SISUA

Semi-supervised Single-cell modeling:

• Free software: MIT license
• Documentation: https://github.com/trungnt13/sisua/tree/master/docs.

Reference:

• Trung Ngo Trong, Roger Kramer, Juha Mehtonen, Gerardo González, Ville Hautamäki, Merja Heinäniemi. "SISUA: SemI-SUpervised Generative Autoencoder for Single Cell Data", ICML Workshop on Computational Biology, 2019. [pdf]

Installation

You only need Python 3.6, the stable version of SISUA installed via pip:

pip install sisua

Install the nightly version on github:

pip install git+https://github.com/trungnt13/sisua@master

For developers, we create a conda environment for SISUA contribution sisua_env

conda env create -f=sisua_env.yml

Getting started

1. The basics:

   • Datasets description
   • Models specification
   • Basic API and work-flow

2. Single-cell analysis:

   • Latent space
   • Imputation of genes expression
   • Prediction of protein markers

3. Advanced technical topics:

   • Probabilistic embedding
   • Hierarchical modeling (coming soon)
   • Causal analysis (coming soon)
   • Cross datasets analysis (coming soon)

4. Benchmarks:

   • Scalability test
   • Fine-tuning networks
   • Data normalization

Roadmap

1. [x] Multi-OMICs single-cell dataset (link)
2. [x] Disentanglement VAE for multi-OMICs data (link)
3. [x] New model: FactorVAE, BetaVAE, MIxture Semi-supervised Autoencoder (MISA) (link)
4. [ ] Better imputation via hierarchical latents model.
5. [ ] Release SISUA 2

Toolkits

We provide binary toolkits for fast and efficient analyzing single-cell datasets:

• sisua-train: train single-cell modeling algorithms, support training multiple systems in parallel.
• sisua-analyze: evaluate, compare, and interpret trained model.
• sisua-embed: probabilistic embedding for semi-supervised training.
• sisua-data: coming soon

Some important arguments:

-model	name of function declared in models scvi: single-cell Variational Inference model dca: Deep Count Autoencoder vae: single-cell Variational Autoencoder movae: SISUA
-ds	name of dataset declared in data. Description of all predefined datasets is in docs. Some good datasets for practicing: pbmc8k_ly cortex pbmcecc_ly pbmcscvi pbmcscvae

Configuration

By default, the data will be saved at your home folder at ~/bio_data, and the experiments' outputs will be stored at ~/bio_log

You can customize these two paths using the environment variables:

• For storing downloaded and preprocessed data: SISUA_DATA
• For the experiments: SISUA_EXP

For example:

import os
os.environ['SISUA_DATA'] = '/tmp/bio_data'
os.environ['SISUA_EXP'] = '/tmp/bio_log'

from sisua.data import EXP_DIR, DATA_DIR

print(DATA_DIR) # /tmp/bio_data
print(EXP_DIR)  # /tmp/bio_log

or you could set the variables in advance:

export SISUA_DATA=/tmp/bio_data
export SISUA_EXP=/tmp/bio_log
python sisua/train.py
# or using the provided toolkit: sisua-train