Beyond Heuristics: A Model-Agnostic Framework for Uncertainty Quantification in QSAR via Adaptive Conformal...

Area of Science:

• Computational chemistry
• Cheminformatics
• Toxicology

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are crucial for predicting chemical properties and toxicity.
• Current Applicability Domain (AD) metrics lack statistically guaranteed confidence estimates, providing only heuristic similarity scores.
• Reliable uncertainty quantification is essential for the regulatory interpretation of QSAR model predictions.

Purpose of the Study:

• To introduce conformal prediction (CP) as a calibration layer to retrofit existing QSAR models into confidence predictors.
• To generate statistically guaranteed prediction intervals (regression) and prediction sets (classification) at a user-specified confidence level.
• To formalize and quantify the relationship between structural novelty and prediction reliability, bridging the gap left by heuristic AD methods.

Main Methods:

• Conformal prediction framework applied as a calibration layer without retraining QSAR models.
• Utilized auxiliary models trained on molecular fingerprints as nonconformity scores, with an ordinal distance strategy for hard-label classifiers.
• Validated on over 100 VEGA QSAR models across diverse endpoints and a large-scale application to the EPA CompTox chemical inventory.

Main Results:

• The CP framework consistently achieved nominal coverage across all tested QSAR models and endpoint types.
• Conformal efficiency metrics (prediction interval width, singleton rate) strongly correlated with traditional AD indices.
• Demonstrated that CP formalizes heuristic chemical similarity into statistically valid prediction intervals or label sets.

Conclusions:

• Conformal prediction offers a statistically rigorous method for quantifying uncertainty in QSAR predictions.
• The CP framework enhances transparency and reliability assessment for QSAR/Quantitative Structure-Property Relationship (QSPR) platforms.
• Enables practical, regulatory-scale deployment of reliable QSAR uncertainty quantification through an open-source pipeline.