BACE1 QSAR Modeling Using RDKit and Machine Learning
A machine learning-based QSAR(Quantitative Structure–Activity Relationship) model developed to predict the inhibitory activity (pIC50) of BACE1 inhibitors using molecular descriptors calculated with RDKit.
Overview:
Beta-secretase 1 (BACE1) is an important therapeutic target in Alzheimer's disease because it initiates the cleavage of amyloid precursor protein, leading to the formation of β-amyloid peptides. Computational approaches such as QSAR can be used to relate molecular properties to biological activity and assist in early-stage drug discovery. This project implements a basic QSAR workflow using publicly available BACE1 bioactivity data from ChEMBL.
Datasets used:
- ChEMBL Database
- Target : BACE1
- Bioactivity: IC50
- Response variable: pIC50
QSAR Pipeline:
ChEMBL Database │ ▼ Download BACE1 Bioactivity Data │ ▼ Filter IC50 Measurements │ ▼ Remove Duplicate Compounds │ ▼ Convert SMILES → Molecular Structures │ ▼ Calculate Molecular Descriptors (MolWt, LogP, TPSA, HBD, HBA, RotBonds) │ ▼ Generate pIC50 Values │ ▼ Train-Test Split │ ▼ Random Forest Regression │ ▼ Model Evaluation (R², MAE, RMSE)
Model performance:
| Metric | Value |
|---|---|
| R² Score | 0.429 |
| MAE | 0.734 |
| RMSE | 0.941 |
Planned enhancements include:
- RDKit descriptor set (~200 descriptors)
- Morgan (ECFP4) fingerprints
- Feature selection
- Hyperparameter optimization
- Cross-validation
- External validation
- Model comparison (Random Forest, XGBoost, SVR)
- Applicability domain analysis
- Virtual screening of novel compoun