Foundation-Assisted Cross-Domain Meta-Learning for Few-Shot Molecular Property Prediction
TL;DR
FACML augments cross-domain episodic learning with property-aware priors from a frozen Structure-Property Multi-Modal foundation model (SPMM), which improves the average ROC-AUC over the strongest baseline by 6.36% and 8.12% under the 1-shot and 10-shot settings, respectively.
Abstract
Molecular property prediction is important for drug discovery, but labeled molecules are often limited for new prediction tasks. The difficulty is amplified in cross-domain few-shot settings, where target tasks differ from available source tasks and the model must adapt from sparse target-domain supervision. Existing few-shot and transferable few-shot methods mainly condition adaptation on the target support set, which can be unstable when the support molecules provide limited coverage of the target distribution. We propose Foundation-Assisted Cross-Domain Meta-Learning for Few-Shot Molecular Property Prediction (FACML), which augments cross-domain episodic learning with property-aware priors from a frozen Structure-Property Multi-Modal foundation model (SPMM). FACML represents each molecule using SPMM-derived SMILES features and graph features, and constructs support Property Vector (PV) prototypes and query PV representations from SPMM-predicted PVs based on molecular SMILES. A property-conditioned relation graph then integrates molecular nodes and PV-based prior nodes to refine episode representations for target-task adaptation. Across cross-dataset few-shot molecular property prediction benchmarks, FACML improves the average ROC-AUC over the strongest baseline by 6.36% and 8.12% under the 1-shot and 10-shot settings, respectively. An auxiliary zero-target-label analysis further examines the use of SPMM-generated molecules when labeled target support molecules are unavailable.