🔬 DeepEntXAI — Explainable Multimodal Deep Learning for Anti-Enterobacteriaceae Bioactivity

A leakage-free, explainable multimodal deep-learning framework for predicting the bioactivity of chemical compounds against Enterobacteriaceae pathogens

---

Enterobacteriaceae (Escherichia coli, Klebsiella, Salmonella, Shigella, Enterobacter, and related genera) are a leading cause of hospital- and community-acquired infections, and their rising antimicrobial resistance is a major global-health threat. Experimentally screening every candidate compound against these pathogens is slow and costly, motivating reliable, interpretable computational models that can prioritise active compounds before the bench.

🧭 Project Overview

DeepEntXAI predicts whether a chemical compound is active or inactive against Enterobacteriaceae from its molecular structure, and — crucially — explains why. The framework fuses four complementary molecular representations in a CNN-LSTM architecture, is evaluated under a strictly leakage-free protocol, and couples every prediction with explainable-AI attributions and a validated compound-ranking for virtual screening.

Scientific-integrity note. An earlier version of this work reported ~99.8 % accuracy that arose from data leakage (PCA / scaling / feature selection fit on the full dataset before splitting, and a random row-wise split). This repository is the rebuilt, leakage-free pipeline. The honest headline is ROC-AUC 0.901 / accuracy 0.840 on a scaffold-disjoint hold-out, with a legitimate high-confidence statement of 98.1 % accuracy on the 22 % most-confident compounds via conformal selective prediction.

🧬 Analyses Performed

Data & Representation

• Genus-wide curation from ChEMBL and PubChem BioAssay across 29 species (49,093 high-confidence compounds).
• Activity-gap labelling — potent actives vs clearly-weak inactives, ambiguous middle removed.
• Multimodal features — Morgan-2048 + RDKit descriptors + MACCS-167 + ChemBERTa-768 (3,083-D), no PCA; selection fit train-only.

Modelling

• Multimodal fusion CNN-LSTM (per-modality encoders → residual Conv1D + squeeze-and-excite attention → bidirectional LSTM → sigmoid).
• Optuna hyper-parameter search; 5-fold ensemble as the headline model; isotonic calibration + OOF threshold tuning.
• Classical baselines (XGBoost, RandomForest) and stacking for comparison.

Leakage-free Evaluation

• Bemis–Murcko scaffold-disjoint split and scaffold-aware cross-validation (0 scaffold / 0 SMILES overlap verified).
• A leakage demonstration quantifying how much a random split inflates results.

Explainability & Screening

• Four XAI methods — permutation importance (modality + descriptor), Integrated Gradients, SHAP, LIME.
• Conformal selective prediction for an honest high-accuracy operating point.
• Validated compound ranking (DEEPENTXAI Score) with virtual-screening metrics.

Robustness (reviewer-driven)

• Highly-imbalanced evaluation, assay/source-bias analysis, provenance-disjoint external validation, and bootstrap + across-fold feature-importance stability.

📊 Key Results (all leakage-free, scaffold-disjoint hold-out)

Metric	Value
ROC-AUC	0.901
Accuracy	0.840
MCC	0.655
Selective accuracy @ 22 % coverage (conformal)	0.981
Ranking — top-1 % that are true actives	100 %
Enrichment Factor @0.1 % at 1:50 imbalance	41.9×

The deep ensemble outperforms classical baselines (XGBoost 0.886, RandomForest 0.872 ROC-AUC) on identical features and splits.

🗂️ Repository Contents

File	Description
0_Codes.ipynb	The complete pipeline in one notebook — the engine (deepentxai/) and all numbered stages (01–15): data curation, multimodal features, scaffold-disjoint split, CNN-LSTM ensemble, conformal prediction, XAI, ranking, and the robustness analyses.
Data_DeepEntXAI.xlsx	The curated activity-gap dataset — 49,093 compounds (smiles, label, n_measurements, frac_active, sources).
README.md	This file.
LICENSE.md	MIT License.

Requirements: numpy · pandas · scikit-learn · tensorflow-cpu · keras · xgboost · rdkit · optuna · shap · lime (Python 3.10; CPU-only, no GPU required).

🔍 Use Cases

• Virtual screening — prioritise active candidates against Enterobacteriaceae.
• Hit triage — act on high-confidence predictions, abstain on the rest (conformal).
• Medicinal-chemistry insight — interpretable descriptor attributions (e.g., carboxylic-acid fr_COO).
• Methodological reference — a template for leakage-free, honestly-reported bioactivity ML.

👩‍🔬 Authors

Nagmi Bano, Dr Shaban Ahmad, Prof Khalid Raza Computational Intelligence and Bioinformatics Lab, Department of Computer Science, Jamia Millia Islamia, New Delhi, India.

ORCID links: add here.

📄 License

Released under the MIT License.