A quantitative framework for evaluating data structure preservation by dimensionality reduction techniques

High-dimensional data is integral to modern systems biology. Computational methods for dimensionality reduction are being rapidly developed for applications to single-cell and multi-modal technologies. In order to understand what these nonlinear transformations do to the underlying biological patterns in our data, we developed a framework of quantitative metrics for global and local distance preservation. See our Cell Reports publication for details and analysis of existing dimensionality reduction methods.

---

Contents

fcc_utils.py

Contain utility functions for manipulating datasets and comparing feature-reduced latent spaces.

Documentation available at Ken-Lau-Lab.github.io/DR-structure-preservation/.

Tutorials

Consult distance_preservation_tutorial.ipynb for info on how to perform global and local structure preservation analyses on low-dimensional embeddings of your own datasets.

---

Required Python Dependencies

Install using pip:

pip install -r requirements.txt

---

Optional Packages

In order to use the "FIt-SNE" implementation of t-SNE, you'll need to download FFTW and compile the code from the FIt-SNE repo.

Clone the scvis and ZIFA packages and install with python setup.py install from their home directories.