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 preprint 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.

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

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

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

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

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

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

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

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

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

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

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.

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)

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)