PREDINHIB_MTB

PREDINHIB_MTB is a machine learning pipeline for predicting compound inhibition and MIC values against Mycobacterium tuberculosis (M. tuberculosis), leveraging curated assay descriptions processed through Large Language Models (LLMs).

🔍 Overview

This repository supports two key tasks:

1. MIC Regression using curated, biologically meaningful datasets.
2. Multi-task Classification of compound activity across different resistance profiles (Non-Resistant, Resistant, MDR).

We use LLMs to extract structured metadata from free-text assay descriptions from ChEMBL, enabling biologically consistent dataset construction and improved model performance.

📁 Repository Structure

.
├── data/
│   ├── chemblapi/           # ChEMBL data for Mycobacterium tuberculosis
│   ├── cv/                  # Repeated 5x5 cross-validation folds
│   ├── llama/               # Llama3.3 input/output
│   └── res_mdr_xdr/         # Data for multi-task classification
│
├── figures/                 # Generated plots and figures
│
├── models/                  # Trained models and configuration files
│
├── pipeline/
│   ├── 0_chembl_download.ipynb         # Download data from ChEMBL
│   ├── 1_data_exploration.ipynb        # Initial dataset inspection
│   ├── 2_data_cleaning.ipynb           # Filtering and cleaning
│   ├── 3_llama_processing/
│   │   └── 3_llama_processing.py       # Llama3.3 metadata extraction
│   ├── 4_data_splitting/
│   │   ├── 4_1_llama_results_exploration.ipynb  # Metadata exploration
│   │   └── 4_2_ml_preparation.ipynb             # ML dataset preparation
│   ├── 5_h37rv_nr_raw/
│   │   ├── 5_0_folds_creation.ipynb             # Stratified fold generation
│   │   ├── fingerprints.ipynb                   # Morgan fingerprint generation
│   │   ├── kpgt_embeddings.ipynb                # KPGT embedding generation
│   │   ├── 5_2_kpgt_embeddings.ipynb             # Alternative KPGT generation
│   │   ├── fp_RF/run_all_rf.sh                   # Random Forest training
│   │   ├── fp_SVR/run_all_svr.sh                  # SVR training
│   │   └── fp_DNN/run_all_dnn.sh                  # DNN training
│   └── 6_multitask/
│       ├── data_curation.ipynb                  # Multi-task dataset curation
│       ├── model_run.ipynb                      # Multi-task training/evaluation
│       └── multitask_test_set.csv               # Test set for classification
│
├── results/                    # Model predictions and CV results
├── utils.py                     # Utility functions
└── predinhib_env.yml            # Conda environment file

🧪 How to Reproduce

1. Environment Setup

conda env create -f predinhib_env.yml
conda activate predinhib

2. Data Preparation

(Optional) Download and process the latest ChEMBL data:

1. 0_chembl_download.ipynb
2. 1_data_exploration.ipynb
3. 2_data_cleaning.ipynb

Note: This step is not needed for pure reproducibility of current results.

3. Assay Metadata Extraction (LLaMA)

Run 3_llama_processing/3_llama_processing.py to extract structured metadata (strain, resistance, mutant type) from assay descriptions.

We recommend running Ollama inside a container for GPU acceleration:

podman pull ollama
podman run -d --name ollama_cuda --privileged --gpus all -v ollama_data:/root/.ollama -p 11434:11434 docker.io/ollama/ollama
podman exec -it ollama bash
ollama pull llama3.3
exit

4. Metadata Exploration and Dataset Preparation

• Explore extracted metadata: 4_data_splitting/4_1_llama_results_exploration.ipynb
• Prepare datasets for ML: 4_data_splitting/4_2_ml_preparation.ipynb

5. Single-task Model Training and Evaluation

• Generate folds: 5_h37rv_nr_raw/5_0_folds_creation.ipynb
• Generate fingerprints: 5_h37rv_nr_raw/fingerprints.ipynb
• Generate KPGT embeddings: 5_h37rv_nr_raw/kpgt_embeddings.ipynb or 5_2_kpgt_embeddings.ipynb

Note on KPGT embeddings:
Clone KPGT, install its environment, and download its pre-trained model from figshare.
Update the paths to the KPGT repository, model, and environment as needed.

• Train and evaluate models using the bash scripts:
  • run_all_rf.sh (Random Forest)
  • run_all_svr.sh (SVR)
  • run_all_dnn.sh (DNN)

6. Multi-task Model Training and Evaluation

Train and evaluate multi-task models using:

• 6_multitask/model_run.ipynb

This focuses on classifying compounds into Non-Resistant, Resistant, and MDR categories.

📊 Models and Benchmarks

We benchmark:

• Random Forest (Morgan fingerprints)
• SVR (KPGT embeddings)
• DNN (KPGT embeddings, best performance)

Task performance:

• MIC regression: R² up to 0.65, MAE ≈ 0.41
• Resistance classification (multi-task): Non-Resistant / Resistant / MDR F1 scores of 0.84, 0.81, and 0.67, respectively. Classification threshold: 10 µM

📬 Contact

For questions or feedback, contact Nuno Alves via GitHub or email: [email protected].

🔒 License

This project is licensed under the MIT License.