structure_plot: Structure Plot

Description

Create a “Structure plot” from a multinomial topic model fit or other model with “loadings” or “weights”. The Structure plot represents the estimated topic proportions of each sample in a stacked bar chart, with bars of different colors representing different topics. Consequently, samples that have similar topic proportions have similar amounts of each color.

Usage

structure_plot(
  fit,
  topics,
  grouping,
  loadings_order = "embed",
  n = 2000,
  colors,
  gap = 1,
  embed_method = structure_plot_default_embed_method,
  ggplot_call = structure_plot_ggplot_call,
  ...
)
structure_plot_default_embed_method(fit, ...)
# S3 method for poisson_nmf_fit
plot(x, ...)
# S3 method for multinom_topic_model_fit
plot(x, ...)
structure_plot_ggplot_call(dat, colors, ticks = NULL, font.size = 9)

Value

A ggplot object.

Arguments

fit: An object of class “poisson_nmf_fit” or “multinom_topic_model_fit”, or an n x k matrix of topic proportions, where k is the number of topics. (The elements in each row of this matrix should sum to 1.) If a Poisson NMF fit is provided as input, the corresponding multinomial topic model fit is automatically recovered using poisson2multinom.
topics: Top-to-bottom ordering of the topics in the Structure plot; topics[1] is shown on the top, topics[2] is shown next, and so on. If the ordering of the topics is not specified, the topics are automatically ordered so that the topics with the greatest total “mass” are at shown at the bottom of the plot. The topics may be specified by number or by name. Note that not all of the topics need to be included, so one may also use this argument to plot a subset of the topics.
grouping: Optional categorical variable (a factor) with one entry for each row of the loadings matrix fit$L defining a grouping of the samples (rows). The samples (rows) are arranged along the horizontal axis according to this grouping, then within each group according to loadings_order. If grouping is not a factor, an attempt is made to convert it to a factor using as.factor. Note that if loadings_order is specified manually, grouping should be the groups for the rows of fit$L before reordering.
loadings_order: Ordering of the rows of the loadings matrix fit$L along the horizontal axis the Structure plot (after they have been grouped). If loadings_order = "embed", the ordering is generated automatically from a 1-d embedding, separately for each group. The rows may be specified by number or by name. Note that loadings_order may include all the rows of fit$L, or a subset.
n: The maximum number of samples (rows of the loadings matrix fit$L) to include in the plot. Typically there is little to no benefit in including large number of samples in the Structure plot due to screen resolution limits. Ignored if loadings_order is provided.
colors: Colors used to draw topics in Structure plot.
gap: The horizontal spacing between groups. Ignored if grouping is not provided.
embed_method: The function used to compute an 1-d embedding from a loadings matrix fit$L; only used if loadings_order = "embed". The function must accept the multinomial topic model fit as its first input (“fit”) and additional arguments may be passed (...). The output should be a named numeric vector with one entry per row of fit$L, and the names of the entries should be the same as the row names of fit$L.
ggplot_call: The function used to create the plot. Replace structure_plot_ggplot_call with your own function to customize the appearance of the plot.
...: Additional arguments passed to structure_plot (for the plot method) or embed_method (for function structure_plot).
x: An object of class “poisson_nmf_fit” or “multinom_topic_model_fit”. If a Poisson NMF fit is provided as input, the corresponding multinomial topic model fit is automatically recovered using poisson2multinom.
dat: A data frame passed as input to ggplot, containing, at a minimum, columns “sample”, “topic” and “prop”: the “sample” column contains the positions of the samples (rows of the L matrix) along the horizontal axis; the “topic” column is a topic (a column of L); and the “prop” column is the topic proportion for the respective sample.
ticks: The placement of the group labels along the horizontal axis, and their names. For data that are not grouped, use ticks = NULL.
font.size: Font size used in plot.

Details

The name “Structure plot” comes from its widespread use in population genetics to visualize the results of the Structure software (Rosenberg et al, 2002).

For most uses of the Structure plot in population genetics, there is usually some grouping of the samples (e.g., assignment to pre-defined populations) that guides arrangement of the samples along the horizontal axis in the bar chart. In other applications, such as analysis of gene expression data, a pre-defined grouping may not always be available. Therefore, a “smart” arrangement of the samples is, by default, generated automatically by performing a 1-d embedding of the samples.

References

Dey, K. K., Hsiao, C. J. and Stephens, M. (2017). Visualizing the structure of RNA-seq expression data using grade of membership models. PLoS Genetics 13, e1006599.

Rosenberg, N. A., Pritchard, J. K., Weber, J. L., Cann, H. M., Kidd, K. K., Zhivotovsky, L. A. and Feldman, M. W. (2002). Genetic structure of human populations. Science 298, 2381–2385.

Examples

Run this code

# \donttest{
set.seed(1)
data(pbmc_facs)

# Get the multinomial topic model fitted to the
# PBMC data.
fit <- pbmc_facs$fit

# Create a Structure plot without labels. The samples (rows of L) are
# automatically arranged along the x-axis using t-SNE to highlight the
# structure in the data.
p1a <- structure_plot(fit)

# The first argument to structure_plot may also be an "L" matrix.
# This call to structure_plot should produce the exact same plot as
# the previous call.
set.seed(1)
p1b <- structure_plot(fit$L)

# There is no requirement than the rows of L sum up to 1. To
# illustrate, in this next example we have removed topic 5 from the a
# structure plot.
p2a <- structure_plot(fit$L[,-5])

# This is perhaps a more elegant way to remove topic 5 from the
# structure plot:
p2b <- structure_plot(fit,topics = c(1:4,6))

# Create a Structure plot with the FACS cell-type labels. Within each
# group (cell-type), the cells (rows of L) are automatically arranged
# using t-SNE.
subpop <- pbmc_facs$samples$subpop
p3 <- structure_plot(fit,grouping = subpop)

# Next, we apply some customizations to improve the plot: (1) use the
# "topics" argument to specify the order in which the topic
# proportions are stacked on top of each other; (2) use the "gap"
# argrument to increase the whitespace between the groups; (3) use "n"
# to decrease the number of rows of L included in the Structure plot;
# and (4) use "colors" to change the colors used to draw the topic
# proportions.
topic_colors <- c("skyblue","forestgreen","darkmagenta",
                  "dodgerblue","gold","darkorange")
p4 <- structure_plot(fit,grouping = pbmc_facs$samples$subpop,gap = 20,
                     n = 1500,topics = c(5,6,1,4,2,3),colors = topic_colors)

# In this example, we use UMAP instead of t-SNE to arrange the
# cells in the Structure plot. Note that this can be accomplished in
# a different way by overriding the default setting of
# "embed_method".
y <- drop(umap_from_topics(fit,dims = 1))
p5 <- structure_plot(fit,loadings_order = order(y),grouping = subpop,
                     gap = 40,colors = topic_colors)

# We can also use PCA to arrange the cells.
y <- drop(pca_from_topics(fit,dims = 1))
p6 <- structure_plot(fit,loadings_order = order(y),grouping = subpop,
                     gap = 40,colors = topic_colors)

# In this final example, we plot a random subset of 400 cells, and
# arrange the cells randomly along the horizontal axis of the
# Structure plot.
p7 <- structure_plot(fit,loadings_order = sample(3744,400),gap = 10,
                     grouping = subpop,colors = topic_colors)
# }

Run the code above in your browser using DataLab