Getting Started#
MIRA versus CODAL#
If you are looking to use the CODAL method for comparative single-cell atlas construction, you’re in the right place. MIRA’s topic modeling methods were originally developed for single-cell multiomic analysis, but work identically for non-multiome single-cell experiments. Now, MIRA’s topic models perform inference using the CODAL objective for disentangling biological effects from technical effects in batched data.
System Requirements#
Linux or MacOS
Python >=3.5
(optional) CUDA-enabled GPU
Data#
Single-cell Multiome
MIRA takes scRNA-seq and scATAC-seq count matrices from a single-cell multiomics experiment or experiments (batched), where each cell is measured using both assays, and measurements are linked by a shared cell barcode. We demonstrated MIRA using SHARE-seq data and commercial 10X genomics multiome data, but MIRA’s assumptions and models are extensible to other multiome protocols.
scRNA-seq or scATAC-seq only
When working with non-multiomic data, some of MIRA’s functionalities are limited. However, one can use MIRA’s topic models to analyze single-mode datasets. Again, MIRA needs a count matrix as input.
Notes on Preprocessing
Since MIRA starts from count matrices, one can use any method for read preprocessing and cell QC. Of note, we find that CellRanger’s ATAC-seq peak-calling method finds fewer and less-specific peaks than MACS2. This contributes to lower resolution manifolds during topic modeling.
Installation#
MIRA can be installed from PyPI:
pip3 install mira-multiome
Installation will take about a minute.
To set up an a new analysis, we recommend starting with a fresh environment:
conda create --name mira-env -c conda-forge -c bioconda scanpy jupyter leidenalg
conda activate mira-env
pip3 install mira-multiome
python -m ipykernel install --user --name mira-env
To use the environment in a jupyter notebook, start the notebook server, then go to Kernel > Change kernel > mira-env.
Installing with GPU support#
Training on a GPU reduces the training time of MIRA topic models. To install MIRA with PyTorch compiled with GPU support, first install MIRA, as above. Then, follow instructions at pytorch.org to find the version of PyTorch that suits your system.
Learning Curve#
If you have experience with Scanpy, we structured MIRA to follow similar conventions so that it would feel familiar and intuitive. In fact, most MIRA analyses seamlessly weave between MIRA and Scanpy functionalities for cleaning, slicing, and plotting the data. In general, the first positional argument of a MIRA function is an AnnData object, and the following keyword arguments change how the function transforms that object.