Abstract
Molecular Property Prediction (MPP) is a critical task in drug discovery. While molecules can be represented as sequences (SMILES) or graphs, existing multi-view methods often fail to effectively integrate these complementary perspectives, especially when one view is noisy. We propose MolFuse, a multi-view learning framework that utilizes a dual learning mechanism and cross-view contrastive fusion to align and integrate sequence and graph representations. By incorporating spatial information (bond angles) and topological structures, MolFuse ensures high-quality feature extraction. Experimental results on MoleculeNet benchmarks demonstrate that MolFuse outperforms state-of-the-art models, achieving significant improvements in prediction accuracy.
TL;DR
MolFuse is a multi-view learning framework that combines molecular sequences (SMILES) and graph structures to predict chemical properties. By using a "dual learning" mechanism and contrastive fusion, it ensures that high-quality information from one perspective helps refine the other, achieving state-of-the-art results across major drug discovery benchmarks.
The Challenge in Molecular Property Prediction
In Computer-Aided Drug Discovery (CADD), predicting whether a molecule is toxic or effective is a high-stakes challenge. Researchers typically represent molecules in two primary ways:
- Sequences: SMILES strings, which represent molecules like sentences.
- Graphs: Structural representations showing how atoms (nodes) connect via bonds (edges).
Current methods suffer from two major gaps. Single-view methods miss out on complementary information available in the alternative representation. Meanwhile, existing multi-view methods often treat all data as equal, failing to account for "noisy" or less informative views for specific molecules. If these views are simply concatenated, inferior data can overshadow high-quality information, leading to suboptimal predictions.
The MolFuse Framework
MolFuse addresses these gaps through a two-stage architecture designed for intelligent feature integration.
Feature Extraction
The model processes molecules through three distinct lenses:
- Sequence View: Utilizes a Bidirectional Gated Recurrent Unit (BiGRU) to encode SMILES strings.
- Graph View: Employs a sophisticated dual-graph approach. It uses a Topological Graph for atoms and bonds and a Spatial Graph to capture bond angles, allowing the model to understand the 3D-like "shape" of the molecule.
Dual Learning & Contrastive Fusion
The core innovation lies in the Dual Learning Module, where the model attempts to reconstruct features of the graph view using the sequence view and vice versa. This forces the different mathematical spaces to align, allowing a high-quality view to "teach" a lower-quality one. Finally, Cross-View Contrastive Fusion ensures the global fused representation remains consistent with the specific details found in both views.
Experimental Setup and Datasets
MolFuse was evaluated on seven datasets from the MoleculeNet benchmark, including physiology-based sets (BBBP, Tox21, ClinTox, SIDER, ToxCast) and biophysics-based sets (BACE, HIV).
Key Results and Performance
MolFuse outperformed 12 state-of-the-art baselines, including Uni-Mol and GraphMVP, on six out of seven datasets.
- Average Performance: Achieved an average ROC-AUC of 78.1%, a 2.2% absolute improvement over previous top models.
- Significant Gains: Notable improvements were seen in ClinTox (+3.5%) and SIDER (+3.2%), where biological interactions are particularly complex.
- Representation Quality: t-SNE visualizations confirmed that MolFuse creates cleaner, more distinct clusters of molecules compared to single-view approaches.
Insights and Future Directions
The "Teacher-Student" dynamic established by the Dual Learning module ensures semantic preservation across views. By incorporating bond angles in the Spatial Graph, MolFuse captures nuanced characteristics without the computational burden of full 3D geometric deep learning. While successful, future work could explore incorporating true 3D conformer data (XYZ coordinates) and improving model interpretability to identify which chemical substructures drive specific predictions.
Summary updated