A federated graph learning method to multi-party collaboration for molecular discovery

Optimizing molecular resources utilization for molecular discovery requires collaborative efforts across research institutions to accelerate progress. However, given the high research value of both successful and unsuccessful molecules conducted by each institution (or laboratory), these findings are typically kept private and confidential until formal publication, with failed ones rarely disclosed. This confidentiality requirement presents a great challenge for most existing methods when handing molecular data with heterogeneous distributions under stringent privacy constraints. Here, we propose FedLG, a federated graph learning method that leverages the Lanczos algorithm to facilitate collaborative model training across multiple parties, achieving reliable prediction performance under strict privacy protection conditions. Compared with various traditional federate learning methods, FedLG method exhibits excellent model performance on all benchmark datasets. With different privacy-preserving mechanism settings, FedLG method demonstrates potential application with high robustness and noise resistance. Comparison tests on datasets from each simulated research institution also show that FedLG method effectively achieves superior data aggregation capabilities and more promising outcomes than localized model training. In addition, we incorporate the Bayesian optimization algorithm into FedLG method to demonstrate its scalability and further enhance model performance. Overall, the proposed method FedLG can be deemed a highly effective method to realize multi-party collaboration while ensuring sensitive molecular information is protected from potential leakage.