kMoL: an open-source machine and federated learning library for drug discovery

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Data privacy and security in AI

Background:

• Machine learning, especially Graph Convolutional Networks (GCNs), is crucial for drug discovery tasks like Quantitative Structure-Activity Relationship (QSAR) and ADME.
• Data privacy concerns hinder collaborative efforts to improve machine learning model performance and robustness in the pharmaceutical industry.

Purpose of the Study:

• Introduce and evaluate kMoL, an open-source machine learning library with integrated federated learning capabilities.
• Address challenges in drug discovery related to data privacy and security.
• Provide guidance on deploying privacy-preserving machine learning models.

Main Methods:

• Developed kMoL, an open-source library featuring state-of-the-art GCN architectures, Bayesian optimization, explainability, and federated learning.
• Evaluated kMoL through local benchmark training and distributed federated learning experiments on diverse datasets.
• Assessed customization, security, adaptability, and performance trade-offs of federated learning strategies.

Main Results:

• kMoL demonstrates extensive customization, advanced security, and adaptability for user-specific models and datasets without extra programming.
• Federated learning experiments provided insights into performance trade-offs, guiding privacy-preserving model deployment.
• kMoL facilitates fast and practical experimentation in drug discovery.

Conclusions:

• kMoL offers an accessible, secure, and open-source platform for collaborative drug discovery, enhancing model development through federated learning.
• The library empowers researchers to build robust machine learning models while maintaining data privacy.
• Results offer valuable insights for implementing privacy-preserving AI in pharmaceutical pipelines.