FL-QSAR

FL-QSAR implements a federated learning-based Quantitative Structure-Activity Relationship (QSAR) framework to enable collaborative, privacy-preserving QSAR model training across multiple institutions.

Key Features:

• Horizontal Federated Learning (HFL): FL-QSAR employs horizontal federated learning (HFL) to enable collaborative QSAR modeling without direct sharing of raw data and can simulate scenarios with and without HFL.
• Privacy-Preserving Collaboration: The framework retains each institution's data locally to address intellectual property and financial concerns while allowing contribution to collective model training.
• Performance Enhancement: Experiments reported that collaboration through FL-QSAR improves QSAR prediction performance compared to using only private datasets and approaches the performance of shared-data methods.
• Comparison with Secure Multiparty Computation (MPC): FL-QSAR has been contrasted with secure multiparty computation frameworks, highlighting differences in efficiency and practicality for QSAR modeling tasks.
• Extensibility: The framework is designed to be adaptable to other deep learning algorithms for drug-related learning tasks beyond QSAR analysis.

Scientific Applications:

• Collaborative QSAR model development: Enables pharmaceutical institutions to jointly develop QSAR models while maintaining control over proprietary datasets.
• Drug discovery and candidate prioritization: Supports development of more accurate and robust QSAR models to accelerate identification of potential therapeutic compounds while safeguarding sensitive information.

Methodology:

Clients train models on local QSAR datasets and contribute model updates (not raw data), which are aggregated iteratively to update a global model.