LGBM-BBB: A Web-Server for Blood-Brain Barrier Permeability Prediction using Winsorized Feature Engineering and Mathew Correlation Coefficient Threshold Optimisation

Technical Summary: LGBM-BBB Web-Server for Blood-Brain Barrier Permeability Prediction

Problem Statement
The development of drugs targeting the central nervous system (CNS) is significantly hindered by the blood-brain barrier (BBB), which blocks approximately 98% of small-molecule therapeutics. While experimental methods (in-vivo and in-vitro) exist to assess permeability, they are time-intensive, low-throughput, and ethically complex. Existing in-silico models often suffer from limitations such as the use of SMOTE (which may introduce noise), reliance on high-dimensional sparse representations (e.g., one-hot encoded vocabulary descriptors), or computationally expensive graph-based approaches. Furthermore, previous studies have lacked comprehensive head-to-head comparisons using standardized data and practical deployment tools for experimental researchers.

Methodology
The study utilized the B3DB dataset, comprising 7,807 compounds with experimentally validated binary permeability classifications. The authors implemented a rigorous pipeline involving data curation, feature engineering, and model optimization:

1. Data Pre-processing and Curation:

   • Compounds were filtered to a molecular weight range of 150–800 Daltons, resulting in 7,298 compounds.
   • To address heterogeneous labeling across 50 distinct data sources, compounds were categorized into four groups (A–D) based on data consistency and threshold reporting, standardizing the classification using a logBB cut-off of -1 where applicable.
   • Winsorization: To handle outliers without data deletion, the 1st and 99th percentile thresholds were applied. This tamed extreme values (particularly in lipophilicity descriptors like SLogP) while preserving the sample size and central distribution.

2. Feature Engineering and Selection:

   • Descriptor Calculation: 1,613 initial descriptors (1D, 2D, and 3D) were generated using Mordred and RDKit, covering constitutional, topological, and geometric properties.
   • Filtering: Features with >5% missing values were removed; remaining gaps were filled via median imputation stratified by class. Near-constant features (variance < 0.01) were discarded.
   • Selection: A two-stage filtering process reduced the feature set to 59 non-redundant descriptors:
     • Distance Correlation: Retained features with a correlation > 0.25 to the target variable.
     • Spearman Correlation: Removed multicollinear features (ρ > 0.85), keeping the feature with the stronger target association.

3. Model Training and Evaluation:

   • Six algorithms were benchmarked: Random Forest, XGBoost, LightGBM (LGBM), CatBoost, Logistic Regression, and Support Vector Classifier (SVC).
   • Data was split into 80% training and 20% testing using stratified sampling, with 10-fold cross-validation for stability assessment.
   • Threshold Optimization: Instead of the default 0.5 threshold, the decision boundary was optimized by evaluating 100 values to maximize the Matthews Correlation Coefficient (MCC) on the test set.

4. Web Deployment:

   • The optimal model was deployed as a web server using Python Flask (backend) and React.js (frontend), hosted on Render cloud infrastructure.

Key Results

• Model Performance: Gradient boosting methods outperformed linear models (Logistic Regression, SVC). LightGBM (LGBM) achieved the best balance of performance and computational speed, reaching 90.1% accuracy and a 0.953 ROC-AUC.
• Metric Comparison: The boosting models showed a significant gain in MCC (approx. 21% higher than linear models), indicating superior handling of class imbalance and complex non-linear interactions.
• Threshold Optimization: Optimizing the threshold for MCC yielded only marginal statistical gains (e.g., LGBM MCC increased from 77.76% to 78.43%), suggesting the model's natural decision boundary was already well-separated.
• Feature Importance: SHAP analysis identified Topological Polar Surface Area (TopoPSA) as the dominant predictor, followed by electrotopological surface area and acidic compound counts. High polar surface area (>90 Ų) was strongly associated with non-permeability.
• Rule-Based Analysis: Compounds passing both Lipinski and Ghose filters showed a 74% permeability rate, compared to 32% for those failing both. However, rule-based filtering alone achieved only ~70% accuracy, underscoring the necessity of machine learning for higher precision.
• Comparison with Literature: The LGBM model achieved performance comparable to or better than previous studies using the same B3DB dataset (e.g., Jia et al. and Nguyen et al.), but without the computational overhead of graph-based inputs or the potential noise of SMOTE.

Key Contributions

1. Robust Feature Engineering: The implementation of winsorization to handle outliers and a rigorous distance correlation-based feature selection process to reduce dimensionality without losing critical information.
2. Benchmarking and Optimization: A systematic comparison of six algorithms with explicit threshold optimization for MCC, demonstrating that ensemble tree-based methods are superior for this specific task.
3. Accessible Tool: The development and public release of LGBM-BBB, a freely accessible web server that allows researchers to input SMILES strings or CSV files for rapid virtual screening without requiring local computational resources or technical expertise.
4. Interpretability: The provision of SHAP-based insights and a confidence score based on the proximity to the MCC threshold, aiding chemists in understanding prediction reliability.

Significance and Claims
The authors claim that this work provides a validated, computationally efficient, and accessible tool to address the critical bottleneck in neurotherapeutic discovery. By avoiding the complexity of SMOTE and graph-based descriptors, the model maintains high predictive power (90.1% accuracy) while being faster and easier to deploy. The study emphasizes that while rule-based filters (Lipinski/Ghose) offer a useful initial heuristic, machine learning is necessary to capture the multidimensional relationships governing BBB permeability. The web server is positioned as a practical resource to enable rapid screening of large chemical libraries, potentially redirecting resources toward promising candidates before expensive lead optimization. The authors modestly note that the model is designed for drug-like molecules and may not generalize to peptides or antibody-drug conjugates, and that species-specific physiological differences in the BBB remain a challenge not fully addressed by the aggregated B3DB dataset.