Feasibility of Active Machine Learning for Multiclass Compound Classification.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Hit-to-lead optimization requires classifying compounds into structural classes for Structure-Activity Relationship (SAR) studies.
• Automating compound classification using machine learning (ML) is desirable but requires substantial labeled training data.
• Acquiring labeled data for compound classification is often expensive, relying on expert input or experimental validation.

Purpose of the Study:

• To investigate the efficacy of active learning (AL) in reducing the number of required training compounds for multiclass compound classification.
• To propose an AL method for interactive, semi-automated multiclass classification with human feedback.

Main Methods:

• An active learning strategy was developed to iteratively select the most informative compounds for training.
• The proposed method integrates human feedback for a semi-automated classification workflow.
• The approach was empirically evaluated on 15 diverse compound classification tasks.

Main Results:

• Active learning successfully addressed multiclass compound classification challenges.
• The proposed AL method achieved classification with substantially fewer training compounds compared to standard techniques.
• Significant data reduction was observed, ranging from 10% to 80% of the data required by conventional methods.

Conclusions:

• Active learning is a viable and efficient strategy for multiclass compound classification in drug discovery.
• This approach can drastically reduce the cost and effort associated with generating training data.
• The integration of human feedback enhances the practicality of AL for real-world applications.