MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Federated learning offers an efficient method to increase training data volume for improved model predictivity, particularly when data generation is resource-intensive.
• The MELLODDY project aimed to explore the potential of federated learning in a pharmaceutical context.

Purpose of the Study:

• To evaluate the effectiveness of federated multipartner machine learning in improving predictive models within the pharmaceutical industry.
• To assess the impact of a novel multitask learning implementation across multiple pharmaceutical partners on model performance.
• To analyze the predictive performance and applicability domain of federated learning using a large-scale, cross-pharma dataset.

Main Methods:

• Utilized a novel federated multitask learning implementation across ten pharmaceutical companies.
• Leveraged a secure and privacy-audited platform for data aggregation and model training.
• Employed a comprehensive dataset of over 2.6 billion experimental activity data points, including small molecules and assay results.
• Developed complementary metrics to evaluate predictive performance in the federated setting.

Main Results:

• Achieved aggregated improvements in classification and regression models for each participating pharmaceutical company.
• Demonstrated increased predictive performance in the labeled space through federated learning.
• Observed an extended applicability domain for federated learning models.
• Noted saturating returns on predictive performance increases with growing collective training data volume.
• Reported markedly higher improvements for pharmacokinetics and safety panel assay-based tasks.

Conclusions:

• Federated learning, particularly with multitask learning, is an effective strategy for enhancing predictive model performance in the pharmaceutical industry.
• The approach successfully utilized a large, confidential cross-pharma dataset, demonstrating scalability and privacy.
• Federated learning expands the utility of machine learning in drug discovery, especially for complex pharmacokinetic and safety predictions.