Improved Machine Learning Predictions of EC50s Using Uncertainty Estimation from Dose-Response Data

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Machine learning models in early-stage drug design often use compressed data representations.
• Curve fitting raw experimental results discards crucial information on fit quality.

Purpose of the Study:

• To integrate a fit-quality metric into machine learning models to assess data reliability.
• To enhance predictive performance in drug design by accounting for curve fit quality.

Main Methods:

• Four machine learning methods were evaluated: random forests (with parametric bootstrap, weighted, and variable output smearing variations) and weighted support vector regression.
• Fit-quality metrics were incorporated into models using 40 diverse datasets from PubChem and BASF.
• Predictive performance was assessed by comparing models with and without fit-quality metrics.

Main Results:

• Including fit-quality metrics significantly improved predictive performance on 31 out of 40 datasets.
• Statistically significant improvements were observed across multiple tested methods.
• The root-mean-squared error was reduced by up to 22% in the best-case scenarios.

Conclusions:

• Accounting for curve fit quality in data processing is a valuable strategy for improving machine learning model performance in drug design.
• This approach enhances the reliability of predictions without necessitating additional experimental data.
• The findings demonstrate a practical method to boost predictive accuracy in early-stage drug discovery pipelines.