Multimodal feature fusion for molecular property classification.

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Accurate molecular property prediction is crucial for drug discovery and materials science.
• Existing unimodal models and simple multimodal approaches often fail to fully leverage complementary chemical information.

Purpose of the Study:

• To develop and evaluate a multimodal feature fusion framework integrating deep chemical language processing (CLP) and molecular fingerprints.
• To systematically investigate principles of effective cross-modal fusion for enhanced molecular characterization.

Main Methods:

• Benchmarking ten CLP architectures and eight fingerprint types via combinatorial search.
• Developing a data-aware fusion strategy focusing on feature integration and complementarity.
• Evaluating the framework across 60 datasets from MoleculeNet and TOXRIC.

Main Results:

• Identified synergistic configurations for multimodal fusion, demonstrating that simple aggregation is suboptimal.
• The proposed strategy effectively couples sequential (SMILES) and structural (fingerprints) representations.
• Achieved consistent and substantial performance gains over state-of-the-art baselines.

Conclusions:

• Effective multimodal fusion requires data-aware design, not just model aggregation.
• The framework provides a coherent, interpretable molecular representation, improving prediction accuracy.
• Offers practical guidelines for designing robust and generalizable multimodal molecular models.