Enabling design of screening libraries for antibiotic discovery by modeling ChEMBL data

Area of Science:

• Microbiology
• Computational Chemistry
• Drug Discovery

Background:

• Identifying novel antibiotics is critical for combating bacterial infections.
• Understanding molecular penetration of bacterial envelopes is a major challenge in antibiotic discovery.
• Existing methods struggle to predict which compounds will reach intracellular bacterial targets.

Purpose of the Study:

• To develop predictive models for identifying molecules that penetrate bacterial cells.
• To create a curated dataset of bacterial cell penetrators and non-penetrators for antibiotic screening.
• To guide the selection of compounds for antibiotic discovery pipelines.

Main Methods:

• Extracted a dataset of chemical compounds and their penetration status from the ChEMBL database.
• Developed and evaluated random forest classification models to distinguish between penetrators and non-penetrators.
• Analyzed physicochemical properties of penetrators versus non-penetrators.

Main Results:

• Random forest models achieved ~87% accuracy in predicting bacterial cell penetration.
• Models showed high precision (~88%) and recall (~97%) for Gram-positive bacteria penetrators.
• Performance varied for Gram-negative bacteria, with better prediction of non-penetrators due to data distribution.
• Limited data on non-penetrators impacted model specificity and negative predictive value.

Conclusions:

• Predictive models can aid in designing antibiotic screening libraries.
• Physicochemical properties alone are insufficient for predicting bacterial cell penetration.
• Further data accumulation, especially for non-penetrators, is necessary to improve model generalizability.
• These preliminary models represent a valuable tool for antibiotic discovery until more comprehensive datasets are available.