Split Training and Validation Data

To get started using PDATK, place each of your patient cohorts into SurvivalExperiment objects and then assemble those into a master CohortList, which holds the training and validation data for use with the various SurvivalModels in this package.

commonGenes <- findCommonGenes(sampleCohortList)
# Subsets all list items, with subset for specifying rows and select for
# specifying columns
cohortList <- subset(sampleCohortList, subset=commonGenes)

ICGCcohortList <- cohortList[grepl('ICGC', names(cohortList), ignore.case=TRUE)]
validationCohortList <- cohortList[!grepl('icgc', names(cohortList),
    ignore.case=TRUE)]

Since we are interested in predicting survival, it is necessary to remove patients with insufficient data to be useful. In general, we want to remove patients who did not have an event in our period of interest. As such we remove samples who dropped out of a study, but did not pass away before the first year.

validationCohortList <- dropNotCensored(validationCohortList)
ICGCcohortList <- dropNotCensored(ICGCcohortList)

We have now split our patient cohorts into training and validation data. For this analysis we will be training using the ICGC cohorts, which includes one cohort with RNA micro-array data and another with RNA sequencing data. When using multiple cohorts to train a model, it is required that those cohorts share samples. As a result we will take as training data all patients shared between the two cohorts and leave the remainder of patients as part of our validationData.

# find common samples between our training cohorts in a cohort list
commonSamples <- findCommonSamples(ICGCcohortList)

# split into shared samples for training, the rest for testing
ICGCtrainCohorts <- subset(ICGCcohortList, select=commonSamples)
ICGCtestCohorts <- subset(ICGCcohortList, select=commonSamples, invert=TRUE)

# merge our training cohort test data into the rest of the validation data
validationCohortList <- c(ICGCtestCohorts, validationCohortList)

# drop ICGCSEQ from the validation data, because it only has 7 patients
validationCohortList <- 
    validationCohortList[names(validationCohortList) != 'ICGCSEQ']

Setup A PCOSP Model Object

We now have patient molecular data, annotated with the number of days survived since treatment and the survival status and are ready to apply a SurvivalModel to this data. In this example, we are applying a Pancreatic Cancer Overall Survival Model, as described in . This class uses the switchBox package to create an ensemble of binary classifiers, whos votes are then tallied into a PCOSP score. A PCOSP score is simply the proportion of models predicting good survival out of the total number of models in the ensemble.

set.seed(1987)
PCOSPmodel <- PCOSP(ICGCtrainCohorts, minDaysSurvived=365, randomSeed=1987)

# view the model parameters; these make your model run reproducible
metadata(PCOSPmodel)$modelParams

## $randomSeed
## [1] 1987
## 
## $RNGkind
## [1] "Mersenne-Twister" "Inversion"        "Rejection"       
## 
## $minDaysSurvived
## [1] 365

Training a PCOSP Model

To simplify working with different SurvivalModel sub-classes, we have implemented a standard work-flow that is valid for all SurvivalModels. This involves first training the model, then using it to predict risk/risk-classes for a set of validation cohorts and finally assessing performance on the validation data.

To train a model, the trainModel method is used. This function abstracts away the implementation of model training, allowing end-users to focus on applying the SurvivalModel to make predictions without a need to understand the model internals. We hope this will make the package useful for those unfamiliar or uninterested in the details of survival prediction methods.

For training a PCOSP model there are two parameters. First, numModels is the number of models to train for use in the ensemble classifier to predict PCOSP scores. To keep computation brief, we are only training 25 models. However, it is recommended to use a minimum of 1000 for real world applications. The second parameter is minAccuracy, which is the minimum model accuracy for a trained model to included in the final model ensemble. Paradoxically, increasing this too high can actually decrease the overall performance of the PCOSP model. We recommend 0.6 as a sweet spot between random chance and over-fitting but encourage experimentation to see what works best with your data.

trainedPCOSPmodel <- trainModel(PCOSPmodel, numModels=15, minAccuracy=0.6)

metadata(trainedPCOSPmodel)$modelParams

## $randomSeed
## [1] 1987
## 
## $RNGkind
## [1] "Mersenne-Twister" "Inversion"        "Rejection"       
## 
## $minDaysSurvived
## [1] 365
## 
## $numModels
## [1] 15
## 
## $minAccuracy
## [1] 0.6

We can see that after training, the additional model parameters are added to the modelParams item in the model metadata. The goal is to ensure that your model training, prediction and validation are fully reproducible by capturing the parameters relevant to a specific model.

Risk Prediction with a PCOSP Model

After training, a model can now be used with new data to make risk predictions and classify samples into ‘good’ or ‘bad’ survival groups. To do this, the standard predictClasses method is used. Similar to trainData, we have abstracted away the implementation details to provide users with a simple, consistent interface for using SurvivalModel sub-classes to make patient risk predictions.

PCOSPpredValCohorts <- predictClasses(validationCohortList,
    model=trainedPCOSPmodel)

The returned CohortList object now indicates that each of the cohorts have predictions. This information is available in the elementMetadata slot of the cohort list and can be accessed with the mcols function from S4Vectors.

mcols(PCOSPpredValCohorts)

## DataFrame with 10 rows and 2 columns
##             mDataType hasPredictions
##           <character>      <logical>
## ICGCMICRO   rna_micro           TRUE
## PCSI          rna_seq           TRUE
## TCGA          rna_seq           TRUE
## KIRBY         rna_seq           TRUE
## UNC         rna_micro           TRUE
## CHEN        rna_micro           TRUE
## COLLISON    rna_micro           TRUE
## ZHANG       rna_micro           TRUE
## OUH         rna_micro           TRUE
## WINTER      rna_micro           TRUE

Predicting risk with a specific model adds a corresponding metadata column to the object colData. In the case of a PCOSP model, the new column is called PCOSP_prob_good and represents the proportion of models in the ensemble which predicted good survival for a given sample.

knitr::kable(head(colData(PCOSPpredValCohorts[[1]])))

Additionally, binary predictions of good or bad survival can be found in the PCOSPpredictions item of each SurvivalExperiments metadata. This contains the raw predictions from the model for each classifier in the ensemble, ordered by classifier accuracy. This data is not important for end users, but is used internally when calculating validation statistics for the model. For users wishing to classify samples rather than estimate risks, we recommend a PCOSP cut-off of >0.5 for good survival prognosis.

knitr::kable(metadata(PCOSPpredValCohorts[[1]])$PCOSPpredictions[1:5, 1:5])

Validating A PCOSP Model

The final step in the standard SurvivalModel work-flow is to compute model performance statistics for the model on the validation data. This can be accomplished using the validateModel method, which will add statistics to the validationStats slot of a SurvivalModel object and the data to the validationData slot.

validatedPCOSPmodel <- validateModel(trainedPCOSPmodel,
    valData=PCOSPpredValCohorts)

knitr::kable(head(validationStats(validatedPCOSPmodel)))

Examining the data.table from the validationStats slot we can see that three model performance statistics have been calculated for all of the validation cohorts. Additionally, aggregate statistics have been calculated by molecular data type and for all cohorts combined. This table can be used to generate model performance plots. We have included several functions for examining model performance.

Plotting Model Performance

PCOSPdIndexForestPlot <- forestPlot(validatedPCOSPmodel, stat='log_D_index')
PCOSPdIndexForestPlot

PCOSPconcIndexForestPlot <- forestPlot(validatedPCOSPmodel, stat='concordance_index')
PCOSPconcIndexForestPlot

cohortROCplots <- plotROC(validatedPCOSPmodel, alpha=0.05)

## Warning: The `<scale>` argument of `guides()` cannot be `FALSE`. Use "none" instead as
## of ggplot2 3.3.4.
## ℹ The deprecated feature was likely used in the PDATK package.
##   Please report the issue at <https://github.com/bhklab/PDATK/issues>.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.

cohortROCplots

Random Label Shuffling Model

The RLSModel class is a SurvivalModel using the same risk prediction algorithm as a PCOSP model, but randomizing the patient prognosis labels in each of the individual KTSP classifiers used in the classification ensemble. Given this random prognosis label shuffling, we expect the classification results for this model to be no better than random chance.

The work-flow for this model follows the standard SurvivalModel sub-class work-flow.

# Merge to a single SurvivalExperiment
ICGCtrainCohorts <- merge(ICGCtrainCohorts[[1]], ICGCtrainCohorts[[2]], 
    cohortNames=names(ICGCtrainCohorts))
RLSmodel <- RLSModel(ICGCtrainCohorts, minDaysSurvived=365, randomSeed=1987)

trainedRLSmodel <- trainModel(RLSmodel, numModels=15)

RLSpredCohortList <- predictClasses(validationCohortList, model=trainedRLSmodel)

## Warning in FUN(X[[i]], ...): 
## [PDATK::NULL] Dropped samples with NA survival data!
## Warning in FUN(X[[i]], ...): 
## [PDATK::NULL] Dropped samples with NA survival data!
## Warning in FUN(X[[i]], ...): 
## [PDATK::NULL] Dropped samples with NA survival data!

validatedRLSmodel <- validateModel(trainedRLSmodel, RLSpredCohortList)

RLSmodelComparisonPlot <- densityPlotModelComparison(validatedRLSmodel,
    validatedPCOSPmodel, title='Random Label Shuffling vs PCOSP',
    mDataTypeLabels=c(rna_seq='Sequencing-based', rna_micro='Array-based',
        combined='Overall'))
RLSmodelComparisonPlot

Random Gene Assignment Model

A RandomGeneAssignmentModel (aliased as RGAModel for user convenience) is similar to a RLSModel except that the gene labels are randomized for each KTSP classifier in the ensemble. In this case, we also expect this model to perform no better than random chance.

RGAmodel <- RGAModel(ICGCtrainCohorts, randomSeed=1987)

trainedRGAmodel <- trainModel(RGAmodel, numModels=15)

RGApredCohortList <- predictClasses(validationCohortList,
    model=trainedRGAmodel)

## Warning in FUN(X[[i]], ...): 
## [PDATK::NULL] Dropped samples with NA survival data!
## Warning in FUN(X[[i]], ...): 
## [PDATK::NULL] Dropped samples with NA survival data!
## Warning in FUN(X[[i]], ...): 
## [PDATK::NULL] Dropped samples with NA survival data!

validatedRGAmodel <- validateModel(trainedRGAmodel, RGApredCohortList)

RGAmodelComparisonPlot <-  densityPlotModelComparison(validatedRGAmodel,
    validatedPCOSPmodel, title='Random Gene Assignment vs PCOSP',
    mDataTypeLabels=c(rna_seq='Sequencing-based', rna_micro='Array-based',
        combined='Overall'))
RGAmodelComparisonPlot

Get the Top Predictive Features

We can get the features which are most predictive from a SurvivalModel object using the getTopFeatures method. By default this retrieves the gene names from the top 10% of models in the SurvivalModel. Users can customize the top N models to extract genes from using the numModels parameter.

topFeatures <- getTopFeatures(validatedPCOSPmodel)
topFeatures

##  [1] "ATOX1"    "NAMPT"    "KRT6A"    "FCGRT"    "SERTAD4"  "ZNF792"  
##  [7] "KRT6C"    "BIRC3"    "PKIB"     "SSBP3"    "PPFIBP2"  "FBLIM1"  
## [13] "ZSCAN29"  "KRT15"    "UBE2C"    "C16orf74" "BLNK"     "MARK3"

Querying Genesets for Enriched Pathways

To perform pathway analysis, which is done internally using the piano::runGSAhyper function we first to pathway data to query against. We recommend using the msigdbr R package to fetch pathways of interest.

Because of the small number of genes included in the sample patient cohorts, this section will not find any enriched pathways and therefore this code is not run.

allHumanGeneSets <- msigdbr()
allGeneSets <- as.data.table(allHumanGeneSets)
geneSets <- allGeneSets[grepl('^GO.*|.*CANONICAL.*|^HALLMARK.*', gs_name),
    .(gene_symbol, gs_name)]

GSEAresultDT <- runGSEA(validatedPCOSPmodel, geneSets)

Build the Model

We can see there are a number of clinical variables related to patient survived in the patient metadata. For our model, we will be using patient sex, age and tumour grade along with the TNM staging of the tumour.

knitr::kable(head(colData(ICGCtrainCohorts)))

clinicalModel <- ClinicalModel(ICGCtrainCohorts,
    formula='prognosis ~ sex + age + T + N + M + grade',
    randomSeed=1987)

Train the Model

trainedClinicalModel <- trainModel(clinicalModel)

Predict the Classes

Clinical annotations are only available for a subset of our patient cohorts, thus we subset our cohort list to ensure the model works as expected. Variable columns with levels of a model parameter which are not present in the original model will be dropped automatically. To prevent this behavior, it is necessary to add any missing levels to the colData of the training data before training the model. The easiest way to achieve this is by converting the columns in question into factors and ensuring all levels in the training and validation data are set for those columns.

hasModelParamsCohortList <-
    PCOSPpredValCohorts[c('ICGCMICRO', 'TCGA', 'PCSI', 'OUH')]

clinicalPredCohortList <- predictClasses(hasModelParamsCohortList,
    model=trainedClinicalModel)

## Warning in .local(object, model, ...): 
## [PDATK::NULL] Rows 120, 155 have levels that are not in the model, skipping these rows...

Validate the Model

validatedClinicalModel <- validateModel(trainedClinicalModel,
    clinicalPredCohortList)

Visualize Comparison with PCOSP Model

To get a general idea of how the ClinicalModel performed relative to the PCOSP model, the barPlotModelComaprison can be used to show the . For this example, we will compare the AUC of the models in each of the validation cohorts. Other statistics in the validationStats slot can be used by changing the stat argument to the plotting function.

clinicalVsPCOSPbarPlot <- barPlotModelComparison(validatedClinicalModel,
    validatedPCOSPmodel, stat='AUC')
clinicalVsPCOSPbarPlot

To reduce the number of plotting functions and simplify meta-analysis of many models of different types, we have included the ModelComparison object. This object aggregates the validation statistics from two models. Plotting methods can then be dispatched on this class, greatly simplifying the process of comparing several SurivalModel sub-classes. You can also compare a model to an existing model comparison, and in this way built up a collection of model performance statistics which can be visualized in a forest plot to enable complex comparisons between many models.

Below, we use this feature to compare the PCOSP and ClinicalModels based on D index and concordance index, as calculated using the survcomp R package.

clinicalVsPCOSP <- compareModels(validatedClinicalModel, validatedPCOSPmodel)

clinVsPCOSPdIndexForestPlot <- forestPlot(clinicalVsPCOSP, stat='log_D_index')
clinVsPCOSPdIndexForestPlot

clinVsPCOSPconcIndexForestPlot <- forestPlot(clinicalVsPCOSP,
    stat='concordance_index')
clinVsPCOSPconcIndexForestPlot

Make the Models

Since there is no training data for the published classifiers, we create empty GeneFuModel objects, then assign the model predictors to the models slot of each respective object.

chenGeneFuModel <- GeneFuModel(randomSeed=1987)
birnbaumGeneFuModel <- GeneFuModel(randomSeed=1987)
haiderGeneFuModel <- GeneFuModel(randomSeed=1987)

For the Haider model, we were unable to get the genes and coefficients. However, the author provided the risk scores. As a result we need to do a bit of work to get the Haider GeneFuModel to work with the package function.

data(existingClassifierData)

models(chenGeneFuModel) <- SimpleList(list(chen=chen))
models(birnbaumGeneFuModel) <- SimpleList(list(birnbuam=birnbaum))
models(haiderGeneFuModel) <- SimpleList(list(haider=NA)) 

Predict the Classes

chenClassPredictions <- predictClasses(PCOSPpredValCohorts[names(haiderSigScores)],
    model=chenGeneFuModel)
birnClassPredictions <- predictClasses(PCOSPpredValCohorts[names(haiderSigScores)],
    model=birnbaumGeneFuModel)

haiderClassPredictions <- PCOSPpredValCohorts[names(haiderSigScores)]
# Manually assign the scores to the prediction cohorts
for (i in seq_along(haiderClassPredictions)) {
  colMData <- colData(haiderClassPredictions[[i]])
  colMData$genefu_score <- NA_real_
  colMData[rownames(colMData) %in% names(haiderSigScores[[i]]), ]$genefu_score <- 
      haiderSigScores[[i]][names(haiderSigScores[[i]]) %in% rownames(colMData)]
  colData(haiderClassPredictions[[i]]) <- colMData
}
# Setup the correct model metadata
mcols(haiderClassPredictions)$hasPredictions <- TRUE
metadata(haiderClassPredictions)$predictionModel <- haiderGeneFuModel

Validate the Models

validatedChenModel <- validateModel(chenGeneFuModel, valData=chenClassPredictions)
validatedBirnbaumModel <- validateModel(birnbaumGeneFuModel, 
    valData=birnClassPredictions)
validatedHaiderModel <- validateModel(haiderGeneFuModel, valData=haiderClassPredictions)

Model Performance Meta-Analysis

genefuModelComparisons <- compareModels(validatedChenModel,
    validatedBirnbaumModel, modelNames=c('Chen', 'Birnbaum'))
genefuModelComparisons <- compareModels(genefuModelComparisons,
    validatedHaiderModel, model2Name='Haider')

allModelComparisons <- compareModels(genefuModelComparisons, validatedPCOSPmodel, 
  model2Name='PCOSP')
# We are only interested in comparing the summaries, so we subset our model comparison
allModelComparisons <- subset(allModelComparisons, isSummary == TRUE)

allDindexComparisonForestPlot <- forestPlot(allModelComparisons,
    stat='log_D_index', colourBy='model', groupBy='mDataType')
allDindexComparisonForestPlot

allConcIndexComparisonForestPlot <- forestPlot(allModelComparisons,
    stat='concordance_index', colourBy='model', groupBy='mDataType')
allConcIndexComparisonForestPlot

From the two forest plots, we can see that the PCOSP model matched our outperformed all of the public classifiers, even when only trained with 100 models. It is likely that using 1000 models, we would see even better separation of the PCOSP model. Indeed, this is was the result in Sandhu et al. (2019).