Predicting Antimicrobial Class Specificity of Small Molecules Using Machine Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Microbiology

Background:

• Antibiotic resistance is a growing threat, necessitating new drug discovery methods.
• Traditional antibiotic development is slow and costly, with declining success rates.
• Machine learning (ML) offers a promising computational approach to accelerate drug discovery.

Purpose of the Study:

• To develop and validate ML models for predicting antimicrobial activity.
• To create a comprehensive antimicrobial knowledge graph (AntiMicrobial-KG) from public data.
• To assess the models' utility in screening compound libraries for novel antibiotics.

Main Methods:

• Constructed the AntiMicrobial-KG by collecting and visualizing public in vitro antibacterial assay data.
• Trained seven classic ML models using six compound fingerprint representations.
• Evaluated model performance, identifying the Random Forest model with MHFP6 fingerprint as optimal (75.9% accuracy, 0.68 Cohen's Kappa).

Main Results:

• The best ML model accurately predicted antimicrobial activity for compounds in the EU-OpenScreen and Enamine libraries.
• Over 30% of active compounds were correctly identified for Gram-positive, Gram-negative, and fungal pathogens in the EU-OpenScreen library.
• The model demonstrated pathogen class specificity predictions for the Enamine library.

Conclusions:

• The developed ML models and AntiMicrobial-KG can significantly accelerate antimicrobial drug discovery.
• This computational strategy efficiently filters compound libraries, reducing costs and identifying promising candidates.
• The approach aids in combating antimicrobial resistance (AMR) by streamlining the search for novel therapeutics.