PredictLogReg {MSclassifR}

Prediction of the category to which a mass spectrum belongs from a multinomial logistic regression model

Description

This function predicts the category (species, phenotypes...) to which a mass spectrum belongs from a set of shortlisted mass-over-charge values of interest and a short-listed multinomial logistic regression model (see LogReg).

Usage

PredictLogReg(peaks,
              model,
              moz,
              tolerance = 6,
              toleranceStep = 2,
              normalizeFun = TRUE,
              noMatch=0,
              Reference = NULL)

Arguments

peaks

a list of MassPeaks objects (see MALDIquant R package).

model

a model or a list of models estimated from a set of shortlisted mass-over-charge values (output of the LogReg function).

moz

a vector with the set of shortlisted mass-over-charge values used to estimate the model Model.

tolerance

a numeric value of accepted tolerance to match peaks to the set of shortlisted mass-over-charge values. It is fixed to 6 Da by default.

toleranceStep

a numeric value added to the tolerance parameter to match peaks to the set of shortlisted mass-over-charge values. It is fixed to 2 Da by default.

normalizeFun

a logical value, if TRUE (default) the maximum intensity will be equal to 1, the other intensities will be expressed in ratio to this maximum.

noMatch

a numeric value used to replace intensity values if there is no match detected between peaks and the set of shortlisted mass-over-charge values moz. It is fixed to 0 by default.

Reference

a factor with a length equal to the number of rows in X and containing the categories of each mass spectrum in X. "NULL" by default.

Details

The PredictLogReg function allows predicting the membership of a mass spectrum to a category from a multinomial regression model. The mass spectrum from the peaks object will be matched to the discriminant mass-over-chage (m/z) values (sel_moz object from the SelectionVar or SelectionVarStat functions) with a tolerance between 2 m/z and defined by the tolerance parameter (by default this value is 6 Da). If a repetition of a same m/z occurs in the selection, only the m/z that is closest in mass peaks (moz) is used. When no match, intensity values are replaced by the noMatch argument. If no m/z values from peaks object matched with the m/z in the object moz, the tolerance will be increased according to a numeric value defined in the toleranceStep parameter and a warning will be notified. Note that it is possible to not perform the SelectionVar function prior to the PredictLogReg function, and to replace the argument moz by all the m/z values present in a mass spectrum.

Value

Returns a dataframe containing probabilities of membership by category for each mass spectrum in peaks. The method used is provided in the method column. The comb_fisher method is the result of the Fisher's method when merging probabilities of membership of used prediction models.The max_vote method is the result of the maximum voting from used prediction models.

If the Reference parameter is not null, the function returns:

Confusion.Matrix

a list of confusion matrix (cross-tabulation with associated statitics) corresponding to the output of the confusionMatrix function of the caret R package.

Gobal.stat

a data.frame with three columns corresponding to the value (value column) of a statistic parameter (Statistic.parameter column) from a method used (model column) obtained with the LogReg function. See LogReg function for the Statistic.parameter column.

Details.stat

a data.frame with four columns corresponding to the same as Gobal.stat dataframe with the class concerned for estimated statistic parameter (class column). All statistic parameters are extracted from the output of the confusionMatrix function of the caret R package.

Correct.ClassificationFreq

a data.frame with predicted class (Prediction column) from a method (Model column) and the reference of the categories of each mass spectrum (Reference column). The Freq column indicates the number of times the category was correctly predicted by the method.

Incorrect.ClassificationFreq

a data.frame with predicted class (Prediction column) from a method (Model column) and the reference of the categories of each mass spectrum (Reference column). The Freq column indicates the number of times the category was not correctly predicted by the method.

References

Kuhn, M. (2008). Building predictive models in R using the caret package. Journal of statistical software, 28(1), 1-26.

Examples

library("MSclassifR")
library("MALDIquant")

###############################################################################
## 1. Pre-processing of mass spectra

# load mass spectra and their metadata
data("CitrobacterRKIspectra","CitrobacterRKImetadata", package = "MSclassifR")
# standard pre-processing of mass spectra
spectra <- SignalProcessing(CitrobacterRKIspectra)
# detection of peaks in pre-processed mass spectra
peaks <- MSclassifR::PeakDetection(x = spectra, averageMassSpec=FALSE)
# matrix with intensities of peaks arranged in rows (each column is a mass-over-charge value)
IntMat <- MALDIquant::intensityMatrix(peaks)
rownames(IntMat) <- paste(CitrobacterRKImetadata$Strain_name_spot)
# remove missing values in the matrix
IntMat[is.na(IntMat)] <- 0
# normalize peaks according to the maximum intensity value for each mass spectrum
IntMat <- apply(IntMat,1,function(x) x/(max(x)))
# transpose the matrix for statistical analysis
X <- t(IntMat)
# define the known categories of mass spectra for the classification
Y <- factor(CitrobacterRKImetadata$Species)

###############################################################################
## 2. Selection of discriminant mass-over-charge values using RFERF
# with 5 to 10 variables,
# without sampling method and trained
# with the Accuracy coefficient metric

a <- MSclassifR::SelectionVar(X,
                              Y,
                              MethodSelection = c("RFERF"),
                              MethodValidation = c("cv"),
                              PreProcessing = c("center","scale","nzv","corr"),
                              NumberCV = 2,
                              Metric = "Kappa",
                              Sizes = c(5:10))

sel_moz=a$sel_moz

###############################################################################
## 3. Perform LogReg from shortlisted discriminant mass-over-charge values

# linear multinomial regression
# without sampling mehod and
# trained with the Kappa coefficient metric

model_lm=MSclassifR::LogReg(X=X,
                            moz=sel_moz,
                            Y=factor(Y),
                            number=2,
                            repeats=2,
                            Metric = "Kappa")
# Estimated model:
model_lm

# nonlinear multinomial regression using neural networks
# with up-sampling method and
# trained with the Kappa coefficient metric

model_nn=MSclassifR::LogReg(X=X,
                            moz=sel_moz,
                            Y=factor(Y),
                            number=2,
                            repeats=2,
                            kind="nnet",
                            Metric = "Kappa",
                            Sampling = "up")
# Estimated model:
model_nn

# nonlinear multinomial regression using random forests
# without down-sampling method and
# trained with the Kappa coefficient metric

model_rf=MSclassifR::LogReg(X=X,
                            moz=sel_moz,
                            Y=factor(Y),
                            number=2,
                            repeats=2,
                            kind="rf",
                            Metric = "Kappa",
                            Sampling = "down")

# Estimated model:
model_rf

# nonlinear multinomial regression using xgboost
# with down-sampling method and
# trained with the Kappa coefficient metric

model_xgb=MSclassifR::LogReg(X=X,
                             moz=sel_moz,
                             Y=factor(Y),
                             number=2,
                             repeats=2,
                             kind="xgb",
                             Metric = "Kappa",
                             Sampling = "down")
# Estimated model:
model_xgb

# nonlinear multinomial regression using svm
# with down-sampling method and
# trained with the Kappa coefficient metric

model_svm=MSclassifR::LogReg(X=X,
                             moz=sel_moz,
                             Y=factor(Y),
                             number=2,
                             repeats=2,
                             kind="svm",
                             Metric = "Kappa",
                             Sampling = "down")
# Estimated model:
model_svm

# Of note, you can also load a model already saved
# (see example in LogReg function) for the next step
###############################################################################
## 4. Probabilities of belonging to each category for the mass spectra
## and associated statitics

# Collect all the estimated models in a list

Models <- list(model_lm$train_mod,
               model_nn$train_mod,
               model_rf$train_mod,
               model_xgb$train_mod,
               model_svm$train_mod)

# Predict classes of mass spectra with 6 Da of tolerance for matching peaks.
prob_cat=MSclassifR::PredictLogReg(peaks = peaks[c(1:5)],
                                   model = Models,
                                   moz = sel_moz,
                                   tolerance = 6,
                                   Reference = Y[c(1:5)])

################################################################################
## 5. Example of meta-classifiers based on several random forest models
## to optimize a Kappa value using the SMOTE method for imbalanced datasets.
## -> a merge of the prediction probabilities using the Fisher's method
## leads generally to robust prediction models.

#Selecting peaks with mda method
a=SelectionVar(X,Y,MethodSelection="mda",Ntree=5*ncol(X))
sel_moz=a$sel_moz

#Building 4 Random Forest models
models=NULL;nbm=4;
for (i in 1:nbm){
  model_rf=MSclassifR::LogReg(X=X,
                             moz=sel_moz,
                             Y=factor(Y),
                             number=5,
                             repeats=5,
                             kind="rf",
                             Metric = "Kappa",
                             Sampling = "smote")
  models <- c(models,list(model_rf$train_mod))
}

#Combining their prediction probabilities
prob_cat=MSclassifR::PredictLogReg(peaks = peaks,model = models,moz = sel_moz,
tolerance = 6,Reference = Y)