Structure Prototype Learning for Federated Molecular Property Prediction under Scaffold Heterogeneity
Tingkai Sun, Xiucai Ye, Tetsuya SakuraiAbstract
Federated molecular property prediction enables collaboration across private chemical libraries, but parameter averaging does not explicitly account for scaffold heterogeneity. FedSPL groups molecules using fixed descriptors and exchanges embedding prototypes alongside model parameters, enabling label-independent prototype matching while retaining supervised property prediction. Across five classification benchmarks and three heterogeneity levels, matched three-seed evaluation using masked, held-out server-model ROC-AUC showed higher mean final AUC than FedAvg in seven of 15 settings. Consistently positive differences on SIDER and mixed outcomes elsewhere indicate data set-dependent benefits, motivating validation-guided use of structural information in federated prototype learning.