Unsupervised machine learning using an imaging mass spectrometry dataset automatically reassembles grey and white matter

Abstract

Current histological and anatomical analysis techniques, including fluorescence in situ hybridisation, immunohistochemistry, immunofluorescence, immunoelectron microscopy and fluorescent fusion protein, have revealed great distribution diversity of mRNA and proteins in the brain. However, the distributional pattern of small biomolecules, such as lipids, remains unclear. To this end, we have developed and optimised imaging mass spectrometry (IMS), a combined technique incorporating mass spectrometry and microscopy, which is capable of comprehensively visualising biomolecule distribution. We demonstrated the differential distribution of phospholipids throughout the cell body and axon of neuronal cells using IMS analysis. In this study, we used solarix XR, a high mass resolution and highly sensitive MALDI-FT-ICR-MS capable of detecting higher number of molecules than conventional MALDI-TOF-MS instruments, to create a molecular distribution dataset. We examined the diversity of biomolecule distribution in rat brains using IMS and hypothesised that unsupervised machine learning reconstructs brain structures such as the grey and white matters. We have demonstrated that principal component analysis (PCA) can reassemble the grey and white matters without assigning brain anatomical regions. Hierarchical clustering allowed us to classify the 10 groups of observed molecules according to their distributions. Furthermore, the group of molecules specifically localised in the cerebellar cortex was estimated to be composed of phospholipids.

Figure 1

Schematic image of IMS data collection and analysis. The data collection has four steps: MALDI-IMS of the sagittal section of a rat brain, peak picking from the mass spectrum, screening of the distributions of biomolecules and the construction of an IMS dataset. Data analysis was undertaken by performing PCA to extract the principal distribution from the IMS dataset. Hierarchical clustering was used to classify molecules by their patterns of distribution.

Figure 2

Data collection from rat brain sagittal sections using MALDI-FT-ICR-IMS. (A) Average mass spectrum of all measured points. The x-axis corresponds to the mass range between m/z 700 to 900, whereas the y-axis shows relative signal intensity. a.u., arbitrary unit. (B) Lineup of all biomolecule distributions from a sagittal section of a rat brain in an IMS dataset, for which the number of distributions was 488. The signal intensity of each spot is shown using a rainbow scale.

Figure 3

PCA of the sagittal section of a rat brain. Rainbow scale, intensity distributions of (A) PC1, (B) PC2 and (C) PC3 were displayed using a rainbow colour scale. The red line indicates the border of the region of interest (ROI). The area of ROI surrounded the whole brain section. Black scale bar = 5 mm.

Figure 4

Hierarchical clustering of an IMS dataset. Heat maps showing (A) the relative values of cosine similarities in red colour gradient and (B) spots having the values larger than 0.5 in blue. Groups consisting of more than seven biomolecules on the diagonal line were marked in yellow (I–VIII) and those of out of the line in green. (C) Representative distributions of each group (I to X). The rainbow colour scale reflects the signal intensity of each sampling point. Scale bar = 5 mm.

Figure 5

The comparison of observed mass spectra and the calculated isotope pattern. The graphs showing mass spectra from m/z 834 to 837, m/z 856 to 859, and m/z 872 to 875 (black lines). The red lines indicated the calculated isotope patterns of [PS(39:0) + H]⁺, [PI(34:0) + NH4]⁺ and [PS(39:0) + K]⁺.

Figure 6

The molecular group distributed within the cerebellar cortex. Representative distributions of molecules belonging to group III. The rainbow colour scale reflects the signal intensity of each sampling point. Scale bar = 5 mm.