A unified framework for molecular property prediction based on hierarchical multi-granularity molecular representation learning 1. The paper presents HMG-MRL, a unified molecular property prediction framework that explicitly integrates three complementary granularities: (i) fine-grained atom–bond structure, (ii) medium-grained functional motifs, and (iii) coarse-grained global molecular attributes (RDKit descriptors), aiming to capture chemical semantics that single-view models often miss. 2. A key design is an atom–bond bipartite graph encoder that treats atoms and bonds as two explicit, learnable node types in one message-passing system—so atom–atom, atom–bond, and bond–bond interactions are modeled jointly within each layer, rather than being handled indirectly or in separated pipelines. 3. For motif-level modeling, the framework builds a diverse motif vocabulary per molecule by combining three decomposition strategies: functional group matching, BRICS fragmentation, and Bemis–Murcko scaffold decomposition. Overlapping and nested motifs are allowed (a single atom/bond can belong to multiple motifs), preserving multiple valid “substructure perspectives.” 4. The motif view is encoded by a Motif Transformer: motif tokens are first made “atom-aware” via cross-attention that pulls context from the learned atom embeddings, then a Transformer with self-attention models global interactions among motifs; a CLS token summarizes motif-level information into a molecule representation. 5. A Global Attribute Encoder incorporates standardized RDKit molecular descriptors (physicochemical priors). It produces an attribute query vector and uses cross-attention to retrieve relevant context from atom embeddings, yielding an attribute-aware molecular representation rather than simply concatenating descriptors at the end. 6. Cross-granularity communication is built into the encoding process (motifs and descriptors attend to atom embeddings), and a cross-granularity alignment loss (contrastive objective) encourages semantic consistency between the atom–bond representation and the motif representation for the same molecule. 7. The final prediction uses attention-based fusion over the three granularity-specific representations (atom–bond Hb, motif Hm, descriptor Hd), learning per-task weights to combine them into a unified molecular embedding for classification (ROC-AUC) or regression (RMSE). 8. Evaluation on nine MoleculeNet benchmarks with scaffold splits (8:1:1) and five seeds shows strong overall performance: best mean results on 8/9 datasets. Reported highlights include ROC-AUC gains on BACE (0.910), BBBP (0.953), ClinTox (0.947), and improved RMSE on ESOL (0.690) and FreeSolv (1.541); Lipophilicity is slightly behind the top baseline. 9. Ablations indicate the gains rely on synergy: removing the bipartite encoder, Motif Transformer, descriptors, cross-granularity communication, or alignment loss all degrades results; descriptor removal hurts regression most, while motif removal impacts classification and regression, supporting the need for both functional substructures and physicochemical priors. 10. Interpretability analyses (t-SNE + clustering on BBBP; attention visualizations and attention-guided masking on FreeSolv/BACE/ClinTox) suggest the model’s attention highlights chemically plausible regions (e.g., polar groups for solvation, nitrogen-containing cores for BACE activity, thiophene/protonatable motifs for toxicity), and masking high-attention atoms/motifs harms performance more than random masking. 💻Code: github.com/AKZstar/HMG-MRL 📜Paper: doi.org/10.1093/bioinformati… #ComputationalBiology #Cheminformatics #MolecularRepresentation #GraphNeuralNetworks #Transformers #DrugDiscovery #ADMET #MoleculeNet