Multimodal Cross-Attention Molecular Property Prediction for Text, Sequence, Graph, and Geometry

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery
• Materials Science

Background:

• Standard Quantitative Structure-Property Relationship (QSPR) models often rely on single-modal molecular representations, limiting their predictive accuracy.
• Accurate QSPR models are crucial for accelerating the pace of drug discovery and material design.
• Integrating diverse molecular data types can potentially overcome the limitations of single-modal approaches.

Purpose of the Study:

• To develop and evaluate a novel multimodal model for molecular property prediction.
• To enhance the accuracy of QSPR models by integrating multiple molecular representations.
• To demonstrate the model's effectiveness in drug discovery and material design applications.

Main Methods:

• Introduced the multimodal cross-attention molecular property prediction (MCMPP) model.
• Integrated various molecular representations: SMILES, ECFP fingerprints, molecular graphs, and 3D conformations.
• Employed a cross-attention mechanism after independent processing by Transformer-Encoder, BiLSTM, GCN, and reduced Unimol+.

Main Results:

• MCMPP demonstrated improved prediction accuracy across four datasets (Delaney, Lipophilicity, SAMPL, BACE).
• The model effectively leveraged complementary information from different molecular modalities.
• MCMPP outperformed other data fusion methods, achieving the highest Pearson correlation coefficient.

Conclusions:

• Multimodal integration significantly enhances molecular property prediction accuracy.
• The MCMPP model offers a powerful tool for accelerating drug discovery and material design.
• Cross-attention mechanisms are effective for fusing diverse molecular data representations.