AIREP: AI-Driven Inhibitor Recommendation for Beta-lactamase TEM

---

Project Overview

AIREP (AI-driven RECommendation of inhibitors) is a machine learning project aimed at predicting potential small molecule inhibitors for the enzyme Beta-lactamase TEM, a key contributor to antibiotic resistance. The model leverages BindingDB experimental data and molecular fingerprints to learn structure-activity relationships and recommend promising candidates from chemical libraries.

---

Motivation

Antibiotic resistance is a growing global health threat. Beta-lactamase TEM enzymes degrade β-lactam antibiotics, rendering many treatments ineffective. Identifying potent inhibitors to block this enzyme is critical. Experimental screening is costly and slow; AI-driven prediction accelerates discovery by narrowing candidates for lab testing.

---

Project Components

• Data Processing:
  Cleans and filters BindingDB data specific to Beta-lactamase TEM, extracting molecular structures (SMILES) and experimental affinities (Kd, IC50, Ki).

• Feature Engineering:
  Converts SMILES to RDKit Morgan fingerprints (2048 bits, radius 2) as input features.

• Modeling:
  Trains a Random Forest regressor using hyperparameter tuning (RandomizedSearchCV). Predicts binding affinity as pAffinity = -log10(affinity in M).

• Evaluation:
  Model performance assessed with RMSE and R² on validation and test splits.

• Recommendation Engine:
  Scores candidate molecules using the trained model and ranks them by predicted affinity.

• Sequence Filtering:
  Filters target-like sequences from provided FASTA files based on keywords and length.

• Web App:
  Flask interface to upload candidate molecules and obtain top inhibitor predictions.

---

Installation

1. Clone the repository: git clone https://github.com/IKKNIGHT/airep.git cd airep

2. Install dependencies: pip install -r requirements.txt

3. Ensure you have the following files in the project root:

   • BindingDB_All.tsv (BindingDB dataset)
   • BindingDBTargetSequences.fasta (target protein sequences)

4. Ensure you're .env file has a secret key for flashing messages

   • for ex. SECRET_KEY=supersecretkey

---

Usage

Training and Running the Web App

Run the script to train the model (if not cached) and start the Flask server:

python3 main.py

• The model trains automatically on Beta-lactamase TEM data.
• Once running, open your browser at http://127.0.0.1:5000.
• Upload candidate SMILES and optionally a FASTA file (not required for Beta-lactamase TEM as local FASTA is used).
• Select top K recommendations and submit.

---

Results and Visualizations

Model Performance

• Validation RMSE: ~0.795
• Test RMSE: ~0.901
• Test R²: ~0.741

These metrics indicate the model predicts binding affinities with reasonable accuracy suitable for guiding experimental screening.

Example Predictions

SMILES	pAffinity	Predicted nM
[O-]C(=O)C1=CS[C@H]2N1C(=O)\C2=C/c1cn2CCOCc2n1	9.052	0.9
[O-]C(=O)C1=CS[C@H]2N1C(=O)\C2=C\c1cnc2COCCn12	9.025	0.9
ICC1CC(C(=O)O1)=C1O[C@@H]2CC(=O)N2C1C(=O)OCc1ccccc1	9.009	1.0

Generated Plots

Validation: Predicted vs Actual

---

Test: Predicted vs Actual

---

Test Residuals

---

Top 20 Feature Importances

---

Top 20 Feature Importances (Table)

---

Project Structure

airep/
│
├── airep.py # Main script (training + Flask app)
├── BindingDB_All.tsv # BindingDB dataset (large TSV)
├── BindingDBTargetSequences.fasta # Target protein sequences
├── cache_.pkl # Cached filtered data by target
├── features_.npz # Cached fingerprint features
├── model_beta-lactamase_tem.joblib # Saved trained model
├── plots/ # Generated evaluation plots
│ ├── pred_vs_actual_val.png
│ ├── pred_vs_actual_test.png
│ ├── residuals_test.png
│ └── feature_importance.png
├── requirements.txt # Python dependencies
└── README.md # This file

---

Future Work

• Expand model to other targets with minimal changes.
• Incorporate deep learning models (e.g., Graph Neural Networks) for improved affinity prediction.
• Integrate with experimental workflows for lab validation of predictions.
• Develop a more interactive web UI with visualization of molecular structures.

---

References

• BindingDB
• RDKit: Open-source cheminformatics software
• Scikit-learn: Machine learning in Python

---

Acknowledgements

Special thanks to the developers of RDKit, Scikit-learn, and the BindingDB team for making data and tools freely available.

---

Contact

For questions or collaboration, please contact:
Isaaq Khanooni — mohammadIsaaqK@gmail.com

---

This project was developed for ISEF competition — combining AI, chemistry, and bioinformatics for drug discovery.