DeepEntXAI: A CNN-LSTM + Explainable AI Framework for Predicting Drug Activity Against Enterobacteriaceae
DeepEntXAI is a hybrid deep learning framework developed to predict the biological activity of chemical compounds against Enterobacteriaceae pathogens (e.g., E. coli, K. pneumoniae, S. enterica) using molecular descriptors. The framework combines a CNN-LSTM architecture with explainable AI (XAI) techniques to achieve state-of-the-art prediction accuracy while ensuring model interpretability. It supports both labelled and unlabelled compound predictions and includes tools for compound prioritisation and descriptor relevance evaluation.
-
Hybrid Deep Learning Model: Combines CNN and LSTM to capture both local and long-range dependencies in molecular descriptor sequences.
-
High Accuracy: Achieved 99.80% accuracy, 99.63% precision, and 99.94% recall on test data.
-
Generalisation: Demonstrated strong generalisation through 5-fold and 10-fold cross-validation with average accuracies of 99.09% ± 0.36% and 99.13% ± 0.27%, respectively.
-
Explainability: Integrates both:
- A hybrid perturbation-based XAI method (accuracy drop, flip ratio, log-loss increase)
- LIME for local interpretability and compound-specific insights.
-
Prediction on Unlabelled Data: Scores new compounds (0–10 scale) to identify high-potential drug candidates.
-
Output-Ready Results: Generates CSV files with prediction probabilities, scaled activity scores, and interpretability rankings.
DeepEntXAI/
├── data/ # Sample_Dataset.csv, SampleInput molecular descriptors (CSV format)
├── models/ # 1_CNN_LSTM_XAI.pth, Trained models and saved weights
├── notebooks/ # Enterobacteriaceae_Code.ipynb, Code with outputs (CNN-LSTM-XAI)
├── README.md # README.md, Project documentation
└── LICENSE # LICENSE.md, License file
-
Source: PubChem and ChEMBL BioAssays relevant to Enterobacteriaceae.
-
Size: 17,097 compounds with 2,997 descriptors each (selected 2000 with PCA).
-
Descriptors: Computed using RDKit, Mordred, and PaDEL.
-
Preprocessing:
- Structure standardisation
- Duplicate removal
- Descriptor curation
- and many more, read the full paper.
- Final active set: 8,097 ligands, 9000 Inactive compounds.
git clone https://github.com/ShabanAhmad/DeepEntXAI.git
cd DeepEntXAIpython -m venv venv
source venv/bin/activate # On Windows use: venv\Scripts\activate
pip install -r requirements.txtDirectly use the saved weight file.Pass any unlabelled data in CSV format. python src/explain_hybrid.py --input data/test.csv --model models/ _weights.h5
python src/explain_lime.py --input data/test.csv --model models/ _weights.h5| Metric | Score |
|---|---|
| Accuracy | 99.80% |
| Precision | 99.63% |
| Recall (Sensitivity) | 99.94% |
| F1-Score | 99.78% |
| Specificity | 99.67% |
| AUC-ROC | 0.998 |
| Cross-Validation (5-fold) | 99.09% ± 0.36% |
| Cross-Validation (10-fold) | 99.13% ± 0.27% |
- Top Feature:
minHsNH3p– High perturbation sensitivity across accuracy, flip ratio, and log-loss. - Low-Importance Features:
geomDiameter,nG12FRing,BCUTZ-1l - Top Compounds (by confidence):
4615770,5309708,70556, with scaled scores ~9.9999 - Least Certain Compounds: Predictions near 0.5 threshold, structurally ambiguous
- Active compound screening
- Compound efficacy assessment
- Drug repurposing
- Prioritisation for Enterobacteriaceae infections (e.g., UTIs, BSIs, sepsis)
- Feature-driven compound design
This project is licensed under the MIT License (read full license file).
We welcome contributions from the community. Please open issues or submit pull requests for improvements, bug fixes, or new features.
For questions or collaboration inquiries, please contact:
Author: \Nagmi Bano1, Dr Shaban Ahmad1,2, Prof Khalid Raza1
Email: \nagmi2300973@st.jmi.ac.in, Shaban@sund.ku.dk, kraza@jmi.ac.in
Institutions: *
1 Department of Computer Science, Jamia Millia Islamia, New Delhi-110025, India.
2 Biomedicine Section, Department of Veterinary and Animal Sciences, Faculty of Health and Medical Sciences, University of Copenhagen, Frederiksberg, Denmark.*