Multi-party collaborative drug discovery via federated learning.

Area of Science:

• Pharmacology and Drug Discovery
• Computational Chemistry
• Artificial Intelligence in Medicine

Background:

• Accurate prediction of drug-target binding affinity (DTA) and drug-drug interactions (DDI) is crucial for drug efficacy and safety.
• Pharmacological data is often siloed across institutions, hindering collaborative model development due to privacy and intellectual property concerns.
• Local learning with isolated data limits predictive model performance.

Purpose of the Study:

• To develop a novel federated learning (FL) framework for collaborative DTA and DDI prediction.
• To enable multiple institutions to train predictive models without sharing sensitive local data.
• To enhance prediction accuracy while ensuring robust data privacy.

Main Methods:

• Proposed a federated learning (FL) framework integrating secure multi-party computation (MPC).
• Employed FL for collaborative model training across distributed datasets.
• Utilized MPC during model aggregation to safeguard data privacy.

Main Results:

• The proposed FL framework achieved prediction performance comparable to centralized learning.
• The method significantly outperformed traditional local learning approaches.
• Demonstrated effective collaboration and data security across multiple institutions.

Conclusions:

• The novel FL framework successfully addresses the challenge of collaborative DTA and DDI prediction with enhanced data privacy.
• This approach accelerates drug discovery by enabling secure data sharing and improved model accuracy.
• The integration of FL and MPC offers a scalable solution for pharmaceutical research.