Secure and Differentially Private Federated Graph Learning for Molecular Property Prediction

Chemical artificial intelligence increasingly relies on molecular property prediction models trained from proprietary compound libraries, bioassay records, and reaction-screening data. However, these data often contain commercially sensitive structures, confidential activity labels, and privacy-relevant experimental metadata, making direct centralization impractical. This paper proposes PrivMol, a privacy-preserving computational chemistry framework for federated molecular representation learning. PrivMol introduces two novel algorithms: Secure Substructure-Aware Federated Optimization and Differentially Private Molecular Gradient Calibration. The first algorithm decomposes molecular graphs into privacy-sensitive and task-relevant substructure regions, enabling local clients to train graph neural networks while transmitting only securely aggregated model updates. The second algorithm adaptively calibrates clipping and perturbation according to atom- and substructure-level contribution scores, reducing unnecessary utility loss on chemically informative fragments while retaining formal differential privacy guarantees. To improve robustness under heterogeneous chemical spaces, PrivMol incorporates local contrastive molecular alignment without exposing raw molecules, labels, scaffolds, substructure masks, or embeddings. Experimental evaluation on widely used public molecular benchmarks, including ESOL, FreeSolv, Lipophilicity, BBBP, BACE, HIV, and Tox21, demonstrates that PrivMol provides a favorable trade-off among prediction accuracy, communication efficiency, empirical leakage resistance, and privacy protection. The study offers a practical route toward secure collaborative chemical intelligence for computer-aided drug discovery, toxicology prediction, and materials informatics.