Contents

1 Introduction

Technological advances continue to spur the exponential growth of biological data as illustrated by the rise of the omics—genomics, transcriptomics, epigenomics, proteomics, etc.—each with there own high throughput technologies. In order to leverage the full power of these resources, methods to integrate multiple data sets and data types must be developed. The reciprocal nature of the genomic, transcriptomic, epigenomic, and proteomic biology requires that the data provides a complementary view of cellular function and regulatory organization; however, the technical heterogeneity and massive size of high-throughput data even within a particular omic makes integrated analysis challenging. To address these challenges, we developed projectR, an R package for integrated analysis of high dimensional omic data. projectR uses the relationships defined within a given high dimensional data set, to interrogate related biological phenomena in an entirely new data set. By relying on relative comparisons within data type, projectR is able to circumvent many issues arising from technological variation. For a more extensive example of how the tools in the projectR package can be used for in silico experiments, or additional information on the algorithm, see Stein-O’Brien, et al and Sharma, et al.

2 Getting started with projectR

2.1 Installation Instructions

For automatic Bioconductor package installation, start R, and run:

BiocManager::install("genesofeve/projectR@projectionDriveR")

2.2 Methods

Projection can roughly be defined as a mapping or transformation of points from one space to another often lower dimensional space. Mathematically, this can described as a function \(\varphi(x)=y : \Re^{D} \mapsto \Re^{d}\) s.t. \(d \leq D\) for \(x \in \Re^{D}, y \in \Re^{d}\) Barbakh, Wu, and Fyfe (2009) . The projectR package uses projection functions defined in a training dataset to interrogate related biological phenomena in an entirely new data set. These functions can be the product of any one of several methods common to “omic” analyses including regression, PCA, NMF, clustering. Individual sections focusing on one specific method are included in the vignette. However, the general design of the projectR function is the same regardless.

2.3 The base projectR function

The generic projectR function is executed as follows:


projectR(data, loadings, dataNames=NULL, loadingsNames=NULL, NP = NULL, full = false)

2.3.1 Input Arguments

The inputs that must be set each time are only the data and loadings, with all other inputs having default values. However, incongruities in the feature mapping between the data and loadings, i.e. a different format for the rownames of each object, will throw errors or result in an empty mapping and should be checked before running. To overcoming mismatched feature names in the objects themselves, the dataNames and loadingNames arguments can be manually supplied by the user.

The arguments are as follows:
data a dataset to be projected into the pattern space
loadings a matrix of continous values with unique rownames to be projected
dataNames a vector containing unique name, i.e. gene names, for the rows of the target dataset to be used to match features with the loadings, if not provided by rownames(data). Order of names in vector must match order of rows in data.
loadingsNames a vector containing unique names, i.e. gene names, for the rows of loadings to be used to match features with the data, if not provided by rownames(loadings). Order of names in vector must match order of rows in loadings.
NP vector of integers indicating which columns of loadings object to use. The default of NP = NA will use entire matrix.
full logical indicating whether to return the full model solution. By default only the new pattern object is returned.

The loadings argument in the generic projectR function is suitable for use with any genernal feature space, or set of feature spaces, whose rows annotation links them to the data to be projected. Ex: the coeffients associated with individual genes as the result of regression analysis or the amplituded values of individual genes as the result of non-negative matrix factorization (NMF).

2.3.2 Output

The basic output of the base projectR function, i.e. full=FALSE, returns projectionPatterns representing relative weights for the samples from the new data in this previously defined feature space, or set of feature spaces. The full output of the base projectR function, i.e. full=TRUE, returns projectionFit, a list containing projectionPatterns and Projection. The Projection object contains additional information from the proceedure used to obtain the projectionPatterns. For the the the base projectR function, Projection is the full lmFit model from the package limma.

3 PCA projection

Projection of principal components is achieved by matrix multiplication of a new data set by previously generated eigenvectors, or gene loadings. If the original data were standardized such that each gene is centered to zero average expression level, the principal components are normalized eigenvectors of the covariance matrix of the genes. Each PC is ordered according to how much of the variation present in the data they contain. Projection of the original samples into each PC will maximize the variance of the samples in the direction of that component and uncorrelated to previous components. Projection of new data places the new samples into the PCs defined by the original data. Because the components define an orthonormal basis set, they provide an isomorphism between a vector space, \(V\), and \(\Re^n\) which preserves inner products. If \(V\) is an inner product space over \(\Re\) with orthonormal basis \(B = v_1,...,v_n\) and \(v \epsilon V s.t [v]_B = (r_1,...,r_n)\), then finding the coordinate of \(v_i\) in \(v\) is precisely the inner product of \(v\) with \(v_i\), i.e. \(r_i = \langle v,v_i \rangle\). This formulation is implemented for only those genes belonging to both the new data and the PC space. The projectR function has S4 method for class prcomp.

3.1 Obtaining PCs to project.

# data to define PCs
library(ggplot2)
data(p.RNAseq6l3c3t)

# do PCA on RNAseq6l3c3t expression data
pc.RNAseq6l3c3t<-prcomp(t(p.RNAseq6l3c3t))
pcVAR <- round(((pc.RNAseq6l3c3t$sdev)^2/sum(pc.RNAseq6l3c3t$sdev^2))*100,2)
dPCA <- data.frame(cbind(pc.RNAseq6l3c3t$x,pd.RNAseq6l3c3t))

#plot pca

setCOL <- scale_colour_manual(values = c("blue","black","red"), name="Condition:")
setFILL <- scale_fill_manual(values = c("blue","black","red"),guide = FALSE)
setPCH <- scale_shape_manual(values=c(23,22,25,25,21,24),name="Cell Line:")

pPCA <- ggplot(dPCA, aes(x=PC1, y=PC2, colour=ID.cond, shape=ID.line,
        fill=ID.cond)) +
        geom_point(aes(size=days),alpha=.6)+
        setCOL + setPCH  + setFILL +
        scale_size_area(breaks = c(2,4,6), name="Day") +
        theme(legend.position=c(0,0), legend.justification=c(0,0),
              legend.direction = "horizontal",
              panel.background = element_rect(fill = "white",colour=NA),
              legend.background = element_rect(fill = "transparent",colour=NA),
              plot.title = element_text(vjust = 0,hjust=0,face="bold")) +
        labs(title = "PCA of hPSC PolyA RNAseq",
            x=paste("PC1 (",pcVAR[1],"% of varience)",sep=""),
            y=paste("PC2 (",pcVAR[2],"% of varience)",sep=""))

3.2 Projecting prcomp objects

# data to project into PCs from RNAseq6l3c3t expression data
data(p.ESepiGen4c1l)
library(ggplot2)

PCA2ESepi <- projectR(data = p.ESepiGen4c1l$mRNA.Seq,loadings=pc.RNAseq6l3c3t,
full=TRUE, dataNames=map.ESepiGen4c1l[["GeneSymbols"]])
## [1] "93 row names matched between data and loadings"
## [1] "Updated dimension of data: 93 9"
pd.ESepiGen4c1l<-data.frame(Condition=sapply(colnames(p.ESepiGen4c1l$mRNA.Seq),
  function(x) unlist(strsplit(x,'_'))[1]),stringsAsFactors=FALSE)
pd.ESepiGen4c1l$color<-c(rep("red",2),rep("green",3),rep("blue",2),rep("black",2))
names(pd.ESepiGen4c1l$color)<-pd.ESepiGen4c1l$Cond

dPCA2ESepi<- data.frame(cbind(t(PCA2ESepi[[1]]),pd.ESepiGen4c1l))

#plot pca
library(ggplot2)
setEpiCOL <- scale_colour_manual(values = c("red","green","blue","black"),
  guide = guide_legend(title="Lineage"))

pPC2ESepiGen4c1l <- ggplot(dPCA2ESepi, aes(x=PC1, y=PC2, colour=Condition)) +
  geom_point(size=5) + setEpiCOL +
  theme(legend.position=c(0,0), legend.justification=c(0,0),
  panel.background = element_rect(fill = "white"),
  legend.direction = "horizontal",
  plot.title = element_text(vjust = 0,hjust=0,face="bold")) +
  labs(title = "Encode RNAseq in target PC1 & PC2",
  x=paste("Projected PC1 (",round(PCA2ESepi[[2]][1],2),"% of varience)",sep=""),
  y=paste("Projected PC2 (",round(PCA2ESepi[[2]][2],2),"% of varience)",sep=""))

4 NMF projection

NMF decomposes a data matrix of \(D\) with \(N\) genes as rows and \(M\) samples as columns, into two matrices, as \(D ~ AP\). The pattern matrix P has rows associated with BPs in samples and the amplitude matrix A has columns indicating the relative association of a given gene, where the total number of BPs (k) is an input parameter. CoGAPS and GWCoGAPS seek a pattern matrix (\({\bf{P}}\)) and the corresponding distribution matrix of weights (\({\bf{A}}\)) whose product forms a mock data matrix (\({\bf{M}}\)) that represents the gene-wise data \({\bf{D}}\) within noise limits (\(\boldsymbol{\varepsilon}\)). That is, \[\begin{equation} {\bf{D}} = {\bf{M}} + \boldsymbol{\varepsilon} = {\bf{A}}{\bf{P}} + \boldsymbol{\varepsilon} ..............(1) \label{eq:matrixDecomp} \end{equation}\] The number of rows in \({\bf{P}}\) (columns in \({\bf{A}}\)) defines the number of biological patterns (k) that CoGAPS/GWCoGAPS will infer from the number of nonorthogonal basis vectors required to span the data space. As in the Bayesian Decomposition algorithm Wang, Kossenkov, and Ochs (2006), the matrices \({\bf{A}}\) and \({\bf{P}}\) in CoGAPS are assumed to have the atomic prior described in Sibisi and Skilling (1997). In the CoGAPS/GWCoGAPS implementation, \(\alpha_{A}\) and \(\alpha_{P}\) are corresponding parameters for the expected number of atoms which map to each matrix element in \({\bf{A}}\) and \({\bf{P}}\), respectively. The corresponding matrices \({\bf{A}}\) and \({\bf{P}}\) are found by MCMC sampling.

Projection of CoGAPS/GWCoGAPS patterns is implemented by solving the factorization in (1) for the new data matrix where \({\bf{A}}\) is the fixed nonorthogonal basis vectors comprising the average of the posterior mean for the CoGAPS/GWCoGAPS simulations performed on the original data. The patterns \({\bf{P}}\) in the new data associated with this amplitude matrix is estimated using the least-squares fit to the new data implemented with the lmFit function in the limma package. The projectR function has S4 method for class Linear Embedding Matrix, LME.


projectR(data, loadings,dataNames = NULL, loadingsNames = NULL,
     NP = NA, full = FALSE)

4.0.1 Input Arguments

The inputs that must be set each time are only the data and patterns, with all other inputs having default values. However, inconguities between gene names–rownames of the loadings object and either rownames of the data object will throw errors and, subsequently, should be checked before running.

The arguments are as follows:

data a target dataset to be projected into the pattern space
loadings a CogapsResult object
dataNames rownames (eg. gene names) of the target dataset, if different from existing rownames of data
loadingsNames loadingsNames rownames (eg. gene names) of the loadings to be matched with dataNames
NP vector of integers indicating which columns of loadings object to use. The default of NP = NA will use entire matrix.
full logical indicating whether to return the full model solution. By default only the new pattern object is returned.

4.0.2 Output

The basic output of the base projectR function, i.e. full=FALSE, returns projectionPatterns representing relative weights for the samples from the new data in this previously defined feature space, or set of feature spaces. The full output of the base projectR function, i.e. full=TRUE, returns projectionFit, a list containing projectionPatterns and Projection. The Projection object contains additional information from the procedure used to obtain the projectionPatterns. For the the the base projectR function, Projection is the full lmFit model from the package limma.

4.1 Obtaining CoGAPS patterns to project.

# get data

AP <- get(data("AP.RNAseq6l3c3t")) #CoGAPS run data
AP <- AP$Amean
# heatmap of gene weights for CoGAPs patterns
library(gplots)
## 
## Attaching package: 'gplots'
## The following object is masked from 'package:stats':
## 
##     lowess
par(mar=c(1,1,1,1))
pNMF<-heatmap.2(as.matrix(AP),col=bluered, trace='none',
          distfun=function(c) as.dist(1-cor(t(c))) ,
          cexCol=1,cexRow=.5,scale = "row",
          hclustfun=function(x) hclust(x, method="average")
      )

4.2 Projecting CoGAPS objects

# data to project into PCs from RNAseq6l3c3t expression data

data('p.ESepiGen4c1l4')
## Warning in data("p.ESepiGen4c1l4"): data set 'p.ESepiGen4c1l4' not found
data('p.RNAseq6l3c3t')

NMF2ESepi <- projectR(p.ESepiGen4c1l$mRNA.Seq,loadings=AP,full=TRUE,
    dataNames=map.ESepiGen4c1l[["GeneSymbols"]])
## [1] "93 row names matched between data and loadings"
## [1] "Updated dimension of data: 93 9"
dNMF2ESepi<- data.frame(cbind(t(NMF2ESepi),pd.ESepiGen4c1l))

#plot pca
library(ggplot2)
setEpiCOL <- scale_colour_manual(values = c("red","green","blue","black"),
guide = guide_legend(title="Lineage"))

pNMF2ESepiGen4c1l <- ggplot(dNMF2ESepi, aes(x=X1, y=X2, colour=Condition)) +
  geom_point(size=5) + setEpiCOL +
  theme(legend.position=c(0,0), legend.justification=c(0,0),
  panel.background = element_rect(fill = "white"),
  legend.direction = "horizontal",
  plot.title = element_text(vjust = 0,hjust=0,face="bold"))
  labs(title = "Encode RNAseq in target PC1 & PC2",
  x=paste("Projected PC1 (",round(PCA2ESepi[[2]][1],2),"% of varience)",sep=""),
  y=paste("Projected PC2 (",round(PCA2ESepi[[2]][2],2),"% of varience)",sep=""))
## $x
## [1] "Projected PC1 (18.32% of varience)"
## 
## $y
## [1] "Projected PC2 (17.12% of varience)"
## 
## $title
## [1] "Encode RNAseq in target PC1 & PC2"
## 
## attr(,"class")
## [1] "labels"

5 Clustering projection

As canonical projection is not defined for clustering objects, the projectR package offers two transfer learning inspired methods to achieve the “projection” of clustering objects. These methods are defined by the function used to quantify and transfer the relationships which define each cluster in the original data set to the new dataset. Briefly, cluster2pattern uses the corelation of each genes expression to the mean of each cluster to define continuous weights. These weights are output as a pclust object which can serve as input to projectR. Alternatively, the intersectoR function can be used to test for significant overlap between two clustering objects. Both cluster2pattern and intersectoR methods are coded for a generic list structure with additional S4 class methods for kmeans and hclust objects. Further details and examples are provided in the followin respecitive sections.

5.1 cluster2pattern

cluster2pattern uses the corelation of each genes expression to the mean of each cluster to define continuous weights.


data(p.RNAseq6l3c3t)


nP<-5
kClust<-kmeans(t(p.RNAseq6l3c3t),centers=nP)
kpattern<-cluster2pattern(clusters = kClust, NP = nP, data = p.RNAseq6l3c3t)
kpattern

cluster2pattern(clusters = NA, NP = NA, data = NA)

5.1.1 Input Arguments

The inputs that must be set each time are the clusters and data.

The arguments are as follows:

clusters a clustering object
NP either the number of clusters desired or the subset of clusters to use
data data used to make clusters object

5.1.2 Output

The output of the cluster2pattern function is a pclust class object; specifically, a matrix of genes (rows) by clusters (columns). A gene’s value outside of its assigned cluster is zero. For the cluster containing a given gene, the gene’s value is the correlation of the gene’s expression to the mean of that cluster.

5.2 intersectoR

intersectoR function can be used to test for significant overlap between two clustering objects. The base function finds and tests the intersecting values of two sets of lists, presumably the genes associated with patterns in two different datasets. S4 class methods for hclust and kmeans objects are also available.

intersectoR(pSet1 = NA, pSet2 = NA, pval = 0.05, full = FALSE, k = NULL)

5.2.1 Input Arguments

The inputs that must be set each time are the clusters and data.

The arguments are as follows:

pSet1 a list for a set of patterns where each entry is a set of genes associated with a single pattern
pSet2 a list for a second set of patterns where each entry is a set of genes associated with a single pattern
pval the maximum p-value considered significant
full logical indicating whether to return full data frame of signigicantly overlapping sets. Default is false will return summary matrix.
k numeric giving cut height for hclust objects, if vector arguments will be applied to pSet1 and pSet2 in that order

5.2.2 Output

The output of the intersectoR function is a summary matrix showing the sets with statistically significant overlap under the specified \(p\)-value threshold based on a hypergeometric test. If full==TRUE the full data frame of significantly overlapping sets will also be returned.

6 Correlation based projection

Correlation based projection requires a matrix of gene-wise correlation values to serve as the Pattern input to the projectR function. This matrix can be user-generated or the result of the correlateR function included in the projectR package. User-generated matrixes with each row corresponding to an individual gene can be input to the generic projectR function. The correlateR function allows users to create a weight matrix for projection with values quantifying the within dataset correlation of each genes expression to the expression pattern of a particular gene or set of genes as follows.

6.1 correlateR


correlateR(genes = NA, dat = NA, threshtype = "R", threshold = 0.7, absR = FALSE, ...)

6.1.1 Input Arguments

The inputs that must be set each time are only the genes and data, with all other inputs having default values.

The arguments are as follows:

genes gene or character vector of genes for reference expression pattern dat
data matrix or data frame with genes to be used for to calculate correlation
threshtype Default “R” indicates thresholding by R value or equivalent. Alternatively, “N” indicates a numerical cut off.
threshold numeric indicating value at which to make threshold
absR logical indicating where to include both positive and negatively correlated genes
... addtion imputes to the cor function

6.1.2 Output

The output of the correlateR function is a correlateR class object. Specifically, a matrix of correlation values for those genes whose expression pattern pattern in the dataset is correlated (and anti-correlated if absR=TRUE) above the value given in as the threshold arguement. As this information may be useful in its own right, it is recommended that users inspect the correlateR object before using it as input to the projectR function.

6.2 Obtaining and visualizing correlateR objects.

# data to

data("p.RNAseq6l3c3t")

# get genes correlated to T
cor2T<-correlateR(genes="T", dat=p.RNAseq6l3c3t, threshtype="N", threshold=10, absR=TRUE)
cor2T <- cor2T@corM
### heatmap of genes more correlated to T
indx<-unlist(sapply(cor2T,rownames))
indx <- as.vector(indx)
colnames(p.RNAseq6l3c3t)<-pd.RNAseq6l3c3t$sampleX
library(reshape2)
pm.RNAseq6l3c3t<-melt(cbind(p.RNAseq6l3c3t[indx,],indx))
## Using indx as id variables
library(gplots)
library(ggplot2)
library(viridis)
## Loading required package: viridisLite
pCorT<-ggplot(pm.RNAseq6l3c3t, aes(variable, indx, fill = value)) +
  geom_tile(colour="gray20", size=1.5, stat="identity") +
  scale_fill_viridis(option="B") +
  xlab("") +  ylab("") +
  scale_y_discrete(limits=indx) +
  ggtitle("Ten genes most highly pos & neg correlated with T") +
  theme(
    panel.background = element_rect(fill="gray20"),
    panel.border = element_rect(fill=NA,color="gray20", size=0.5, linetype="solid"),
    panel.grid.major = element_blank(),
    panel.grid.minor = element_blank(),
    axis.line = element_blank(),
    axis.ticks = element_blank(),
    axis.text = element_text(size=rel(1),hjust=1),
    axis.text.x = element_text(angle = 90,vjust=.5),
    legend.text = element_text(color="white", size=rel(1)),
    legend.background = element_rect(fill="gray20"),
    legend.position = "bottom",
    legend.title=element_blank()
)
## Warning: Using `size` aesthetic for lines was deprecated in ggplot2 3.4.0.
## ℹ Please use `linewidth` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
## Warning: The `size` argument of `element_rect()` is deprecated as of ggplot2 3.4.0.
## ℹ Please use the `linewidth` argument instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.

6.3 Projecting correlateR objects.

# data to project into from RNAseq6l3c3t expression data
data(p.ESepiGen4c1l)


cor2ESepi <- projectR(p.ESepiGen4c1l$mRNA.Seq,loadings=cor2T[[1]],full=FALSE,
    dataNames=map.ESepiGen4c1l$GeneSymbols)
## [1] "9 row names matched between data and loadings"
## [1] "Updated dimension of data: 9 9"

7 Differential features identification.

7.1 projectionDriveR

Given loadings that define the weight of features (genes) in a given latent space (e.g. PCA, NMF), and the use of these patterns in samples, it is of interest to look at differential usage of these features between conditions. These conditions may be defined by user-defined annotations of cell type or by differential usage of a (projected) pattern. By examining differences in gene expression, weighted by the loadings that define their importance in a specific latent space, a unique understanding of differential expression in that context can be gained. This approach was originally proposed and developed in (Baraban et al, 2021), which demonstrates its utility in cross-celltype and cross-species interpretation of pattern usages.


projectionDriveR(cellgroup1, cellgroup2, loadings, loadingsNames = NULL,
                 pvalue, pattern_name, display = T, normalize_pattern = T, mode = "CI")

7.1.1 Input Arguments

The required inputs are two feature by sample (e.g. gene by cell) matrices to be compared, the loadings that define the feature weights, and the name of the pattern (column of feature loadings). If applicable, the expression matrices should already be corrected for variables such as sequencing depth.

The arguments for projectionDriveR are:

cellgroup1 Matrix 1 with features as rows, samples as columns.
cellgroup2 Matrix 2 with features as rows, samples as columns.
loadings Matrix or dataframe with features as rows, columns as patterns. Values define feature weights in that space
loadingsNames Vector of names corresponding to rows of loadings. By default the rownames of loadings will be used
pattern_name the column name of the loadings by which the features will be weighted
pvalue Determines the significance of the confidence interval to be calculated between the difference of means. Default 1e-5 display Boolean. Whether or not to plot the estimates of significant features. Default = T
normalize_pattern Boolean. Whether or not to normalize the average feature weight. Default = T
mode ‘CI’ or ‘PV’. Specifies whether to run projectionDriveR in confidence interval mode or to generate p values. Default = “CI”

7.1.2 Output

The output of projectionDriveR in confidence interval mode (‘CI’) is a list of length six mean_ci holds the confidence intervals for the difference in means for all features, weighted_mean_ci holds the confidence intervals for the weighted difference in means for all features, and normalized_weights are the weights themselves. In addition, sig_genes is a list of three vectors of gene names that are significantly different at the threshold provided generated from the mean confidence intervals (unweighted_sig_genes), the weighted mean confidence intervals (weighted_sig_genes) and genes shared between the two (significant_shared_genes) . plotted_ci returns the ggplot figure of the confidence intervals, see plotConfidenceIntervals for documentation. meta_data contains matrix names and pvalue thresholds. The output of projectionDriveR in p value mode (‘PV’) is a list of length nine. meta_data, sig_genes and normalized_weights are similar between modes. mean_stats and weighted_mean_stats contains summary information for welch t-tests. difexpgenes and weighted_difexpgenes are filtered dataframes containing differentially expressed genes at a FC and pvalue cut off of 0.2 and 1e-5 respectively. fgseavecs contain unweighted and weighted named vectors of welch-t test estimates that can be used with fgsea. plt returns the volcano ggplot figure. See pdVolcano for documentation. FC and pvalue can be manually altered by calling pdVolcano on projectionDriveR result.

7.1.3 Identifying differential features associated with learned patterns

options(width = 60)
library(dplyr, warn.conflicts = F)

#size-normed, log expression
data("microglial_counts")

#size-normed, log expression
data("glial_counts")

#5 pattern cogaps object generated on microglial_counts
data("cr_microglial")
microglial_fl <- cr_microglial@featureLoadings

#the features by which to weight the difference in expression 
pattern_to_weight <- "Pattern_1"
drivers_ci <- projectionDriveR(microglial_counts, #expression matrix
                                       glial_counts, #expression matrix
                                       loadings = microglial_fl, #feature x pattern dataframe
                                       loadingsNames = NULL,
                                       pattern_name = pattern_to_weight, #column name
                                       pvalue = 1e-5, #pvalue before bonferroni correction
                                       display = T,
                                       normalize_pattern = T, #normalize feature weights
                                       mode = "CI") #confidence interval mode