Filter peaks based on noise modeling — filterNoisePeaks • Spectra

The filterNoisePeaks() cleans spectra removing noise peaks. The method to identify noise peaks can be selected with the method parameter. The currently supported method is:

method = "uniform_intensity": this approach assumes the intensity of noise peaks (e.g. due to electronic noise) to follow a uniform distribution while real fragment peaks having larger intensities deviating from this background distribution (Dalla Valle 2025). This method estimates noise and filters peaks using only the intensity values of a spectrum. The method was developed and tested for fragment spectra and might not be directly applicable to MS1 data. See also the details section for more information. The noise filter is only applied to spectra with an MS level matching parameter msLevel., a minimum number of peaks of minPeakCount, if at least 4 peaks are available to fit the background model, and at least one peak passes the filter. In all other cases the spectrum is returned unfiltered.

Usage

filterNoisePeaks(
  object,
  method = c("uniform_intensity"),
  minPeakCount = 30L,
  intensityQuantile = 0.75,
  msLevel. = 2L,
  ...
)

Arguments

object: A Spectra object.
method: character(1) with the noise filtering method. Currently only method = "uniform_intensity" (the default) is supported.
minPeakCount: integer(1) with the minimal required number of peaks of a spectrum to apply the filter.
intensityQuantile: numeric(1) defining the intensity quantile below which peak intensities are used to model noise (defaults to intensityQuantile = 0.75).
msLevel.: integer(1) defining the MS level of the spectra which should be filtered. Defaults to msLevel. = 2L.
...: Ignored.

Value

A Spectra object with filtered/cleaned spectra. For method = "uniform_intensity" the filter is only applied to spectra if:

the MS level of a spectrum matches parameter msLevel.
a spectrum has at least minPeakCount peaks
at least 4 data points (peaks) are available to fit the background model
at least one peak passes the filter

The original, unfiltered spectrum is returned otherwise.

Details

For method = "uniform_intensity": the method fits (robustly) a linear model explaining the peak intensity by the rank of the peak (in the intensity-ordered peak list) for peaks with an intensity below the intensityQuantile intensity quantile. Peaks are retained if their measured intensity is considerably larger than the expected intensity following the background distribution (i.e., larger than 3 times the standard deviation of the noise model residuals). True fragment peaks are thus required to have intensities larger than expected from the (uniform) background distribution. See examples below for a graphical illustration of the approach.

The implementation of this method based on the original function from (Dalla Valle 2025). Source code available here

References

Nicola Dalla Valle, Mar Garcia-Aloy, Peter Robatscher, Pietro Franceschi, Michael Oberhuber; Improving Spectral Similarity and Molecular Network Reliability through Noise Signal Filtering in MS/MS Spectra. Anal. Chem. 29 July 2025; 97 (29): 15873–15882. https://doi.org/10.1021/acs.analchem.5c02109

Author

Johannes Rainer, Nicola Dalla Valle

Examples


## Load an example data set with DDA MS2 spectra
library(MsDataHub)
s <- Spectra(MsDataHub::PestMix1_DDA.mzML()) |>
    filterMsLevel(2L)
#> see ?MsDataHub and browseVignettes('MsDataHub') for documentation
#> loading from cache

## Select two example MS2 spectra.
s <- s[c(1, 271)]

res <- filterNoisePeaks(s)

## The spectra **before** filtering
plotSpectra(s)


## The spectra **after** filtering
plotSpectra(res)

## Only the second spectrum, with a large number of (presumably noise) peaks
## was filtered.

########
## Step-wise explanation of the *uniform_intensity* approach
library(MASS)
set.seed(123)

## Create an example spectrum with 40 peaks with intensities following a
## uniform distribution and 4 peaks with higher intensities. m/z values
## are random.
x <- cbind(
    mz = abs(rnorm(44, mean = 20, sd = 10)),
    intensity = c(runif(40, min = 10, max = 20), 28, 87, 59, 49))

## Plot the example spectrum
plot(x[, "mz"], x[, "intensity"], type = "h",
    xlab = "m/z", ylab = "intensity")

## The individual steps of the *uniform_intensity* cleaning approach:
i <- log10(x[, "intensity"])

## - identify the peaks potentially representing noise peaks:
n <- which(i < quantile(i, 0.75))
o <- rank(i)

## - fit a (robust) model explaining intensity by intensity rank
m <- rlm(y ~ x, list(x = o[n], y = i[n]))

## - predict intensity values based on this background model
p <- predict(m, list(x = o, y = i))

## - select the peaks to keep: intensity more than 3 sd higher than expected
keep <- (i - p) > sd(residuals(m)) * 3
sum(keep)
#> [1] 5

## Visualizing the data and filtering:
## - plot the intensities by their rank
plot(o, i, xlab = "order", ylab = expression(log[10]~intensity))

## - highlight the points on which the background model if fitted
points(o[n], i[n], pch = 16)

## - indicate the background model
abline(m)

## - label the peaks to keep (above background noise)
points(o[keep], i[keep], pch = 16, col = "blue", cex = 2)


## Repeating for a spectrum with only intensities from a uniform distribution
x <- cbind(
    mz = abs(rnorm(44, mean = 20, sd = 10)),
    intensity = c(runif(44, min = 10, max = 30)))

i <- log10(x[, "intensity"])
n <- which(i < quantile(i, 0.75))
o <- rank(i)
m <- rlm(y ~ x, list(x = o[n], y = i[n]))
p <- predict(m, list(x = o, y = i))
keep <- (i - p) > sd(residuals(m)) * 3
sum(keep)
#> [1] 0

plot(o, i, xlab = "order", ylab = expression(log[10]~intensity))
points(o[n], i[n], pch = 16)
abline(m)
points(o[keep], i[keep], pch = 16, col = "blue", cex = 2)