A Bayesian active learning platform for scalable combination drug screens

Area of Science:

• Computational biology
• Pharmacology
• Genomics

Background:

• Large-scale drug combination screening is computationally challenging due to the vast number of potential drug pairs.
• Current methods often rely on fixed experimental designs and predictive modeling for unobserved combinations.
• Discovering synergistic drug combinations is crucial for effective cancer therapy.

Purpose of the Study:

• To introduce BATCHIE, a novel adaptive experimental design for large-scale drug combination screening.
• To demonstrate BATCHIE's ability to efficiently identify highly effective and synergistic drug combinations.
• To validate BATCHIE's predictive power and identify promising drug combinations for pediatric cancers.

Main Methods:

• BATCHIE employs information theory and probabilistic modeling for dynamic, batch-wise experimental design.
• The approach iteratively selects the most informative experiments based on prior results.
• Retrospective and prospective experiments were conducted on drug libraries and cancer cell lines.

Main Results:

• BATCHIE rapidly discovered highly effective synergistic combinations in retrospective analyses.
• In a prospective screen of 1.4M combinations, BATCHIE identified synergies after exploring only 4% of possibilities.
• The method pinpointed effective combinations for Ewing sarcomas, including PARP plus topoisomerase I inhibition.

Conclusions:

• Adaptive experimental design enables efficient and unbiased large-scale drug combination screening.
• BATCHIE significantly reduces the experimental burden for discovering synergistic drug therapies.
• The findings support the translation of BATCHIE for identifying novel combination treatments for cancers.