In RaikOtto/ArtDeco: Alikeness Based on Regression on Transciptome DECOnvolution

artdeco R package

A R package that determines the differentiation stage of pancreatic neuroendocrinal neoplasia (PANnen) based on bulk RNA-seq or on mRNA array exepriments of cancer-tissue.

1 How to make it work: Quickstart

Installing artdeco

Start an R session e.g. using RStudio

if (!requireNamespace("BiocManager", quietly=TRUE)) install.packages("BiocManager")

BiocManager::install("GEOquery")

Test run

97 representative NEN samples are utilized for this vignette from the GSE733338 GSE package. In the following, these will be analyzed by the artdeco algorithm to demonstrate the utilization of artdeco.

library("artdeco")

data("visualization_data")
transcriptome_data = visualization_data

# testrun
deconvolution_results = Deconvolve_transcriptome(
    transcriptome_data = transcriptome_data,
    deconvolution_algorithm = "nmf"
)

# show results
deconvolution_results[1:2,]

The call returns a dataframe that contains the differentiation stage similarity predictions.

Explanation test run results

Percent-wise similarities of the testdata samples to all scRNA derived training samples are shown, both as written interpretation e.g. "significant similarity (to this cell type)" or "no significant similarity (to this cell type)" as well as percent scalars.

Explanation percent similarities

The percent value quantifies to what extend a given transcriptome can be reconstructed utilizing a given reference cell type e.g. by the insulin producing 'beta' cell type as basis. The nu-SVM regression based similarity determination has the limitation that the returned unit-less scalar is challenging to interpret: It is not clearly decidable whether a scalar indicates a high or low similarity. Thus, artdeco interprets the returned scalar. Measurements may be calculated based on a trained model (absolute) or relative to all analyzed within a given run (relative). The default 'absolute' mode interprets the returned scalar for a given sample relative to the maximal similarity measured for each model subtype during training. I.e. the returned scalar is divided by the maximal similarity observed for training set samples. E.g. when a scRNA beta cell received the similarity 5.7 to the beta set, than all subsequent beta similarities are divided by 5.7 to show how similar they are relative to the self-identification of cell subtypes. The 'relative' mode divides by the maximal similarity measured for a given set of query data, i.e. the maximally measured similarity is taken as divisor.

2 Visualization of similarity predictions

It is important to interpret the simialrity predictions given the clustering of the data. In order to detect clusters, a correlation heatmap can be used along with an overlay of similarity predictions. The rational is, that samples that show a comparable differentiation state should cluster together, which is why a heatmap creator is included in the artdeco package and can be utilized as follows. First the analyzed transcriptome data is loaded and afterwards the similarity predictions from step one are being passed on to the visualization function.

library(devtools)
install_github("vqv/ggbiplot")

create_heatmap_deconvolution(
    deconvolution_results = deconvolution_results,
    visualization_data = transcriptome_data,
    utilize_sadanandam_genes = TRUE
)

create_PCA_deconvolution(
    visualization_data = transcriptome_data,
    deconvolution_results = deconvolution_results,
    utilize_sadanandam_genes = TRUE
)

3 Adding and removing models

Adding a new model requires 1) training data and 2) a labeling vector of the training data specifying the cell type of each training sample.

data(Lawlor) # Data from Lawlor et al.
Lawlor[1:5,1:5]
data(meta_data)

# extract the training sample subtype labels
subtype_vector = as.character(meta_data$Subtype) 
subtype_vector[1:6] # show subtype definition

add_deconvolution_training_model_NMF(
    transcriptome_data = Lawlor,
    model_name = "my_model",
    subtype_vector = subtype_vector,
    rank_estimate = 0,
    exclude_non_interpretable_NMF_components = FALSE,
    training_nr_marker_genes = 100,
    parallel_processes = 1,
    nrun = 1
)

Note that the training data in the current version (1.0.0) of artdeco has to have HGNC symbols as row names.

Removing models is fairly straight forward; list the models after specifying the Machine-Learning algorihm i.e. 'NMF' or 'bseqsc' etc..

show_models(lib_name = "NMF")

model_to_be_removed = "my_model"
remove_model_NMF(
    model_name = model_to_be_removed,
    test_mode = TRUE
)

4 Important: data format

Please note that the expression data always has to have HGNC symbols as row names and that the column names contain the samples.

data("Lawlor") # data from Lawlor et al.
Lawlor[1:5,1:5]

Contact: [e-mail]

RaikOtto/ArtDeco documentation built on Oct. 30, 2021, 6:20 p.m.