Development and Evaluation of Conformal Prediction Methods for Quantitative Structure-Activity Relationship.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Quantitative structure-activity relationship (QSAR) models predict compound biological activity using molecular descriptors.
• Accurate activity estimation is crucial, but quantifying prediction uncertainty (e.g., prediction intervals) remains a challenge in QSAR.
• Most high-performance machine learning (ML) algorithms require separate methods for uncertainty estimation.

Purpose of the Study:

• To develop computationally efficient conformal prediction (CP) algorithms for QSAR modeling.
• To enable reliable prediction intervals for widely used ML models in QSAR.
• To address the challenge of quantifying prediction uncertainty in QSAR.

Main Methods:

• Proposed computationally efficient conformal prediction (CP) algorithms tailored for random forests, deep neural networks, and gradient boosting.
• Utilized a novel approach for deriving nonconformity scores from ensemble-based prediction uncertainty estimates.
• Implemented algorithms designed to be agnostic to prediction modes and produce valid prediction intervals.

Main Results:

• Demonstrated the validity and efficiency of the proposed CP algorithms on diverse QSAR datasets and simulation studies.
• Showcased the ability of the algorithms to generate reliable prediction intervals for ML-based QSAR models.
• Validated the performance of the novel nonconformity score derivation method.

Conclusions:

• The developed CP algorithms offer a robust and efficient solution for uncertainty quantification in QSAR modeling.
• The software implementation facilitates integration into existing ML workflows for QSAR.
• This work advances the reliability and interpretability of QSAR predictions through valid prediction intervals.