DrivAER for manifold interpretation in scRNA-seq data

DrivAER is a method for identification of Driving transcriptional programs based on AutoEncoder derived Relevance scores. DrivAER infers relevance scores for transcriptional programs with respect to specified outcomes of interest in single-cell RNA sequencing data, such as psuedotemporal ordering or disease status.

See our manuscript for more details.

Workflow (a) DrivAER iteratively subjects annotated gene sets to unsupervised dimension reduction via DCA. (b) For each gene set the generated two-dimensional data manifold coordinates are used as (c) input features in a random forest model to predict the outcome of interest (i.e. pseudotemporal ordering). (d) The random forest prediction accuracy represents the relevance score.

Reproducibility

To reproduce the analysis and figures presented in our manuscript please see the Reproducibility folder.

Check out our live, interactive tutorial!

The following Google colab notebooks allow you to interactively explore DrivAER and can be run within your browser. We have prepared two analysis examples:

1. Blood development
2. Interferon stimulation

Installation

via pip

pip install git+https://github.com/lkmklsmn/DrivAER.git

via git

git clone https://github.com/lkmklsmn/DrivAER
cd DrivAER
python setup.py install

Input

1. Raw count expression matrix
2. Outcome of interest (pseudotemporal ordering/cell grouping etc)
3. Gene set annotation

Output

1. Relevance scores for each annotated transcriptional program
2. Data manifolds derived from each transcriptional program
3. Various visualizations (heatmap, DCA embedding, barplots)

Usage

Step 1: Get Gene Set Annotations

DrivAER supports annotations in gmt and csv format, as well as user-defined annotation file.

1.1 From gmt format

The gmt format files for Broad's MSigDB can be downloaded from the Broad Website.

Gene set	Source	Gene1	Gene2	Gene3
set1	source	gene1	gene2	gene3
set2	source	gene1	gene2	gene3
set3	source	gene1	gene2	gene3

import DrivAER as dv
C3_mouse = dv.get_anno(filename = "C3.gmt", filetype = "gmt", conv_mouse = True)

1.2 From tsv format

The tsv format files can be downloaded from Trandcription Factor sites, such as TRRUST. DrivAER provides built-in annotations from TRRUST.

Transcription factor	Target	Type	Source
set1	gene1	XX	XX
set1	gene2	XX	XX
set1	gene3	XX	XX
set2	gene1	XX	XX

trrust_human = dv.get_anno(filename = "trrust_human.tsv", filetype = "tsv", conv_mouse = False)

1.3 From user-defined gene set annotations

Users can create your own gene set annotations. The format is a pandas series. Index are trandcription factor names or gene set names. Each row contains a list of corresponding genes.

Step 2: Calculate relevance scores

res = dv.calc_relevance(count = your_count, pheno = your_pt, tf_targets = C3_mouse, min_targets=5,
               ae_type="nb-conddisp", epochs=100, early_stop=3, hidden_size=(8, 2, 8), verbose=False)

Additionally, users can replace the DCA with other dimension reduction methods. The commands calc_relevance_pca, calc_relevance_tsne, calc_relevance_umap will perform dimension reduction based on PCA, tSNE and UMAP, respectively.

Step 3: Generate visualizations

dv.rank_plot(result, save, path)
dv.embedding_plot(result, tf_name, pheno, datatype, save, path)
dv.gene_plot(result, count, tf_name, gene, save, path)