sci-VAE

Sci-VAE is an implementation of a variational autoencoder in Keras that was developed to use VAEs to integrate biological data. The implementation allows for customisations to the VAE to be passed in via CLI (and a JSON file) or in python and R scripts (see examples).

The VAE implementation expects a data matrix with features as columns (no headers) and rows as training data (no row IDs). The first thing the VAE will do is transform your data between 0 and 1 so you don’t need to do this prior to running the VAE.

I show several examples, using MNIST, IRIS dataset and then also a publicly available histone modification and RNAseq dataset from encode (for those bioinformaticians out there).

There are also some useful visualisations that I was having to repeat often when inspecting the latent space so check out the Vis functions if you’re interested (these are also in the examples).

Lastly, there is a optimisation library that allows you to optimise the VAE architecture based on building a separable latent space based on classification. If you choose to use this you’ll also need to pass in labels into the VAE. Check out some tests for how to run this - it uses an evolutionary algorithm.