M-GLC: Motif-Driven Global-Local Context Graphs for Few-shot Molecular Property Prediction

TL;DR

M-GLC tackles few-shot molecular property prediction by enriching relational context with a motif-driven tri-partite graph that includes motif, molecule, and property nodes. It combines a global GNN over the full context graph with local-focus subgraph encoders that isolate task-relevant neighborhoods, paired with structure-aware edge-weighted aggregation to balance heterogeneous signals. Across five MoleculeNet benchmarks, M-GLC achieves state-of-the-art ROC‑AUC results in both 5-shot and 10-shot settings, with ablations confirming the contribution of motifs, normalization, and local subgraphs. This approach demonstrates that incorporating shared substructures as structural priors and focusing on local context can substantially improve generalization under scarce supervision, with notable gains on datasets featuring sparse or imbalanced signals.

Abstract

Molecular property prediction (MPP) is a cornerstone of drug discovery and materials science, yet conventional deep learning approaches depend on large labeled datasets that are often unavailable. Few-shot Molecular property prediction (FSMPP) addresses this scarcity by incorporating relational inductive bias through a context graph that links molecule nodes to property nodes, but such molecule-property graphs offer limited structural guidance. We propose a comprehensive solution: Motif Driven Global-Local Context Graph for few-shot molecular property prediction, which enriches contextual information at both the global and local levels. At the global level, chemically meaningful motif nodes representing shared substructures, such as rings or functional groups, are introduced to form a global tri-partite heterogeneous graph, yielding motif-molecule-property connections that capture long-range compositional patterns and enable knowledge transfer among molecules with common motifs. At the local level, we build a subgraph for each node in the molecule-property pair and encode them separately to concentrate the model's attention on the most informative neighboring molecules and motifs. Experiments on five standard FSMPP benchmarks demonstrate that our framework consistently outperforms state-of-the-art methods. These results underscore the effectiveness of integrating global motif knowledge with fine-grained local context to advance robust few-shot molecular property prediction.

Figures (7)

• Figure 1: (a) Framework of M-GLC: Motif-driven Global–Local Context for few-shot molecular property prediction. Our work focuses on extracting informative context signals. (b) Firstly, we construct a tri-partite heterogeneous context graph with motifs, molecules, and properties to inject domain-specific structural priors. A global GNN encoder $\psi$ processes the full context graph to produce node embeddings that summarize global relational structure. (c) Then, we select task-relevant local-focus subgraphs centered on the target molecule and its target property. Finally, the dedicated subgraph encoders $\phi_{\text{mol}}$ and $\phi_{\text{prop}}$ map molecule-centric and property-centric subgraphs to subgraph-level embeddings, yielding a cleaner and more stable context representation for fast adaptation.
• Figure 2: The top–8 motifs by frequency in ChEMBL gaulton2012chembl, obtained via the bridge extraction method jin2020hierarchical. See Appendix \ref{['visualization']} for more high-frequency motifs.
• Figure 3: (a) Molecule–property context graph used in previous work schimunek2023contextwang2024pinwang2021propertyzhuang2023graph; missing (unknown) labels weaken the context information. (b) Our tri‑partite context graph introduces domain‑specific motif priors that enhance context informativeness. However, it also increases motif–molecule signal imbalance and heterogeneity, which we mitigate via structure‑aware aggregation and local‑focus subgraph.
• Figure 4: (a) In the previous context graph, edge weights are the same within each node type and influence message passing equally. (b) In our proposed tri-partite context graph, edge weights are uneven. Even among nodes of the same type, such as motifs or molecules, the contributions vary significantly. Some motif nodes have minimal influence (e.g., node 0 in the motif set), while a few molecule nodes dominate the overall propagation weight distribution (e.g., node 5 in the molecule set). Node indices correspond to positions shown in the figure. (c) After normalization, the imbalance observed in (b) is reduced, resulting in a more balanced flow of information.
• Figure 5: We project query‑molecule context embeddings at test time to 2D with PCA. (a1) and (b1) use the local‑focus subgraph; (a2) and (b2) remove it and read out node‑level embeddings. With the subgraph, embeddings form well‑separated clusters with clear structure within each label, and high‑confidence positives and negatives concentrate into tight groups while potential misclassifications converge into a small, isolated cluster. Without the subgraph, clusters are diffuse; many positive samples disperse throughout the negative region, indicating stronger noise and single‑node bias.