Using SIGHTS R-package

Abstract

Identifying rare biological events in high-throughput screens requires using the best available normalization and statistical inference procedures. It is not always clear, however, which algorithms are best suited for a particular screen. The Statistics and dIagnostics Graphs for High Throughput Screening (SIGHTS) R package is designed for statistical analysis and visualization of HTS assays. It provides graphical diagnostic tools to guide researchers in choosing the most appropriate normalization algorithm and statistical test for identifying active constructs.

---

1 Introduction

The sights package provides numerous normalization methods that correct the three types of bias that affect High-Throughput Screening (HTS) measurements: overall plate bias, within-plate spatial bias, and across-plate bias. Commonly-used normalization methods such as Z-scores (or methods such as percent inhibition/activation which use within-plate controls to normalize) correct only overall plate bias. Methods included in this package attempt to correct all three sources of bias and typically give better results.

Two statistical tests are also provided: the standard one-sample t-test and the recommended one-sample Random Variance Model (RVM) t-test, which has greater statistical power for the typically small number of replicates in HTS. Correction for the multiple statistical testing of the large number of constructs in HTS data is provided by False Discovery Rate (FDR) correction. The FDR can be described as the proportion of false positives among the statistical tests called significant.

Included graphical and statistical methods provide the means for evaluating data analysis choices for HTS assays on a screen-by-screen basis. These graphs can be used to check fundamental assumptions of both raw and normalized data at every step of the analysis process.

Citing Methods

Please cite the sights package and specific methods as appropriate.

References for the methods can be found in this vignette, on their specific help pages, and in the manual. They can also be accessed by help(sights_method_name) in R. For example:

# Help page of SPAWN with its references
help(normSPAWN)

The package citation can be accessed in R by:

citation("sights")
>> 
>> Garg E, Murie C, Nadon R (2016). _sights: Statistics and
>> dIagnostic Graphs for HTS_. R package version 1.12.0.
>> 
>> A BibTeX entry for LaTeX users is
>> 
>>   @Manual{,
>>     title = {sights: Statistics and dIagnostic Graphs for HTS},
>>     author = {Elika Garg and Carl Murie and Robert Nadon},
>>     year = {2016},
>>     note = {R package version 1.12.0},
>>   }

2 Getting Started

2.1 Installation and loading

1. Please install the package directly from Bioconductor and load it. Note that SIGHTS requires a minimum R version of 3.3.

if (!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager")
BiocManager::install("sights")
library("sights")

2. This should also install and load the packages that SIGHTS imports: ggplot2 (Wickham, 2009), reshape2 (Wickham, 2007), qvalue (Storey, 2015), MASS (Venables and Ripley, 2002), and lattice (Sarkar, 2008).
   Otherwise, you can install/update these packages manually.

# Installing packages
BiocManager::install(c("ggplot2", "reshape2", "lattice", "MASS", "qvalue"))

# Updating packages
BiocManager::install("BiocUpgrade")
BiocManager::install()

2.2 Importing and exporting data

All SIGHTS normalization functions require that the data be arranged such that each plate is a column and each row is a well. The arrangement within each plate should be by-row first, then by-column. For more details and example, see help("ex_dataMatrix"). This required arrangement can be done in Microsoft Excel before importing the data into R, although advanced users may prefer to do so in R as needed.

1. The datasets within SIGHTS can be loaded by:

data("ex_dataMatrix")
help("ex_dataMatrix")
## Required data arrangement (by-row first) is explained.
data("inglese")

2. Your own data can be imported by giving the path of your file:

• If it is a .csv or .txt file, run

read.csv("~/yourfile.csv", header = TRUE, sep = ",")
## '~' is the folder location of your file 'yourfile.csv'.

## Use header=TRUE if you have column headers (recommended); otherwise, use
## header=FALSE.

## N.B. Be sure to use a forward slash ('/') to separate folder names.

• If it is a Microsoft Excel file, you can import it directly by installing another package:

install.packages("xlsx")
## This installs the xlsx package which enables import/export of Excel files.
library("xlsx")
read.xlsx("~/yourfile.xlsx", sheetIndex = 1)  # or
read.xlsx("~/yourfile.xlsx", sheetName = "one")
## sheetIndex is the sheet number where your data is stored in
## 'yourfile.xlsx'; sheetName is the name of that sheet.

3. Similarly any object saved in R (e.g. normalized results) can be exported as .csv or .xlsx files:

write.csv(object_name, "~/yourresults.csv")
## As a .csv file
write.xlsx(object_name, "~/yourresults.xlsx")
## As a Microsoft Excel file (requires the 'xlsx' package)

2.3 Information about data

1. There are two datasets provided within SIGHTS:

• CMBA data (Murie et al., 2015), see help("ex_dataMatrix")
• Inglese et. al. data (Inglese et al., 2006), see help("inglese")

2. Some basic information about data (including your own data after importing) can be accessed by various functions. For example, information about the Inglese et al. data set can be obtained as follows:

View(inglese)
## View the entire dataset
edit(inglese)
## Edit the dataset
head(inglese)
## View the top few rows of the dataset
str(inglese)
## Get information on the structure of the dataset
summary(inglese)
## Get a summary of variables in the dataset
names(inglese)
## Get the variable names of the dataset

2.4 Information about methods

1. There are several methods provided within SIGHTS:

• Normalization:
  • Z, Robust Z (see Malo et al. (2006)),
  • Loess (Baryshnikova et al., 2010),
  • Median Filter (Bushway et al., 2011),
  • R (Wu et al., 2008), and
  • SPAWN (Murie et al., 2015).
• Statistical testing:
  • one-sample t-test,
  • one-sample RVM t-test (Malo et al., 2006; Wright and Simon, 2003), and
  • FDR correction (Storey, 2002).
• Plotting:
  • 3d plot,
  • heatmap,
  • auto-correlation plot,
  • scatter plot,
  • boxplot,
  • inverse-gamma fit plot, and
  • histograms.

See help("normSights"), help("statSights"), help("plotSights"), and the help pages of individual methods for more information.

2. Information about the package functions can be accessed by:

ls("package:sights")
## Lists all the functions and datasets available in the package
lsf.str("package:sights")
## Lists all the functions and their usage
args(plotSights)
## View the usage of a specific function
example(topic = plotSights, package = "sights")
## View examples of a specific function

2.5 Quick reference

1. Normalization - All normalization functions are accessible either via normSights() or their individual function names (e.g. normSPAWN()).

2. Statistical tests - All statistical testing functions are accessible either via statSights() or their individual function names (e.g. statRVM()).

3. Plots - All plotting functions are accessible either via plotSights() or their individual function names (e.g. plotAutoco()).

The results of these functions can be saved as objects and called by their assigned names. For example:

library(sights)
data("inglese")
# Normalize
spawn_results <- normSPAWN(dataMatrix = inglese, plateRows = 32, plateCols = 40, 
    dataRows = NULL, dataCols = 3:44, trimFactor = 0.2, wellCorrection = TRUE, 
    biasMatrix = NULL, biasCols = 1:18)
## Or
spawn_results <- normSights(normMethod = "SPAWN", dataMatrix = inglese, plateRows = 32, 
    plateCols = 40, dataRows = NULL, dataCols = 3:44, trimFactor = 0.2, wellCorrection = TRUE, 
    biasMatrix = NULL, biasCols = 1:18)
## Access
summary(spawn_results)
# Apply statistical test
rvm_results <- statRVM(normMatrix = spawn_results, repIndex = rep(1:3, each = 3), 
    normRows = NULL, normCols = 1:9, testSide = "two.sided")
## Or
rvm_results <- statSights(statMethod = "RVM", normMatrix = spawn_results, repIndex = c(1, 
    1, 1, 2, 2, 2, 3, 3, 3), normRows = NULL, normCols = 1:9, ctrlMethod = NULL, 
    testSide = "two.sided")
## Access
head(rvm_results)
# Plot
autoco_results <- plotAutoco(plotMatrix = spawn_results, plateRows = 32, plateCols = 40, 
    plotRows = NULL, plotCols = 1:9, plotName = "SPAWN_Inglese", plotSep = TRUE)
## Or
autoco_results <- plotSights(plotMethod = "Autoco", plotMatrix = spawn_results, 
    plateRows = 32, plateCols = 40, plotRows = NULL, plotCols = c(1, 2, 3, 4, 
        5, 6, 7, 8, 9), plotName = "SPAWN_Inglese", plotSep = TRUE)
## Access
autoco_results
autoco_results[[1]]

3 Navigating through SIGHTS

We recommend the following workflow:

• Visualize the raw data to identify bias, if any:

Types of bias	Expectation, in absence of bias	Identification, in presence of bias
Plate bias	Replicate plates have similar overall distributions.	Boxplots show different medians and/or variability among replicate plates.
Within-plate spatial bias	Data within a plate is not affected by well position.	Heatmaps and 3-D plots show row and/or column effects. Auto-correlation plots show non-zero correlations at various lags; typical patterns include cyclical and/or decreasing correlation values.
Across-plate bias	Assuming few true ‘hits’ within the screen, the majority of data points should be uncorrelated across replicate plates. Only the hits should be correlated.	Scatter plots of replicate plates show strong correlation.

• Try different normalization methods and visualize the results, comparing them to raw data
• Normalize the raw data using the method that best minimizes bias
• Conduct statistical tests on the normalized data and visualize the p-value distribution

• Apply FDR correction

  We will use the Inglese et. al. dataset (Inglese et al., 2006) to demonstrate application of SIGHTS and interpretation of results.

library("sights")
data("inglese")

• Each column represents one plate of 1536-well plates
• 1^st 4 columns and last 4 columns have controls - these have been removed as they are not necessary for the normalization methods
• As required, data are arranged by row so that the well index goes from (A01, A02, A03, …)

• There are 14 concentrations with three replicates each

  We will analyze data from two of the concentrations separately:

1. Lowest Concentration (Three Replicate Plates: Exp1R1-Exp1R3)
   For these lowest concentration plates, the concentration is so low that even active molecules (as determined by a titration series) do not show activity. We use these data to show what normalized null data should look like.

2. 9^th Concentration (Three Replicate Plates: Exp9R1-Exp9R3)
   For these higher concentration plates, some compounds show activity levels. We use these data to illustrate what data for a typical experiment might look like.

3.1 Choose Normalization Method

SIGHTS has three graphical methods for visually detecting spatial bias within plates: standard “Heatmap” and “3d” plots, and autocorrelation plots (“Autoco”) (Murie et al., 2015).

3.1.1 Replicates of Lowest Concentration (Exp1R1-Exp1R3)

1. Raw Data
   The following plots will show that the raw data are affected by the three types of bias described above:

• overall plate bias
• within-plate spatial bias
• across-plate bias

sights::plotSights(plotMethod = "Box", plotMatrix = inglese, plotCols = 3:5, 
    plotName = "Raw Exp1")
>> Number of plots = 1
>> Number of plates = 3
>> Number of plate wells = 1280

Boxplots: There is overall plate bias, since the median values of the three replicate plates differ.

sights::plotSights(plotMethod = "Heatmap", plotMatrix = inglese, plateRows = 32, 
    plateCols = 40, plotRows = NULL, plotCols = 3, plotName = "Raw Exp1")
>> Number of plots = 1
>> Number of plates = 1
>> Number of plate wells = 1280
sights::plotSights(plotMethod = "3d", plotMatrix = inglese, plateRows = 32, 
    plateCols = 40, plotRows = NULL, plotCols = 3, plotName = "Raw Exp1")
>> Number of plots = 1
>> Number of plates = 1
>