Conformal Regression for Quantitative Structure-Activity Relationship Modeling-Quantifying Prediction Uncertainty.

Area of Science:

• Computational chemistry
• Machine learning
• cheminformatics

Background:

• Predictive modeling requires associated confidence for effective decision-making.
• Conformal regression offers a framework for guaranteed prediction confidence levels.

Purpose of the Study:

• To apply conformal regression for modeling molecular properties and bioactivity.
• To optimize prediction interval efficiency (narrowness) using various scaling methods.

Main Methods:

• Conformal regression framework applied to molecular datasets.
• Evaluation of different algorithms for estimating prediction uncertainty to normalize ranges.
• Comparison of approaches across 29 publicly available datasets.

Main Results:

• The most efficient conformal regressors were achieved using the natural exponential of the ensemble standard deviation from random forests.
• This method yielded an average prediction range of 1.65 pIC50 units at 80% confidence for bioactivity modeling.
• The choice of nonconformity function significantly impacted prediction range width.

Conclusions:

• Conformal regression is a robust method for generating bioactivity predictions with associated confidence.
• Optimizing prediction interval scaling is crucial for regressor efficiency.
• The findings support the use of conformal regression in drug discovery and chemical research.