3D DESI mass spectrometry imaging of 52 serial sections from a human colorectal adenocarcinoma (MTBLS415)

DOI: 10.1039/C6SC03738K
创建时间: 2025-01-10 12:39:05

Visual inspection of tumour tissues does not reveal the complex metabolic changes that differentiate cancer and its sub-types from healthy tissues. Mass spectrometry imaging, which quantifies the underlying chemistry, represents a powerful tool for the molecular exploration of tumour tissues. A 3-dimensional topological description of the chemical properties of the tumour permits the formulation of hypotheses about the biological composition and interactions and the possible causes of its heterogeneous structure. The large amount of information contained in such datasets requires powerful tools for its analysis, visualisation and interpretation. Linear methods for unsupervised dimensionality reduction, such as PCA, are inadequate to capture the complex non-linear relationships present in these data. For this reason, a deep unsupervised neural network based technique, parametric t-SNE, is adopted to map a 3D-DESI-MS dataset from a human colorectal adenocarcinoma biopsy onto a 2-dimensional manifold. This technique allows the identification of clusters not visible with linear methods. The unsupervised clustering of the tumour tissue results in the identification of sub-regions characterised by the abundance of identified metabolites, making possible the formulation of hypotheses to account for their significance and the underlying biological heterogeneity in the tumour. Intra-tumour phenotypic heterogeneity in human cancer has been associated with tumour progression, treatment resistance and metastasis development.1 3D Mass Spectrometry Imaging (MSI), being able to capture the different molecular patterns present in sub-regions of the tumour tissue, represents a highly promising approach for probing tumour and tumour-microenvironment heterogeneity.2–5 Determining the regions of tumour similarity and heterogeneity is not only crucial to investigate the nature of the diversity of tumours and to classify those into sub-groups, but can provide, through a topological mapping of the heterogeneity, an invaluable tool to understand the possible interactions between those different cell clusters.6 The study of biological interactions in three dimensions is essential,7–9 since biochemical mechanisms occur in a 3-dimensional environment whose complexity and richness may not be captured by the analysis of only a 2-dimensional sample of the tissue. From the point of view of statistical modelling, the lack of the standard state (the ‘normal’ cell type for this tissue) and a comprehensive compendium of the possible tumour cell types represents the biggest obstacle in the identification of tumour sub-types, requiring the employment of unsupervised learning techniques. Supervised classification of DESI imaging data from brain tumours was used by Eberlin et al. 10 for the identification of molecular patterns related to different types of tumours, but the main limitation of this approach is represented by the impossibility of identification of new tumour sub-types. In a similar vein, previous work has applied unsupervised analysis to MSI datasets to study intra-tumour heterogeneity. In Balluff et al.,11 a set of clustering algorithms were applied to matrix-assisted laser desorption ionization (MALDI) imaging data from gastric and breast carcinoma patients. An agreement-based procedure12 was employed to extract the final segmentation of the images, exploiting the assumption that different algorithms should retrieve real clusters consistently. The main difficulty of this procedure is represented by the selection of the clustering algorithms that should be compared, since some of those could provide similar results as they are founded on a similar concept of a cluster. An example is represented by PCA and k-means, which tend to capture the same kind of structures.13 This would result in an over-optimistic evaluation of the robustness of clusters. A similar difficulty is shown in Lou et al.,14 where similarly, the clusters are defined on the basis of consistency across a set of different algorithms. A further challenge is represented by the selection of the optimal number of clusters. For this reason, the challenges typical of unsupervised analysis, such as determining the correct number of clusters and the assessment of their validity15,16 can be reduced through inspection of the data structure.17 In order to make the visualisation of high-dimensional data (such as MSI) more straightforward, dimensionality reduction techniques are required. Several methods are currently available for unsupervised dimensionality reduction.18 Among these, linear techniques, such as Principal Component Analysis (PCA), are widely used to explore the internal relationships of mass spectrometry data.3,19 Unfortunately, these techniques can be inadequate to detect complex relationships between data, suggesting the application of non-linear methods.20,21 Such techniques, however, often make it difficult to extend the non-linear models to unseen data without introducing some degree of approximation.22 This aspect is critical in the case of 3D MSI data, where the datasets can consist of hundreds of thousands of spectra. For example, multi-dimensional scaling (MDS), implemented in Cornett et al. 23 or hierarchical clustering, would hardly be feasible because of the necessity of a complete pairwise distance matrix. Self-organizing maps (SOM), and in particular the extension, generalized self-organizing maps (GSOM), were used in Wijetunge et al. 20 to extract similar ion images from MALDI data. However, a limitation of SOM-based techniques for data dimensionality reduction is exemplified by the fact that high dimensional data are projected on a fixed grid, hence losing the possibility to project and separate ambiguous objects in different regions of the low-dimensional space. This limitation is overcome by Stochastic Neighbour Embedding (SNE) that makes the high dimensional data fixed and determines a continuous mapping for the low-dimensional embedding.24 Application of t-SNE to mass spectrometry imaging data can be found in Fonville et al. 25 and Abdelmoula et al. 26 However, it should be stressed that in those two works, a non-parametric t-SNE was employed. This has two important consequences: (1) the difficulty of projecting unseen data in the low-dimensional space without any approximation or ad hoc assumption, (2) the possibility of obtaining different results as the t-SNE cost function is not convex. The challenges described above can be addressed simultaneously through the application of deep learning based techniques. Firstly, a parametric model can naturally project the unseen high-dimensional data to the low-dimensional space and, secondly, the use of autoencoders trained to reconstruct the original data makes the initial parameters (weights and biases) used during the fine-tuning more stable across different runs. This, in the case of t-SNE, is equivalent to having more stable low-dimensional representations of the high dimensional data.27

160TopL,130TopR,150BottomL,140BottomR-profile

创建时间: 2025-01-10 17:31:59
空间分辨率: 17μm,   142x136
扫描: DESI ()
物种: Homo sapiens (NA) - normal

439TopL, 409TopR, 429BottomL, 419BottomR-profile

创建时间: 2025-01-10 17:19:38
空间分辨率: 17μm,   157x136
扫描: DESI ()
物种: Homo sapiens (colorectal adenocarcinoma) - Cancer

520TopL, 490TopR, 510BottomL, 500BottomR-profile

创建时间: 2025-01-10 17:17:40
空间分辨率: 17μm,   147x131
扫描: DESI ()
物种: Homo sapiens (colorectal adenocarcinoma) - Cancer

80TopL, 50TopR, 70BottomL, 60BottomR-profile

创建时间: 2025-01-10 12:39:05
空间分辨率: 17μm,   137x136
扫描: DESI ()
物种: Homo sapiens (colorectal adenocarcinoma) - Cancer