Fusing dual-event data sets for Mycobacterium tuberculosis machine learning models and their evaluation.

Area of Science:

• Medicinal Chemistry
• Computational Biology
• Drug Discovery

Background:

• Tuberculosis (TB) treatment requires novel drugs to combat resistance and improve efficacy.
• High-throughput screening (HTS) generates vast datasets for drug discovery.
• Machine learning (ML) offers powerful tools for analyzing HTS data.

Purpose of the Study:

• To develop and validate machine learning models for identifying novel anti-tuberculosis compounds.
• To evaluate the effectiveness of data-fusion approaches in enhancing ML model performance.
• To assess the predictive power of different ML models using diverse screening datasets.

Main Methods:

• Phenotypic high-throughput screening against Mycobacterium tuberculosis (Mtb).
• Development of ML models (Bayesian, SVM, RP Forest) incorporating dose-response and cytotoxicity data.
• Cheminformatics data-fusion approach to combine multiple screening datasets.
• External validation using a set of 1924 commercially available molecules and a published Mtb lead set.

Main Results:

• Combined datasets achieved a high external validation ROC of 0.83 (Bayesian or RP Forest).
• Models with lower cross-validation scores sometimes outperformed others on test sets.
• No single ML model was sufficient for identifying all compounds of interest.
• Data set fusion proved a valuable strategy for ML model construction in Mtb drug discovery.

Conclusions:

• Integrating multiple screening datasets via data fusion significantly enhances ML model performance for anti-TB drug discovery.
• Diverse ML models and validation strategies are crucial for robust compound identification.
• Further research should address limitations in chemical space and target coverage for HTS data.