scyan 1.2.0

Single-cell Cytometry Annotation Network

Scyan stands for Single-cell Cytometry Annotation Network. Based on biological knowledge prior, it provides a fast cell population annotation without requiring any training label. Scyan is an interpretable model that also corrects batch-effect and can be used for debarcoding, cell sampling, and population discovery.

Documentation

The complete documentation can be found here. It contains installation guidelines, tutorials, a description of the API, etc.

Overview

Scyan is a Bayesian probabilistic model composed of a deep invertible neural network called a normalizing flow (the function $f_{\phi}$). It maps a latent distribution of cell expressions into the empirical distribution of cell expressions. This cell distribution is a mixture of gaussian-like distributions representing the sum of a cell-specific and a population-specific term. Also, interpretability and batch effect correction are based on the model latent space — more details in the article's Methods section.

Getting started

Scyan can be installed on every OS with pip or poetry.

On macOS / Linux, python>=3.8,<3.11 is required, while python>=3.8,<3.10 is required on Windows. The preferred Python version is 3.9.

Install from PyPI (recommended)

pip install scyan

Install locally (if you want to contribute)

Advice (optional): We advise creating a new environment via a package manager (except if you use Poetry, which will automatically create the environment). For instance, you can create a new conda environment:

conda create --name scyan python=3.9
conda activate scyan

Clone the repository and move to its root:

git clone https://github.com/MICS-Lab/scyan.git
cd scyan

Choose one of the following, depending on your needs (it should take at most a few minutes):

pip install .                           # pip minimal installation (library only)
pip install -e '.[dev,docs,discovery]'  # pip installation in editable mode
poetry install -E 'dev docs discovery'  # poetry installation in editable mode

Basic usage / Demo

import scyan

adata, marker_pop_matrix = scyan.data.load("aml")

model = scyan.Scyan(adata, marker_pop_matrix)
model.fit()
model.predict()

This code should run in approximately 40 seconds (once the dataset is loaded). For more usage demo, read the tutorials or the complete documentation.

Technical description

Scyan is a Python library based on:

• Pytorch, a deep learning framework
• AnnData, a data library that works nicely with single-cell data
• Pytorch Lighning, for model training
• Hydra, for project configuration (optional)
• Weight & Biases, for model monitoring (optional)

Project layout

.github/      # Github CI and templates
config/       # Hydra configuration folder (optional use)
data/         # Data folder containing adata files and csv tables
docs/         # The folder used to build the documentation
scripts/      # Scripts to reproduce the results from the article
tests/        # Folder containing tests
scyan/                    # Library source code
    data/                 # Folder with data-related functions and classes
        datasets.py       # Load and save datasets
        tensors.py        # Pytorch data-related classes for training
    module/               # Folder containing neural network modules
        coupling_layer.py # Coupling layer
        distribution.py   # Prior distribution (called U in the article)
        mmd.py            # Maximum Mean Discrepancy implementation
        real_nvp.py       # Normalizing Flow
        scyan_module      # Core module
    plot/                 # Plotting tools
        ...
    tools/
        ...               # Preprocessing tools and more
    model.py              # Scyan model class
    utils.py              # Misc functions
.gitattributes
.gitignore
CONTRIBUTING.md   # To read before contributing
LICENSE
mkdocs.yml        # The docs configuration file
poetry.lock
pyproject.toml    # Dependencies, project metadata, and more
README.md
setup.py          # Setup file, see `pyproject.toml`

Cite us

Our paper is not published yet. Meanwhile, you can read our preprint on arXiv.