Motivation

Motivation

Too many plots!

Enter, Shiny!

My motivation

(First step) Help collaborator: app to display one data set

• Explore, visualize, share data
• Generate hypotheses

(Now) Help everyone: general app for uploading/analyzing data sets

• Implement existing R tools to analyze data
• Explore, visualize, share data
• Generate hypotheses

Collaborator: "How can I share my data?"

https://kcvi.shinyapps.io/START_HT/

Developer/Statistician/Data Scientist:
"How can I get people to use my new method/package?"

http://qvalue.princeton.edu/

Developer/Statistician/Data Scientist:
Look at this analysis I did!

https://gallery.shinyapps.io/TSupplyDemand/

Everyone: Teach! (learnr)

https://minnier.shinyapps.io/ODSI_continuousData/

Shiny basics

# what you see, client side
ui <- fluidPage(
)

# computations here!
server <- function(input, output) {
}

# run the app!
shinyApp(ui = ui, server = server)

Shiny basics

# what you see, client side
ui <- fluidPage(
  # Title, panels
  plotOutput("myplot"), #from output$myplot
  selectInput("foo",choices=allthechoices)
)

# computations here!
server <- function(input, output) {
}

# run the app!
shinyApp(ui = ui, server = server)

Shiny basics

# what you see, client side
ui <- fluidPage(
  # Title, panels
  plotOutput("myplot"), #from output$myplot
  selectInput("foo",choices=allthechoices)
)

# computations here!
server <- function(input, output) {
  # reactive plots
  output$myplot <- renderPlot({ ggplot(yay) + geom_point() })
  # use input$foo here, maybe observe changes in UI
  observe({ dostuff(input$foo) })
}

# run the app!
shinyApp(ui = ui, server = server)

An example: START

Very quick tour

Very quick tour

Data distributions, interactive

Analysis results, interactive

Data exploration, interactive

Who is your audience?

Example: START app

• likely biologists
• may not be used to coding or large data sets

Challenges: User

Challenges, Input/Ouput Data

• Input data ~ multiple formats: Special input formats
• Interface with other software? Export special formats

Challenges, Input/Ouput Data

• Input data ~ multiple formats: Special input formats

https://biit.cs.ut.ee/clustvis/

Challenges, Input/Ouput Data

• Input data ~ multiple formats: Special input formats

https://biit.cs.ut.ee/clustvis/

Challenges, Analysis

Many possible ways to analyze data:

• RNA-seq data has many analysis pipelines (edgeR, DESeq2, limma)
• Complex study designs
  • repeated measures
  • multiple pairwise comparisons vs overall group comparisons
• How to allow for so many possibilities?

Challenges, Analysis

Many possible ways to analyze data:

• RNA-seq data has many analysis pipelines (edgeR, DESeq2, limma)
• Complex study designs
  • repeated measures
  • multiple pairwise comparisons vs overall group comparisons
• How to allow for so many possibilities?

Balance flexibility/simplicity/CORRECTness

• Too many options $\to$ high likelihood of incorrect use of analysis tools

Challenges, Analysis

Many possible ways to analyze data:

• Restrict analyses performed
• Allow for uploading of analysis results

Challenges, Analysis: QA/QC

Present QA plots up front

https://github.com/joey711/shiny-phyloseq

Challenges, Analysis: QA/QC

Does this data look "weird"?

Responsibilities

- Are resulting analyses correct?

- Are we encouraging "torturing the data until it speaks"?

- Misuse of p-values?

Responsibilities: Visual inference

Plots = statistics

-- Di Cook's keynote at rstudio::conf 2018

Solution?: Rorschach image tests

Generate null data sets and null plots using permutation and simulation

Can you pick out your data from a line up?

R package nullabor

Visual inference: Apophenia

http://visiphilia.org/2014/11/03/nullabor

Room for misuse, possible solutions:

Documentation

- Use instructions and data inputs wisely.

- Provide examples of when advanced analyses are needed.

- Provide examples of statistical write ups based on app's results.

Analysis

- Avoid too many hypothesis tests.

- Avoid p-values (Bayesian posterior probabilities?).

- With collaborators: stay involved!

Landing page - Where am I?

• Show what the app is for/can do
• Point to instructions and guidelines

Bonus: Jasmine Dumas has a shinyLP package to help!

Instructions, Please

• Extremely difficult.
• Not just a vignette of a package (users may not be coders)
• Written instructions vs. video demos, gif instructions

No one will use it if they don't know how to use it!

Instructions

Tauno Metsalu and Jaak Vilo: "Clustvis"

Instructions

Tauno Metsalu and Jaak Vilo: "Clustvis"

Debugging Tips

Debugging is notoriously a pain here

Thanks, stack trace:

Old fashioned methods work best

• browser() - I didn't say it was fun
• print() - every single function/observe statement gets a print

New methods are useful

runApp(..., display.mode="showcase")

But, doesn't show why something broke.

Final Tips

Put on your Software Development Hat!

• Test it first
  • Unit testing can save you so much headache (packages: testthat, shinytest)
  • Run tests on example data
  • Create .Md file of test plots
• Security
  • Are users uploading sensitive data?
  • Security review of public apps vs commercial version
• Speed: Why so slow?
  • Rprofiler in Rstudio
  • Reduce start up time
  • Newer tools: future and promises packages for async programming

Be Creative!

papr by Lucy D'Agostino McGowan, Nick Strayer, Jeff Leek

Have fun!

homebrewR by David Stephens

Thank you! Go forth and shiny!

Contact: @ datapointier, email: minnier-[at]-ohsu.edu

Github: https://github.com/jminnier/STARTapp

Slides available at http://bit.ly/shiny-csp18

Code for slides available at https://github.com/jminnier/talks

Slides created via the R package xaringan by Yihui Xie with the Rladies theme by Alison Hill.