Comparing and Validating Machine Learning Models for Mycobacterium tuberculosis Drug Discovery | JoVE Visualize

Area of Science:

Computational chemistry and drug discovery
Machine learning in pharmaceutical research
Tuberculosis (TB) treatment development

Background:

Tuberculosis (TB) remains a significant global health challenge, with millions of cases and deaths annually.
Discovering new anti-TB drugs is crucial, but current methods are often inefficient and costly.
Machine learning (ML) offers a promising approach to enhance the efficiency of identifying novel anti-TB compounds.

Purpose of the Study:

To develop and evaluate machine learning models for predicting the activity of small molecules against Mycobacterium tuberculosis (Mtb).
To compare the performance of various ML algorithms, including deep learning and Bayesian models, using curated Mtb data.
To assess the utility of these models in prioritizing compounds for further in vitro and in vivo testing.

Main Methods:

Curated a comprehensive dataset of 18,886 small molecules with known activity against Mtb at different thresholds (10 μM, 1 μM, 100 nM).
Developed and evaluated multiple ML models, including Bayesian and deep learning approaches, using 5-fold cross-validation.
Utilized an independent evaluation set of 153 compounds published in 2017 to test model performance and generalizability.

Main Results:

A Bayesian model utilizing combined in vitro and in vivo Mtb data at a 100 nM cutoff achieved high cross-validation metrics (accuracy=0.88, recall=0.91).
The best-performing Bayesian model demonstrated comparable performance on the external test set (accuracy=0.83, recall=1.00).
Bayesian ML models generally performed equivalently to or better than deep neural networks on external Mtb datasets.

Conclusions:

Machine learning models, particularly Bayesian approaches, can effectively prioritize compounds for tuberculosis drug discovery.
These models provide a more efficient strategy for identifying potential anti-Mtb agents compared to traditional methods.
The developed models and curated datasets can accelerate the search for new treatments against Mycobacterium tuberculosis.