Splitting chemical structure data sets for federated privacy-preserving machine learning.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Machine learning (ML) is increasingly used in drug design, necessitating robust test sets for reliable performance evaluation.
• Federated learning (FL) presents unique challenges for test set creation due to privacy constraints, preventing direct data sharing.

Purpose of the Study:

• To evaluate methods for splitting datasets in privacy-preserving federated machine learning for drug discovery.
• To assess the quality of splits generated by different methods against criteria like prediction bias and data imbalance.

Main Methods:

• Three data splitting techniques were investigated: locality-sensitive hashing (LSH), sphere exclusion clustering, and scaffold-based binning.
• Methods were evaluated based on prediction performance bias, label and data imbalance, and similarity between training and test sets compared to random splitting.

Main Results:

• Sphere exclusion clustering and scaffold-based binning demonstrated high-quality data splitting capabilities.
• Locality-sensitive hashing was found to be computationally less expensive in a federated setting.

Conclusions:

• Sphere exclusion clustering and scaffold-based binning are effective for generating meaningful test sets in federated drug discovery.
• The choice of splitting method involves a trade-off between splitting quality and computational cost.