Deep Latent Variable Path Modelling (DLVPM)

Deep Latent Variable Path Modelling (DLVPM) is a framework for path / structural equation modelling using deep neural networks. The method links heterogeneous datasets through sets of orthogonal deep latent variables (DLVs), enabling structured multimodal learning.

Full documentation: https://deep-lvpm.readthedocs.io/en/latest/

Published in Deep Latent Variable Path Modelling in Nature Machine Intelligence.

If you find this project useful, consider starring the repository on GitHub.

The animation above shows model training for a three-factor DLVPM model linking omics and imaging data from lung cancer patients. This dataset is included with the package.

---

Installation

uv venv  # create environment
uv pip install .

---

Tutorials

Three runnable tutorials are included:

Integrate five TCGA lung cancer modalities

uv run -m tutorial.run_tcga

Associate MNIST images with digit labels

uv run -m tutorial.run_mnist

Demonstrate a Siamese encoder on CIFAR-10.

uv run -m tutorial.run_siamese

---

Testing

Run the test suite with:

uv run -m tests.run_tests