NECTAR: A New Algorithm for Characterizing and Correcting Noise in QToF-Mass Spectrometry Imaging Data

Abstract

A typical mass spectrometry imaging experiment yields a very high number of detected peaks, many of which are noise and thus unwanted. To select only peaks of interest, data preprocessing tasks are applied to raw data. A statistical study to characterize three types of noise in MSI QToF data (random, chemical, and background noise) is presented through NECTAR, a new NoisE CorrecTion AlgoRithm. Random noise is confirmed to be dominant at lower m/z values (∼50-400 Da) while systematic chemical noise dominates at higher m/z values (>400 Da). A statistical approach is presented to demonstrate that chemical noise can be corrected to reduce its presence by a factor of ∼3. Reducing this effect helps to determine a more reliable baseline in the spectrum and therefore a more reliable noise level. Peaks are classified according to their spatial S/N on the single ion images, and background noise is thus removed from the list of peaks of interest. This new algorithm was applied to MALDI and DESI QToF data generated from the analysis of a mouse pancreatic tissue section to demonstrate its applicability and ability to filter out these types of noise in a relevant data set. PCA and t-SNE multivariate analysis reviews of the top 4000 peaks and the final 744 and 299 denoised peak list for MALDI and DESI, respectively, suggests an effective removal of uninformative peaks and proper selection of relevant peaks.

Figure 1

Diagram of the different noise contributions found in the literature affecting the MSI data. Instrumental noise mostly affects the shape of the peaks, while random and chemical noise are direct noise contributions affecting the detection limit of the spectrum. Background noise refers to uninformative peaks, most likely related to the presence of matrix or solvent.

Figure 2

(A) H&E staining was performed post MALDI MSI analysis on same tissue section. The stained section was imaged at 40× magnification (0.226 μm/pixel) with the Aperio CS2 digital pathology scanner (Leica Biosystems) and visualized ImageScope software (Leica Biosystems version 12.3.2.8013). (B) Example of a compound that is related to the matrix and thus not relevant for the study of the biological tissue sample. (C) If no background is considered when selecting the peak, it could be wrongly selected at first. (D) An estimation of the S/N on background for the same number of pixels in tissue is performed to obtain the S/N ratio between tissue and background.

Figure 3

Characterization of the different types of noise in QToF-MSI instruments. Noise level is determined (red), and baseline correction is applied to the spectrum. Random noise is identified, and chemical noise is modeled (red) and subtracted from the original spectrum.

Figure 4

Noise contribution for 6 selected peaks in the MALDI spectrum. Random noise is dominant for peaks at lower m/z values, while systematic chemical noise is dominant for peaks at higher m/z values. Dashed lines correspond to data before correction, and solid lines correspond to data after correction (baseline and chemical correction).

Figure 5

Relative noise level in the MALDI data spectrum obtained with two different window sizes (4 and 20 Da) before and after applying the noise correction.

Figure 6

Histogram of the MALDI noise distribution before and after baseline and chemical noise correction at different areas of the spectrum: (A) full spectrum, (B) m/z 50–350, (C) m/z 400–1000. After final correction, a Gaussian distribution of noise is obtained (in red).

Figure 7

Classification of peaks according to their signal spatial distribution. The central figure represents the logarithmic mean intensity versus the fold change of the mean intensity of the tissue background on the single ion images. Triangles correspond to the NECTAR selected peaks, while crosses correspond to the top 4000 most intense peaks. Blue and orange symbols represent noise peaks (S/N_T – S/N_B < 0 and S/N_T/S/N_B < 5), while blue and green are peaks of interest (S/N_T – S/N_B > 0 and S/N_T/S/N_B > 5). Information related to the single ion images is given in Table 2. (A) Example of a detected very low intense peak. (B,C) Examples of background noise originated in the matrix. (D) Clear detection. (E) Artifact produced most likely during the sample preparation.

Figure 8

Multivariate analysis comparison of the top 4000 peaks and the NECTAR denoised list for the MALDI data set. A greater cumulative variance is obtained with the first 10 PCs for the denoised list than with the top 4000. T-SNE clustering demonstrates a proper removal of uninformative peaks and selection of relevant peaks.