Evaluation of categorical matrix completion algorithms: toward improved active learning for drug discovery

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning applications in pharmacology

Background:

• High-throughput screening (HTS) and high-content screening (HCS) are vital in early drug development for identifying therapeutic effects.
• Current screening methods primarily focus on desired effects, neglecting the identification of potential undesirable side effects due to the vast search space.
• Active machine learning (ML) presents a promising solution to address the limitations of traditional screening approaches.

Purpose of the Study:

• To introduce an improved matrix completion method, Impute by Committee, for handling categorical data.
• To evaluate the performance of Impute by Committee against existing methods in modeling compound-target interactions.
• To assess the efficacy of these methods in driving active learning for predictive model development.

Main Methods:

• Development of the Impute by Committee method for matrix completion with categorical values.
• Comparison of Impute by Committee with existing imputation techniques using latent similarities between compounds and targets.
• Application of active learning strategies to guide experimental selection for training predictive models.

Main Results:

• Impute by Committee demonstrated superior performance in both matrix completion accuracy and the efficiency of training predictive models compared to random selection.
• The method effectively modeled compound-target effects by leveraging latent similarities.
• An adaptive switching strategy for active learning further enhanced the performance of the Impute by Committee method.

Conclusions:

• Impute by Committee offers a more accurate and efficient approach to matrix completion for modeling compound-target interactions.
• This method significantly reduces the number of experiments required to train accurate predictive models in drug discovery settings.
• The developed active learning strategies improve the overall efficiency and effectiveness of computational approaches in identifying potential therapeutics.