Active learning strategies with COMBINE analysis: new tricks for an old dog

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• The COMBINE method is valuable for studying congeneric series and ligand-protein interactions in Quantitative Structure-Activity Relationships (QSAR).
• COMBINE's adoption is limited by challenges in quantifying prediction uncertainty and the requirement for extensive datasets.
• Active learning (AL) offers a solution by leveraging uncertainty to improve model performance and reduce training set size.

Purpose of the Study:

• To address the limitations of the COMBINE method by integrating active learning strategies.
• To develop and evaluate novel uncertainty estimators for active learning in QSAR.
• To demonstrate the efficacy of the proposed AL-COMBINE approach across diverse chemical datasets.

Main Methods:

• Implementation of active learning workflows to enhance the COMBINE method.
• Development of two uncertainty estimators: 'pool of regressors' and 'distance to the training set'.
• Validation of the AL-COMBINE strategy on three distinct datasets: HIV-1 protease inhibitors, Taxol derivatives, and BRD4 inhibitors.

Main Results:

• The AL-COMBINE strategy achieved success in 80% of cases for Taxol derivatives and BRD4 inhibitors datasets.
• The method outperformed random selection in the HIV-1 protease inhibitors time-split dataset.
• Demonstrated effectiveness in improving QSAR model performance with reduced sample sizes.

Conclusions:

• Active learning, when combined with the COMBINE method (AL-COMBINE), effectively addresses prediction uncertainty and data requirements in QSAR.
• The proposed uncertainty estimators are valuable for guiding active learning in cheminformatics.
• AL-COMBINE presents a promising approach for developing robust and efficient QSAR models, particularly when dealing with limited data.