StackPR

StackPR identifies progesterone receptor (PR) antagonists from SMILES using a stacked ensemble of machine learning models for large-scale in silico discovery.

Key Features:

• Stacked ensemble learning: Integrates multiple selected baseline models into a final meta-predictor using a stacking strategy.
• SMILES-based input: Operates on SMILES notation without requiring 3D structural information.
• Machine learning algorithms: Builds models using logistic regression, partial least squares, k-nearest neighbor, support vector machine, extremely randomized trees, and random forest.
• Molecular descriptors: Utilizes twelve conventional molecular descriptors as input features.
• Baseline model generation: Produces 72 baseline models from combinations of descriptors and algorithms.
• Model selection: Applies a genetic algorithm combined with a self-assessment-report approach to select the most effective baseline models.
• Validation: Validates models using tenfold cross-validation.
• Predictive performance: Achieves accuracy of 0.966 and a Matthews correlation coefficient (MCC) of 0.925 on independent test datasets.
• Feature analysis: Employs the SHapley Additive exPlanation (SHAP) algorithm and molecular docking for feature importance analysis.
• Structural determinants: Identifies aliphatic hydrocarbons and nitrogen-containing substructures as critical for PR antagonist activity.

Scientific Applications:

• In silico screening: Large-scale identification and prioritization of candidate PR antagonists from compound libraries.
• Experimental prioritization: Prioritizes computational hits for follow-up experimental validation.
• Cancer research: Supports studies of progesterone receptor modulation relevant to oncology.

Methodology:

Uses SMILES and twelve conventional molecular descriptors to generate 72 baseline models with six algorithms (logistic regression, partial least squares, k-nearest neighbor, support vector machine, extremely randomized trees, random forest); selects models via a genetic algorithm combined with a self-assessment-report approach; integrates selected models into a stacked meta-predictor and validates performance by tenfold cross-validation.