Error Models for Uncertainty Quantification in Molecular Machine Learning
- 1Novartis Biomedical Research
- 2Novartis Campus
- 3Institut de Quimica Computacional i Catalisi, Facultat de Ciencies, Universitat de Girona, Maria Aurelia Capmany 69, Girona, Spain, 17003
Abstract
Molecular property optimization is essential in drug discovery and medicinal chemistry. Machine learning (ML) models have demonstrated successful predictive ability and a certain degree of generalization capabilities to predict several molecular properties, including absorption, distribution, metabolism, and excretion (ADME) endpoints. However, quantifying the uncertainty of individual predictions remains a critical challenge. Assessing whether a given prediction is likely to be correct has the potential to fundamentally change how ML-based predictions are used in molecular design. In this work, error models (EMs), which are ML systems trained to predict the error of the primary property models, were systematically developed and analyzed to assess the reliability of property predictions. EMs were shown to discriminate between predictions of varying error levels, highlighting their utility for uncertainty quantification (UQ) and decision-making. Using both proprietary and public datasets covering diverse ADME and physicochemical endpoints, EM-derived uncertainty estimates outperformed standard UQ metrics, such as the distance to the training set or ensemble variance. Overall, random forest-based EMs built with the distance to the training set, ensemble variance, and predicted property values provided a robust and interpretable default configuration. While alternative feature sets, algorithms, and data configurations occasionally improved performance for specific ADME datasets, these gains were generally modest compared with the improvements achieved over standard UQ metrics. Feature importance and explainability of EMs were analyzed to better understand sources of uncertainty. Such explainability analyses further showed that the three EM input features provide complementary information for reliability assessment and enable compound-level interpretation of uncertainty estimates. Taken together, EMs are presented as a tool that bridges prospective uncertainty estimation and error analyses, supporting the identification of model failures and enabling more informed decision-making in molecular ML.
Related
- NephroTox AI: an explainable machine learning framework with uncertainty quantification for drug-induced nephrotoxicity prediction
- Towards Out-of-Distribution Uncertainty Quantification for Machine-Learned Interatomic Potentials via Pre-Training
- AdaptNTK: Adaptive Uncertainty Quantification and Active Learning for Neural Network Potentials
- An integrated machine learning framework for accurate, real-time and non-destructive vitamin C quantification from Vis/NIR spectral data
- Machine Learning of Band Gaps and Thermodynamic Stability in A2BB′X6 (X = F, Cl, Br, I) Double Perovskites with Uncertainty-Aware Multi-Objective Screening
BibTeX
@misc{parrondopizarro2026error,
title = {{Error Models for Uncertainty Quantification in Molecular Machine Learning}},
author = {Raquel Parrondo-Pizarro and Jessica Lanini and Raquel Rodríguez-Pérez},
howpublished = {ChemRxiv},
year = {2026},
month = aug,
doi = {10.26434/chemrxiv.15008122/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008122/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Error Models for Uncertainty Quantification in Molecular Machine Learning
AU - Raquel Parrondo-Pizarro
AU - Jessica Lanini
AU - Raquel Rodríguez-Pérez
PY - 2026
DA - 2026/08/31
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008122/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008122/v1
AB - Molecular property optimization is essential in drug discovery and medicinal chemistry. Machine learning (ML) models have demonstrated successful predictive ability and a certain degree of generalization capabilities to predict several molecular properties, including absorption, distribution, metabolism, and excretion (ADME) endpoints. However, quantifying the uncertainty of individual predictions remains a critical challenge. Assessing whether a given prediction is likely to be correct has the potential to fundamentally change how ML-based predictions are used in molecular design. In this work, error models (EMs), which are ML systems trained to predict the error of the primary property models, were systematically developed and analyzed to assess the reliability of property predictions. EMs were shown to discriminate between predictions of varying error levels, highlighting their utility for uncertainty quantification (UQ) and decision-making. Using both proprietary and public datasets covering diverse ADME and physicochemical endpoints, EM-derived uncertainty estimates outperformed standard UQ metrics, such as the distance to the training set or ensemble variance. Overall, random forest-based EMs built with the distance to the training set, ensemble variance, and predicted property values provided a robust and interpretable default configuration. While alternative feature sets, algorithms, and data configurations occasionally improved performance for specific ADME datasets, these gains were generally modest compared with the improvements achieved over standard UQ metrics. Feature importance and explainability of EMs were analyzed to better understand sources of uncertainty. Such explainability analyses further showed that the three EM input features provide complementary information for reliability assessment and enable compound-level interpretation of uncertainty estimates. Taken together, EMs are presented as a tool that bridges prospective uncertainty estimation and error analyses, supporting the identification of model failures and enabling more informed decision-making in molecular ML.
ER -