In hdsu-bioquant/ButchR: Non-Negative Matrix Factorization Suite

ShinyButchR

ShinyButchR is an interactive R/Shiny app to perform NMF on an input matrix, ShinyButchR uses ButchR to run the matrix decomposition and generate diagnostic plots and visualizations that helps to find association of the NMF signatures to known biological factors.

How to use ShinyButchR

The app can be used directly from the official ShinyButchR website, a pre-build image can be pulled from Docker hdsu/shinybutchr:

docker run --rm -p 3838:3838 hdsu/shinybutchr

or also the Github repository can be pulled from hdsu-bioquant/ShinyButchR.

Loading screen

When you start the app the first displayed screen is a loading screen, while the app loads ButchR and TensorFlow (Figure 1).
Depending on the system this step could take from a few second to around one minute. If you are using ShinyButchR from the ShinyButchR website, and the app was idle for more than one hour; a new instance has to be created by shinyapps.io, this process can take around one to one and a half minutes.

Data and annotation screen

When the app is loaded, you are prompted to either run an analysis using the demo dataset, or to upload your own data (Figure 2). For this tutorial we are using the demo dataset which is a the publicly available RNA-seq dataset of sorted blood cell populations, comprising 12 cell populations and 45 samples (Corces et al., 2016).

The options available in the Data and annotation screen are (Figure 3): - a Upload a non-negative matrix in a csv or RDS file format using the file browser. You can also clear previously loaded datasets by clicking in the clear button. 2. Upload a table with annotation data in a csv or RDS file, the first column should match the column names in the input matrix. This file is optional and you can run the NMF based analysis without the annotation file. 3. NMF parameters: select the minimum and maximum factorization rank (minimum value allowed is 2 and maximum value allowed is 30); Select a factorization method, from "NMF" and "GRNMF_SC"; Number of initializations, number of times that the decomposition should be run with random initializations in ShinyButchR the limit in 10; Convergence threshold: in ShinyButchR the convergence of the NMF is reach when every sample (in the columns of the input matrix), is assigned to the same signature (on the basis of maximum exposure) after n iterations. 4. Start NMF: after uploading a non-negative matrix (and optionally a table with annotation data), the NMF-based analysis run by clicking the Submit button.

To load the demo dataset, please click on the demo button inside the Matrix upload box (Figure 4.1). After loading the demo dataset, the first five rows of the input matrix and annotation data are displayed in the Matrix upload and Annotation upload boxes.

The default parameters to run NMF for the demo are: minimum factorization rank: 2; maximum factorization rank: 10; factorization method: NMF; number of initializations: 2; convergence threshold: 30. Now the data is ready to be run the workflow, please click the Submit button in the Start NMF box (Figure 4.2).

Data exploration with ShinyButchR

Start NMF

The matrix decomposition starts after clicking the Submit button in the Start NMF box (Figure 5). The running time of the decomposition depends on the initial NMF parameters. For the demo default parameters, it takes around one minute.

NMF results

When the NMF decomposition finishes you are prompted to the NMF plots screen, where you can explore all the results:

Heatmat of the H matrix

One of the big advantages of the lower-dimensional representation obtained by NMF is the possibility of representing the resulting H matrix as a heatmap. In this representation the rows of the matrix are the Signatures identified by NMF and the columns represent the samples in the original matrix. The differential exposure of the samples to the signature can be seen in this type of plot, allowing to understand which samples are closely related (Figure 6).

In case you have more than one type of annotation for the samples in the input matrix, it is possible to display them all by clicking on the Select columns to use input box (Figure 7).

In ShinyButchR you can change the factorization rank in the Select K input box (Figure 8.1), to display a heatmap of the H matrix based in any of the factorization ranks included in the original range of ranks used to run the analysis.

UMAP based on the H matrix

To identify clusters on the samples, a UMAP is run using the H matrix as input. It is possible to select the factorization rank to change the H matrix used to run UMAP. And also, the coloring of the samples on the scatter plot is customizable depending on the uploaded annotation table (Figure 8.2).

Recovery plots based on the H matrix

The association of the recovered signatures to known biological and clinical variables is one of the most important steps in an NMF-based workflow. In ShinyButchR a recovery plot is made for all the associated categorical variables available in the annotation table. (Figure 9.1).

Determining the optimal factorization rank

Based on the results of the factorization quality metrics, the optimal number of signatures (k) is determined by minimizing the Frobenius error, the coefficient of variation and the mean Amari distance, while maximizing the sum and mean silhouette width and the cophenic coefficient (Figure 9.2).

Riverplot visualization

To visualize the stability and hierarchical relationships between signatures, a Riverplot or Sankey diagram is displayed showing the similarity between signatures across factorization ranks. The cutoff of displayed similarities can be changed with the slider, as well as the range of factorization ranks to use in the visualization (Figure 10).

Download results

ShinyButchR allows to download the results as and RDS object of class ButchR_NMF compatible with the latest version of ButchR (Figure 11.1). To use this object you can start a new R session and read it like this:

library(ButchR)
nmf_exp <- readRDS("~/path/to/NMF_object.RDS")

It is also possible to download csv and RDS files of the H and W matrices for a selected factorization rank (Figure 11.2).

hdsu-bioquant/ButchR documentation built on Jan. 28, 2023, 6:06 p.m.