Applicability Domain of Active Learning in Chemical Probe Identification: Convergence in Learning from Non-Specific...

Area of Science:

• Chemical biology
• Computational chemistry
• Drug discovery

Background:

• Identifying specific chemical probes is crucial for understanding biological systems.
• Computational methods are needed to optimize candidate compound selection for experimental evaluation.
• Active learning (AL) shows promise in efficiently converging on predictive models with reduced datasets.

Purpose of the Study:

• To evaluate the applicability of active learning for identifying selective chemical probes.
• To assess AL's ability to predict inhibitory bioactivity profiles of selective compounds using non-selective ligand-target data.
• To understand the factors influencing AL's predictive capability in chemogenomic modeling.

Main Methods:

• Active learning virtual screening was employed.
• Chemogenomic features from non-selective ligand-target pairs were used for training.
• Multiple molecule representations and controls were compared.
• Matrix metalloproteinase family was used as an experimental model.
• Feature weight analyses and custom visualization were utilized.

Main Results:

• Active learning successfully predicted probe bioactivity using approximately 20% of non-probe data.
• Prediction performance was consistent with prior chemogenomic AL studies, despite increased experimental difficulty.
• Feature weight analysis and visualization clarified AL's decision-making process in classification.
• The study identified key factors contributing to predictive capability.

Conclusions:

• Active learning is applicable and effective for selective chemical probe identification in chemogenomics.
• The findings provide insights into optimizing computational probe design and discovery strategies.
• This approach can guide tactical decisions in developing specific and effective chemical probes.
• Understanding AL's decision-making enhances expectations for chemogenomic modeling accuracy.