Page 28 - TD-3-1

Tumor Discovery AI uncovers tumor spatial organization

1. Introduction neural network (GNN) architectures aimed at reducing
the dimensionality of ST data and identifying spatial
Spatiotemporal molecular medicine involves regions. Notably, SpaGCN presents a comprehensive
18
comprehending medicine across various dimensions, framework that partitions data based on the aggregation
layers, perspectives, and dynamics through the integration of gene expression, histology, and spatial location. In the
of spatialization and temporalization of clinical cell clustering for ST data technique, a sequence of graph
phonemes with spatiotemporal molecular omics. This convolutional networks is embedded within the deep
1
emerging discipline strives to elucidate the pathogenesis, graph infomax module to derive embeddings for cell
epidemiology, historical context, patient symptoms and nodes. STAGATE devises a graph attention autoencoder
20
19
signs, clinical measurements, and therapeutic approaches. framework for spatial domain identification. This
Spatially resolved transcriptomics plays a crucial role
in connecting and correlating information between approach constructs a cell-type-aware spatial neighbor
network to characterize the spatial similarity at domain
histological sections and molecular profiles, such as cell- boundaries. Similarly, GraphST utilizes a graph self-
21
cell interactions, transcription factor distribution, spatial supervised contrastive learning mechanism for the unified
positioning, and mRNA expression within the cell through analysis of ST data, encompassing spatial clustering,
artificial intelligence, computerized programming, and
visualization techniques. Spatial transcriptomics (ST) multisample integration, and cell-type deconvolution.
2
technologies document transcriptomic expression data and The contrastive learning framework within the spatial
their respective two-dimensional tissue locations at spatial clustering module enables the acquisition of informative
3
resolution. In the past decade, various ST technologies and discriminative spot representations by minimizing the
have been introduced to capture mRNA molecules and embedding distance between spatially adjacent spots. All
the algorithms mentioned above can be improved in terms
their locations, each differing in resolution, sensitivity,
sequencing depth, and scale. Some notable technologies of clustering performance.
include seqFISH+, MERFISH, omsFISH, 10× Visium, Motivated by the successful application of diverse
4
7
5
6
Slide-seqV2, 10× Xenium, and Stereo-seq. 10 artificial intelligence algorithms in analyzing ST datasets,
8
9
Partitioning spatial domains in a tumor’s spatial this paper introduces a variational graph autoencoder
transcriptome data can shed light on the composition (VGAE) enriched with multiple layers of GNNs for
of the tumor microenvironment and the localized identifying spatial domains in tumor tissues (Figure 1).
development of tumor cells. The primary goal of spatial Initially, gene expression profiles and spatial location
clustering is to segment tumor tissues into distinct cell information are transformed into a gene feature matrix and
subpopulations, facilitating downstream analyses of a cell adjacency matrix, respectively. The VGAE consists of
differential gene expression, biological function grouping, two primary components: an encoder and a decoder. The
and cell-cell communication. Conventional methods encoder employs GNNs to extract matrix features, reduce
11
classify cell types based on gene expression profiles using the feature dimensions, and generate latent embeddings. In
algorithms such as K-means and Louvain. Giotto contrast, the decoder reconstructs two matrices based on
13
12
identifies spatial domains through the construction of a these latent embeddings. Clustering methods were, then,
hidden Markov random field model, leveraging both the applied to latent embeddings to detect spatial domains
feature matrix and spatial coordinates. stLearn utilizes a within the tumor tissue. Various loss functions are utilized
14
standard Louvain clustering workflow to handle the cell to enhance the quality and performance of the latent
adjacency matrix, subsequently using spatial positions to embeddings by minimizing the overall errors, including
identify subgroups within broader clusters. BayesSpace the reconstruction error and Kullback-Leibler (KL)
15
employs a fully Bayesian statistical model to enhance divergence error. The presented spatial clustering method
resolution and conduct clustering analysis, promoting the is compared with existing approaches to demonstrate
categorization of nearby spots into the same group based its effectiveness. Additional details are provided in the
on a predefined spatial prior. Despite the potential of spatial “Methods” section. In addition, we employ this clustering
location information to enhance clustering accuracy, the framework on different tumor tissues to investigate their
aforementioned spatial clustering algorithms have yet to respective tumor microenvironments.
achieve optimal performance. 16
2. Methods
Graph deep learning (GDL) has emerged as a
promising approach for integrating gene expression 2.1. VGAE
profiles and spatial location information to address spatial The VGAE is a machine learning model that is primarily
clustering tasks. Numerous studies have proposed graph utilized in the domains of graph representation learning
17

Volume 3 Issue 1 (2024) 2 https://doi.org/10.36922/td.2049