Here's the README file content based on the manuscript and reproducibility guidelines:

```markdown
# Machine Learning-Assisted Genomic Profiling of BCG vs. Wild-Type Mycobacterium bovis

## Description
This repository contains code for differentiating BCG vaccine strains from non-BCG wild-type Mycobacterium bovis using machine learning approaches. The pipeline implements both random forest and 1D CNN classifiers trained on genomic features extracted from 72 bacterial genomes.

## Dataset Information
### Source
- 72 complete genome assemblies from NCBI:
  - 28 BCG vaccine strains
  - 44 non-BCG wild-type M. bovis
- Accession numbers: See Supplementary Excel File 1

### Characteristics
- Assembly quality: ≤200 contigs/genome, N50 ≥50kb
- Validation: ANI ≥98% against reference genomes
- Exclusion criteria: Contamination <1%, assembly gaps <5%

## Code Information
### Main Scripts
1. `preprocessing_pipeline.sh` - deltaBS workflow for data preparation
2. `random_forest_classifier.R` - R script for RF model development
3. `cnn_classifier.py` - Python CNN implementation

### Key Features
- DeltaBS-based feature extraction
- Iterative feature selection with permutation testing
- CNN architecture with dropout regularization (0.3-0.5)

## Usage Instructions
### Step-by-Step Implementation
1. Clone repository:
   ```bash
   git clone https://gist.github.com/szypanther/ffa7e7a6d869020cc53eb809e4794f0d
   ```

2. Install dependencies (see Requirements section)

3. Run preprocessing:
   ```bash
   bash preprocessing_pipeline.sh 
   ```
4. Execute classifiers:
   ```bash
   Rscript random_forest_classifier.R
   python cnn_classifier.py
   ```

## Requirements
### Software
- Numpy version:  1.21.6
- Joblib version:  1.3.2
- R ≥4.1.2
- Python ≥3.8
- deltaBS
### Libraries
```plaintext
R: randomForest, caret
Python: TensorFlow 2.15, scikit-learn 1.2, pandas 1.5
Bioinformatics: Prokka 1.14.6, Roary 3.13.0, HMMER 3.3.2
```

## Methodology
### Workflow Overview
1. Genome annotation (Prokka)
2. Pan-genome analysis (Roary)
3. Bitscore calculation (DeltaBS)
4. Feature selection:
   - RF: OOB error minimization
   - CNN: Gradient activation mapping
5. Model training:
   - RF: 10,000 trees, mtry=n/10
   - CNN: Two Conv1D layers (32/64 filters)

## Citations
If using this work, please cite:
```bibtex
@article{shi2025ml_bcg,
  title={Machine Learning-Assisted genomic profiling to identify differences between BCG vaccine strains and non-BCG wild-type Mycobacterium bovis},
  author={Shi, Yunyun et al.},
  journal={PeerJ},
  year={2025}
}
```

## License
This code is made available under the MIT License. Contributions via pull requests are welcome.

## Contact
Zhiyong Shen (szypanther@gmail.com)
School of Basic Medical Sciences
Hunan University of Medicine
```