Are bigger data sets better for machine learning? Fusing single-point and dual-event dose response data for...

Area of Science:

• Drug discovery
• Computational biology
• Machine learning

Background:

• Tuberculosis (TB) remains a significant global health challenge requiring novel treatments.
• Extensive phenotypic screening data for Mycobacterium tuberculosis (Mtb) is publicly available.
• Machine learning (ML) can leverage existing data to build predictive models for drug discovery.

Purpose of the Study:

• To investigate if combining large single-point screening data with smaller, high-quality dual-event dose-response data improves ML models for TB drug discovery.
• To assess the predictive performance of various ML models using different data combinations.

Main Methods:

• Bayesian machine learning was employed to evaluate 12 models.
• Models were trained and tested using diverse datasets, including single-point, dual-event, and combined data.
• Performance was assessed using receiver operator characteristic (ROC) curves for internal and external validation.

Main Results:

• Combining single-point and dual-event data did not reduce model predictive ability (internal ROC: 0.83-0.91, external ROC: 0.62-0.83).
• Models trained on 1200-5000 compounds demonstrated predictive performance comparable to models trained on 25,000-35,000 compounds.
• Dual-event models showed lower performance (internal ROC: 0.6-0.83, external ROC: 0.54-0.83).

Conclusions:

• Integrating diverse TB screening data enhances ML model development.
• Smaller, well-curated datasets can yield highly predictive models, challenging the need for massive screening efforts.
• Results inform strategies for efficient high-throughput screening versus focused compound testing.