ChemGraphy: A Graph-Aware Multimodal Knowledge and Reasoning System for Molecular Intelligence

Abstract

We introduce ChemGraphy, a multimodal drug discovery platform that unifies structural molecular encoding, literature retrieval, and language-model generation within a cross-attention fusion framework. The system accepts natural-language queries, identifies molecular entities, retrieves evidence from a ChEMBL-derived corpus, encodes molecular graphs using a six-layer Graphormer-based graph neural network (GNN), and supplies the resulting representations to an instruction-tuned language model. Its central architectural component, CrossAttentionFusion (CAF), projects 768-dimensional molecu-lar representations, 384-dimensional query representations, and retrieved-context vectors into a shared 512-dimensional space and integrates them using eight-head cross-attention followed by self-attention refinement. On four MoleculeNet benchmarks, the graph encoder provides molecular-property signals across solubility, hydration free energy, lipophilicity, and beta-secretase classification. On BACE, the reported CAF ablation reaches a ROC-AUC of 0.8324 ± 0.0466, compared with 0.7557 ± 0.0174 for naive feature concatenation under the same ablation protocol. Additional controlled experiments evaluate CAF retrieval depth, hidden dimension, scaffold-based generalization, attention perturbation, and external ADMET benchmarks. Retrieval depth is relatively stable on BACE and improves ESOL performance up to $k=5$; scaffold splitting produces lower performance than random splitting, exposing a measurable scaffold-shift gap. A token-removal study finds a significant inverse association between retained attention and perturbation impact, so attention is not treated as a causal explanation. External evaluation on TDC BBB and HIA benchmarks provides independent screening evidence while retaining the limitations of the descriptor-based ADMET module.