MSclassifR (version 0.5.0)

PredictLogReg: Prediction of the category to which a mass spectrum belongs

Description

Predicts the category (species, phenotype, etc.) of each spectrum in a list of MALDIquant MassPeaks using one or more trained models (e.g., multinomial logistic regression from LogReg). Peaks are matched to a given shortlist of discriminant m/z values (moz) within a tolerance; unmatched positions are filled with noMatch. If several models are supplied, the function also produces meta-predictions: per-class Fisher combinations (if 'metap' is available) and a majority-vote fraction across models.

Usage

PredictLogReg(
  peaks,
  model,
  moz,
  tolerance = 6,
  toleranceStep = 2,
  normalizeFun = TRUE,
  noMatch = 0,
  Reference = NULL,
  chunk_size = 10000L,
  ncores = 1L,
  verbose = FALSE
)

Value

If Reference is missing (or has < 2 levels), a data.frame with:

• name: spectrum name (from MassPeaks metaData fullName/file when available)

• method: model identifier (from model$method; suffixed with "_i" if needed)

• one column per class with predicted probabilities

• pred_max_p: predicted class (argmax of probabilities)

If Reference is provided with at least two levels, a list with:

• Prob.results: the predictions data.frame as above

• Confusion.Matrix: a list of caret::confusionMatrix objects (one per method)

Arguments

peaks

a list of MALDIquant::MassPeaks objects (one per spectrum).

model

a model or a list of models estimated from a shortlist of m/z (e.g., the output of LogReg). Each model must support predict(..., type = "prob"). If a single model is supplied, it is wrapped into a one-element list.

moz

a numeric vector of shortlisted m/z values used for prediction (typically the selection from SelectionVar/SelectionVarStat_fast).

tolerance

numeric; accepted m/z tolerance (in Da) for matching peaks to moz. Default 6.

toleranceStep

numeric; if a spectrum yields no matches at the initial tolerance, the function retries by increasing tolerance in steps of toleranceStep until at least one match is found (bounded internally). Default 2.

normalizeFun

logical; if TRUE (default), per-spectrum max normalization is applied after matching (row is divided by its maximum).

noMatch

numeric; intensity used when no peak matches a given m/z. Default 0.

Reference

optional factor of true categories, length equal to length(peaks). If provided and has at least two distinct levels, the function returns, in addition to the predictions, per-model confusion matrices (caret::confusionMatrix).

chunk_size

integer; number of spectra per prediction batch (rows of X). Large datasets can be processed in chunks to limit memory. Default 10000.

ncores

integer; number of cores used while building X from peaks on the R side (the C++ matching itself is single-threaded here). Default 1.

verbose

logical; print progress messages. Default FALSE.

Details

• Matching and normalization: peak-to-moz matching is performed by build_X_from_peaks_fast (C++-backed; nearest-within-tolerance). If no m/z from a spectrum match the shortlist initially, the tolerance is increased by toleranceStep in a small number of attempts until at least one match is found. If normalizeFun = TRUE, each row is divided by its maximum (guarded to avoid divide-by-zero).

• Multiple models: when several models are supplied, the output contains one set of probabilities per model (with method column identifying it). Two additional rows per spectrum can be appended:

  • comb_fisher: per-class Fisher combined p-values computed via metap::sumlog (if available).

  • max_vote: per-class fraction of models casting the top-probability vote for that class.

• Models: this function is agnostic of the modeling engine as long as predict(type = "prob") is implemented (e.g., caret multinom/nnet/ranger/xgb, glmnet, etc.).

References

Kuhn, M. (2008). Building predictive models in R using the caret package. Journal of Statistical Software, 28(1), 1–26.

Alexandre Godmer, Yahia Benzerara, Emmanuelle Varon, Nicolas Veziris, Karen Druart, Renaud Mozet, Mariette Matondo, Alexandra Aubry, Quentin Giai Gianetto (2025). MSclassifR: An R package for supervised classification of mass spectra with machine learning methods. Expert Systems with Applications, 294, 128796. doi:10.1016/j.eswa.2025.128796

See Also

LogReg; SelectionVar; SelectionVarStat_fast; build_X_from_peaks_fast

Examples

# \donttest{
library(MSclassifR)
library(MALDIquant)

## 1) Preprocess and detect peaks
data("CitrobacterRKIspectra", "CitrobacterRKImetadata", package = "MSclassifR")
spectra <- SignalProcessing(CitrobacterRKIspectra)
peaks   <- MSclassifR::PeakDetection(x = spectra, averageMassSpec = FALSE)

## 2) Build X and Y (sample-by-peak intensities + labels)
##    Option A: if you prefer the helper and a sparse return:
Y <- factor(CitrobacterRKImetadata$Species)
xy <- build_XY_from_peaks(peaks, labels = Y, normalize = "max", sparse = FALSE)
X <- xy$X
Y <- xy$Y

##    Option B: via MALDIquant::intensityMatrix (as in the original examples)
##IntMat <- MALDIquant::intensityMatrix(peaks)
##rownames(IntMat) <- paste(CitrobacterRKImetadata$Strain_name_spot)
##IntMat[is.na(IntMat)] <- 0
##IntMat <- t(apply(IntMat, 1, function(x) x / max(x)))  # per-spectrum max norm
##X <- t(IntMat)                                         # features in columns
##Y <- factor(CitrobacterRKImetadata$Species)

## 3) Select discriminant m/z with "cvp" method
a <- MSclassifR::SelectionVar(
  X, Y,
  MethodSelection = "cvp",
  MethodValidation = "cv",
  PreProcessing = c("center","scale","nzv","corr"),
  NumberCV = 2,
  Metric = "Kappa"
)
sel_moz <- a$sel_moz

## 4) Train several models on the shortlisted m/z
model_lm  <- MSclassifR::LogReg(X = X, moz = sel_moz, Y = Y, number = 2,
 repeats = 2, Metric = "Kappa", kind = "linear")
model_nn  <- MSclassifR::LogReg(X = X, moz = sel_moz, Y = Y, number = 2,
 repeats = 2, Metric = "Kappa", kind = "nnet", Sampling = "up")
model_rf  <- MSclassifR::LogReg(X = X, moz = sel_moz, Y = Y, number = 2,
 repeats = 2, Metric = "Kappa", kind = "rf",  Sampling = "down")
model_svm <- MSclassifR::LogReg(X = X, moz = sel_moz, Y = Y, number = 2,
 repeats = 2, Metric = "Kappa", kind = "svm", Sampling = "up")

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

## 5) Predict classes for a subset of peaks; 6 Da tolerance for matching
prob_cat <- MSclassifR::PredictLogReg(
  peaks     = peaks[1:5],
  model     = Models,
  moz       = sel_moz,
  tolerance = 6,
  Reference = Y[1:5]
)
prob_cat

## 6) Meta-classifier strategy (several RF models + SMOTE + Fisher combine)
a2 <- MSclassifR::SelectionVar(X, Y, MethodSelection = "mda", Ntree = 5 * ncol(X))
sel_moz2 <- a2$sel_moz
models2 <- vector("list", 4L)
for (i in seq_along(models2)) {
  models2[[i]] <- MSclassifR::LogReg(
    X = X, moz = sel_moz2, Y = Y,
    number = 5, repeats = 5,
    kind = "rf", Metric = "Kappa",
    Sampling = "smote"
  )$train_mod
}
prob_cat2 <- MSclassifR::PredictLogReg(
  peaks = peaks,
  model = models2,
  moz   = sel_moz2,
  tolerance = 6,
  Reference = Y
)
# }