretraining yields more substantial improvements on regression tasks compared to classification tasks, suggesting that it's particularly effective at learning continuous molecular property relationships.

Key Results

  • State-of-the-Art Performance: MolHFCNet consistently achieves top-tier results across multiple benchmark datasets for molecular property prediction, outperforming existing models in both regression (e.g., lowest RMSE on FreeSolv and ESOL) and classification (e.g., highest average ROC-AUC).
  • Improved Binding Affinity Prediction: MolHFCNet-CFC demonstrates competitive and often superior performance on binding affinity prediction tasks, outperforming advanced GeoSSL models.
  • Efficiency: The n-HFC module's computational complexity is comparable to a single standard graph layer, enabling deeper architectures without prohibitive costs.
  • Visualization: t-SNE visualizations show that MolHFCNet representations effectively capture scaffold information, with the CFC variant demonstrating more compact and better-separated clusters.

Limitations and Future Work

  • Limitations: While the paper highlights efficiency, the computational cost of some advanced 3D graph layers might still be a constraint for certain applications. The performance gains from pretraining are noted to be more subtle for classification tasks compared to regression.
  • Future Work: The authors suggest several promising directions:
  • Enhancing model robustness with more advanced 3D graph layers.
  • Further integrating 2D and 3D representations.
  • Expanding applications to molecular generation and optimization tasks.

Summary updated

BibTeX
@inproceedings{nguyen2025molhfcnet,
  author = {Duy-Long Nguyen and Duc-Luong Ho-Viet and Anh-Thu Ngo-Tran and Quang H. Nguyen and Binh P. Nguyen},
  title = {{MolHFCNet: Enhancing Molecular Graph Representations with Hierarchical Feature Combining and Hybrid Pretraining}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1035},
}
Figure