OPERA

OPERA predicts physicochemical and environmental fate properties of chemicals using QSAR/QSPR models for regulatory and research assessments.

Key Features:

• Source Data: Leverages curated datasets from the PHYSPROP database covering 13 common physicochemical and environmental fate properties.
• Data Quality and Curation: Implements an automated workflow for extensive curation and standardization of chemical structures prior to descriptor calculation.
• Descriptor Calculation and Selection: Calculates molecular descriptors with PaDEL and selects descriptors using genetic algorithms, typically using 2–15 descriptors (average 11) per model.
• Modeling Approach: Constructs models using a weighted k-nearest neighbor algorithm to prioritize interpretability and local similarity.
• Modeling Principles: Adheres to the five OECD QSAR principles, including defined endpoints, unambiguous algorithms, defined applicability domains, and measures of goodness-of-fit, robustness, and predictivity.
• Dataset Size and Validation: Handles datasets from ~150 to 14,050 chemicals (average 3,222) with models trained on 75% randomly selected training sets and validated using fivefold cross-validation and 25% test sets.
• Performance Metrics: Reports cross-validation Q² values ranging 0.72–0.95 (average 0.86) and R² on test sets ranging 0.71–0.96 (average 0.82).
• Regulatory Validation: Models have been validated by the European Commission's Joint Research Center as OECD compliant.

Scientific Applications:

• Regulatory assessment: Provides predicted physicochemical and environmental fate properties to support chemical safety and regulatory decision making.
• Large-scale screening: Has been applied to over 750,000 chemicals with predicted data integrated into the U.S. EPA CompTox Chemistry Dashboard.
• Environmental fate analysis: Used to estimate properties relevant to environmental persistence, bioaccumulation, and transport for research and policy.

Methodology:

Automated curation and standardization of chemical structures, calculation of molecular descriptors with PaDEL, descriptor selection via genetic algorithms, modeling with a weighted k-nearest neighbor algorithm, and validation using fivefold cross-validation with 75% training / 25% test splits, reported under OECD QSAR principles.