pySCENIC is a lightning-fast python implementation of the SCENIC pipeline (Single-Cell rEgulatory Network Inference and Clustering) which enables biologists to infer transcription factors, gene regulatory networks and cell types from single-cell RNA-seq data.
The pioneering work was done in R and results were published in Nature Methods [1].
pySCENIC can be run on a single desktop machine but easily scales to multi-core clusters to analyze thousands of cells in no time. The latter is achieved via the dask framework for distributed computing [2].
The pipeline has three steps:
- First transcription factors (TFs) and their target genes, together defining a regulon, are derived using gene inference methods which solely rely on correlations between expression of genes across cells. The arboreto package is used for this step.
- These regulons are refined by pruning targets that do not have an enrichment for a corresponding motif of the TF effectively separating direct from indirect targets based on the presence of cis-regulatory footprints.
- Finally, the original cells are differentiated and clustered on the activity of these discovered regulons.
Note
The most impactfull speed improvement is introduced by the arboreto package in step 1. This package provides an alternative to GENIE3 [3] called GRNBoost2. This package can be controlled from within pySCENIC.
Quick Start
- Installation
- Tutorial
- Command Line Interface
- Frequently Asked Questions
- See notebooks
- Report an issue
- Releases at PyPI
All the functionality of the original R implementation is available and in addition:
- You can leverage multi-core and multi-node clusters using dask and its distributed scheduler.
- We implemented a version of the recovery of input genes that takes into account weights associated with these genes.
- Regulons, i.e. the regulatory network that connects a TF with its target genes, with targets that are repressed are now also derived and used for cell enrichment analysis.
The lastest stable release of the package itself can be installed via pip install pyscenic
.
Caution!
pySCENIC needs a python 3.6 or greater interpreter.
You can also install the bleeding edge (i.e. less stable) version of the package directly from the source:
git clone https://github.com/aertslab/pySCENIC.git
cd pySCENIC/
pip install .
To successfully use this pipeline you also need auxilliary datasets:
- Databases ranking the whole genome of your species of interest based on regulatory features (i.e. transcription factors). Ranking databases are typically stored in the feather format and can be downloaded from cisTargetDBs.
- Motif annotation database providing the missing link between an enriched motif and the transcription factor that binds this motif. This pipeline needs a TSV text file where every line represents a particular annotation.
Annotations | Species |
---|---|
HGNC annotations | Homo sapiens |
MGI annotations | Mus musculus |
Flybase annotations | Drosophila melanogaster |
Caution!
These ranking databases are 1.1 Gb each so downloading them might take a while. An annotations file is typically 100Mb in size.
For this tutorial 3,005 single cell transcriptomes taken from the mouse brain (somatosensory cortex and hippocampal regions) are used as an example [4]. The analysis is done in a Jupyter notebook.
Caution!
If you run this from a python script instead of a Jupyter notebook, please enclose the code in
a if __name__ == '__main__':
construct.
First we import the necessary modules and declare some constants:
import os
import glob
import pickle
import pandas as pd
import numpy as np
from dask.diagnostics import ProgressBar
from arboreto.utils import load_tf_names
from arboreto.algo import grnboost2
from pyscenic.rnkdb import FeatherRankingDatabase as RankingDatabase
from pyscenic.utils import modules_from_adjacencies, load_motifs
from pyscenic.prune import prune2df, df2regulons
from pyscenic.aucell import aucell
import seaborn as sns
DATA_FOLDER="~/tmp"
RESOURCES_FOLDER="~/resources"
DATABASE_FOLDER = "~/databases/"
SCHEDULER="123.122.8.24:8786"
DATABASES_GLOB = os.path.join(DATABASE_FOLDER, "mm9-*.feather")
MOTIF_ANNOTATIONS_FNAME = os.path.join(RESOURCES_FOLDER, "motifs-v9-nr.mgi-m0.001-o0.0.tbl")
MM_TFS_FNAME = os.path.join(RESOURCES_FOLDER, 'mm_tfs.txt')
SC_EXP_FNAME = os.path.join(RESOURCES_FOLDER, "GSE60361_C1-3005-Expression.txt")
REGULONS_FNAME = os.path.join(DATA_FOLDER, "regulons.p")
MOTIFS_FNAME = os.path.join(DATA_FOLDER, "motifs.csv")
The scRNA-Seq data is downloaded from GEO: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE60361 and loaded into memory:
ex_matrix = pd.read_csv(SC_EXP_FNAME, sep='\t', header=0, index_col=0).T
ex_matrix.shape
(3005, 19970)
and the list of Transcription Factors (TF) for Mus musculus are read from file. The list of known TFs for Mm was prepared from TFCat (cf. notebooks section).
tf_names = load_tf_names(MM_TFS_FNAME)
Finally the ranking databases are loaded:
db_fnames = glob.glob(DATABASES_GLOB)
def name(fname):
return os.path.basename(fname).split(".")[0]
dbs = [RankingDatabase(fname=fname, name=name(fname)) for fname in db_fnames]
dbs
[FeatherRankingDatabase(name="mm9-tss-centered-10kb-10species"), FeatherRankingDatabase(name="mm9-500bp-upstream-7species"), FeatherRankingDatabase(name="mm9-500bp-upstream-10species"), FeatherRankingDatabase(name="mm9-tss-centered-5kb-10species"), FeatherRankingDatabase(name="mm9-tss-centered-10kb-7species"), FeatherRankingDatabase(name="mm9-tss-centered-5kb-7species")]
In the initial phase of the pySCENIC pipeline the single cell expression profiles are used to infer co-expression modules from.
Run GENIE3 or GRNBoost from arboreto to infer co-expression modules
The arboreto package is used for this phase of the pipeline. For this notebook only a sample of 1,000 cells is used for the co-expression module inference is used.
adjacencies = grnboost2(ex_matrix, tf_names=tf_names, verbose=True)
Regulons are derived from adjacencies based on three methods.
The first method to create the TF-modules is to select the best targets for each transcription factor:
- Targets with importance > the 50th percentile.
- Targets with importance > the 75th percentile
- Targets with importance > the 90th percentile.
The second method is to select the top targets for a given TF:
- Top 50 targets (targets with highest weight)
The alternative way to create the TF-modules is to select the best regulators for each gene (this is actually how GENIE3 internally works). Then, these targets can be assigned back to each TF to form the TF-modules. In this way we will create three more gene-sets:
- Targets for which the TF is within its top 5 regulators
- Targets for which the TF is within its top 10 regulators
- Targets for which the TF is within its top 50 regulators
A distinction is made between modules which contain targets that are being activated and genes that are being repressed. Relationship between TF and its target, i.e. activator or repressor, is derived using the original expression profiles. The Pearson product-moment correlation coefficient is used to derive this information.
In addition, the transcription factor is added to the module and modules that have less than 20 genes are removed.
modules = list(modules_from_adjacencies(adjacencies, ex_matrix))
# Calculate a list of enriched motifs and the corresponding target genes for all modules.
with ProgressBar():
df = prune2df(dbs, modules, MOTIF_ANNOTATIONS_FNAME)
# Create regulons from this table of enriched motifs.
regulons = df2regulons(df)
# Save the enriched motifs and the discovered regulons to disk.
df.to_csv(MOTIFS_FNAME)
with open(REGULONS_FNAME, "wb") as f:
pickle.dump(regulons, f)
Clusters can be leveraged in the following way:
# The clusters can be leveraged via the dask framework:
df = prune2df(dbs, modules, MOTIF_ANNOTATIONS_FNAME, client_or_address=SCHEDULER)
Caution!
The nodes of the clusters need to have access to a shared network drive on which the ranking databases are stored.
Reloading the enriched motifs and regulons from file should be done as follows:
df = load_motifs(MOTIFS_FNAME)
with open(REGULONS_FNAME, "rb") as f:
regulons = pickle.load(f)
We characterize the different cells in a single-cell transcriptomics experiment via the enrichment of the previously discovered regulons. Enrichment of a regulon is measured as the Area Under the recovery Curve (AUC) of the genes that define this regulon.
auc_mtx = aucell(ex_matrix, regulons, num_workers=4)
sns.clustermap(auc_mtx, figsize=(8,8))
A command line version of the tool is included. This tool is available after proper installation of the package via pip
.
{ ~ } » pyscenic ~
usage: pySCENIC [-h] {grnboost,ctx,aucell} ...
Single-CEll regulatory Network Inference and Clustering
positional arguments:
{grnboost,ctx,aucell}
sub-command help
grnboost Derive co-expression modules from expression matrix.
ctx Find enriched motifs for a gene signature and
optionally prune targets from this signature based on
cis-regulatory cues.
aucell Find enrichment of regulons across single cells.
optional arguments:
-h, --help show this help message and exit
Arguments can be read from file using a @args.txt construct.
Yes you can. The code snippet below shows you how to create your own databases:
from pyscenic.rnkdb import DataFrameRankingDatabase as RankingDatabase
import numpy as np
import pandas as pd
# Every model in a database is represented by a whole genome ranking. The rankings of the genes must be 0-based.
df = pd.DataFrame(
data=[[0, 1],
[1, 0]],
index=['Model1', 'Model2'],
columns=['Symbol1', 'Symbol2'],
dtype=np.int32)
RankingDatabase(df, 'custom').save('custom.db')
For more information, please visit LCB and SCENIC.
GNU General Public License v3
We are grateful to all providers of TF-annotated position weight matrices, in particular Martha Bulyk (UNIPROBE), Wyeth Wasserman and Albin Sandelin (JASPAR), BioBase (TRANSFAC), Scot Wolfe and Michael Brodsky (FlyFactorSurvey) and Timothy Hughes (cisBP).
[1] | Aibar, S. et al. SCENIC: single-cell regulatory network inference and clustering. Nat Meth 14, 1083–1086 (2017). |
[2] | Rocklin, M. Dask: parallel computation with blocked algorithms and task scheduling. conference.scipy.org |
[3] | Huynh-Thu, V. A. et al. Inferring regulatory networks from expression data using tree-based methods. PLoS ONE 5, (2010). |
[4] | Zeisel, A. et al. Cell types in the mouse cortex and hippocampus revealed by single-cell RNA-seq. Science 347, 1138–1142 (2015). |