latent 0.1

Latent spaces for single cells

LatentLego - flexible latent spaces for single cells

LatentLego is a python library for building autoencoders for single-cell genomics. It is build on top of TensorFlow and Keras and provides modular components to build common single-cell autoencoder architectures. All individual comonents are valid Keras Layer or Model objects and therefore support model construction using the standard Keras (functional/sequential) API. There is also a model zoo of preconstructed models that understand builtin Keras methods like .fit() or .predict() and .compile() with custom loss functions.

Installation

You can install LatentLego from PyPI using pip

pip install latent

or install the development version from GitHub

pip install git+https://github.com/quadbiolab/latent-lego

Overview

LatentLego is designed to be a flexible kit for defining autoencoders by providing common building blocks that can be freely assembled. Here's an overview:

• latent.models - Here you can find Keras-compatible implementations of common single-cell autoencoder architectures, such as the ZINBAutoencoder (Erslan 2019) or the NegativeBinomialVAE (Lopez 2018). I also implemented some more experimental architectures like the TopologicalAutoencoder (Moor 2019).

• latent.modules - Here live the lower level building blocks of autoencoders, namely encoder and decoder models that can be freely assembled to produce autoencoder models. E.g., using a VariationalEncoder with a PoissonDecoder produces a PoissonVAE, a variational autoencoder with a poisson reconstruction loss.

• latent.layers - These are the most basic building blocks of encoder and decoder models. In addition to the DenseBlock/DenseStack layers that are the core component of encoders and decoders, I implemeneted some other useful utility layers.

• latent.losses, latent.activations, latent.callbacks - Keras-style components that are useful for single-cell data, like a NegativeBinomial loss or a KLDivergenceScheduler callback function. All of these should work nicely within the TensorFlow/Keras framework.