Abstract
Molecular Property Prediction (MPP) is a critical component of Computer-Aided Drug Discovery. However, existing methods often struggle to integrate disparate molecular representations effectively, particularly when one view contains noisy or inferior information. This paper presents MolFuse, a novel framework that bridges sequence-based (SMILES) and graph-based representations. MolFuse utilizes a dual-learning mechanism to refine features across views and a cross-view contrastive fusion strategy to maximize global coherence. Experimental results on the MoleculeNet benchmark demonstrate that MolFuse achieves state-of-the-art performance, with an average ROC-AUC of 78.1%, effectively addressing the "weakest link" problem in multi-view molecular modeling.
Introduction
In Computer-Aided Drug Discovery (CADD), predicting molecular properties such as toxicity or bioactivity is foundational. Molecules are typically represented in two primary formats: sequences (SMILES strings) and graphs (atoms as nodes, bonds as edges). Current research faces two major hurdles: information silos, where single-view methods miss complementary features, and the "weakest link" problem, where multi-view methods fail to account for varying data quality between views. For a specific property, one view might be highly informative while the other is noisy; simple concatenation often leads to suboptimal performance as the model is misled by low-quality information.
Methodology
MolFuse addresses these challenges through a two-stage feature learning framework designed to extract, refine, and fuse molecular data.
Multi-Granularity Extraction
The framework first extracts features from three distinct perspectives:
- Sequence View: A Bidirectional Gated Recurrent Unit (BiGRU) processes SMILES strings to capture the contextual "language" of the molecule.
- Graph View (Topological & Spatial): MolFuse employs a dual-graph approach. It processes an atomic graph for topological structure and a chemical bond graph to capture spatial and angular characteristics. This ensures the model understands both the atomic composition and the geometric arrangement.
Dual Learning and Contrastive Fusion
To mitigate the disparity in view quality, MolFuse implements two key modules:
- Dual Learning Module: This module attempts to reconstruct features of one view using the other (e.g., predicting graph features from sequence features). This cross-view reconstruction forces the sharing of reliable semantic information and smooths out inconsistencies.
- Cross-View Contrastive Fusion: The model generates a global fused feature and employs contrastive learning to maximize mutual information between this global representation and the individual views, ensuring the final output is coherent and representative of the best features from both worlds.
Experimental Results
The model was evaluated on seven binary classification tasks from the MoleculeNet benchmark using scaffold splitting, a rigorous method that tests generalization to new molecular structures.
Performance Analysis
MolFuse outperformed several state-of-the-art (SOTA) models, including Uni-Mol and GraphMVP. It achieved an average ROC-AUC of 78.1% across all datasets, a 2.2% absolute improvement over previous benchmarks. Notably, on the ClinTox dataset, MolFuse reached 95.5% ROC-AUC, a significant increase from Uni-Mol’s 92.0%.
Key Insights
- Geometric Modeling: Explicitly modeling bond angles in a spatial graph allows MolFuse to capture nuanced shape characteristics often overlooked by standard GNNs.
- Mutual Guidance: The dual learning loss acts as a quality filter, allowing a "low quality" view to borrow reliable semantics from a higher-quality counterpart.
- Architectural Synergy: Ablation studies confirmed that removing either the Dual Learning or Contrastive Fusion modules leads to a noticeable drop in performance, proving their synergistic necessity.
Conclusion
MolFuse represents a significant advancement in multi-view molecular representation. By addressing the "weakest link" problem through dual learning and contrastive fusion, it sets a new benchmark for molecular property prediction. Future work will explore reducing computational complexity, integrating 3D conformer data, and extending the framework to large-scale self-supervised pre-training on unlabeled molecular datasets.
Summary updated