Functions for Shiny app development

Usage

htShiny() can export heatmaps as a stand-alone Shiny app. InteractiveComplexHeatmap also provides two functions for integrating the interactive heatmap widgets into other Shiny apps. The two functions are:

• InteractiveComplexHeatmapOutput(): for building UI on the client side.
• makeInteractiveComplexHeatmap(): for processing on the sever side.

With one interactive heatmap widget

The usage is simple. Following is an example that you can directly copy and paste to your R session.

library(ComplexHeatmap)
library(InteractiveComplexHeatmap)
library(shiny)

data(rand_mat) # simply a random matrix
ht1 = Heatmap(rand_mat, name = "mat",
show_row_names = FALSE, show_column_names = FALSE)
ht1 = draw(ht1)

ui = fluidPage(
h3("My first interactive ComplexHeatmap Shiny app"),
p("This is an interactive heatmap visualization on a random matrix."),
InteractiveComplexHeatmapOutput()
)

server = function(input, output, session) {
makeInteractiveComplexHeatmap(input, output, session, ht1)
}

shinyApp(ui, server)

With multiple interactive heatmap widgets

You can also put multiple interactive heatmaps widgets in a single Shiny app, but this time you must assign a “heatmap ID” for each one, so that makeInteractiveComplexHeatmap() can find the correct heatmap to respond. The heatmap ID should start with letters.

mat2 = matrix(sample(letters[1:10], 100, replace = TRUE), 10)
ht2 = draw(Heatmap(mat2, name = "mat2"))

ui = fluidPage(
h3("The first heatmap"),
InteractiveComplexHeatmapOutput("heatmap_1"),
hr(),
h3("The second heatmap"),
InteractiveComplexHeatmapOutput("heatmap_2")
)
server = function(input, output, session) {
makeInteractiveComplexHeatmap(input, output, session, ht1, "heatmap_1")
makeInteractiveComplexHeatmap(input, output, session, ht2, "heatmap_2")
}
shinyApp(ui, server)

Similarly, multiple interactive heatmap widgets can be arranged by a list of tabs:

ui = tabsetPanel(
tabPanel("Numeric", InteractiveComplexHeatmapOutput("heatmap_1")),
tabPanel("character", InteractiveComplexHeatmapOutput("heatmap_2"))
)
server = function(input, output, session) {
makeInteractiveComplexHeatmap(input, output, session, ht1, "heatmap_1")
makeInteractiveComplexHeatmap(input, output, session, ht2, "heatmap_2")
}
shinyApp(ui, server)

Customize the widgets

There are three main components in the interactive heatmap UI, i.e., the orignal heatmap, the sub-heatmap and an output that shows information of the clicked cell or the selected sub-heatmap. The original heatmap and sub-heatmap components can be resized by dragging the two boxes, but still, InteractiveComplexHeatmapOutput() provides arguments of width1, width2, height1 and height2 to control the initial sizes of the two components. They can be manually set to make sure the heatmap is well aligned, e.g. in htShinyExample(2.2).

The initial style of the brush can be specified by brush_opt argument. The value should be a list and the value will be sent to shiny::brushOpts(). Note, the style of the brush can also be manually adjusted in the Shiny app.

The layout

The layout of the three components are controlled by argument layout. It supports following values:

• "(1-2)|3": Original heatmap and sub-heatmap are in the same row, and output is in a second row. This is the default layout.
• "1|(2-3)": Original heatmap is in a single row, while sub-heatmap and output are in a second row.
• "1-2-3": All three components are in the same row.
• "1|2|3": Each component is in a single row.
• "1-(2|3)": Being different from the other four layouts, this is a two-column layout. Original heatmap is in a single column. Sub-heatmap and output are vertically aligned and the two are in the second column. An example can be found at htShinyExample(4.1).

Note the values for layout are in a special format to help to understand the layout, where the three code 1, 2 and 3 correspond to original heatmap, sub-heatmap and output respectively, symbol "-" corresponds to horizontal alignment and "|" corresponds to vertical alignment. With different layouts, different default values are assigned to widths and heights of the three components to make sure they are well aligned.

Action on single heatmap cells

By default, to get the information of a single cell in the heatmap, a "click" action is used. In InteractiveComplexHeatmapOutput(), the action can also be set to "hover" or "dblclick", then hovering or double clicking will trigger the response on the sever side. The example in htShinyExample(1.9) demonstrates usages of these three actions.

Which action to respond

The argument response can be set as a vector with values in "click", "hover", "dblclick", "brush" and "brush-output" to only respond to one or multiple events on heatmap. E.g. if response is only set to "click", there will be no response for the “brush event” in the interactive heatmap, also the sub-heatmap component is removed from the app. Please go to Section “Only respond to one event” for examples.

A brush on heatmap by default triggers two responses, one in the sub-heatmap and one in the output. If "brush-output" is included in response instead of "brush", you can still brush on the heatmap, but there is only response in the output, and the sub-heatmap component is removed from the app.

Separately specify the three UI components

InteractiveComplexHeatmapOutput() contains all three UI components. The three components can be separately specified as three individual functions: originalHeatmapOutput(), subHeatmapOutput() and HeatmapInfoOutput(). This provides flexibility for the UI arrangement, e.g. to integrate with package shinydashboard where each UI component is wrapped within an individual box.

body = dashboardBody(
fluidRow(
box(
title = "Original heatmap", width = 4, solidHeader = TRUE, status = "primary",
originalHeatmapOutput("ht", title = NULL)
),
box(
title = "Sub-heatmap", width = 4, solidHeader = TRUE, status = "primary",
subHeatmapOutput("ht", title = NULL)
),
box(
title = "Output", width = 4, solidHeader = TRUE, status = "primary",
HeatmapInfoOutput("ht", title = NULL)
)
)
)

ui = dashboardPage(
dashboardSidebar(),
body
)

server = function(input, output, session) {
makeInteractiveComplexHeatmap(input, output, session, ht, "ht")
}

shinyApp(ui, server)

Please note, since now the three components are generated independently, to correctly connect the three components as well as the server side, the heatmap ID must be explicitely specified in all functions. More examples on integrating with shinydashboard can be found from htShinyExample(10.1) to htShinyExample(10.5). In vignette “A Shiny app for visualizing DESeq2 results”, we demonstrate a complex app where the three UI components are specified separately to work with shinydashboard.

One thing that needs to be noted when integrating with shinydashboard is that the width (via width argument in box()) is actually measured as the relative fraction to parent div block, thus it is suggested to set the minimal width for parent block by adding a self-defined CSS code:

body = dashboardBody(
fluidRow(
box(
title = "Original heatmap", width = 4, solidHeader = TRUE, status = "primary",
originalHeatmapOutput("ht", title = NULL)
),
box(
title = "Sub-heatmap", width = 4, solidHeader = TRUE, status = "primary",
subHeatmapOutput("ht", title = NULL)
),
box(
title = "Output", width = 4, solidHeader = TRUE, status = "primary",
HeatmapInfoOutput("ht", title = NULL)
),
tags$style(" .content-wrapper, .right-side { overflow-x: auto; } .content { min-width:1500px; } ") ) ) Work with R Markdown documents It is very straightforward to integrate InteractiveComplexHeatmap in an interactive R Markdown document, just in the same way of integrating normal Shiny widgets. Following is an example and you can run a real interactive document with heatmaps by htShinyExample(7.1). --- title: "InteractiveComplexHeatmap in an Rmarkdown document" author: "Zuguang Gu" date: "16/12/2020" output: html_document runtime: shiny --- {r, echo = FALSE} library(InteractiveComplexHeatmap) m = matrix(rnorm(100*100), 100) ht = Heatmap(m)  {r, echo = FALSE} ui = fluidPage( InteractiveComplexHeatmapOutput() ) server = function(input, output, session) { makeInteractiveComplexHeatmap(input, output, session, ht) } shiny::shinyApp(ui, server)  More simply, you can directly use htShiny() in the chunk: --- title: "InteractiveComplexHeatmap in an Rmarkdown document" author: "Zuguang Gu" date: "16/12/2020" output: html_document runtime: shiny --- {r, echo = FALSE} library(InteractiveComplexHeatmap) m = matrix(rnorm(100*100), 100) ht = Heatmap(m)  {r, echo = FALSE} htShiny(ht)  Self-define the output Both the click and brush actions on the heatmap trigger an output below the heatmaps. The output gives the information of which row(s) and columns(s) are selected by users. The reponse for the two actions can be self-defined. In makeInteractiveComplexHeatmap(), there are two arguments click_action and brush_action which accept self-defined functions and define how to respond after the heatmap is clicked or brushed. The input for the two functions should accept two arguments, one is a DataFrame object which contains the information of which row(s) and columns(s) selected by users, and the second argument should always be output which is used in the Shiny app. click_action and brush_action can also be functions with four arguments which also includes input and session, in a form of function(df, input, output, session) {...}. To use click_action or brush_action, a htmlOutput (or other similar *Output) should be first set up in the UI, then the Shiny application knows where to update the output. The output UI can replace the default output by directly assigning to argument output_ui in InteractiveComplexHeatmapOutput(). ui = fluidPage( InteractiveComplexHeatmapOutput(output_ui = htmlOutput("info")) ) Or to create a new output UI independent to the interactive heatmap widget: ui = fluidPage( InteractiveComplexHeatmapOutput(), htmlOutput("info") ) The click_action or brush_action is basically defined as follows (assume the ID set in htmlOutput() is "info"): function(df, output) { output[["info"]] = renderUI({ # or output$info = ...
if(is.null(df)) { # have not clicked or brushed into the heatmap body
...
} else {
...
}
})
}

If users didn’t click or brush inside the heatmap body (e.g. clicked in the dendrograms), df that is passed to the functions will be NULL. Users might need to perform a sanity check here and print specific output when the heatmap was not selected.

The format of df is slightly different between click and brush. If it is a click action, df has the same format as the returned object of selectPosition() function, which looks like follows. It always has one row.

## DataFrame with 1 row and 6 columns
##       heatmap                  slice row_slice column_slice row_index
##   <character>            <character> <numeric>    <numeric> <integer>
## 1       mat_a mat_a_heatmap_body_1_2         1            2         9
##   column_index
##      <integer>
## 1            1

If it is a brush action, df has the same format as the returned object of selectArea() function, which looks like in the following chunk. Each line contains row and column indices of the selected sub-matrix in a specific heatmap slice of a specific heatmap.

## DataFrame with 4 rows and 6 columns
##       heatmap                  slice row_slice column_slice     row_index
##   <character>            <character> <numeric>    <numeric> <IntegerList>
## 1       mat_a mat_a_heatmap_body_1_2         1            2     7,5,2,...
## 2       mat_a mat_a_heatmap_body_2_2         2            2           6,3
## 3       mat_b mat_b_heatmap_body_1_1         1            1     7,5,2,...
## 4       mat_b mat_b_heatmap_body_2_1         2            1           6,3
##    column_index
##   <IntegerList>
## 1     2,4,1,...
## 2     2,4,1,...
## 3     1,2,3,...
## 4     1,2,3,...

Note as demonstrated above, the values in column row_index and column_index might be duplicated due to that the selected heatmap slices are in a same row slice or column slice, e.g., in previous example, the first and the third rows correspond to selection in the first row slice, but in the two column slices respectively, so they have the same value for row_index. thus, to safely get the row indices and column indices of the selected heatmap, users might need to perform:

unique(unlist(df$row_index)) unique(unlist(df$column_index))

Note again, if users want to use the values in input or session, click_action and brush_action can also be specified as functions with four arguments:

function(df, input, output, session) {
output[["info"]] = renderUI({  # or output$info = ... if(is.null(df)) { # have not clicked into the heatmap body ... } else { ... } }) } If action in InteractiveComplexHeatmapOutput() is set to "hover" or "dblclick", the corresponding argument for action is hover_action or dblclick_action. The usage is exactly the same as click_action. Examples of self-defining output In this section, I will demonstrate several examples of implementing self-defined output. In the first example, I replace the default ui with a new htmlOutput("info"). On the sever side, I define a click_action to print a styled text and a brush_action to print the table of the selected rows and columns from the heatmap. This following example can be run by htShinyExample(5.2). library(GetoptLong) # for the qq() function which does variable intepolation data(rand_mat) ht = Heatmap(rand_mat, show_row_names = FALSE, show_column_names = FALSE) ht = draw(ht) ui = fluidPage( InteractiveComplexHeatmapOutput(output_ui = htmlOutput("info")), ) click_action = function(df, output) { output[["info"]] = renderUI({ if(!is.null(df)) { HTML(qq("<p style='background-color:#FF8080;color:white;padding:5px;'>You have clicked on heatmap @{df$heatmap}, row @{df$row_index}, column @{df$column_index}</p>"))
}
})
}

suppressPackageStartupMessages(library(kableExtra))
brush_action = function(df, output) {
row_index = unique(unlist(df$row_index)) column_index = unique(unlist(df$column_index))
output[["info"]] = renderUI({
if(!is.null(df)) {
HTML(kable_styling(kbl(m[row_index, column_index, drop = FALSE], digits = 2, format = "html"), full_width = FALSE, position = "left"))
}
})
}

server = function(input, output, session) {
makeInteractiveComplexHeatmap(input, output, session, ht,
click_action = click_action, brush_action = brush_action)
}

shinyApp(ui, server)

The second example gives another scenario where the output needs to be self-defined. In this example, an gene expression matrix is visualized and clicking on the heatmap will print the corresponding gene and some other annotations related to this gene (e.g. the corresponding gene symbol, RefSeq IDs and UniProt IDs). Run htShinyExample(5.3) to see how this is implemented.

htShinyExample(5.4) gives an example where the heatmap visualizes correlations of a list of Gene Ontology terms (The plot is generated by the simplifyEnrichment package). In this example, the click and brush actions are self-defined so that the selected GO IDs as well as their detailed descriptions are printed.

htShinyExample(5.5) visualizes an correlation heatmap where clicking on the cell generates a scatter plot of the two corresponding variables. In this example, I set response = "click" in InteractiveComplexHeatmapOutput(), so that the sub-heatmap is removed from the app and the scatterplot (the output) is directly placed on the right of the original correlation heatmap.

htShinyExample(5.6) visualizes an a heatmap of pairwise Jaccard coefficients for multiple lists of genomic regions. Clicking on the heatmap cell draws a Hilbert curve (draw by the HilbertCurve package) which shows how the two corresponding sets of genomic regions overlap.

Instead of occupying static space, the output component can be floated to the mouse positions by setting output_ui_float = TRUE in InteractiveComplexHeatmapOutput() so that clicking, hovering or brushing from the heatmap opens a frame that contains the output. There are two examples: htShinyExample(9.1) and htShinyExample(9.2). The demonstration is as follows:

The self-defined output can also be floated if the self-defined UI replaces the default UI by setting InteractiveComplexHeatmapOutput(..., output_ui = new_output_ui):

Compact mode

In InteractiveComplexHeatmapOutput(), argument compact can be set to TRUE, so there is only the original heatmap and the output is floating at the mouse positions if hovering/clicking on heatmap. The calling

InteractiveComplexHeatmapOutput(..., compact = TRUE)

is actually identical to

InteractiveComplexHeatmap(..., response = c(action, "brush-output"), output_ui_float = TRUE)

Self-defined output can still be used here, e.g.

new_output_ui = ...
InteractiveComplexHeatmap(..., compact = TRUE, output_ui = new_output_ui)

See examples with htShinyExample(1.11).

Dynamically generate interactive heatmap widget

In previous examples, the heatmaps are already generated before making the interactive app. There are also scenarios where the heatmaps are generated on the fly, e.g. when the matrix is dynamically generated in the middle of an analysis. There might be following scenarios:

• The heatmap is based on a subset of matrix which is filtered by users, e.g., the expression matrix for differentially expressed genes filtered by different cutoffs.
• The annotations are dynamically provided by users.
• The heatmap parameters are changed by users, e.g., the clustering method or the splitting variable.
• If there are multiple heatmaps, which heatmaps are going to be drawn is dynamically selected.

In InteractiveComplexHeatmap, there are three ways to dynamically generate the interactive heatmap widgets which I will explain one by one.

Directly use makeInteractiveComplexHeatmap()

I first demonstrate use of makeInteractiveComplexHeatmap(). In the following example, the matrix is reordered by a user-selected column:

ui = fluidPage(
sliderInput("column", label = "Which column to order?",
value = 1, min = 1, max = 10),
InteractiveComplexHeatmapOutput()
)

server = function(input, output, session) {
m = matrix(rnorm(100), 10)
rownames(m) = 1:10
colnames(m) = 1:10

observeEvent(input$column, { order = order(m[, input$column])
ht = Heatmap(m[order, , drop = FALSE],
cluster_rows = FALSE, cluster_columns = FALSE)
makeInteractiveComplexHeatmap(input, output, session, ht)
})
}
shiny::shinyApp(ui, server)

A similar but slightly complex example is as follows. It can be run by htShinyExample(6.2).

The use is very natural. makeInteractiveComplexHeatmap() is put inside an observeEvent() or an observe() so that every time input$column changes, it triggers an update of the interactive heatmap widgets. In the following code block defined in server function:  ... observeEvent(input$column, {
order = order(m[, input$column]) ht = Heatmap(m[order, , drop = FALSE], cluster_rows = FALSE, cluster_columns = FALSE) makeInteractiveComplexHeatmap(input, output, session, ht) }) ... makeInteractiveComplexHeatmap() internally creates a list of responses by observeEvent(). Every time when input$column triggers the update of makeInteractiveComplexHeatmap(), all the calls of observeEvent() will be re-executed. Re-executing observeEvent() only adds the observations to the current observation list while not overwrites them, thus, repeatedly executing makeInteractiveComplexHeatmap() will make a same observatin running multiple times. To solve this issue, makeInteractiveComplexHeatmap() saves all the observations returned by observeEvent() and it tries to first destroies all the avaiable observations that have been created. However, if user-defined reponses via click_action and brush_action use observe() or observeEvent(), they must manually recorded so that they can also be destroied when updating makeInteractiveComplexHeatmap(). See the following example:

click_action = function(df, input, output, session) {
obs = observeEvent(input$foo, { ... }) record_observation(obs) } Use InteractiveComplexHeatmapModal() and InteractiveComplexHeatmapWidget() In the first example, the interactive heatmap is already generated when the Shiny app is loaded. There is a second scenario where the complete interactive heatmap widget is dynamically generated and inserted into the HTML document. There are two other functions InteractiveComplexHeatmapModal() and InteractiveComplexHeatmapWidget() which have very similar behaviors. These two functions are normally put inside e.g. shiny::observeEvent() or shiny::observe() and they generate UI as well as render the interactive heatmaps. First I will introduce the usage of InteractiveComplexHeatmapModal(). In the following example, there is only an action button in the UI, and in the server function, InteractiveComplexHeatmapModal() is called when receiving an input$show_heatmap signal. This example can also be run by htShinyExample(6.3).

ui = fluidPage(
actionButton("show_heatmap", "Generate_heatmap"),
)

server = function(input, output, session) {
m = matrix(rnorm(100), 10)
ht = Heatmap(m)

observeEvent(input$show_heatmap, { InteractiveComplexHeatmapModal(input, output, session, ht) }) } shiny::shinyApp(ui, server) As shown in the following figure, InteractiveComplexHeatmapModal() will open an “modal frame” which includes the interactive heatmap. In the next example which is also available in htShinyExample(6.4), a different heatmap is generated according to user’s selection. ui = fluidPage( radioButtons("select", "Select", c("Numeric" = 1, "Character" = 2)), actionButton("show_heatmap", "Generate_heatmap"), ) get_heatmap_fun = function(i) { mat_list = list( matrix(rnorm(100), 10), matrix(sample(letters[1:10], 100, replace = TRUE), 10) ) Heatmap(mat_list[[i]]) } server = function(input, output, session) { observeEvent(input$show_heatmap, {
i = as.numeric(input$select) InteractiveComplexHeatmapModal(input, output, session, get_heatmap = get_heatmap_fun(i)) }) } shiny::shinyApp(ui, server) The usage of InteractiveComplexHeatmapWidget() is very similar as InteractiveComplexHeatmapModal(), except that now for InteractiveComplexHeatmapWidget(), user needs to allocate a place defined by shiny::htmlOutput() in UI, and later the interactive heatmap widget is put there. I modify the previous example with InteractiveComplexHeatmapWidget(). Now in the UI, I add one line where I specify htmlOutput() with ID "heatmap_output", and this ID is set in InteractiveComplexHeatmapWidget() correspondingly. ui = fluidPage( actionButton("show_heatmap", "Generate_heatmap"), htmlOutput("heatmap_output") ) server = function(input, output, session) { m = matrix(rnorm(100), 10) ht = Heatmap(m) observeEvent(input$show_heatmap, {
InteractiveComplexHeatmapWidget(input, output, session, ht, output_id = "heatmap_output")
})
}
shiny::shinyApp(ui, server)

The app looks like follows:

InteractiveComplexHeatmapModal() and InteractiveComplexHeatmapWidget() all accept an argument js_code where customized JavaScript code can be inserted after the interactive UI. This is sometimes useful. In previous example where the heatmap widget is triggered by clicking on the action button, every time when clicking on the button, the widget is regenerated although the heatmaps are actually the same. Actually we can change the behavior of the button that from the second click it just switches the visibility of the heatmap widget. See more examples in htShinyExample(6.5) and htShinyExample(6.7).

ui = fluidPage(
actionButton("show_heatmap", "Generate_heatmap"),
htmlOutput("heatmap_output")
)

server = function(input, output, session) {
m = matrix(rnorm(100), 10)
ht = Heatmap(m)

observeEvent(input$show_heatmap1, { InteractiveComplexHeatmapWidget(input, output, session, ht, output_id = "heatmap_output", close_button = FALSE, js_code = "$('#show_heatmap').click(function() {
$('#heatmap_output').toggle('slow'); }).text('Show/hide heatmap'). attr('id', 'show_heatmap_toggle'); " ) }) } shiny::shinyApp(ui, server) Implement interactivity from scratch InteractiveComplexHeatmap provides rich tools for interactively working with heatmaps. However, some people might want to develop their own tools while they only need the information of which cells are selected. Next I demonstrate it with a simple example. The following example is runnable. ui = fluidPage( actionButton("action", "Generate heatmap"), plotOutput("heatmap", width = 500, height = 500, click = "heatmap_click", brush = "heatmap_brush"), verbatimTextOutput("output") ) server = function(input, output, session) { ht_obj = reactiveVal(NULL) ht_pos_obj = reactiveVal(NULL) observeEvent(input$action, {
m = matrix(rnorm(100), 10)
rownames(m) = 1:10
colnames(m) = 1:10

output$heatmap = renderPlot({ ht = draw(Heatmap(m)) ht_pos = htPositionsOnDevice(ht) ht_obj(ht) ht_pos_obj(ht_pos) }) }) observeEvent(input$heatmap_click, {
pos = getPositionFromClick(input$heatmap_click) selection = selectPosition(ht_obj(), pos, mark = FALSE, ht_pos = ht_pos_obj(), verbose = FALSE) output$output = renderPrint({
print(selection)
})
})

observeEvent(input$heatmap_brush, { lt = getPositionFromBrush(input$heatmap_brush)

selection = selectArea(ht_obj(), lt[[1]], lt[[2]], mark = FALSE, ht_pos = ht_pos_obj(),
verbose = FALSE)
output\$output = renderPrint({
print(selection)
})
})
}
shinyApp(ui, server)

In this simple Shiny app, a click or a brush on heatmap prints the corresponding data frame that contains the information of the selected cells.

There are several points that need to be noticed:

1. draw() and htPositionsOnDevice() need to be put inside the renderPlot().
2. To get the position of the click on heatmap, getPositionFromClick() should be used. With knowing the position of the click, it can be sent to selectPosition() to correspond to the original matrix.
3. Similarly, to get the positions of the area that was brushed on heatmap, getPositionFromBrush() should be used.

This method also works for complex heatmaps, e.g. with row or column splitting, or with multiple heatmaps.