PaperlayerView exact revision on arXiv

Source-linked AI summary

An Explainable Machine Learning Framework for Predicting Blood-Brain Barrier Permeability Using Molecular Descriptors

Fatemeh Mahmoudi

arXiv:2609.10012v1cs.LGcond-mat.mtrl-sci

TL;DR

BBB permeability is important for CNS drug development because the selective barrier restricts therapeutic delivery, motivating efficient computational prediction. The study compares four classifiers using RDKit molecular descriptors and SHAP interpretation, with optimized XGBoost achieving the strongest reported performance. The framework is positioned as an interpretable approach for early-stage screening, but its scope is limited by two-dimensional descriptors and the absence of external validation.

  • Problem

    BBB permeability prediction matters because the selective BBB restricts delivery of therapeutic agents to the CNS, while experimental assessment is costly and time-consuming.

  • Method

    The study evaluates four classifiers using fifteen RDKit-derived physicochemical descriptors from 2,039 BBBP compounds, optimizes XGBoost, and applies SHAP for interpretability.

  • Results

    88.97% accuracy, 93.13% F1-score, and 0.9282 ROC–AUC were achieved by the optimized XGBoost classifier, the best-performing evaluated model.

  • Takeaways & Limitations

    The framework provides an accurate and interpretable approach for BBB permeability prediction and may support early-stage CNS drug-candidate screening.

  • Takeaways & Limitations

    The study uses only two-dimensional molecular descriptors and does not include external validation, limiting the assessed representation scope and generalizability evidence.

Abstract

from arXiv · show

Blood-brain barrier (BBB) permeability is a critical determinant in the development of central nervous system therapeutics because it directly influences the ability of drug candidates to reach their target sites within the brain. In this study, an explainable machine learning framework was developed to predict BBB permeability using molecular descriptors generated from the MoleculeNet BBBP dataset with the RDKit cheminformatics toolkit. Fifteen physicochemical descriptors extracted from 2,039 compounds were used to train four supervised machine learning algorithms, including Logistic Regression, Support Vector Machine (SVM), Random Forest, and Extreme Gradient Boosting (XGBoost). Hyperparameter optimization was performed using GridSearchCV, while model interpretability was investigated using SHapley Additive exPlanations (SHAP). Among the evaluated models, the optimized XGBoost classifier achieved the best predictive performance, with an accuracy of 88.97%, a precision of 88.92%, a recall of 97.76%, an F1-score of 93.13%, and a ROC-AUC of 0.9282. Stratified five-fold cross-validation further demonstrated the robustness of the proposed model, yielding a mean ROC-AUC of 0.8982 +/- 0.0130. Feature importance and SHAP analyses consistently identified TPSA, HBD, and LogP as the most influential molecular descriptors governing BBB permeability prediction. Overall, the proposed framework provides an accurate, interpretable, and computationally efficient approach for BBB permeability prediction and may serve as a valuable tool for the early-stage screening of CNS drug candidates.

1. Introduction

BBB permeability prediction matters for CNS drug discovery because the BBB restricts therapeutic delivery, while experimental assays are costly and slow. This study combines molecular-descriptor machine learning with SHAP interpretability to support efficient early-stage screening.

  • BBB permeability is critical in CNS drug discovery because the selective barrier restricts delivery of many therapeutic agents.
  • Computational prediction can prioritize candidate molecules before expensive, labor-intensive, and time-consuming experimental validation.
  • Many accurate machine-learning models remain black boxes, limiting insight into the molecular features driving BBB permeability predictions.
  • SHAP-based explainability reveals descriptor contributions and can provide mechanistic insight into physicochemical properties governing BBB permeability.
  • The framework evaluates four supervised algorithms with RDKit descriptors, optimizes XGBoost, applies SHAP, and uses stratified five-fold cross-validation.
  • The proposed approach is presented as accurate, interpretable, and computationally efficient for early-stage BBB permeability prediction and drug-candidate screening.

2. Materials and Methods

The study uses RDKit-derived physicochemical descriptors from experimentally annotated BBBP molecules to compare four classifiers. XGBoost is optimized with GridSearchCV, interpreted with SHAP, and assessed for robustness using stratified cross-validation.

  • Dataset and descriptors: The BBBP dataset retained 2,039 experimentally annotated compounds represented by SMILES strings and binary permeability labels.
  • Dataset and descriptors: Fifteen RDKit-derived descriptors characterize molecular weight, lipophilicity, polarity, hydrogen bonding, topology, rings, and related structural properties.
  • Machine-learning models: Logistic Regression, SVM, Random Forest, and XGBoost were compared using an 80:20 reproducible train–test split.
  • Evaluation and optimization: Performance was assessed with accuracy, precision, recall, F1-score, ROC–AUC, confusion matrices, and ROC curves, with ROC–AUC as the primary indicator.
  • Evaluation and optimization: GridSearchCV systematically explored XGBoost hyperparameters using stratified five-fold cross-validation for parameter selection.
  • Interpretability and robustness: SHAP values quantified global and local descriptor contributions through summary and waterfall plots.
  • Interpretability and robustness: Stratified five-fold validation preserved class distributions across folds to estimate stability, robustness, generalization, and resistance to overfitting.

3. Results

Random Forest feature-importance analysis ranks molecular descriptors associated with BBB permeability prediction. TPSA is the leading descriptor, followed by heteroatom count, LogP, HBD, and molecular weight.

  • Molecular Descriptor Analysis: The analysis used Random Forest feature-importance scores summarized in Table 2 and visualized in Figure 1.
  • Molecular Descriptor Analysis: TPSA ranked first in Random Forest feature importance with a score of 0.1569, followed by NumHeteroatoms, LogP, HBD, and ExactMolWt.The reported subsequent scores were 0.1057, 0.1003, 0.0896, and 0.0735, respectively.

3.2 Model Performance

The optimized XGBoost classifier delivered the strongest overall BBB permeability performance among the evaluated models. Ensemble methods outperformed the conventional baseline and alternative classifiers on the reported comparison.

  • Model Performance: 88.97% accuracy, 88.92% precision, 97.76% recall, 93.13% F1-score, and 0.9282 ROC–AUC were achieved by optimized XGBoost.
  • Model Performance: 83.58% accuracy and 0.8866 ROC–AUC were achieved by Logistic Regression as the baseline classifier.
  • Model Performance: SVM achieved the highest recall at 98.08%, while ensemble learning algorithms consistently outperformed conventional machine-learning models overall.

3.3 Feature importance Analysis

Feature-importance and SHAP analyses identified TPSA, HBD, NumHeteroatoms, and LogP as the descriptors most strongly associated with BBB permeability predictions.

  • 3.3 Feature importance Analysis: TPSA was the most influential descriptor, followed by NumHeteroatoms, LogP, HBD, and Exact Molecular Weight in Random Forest analysis.The reported importance scores were 0.1569, 0.1057, 0.1003, 0.0896, and 0.0735, respectively.
  • 3.3 Feature importance Analysis: Lower TPSA and fewer HBD generally increased predicted BBB permeability, whereas higher LogP tended to increase predicted BBB penetration.SHAP ranked TPSA, HBD, NumHeteroatoms, and LogP as the leading contributors in the optimized XGBoost model.
  • 3.3 Feature importance Analysis: For a representative compound, descriptor contributions shifted the baseline prediction from 0.766 to a final model output of 0.883.HBD, MolecularWeight, TPSA, NumHeteroatoms, ExactMolWt, and RotatableBonds positively influenced the prediction, while LogP slightly reduced it.

3.5. Receiver Operating Characteristic Analysis

ROC analysis showed that the optimized XGBoost classifier had the strongest discriminative performance among the evaluated models for BBB permeability prediction.

  • 3.5. Receiver Operating Characteristic Analysis: The optimized XGBoost model achieved the highest ROC–AUC at 0.9282, ahead of XGBoost, Random Forest, SVM, and Logistic Regression.The corresponding ROC–AUC values were 0.9192, 0.9163, 0.9117, and 0.8866, respectively.

3.6. Confusion Matrix Analysis

The optimized XGBoost classifier correctly identified most BBB-permeable and non-permeable compounds, with relatively few classification errors.

  • 3.6. Confusion Matrix Analysis: The confusion matrix recorded 302 true positives and 60 true negatives, alongside 10 false negatives and 36 false positives.The results were described as demonstrating strong overall classification performance and a favorable balance between sensitivity and specificity.
  • 3.6. Confusion Matrix Analysis: The model’s relatively small number of false negatives supported high sensitivity during early-stage BBB-permeability screening.The cited analysis links this error profile to a reduced likelihood of overlooking potentially BBB-permeable compounds.

3.8 Cross-Validation Results

Stratified five-fold cross-validation indicated that the optimized XGBoost model maintained stable predictive performance across validation folds.

  • 3.8 Cross-Validation Results: The optimized XGBoost model achieved a mean ROC–AUC of 0.8982 ± 0.0130 across stratified five-fold cross-validation.The small standard deviation and stable fold performance were reported as evidence of consistent generalization across data partitions.

3.9. Overall Model Performance

Ensemble methods outperformed conventional machine learning approaches for BBB permeability prediction, with optimized XGBoost providing the strongest overall discrimination and gradient boosting capturing nonlinear descriptor relationships.

  • 0.9282 ROC-AUC was achieved by optimized XGBoost, the highest among evaluated models for distinguishing BBB-permeable from non-permeable compounds.The result indicates superior discrimination after hyperparameter optimization.
  • Random Forest and XGBoost produced superior overall classification performance, whereas Logistic Regression had the lowest predictive accuracy.
  • Support Vector Machine achieved the highest recall, demonstrating excellent sensitivity among the evaluated models.
  • Gradient boosting algorithms were particularly effective for capturing nonlinear relationships between molecular descriptors and BBB permeability.

4. Discussion

The discussion links strong BBB prediction performance to ensemble learning and descriptor-based modeling, while emphasizing physicochemical interpretability and acknowledging limits in representation and external validation.

  • 0.9282 ROC-AUC and 0.9313 F1-score were achieved by optimized XGBoost using fifteen RDKit-generated molecular descriptors.The findings align with prior reports that ensemble and gradient-boosting methods perform strongly on BBB permeability prediction.
  • 0.1569 feature importance was assigned to TPSA, followed by NumHeteroatoms at 0.1057, LogP at 0.1003, HBD at 0.0896, and Exact Molecular Weight at 0.0735.Lower TPSA generally facilitates membrane permeation, while appropriate lipophilicity improves passive diffusion.
  • Hydrogen-bonding capacity directly affects BBB penetration, explaining the substantial contribution of HBD in the model.
  • SHAP provided global descriptor rankings and local explanations showing how individual descriptors collectively affected a representative compound’s prediction.
  • 0.8982 mean ROC-AUC was obtained in stratified five-fold cross-validation, while the confusion matrix showed a favorable sensitivity-specificity balance.The relatively small number of false negatives reduced the likelihood of overlooking potentially BBB-permeable compounds during early screening.
  • The study used only two-dimensional RDKit descriptors and did not perform external validation with an independent dataset.Future work may incorporate richer molecular representations and independent validation datasets.

5. Conclusion

The study concludes that combining machine learning with explainable artificial intelligence yields an accurate and interpretable BBB permeability framework, while recognizing representation and validation boundaries.

  • The framework systematically evaluated Logistic Regression, Support Vector Machine, Random Forest, and XGBoost for BBB permeability prediction.
  • 88.97% accuracy, 93.13% F1-score, and 0.9282 ROC-AUC made optimized XGBoost the best-performing evaluated classifier.
  • TPSA, HBD, LogP, and the number of heteroatoms were consistently identified as the most influential descriptors governing BBB permeability.
  • Global and local SHAP interpretations improved transparency by explaining molecular descriptor contributions to model predictions.
  • The study relied exclusively on two-dimensional descriptors and lacked external validation, limiting the demonstrated scope of generalizability.Future studies may add fingerprints, graph neural networks, multimodal representations, and independent validation datasets.
  • The framework may serve as a computational tool for accelerating early-stage screening and prioritization of CNS drug candidates.
Loading 2609.10012v1…