Global Bayesian models for the prioritization of antitubercular agents

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Tuberculosis (TB) remains a significant global health threat, necessitating the development of novel therapeutic agents.
• Existing treatments face challenges such as drug resistance and long treatment durations, highlighting the urgent need for new anti-TB compounds.

Purpose of the Study:

• To develop and validate robust Bayesian models for predicting the minimum inhibitory concentration (MIC) of compounds against Mycobacterium tuberculosis.
• To identify key molecular features and substructures associated with antitubercular activity.

Main Methods:

• Utilized a dataset of 3779 compounds with measured MIC values against Mycobacterium tuberculosis H37Rv.
• Explored various training sets and 15 fingerprint types, resulting in 90 distinct models.
• Identified the optimal model using Extended-Ring Class Fingerprints (ECFP_12) and global descriptors on a Functional Class Fingerprints (FCFP_4) derived training set.

Main Results:

• The best Bayesian model achieved high accuracy on the training set (total accuracy: 0.968) and good predictive performance on independent test sets (total accuracy: 0.869 and 0.73).
• The model identified conserved substructures influencing antitubercular activity, distinguishing between beneficial and detrimental molecular features.
• Demonstrated strong discriminant and predictive abilities for identifying potential anti-TB drug candidates.

Conclusions:

• The developed Bayesian model is a valuable tool for accelerating the discovery of novel antituberculosis compounds.
• The model's ability to predict compound activity and identify key substructures supports virtual screening and combinatorial library design strategies.
• This approach offers a promising avenue for expanding the pipeline of effective anti-TB therapeutics.