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

Graph convolutional networks (GCN) are effective for learning molecular representations, but their reliance on local message passing and simple feature concatenation limits their ability to capture global physicochemical properties. We present KROnecker-product based multimodal fusion with Variable sElection for eXpressive molecular representation learning (KROVEX), a method that integrates graph embeddings with molecular descriptors through a Kronecker-product to explicitly model second-order interactions. Informative descriptors are identified using a two-stage procedure that combines iterative sure independence screening with Elastic Net regularization. The proposed approach was evaluated on two benchmark datasets (FreeSolv and ESOL) as well as two self-curated datasets with vapor pressure and aqueous solubility as the target property. Overall, our method outperformed not only GCN but also fusion-based baselines such as EGCN, D-MPNN, and BAN under both the random and scaffold split. More importantly, the fusion operates at the final embedding level, enabling consistent performance across different GNN backbones (e.g., GAT and GIN). KROVEX achieves state-of-the-art performance on vapor pressure prediction, establishing a new benchmark for this safety-critical property essential for environmental monitoring and industrial process design. Ablation studies further demonstrated that (1) statistically guided descriptor selection yields more informative features than predefined descriptors, and (2) Kronecker-product fusion provides greater improvements than simple concatenation as the number of descriptors increases. These results demonstrate that parsimonious descriptor selection combined with multimodal graph fusion enhances predictive performance and interpretability, providing a generalizable framework for molecular property prediction.