ProQSAR: A modular and reproducible framework for small-data QSAR modeling with fit-and-use models

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Quantitative structure-activity relationship (QSAR) models are vital for drug discovery and toxicology.
• Practical QSAR adoption is hindered by inconsistent tools, validation, and reproducibility issues.

Purpose of the Study:

• Introduce ProQSAR, a modular and reproducible workbench for end-to-end QSAR development.
• Enable independent use of QSAR components for flexibility.
• Formalize QSAR workflows for improved adoption.

Main Methods:

• ProQSAR integrates modules for data standardization, feature generation, splitting (scaffold/cluster-aware), preprocessing, outlier handling, scaling, feature selection, and model training/tuning.
• Includes statistical comparison, conformal calibration, and applicability-domain assessment.
• End-to-end pipeline generates versioned artifacts and reports for deployment and audit.

Main Results:

• Achieved state-of-the-art descriptor-based performance on MoleculeNet benchmarks, including lowest mean RMSE on regression tasks (e.g., FreeSolv RMSE 0.494).
• Demonstrated superior performance on QM7 and competitive results on QM8 quantum mechanical benchmarks.
• Achieved top ROC-AUC on ClinTox (91.4%) and competitive classification performance overall.
• All predictions include cross-conformal prediction and applicability-domain flags for calibrated, decision-supportive outputs.

Conclusions:

• ProQSAR enforces best-practice validation and statistical comparisons.
• Integrates uncertainty quantification and applicability-domain diagnostics for risk-aware predictions.
• Offers a composable API and a one-click pipeline for generating deployment-ready artifacts and reports.