Batched Bayesian Optimization for Drug Design in Noisy Environments.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in scientific research

Background:

• Drug design requires identifying bioactive compounds from large candidate libraries.
• Experimental assays are costly and can only be performed on a subset of compounds.
• Assay results are often affected by experimental noise, complicating compound selection.

Purpose of the Study:

• To evaluate the impact of noise on various batched Bayesian optimization techniques for drug screening.
• To introduce and assess a novel retest policy for mitigating noise in active learning.
• To enhance the efficiency of identifying active drug compounds in early-stage drug design.

Main Methods:

• Comparison of different batched Bayesian optimization algorithms under varying noise levels.
• Implementation and testing of a retest policy to address noisy assay data.
• Experimental validation of proposed methods in the context of drug candidate selection.

Main Results:

• Batched Bayesian optimization demonstrates robustness and effectiveness even with significant assay noise.
• The proposed retest policy effectively reduces the negative impact of noise.
• The retest policy facilitates the identification of more active compounds within a fixed experimental budget.

Conclusions:

• Active learning strategies, particularly batched Bayesian optimization, are valuable for efficient drug discovery.
• Noise in high-throughput screening assays can be effectively managed with appropriate strategies like retesting.
• The developed retest policy enhances the performance of active learning in noisy environments, accelerating the identification of promising drug candidates.