QSAR Workflow (RDKit + Deep Learning)

A reproducible QSAR modeling workflow combining RDKit molecular descriptors, MACCS fingerprints, outlier detection, applicability domain analysis, and a Keras-based neural network for regression tasks.

This repository contains a cleaned and fixed version of an original QSAR script, with improved stability, validation, and error handling.

---

📌 Overview

Author: Divya Karade
Model type: Regression (Neural Network)
Use case: Binding affinity / docking score prediction
Descriptors: RDKit 2D descriptors + MACCS keys
Outlier detection: PCA + Isolation Forest
Applicability Domain (AD): Descriptor range-based check
Frameworks: RDKit, scikit-learn, TensorFlow / Keras

---

🧪 Workflow Steps

1. Data Loading

   • Reads molecular data from a CSV file (SpikeRBD_DD.csv)
   • Requires SMILES strings and a numeric target value

2. Descriptor Generation

   • RDKit 2D molecular descriptors
   • MACCS fingerprints (167 bits)

3. Feature Cleaning

   • Removes NaN and infinite values
   • Clips extreme descriptor values
   • Ensures numeric consistency

4. Train/Test Split

   • 70/30 split with fixed random seed
   • Standard scaling applied to features

5. Outlier Detection

   • PCA (2 components) for dimensionality reduction
   • Isolation Forest to remove anomalous samples

6. Applicability Domain (AD)

   • Checks whether input molecule descriptors fall within the min–max range of the training set

7. Model Training

   • Fully connected neural network (Keras)
   • Early stopping based on validation R²
   • Custom metrics: RMSE and R²

8. Evaluation

   • MAE, MSE, RMSE, and R² on training and test sets

9. Prediction

   • Predicts binding affinity for new SMILES
   • Outputs applicability domain status

---

📂 Input Data Format

Training Dataset (SpikeRBD_DD.csv)

Required columns:

• smiles — SMILES representation of molecules
• DockingScore — Target regression value

Example:

smiles, DockingScore
CCO,-6.5
CCN,-7.1

---

🧬 Predicting New Molecules

New molecules can be provided as a Python list:

smiles = ["CCO"] # Example: ethanol

The script performs the following steps:

• Validates SMILES strings
• Generates RDKit 2D descriptors and MACCS fingerprints
• Checks applicability domain
• Predicts binding affinity using the trained neural network

📈 Model Architecture

Input Layer: Descriptor dimension Dense Layer: 600 neurons (ReLU) Dense Layer: 100 neurons (ReLU) Dense Layer: 100 neurons (ReLU) Output Layer: 1 neuron (Linear)

Training configuration

• Loss: Mean Squared Error (MSE)
• Optimizer: Adam
• Batch size: 400
• Epochs: Up to 200 (early stopping enabled)

---

📊 Evaluation Metrics

The following metrics are reported for both training and test sets:

• Mean Absolute Error (MAE)
• Mean Squared Error (MSE)
• Root Mean Squared Error (RMSE)
• Coefficient of Determination (R²)

---

🛠 Requirements

• Python ≥ 3.8
• RDKit
• TensorFlow / Keras
• scikit-learn
• pandas
• numpy
• matplotlib

Installation (example)

pip install numpy pandas scikit-learn tensorflow matplotlib conda install -c conda-forge rdkit

⚠️ Notes & Limitations

• Applicability domain is based on descriptor range checks, not distance-based confidence metrics.
• Isolation Forest contamination rate is fixed at 10%.
• Designed for QSAR regression, not classification.
• Best used with chemically consistent datasets.

---

📜 License

This code is provided for research and educational purposes.
Please cite appropriately if used in academic or industrial work.

---

👩‍🔬 Author

Divya Karade
Cheminformatician | QSAR | AI/ML