Package 'netDx'

Description

wrapper function for getting BiocFileCache associated with netDx package

Usage

.get_cache()

Value

BiocFileCache object associated with netDx

Description

built-in similarity functions

Usage

allowedSims()

Description

takes average of normdiff of each row in x

Usage

avgNormDiff(x)

Arguments

Value

symmetric matrix of size ncol(dat) (number of patients) containing pairwise patient similarities

Examples

data(xpr)
sim <- avgNormDiff(xpr[,seq_len(2)])

Description

Run nested cross-validation on data

Usage

buildPredictor(
  dataList,
  groupList,
  outDir = tempdir(),
  makeNetFunc = NULL,
  sims = NULL,
  featScoreMax = 10L,
  trainProp = 0.8,
  numSplits = 10L,
  numCores,
  JavaMemory = 4L,
  featSelCutoff = 9L,
  keepAllData = FALSE,
  startAt = 1L,
  preFilter = FALSE,
  impute = FALSE,
  preFilterGroups = NULL,
  imputeGroups = NULL,
  logging = "default",
  debugMode = FALSE
)

Arguments

Details

wrapper function to run netDx with nested cross-validation, with an inner loop of X-fold cross-validation and an outer loop of different random splits of data into train and blind test. The user needs to supply a custom function to create PSN, see createPSN_MultiData(). This wrapper provides flexibility for designs with one or several heterogeneous data types, and one or more ways of defining patient similarity. For example, designs it handles includes 1) Single datatype, single similarity metric: Expression data -> pathways 2) Single datatype, multiple metrics: Expression data -> pathways (Pearson corr) and single gene networks (normalized difference) 3) Multiple datatypes, multiple metrics: Expression -> Pathways; Clinical -> single or grouped nets

Value

symmetric matrix of size ncol(dat) (number of patients) containing pairwise patient similarities

(list) "inputNets": data.frame of all input network names. Columns are "NetType" (group) and "NetName" (network name). "Split<i>" is the data for train/test split i (i.e. one per train/test split). Each "SplitX" entry contains in turn a list of results for that split. Key-value pairs are: 1) predictions: real and predicted labels for test patients 2) accuracy: percent accuracy of predictions 3) featureScores: list of length g, where g is number of patient classes. scores for all features following feature selection, for corresponding class. 4) featureSelected: list of length g (num patient classes). List of selected features for corresponding patient class, for that train/test split. Side effect of generating predictor-related data in <outDir>.

Examples

library(curatedTCGAData)
library(MultiAssayExperiment)
curatedTCGAData(diseaseCode="BRCA", assays="*",dry.run=TRUE,version="1.1.38")

# fetch mrna, mutation data
brca <- curatedTCGAData("BRCA",c("mRNAArray"),FALSE,version="1.1.38")

# get subtype info
pID <- colData(brca)$patientID
pam50 <- colData(brca)$PAM50.mRNA
staget <- colData(brca)$pathology_T_stage
st2 <- rep(NA,length(staget))
st2[which(staget %in% c("t1","t1a","t1b","t1c"))] <- 1
st2[which(staget %in% c("t2","t2a","t2b"))] <- 2
st2[which(staget %in% c("t3","t3a"))] <- 3
st2[which(staget %in% c("t4","t4b","t4d"))] <- 4
pam50[which(!pam50 %in% "Luminal A")] <- "notLumA"                         
pam50[which(pam50 %in% "Luminal A")] <- "LumA"
colData(brca)$ID <- pID
colData(brca)$STAGE <- st2                                                 
colData(brca)$STATUS <- pam50

# keep only tumour samples
idx <- union(which(pam50 == "Normal-like"), which(is.na(st2)))
cat(sprintf("excluding %i samples\n", length(idx)))
                                                                           
tokeep <- setdiff(pID, pID[idx])
brca <- brca[,tokeep,]

pathList <- readPathways(fetchPathwayDefinitions(month=10,year=2020))
brca <- brca[,,1] # keep only clinical and mRNA data

# remove duplicate arrays
smp <- sampleMap(brca)
samps <- smp[which(smp$assay=="BRCA_mRNAArray-20160128"),]
notdup <- samps[which(!duplicated(samps$primary)),"colname"]
brca[[1]] <- brca[[1]][,notdup]

groupList <- list()
groupList[["BRCA_mRNAArray-20160128"]] <- pathList[seq_len(3)]
groupList[["clinical"]] <- list(
age="patient.age_at_initial_pathologic_diagnosis",
 stage="STAGE")
makeNets <- function(dataList, groupList, netDir,...) {
    netList <- c()
    # make RNA nets: group by pathway
    if (!is.null(groupList[["BRCA_mRNAArray-20160128"]])) {
    netList <- makePSN_NamedMatrix(dataList[["BRCA_mRNAArray-20160128"]],
                rownames(dataList[["BRCA_mRNAArray-20160128"]]),
                groupList[["BRCA_mRNAArray-20160128"]],
                netDir,verbose=FALSE,
                writeProfiles=TRUE,...)
    netList <- unlist(netList)
    cat(sprintf("Made %i RNA pathway nets\n", length(netList)))
    }

    # make clinical nets,one net for each variable
    netList2 <- c()
    if (!is.null(groupList[["clinical"]])) {
    netList2 <- makePSN_NamedMatrix(dataList$clinical,
        rownames(dataList$clinical),
        groupList[["clinical"]],netDir,
        simMetric="custom",customFunc=normDiff, # custom function
        writeProfiles=FALSE,
        sparsify=TRUE,verbose=TRUE,...)
    }
    netList2 <- unlist(netList2)
    cat(sprintf("Made %i clinical nets\n", length(netList2)))
    netList <- c(netList,netList2)
    cat(sprintf("Total of %i nets\n", length(netList)))
    return(netList)
}

# takes 10 minutes to run
#out <- buildPredictor(dataList=brca,groupList=groupList,
#   makeNetFunc=makeNets, ### custom network creation function
#   outDir=paste(tempdir(),"pred_output",sep=getFileSep()), ## absolute path
#   numCores=16L,featScoreMax=2L, featSelCutoff=1L,numSplits=2L)

Description

Performs feature selection using multiple resamplings of the data

Usage

buildPredictor_sparseGenetic(
  phenoDF,
  cnv_GR,
  predClass,
  group_GRList,
  outDir = tempdir(),
  numSplits = 3L,
  featScoreMax = 10L,
  filter_WtSum = 100L,
  enrichLabels = TRUE,
  enrichPthresh = 0.07,
  numPermsEnrich = 2500L,
  minEnr = -1,
  numCores = 1L,
  FS_numCores = NULL,
  ...
)

Arguments

Details

This function is used for feature selection of patient networks, using multiple resamplings of input data. It is intended for use in the scenario where patient networks are sparse and binary. This function should be called after defining all patient networks. It performs the following steps: For i = 1..numSplits randomly split patients into training and test (optional) filter training networks to exclude random-like networks compile features into database for cross-validation score networks out of 10 end using test samples from all resamplings, measure predictor performance.

In short, this function performs all steps involved in building and evaluating the predictor.

Value

(list) Predictor results 1) phenoDF (data.frame): subset of phenoDF provided as input, but limited to patients that have at least one event in the universe of possibilities e.g. if using pathway-level features, then this table excludes patients with zero CNVs in pathways 2) netmat (data.frame): Count of genetic events by patients (rows) in pathways (columns). Used as input to the feature selection algorithm 3) pathwayScores (list): pathway scores for each of the data splits. Each value in the list is a data.frame containing pathway names and scores. 4) enrichedNets (list): This entry is only found if enrichLabels is set to TRUE. It contains the vector of features that passed label enrichment in each split of the data. 5 - 9) Output of RR_featureTally: 5) cumulativeFeatScores: pathway name, cumulative score over N-way data resampling. 6) performance_denAllNets: positive,negative calls at each cutoff: network score cutoff (score); num networks at cutoff (numPathways) ; total +, ground truth (pred_tot); + calls (pred_ol); + calls as pct of total (pred_pct); total -, ground truth (other_tot) ; - calls (other_ol) ; - calls as pct of total (other_pct) ; ratio of pred_pct and other_pct (rr) ; min. pred_pct in all resamplings (pred_pct_min) ; max pred_pct in all resamplings (pred_pct_max) ; min other_pct in all resamplings (other_pct_min); max other_pct in all resamplings (other_pct_max) 7) performance_denEnrichedNets: positive, negative calls at each cutoff label enrichment option: format same as performance_denAllNets. However, the denominator here is limited to patients present in networks that pass label enrichment 8) resamplingPerformance: breakdown of performance for each of the resamplings, at each of the cutoffs. This is a list of length 2, one for allNets and one for enrichedNets. The value is a matrix with (resamp * 7) columns and S rows, one row per score. The columns contain the following information per resampling: 1) pred_total: total num patients of predClass 2) pred_OL: num of pred_total with a CNV in the selected net 3) pred_OL_pct: 2) divided by 1) (percent) 4) other_total: total num patients of other class(non-predClass) 5) other_OL: num of other_total with CNV in selected net 6) other_OL_pct: 5) divided by 4) (percent) 7) relEnr: 6) divided by 3).

Examples

suppressMessages(require(GenomicRanges))
suppressMessages(require(BiocFileCache))

# read CNV data
phenoFile <- system.file("extdata","AGP1_CNV.txt",package="netDx")
pheno   <- read.delim(phenoFile,sep="\t",header=TRUE,as.is=TRUE)
colnames(pheno)[1] <- "ID"
pheno <- pheno[!duplicated(pheno$ID),]

# create GRanges with patient CNVs
cnv_GR    <- GRanges(pheno$seqnames,IRanges(pheno$start,pheno$end),
                        ID=pheno$ID,LOCUS_NAMES=pheno$Gene_symbols)

# get gene coordinates
geneURL <- paste("http://download.baderlab.org/netDx/",
	"supporting_data/refGene.hg18.bed",sep="")
cache <- rappdirs::user_cache_dir(appname = "netDx")
bfc <- BiocFileCache::BiocFileCache(cache,ask=FALSE)
geneFile <- bfcrpath(bfc,geneURL)
genes <- read.delim(geneFile,sep="\t",header=FALSE,as.is=TRUE)
genes <- genes[which(genes[,4]!=""),]
gene_GR     <- GRanges(genes[,1],IRanges(genes[,2],genes[,3]),
   name=genes[,4])

# create GRangesList of pathways
pathFile <- fetchPathwayDefinitions("February",2021,verbose=TRUE)
pathwayList <- readPathways(pathFile)
path_GRList <- mapNamedRangesToSets(gene_GR,pathwayList)

#### uncomment to run - takes 5 min
#out <- buildPredictor_sparseGenetic(pheno, cnv_GR, "case",
#                             path_GRList,outDir,
#                             numSplits=3L, featScoreMax=3L,
#                             enrichLabels=TRUE,numPermsEnrich=20L,
#                             numCores=1L)
#summary(out)
#head(out$cumulativeFeatScores)

Description

Return feature selected nets based on given criteria

Usage

callFeatSel(netScores, fsCutoff, fsPctPass)

Arguments

Details

given the output of genNetScores.R and criteria for defining feature-selected (FS) nets, returns subset of nets that pass criteria. Net must score <fsCutoff> for at least <fsPctPass> considered feature-selected.

Value

(char) names of nets that pass feature-selection

Examples

data(featScores)
passed <- lapply(featScores, function(x) {
   callFeatSel(x,10,0.7) # score 10/10 in >=70% of trials
})
print(passed)

Description

Wrapper to call selected features

Usage

callOverallSelectedFeatures(
  featScores,
  featureSelCutoff,
  featureSelPct,
  cleanNames = TRUE
)

Arguments

Details

Calls features that are consistently high-scoring for predicting each class. The context for this is as follows: The original model runs feature selection over multiple splits of data into train/test samples, and each such split generates scores for all features. This function identifies features with scores that exceed a threshold for a fraction of train/test splits; the threshold and fraction are both user-specified. This function is called by the wrapper getResults(), which returns both the matrix of feature scores across splits and list of features that pass the user-specified cutoffs.

Value

(list) Feature scores for all splits, plus those passing selection for overall predictor featScores: (matrix) feature scores for each split selectedFeatures: (list) features passing selection for each class; one key per class

Examples

pathways <- paste("PATHWAY_",1:100,sep="")
highrisk <- list()
lowrisk <- list()
for (k in 1:10) { 
	highrisk[[k]] <- data.frame(PATHWAY_NAME=pathways, 
	        SCORE=floor(runif(length(pathways),min=0,max=10)),
			stringsAsFactors=FALSE);
    lowrisk[[k]] <- data.frame(PATHWAY_NAME=pathways, 
	        SCORE=floor(runif(length(pathways),min=0,max=10)),
			stringsAsFactors=FALSE);
}
names(highrisk) <- sprintf("Split%i",1:length(highrisk))
names(lowrisk) <- sprintf("Split%i",1:length(lowrisk))
callOverallSelectedFeatures(list(highrisk=highrisk,lowrisk=lowrisk), 5,0.5)

Description

internal test function to check validity of makeNetFunc and sims

Usage

checkMakeNetFuncSims(makeNetFunc, sims, groupList)

Arguments

Details

User must provide either makeNetFunc or sims. This function confirms this.

Value

(list) cleaned values for makeNetFunc and Sims

Description

checks if provided similarity functions are valid. Returns error if not

Usage

checkSimValid(sims)

Arguments

Value

TRUE if all pass check. Else throws error.

Description

Clean pathway name so it can be a filename.

Usage

cleanPathwayName(curP)

Arguments

Value

(char) Cleaned pathway name

Examples

cleanPathwayName('7-(3-AMINO-3-CARBOXYPROPYL)-WYOSINE BIOSYNTHESIS%HUMANC')

Description

Subset of CNV locations for TCGA breast tumour. Each range is associated with a patient (ID)

Usage

data(cnv_GR)

Source

The Cancer Genome Atlas. (2012). Nature 490:61-70.

References

The Cancer Genome Atlas. (2012). Nature 490:61-70.

Examples

data(cnv_GR)
head(cnv_GR)

Description

Vector of pathways that pass class enrichment

Usage

data(cnv_netPass)

Examples

data(cnv_netPass)
head(cnv_netPass)

Description

List of pathway-level feature selection scores

Usage

data(cnv_netScores)

Examples

data(cnv_netScores)
summary(cnv_netScores)
head(cnv_netScores[[1]])

Description

Binary matrix of patient occurrence in networks

Usage

data(cnv_patientNetCount)

Examples

data(cnv_patientNetCount)
head(cnv_patientNetCount)

Description

data.frame of patient labels and status for CNV example

Usage

data(cnv_pheno)

Examples

data(cnv_pheno)
head(cnv_pheno)

Description

list of train/test statuses for CNV example

Usage

data(cnv_TTstatus)

Examples

data(cnv_TTstatus)
head(cnv_TTstatus)

Description

compare intra-cluster shortest distance to overall shortest distance of the network

Usage

compareShortestPath(net, pheno, plotDist = FALSE, verbose = TRUE)

Arguments

Details

Uses Dijkstra's algorithm for weighted edges. Pairwise nodes with infinite distances are excluded before computing average shortest path for a network. This function requires the igraph package to be installed.

Value

(list) Two lists, 'avg' and 'all'. keys are cluster names. values for 'avg' are mean shortest path ; for 'all', are all pairwise shortest paths for subnetworks that contain only the edges where source and target both belong to the corresponding cluster. In addition, there is an 'overall' entry for the mean shortest distance for the entire network.

Examples

data(silh); 
colnames(silh$net)[3] <- 'weight'
compareShortestPath(silh$net, silh$groups)

Description

Create GeneMANIA database

Usage

compileFeatures(
  netDir,
  outDir = tempdir(),
  simMetric = "pearson",
  netSfx = "txt$",
  verbose = TRUE,
  numCores = 1L,
  P2N_threshType = "off",
  P2N_maxMissing = 100,
  JavaMemory = 4L,
  altBaseDir = NULL,
  debugMode = FALSE,
  ...
)

Arguments

Details

Creates a generic_db for use with GeneMania QueryRunner. The database is in tab-delimited format, and indexes are built using Apache lucene. NOTE: This pipeline expects input in the form of interaction networks and not profiles. Profile tables have patient-by-datapoint format (e.g. patient-by-genotype) Interaction networks have pairwise similarity measures: <PatientA> <PatientB><similarity> Documentation: https://github.com/GeneMANIA/pipeline/wiki/GenericDb

Value

(list). 'dbDir': path to GeneMANIA database 'netDir': path to directory with interaction networks. If profiles are provided, this points to the INTERACTIONS/ subdirectory within the text-based GeneMANIA generic database If the DB creation process results in an erorr, these values return NA

Examples

data(xpr,pheno)
pathwayList <- list(pathA=rownames(xpr)[1:10],
pathB=rownames(xpr)[21:50])

dataList <- list(rna=xpr)  #only one layer type
groupList <- list(rna=pathwayList) # group genes by pathways

makeNets <- function(dataList, groupList, netDir,...) {
    netList <- makePSN_NamedMatrix(dataList[['rna']],
				rownames(dataList[['rna']]),
         groupList[['rna']],netDir,verbose=FALSE,
				writeProfiles=TRUE,...)
    unlist(netList)
}
tmpDir <- tempdir(); netDir <- paste(tmpDir,"nets",
sep=getFileSep())
if (file.exists(netDir)) unlink(netDir,recursive=TRUE)
dir.create(netDir,recursive=TRUE)

pheno_id <- setupFeatureDB(pheno,netDir)
netList <- createPSN_MultiData(dataList=dataList, groupList=groupList,
    pheno=pheno_id,netDir=netDir,makeNetFunc=makeNets,verbose=TRUE)

outDir <- paste(tmpDir,'dbdir',sep=getFileSep()); 
dir.create(outDir)
dbDir <- compileFeatures(netDir,outDir)

Description

Tally the score of networks through cross-validation

Usage

compileFeatureScores(fList, filter_WtSum = 100, verbose = FALSE)

Arguments

Value

(data.frame) Feature name and score; includes features that occur at least once in fList.

Examples

netDir <- system.file("extdata","GM_NRANK",package="netDx")
netFiles <- sprintf('%s/%s', netDir,dir(netDir,pattern='NRANK$'))
pTally <- compileFeatureScores(netFiles,verbose=TRUE)
print(head(pTally))

Description

Sample table of True/False Positives and Negatives for various feature selection cutoffs tp: true positive rate, fp: false positive rate, tn: true negative rate, fn: false negative rate

Usage

data(confmat)

Examples

data(confmat)
head(confmat)

Description

Make confusion matrix

Usage

confusionMatrix(model)

Arguments

Details

Creates a confusion matrix, a square matrix which indicates the fraction of times patients in a class are correctly classified, versus misclassified as each of the other classes. Here, the confusion matrix is computed once per train-test split and the average is displayed. For this reason, the fractions may not cleanly add up to 100

Value

(list) confusion matrix for all train/test splits and final averaged matrix Side effect of plotting the averaged matrix.

Examples

data(toymodel)
confusionMatrix(toymodel)

Description

Convert profiles to interaction networks before integration

Usage

convertProfileToNetworks(
  netDir,
  outDir = tempdir(),
  simMetric = "pearson",
  numCores = 1L,
  JavaMemory = 4L,
  GM_jar = NULL,
  P2N_threshType = "off",
  P2N_maxMissing = 100,
  netSfx = "txt$",
  debugMode = FALSE
)

Arguments

Details

In preparation for network integration. When using GeneMANIA's built-in functionality to create PSN using ProfileToNetworkDriver, this step needs to run to process profiles to networks. These are currently used for Pearson correlation-based networks and those using mutual information.

Value

No value. Side effect of creating interaction networks in outDir.

Description

Wrapper that converts an input list into a MultiAssayExperiment object

Usage

convertToMAE(dataList)

Arguments

Details

This function takes in a list of key-value pairs (keys: data types, values: matrices/dataframes) and calls the necessary functions from the MultiAssayExperiment package to incorporate the values from the input list into a MultiAssayExperiment object, transforming the values according to the keys. If duplicate sample names are found in the assay data, only the first instance is kept.

Value

MAE (MultiAssayExperiment) data from input list incorporated into a MultiAssayExperiment object, compatible with further analysis using the netDx algorithm.

Examples

data(xpr, pheno)

# Generate random proteomic data
prot <- matrix(rnorm(100*20), ncol=20)
colnames(prot) <- sample(pheno$ID, 20)
rownames(prot) <- sprintf("protein%i",1:100)	

myList <- list(rna = xpr, proteomic = prot, pheno = pheno)

MAE <- convertToMAE(myList)

Description

Counts the number of (+,+) and (+,-) interactions in a single network

Usage

countIntType(inFile, plusID, minusID)

Arguments

Value

(numeric of length 2) Number of (+,+) interactions, and non-(+,+) interactions (i.e. (+,-) and (-,-) interactions)

Examples

d <- tempdir()
# write PSN
m1 <- matrix(c("P1","P1","P2","P2","P3","P4",1,1,1),byrow=FALSE,ncol=3)
write.table(m1,file=paste(d,"net1.txt",sep=getFileSep()),
sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)

countIntType(paste(d,"net1.txt",sep=getFileSep()),c("P1","P2","P3"),
c("P4","P5"))

Description

Counts number of (+,+) and (+,-) interactions in a set of networks

Usage

countIntType_batch(
  inFiles,
  plusID,
  minusID,
  tmpDir = tempdir(),
  enrType = "binary",
  numCores = 1L
)

Arguments

Value

(matrix) two columns, one row per network If enrType="binary", number of (+,+) and other interactions Otherwise if enrType="corr" mean edge weight of (+,+) edges and of other edges

Examples

d <- tempdir()
# write PSN
m1 <- matrix(c("P1","P1","P2","P2","P3","P4",1,1,1),byrow=FALSE,ncol=3)
write.table(m1,file=paste(d,"net1.txt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)
m2 <- matrix(c("P3","P4",1),nrow=1)
write.table(m2,file=paste(d,"net2.txt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)

countIntType_batch(paste(d,c("net1.txt","net2.txt"),sep=getFileSep()),
	c("P1","P2","P3"),c("P4","P5"))

Description

Count number of patients in a network

Usage

countPatientsInNet(netDir, fList, ids)

Arguments

Details

This functionality is needed to count patient overlap when input data is in a form that results in highly missing data, rather than when the same measures are available for almost all patients. An example application is when patient networks are based on unique genomic events in each patients (e.g. CNVs or indels), rather than 'full-matrix' data (e.g. questionnaires or gene expression matrices). The former scenario requires an update in the list of eligible networks each time some type of patient subsetting is applied (e.g. label enrichment, or train/test split). A matrix with patient/network membership serves as a lookup table to prune networks as feature selection proceeds

Value

(matrix) Size P by N, where P is num patients and N is number of networks networks; a[i,j] =1 if patient i in network j, else 0

Examples

d <- tempdir()
pids <- paste("P",1:5,sep="")
m1 <- matrix(c("P1","P1","P2","P2","P3","P4",1,1,1),
	byrow=FALSE,ncol=3)
write.table(m1,
file=paste(d,"net1.txt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)
m2 <- matrix(c("P3","P4",1),nrow=1)
write.table(m2,
file=paste(d,"net2.txt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)
x <- countPatientsInNet(d,c("net1.txt","net2.txt"), pids)

Description

Create PSN from provided similarities

Usage

createNetFuncFromSimList(
  dataList,
  groupList,
  netDir,
  sims,
  verbose = TRUE,
  ...
)

Arguments

Details

Called by CreatePSN_MultiData(), this is the function that converts user-provided simlarity metrics to internal netDx function calls to generate nets.

Description

Wrapper to create custom input features (patient similarity networks)

Usage

createPSN_MultiData(
  dataList,
  groupList,
  pheno,
  netDir = tempdir(),
  filterSet = NULL,
  verbose = TRUE,
  makeNetFunc = NULL,
  sims = NULL,
  ...
)

Arguments

Value

(char) vector of network names. Side effect of creating the nets

Examples

library(curatedTCGAData)
library(MultiAssayExperiment)
curatedTCGAData(diseaseCode='BRCA', assays='*',dry.run=TRUE,version="1.1.38")

# fetch mrna, mutation data
brca <- curatedTCGAData('BRCA',c('mRNAArray'),FALSE,version="1.1.38")

# get subtype info
pID <- colData(brca)$patientID
pam50 <- colData(brca)$PAM50.mRNA
staget <- colData(brca)$pathology_T_stage
st2 <- rep(NA,length(staget))
st2[which(staget %in% c('t1','t1a','t1b','t1c'))] <- 1
st2[which(staget %in% c('t2','t2a','t2b'))] <- 2
st2[which(staget %in% c('t3','t3a'))] <- 3
st2[which(staget %in% c('t4','t4b','t4d'))] <- 4
pam50[which(!pam50 %in% 'Luminal A')] <- 'notLumA'                         
pam50[which(pam50 %in% 'Luminal A')] <- 'LumA'
colData(brca)$ID <- pID
colData(brca)$STAGE <- st2                                                 
colData(brca)$STATUS <- pam50

# keep only tumour samples
idx <- union(which(pam50 == 'Normal-like'), which(is.na(st2)))
cat(sprintf('excluding %i samples\n', length(idx)))
                                                                           
tokeep <- setdiff(pID, pID[idx])
brca <- brca[,tokeep,]

pathList <- readPathways(fetchPathwayDefinitions("October",2020))

brca <- brca[,,1] # keep only clinical and mRNA data

# remove duplicate arrays
smp <- sampleMap(brca)
samps <- smp[which(smp$assay=='BRCA_mRNAArray-20160128'),]
notdup <- samps[which(!duplicated(samps$primary)),'colname']
brca[[1]] <- brca[[1]][,notdup]

groupList <- list()
groupList[['BRCA_mRNAArray-20160128']] <- pathList[seq_len(3)]
makeNets <- function(dataList, groupList, netDir,...) {
    netList <- c()
    # make RNA nets: group by pathway
    if (!is.null(groupList[['BRCA_mRNAArray-20160128']])) {
    netList <- makePSN_NamedMatrix(dataList[['BRCA_mRNAArray-20160128']],
                rownames(dataList[['BRCA_mRNAArray-20160128']]),
                groupList[['BRCA_mRNAArray-20160128']],
                netDir,verbose=FALSE,
                writeProfiles=TRUE,...)
    netList <- unlist(netList)
    cat(sprintf('Made %i RNA pathway nets\n', length(netList)))
    }

    cat(sprintf('Total of %i nets\n', length(netList)))
    return(netList)
}

exprs <- experiments(brca)
datList2 <- list()
for (k in seq_len(length(exprs))) {
 tmp <- exprs[[k]]
 df <- sampleMap(brca)[which(sampleMap(brca)$assay==names(exprs)[k]),]
 colnames(tmp) <- df$primary[match(df$colname,colnames(tmp))]
 tmp <- as.matrix(assays(tmp)[[1]]) # convert to matrix
 datList2[[names(exprs)[k]]]<- tmp
}
pheno <- colData(brca)[,c('ID','STATUS')]
netDir <- tempdir()
pheno_id <- setupFeatureDB(colData(brca),netDir)
createPSN_MultiData(dataList=datList2,groupList=groupList,
 pheno=pheno_id,
 netDir=netDir,makeNetFunc=makeNets,numCores=1)

Description

Convert MultiAssayExperiment object to list and data.frame

Usage

dataList2List(dat, groupList)

Arguments

Details

Used by internal routines in netDx

Value

(list) Keys are: 1) assays: list of matrices, each corresponding to data from a particular layer 2) pheno: (data.frame) sample metadata

Examples

data(xpr,pheno)
require(MultiAssayExperiment)
objlist <- list("RNA"=SummarizedExperiment(xpr))
mae <- MultiAssayExperiment(objlist,pheno)
groupList <- list(RNA=rownames(xpr))
dl <- dataList2List(mae,groupList)
summary(dl)

Description

Score networks based on their edge bias towards (+,+) interactions

Usage

enrichLabelNets(
  netDir,
  pheno_DF,
  outDir,
  numReps = 50L,
  minEnr = -1,
  outPref = "enrichLabelNets",
  verbose = TRUE,
  setSeed = 42L,
  enrType = "binary",
  numCores = 1L,
  predClass,
  tmpDir = tempdir(),
  netGrep = "_cont.txt$",
  getShufResults = FALSE,
  ...
)

Arguments

Details

Determines which networks are statistically enriched for interactions between the class of interest. The resulting ENR score and corresponding p-value serve as a filter to exclude random-like interaction networks before using feature selection. This filter is known to be important when patient networks are sparse and binary; e.g. networks based on shared overlap of CNV locations. If the filter is not applied, GeneMANIA WILL promote networks with slight bias towards (+,+) edges , even if these are small and random-like.

The measure of (+,+)-enrichment is defined as: ENR(network N) = ((num (+,+) edges) - (num other edges))/(num edges). A p-value for per-network ENR is obtained non-parametrically by measuring a null distribution for ENR following multiple permutations of case-control labels.

Value

(data.frame) networks stats from clique-filtering, one record per network

Examples

data(npheno)
netDir <- system.file("extdata","example_nets",package="netDx")
x <- enrichLabelNets(netDir,npheno,".",predClass="case",netGrep="txt$",
	numReps=5)
print(x)

Description

List with one entry per patient label ("SURVIVEYES" and "SURVIVENO"). Each entry contains scores obtained through feature-selection acros 100 train/test splits. Scores range from 0 to 10. Scores in data.frame format, with rows corresponding to features and columns to a particular train/test split.

Usage

data(featScores)

Examples

data(featScores)
head(featScores)

Description

fetch pathway definitions from downloads.baderlab.org

Usage

fetchPathwayDefinitions(month = NULL, year = NULL, day = 1, verbose = FALSE)

Arguments

Details

Fetches genesets compiled from multiple curated pathway databases. Downloaded from: https://download.baderlab.org/EM_Genesets/ The file contains pathways from HumanCyc, NetPath, Reactome, NCI Curated Pathways and mSigDB. For details see Merico D, Isserlin R, Stueker O, Emili A and GD Bader. (2010). PLoS One. 5(11):e13984.

Value

(char) Path to local cached copy of GMT file or initial download is required

Examples

fetchPathwayDefinitions("October",2021)
fetchPathwayDefinitions("October",2021)
fetchPathwayDefinitions(month=10,year=2021)

Description

data.frame object with columns of (gene) RefSeq ID (name), chromosome (chrom), strand, transcription start site (txStart), transcription end site (txEnd), and gene symbol (name2)

Usage

data(genes)

Examples

data(genes)
head(genes)

Description

Counts the relative correlation of (+,+) and (+,-)(-,-) interactions

Usage

getCorrType(inFile, plusID, minusID)

Arguments

Value

(numeric) mean edge weight for (+,+) and other edges

Description

write enrichment map for consensus nets

Usage

getEMapInput(
  featScores,
  namedSets,
  netInfo,
  pctPass = 0.7,
  minScore = 1,
  maxScore = 10,
  trimFromName = c(".profile", "_cont"),
  verbose = FALSE
)

Arguments

Value

(list) Length two. 1) nodeAttrs: data.frame of node attributes 2) featureSets: key-value pairs of selected feature sets (e.g. if pathway features are used, keys are pathway names, and values are member genes).

Examples

inDir <- system.file("extdata","example_output",package="netDx")
outDir <- paste(tempdir(),'plots',sep='/')
if (!file.exists(outDir)) dir.create(outDir)
featScores <- getFeatureScores(inDir,predClasses=c('LumA','notLumA'))
gp <- names(featScores)[1]
pathwayList <- readPathways(fetchPathwayDefinitions("October",2020))
pathwayList <- pathwayList[seq_len(5)]
netInfoFile <- system.file("extdata","example_output/inputNets.txt",package="netDx")
netInfo <- read.delim(netInfoFile,sep='\t',h=FALSE,as.is=TRUE)
emap_input <- getEMapInput(featScores[[gp]],pathwayList,netInfo)
summary(emap_input)

Description

Wrapper to generate multiple EnrichmentMaps (perhaps one per class)

Usage

getEMapInput_many(featScores, namedSets_valid, netTypes, outDir, ...)

Arguments

Value

(list) of length g, where g is the number of groups in featScores. Values are lists, corresponding to the output of getEmapInput.R

Examples

data(featScores)

pathwayList <- readPathways(fetchPathwayDefinitions("October",2020))
pathwayList <- pathwayList[seq_len(5)]

netInfoFile <- system.file("extdata","example_output/inputNets.txt",package="netDx")
netTypes <- read.delim(netInfoFile,sep='\t',h=FALSE,as.is=TRUE)
outDir <- paste(tempdir(),'plots',sep='/')
if (!file.exists(outDir)) dir.create(outDir)
EMap_input <- getEMapInput_many(featScores,pathwayList,
     netTypes,outDir=outDir)

Description

Get ENR for all networks in a specified directory

Usage

getEnr(
  netDir,
  pheno_DF,
  predClass,
  netGrep = "_cont.txt$",
  enrType = "binary",
  ...
)

Arguments

Details

For each network, compute the number of (+,+) and other (+,-),(-,+),(-,-) interactions. From this compute network ENR. The measure of (+,+)-enrichment is defined as: ENR(network N) = ((num (+,+) edges) - (num other edges))/(num edges). A network with only (+,+) interactions has an ENR=1 ; a network with no (+,+) interactions has an ENR=-1; a network with a balance of the two has ENR=0.

Value

(list): 1) plusID (char) vector of + nodes 2) minusID (char) vector of - nodes 3) orig_rat (numeric) ENR for data networks 4) fList (char) set of networks processed 5) orig (data.frame) output of countIntType_batch for input networks

Examples

d <- tempdir()
options(stringsAsFactors=FALSE)
pids <- paste("P",seq_len(5),sep="")
pheno <- data.frame(ID=pids,STATUS=c(rep("case",3),rep("control",2)))

# write PSN
m1 <- matrix(c("P1","P1","P2","P2","P3","P4",1,1,1),byrow=FALSE,ncol=3)
write.table(m1,file=paste(d,"net1.nettxt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)
m2 <- matrix(c("P3","P4",1),nrow=1)
write.table(m2,file=paste(d,"net2.nettxt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)

# compute enrichment
x <- countPatientsInNet(d,dir(d,pattern=c("net1.nettxt","net2.nettxt")), pids)
getEnr(d,pheno,"case","nettxt$")

Description

Compile network scores into a matrix

Usage

getFeatureScores(inDir, predClasses, getFullCons = TRUE)

Arguments

Details

Given network scores over a set of train/test splits, compiles these into a matrix for downstream analysis. See the section on 'Output Files'

Value

(list) one key per patient class. Value is matrix of network scores across all train/test splits. Each score is the output of the inner fold of CV.

Examples

inDir <- system.file("extdata","example_output",package="netDx")
netScores <- getFeatureScores(inDir, predClasses = c('LumA','notLumA'))

Description

Returns OS-specific file separator

Usage

getFileSep()

Value

(char) "\" if Windows, else "/"

Examples

getFileSep()

Description

download and update GeneMANIA jar file

Usage

getGMjar_path(verbose = FALSE)

Arguments

Value

(char) Path to local cached copy of GeneMANIA jar file.. or initial download is required

Examples

getGMjar_path()

Description

compile net score across a set of predictor results

Usage

getNetConsensus(scorelist)

Arguments

Details

used to compare how individual nets score for different predictor configurations

Value

(data.frame) Rownames are union of all nets in the input list. Columns show net scores for each key of the input list. Where a net is not found in a given list, it is assigned the value of NA

Examples

pathways <- paste("PATHWAY_",1:100,sep="")
highrisk <- list()
for (k in 1:10) { 
	highrisk[[k]] <- data.frame(PATHWAY_NAME=pathways, 
	SCORE=runif(length(pathways),min=0,max=10),
			stringsAsFactors=FALSE);
}
names(highrisk) <- sprintf("Split%i",1:length(highrisk))
x <- getNetConsensus(highrisk)

Description

Get relative proportion of patient classes that contribute to a set of networks

Usage

getOR(pNetworks, pheno_DF, predClass, netFile, verbose = TRUE)

Arguments

Details

Feature selected networks should have the property of being enriched in the class of interest; e.g. be enriched in 'case' relative to 'control'. When given a list of networks N, this method computes the number and proportion of patients that overlap N. A high relative fraction of the predicted class indicates successful feature selection. To create a ROC or precision-recall curve, several calls can be made to this function, one per cutoff.

Value

List. 1) stats: statistics on group overlap with , This is a 2xK matrix, where rows are classes (predClass,other), and columns are: total samples, samples overlapping nets, 2) relEnr: relative enrichment of predClass over other

Examples

d <- tempdir()
options(stringsAsFactors=FALSE)
pids <- paste("P",seq_len(5),sep="")
pheno <- data.frame(ID=pids,STATUS=c(rep("case",3),rep("control",2)))

# write PSN
m1 <- matrix(c("P1","P1","P2","P2","P3","P4",1,1,1),byrow=FALSE,ncol=3)
write.table(m1,file=paste(d,"net1.txt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)
m2 <- matrix(c("P3","P4",1),nrow=1)
write.table(m2,file=paste(d,"net2.txt",sep=getFileSep()),sep="\t",
col.names=FALSE,row.names=FALSE,quote=FALSE)

# compute enrichment
x <- countPatientsInNet(d,dir(d,pattern=c("net1.txt","net2.txt")), pids)
getOR(x,pheno,"case",colnames(x)) # should give large RelEnr

Description

Calculates patient-level classification accuracy across train/test splits

Usage

getPatientPredictions(predFiles, pheno, plotAccuracy = FALSE)

Arguments

Details

Takes all the predictions across the different train/test splits, and for each patient, generates a score indicating how many times they were classified by netDx as belonging to each of the classes. The result is that we get a measure of individual classification accuracy across the different train/test splits.

Value

(list) of length 2. 1) (data.frame) rows are patients, (length(predFiles)+2) columns. Columns seq_len(length(predFiles)): Predicted labels for a given split (NA if patient was training sample for the split). Column (length(predFiles)+1): split, value is NA. Columns are : ID, REAL_STATUS, predStatus1,... predStatusN. Side effect of plotting a dot plot of and the value is '

Examples

inDir <- system.file("extdata","example_output",package="netDx")
data(pheno)
all_rngs <- list.dirs(inDir, recursive = FALSE)
all_pred_files <- unlist(lapply(all_rngs, function(x) {
    paste(x, 'predictionResults.txt', 
	sep = getFileSep())}))
pred_mat <- getPatientPredictions(all_pred_files, pheno)

Description

Process GM PRANK files to get the ROC curve for the query

Usage

getPatientRankings(pFile, pheno_DF, predClass, plotIt = FALSE, verbose = FALSE)

Arguments

Value

(list) 1) predLbl: GeneMANIA scores (predicted labels). Higher score for higher ranked patient. 2) realLbl: binary value indicating if patient label matches predictor label (real labels) 3) fullmat: pheno_DF merged with similarity scores ('similarityScore') and real label ('isPredClass') 4) roc: output of ROCRs performance(,'tpr','fpr') - ROC curve 5) auc: output of ROCRs auc() 6) precall: output of ROCRs performance(, 'prec','rec') 7) f: output of ROCRs performance(,'f') If < 2 patients in PRANK file, roc,auc, precall, f are all returned as NA.

Examples

data(pheno)
prankFile <- system.file("extdata",
paste("GM_PRANK","CV_1.query-results.report.txt.PRANK",sep=getFileSep()),
	package="netDx")
x <- getPatientRankings(prankFile, pheno, 'LumA')

Description

performance metrics for model

Usage

getPerformance(res, predClasses)

Arguments

Value

(list) 1) rocCurve: ROCR performance object for ROC curve 2) prCurve: ROCR performance object for PR curve 3) auroc: Area under ROC curve 4) aupr: Area under PR curve 5) accuracy: Accuracy

Description

get the integrated patient similarity network made of selected features

Usage

getPSN(
  dat,
  groupList,
  makeNetFunc = NULL,
  sims = NULL,
  selectedFeatures,
  plotCytoscape = FALSE,
  aggFun = "MEAN",
  prune_pctX = 0.3,
  prune_useTop = TRUE,
  numCores = 1L,
  calcShortestPath = FALSE
)

Arguments

Details

An integrated patient similarity network can be built using combined top features for each patient class. Such a network is created by taking the union of selected features for all patient labels, and aggregating pairwise edges for all of them using a user-specified function (aggFun). The network is then pruned prior to visualization, using a user-specified fraction of strongest edges (prune_pctX, prune_useTop). In addition, the user may quantify the distance between patients of the same class, relative to those of other classes, using Dijkstra distance (calcShortestPath flag).

Value

(list) information about the integrated network similarity network 2) patientDistNetwork_pruned (matrix) the network plotted in Cytoscape. Also note that this is a dissimilarity network, so that more similar nodes have smaller edge weights 3) colLegend (data.frame): legend for the patient network plotted in Cytoscape. Columns are node labels (STATUS) and colours (colour) 6) outDir (char) value of outDir parameter

Description

Returns overlapping named ranges for input ranges

Usage

getRegionOL(gr, rngList)

Arguments

Details

Given a set of query GRanges, and a subject list-of-GRanges, updates the query with a column 'LOCUS_NAMES' containing the names of ranges overlapped by the query. One application is to map structural variants, such as CNVs, to genes in pathways of interest. In this scenario gr would contain the patient CNVs, and rngList would be a list of GenomicRanges objects, one per cellular pathway.

Value

(GRanges) query ranges with the added column 'LOCUS_NAMES'. Where a range overlaps with multiple loci, the names are reported as a comma-separated vector

Examples

data(cnv_GR,pathway_GR)
x <- getRegionOL(cnv_GR,pathway_GR)

Description

Compiles performance and selected features for a trained model.

Usage

getResults(res, status, featureSelCutoff = 1L, featureSelPct = 0)

Arguments

Details

This function is run after training a model using buildPredictor(). It takes patient input data, model output, and returns performance and selected features.

Value

list of results. - selectedFeatures (list of character vectors): list, one per class - performance (list of mixed datatypes) including mean accuracy (meanAccuracy), split-level accuracy (splitAccuracy), split-level AUROC (auroc), split-level AUPR (splitAUR) Side effect of plotting ROC curve if binary classifier

Examples

data(toymodel) # load example results from binary breast classification
patlabels <- names(toymodel$Split1$featureSelected)
getResults(toymodel,patlabels,2,0.5)

Description

Measures of patient similarity

Usage

getSimilarity(x, type = "pearson", customFunc, ...)

Arguments

Value

symmetric matrix of size N, where N is number of samples

Examples

data(xpr) 
x <- getSimilarity(xpr) # similarity by Pearson corr
mySim <- function(x) cor(x,method='kendall')
x <- getSimilarity(xpr,customFunc=mySim) # custom similarity

Description

Wrapper to create input files for Enrichment Map

Usage

makeInputForEnrichmentMap(
  model,
  results,
  pathwayList,
  EMapMinScore = 0L,
  EMapMaxScore = 1L,
  EMapPctPass = 0.5,
  outDir
)

Arguments

Details

An Enrichment Map is a network-based visualization of top-scoring pathway features and themes. It is generated in Cytoscape. This script generates the input files needed for Cytoscape to create an Enrichment Map visualization.

Value

(list) 1) GMTfiles (char): GMT files used to create EnrichmentMap in Cytoscape. 2) NodeStyles (char): .txt files used to assign node attributes in Cytoscape. Importantly, attributes include node fill, which indicates the highest consistent score for a given feature.

Description

Create patient networks from full matrix of named measurements

Usage

makePSN_NamedMatrix(
  xpr,
  nm,
  namedSets,
  outDir = tempdir(),
  simMetric = "pearson",
  verbose = TRUE,
  numCores = 1L,
  writeProfiles = TRUE,
  sparsify = FALSE,
  useSparsify2 = FALSE,
  cutoff = 0.3,
  sparsify_edgeMax = Inf,
  sparsify_maxInt = 50,
  minMembers = 1L,
  runSerially = FALSE,
  ...
)

Arguments

Details

Creates patient similarity networks when full matrices of data are provided (e.g. gene expression, questionnaire results). To generate networks from sparse data such as CNVs or indels, use makePSN_RangeSets instead. The rows of the data matrix (xpr) must be named (nm); one network is create for each named set (namedSets). There are two options for the way in which networks are created, depending on the value of writeProfiles. 1. writeProfiles=TRUE: GeneMANIA is used to generate interaction networks and sparsify networks. This only works if the desired measure of similarity is network-level Pearson correlation; an example is networks at the level of pathways. In this case, the user does not explicitly specify a similarity measure and simMetric is ignored. 2. writeProfiles=FALSE: GeneMANIA is not used to generate interaction networks. Rather, netDx uses simMetric to create interaction networks. Networks can be sparsified by excluding weak connections (cutoff).

Value

(char) Basename of files to which networks are written. Side effect of writing interaction networks in outDir

Examples

data(xpr,pheno,pathwayList); 
# you may get a warning message that the output directory already
# exists; ignore it
out <- makePSN_NamedMatrix(xpr,rownames(xpr),pathwayList, 
	'.',writeProfiles=TRUE)

Description

Create patient similarity interaction networks based on range sets

Usage

makePSN_RangeSets(
  gr,
  rangeSet,
  netDir = tempdir(),
  simMetric = "coincide",
  quorum = 2L,
  verbose = TRUE,
  numCores = 1L
)

Arguments

Details

Creates patient similarity networks when data consist of genomic events associated with patients. Examples include CNV or indel data for patients. To generate networks from full matrices such gene expression data, use makePSN_NamedMatrix instead. Genomic ranges corresponding to events in patients (gr) should be named. One network is created per named range set (rangeSet). Each set reflects a group of related loci ; for example, genomic ranges associated with genes in the same cellular pathway. Currently, the only similarity measure supported is binary; two patients are related in a network N if they both overlap elements of set N.

Value

Vector of network filenames

Examples

data(pathway_GR,cnv_GR)
### # example commented out to avoid build errors because of parallel
### # execution. Uncomment to run.
### netList <- makePSN_RangeSets(cnv_GR,pathway_GR,'.')

Description

Randomly select patients for queries for feature selection

Usage

makeQueries(incPat, featScoreMax = 10L, verbose = TRUE)

Arguments

Value

(list) of length featScoreMax, containing names of patients in query file for each fold

Examples

data(pheno)
x <- makeQueries(pheno$ID)

Description

Convert a network in source-target-weight format to symmetric matrix

Usage

makeSymmetric(x, verbose = FALSE)

Arguments

Details

A common format for network representation is to use a three column table listing source node, target node, and weight. This is the format netDx uses for network integration and visualization in Cytoscape. However, some functionality requires a square symmetric adjacency matrix. This function takes as input the three-column format and converts to the adjacency matrix. NOTE: Symmetric attribute is assumed, and the function automatically sets a[i,j] = a[j,i]. Diagonal is assumed to have value of 1.0. Finally missing edges will be assigned NA values.

Value

(matrix) symmetric adjacency matrix

Examples

src <- c("A","B"); tgt <- c("C","C")
cur <- data.frame(source=src,target=tgt,weight=c(0.3,0.8))
makeSymmetric(cur)

Description

Map named ranges to corresponding set of named ranges

Usage

mapNamedRangesToSets(gr, rangeList, verbose = FALSE)

Arguments

Details

Example application is when we have named ranges each corresponding to genes or regulatory elements, and we wish to group these ranges based on metabolic pathway.

Value

RangeList. keys are names of rangeList, values are GRanges

Examples

data(genes,pathwayList); 
gene_GR<-GenomicRanges::GRanges(genes$chrom,
  IRanges::IRanges(genes$txStart,genes$txEnd),
		name=genes$name2)
path_GRList <- mapNamedRangesToSets(gene_GR,pathwayList)

Description

Converts matrix index (1 to m*n) to row (m) and column (n) number

Usage

matrix_getIJ(dimMat, idx)

Arguments

Value

(matrix) n-by-2, first column has row indices ; second column has col indices

Description

data.frame with patient ID and tumour subtype (STATUS)

Usage

data(MB.pheno)

Source

Northcott et al. (2011). J Clin Oncol. 29 (11):1408.

References

Northcott et al. (2011). J Clin Oncol. 29 (11):1408.

Examples

data(MB.pheno)
head(MB.pheno)

Description

Output of getResults() generated by running toymodel. toymodel is itself the output of buildPredictor() run on a simple dataset for binary breast tumour classification using two genomic data sources. BRCA data were downloaded using curatedTCGAData for mRNA and miRNA expression. buildPredictor()] was run by scoring features out of 2, with selected features passing 1 out of 2. Tumours were labelled either "Luminal.A" or "other". See details of getResults() for output format.

Usage

data(modelres)

Examples

data(modelres)
head(modelres)

Description

moves interaction networks when compiling database for sparse genetic workflow

Usage

moveInteractionNets(netDir, outDir, pheno, fileSfx = "_cont.txt")

Arguments

Value

No value. Side effect of moving interaction nets to target directory and creating network-related metadata files used to compile feature database

Description

Similarity metric of normalized difference

Usage

normDiff(x)

Arguments

Details

Similarity metric used when data for a network consists of exactly 1 continuous variable (e.g. a network based only on 'age'). When number of variables is 2-5, use avgNormDiff() which takes the average of normalized difference for individual variables

Value

symmetric matrix of size ncol(dat) (number of patients) containing pairwise patient similarities

Examples

sim <- normDiff(rnorm(10))

Description

data.frame with patient ID ("ID") and label ("STATUS"). 100 "cases" and 100 "controls"

Usage

data(npheno)

Examples

data(npheno)
head(npheno)

Description

List object. Keys are pathway names, values are GRanges objects with coordinates of corresponding genes. Small subset of pathways sufficient for package examples.

Usage

data(pathway_GR)

Examples

data(pathway_GR)
head(pathway_GR)

Description

List where keys are pathway names and values are character vectors comprising of member genes for corresponding pathways

Usage

data(pathwayList)

Examples

data(pathwayList)
head(pathwayList)

Description

Computes variety of predictor evaluation measures based on the confusion matrix

Usage

perfCalc(dat)

Arguments

Value

(list) stats (data.frame): score, f1, ppv, precision and recall. One row per cutoff for feature selection auc (numeric between 0 and 1): AUC of overall ROC curve prauc (numeric between 0 and 1): AUC of overall precision-recall curve

Examples

data(confmat)
x <- perfCalc(confmat)

Description

data.frame with patient ID (ID), sample type (Type), tumour subtype (STATUS). From TCGA 2012 breast cancer paper (see reference).

Usage

data(pheno)

Source

The Cancer Genome Atlas. (2012). Nature 490:61-70.

References

The Cancer Genome Atlas. (2012). Nature 490:61-70.

Examples

data(pheno)
head(pheno)

Description

Patient ID and tumour status in "pheno", subsample of gene expression in "xpr" and CNV data in "cnv_GR"

Usage

data(pheno_full)

Source

The Cancer Genome Atlas. (2012). Nature 490:61-70.

References

The Cancer Genome Atlas. (2012). Nature 490:61-70.

Examples

data(pheno_full)
head(pheno_full)

Description

Create a network where nodes are predictive pathways passing certain cutoff and edges indicate similarity in gene-sets. Pathways are then clustered to identify themes of predictive pathways. Generates one such network for each patient label.

Usage

plotEmap(
  gmtFile,
  nodeAttrFile,
  netName = "generic",
  scoreCol = "maxScore",
  minScore = 1,
  maxScore = 10,
  nodeFillStops = c(7, 9),
  colorScheme = "cont_heatmap",
  imageFormat = "png",
  verbose = FALSE,
  createStyle = TRUE,
  groupClusters = FALSE,
  hideNodeLabels = FALSE
)

Arguments

Value

No value. Side effect of plotting the EnrichmentMap in an open session of Cytoscape.

Examples

#refer to getEMapInput_many.R for working getEMapInput_many() example
data(featScores)
pathwayList <- readPathways(fetchPathwayDefinitions("October",2020))
pathwayList <- pathwayList[seq_len(5)]
netInfoFile <- system.file("extdata",
	paste("example_output","inputNets.txt",sep=getFileSep()),
	package="netDx")
netTypes <- read.delim(netInfoFile,sep='\t',header=FALSE,as.is=TRUE)
outDir <- paste(tempdir(),'plots',sep=getFileSep())
if (!file.exists(outDir)) dir.create(outDir)
EMap_input <- getEMapInput_many(featScores,pathwayList,
     netTypes,outDir=outDir)
outDir <- paste(getwd(),'plots',sep=getFileSep())
if (!file.exists(outDir)) dir.create(outDir)
gmtFile <- EMap_input[[1]][1]
nodeAttrFile <- EMap_input[[1]][2]

# not run because requires Cytoscape to be installed and open
# plotEmap(gmtFile = gmtFile, nodeAttrFile = nodeAttrFile, 
#\t\tnetName='HighRisk')

Description

Visualize integrated patient similarity network based on selected features

Usage

plotIntegratedPatientNetwork(
  dataList,
  groupList,
  makeNetFunc = NULL,
  sims = NULL,
  setName = "predictor",
  prune_pctX = 0.05,
  prune_useTop = TRUE,
  aggFun = "MAX",
  calcShortestPath = FALSE,
  showStats = FALSE,
  outDir = tempdir(),
  numCores = 1L,
  nodeSize = 50L,
  edgeTransparency = 40L,
  nodeTransparency = 155L,
  plotCytoscape = FALSE,
  verbose = FALSE
)

Arguments

Details

Generates a Cytoscape network where nodes are patients and edges are weighted by aggregate pairwise patient similarity. Integrated PSN plotting is intended to run after feature selection, which identifies the subset of input networks predictive for each class of interest. The method of generating the network is as follows: All networks feature-selected in either patient ategory are concatenated; where a network is feature-selected in both categories, it is included once. The similarity between two patients in the integrated network is the mean of corresponding pairwise similarities. Dissimilarity is defined as 1-similarity, and Dijkstra distances are computed on this resulting network. For visualization, only edges representing the top fraction of distances (strongest edge weights) are included.

Value

(list) information about the integrated network similarity network 2) patientDistNetwork_pruned (matrix) the network plotted in Cytoscape. Also note that this is a dissimilarity network, so that more similar nodes have smaller edge weights 3) colLegend (data.frame): legend for the patient network plotted in Cytoscape. Columns are node labels (STATUS) and colours (colour) 6) outDir (char) value of outDir parameter

Description

Plots various measures of predictor performance for binary classifiers

Usage

plotPerf(resList = NULL, inFiles, predClasses, plotSEM = FALSE)

Arguments

Details

Plots individual and average ROC/PR curves. mean+/-SEM performance for a predictor run using nested cross-validation or a similar repeated design. predictionResults.txt contains a (data.frame)

Value

(list) each key corresponds to an input file in inDir. Value is a list with: 1) stats: 'stats' component of perfCalc 2) rocCurve: ROCR performance object for ROC curve 3) prCurve: ROCR performance object for PR curve 4) auroc: Area under ROC curve 5) aupr: Area under PR curve 6) accuracy: Accuracy

Side effect of plotting in a 2x2 format: 1) mean+/-SEM or (mean+/-SD) AUROC 2) mean+/-SEM or (mean+/-SD) AUPR 3) ROC curve for all runs plus average 4) PR curve for all runs plus average

Examples

inDir <- system.file("extdata","example_output",package='netDx')
inFiles <- paste(rep(inDir,3), sprintf("rng%i",seq_len(3)),"predictionResults.txt",
 sep=getFileSep())
resList <- list()
for (k in seq_len(length(inFiles))) {
	resList[[k]] <- read.delim(inFiles[k],sep="\t",header=TRUE,as.is=TRUE)
}
plotPerf(resList, predClasses = c('LumA','notLumA'))

Description

Plots a set of ROC/PR curves with average.

Usage

plotPerf_multi(
  inList,
  plotTitle = "performance",
  plotType = "ROC",
  xlab = "TPR",
  ylab = "FPR",
  meanCol = "darkblue",
  xlim = c(0, 1),
  ylim = c(0, 1)
)

Arguments

Details

Plots average curves with individual curves imposed.

Value

No value. Side effect of plotting ROC and PR curves

Examples

inDir <- system.file("extdata","example_output",package="netDx")
all_rng <- list.files(path = inDir, pattern = 'rng.')
fList <- paste(inDir,all_rng,'predictionResults.txt',sep=getFileSep())
rocList <- list()
for (k in seq_len(length(fList))) {
  dat <- read.delim(fList[1],sep='\t',header=TRUE,as.is=TRUE)
  predClasses <- c('LumA', 'notLumA')
  pred_col1 <- sprintf('%s_SCORE',predClasses[1])
  pred_col2 <- sprintf('%s_SCORE',predClasses[2])
  idx1 <- which(colnames(dat) == pred_col1)
  idx2 <- which(colnames(dat) == pred_col2)
 pred <- ROCR::prediction(dat[,idx1]-dat[,idx2], 
		dat$STATUS==predClasses[1])
 rocList[[k]] <- ROCR::performance(pred,'tpr','fpr')
}
plotPerf_multi(rocList,'ROC')

Description

Once a model is trained, this function is used to classify new patients using selected features

Usage

predict(
  trainMAE,
  testMAE,
  groupList,
  selectedFeatures,
  makeNetFunc = NULL,
  sims = NULL,
  outDir,
  verbose = FALSE,
  numCores = 1L,
  JavaMemory = 4L,
  debugMode = FALSE
)

Arguments

Value

(data.frame) predicted patient similarities and labels columns are: 1) ID, 2) STATUS (ground truth), 3) <label>_SCORE: similarity score for the corresponding label, 4) PRED_CLASS: predicted class

Description

assign patient class when ranked by multiple GM predictors

Usage

predictPatientLabels(resSet, verbose = TRUE)

Arguments

Value

data.frame: ID, similarityScore, PRED_CLASS

Examples

data(predRes); predClass <- predictPatientLabels(predRes)

Description

List of lists. First level is a list of size 4, with one key entry for each tumour type in example medulloblastoma dataset (WNT,SHH,Group3,Group4). Each list in the second level is of length 8, with structure corresponding to the output of getPatientRankings().

Usage

data(predRes)

Examples

data(predRes)
summary(predRes)
summary(predRes[[1]])

Description

Prune network by retaining strongest edges

Usage

pruneNet(net, vertices, pctX = 0.1, useTop = TRUE)

Arguments

Value

(data.frame) pruned network. Three columns: AliasA, AliasB, and weight

Description

Prune network by retaining strongest edges

Usage

pruneNet_pctX(net, vertices, pctX = 0.1, useTop = TRUE)

Arguments

Value

(data.frame) pruned network. Three columns: AliasA, AliasB, and weight

Description

Prune interaction networks to keep only the networks and patients requested

Usage

pruneNets(
  oldDir,
  newDir = tempdir(),
  filterNets = "*",
  filterIDs = "*",
  netSfx = "_cont.txt$",
  verbose = TRUE
)

Arguments

Details

This function is crucial for patient data that is highly sparse; examples include patient CNVs indels, as opposed to full matrix measures (gene expression, questionnaire data). Each step where the pool of patients is subset - e.g. limiting feature selection only to patients in training set - changes the set of networks that are eligible. Some networks may only contain test patients, while others may contain a single edge between a training and a test patient. Upon subsetting, such networks are no longer eligible for downstream use, such as feature selection. This function rewrites those subnetworks of the original networks that consist of eligible patients.

Value

(no value). Side effect of writing pruned network files to newDir

Examples

data(npheno)
netDir <- system.file("extdata","example_nets",package='netDx')
pruneNets(netDir,tempdir(),filterIDs=npheno[seq_len(10),],
 netSfx='txt$')

Description

make PSN for built-in similarity functions

Usage

psn__builtIn(settings, verbose, ...)

Arguments

Value

(char) names of networks created. Side effect of network creation.

Description

wrapper for PSNs using Pearson correlation

Usage

psn__corr(settings, verbose, ...)

Arguments

Value

(char) names of networks created. Side effect of network creation.

Description

make PSN for custom similarity functions

Usage

psn__custom(settings, fn, verbose, ...)

Arguments

Value

(char) names of networks created. Side effect of network creation.

Description

Generate random alphanumerical string of length 10

Usage

randAlphanumString(numStrings = 1L)

Arguments

Details

Used to create multiple temporary directories during an R session

Value

vector of length n, each with 10-char alphanumerical strings

Examples

randAlphanumString()

Description

Parse GMT file and return pathways as list

Usage

readPathways(
  fname,
  MIN_SIZE = 10L,
  MAX_SIZE = 200L,
  EXCLUDE_KEGG = TRUE,
  IDasName = FALSE,
  verbose = TRUE,
  getOrigNames = FALSE
)

Arguments

Details

The GMT file format currently supported should match the ones found at http://downloads.baderlab.org. The original GMT file format is: <set name><set description><member 1><member 2>...<member N>, one row per set, with values tab-delimited. The version at baderlab.org has additional unique formatting of the <set name> column as follows: <pathway_full_name> This function requires the specific formatting of the first column to assign the key name of the output list (see useIDasName argument).

Value

Depends on value of getOrigNames. If FALSE (Default), list with pathway name as key, vector of genes as value. If TRUE, returns list of length two, (1) geneSets: pathway-gene mappings as default, (2) pNames: data.frame with original and cleaned names.

Examples

pathFile <- fetchPathwayDefinitions("October",2020)
pathwayList    <- readPathways(pathFile)

Description

find/replace pattern in all files of specified file type in specified directory. Needed to modify number format when intefacing with GeneMANIA, on French locale machines. Without this step, CacheBuilder throws error with commas.

Usage

replacePattern(pattern = ",", target = ".", path = getwd(), fileType = "txt$")

Arguments

Value

No value. Files have patterns replaced in place.

Description

Computes positive and negative calls upon changing stringency of feature selected networks (binary networks only)

Usage

RR_featureTally(
  netmat,
  phenoDF,
  TT_STATUS,
  predClass,
  pScore,
  outDir = tempdir(),
  enrichLabels = TRUE,
  enrichedNets,
  maxScore = 30L,
  verbose = FALSE
)

Arguments

Details

This function computes predictor performance in the context of binary networks, where + and - calls are based on membership (or lack thereof) in feature selected networks. An example would be networks based on CNV occurrence in cellular pathways; in this use case, a + is based on patient membership in feature-selected networks. This function takes the output data from a feature selection exercise and computes the number and fraction of positive and negative calls at each level of feature selection stringency. The output of this function can then be used to compute performance measures such as the ROC or precision-recall curve.

Value

(list) 1) cumulativeFeatScores: pathway name, cumulative score over N-way data resampling. 2) performance_denAllNets: positive,negative calls at each cutoff: network score cutoff (score); num networks at cutoff (numPathways) ; total +, ground truth (pred_tot); + calls (pred_ol); + calls as pct of total (pred_pct); total -, ground truth (other_tot) ; - calls (other_ol) ; - calls as pct of total (other_pct) ; ratio of pred_pct and other_pct (rr) ; min. pred_pct in all resamplings (pred_pct_min) ; max pred_pct in all resamplings (pred_pct_max) ; min other_pct in all resamplings (other_pct_min); max other_pct in all resamplings (other_pct_max) 3) performance_denEnrichedNets: positive, negative calls at each cutoff label enrichment option: format same as performance_denAllNets. However, the denominator here is limited to patients present in networks that pass label enrichment 4) resamplingPerformance: breakdown of performance for each of the resamplings, at each of the cutoffs. This is a list of length 2, one for allNets and one for enrichedNets. The value is a matrix with (resamp * 7) columns and S rows, one row per score. The columns contain the following information per resampling: 1) pred_total: total num patients of predClass 2) pred_OL: num of pred_total with a CNV in the selected net 3) pred_OL_pct: 2) divided by 1) (percent) 4) other_total: total num patients of other class(non-predClass) 5) other_OL: num of other_total with CNV in selected net 6) other_OL_pct: 5) divided by 4) (percent) 7) relEnr: 6) divided by 3).

Examples

data(cnv_patientNetCount) # patient presence/absence in nets
data(cnv_pheno)		# patient ID, label
data(cnv_netScores)	# network scores for resampling
data(cnv_TTstatus)	# train/test status
data(cnv_netPass) 	# nets passing label enrichment

d <- tempdir()
out <- RR_featureTally(cnv_patientNetCount,
		cnv_pheno,cnv_TTstatus,"case",cnv_netScores,
		outDir=d,enrichLabels=TRUE,enrichedNets=cnv_netPass,
		maxScore=30L)
print(summary(out))

Description

Run GeneMANIA cross-validation with a provided subset of networks

Usage

runFeatureSelection(
  trainID_pred,
  outDir,
  dbPath,
  numTrainSamps = NULL,
  incNets = "all",
  orgName = "predictor",
  fileSfx = "CV",
  verbose = FALSE,
  numCores = 2L,
  JavaMemory = 6L,
  verbose_runQuery = FALSE,
  debugMode = FALSE,
  ...
)

Arguments

Details

Creates query files, runs GM for 10-fold cross validation.

Value

No value. Side effect of generating feature scores.

Examples

data(MB.pheno)
dbPath <- system.file("extdata","dbPath",package="netDx")
runFeatureSelection(MB.pheno$ID[which(MB.pheno$STATUS%in% 'WNT')],
		tempdir(),dbPath,103L)

Description

Run a query

Usage

runQuery(
  dbPath,
  queryFiles,
  resDir,
  verbose = TRUE,
  JavaMemory = 6L,
  numCores = 1L,
  debugMode = FALSE
)

Arguments

Value

(char) path to GeneMANIA query result files with patient similarity rankings (*PRANK) and feature weights (*NRANK) of results file

Examples

dbPath <- system.file("extdata","dbPath",package="netDx")
queryFile <- system.file("extdata","GM_query.txt",package="netDx")
runQuery(dbPath, queryFile,tempdir())

Description

Creates all the input files for the collection of features used in feature selection.

Usage

setupFeatureDB(pheno, prepDir = tempdir())

Arguments

Value

(data.frame) internal numerical id for patients (INTERNAL_ID) and user-provided ID (ID)

Examples

data(xpr,pheno)
pathwayList <- list(pathA=rownames(xpr)[1:10],pathB=rownames(xpr)[21:50])

dataList <- list(rna=xpr)  #only one layer type
groupList <- list(rna=pathwayList) # group genes by pathways

makeNets <- function(dataList, groupList, netDir,...) {
    netList <- makePSN_NamedMatrix(dataList[['rna']],
		rownames(dataList[['rna']]),
     groupList[['rna']],netDir,verbose=FALSE,
			writeProfiles=TRUE,...)
    unlist(netList)
}
tmpDir <- tempdir(); netDir <- paste(tmpDir,"nets",sep=getFileSep())
dir.create(netDir,recursive=TRUE)

pheno_id <- setupFeatureDB(pheno,netDir)

Description

List with two entries. net: Network specification. "X" and "Y" are source and target columns respectively. "DIST" specifies weights. groups: Node labsls. A data.frame with columns "ID" and "GROUP"

Usage

data(silh)

Examples

data(silh)
summary(silh)
silh$net
silh$groups

Description

Computes Euclidean distance between patients. A scaled exponential similarity kernel is used to determine edge weight. The exponential scaling considers the K nearest neighbours, so that similarities between non-neighbours is set to zero. Alpha is a hyperparameterthat determines decay rate of the exponential. For details, see Wang et al. (2014). Nature Methods 11:333.

Usage

sim.eucscale(dat, K = 20, alpha = 0.5)

Arguments

Value

symmetric matrix of size ncol(dat) (number of patients) containing pairwise patient similarities

Examples

data(xpr)
sim <- sim.eucscale(xpr)

Description

Computes Pearson correlation between patients. A scaled exponential similarity kernel is used to determine edge weight. The exponential scaling considers the K nearest neighbours, so that similarities between non-neighbours is set to zero. Alpha is a hyperparameter that determines decay rate of the exponential. For details see Wang et al. (2014). Nature Methods 11:333.

Usage

sim.pearscale(dat, K = 20, alpha = 0.5)

Arguments

Value

symmetric matrix of size ncol(dat) (number of patients) containing pairwise patient similarities

Examples

data(xpr)
sim <- sim.pearscale(xpr)

Description

simple capitalization

Usage

simpleCap(x)

Arguments

Details

used to format feature names so they are not in all-caps

Value

(char) Changes case so start of each word is in upper-case, and the rest is in lowercase

Examples

simpleCap('this IS a TEST sEnTenCe')

Description

This function applies the random walk with restart propagation algorithm to a matrix of patients profiles

Usage

smoothMutations_LabelProp(mat, net, numCores = 1L)

Arguments

Details

A network is an undirected graph G defined by a set of nodes corresponding to genes, and edges connecting nodes with an experimental evidence of interaction. A priori nodes are genes for which an information is known. A novel node is a candidate for being associated to the nodes above based on their information. A node prediction task leads to detect novel nodes and propagation techniques are largely applied for the purpose. Network-based propagation algorithms for node prediction transfer the information from a priori nodes to any other node in a network. Each node gets an imputation value which assesses how much information got. The prediction is based on the guilty-by-association principle. A node with a high imputation value has a high probability to be associated to a priori nodes. E.g. in a house where room A has one heater, if room B is the second hottest room it means that B is close to A and that there is a high probability that they share a door or wall. These algorithms exploit the global topology of the network. However, when they are applied to detect if unknown nodes are functionally associated to known ones, they may suffer of a drawback depending by the context. In biology, two functionally related fragments interact physically (direct interaction) or interact indirectly thanks to one or very few mediators. Therefore, exploring too far similarities between nodes can introduce noise in the prediction. We apply a random walk with restart propagation algorithm which resolution is set to 0.2 for giving high values only to the close neighbours of the a priori nodes.

Value

(data.frame) Continuous matrix of patient profiles in which each gene has the final propagation score

Examples

suppressWarnings(suppressMessages(require(MultiAssayExperiment)))
require(doParallel)

# load mutation and phenotype data
genoFile <- system.file("extdata","TGCT_mutSmooth_geno.txt",package="netDx")
geno <- read.delim(genoFile,sep="\t",header=TRUE,as.is=TRUE)
phenoFile <- system.file("extdata", "TGCT_mutSmooth_pheno.txt",
			package="netDx")
pheno <- read.delim(phenoFile,sep="\t",header=TRUE,as.is=TRUE)
rownames(pheno) <- pheno$ID

# load interaction nets to smooth over
require(BiocFileCache)
netFileURL <- paste("https://download.baderlab.org/netDx/",
	"supporting_data/CancerNets.txt",sep="")
cache <- rappdirs::user_cache_dir(appname = "netDx")
bfc <- BiocFileCache::BiocFileCache(cache,ask=FALSE)
netFile <- bfcrpath(bfc,netFileURL)
cancerNets <- read.delim(netFile,sep="\t",header=TRUE,as.is=TRUE)
# smooth mutations
prop_net <- smoothMutations_LabelProp(geno,cancerNets,numCores=1L)