Abstract
Progesterone receptors (PRs) are implicated in various cancers since their presence/absence can determine clinical outcomes. The overstimulation of progesterone can facilitate oncogenesis and thus, its modulation through PR inhibition is urgently needed. To address this issue, a novel stacked ensemble learning approach (termed StackPR) is presented for fast, accurate, and large-scale identification of PR antagonists using only SMILES notation without the need for 3D structural information. We employed six popular machine learning (ML) algorithms (i.e., logistic regression, partial least squares, k-nearest neighbor, support vector machine, extremely randomized trees, and random forest) coupled with twelve conventional molecular descriptors to create 72 baseline models. Then, a genetic algorithm in conjunction with the self-assessment-report approach was utilized to determine m out of the 72 baseline models as means of developing the final meta-predictor using the stacking strategy and tenfold cross-validation test. Experimental results on the independent test dataset show that StackPR achieved impressive predictive performance with an accuracy of 0.966 and Matthew's coefficient correlation of 0.925. In addition, analysis based on the SHapley Additive exPlanation algorithm and molecular docking indicates that aliphatic hydrocarbons and nitrogen-containing substructures were the most important features for having PR antagonist activity. Finally, we implemented an online webserver using StackPR, which is freely accessible at http://pmlabstack.pythonanywhere.com/StackPR . StackPR is anticipated to be a powerful computational tool for the large-scale identification of unknown PR antagonist candidates for follow-up experimental validation.
Introduction
The progesterone receptor (PR) is a critical biomarker and therapeutic target in the progression of breast, ovarian, and endometrial cancers. Modulating this receptor with antagonists is a primary strategy for cancer treatment; however, traditional drug discovery is a costly and time-intensive venture. While computational methods like molecular docking exist, they often require high-resolution 3D structural data, which may not be available for all compounds and can be computationally prohibitive for large-scale screening.
To overcome these hurdles, researchers have developed StackPR, a novel machine learning framework designed to identify PR antagonists with high precision using only 1D chemical strings (SMILES). This approach enables rapid, large-scale virtual screening of chemical libraries without the need for complex structural modeling.
Methodology: The Stacking Approach
The core of StackPR lies in its stacked ensemble learning strategy. Rather than relying on a single algorithm, the framework builds a hierarchy of models to capture diverse chemical features.
1. Baseline Model Construction
The researchers first generated 72 baseline models by pairing six machine learning algorithms (Logistic Regression, Partial Least Squares, K-Nearest Neighbor, Support Vector Machine, Extremely Randomized Trees, and Random Forest) with twelve different molecular fingerprints, such as PubChem, MACCS, and Circle descriptors.
2. Feature Selection and Optimization
To refine the ensemble, a Genetic Algorithm coupled with a self-assessment-report (GA-SAR) approach was used to identify the 10 most effective baseline models from the original 72. This step minimizes noise and redundancy, ensuring the final predictor is both efficient and accurate.
3. Meta-Predictor Development
The outputs of these 10 optimal models were fed into a final "meta-predictor" based on a Random Forest algorithm. This stacking method allows the final model to learn from the collective strengths and mitigate the individual biases of the underlying sub-models.
4. Dataset and Validation
The study utilized data from the ChEMBL database (Target ID: CHEMBL208). After rigorous curation based on $IC_{50}$ thresholds, a dataset of 463 active and 260 inactive compounds was established. This was split into a training set (TRN515) for cross-validation and an independent test set (IND145) to evaluate the model's performance on unseen data.
Results and Discussion
Predictive Performance
StackPR demonstrated exceptional robustness on the independent test set, achieving an accuracy of 0.966 and a Matthew’s Correlation Coefficient (MCC) of 0.925. Notably, the model achieved a sensitivity of 0.978, indicating it is highly effective at identifying true antagonists.
As shown in the performance analysis of the baseline models, while individual combinations like ET-Circle performed well, the final StackPR meta-predictor outperformed the best individual baseline model by approximately 4.3% in MCC. This confirms that the stacking strategy successfully integrates diverse chemical information to improve predictive power.
Chemical Interpretability and Discovery
To move beyond a "black box" model, the team employed SHAP (SHapley Additive exPlanation) analysis. This revealed that aliphatic hydrocarbons and nitrogen-containing substructures are the most critical features for PR antagonist activity, aligning with the pharmacology of known drugs like mifepristone.
Furthermore, molecular docking simulations confirmed that the top-scoring candidate identified by the model (CHEMBL259879) possesses a higher binding affinity ($-12.67 \text{ kcal/mol}$) than the established modulator asoprisnil, marking it as a promising candidate for future experimental study.
Conclusion
StackPR represents a significant advancement in the computational screening of progesterone receptor antagonists. By leveraging a stacking ensemble strategy and 1D SMILES notation, it provides a fast and highly accurate alternative to traditional 3D-based methods. While the tool is primarily for prioritization and requires follow-up wet-lab validation, its high sensitivity makes it an ideal first step in drug discovery. The authors have made the tool freely available as an online webserver to support the broader research community.
Summary updated