In mamouzgar/hsslda_dep: LDA with Hybrid Subset Selection.

LDA with Hybrid Subset Selection (HSS)

Authors: Meelad Amouzgar and David Glass

Bendall lab @ Stanford University

Not permitted for distribution outside of the Bendall lab @ Stanford

knitr::opts_chunk$set(echo = TRUE)
knitr::opts_knit$set(root.dir = getwd())

Introduction:

Linear Discriminant Analysis (LDA) is a classification algorithm that we've repurposed for supervised dimensionality reduction of single-cell data. LDA identifies linear combinations of predictors that optimally separate a priori labels, enabling users to tailor visualizations to separate specific aspects of cellular heterogeneity. We implement LDA with feature selection by Hybrid Subset Selection (HSS) in this R package, called hsslda.

Installation instructons:

You can install hsslda using:

remotes::install_github("mamouzgar/hsslda", build_vignettes = FALSE)

The github page includes an introduction to the package. Or if you'd like the introductory vignette seen below included in your Rstudio installation of hsslda, you can use:

remotes::install_github("mamouzgar/hsslda", build_vignettes = TRUE)

library(magrittr)

library(hsslda)

Here we read-in the example data

# load("/Users/mamouzgar/phd-projects/Rpackage/hsslda/data/TcellHartmann2020_sampleData.rda")

data(TcellHartmann2020_sampleData)

colnames(TcellHartmann2020_sampleData)
head(TcellHartmann2020_sampleData,3)

How to run HSS-LDA:

You can run hybrid subset selection on a dataset using runHSS().

runHSS takes 3 inputs:

1) x: a dataframe of cells (rows) and markers/genes (columns)

2) y: a vector of your class labels of interest.

3) score.method: A scoring metric to use for feature selection, which can include: 'euclidean', 'silhouette', 'pixel.entropy', 'pixel.density', or 'custom'

4) (optional) downsample: Boolean, defaults to TRUE. Downsamples data to improve runtime. Set to FALSE to use all input data.

Here we will run HSS using the default scoring method: euclidean distance.Note that scoring metrics like silhouette or pixel class entropy (PCE) score will require additional package dependencies.

channels = c('GLUT1', 'HK2', 'GAPDH', 'LDHA', 'MCT1', 'PFKFB4', 'IDH2', 'CyclinB1', 'GLUD12', 'CS', 'OGDH', 'CytC', 'ATP5A', 'S6_p', 'HIF1A')
train.x = TcellHartmann2020_sampleData[channels]
train.y = TcellHartmann2020_sampleData[['labels']]
hss.result = runHSS(x = train.x, y = train.y, score.method = 'euclidean', downsample = FALSE)

Output summary:

The final hss.result object outputted from runHSS contains a few elements, the most important ones are highlighted below:

1) HSSscores: a table of all scores for each subsetted model.

2) ElbowPlot: visualizes the elbow plot and calculated elbow point for the optimal feature set.

3) HSS-LDA-model: The final LDA model using the optimal feature set.

4) HSS-LDA-result: The final result table containing all initially inputted markers, class labels, and linear discriminants from the opimal model.

5) downsampled.cells.included.in.analysis: A boolean vector of rows included in the downsampling analysis. If downsampling was not performed, all values are TRUE. The final output results includes all data projected onto the HSS-LD axes.

You can visualize the elbow plot, which is ggplot configurable:

hss.result$ElbowPlot

The final output object contains a merged dataframe of your markers, class labels, and newly generated LD axes

lda.df = hss.result$`HSS-LDA-result`
head(lda.df, 3)

viridis.colors = c("#440154FF", "#414487FF", "#2A788EFF", "#22A884FF", "#7AD151FF", "#FDE725FF")
names(viridis.colors) = unique(lda.df$labels)
myColScale = ggplot2::scale_color_manual(values = viridis.colors)

ggplot2::ggplot(lda.df,ggplot2::aes(x = HSS_LD1, y = HSS_LD2, color = labels)) +
  ggplot2::geom_point() +
  myColScale

The final HSS-LDA model is also saved, and can be used like any R model.

hss.result$`HSS-LDA-model`

newdata = train.x

You can use this final model to apply the same dimensionality reduction to new data using makeAxes.

lda.df.newData = makeAxes(df = newdata, co =hss.result$`HSS-LDA-model`$scaling)

You can also add your own custom separation metric to perform feature selection with by changing score.method = 'custom' and adding custom function. The custom function must take in as input:

1) x: dataframe of all your data.

2) y: vector of class labels matching training data rows.

3) custom.score.method: the function for your custom separation metric.

You can then input your custom function into the hsslda::runHSS() function as an argument into custom.score.method, and indicate score.method = 'custom'

hss.resultCustom = runHSS(x = train.x, y = train.y, score.method = 'custom', custom.score.method = myCustomFunction)

And that's how you use LDA with Hybrid Subset Selection!

Questions, comments, or general feedback:

[e-mail]

mamouzgar/hsslda_dep documentation built on Dec. 21, 2021, 1:46 p.m.