MarbliX

MarbliX is a framework designed to integrate diverse biomedical data modalities, such as histopathology images and genomic data, into compact binary representations called monograms. This framework is adaptable to other modalities, enabling scalable and interpretable multimodal search, classification, and patient similarity analysis for a variety of applications in biomedical research.

---

Project Structure

MarbliX/
├── data/
│   ├── embeddings/
│   │   ├── images/                  # Directory containing image embedding .npy files
│   │   └── sequences/              # Directory containing sequence embedding .npy files
│   └── metadata/
│       └── samples.csv             # CSV file listing sample metadata (file names + labels)
├── src/
│   └── marblix.py                  # Main script containing model training and feature extraction
├── requirements.txt
└── README.md

---

Prerequisites

Ensure Python 3.8+ is installed. Then install the required packages:

pip install -r requirements.txt

---

Data Format

Place your data under the data/ folder in the following format:

1. data/metadata/samples.csv

This CSV file should have at least the following columns:

file_name	label
sample_001	low-risk
sample_002	high-risk
...	...

Each file_name should correspond to a sample, and match the .npy files in the embeddings folders.

---

2. data/embeddings/images/

Contains NumPy .npy files for image embeddings. Each file should be named:

sample_001-image-features.npy
sample_002-image-features.npy
...

3. data/embeddings/sequences/

Contains NumPy .npy files for sequence embeddings. Each file should be named:

sample_001-seq-features.npy
sample_002-seq-features.npy
...

---

Running MarbliX

After placing your data and updating the paths in the code, you can run the training pipeline with:

python src/marblix.py

This will:

• Load and normalize image and sequence embeddings

• Train:

  • Image -> Sequence Hybrid Autoencoder
  • Sequence -> Image Hybrid Autoencoder
  • Multimodal Triplet Model

• Extract real-valued and binary multimodal features

• Optionally visualize binary feature matrices

---

Output

• Console output includes training metrics and model summaries.
• Final multimodal (features) and binary (binary_features) features are computed for test samples.
• Visualization of one binary matrix (8x8) is displayed using matplotlib.

You can extend the code to save outputs by adding:

np.save('multimodal_features.npy', multimodal_features)
np.save('multimodal_binary_features.npy', multimodal_binary_features)