Deep-learning-based virtual screening of antibacterial compounds.

Area of Science:

• Medicinal Chemistry
• Computational Biology
• Microbiology

Background:

• Rising multidrug-resistant bacteria necessitate novel antibiotic development.
• Existing antibiotic discovery methods face limitations in speed and scope.

Purpose of the Study:

• To integrate high-throughput screening (HTS) with deep learning (DL) for efficient antibacterial discovery.
• To develop and validate a DL model (GNEprop) for predicting antibacterial activity.
• To identify novel antibacterial compounds with high potency and selectivity.

Main Methods:

• Screened ~2 million small molecules using HTS against a sensitized Escherichia coli strain.
• Trained a DL model (GNEprop) on HTS data to predict antibacterial activity.
• Performed virtual screening of >1.4 billion compounds using the trained DL model.

Main Results:

• Identified thousands of initial hits from HTS.
• Achieved a 90-fold improved hit rate in virtual screening compared to HTS.
• Discovered 82 compounds with antibacterial activity, many dissimilar to known antibiotics.
• Validated specific biological targets for newly identified compounds.

Conclusions:

• The integrated HTS and DL approach significantly enhances antibacterial discovery efficiency.
• GNEprop model demonstrates robust predictive power and generalization capabilities.
• Identified novel antibacterial candidates offer promising avenues for combating resistant bacterial infections.