HD-SKRG: a hierarchical dual-view and structure-knowledge relation graph enhancement network for few-shot molecular...

Few-shot molecular property prediction aims to adapt to new property tasks using only a few labeled molecules. However, existing methods are still limited by insufficient molecular representations and biased intra-task relation construction under scarce supervision. To address these limitations, this paper proposes HD-SKRG, a hierarchical dual-view and structure-knowledge relation graph enhancement network. HD-SKRG improves few-shot prediction from two complementary aspects: molecular representation learning and intra-task relation propagation. For molecular representation, HD-SKRG builds atom-level and functional-group-level graphs, injects elemental knowledge into atom representations, and transfers local atomic information to functional-group representations. A frequency-aware aggregation module further produces molecular-level knowledge representations. For intra-task relation propagation, the structure relation graph provides the main feature-propagation pathway, whereas the knowledge relation graph supplies semantically related neighbors and refines relation weights. This design reduces graph-construction bias caused by relying on structural similarity alone. The dual-view encoders are pretrained by cross-view contrastive learning, and the full model is meta-trained under the MAML framework. Experiments on Tox21, SIDER, MUV, and ToxCast under 1-shot and 10-shot settings show that, HD-SKRG achieves the best results in five of eight settings and the second-best results in the remaining three. Ablation studies further confirm the contribution of each module.