PatHoST

Pathogen-Host Spatial Transcriptomics

PatHoST is a Python framework for spatial transcriptomics data analysis, designed to study pathogen-host interactions using machine learning and deep learning techniques by imputing the presence of bacteria on the tissues

📋 Description

PatHoST provides tools for:

• Ground Truth Projection: Alignment and projection of ground truth data onto training datasets
• Spatial Smoothing: Algorithms for data smoothing based on spatial proximity and PCA similarity
• Bacterial Classification: Machine learning and deep learning models for classifying pathogen presence/absence in spatial data
• Performance Evaluation: Clustering metrics and model stability assessment

🏗️ Model Architecture

The library includes various neural network architectures:

Model	Description
Encoder / Encoder_dropout	Feed-forward encoder for latent space compression
Decoder / Decoder_dropout	Decoder for data reconstruction
Autoencoder / Autoencoder_dropout	Standard autoencoder with/without dropout
Autoencoder_Bacterial_Predictor	Autoencoder with classification head for bacterial prediction
Encoder_Bacterial_Predictor_dropout	Encoder with integrated predictor

📁 Project Structure

PatHoST/
├── src/
│   ├── __init__.py
│   ├── models.py        # Neural network architecture definitions
│   ├── train.py         # Training workflow for classification
│   ├── predict.py       # Prediction functions
│   ├── datasets.py      # Dataset and data loading utilities
│   ├── spatial.py       # Spatial analysis functions
│   ├── metrics.py       # Evaluation metrics and clustering
│   └── utils.py         # General utilities
├── notebooks/
│   ├── Data_prep.ipynb  # Data preparation
│   ├── Training.ipynb   # Model training
│   └── Evaluation.ipynb # Results evaluation
└── requirements.txt     # Dependencies

🚀 Installation

Prerequisites

• Python 3.13+
• CUDA (optional, for GPU acceleration)

Setup with Conda

# Create a new conda environment
conda create --name pathost python=3.13
conda activate pathost

# Install dependencies
pip install git+[https://github.com/Allen13x/PatHoST.git](https://github.com/Allen13x/PatHoST.git)

Setup with pip

# Create a virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or
venv\Scripts\activate     # Windows

# Install main dependencies
pip install git+[https://github.com/Allen13x/PatHoST.git](https://github.com/Allen13x/PatHoST.git)

📖 Usage

Spatial Smoothing

from PatHoST.spatial import Smooth

# Apply Mean-based smoothing
smoothed_data, distance_matrix = Smooth(
    spdata=spatial_data,
    key='sample',
    data='dataY',
    X='dataX',
    mode='Mean',
    Dist=200
)

# Or apply Local PCA-based smoothing
smoothed_data, D = Smooth(
    spdata=spatial_data,
    key='sample',
    data='dataY',
    X='dataX',
    mode='Local',
    Dist=200
)

Training the Bacterial Classifier

from PatHoST.train import BacterialClassifierWorkflow

# Define parameters
train_keys = ['sample_1', 'sample_2']
Xdata = ['dataX']
Ydata = ['bacterial_signal']
models = ['Logistic', 'RandomForest', 'XGBoost', 'Autoencoder_dropout']

# Run training workflow
results = BacterialClassifierWorkflow(
    spdata_i=spatial_data,
    train_keys=train_keys,
    Xdata=Xdata,
    Ydata=Ydata,
    spots='overlapping_spots',
    signal_cutoff=[0.5],
    models=models,
    embed_size=128,
    batch_size=32,
    n_epochsAE=1000,
    n_epochs=100
)

Prediction

from PatHoST.predict import BacterialClassifierPrediction
from PatHoST.models import Autoencoder_Bacterial_Predictor_dropout

# Load model
n_genes = spatial_data['sample']['dataX'].shape[1]  # Number of genes/features in expression matrix
model = Autoencoder_Bacterial_Predictor_dropout(n_input=n_genes, embed_size=128, h_dim=128)
model.load_state_dict(trained_weights)

# Run prediction
predictions = BacterialClassifierPrediction(
    spdata=spatial_data,
    key='test_sample',
    x='dataX',
    model=model,
    type='Autoencoder_dropout'
)

Evaluation with Clustering Metrics

from PatHoST.metrics import evaluate_all_methods

# Evaluate all clustering methods
results_df, labels_df = evaluate_all_methods(
    pred_scores=prediction_scores,
    true_labels=ground_truth,  # optional
    fcmin=0.3,
    fcmax=0.7
)

print(results_df)

📊 Supported Models

Classical Machine Learning

• Logistic Regression: Linear classification
• Random Forest: Decision tree ensemble
• XGBoost: Optimized gradient boosting

Deep Learning

• Autoencoder with Dropout: Latent representation learning with regularization
• Encoder with Predictor: Direct classification from latent space

🔧 Main Dependencies

Package	Usage
torch	Deep learning framework
scanpy	Single-cell analysis
anndata	Transcriptomics data structure
scikit-learn	Classical machine learning
xgboost	Gradient boosting
numpy, pandas	Data manipulation
matplotlib, seaborn	Visualization

📓 Notebooks

The notebooks/ folder contains interactive tutorials:

1. Data_prep.ipynb: Spatial data preparation and preprocessing
2. Training.ipynb: Model training for classification
3. Evaluation.ipynb: Performance evaluation and results visualization

📄 License

This project is distributed under the MIT License. See the LICENSE file for more details.

📧 Contact

For questions or suggestions, please open an issue on GitHub.

---

PatHoST - Advanced pathogen-host interaction analysis through spatial transcriptomics