Multimodal graph fusion with statistically guided parsimonious descriptor selection for molecular property prediction

Area of Science:

• Computational chemistry
• Machine learning for drug discovery

Background:

• Graph convolutional networks (GCNs) excel at molecular representation but struggle with global physicochemical properties.
• Existing fusion methods often lack explicit modeling of higher-order interactions.

Purpose of the Study:

• To introduce KROVEX (KROnecker-product based multimodal fusion with Variable sElection) for expressive molecular representation learning.
• To improve the prediction of molecular properties by integrating graph embeddings and molecular descriptors.

Main Methods:

• KROVEX utilizes Kronecker products to fuse graph embeddings with selected molecular descriptors, explicitly modeling second-order interactions.
• A two-stage descriptor selection process involving iterative sure independence screening and Elastic Net regularization identifies informative features.
• The method was validated on FreeSolv, ESOL, and custom datasets for vapor pressure and aqueous solubility prediction.

Main Results:

• KROVEX outperformed GCNs and other fusion baselines (EGCN, D-MPNN, BAN) on both random and scaffold splits.
• The method achieved state-of-the-art performance in vapor pressure prediction, a critical property for industrial applications.
• Ablation studies confirmed the benefits of statistically guided descriptor selection and Kronecker-product fusion over simple concatenation.

Conclusions:

• KROVEX provides a generalizable framework for molecular property prediction, enhancing both performance and interpretability.
• Parsimonious descriptor selection combined with multimodal graph fusion is key to improved predictive accuracy.
• The fusion strategy's effectiveness across different GNN backbones (GAT, GIN) highlights its versatility.