SEDIST: Spatially Enhanced Domain Identification through Spatial Transcriptomics

SEDIST: Spatially Enhanced Domain Identification through Spatial Transcriptomics | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article SEDIST: Spatially Enhanced Domain Identification through Spatial Transcriptomics Fahad Rahman, Abdullah Al Sakib, Riasat Azim, Fahim Hafiz, Shabrina Airen Esha, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6619020/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Motivation: The recent advancement of spatially resolved transcriptomics provides a way to investigate cellular heterogeneity and tissue micro-environment using gene expression profiles with spatial context, often accommodated with histological data. However, precise spatial domain identification, multisample integration, and cellular decomposition still remain challenging. Results Here we introduce SEDIST, a novel method that integrates autoencoders and graph neural networks with self-supervised contrastive learning to utilize spatial transcriptomics data effectively. This approach enhances the model’s ability to extract informative and discriminative spot representations by minimizing the embedding distance between spatially adjacent spots through self-supervised contrastive learning. Highly expressed genes in different domain layers are also identified and 3D embedded plot of clusters along with diffusion map are also graphically represented as downstream analysis. This comprehensive integration of spatial context leads to superior performance compared to existing state-of-the-art (SOTA) methods in the human brain dorsolateral prefrontal cortex (DLPFC) and the BRCA (human breast cancer) 10X Visium dataset, as evidenced by achieving an Adjusted Rand Index (ARI) of 0.6466 and 0.6185, respectively. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1 Introduction Recent technological advances in spatial transcriptomics(ST) have brought new opportunities for biomarker identification, spatial heterogeneity analysis and characterization of gene expression while taking into consideration the spatial location of gene expression [ 1 ]. The spatial domains and their associations are often responsible for the tissue microenvironments and their biological functions. The ways in which complex biological tissue functions are closely related to the spatial context of the cells. Understanding gene functions is crucial to discovering activities in different organs, and understanding disease pathways. Traditional techniques such as bulk RNA sequencing (RNA-seq) and single-cell RNA sequencing (scRNA-seq) have significantly advanced our knowledge by identifying which genes are expressed. However, these methods lack spatial resolution, meaning that they do not provide information about where gene expression occurs in the tissue. This limitation is critical, as the spatial context of gene expression plays a key role in cellular function, organization, and communication. Spatial transcriptomics is an emerging technology that overcomes this limitation by enabling the simultaneous measurement of gene expression levels and their spatial locations within tissues. By integrating spatial data with gene expression profiles, this technique allows researchers to gain deeper insights into cell-to-cell interactions, tissue architecture and disease mechanisms at the cellular level. As a result, spatial transcriptomics offers a more comprehensive understanding of biological processes and has the potential to advance both basic research and clinical applications. Experimental technology in spatial transcriptomics can be broadly categorized into two types: (i) imaging-based technologies and (ii) sequencing-based technologies. Imaging-based technologies use highly multiplexed adaptations of single-molecule fluorescence in situ hybridization (smFISH), such as sequential FISH (seqFISH) and MERFISH along with histology, molecular, and imaging tools without relying on sequencing methods [ 3 ].Molecular cartography, MERSCOPE, and 10x Xenium are examples of imaging-based methods that are adopted from sequential FISH (seqFISH), MERFISH, and 10x Genomics technology. These imaging-based methods provide single-cell or subcellular level resolution such as Molecular Cartography (275nm), MERSCOPE (100nm), and Xenium (50nm). However, they can only provide the profile of a few hundred to a few thousand genes due to various technical limitations [ 2 ]. Conversely, the sequencing-based methodology includes 10x Visium, GeoMx DSP and Stereo-seq. These methods combine spatial barcoding with next-generation sequencing technology to preserve spatial context. Although these methods can profile the whole transcriptome, the spatial resolution varies across them such as 10x Visium (100 µm), GeoMx DSP (10 µm) and Stereo-seq (0.5 µm). Regardless of the spatial transcriptomics methodology employed, the initial step in any spatial expression analysis workflow is to identify and define the cell types within the 2D tissue sample. The most essential task of spatial transcriptomics is the identification of domains and cell types. Further downstream analysis and understanding of the tissue microenvironment often rely on the identification of the cell types. Several methodologies have been constructed to facilitate such clustering for single-cell spatial transcriptomics data. These methods usually convert the high-dimensional data into low low-dimensional space and cluster the cells. Furthermore, clusters are categorized into known or novel cell types by analyzing the expression of a certain group of marker genes. stPlus proposed an autoencoder-based framework with a tailored loss function to enhance the spatial transcriptomics data by leveraging scRNA-seq data as a reference [ 3 ]. This proposed method outperformed other methods, such as Seurat [ 4 ], SpaGE [ 5 ], and Liger [ 6 ] in predicting gene expression in spatial transcriptomics data. Furthermore, it enables better identification of cell populations and more accurate segmentation of tissue sections. This demonstrates that integrating additional data with spatial transcriptomics data enhances the accuracy of the analysis. FICT proposed a novel clustering method designed for spatial transcriptomics data by integrating both gene expression and spatial information to improve cell-type assignment. It demonstrated superior performance on both simulated and real datasets such as MERFISH, osmFISH, and seqFISH, than other methods that use only gene expression or partial spatial information [ 7 ]. STAGATE analyzed spatial transcriptomics data by identifying spatial domains and tissue architectures, particularly in brain tissues [ 8 ]. To better identify genes that are spatially expressed and reduce noise, STAGATE integrates spatial information with gene expression profiles using a graph attention method. It builds a spatial neighbor network and learns low-dimensional representations using gene expression and geographic data via a graph attention auto-encoder. The approach outperforms SOTA techniques such as SCANPY [ 9 ] and SEDR [ 10 ] in uncovering tissue architecture and laminar organization across many datasets. GraphST used spatial location and spatial gene expression for spatial clustering, multi-sample integration, and integration of spatial transcriptomics and scRNA-seq [ 11 ]. The authors proposed augmentation-based contrastive learning by creating a corrupted graph by random shuffling of gene expression across spots. This allows the self-supervised contrastive model to capture the local context surrounding the environment. Then GNN encoder is used to learn the latent representation of the corrupted and uncorrupted data. The local context in the latent representation helps to determine accurate gene expression and graphs that are reconstructed using a decoder. This graph self-supervised contrastive learning approach in clustering achieves 10% higher accuracy than other methods such as STAGATE [ 8 ], BayesSpace [ 12 ], SpaGCN [ 13 ], DeepST[ 14 ], GraphST[ 11 ] and ConST [ 15 ]. scFeatures [ 16 ] published an R package designed to generate a multi-view molecular representation of samples by creating interpretable features from scRNA-seq and spatial transcriptomics data, to enhance the understanding of disease outcomes. The findings indicate that different feature types are useful for different datasets. Another study, NICHES [ 17 ], integrates the detailed gene expression profiles from scRNA-seq with the spatial context provided by spatial transcriptomics. Constructing matrices that capture ligand-receptor interactions using the integrated data and visualizing these interactions creates a detailed map of how cells communicate within their local environment, offering a deeper understanding of cell-cell interactions in tissues. Similar approaches of integrating spatial transcriptomics data with scRNA-seq data, interactive analysis, and visualization of the integrated data have been discussed in [ 18 ][ 19 ][ 20 ]. Another approach to analyzing spatial transcriptomics data is utilizing histology information along with gene expression and spatial data. stLearn [ 21 ] integrates gene expression, spatial location, and histology data to cluster similar regions within tissue sections, providing a basic understanding of tissue organization. While it combines these data sources to identify spatial domains, it treats spatial domain identification and the detection of spatially variable genes (SVGs) as separate tasks. This limits stLearn's ability to link gene expression patterns with histology, making it difficult to understand how differences in cell types relate to the overall organization of the tissue. SpaGCN [ 13 ], on the other hand, addresses this issue by integrating spatial domain identification and SVG detection into a unified approach while utilizing scRNA-seq data, spatial location, and histology. In another study, DeepST [ 14 ] integrated the spatial location, gene expression, and tissue image to enhance the identification of spatial domains in samples. Image (H&E staining tiles) is segmented according to the coordinates of the spots, and consecutively, a pre-trained convolutional neural network (CNN) is used to transform each spot into 2048 latent variables, creating an image matrix. Additionally, graph representations of the spatial gene expression are used to determine the adjacency matrix between spots. These image matrices, the adjacency matrix, and gene expression are used to determine enhanced gene expression. A denoising autoencoder with a variable graph autoencoder is used to determine the final reconstructed spatial gene expression with higher accuracy compared to SpaGCN, stLearn, SEDR, BayesSpace, DeepST, ConSt, and GraphST. Furthermore, we also benchmarked some recent spatial transcriptomics methods such as: GRAS4T[ 22 ], GAAEST [ 23 ], and MNMST [ 24 ] with the SEDIST. By addressing limitations in feature extraction, spatial integration, and computational inefficiencies, we proposed this deep learning-based method named SEDIST to accurately identify spatial domains from spatial transcriptomics data. Initially, we extract features of tissue image tiles from the morphological image utilizing an autoencoder architecture. A feature vector was created for each spot of the tissue, and integration with the gene expressions and spatial locations was performed to create enhanced representations of the spatial domains. Furthermore, we converted these integrated representations into a graph where the graph undergoes a shuffling process along with randomness for enhanced robustness of the graph data. Moreover, this modified graph is passed through a graph autoencoder to learn an enhanced matrix representation of the graph. To achieve the final matrix representation, a self-contrastive mechanism is applied to project the gene expression data into a spatial space by emphasizing the spot similarities. After that, the final representation is used to cluster the spatial domains using Mclust. In order to ensure alignment with ground truth data, the number of clusters is decided by manually annotating tissue slices. Utilizing human dorsolateral prefrontal cortex (DLPFC) data from the LIBD repository and Human Breast Cancer (BRCA) dataset, SEDIST is tested extensively and provides better accuracy in identifying spatial domains compared to other SOTA methods. Another extensive feature of this method named SEDIST is capable of presenting the 2D embedding in a 3D spatial space. SEDIST effectively maintains consistency across multiple datasets, such as Visium 10x, standard Visium. This method’s consistent performance across different datasets demonstrates its flexibility and robustness. Additionally, balances clustering accuracy with computational efficiency, optimizing execution time and memory usage for large-scale datasets and resource-constrained environments. Furthermore, it showed more computational effectiveness compared to the other recent methods, such as GraphST, SpaGCN, DeepST, GAAEST, GRAS4T, and MNMST. In summary, the Findings we present in this paper underscore the robustness of SEDIST in accurately identifying spatial domains, while also highlighting its scalability for processing additional spatial omics data. 2 Materials and Methods 2.1 Overview of SEDIST The proposed method SEDIST, consists of four levels of architecture, as shown in Fig. 1 . Initially, an H&E-stained image undergoes segmentation, resulting in its division into multiple sections, which are known as spots. Subsequently, an autoencoder is applied to discern the underlying patterns, generating an image feature vector. By utilizing spatial locations, a graph is constructed using the distance measure. Subsequently, the feature vector, along with the image feature vector and corresponding gene expressions of specific spots undergoes feature extraction through a scalar function and principal component analysis (PCA). Here, log normalization is applied to the datasets as a scalar function. Such preprocessing step is crucial before applying PCA as it ensures that no single feature disproportionately influences the PCA due to its magnitude, thereby giving equal importance to all the features. From the resultant combined features and undirected neighborhood graph, a modified graph is constructed which subsequently goes undergoes a shuffling process to introduce randomness and enhance robustness. Spatial information is then incorporated to spatially map the graph, aligning it with the original spatial distribution of spots. After that, a graph encoder is subsequently deployed to unveil latent features from graphs and transform them into two enhanced matrix representations. We implemented self-supervised contrastive learning (SSCL) to those matrix representations to increase the correctness of adjacent or neighborhoods and converted them into a final matrix representation. Finally, the modified matrix representation is utilized using the nonspatial assignment algorithm Mclust[ 25 ] for the clustering process, which is specialized in clustering spatial domains through expectation maximization . 2.2 Data Collection and Preprocessing In SEDIST, the human dorsolateral prefrontal cortex (DLPFC) and BRCA (human breast cancer) dataset is used which is collected from the LIBD repository [ 1 ] and SEDR [ 10 ]. The DLPFC dataset comprises 12 tissue slices obtained through the 10x Visium platforms. Each tissue slice has a varying number of spots, ranging from 3460 to 4789, and captures information from 33,538 genes. Manual annotations were applied to categorize each slice into distinct regions, including DLPFC layers and white matter. DLPFC datasets consisted of cell type composition, spatial organization, and gene expression profiles within each spatial spot, utilizing deconvolution techniques in spatial transcriptomics (ST) using scRNA-seq reference data were employed. In one instance, we utilized slice #151673 from the human brain sample DLPFC dataset. The corresponding scRNA-seq data for the dorsolateral prefrontal cortex (BA9) tissue, obtained from postmortem archives, were generated using the 10x Genomics Chromium platform [ 26 ]. This scRNA-seq dataset comprises 78,886 cells and captures gene expression from 30,062 genes. The BRCA dataset consists of 3799 cells having 20 layers. All data utilized in the experiment and validation process are presented in Supplementary Table S2 . To facilitate a comprehensive comparison, additional slices from the DLPFC dataset were utilized for assessing the performance of the proposed method against other existing state-of-the-art (SOTA) methods. 2.3 Image Matrix generation Using H & E staining, the morphological image is divided into smaller sections or images according to the spatial coordinates of each spot. Spot tiles are augmented and transformation method such as rotation, normalization, sharpness adjustment are applied [ 27 ]. Subsequently, an autoencoder is applied to extract each spot tile’s high-level features. The dense layer in the autoencoder began at 1024. Each dense layer of the autoencoder contains a number of fixed morphological features. Extracted a total of 14400 high-level morphological features from the autoencoder’s output in this instance [ 28 ][ 29 ]. Finally, the feature weights of morphological similarity \(\:{MS}_{ij}\) between spot \(\:{S}_{i}\) and adjacent \(\:{S}_{j}\) were calculated using the cosine distance as follows Eq. 1 : $$\:{MS}_{ij}=1-\frac{{S}_{i}\cdot\:{S}_{j}}{{\Vert\:{S}_{i}\Vert\:}_{2}{\Vert\:{S}_{j}\Vert\:}_{2}}$$ 1 2.4 Graph Construction Using a predetermined number of K = 3 neighbors, we created an undirected neighborhood graph \(\:G=\left(V,\:E\right)\:\) to employ the spatial data fully. The set of edges that connect the neighborhood spots represented by E to V, the collection of spots, is represented in graph G. The adjacency matrix of graph G is defined as follows: \(\:\text{}\text{(}A\in\:{\mathbb{R}}^{{N}_{\text{spot}}\times\:{N}_{\text{spot}}}\text{)},\:\) where \(\:{N}_{\text{spot}}\) is the number of spots. If spot i is the neighbor of spot j , then \(\:{a}_{ij}=1\) ; otherwise, 0. As a result, the Euclidean distance, which is computed using the spatial position data, determines a spot’s neighbors based on how close it is to [ 30 ]. Finally, we select k spots as neighbors from the top-ranked nearest neighbors. In our experiments, the best performance across most tested datasets was achieved when k = 3 . It also aligns with the parameter setting utilized by GraphST in its implementation. 2.5 Data Augmentation Primarily, another neighborhood graph is constructed from augmented data. More specifically, the combined features of the image matrix and arbitrarily rearranged gene expression vectors X among the spots are added to create the modified graph while preserving the original graph’s topological structure [ 31 ]. \(\:{G}^{{\prime\:}}=\left({V}^{{\prime\:}},\:{E}^{{\prime\:}}\right)\:\:\) subsequently represents the modified graph, and X’ represents the shuffled gene expression. 2.6 Latent Representation Encoding In order to learn spot representations that incorporate the informative portions of the gene expression profiles and spatial locations we leveraged a GNN-based encoder [ 32 ]. The normalized gene expression profiles X and the neighborhood graph G are fed into the encoder, which then delivers the modified gene expressions H to the decoder. To be more precise, a graph convolutional network (GCN) is used as an encoder to combine the representations of a spot’s neighbors in order to learn a latent representation for it. In mathematical terms, the encoder’s l-th layer representations are defined by Eq. 2 : $$\:{Z}_{s}^{l}=\sigma\:\left(\stackrel{\sim}{A}{Z}_{s}^{l-1}{W}_{e}^{l-1}+{b}_{e}^{l-1}\right)\:\:\:\:\:\:\:\:\left(2\right)\:$$ \(\:{Z}_{s}\:\) is denoted as the final output of the encoder while \(\:{Z}_{s}^{l}\) denotes the l-th layer output representation . \(\:\stackrel{\sim}{A}={D}^{-\frac{1}{2}}A{D}^{-\frac{1}{2}}\) indicates the location of the diagonal elements of the diagonal matrix in the normalized neighboring matrix and denotes a trainable weight matrix W and a bias vector \(\:b\) , respectively. The Rectified Linear Unit (ReLU) is applied as a nonlinear activation function which is defined by \(\:\sigma\:\) . Subsequently, a decoder receives the latent representations and reverses it, restoring it to the original gene expression space. Similar to the encoder, the decoder reconstructs the gene expression using a symmetric architecture. The decoder is specified in particular as follows in Eq. 3 : $$\:{H}_{s}^{t}=\sigma\:(\stackrel{\sim}{A}{H}_{s}^{t-1}{W}_{d}^{t-1}+{b}_{d}^{t-1})\:\:\:\:\:\left(3\right)$$ Here, \(\:{H}_{s}^{t}\) is set as the encoder’s output representation at the t -th layer and \(\:\:{H}_{s}^{t-1}\) , represents the reconstructed gene expression profiles at the ( t-1) th layer. All nodes in the network share the trainable weight matrix, denoted by \(\:{\:W}_{d}\) , and the bias vector, denoted by \(\:\:{b}_{d}\) . The model is trained by minimizing the self-reconstruction loss of gene expressions, following a specified process to fully leverage the gene expression profiles. The loss function is calculated using Eq. 4 : $$\:{\text{L}}_{\text{recon\:}}=\sum\:_{i=1}^{{N}_{\text{spot\:}}}{\Vert\:{x}_{i}-{h}_{i}\Vert\:}_{F}^{2}\:\:\:\:\left(4\right)$$ For spot i , the original normalized gene expression is represented by \(\:{\:x}_{i}\) , and the reconstructed gene expression is represented by \(\:{\:h}_{i}\) . The \(\:{\Vert\:.\Vert\:}_{F}\) refers to the Frobenius norm, which is used for measuring the magnitude between two matrices. 2.7 Self-supervised contrastive learning SEDIST employs a Self-Supervised Contrastive Learning (SSCL) technique to make sure the model captures the local spatial context of spots, hence improving the representation’s discriminative and informative qualities. In particular, using the original and corrupted graph G and G’ as input, the GNN-based encoder initially creates two distinguished representation matrices for the specific spots. Inspired by deep graph informax (DGI), which is an unsupervised learning framework designed to learn node representations in graph-structured data, we combine the representations of the neighbors to describe a spot’s local context or its neighborhood microenvironment [ 33 ]. In contrast to DGI's use of global neighbors, the readout function here is redefined by applying a sigmoid to the mean of the immediate neighbors' representations. This approach assumes that each point in the spatial data is characterized by its local environment, which typically reflects the cell types and gene expression profiles of its local environment. For each spot i in the graph, the local context vector and its representation \(\:{P}_{i}\) create a positive pair, and the local context vector g and its equivalent representation \(\:{P}_{i}^{{\prime\:}}\) from the corrupted graph form a negative pair. Maximizing positive pairs’ mutual information while limiting negative pairs’ mutual information is the fundamental principle of SSCL [ 35 ]. Physically adjacent spots will exhibit similar representations through the application of contrastive learning, whereas non-adjacent spots will display dissimilar representations. Subsequently, SSCL is modeled by utilizing binary cross-entropy (BCE). The contrastive loss has the following formal definition as shown in Eq. 5 : $$\:{\mathcal{A}}_{1}=\sum\:_{i=1}^{{N}_{\text{spot\:}}}\left({\mathbb{E}}_{\left(X,A\right)}\left[\text{log}{\Phi\:}\left({p}_{i},{q}_{i}\right)\right]\right)\:\:\:\:\:\:\:\:\:\:\:\:$$ $$\:{\mathcal{B}}_{1}={\mathbb{E}}_{\left({X}^{{\prime\:}},{A}^{{\prime\:}}\right)}\left[\text{log}\left(1-{\Phi\:}\left({p}_{i}^{{\prime\:}},{q}_{i}\right)\right)\right]\:\:\:\:\:\:\:\:\:\:$$ $$\:{\mathcal{L}}_{SSCL}=-\frac{1}{2{N}_{\text{spot\:}}}\left({\mathcal{A}}_{1}+{\mathcal{B}}_{1}\right)\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(5\right)$$ Where \(\:\varPhi\:\left(\cdot\:\right)\:\) dual neural network discriminator \(\:\mathcal{D}:{\mathbb{R}}^{d}\times\:{\mathbb{R}}^{d}\to\:\mathbb{R}\mathbb{\:}\) that separates the positive pairs from the negative pairs. \(\:\varPhi\:({p}_{i},{q}_{i})\) denotes the probability score assigned to the positive pair \(\:({p}_{i},{q}_{i})\) . Since the original graph \(\:G\:\) and the modified graph \(\:{G}^{{\prime\:}}\) share the same topological structure, a symmetric contrastive loss is employed, denoted as \(\:\mathcal{B}\) . This loss, defined in Eq. 5, is applied to the corrupted graph used in the modified loss \(\:\:{L}_{SSCL}\) to improve the model’s stability and balance. The detailed explanation of the contrastive loss of the original and modified graphs is provided in supplementary equation S1 and supplementary equation S2 . 2.8 Loss Function for SSCL The self-reconstruction loss and contrastive loss are minimized to train the ST data representation-learning module. In a nutshell, this module’s total training loss is defined as \(\:\mathcal{L}\) in Eq. 6 [ 34 ]: $$\:\mathcal{L}={\lambda\:}_{1}{\mathcal{L}}_{\text{recon\:}}+{\lambda\:}_{2}\left({\mathcal{L}}_{SSCL}+{\mathcal{L}}_{SSC{L}_{corrupt}}\right)\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(6\right)$$ The weight parameters that balance the effects of the contrastive loss and the reconstructive loss are \(\:{\lambda\:}_{1}\) and \(\:{\:\lambda\:}_{2}\) . We predefined \(\:{\lambda\:}_{1}\) and \(\:{\lambda\:}_{2}\) as 10 and 1 empirically. This module is trained without the use of ST data, and the optimization is achieved using the Adam optimizer. For spatial clustering, the training epoch and learning rate are set to 600 and 0.0025, respectively. 2.9 Location-specific gene expression analysis Spatial transcriptomics combines gene expression analysis with spatial information, enabling researchers to map where specific genes are active within tissues, offering deeper insights into cellular organization and function [ 35 ][ 12 ]. This behavior can be mapped as explained in Eq. 7 : $$\:S\left(x,y\right)=\sum\:_{i=1}^{N}{G}_{i}\cdot\:\partial\:\left(x-{x}_{i}\right)\cdot\:\partial\:\left(y-{y}_{i}\right)\:\:\:\:\:\:\:\:\:\:\:\left(7\right)$$ Here, \(\:S\left(x,y\right)\) shows how genes ( \(\:{G}_{i}\) ) are expressed in specific positions in the cell, and \(\:{(x}_{i},{y}_{i})\) is the coordinates of each cell. The equation ensures precise measurement of gene activity at specific cellular locations, facilitating an understanding of how genes interact across different cellular regions and providing insights into cellular architecture [ 36 ]. Incorporating spatial location allows for the observation of gene interactions at specific sites, offering a more detailed view of cellular organization and function. The spatial transcriptomics method increases our understanding of cell architecture and opens up new possibilities for studying cellular interactions in various biological processes [ 37 ][ 38 ]. 2.10 Evaluation Matrix The Adjusted Rand Index (ARI) works as a pivotal metric within the research framework, providing a robust assessment of the clustering algorithm’s efficacy in capturing meaningful cellular and molecular distinctions [ 39 ]. ARI evaluates the agreement between true and predicted cell or gene cluster assignments, offering a comprehensive understanding of the spatial and transcriptional landscape within tissues [ 40 ]. Here, the ARI is used to evaluate the performance of various clustering techniques including SEDIST on datasets containing spot-type labels. The ARI is calculated using the following Eq. 8 : $$\:ARI=\frac{\sum\:_{ij}\left(\begin{array}{c}{n}_{ij}\\\:2\end{array}\right)-\left[\sum\:_{i}\left(\begin{array}{c}{a}_{i}\\\:2\end{array}\right)\sum\:_{j}\left(\begin{array}{c}{b}_{j}\\\:2\end{array}\right)\right]/\left(\begin{array}{c}n\\\:2\end{array}\right)}{\frac{1}{2}[\sum\:_{i}\left(\begin{array}{c}{a}_{i}\\\:2\end{array}\right)+\sum\:_{j}\left(\begin{array}{c}{b}_{j}\\\:2\end{array}\right)]-\left[\sum\:_{i}\left(\begin{array}{c}{a}_{i}\\\:2\end{array}\right)\sum\:_{j}\left(\begin{array}{c}{b}_{j}\\\:2\end{array}\right)\right]/\left(\begin{array}{c}n\\\:2\end{array}\right)}\:\:\:\:\:\:\:\:\left(8\right)$$ Here, \(\:{n}_{ij}\) represents the number of data points shared between clusters i and j , \(\:{a}_{i}\) denotes the total number of data points in cell or gene cluster i . \(\:{b}_{j}\) represents the total number of data points in cluster j , and N is the total number of data points. The resulting normalized index ranges from 0 to 1, with a higher ARI indicating a more accurate agreement between true and predicted cluster assignments [ 41 ]. 3 Result Analysis In this section, we discuss SEDIST's overall performance. The proposed method showed enhanced capability of identifying spatial domains in most of the tissue slices, while achieving around a maximum of 0.6466 ARI. Additionally, we discuss the exploratory analysis capability of this method such as 3D embedding, diffusion map, draw map and identification of top genes within each cell. 3.1 Performance Comparison with State-of-the-Art (SOTA) Methods on DLPFC Datasets The DLPFC benchmarking dataset, annotated by Maynard et al. for cortical layers and white matter using gene markers and cytoarchitecture (Fig. 2 A), was used to compare the performance of SEDIST with other state-of-the-art (SOTA) spatial clustering methods DeepST, GraphST, conST, GRAS4T, GAAEST, MNMST, SpaGCN and SpaceFlow. All models were run under a standardized development environment with PyTorch version 2.3.0 and the hardware configuration specified in Section 3.5 to ensure fair and consistent evaluation. Our results demonstrate that SEDIST consistently aligns well with manual annotations of the DLPFC dataset and outperforms several existing methods across most of the 12 dataset slices. Our results show that SEDIST achieved the highest Adjusted Rand Index (ARI) score of 0.6466 on slice 151673 by outperforming GraphST (0.6123), DeepST (0.5855), and MNMST (0.6012). These findings highlight the effectiveness of SEDIST in accurately identifying spatial domains. In Fig. 2 B, the visual comparison further emphasizes SEDIST's strong ARI of 0.6466 on slide 151673 for layer boundary delineation. It is evident that SEDIST outperforms DeepST, GraphST, conST, GRAS4T, GAAEST, MNMST, SpaGCN and SpaceFlow on this particular slice. Among these, only SEDIST and GraphST were successful in clearly distinguishing between cortical layers L1 and L2. Across the 12 slices listed in Supplementary Table S1 , SEDIST outperformed all other methods in 8 slices, specifically 151508, 151510, 151670, 151671, 151672, 151673, 151674, and 151676, demonstrating a consistent trend of strong performance. For example, on slice 151670 (Fig. 2 H), SEDIST achieved an ARI of 0.6244 which is higher than GraphST (0.5004) and DeepST (0.3048). Similarly on slice 151672 (Fig. 2 J), SEDIST obtained an ARI of 0.6273 by outperforming GraphST (0.5733), DeepST (0.4512) and MNMST (0.5738). Another example is slice 151674 (Fig. 2 L), where SEDIST reached an ARI of 0.5993, again surpassing GraphST (0.5561) and MNMST (0.5361). These examples highlight the SEDIST of ability to consistently identify spatial domains more effectively than other state-of-the-art approaches across diverse samples. However, SEDIST showed moderate performance in some other slices. For example, in slice 151507 (Fig. 2 C) SEDIST scored an ARI of 0.4361, which was lower than DeepST (0.5529) and MNMST (0.5203). Similar trends were observed in slice 151509 (Fig. 2 E), where DeepST slightly outperformed SEDIST. These variations highlight the importance of evaluating models across diverse biological conditions. While SEDIST excels in most cases, its performance can vary depending on the specific structural and biological features of the tissue slices. Figures 2 C through 2 N provide detailed visualizations of the spatial clustering outcomes across all slices and demonstrate the relative strength of SEDIST in real-world spatial domain identification tasks. 3.2 Exploring Spatial Organization and Gene Expression from SEDIST Visualizations and Gene Profiling In this proposed method, we also employed a 3D embedding technique for clustering gene expression data, enabling the integration of spatial information with gene expression profiles, and facilitating a more comprehensive understanding of transcriptomic patterns as illustrated in Fig. 3 A. The 3D embedding helps us visualize how genes are expressed in different tissue areas due to their projection in 3D space. We additionally observe from Fig. 3 A that the ground truth of the tissue slides closely matches with the predicted 3D embedding by our method, especially Layer 1, 2, 3, 5 & 6 as well as white matter. Furthermore, UMAP[ 42 ] and partition-based graph abstraction in SEDIST revealed a clear, structured arrangement of cortical layers from L1 to L6 and white matter (Fig. 3 B). Unlike the less-defined layer separation in 3D embedding (Fig. 3 A), the UMAP results show the delineation of each layer. This demonstrates how UMAP outperforms other methods in providing a more accurate and detailed organization of cortical structures. We also generated a diffusion map to understand the cell-cell connections among different tissue layers in terms of physical space and gene expression as shown in Fig. 3 C. In Fig. 3 C, we can observe the ground truth and the predicted diffusion map shows a similar pattern. We observe the continuum of cellular states in both maps as we move from one cortical layer to another. The subtle transitions between layers 1–6 and WM as well as their distributions are also visible in this figure. For example, the cells are distributed within a similar range along DC1 (Diffusion Component 1), indicating these layers are within the same tissue sections. However, we observe in Fig. 3 C that WM and Layer 6 exhibit a broad spread along DC2 (Diffusion Component 2), indicating significant heterogeneity within these layers. In contrast, layers 1–5 are narrowly clustered and overlapped along DC2, indicating a higher degree of homogeneity in their secondary characteristics. This pattern highlights distinct gene expression profiles and transitions across different cortical layers. Furthermore, we analyzed to facilitate the visualization of cells clustering in two-dimensional space FR1 (Functional Regions 1) & FR2 (Functional Regions 2) based on the gene expression patterns as illustrated in Fig. 3 D. Unlike diffusion maps, which cluster cells based on their gene expression and spatial location, focus only on gene expression. Figure 3 D demonstrates that in both ground truth and predictions of SEDIST, layer 1 is distinctly clustered (blue color in ground truth and brown color in SEDIST), suggesting that layer 1 is well-defined in terms of gene expression and our method can distinguish it accurately. Additionally, the WM is separated in both visualizations, indicating that SEDIST can identify the transition between cortical layers and white matter. However, the predicted drawgraph has some discrepancies in layers 2–6 compared to the ground truth. While these layers overlap in the ground truth, SEDIST predicts more distinct clusters (green, orange, and blue), indicating deviations in the clustering. The combined integration of UMAP, Diffmap[ 43 ], and Drawmap [ 44 ] in SEDIST results enhanced the interpretability of our proposed method and provided a comprehensive view of the spatial and transcriptional layout within cortical layers. Furthermore, SEDIST can figure out spatially variable gene influence across the tissue slice to understand the intricate cell behavior as illustrated in Fig. 3 E. In Fig. 3 E, we demonstrate these genes’ (MBP, PLP1, NEFL, COX6C, IGKC, HPCAL1 and PCP4) spatial distribution across the overall tissue slices. We observed varying levels of expression across the tissue for these genes, with a few spots of higher expression (yellow color), indicating specific cellular processes influenced by these genes. The consistent low expression (black color) of these genes may provide insights into their role in maintaining cellular homeostasis. We can understand how gene expression varies across tissue by visualizing such a spatial distribution graph for particular genes, which is crucial for understanding complex biological processes, cell behavior, cell-cell communications, and advancing therapeutic interventions. Additionally, we discuss how SEDIST can perform a comprehensive analysis of individual cell gene expression to identify the main genetic drivers influencing cellular behavior within spatial transcriptomics. Figure S3 shows a box plot of the top 20 genes by percentage of total counts in the dataset identified by SEDIST. Here, we calculated each gene expression level across all the spots in the tissue and showed the top genes that showed the highest mean expression level across all the cells. In this result, we observe that a significant portion of the top-expressed genes are Mitochondrial (e.g., MT-CO2, MT-ND4, MT-CYB), indicating cells undergoing metabolic activity since mitochondrial genes are involved in energy production. Mitochondrial genes usually perform different metabolic functions in tissues. However, several MT genes exhibit outliers, indicating that certain cells have unusually high metabolic activity or are in a low-energy state. These variations might reflect differences in cellular states, such as heightened metabolic activity in certain regions or phases of the tissue. A recent study showed that mitochondrial genes were down-regulated in the tissue of aging mice whereas some mitochondrial genes were up-regulated for specific biological roles [ 45 ]. Another study discussed how mitochondrial gene expression levels were up-regulated in the prefrontal cortex due to chronic stress [ 46 ]. Such studies indicate that the overall mean expression level of mitochondrial genes across cortical layers can be up-regulated which we also demonstrated in Supplementary Figure S3 . It is also observed that GAPDH[ 47 ], ACTB[ 48 ], and RPL41[ 49 ] genes (a housekeeping gene) among the top 20 genes which are expected because these genes are often highly expressed in many cell types. studies suggest that certain GAPDH genes can have high levels of expression due to stressful conditions and other biological reasons [ 50 ][ 51 ]. For example, a study asserted that housekeeping genes are usually expressed stably across tissues and used as a reference. Genes like FTH1 [ 52 ], CST3 [ 53 ], and CLU[ 54 ] in the top genes suggest that some cells in the tissue slice may be under stress or actively maintaining homeostasis. This could be due to the tissue type, experimental conditions, or cell state (e.g., proliferation, differentiation, or apoptosis). The expression of CALM1[ 55 ] and TMSB10[ 56 ] suggests active calcium signaling and cytoskeletal dynamics, which are important for many cellular processes. However, under various conditions, they can also show high expression levels[ 51 ]. This analysis helps to identify the genes that are most actively expressed in cells and gives different biological insights, such as the metabolic state of cells, possible differences between cell types, and gene expression roles in governing cellular functions and interactions within the tissue microenvironment. 3.3 Insights from SEDIST: Exploratory Analysis of Gene Expression and Spatial Architecture We also examined the complexities of feature representation in PCA through graphical visualization (See Supplementary Figure S1 ). We selected the top 20 principal components (PCA20) to retain the most significant variance in the dataset while reducing dimensionality and minimizing redundancy in the feature space. This allowed our analysis to focus on the most informative features (essential gene expression patterns) while avoiding data overload. To interpret the distinctions between normalized and non-normalized gene expression, SEDIST utilized variance versus mean expression plots (See Supplementary Figure S1 ), which was essential for ensuring the accuracy and reliability of our gene expression analyses, as normalization significantly affects the interpretation of gene expression data. Moreover, the top 10 genes in each layer were identified within slice 151673 as shown in Fig. 4 . These genes, such as MBP, PLP1, and CNP dominated in the cortical layer 1, contributing to processes like myelination. Other layers showed distinct gene profiles such as KRT17 and DIRAS2 in layer 2, and NEFL and NRGN in the third, each demonstrating valuable insights into the spatial organization and unique biological functions within the cortical layers. 3.4 Performance of SEDIST on Visium Dataset In this study, we also applied the SEDIST to spatial transcriptomics data from 10x Genomics Visium for human breast cancer, a cancer known for its high levels of variability within and between tumors [ 10 ]. To check the robustness of SEDIST we performed spatial transcriptomics analysis on Visium dataset along with Visium 10x. This was performed to ensure the robustness. We tested it on the Human Breast Cancer Block A (Section 1) dataset. This BRCA-1 H&E slice has a total number of 20 different regions based on the pathological features, with 4 major morphotypes. Our results, shown in Fig. 5 , highlight SEDIST’s ability to identify spatial domains and provide different visual perspectives of its clustering performance. SEDIST achieved an Adjusted Rand Index (ARI) score of 0.6185, which indicates its accuracy in dividing tissue layers. Most of the existing literature has manually separated 20 different regions of the tissue slice, while SEDIST successfully identified all 20 clusters with an ARI score of 0.6185. 3.5 Development Environment and Computational Benchmarking SEDIST was developed and tested on a MacBook Pro with the following hardware configuration: Model: MacBook Pro M2 Pro 14-Inch, Processor: 10-Core CPU, Graphics: 16- Core GPU, Memory: 16GB RAM, Storage: 512GB SSD. The powerful combination of a 10-core CPU and a 16-core GPU provides significant computational capabilities for both the training and testing phases. The 16GB RAM ensures efficient handling of data and model parameters, and the 512GB SSD offers ample storage for datasets and model checkpoints. This hardware configuration offers a robust and efficient environment for developing, training, and evaluating SEDIST. In addition to evaluating the clustering performance of SEDIST against state-of-the-art (SOTA) methods based on Adjusted Rand Index (ARI), we conducted a comparative analysis of their computational efficiency in terms of time and memory usage. Table 1 summarizes the execution time (in seconds) and memory consumption (in megabytes) for SEDIST, GraphST, DeepST, SpaGCN, GRAS4T, GAAEST, and MNMST. SEDIST demonstrated competitive time efficiency, completing the analysis in 121.45 seconds. This was faster than DeepST (261.45 seconds), SpaGCN (161.45 seconds), GRAS4T (189.07 seconds), and MNMST (776.2 seconds). However, SEDIST required slightly more time than GraphST (90.45 seconds) and GAAEST (109.3 seconds). Among the evaluated methods, MNMST was the slowest, requiring 776.2 seconds, which is more than six times the execution time of SEDIST. Conversely, GraphST and GAAEST were the fastest methods, with execution times of 90.45 and 109.3 seconds, respectively. Despite these differences, SEDIST’s execution time strikes a balance between computational speed and its superior clustering accuracy, making it a viable choice for applications where both precision and efficiency are critical. In terms of memory usage, SEDIST required 2195.3 MB, demonstrating a favorable balance between computational demands and performance. It used significantly less memory than GAAEST (5646.06 MB) and DeepST (3295.56 MB), both of which exhibited the highest memory consumption among the methods evaluated. SpaGCN and GraphST also consumed more memory than SEDIST, requiring 2469.46 MB and 2288.67 MB, respectively. Table 1 Time and Memory benchmark for different State-of-the-Art methods against SEDIST. SEDIST GraphST DeepST SpaGCN GRAS4T GAAEST MNMST Time (sec) 121.45 90.45 261.45 161.45 189.07 109.3 776.2 Memory (mb) 2195.3 2288.67 3295.56 2469.46 1663.48 5646.06 MB 1761.28 GRAS4T and MNMST, in contrast, exhibited lower memory consumption than SEDIST, with GRAS4T requiring the least memory at 1663.48 MB, followed by MNMST at 1761.28 MB. While these methods demonstrated efficiency in memory usage, their clustering accuracy was generally lower compared to SEDIST, as discussed earlier. The results presented in Table 1 highlight SEDIST’s balanced computational profile. While slightly slower than GraphST and GAAEST, SEDIST outperformed these methods in clustering accuracy, as shown in Fig. 2 . Moreover, SEDIST's memory usage is significantly more efficient than that of GAAEST and DeepST, making it a more practical choice for large-scale datasets or resource-constrained environments. This analysis underscores the trade-offs between computational efficiency and clustering performance among the methods. SEDIST’s ability to deliver superior clustering accuracy, coupled with reasonable time and memory requirements, further establishes its robustness and applicability in spatial clustering tasks. 4 Discussion Deciphering the activity within a complex cellular microenvironment is crucial. Also, the identification of spatial domains is significant for understanding these interactions inside a multicellular organism. In this study, we introduce SEDIST, a novel deep learning-based approach designed to identify spatial domains by integrating gene expression profiles with spatial information. Initially, the morphological image is segmented into spots based on H&E staining, from which high-level morphological features are extracted using an autoencoder. By incorporating spatial locations along with gene expression profiles and morphological features, we constructed and enhanced an undirected neighborhood graph, then applied a graph encoder to unveil latent features. Finally, we utilized self-supervised contrastive learning to refine these features and employed the MCLUST method to cluster the spatial domains based on their morphological and spatial characteristics. To validate the performance of SEDIST, a comprehensive comparative analysis is conducted with state-of-the-art method such as GraphST, DeepST, ConST, SpaceFlow, SpaGCN. The performance of identifying the spatial domain is evaluated using Adjusted Rand Index (ARI) and SEDIST provides more stable and comparatively better performances in most of the cases. For slice 151673, SEDIST achieves an ARI of 0.6466, outperforming stLearn (0.3031), SpaGCN (0.4557), DeepST (0.5855), and GraphST (0.6123). Furthermore, the proposed method demonstrates superior performance most of the cases across multiple slices, including 151507, 151508, 151509, 151669, 151670, 151671, 151672, 151674, 151675, and 151676, particularly within the Dorsolateral Prefrontal Cortex (DLPFC). This consistent improvement highlights the robustness and effectiveness of SEDIST in comparison to existing approaches. Another key strength of SEDIST is that it employs a 3D embedding technique for clustering gene expression data, where this approach integrates spatial information with gene expression profiles, allowing for a more comprehensive understanding of transcriptomics patterns and improving the accuracy of downstream analysis. By leveraging the spatial context, the method enhances the interpretability and precision of gene expression clustering. The integration of diffusion-based visualization approaches significantly improved the interpretability, also offering a holistic view on the spatial and transcriptional organization within cortical layers. To understand tissue organization and biological functions, it is important to accurately identify spatial domains and extract spatially expressed genes. SEDIST allows for the identification and visualization of highly expressed genes within spatial domains. Moreover, the top 10 highly expressed genes were shown within slice 151673. Several genes were identified as playing critical roles in key biological processes, including cell proliferation, axon recognition, axon-glia signaling, differentiation, membrane component transport, myelin compaction, and node formation. Distinct gene expression patterns revealed functional specializations and characteristics, providing valuable insights into the organization and biological activity of cortical regions In terms of model stability, SEDIST consistently identifies spatial domains that align with ground truth data, outperforming existing methods. However, some discrepancies were observed between our reproduced results and those reported in the original literature, likely due to differences in hardware and the convergence challenges inherent to unsupervised methods. To address this limitation, all comparative methods were evaluated under the same hardware and software conditions, ensuring a fair and consistent comparison. In summary, SEDIST is a novel deep learning-based approach that modifies feature representation for each spot to identify spatial domains. As the generation of spatial transcriptomics (ST) data continues to grow, it is anticipated that SEDIST will contribute to revealing new insights into cellular organization within a spatial context. Therefore, optimizing memory efficiency through mini-batch processing and parallel computing in distributed learning systems can be an important enhancement to spatial transcriptomics (ST) studies. Additionally, integrating spatial omics and scRNA-seq data can further enhance the resolution of spatial transcriptomics results and facilitate automatic annotation of spatial domains. Declarations Clinical trial number: Not applicable. Ethics approval and consent to participate: Not applicable. Consent for publication: Not applicable. Availability of data and material: All the data and codes used in the methods are available at https://github.com/SEDIST-014/SEDIST. Competing interests: None declared. Funding: Funded by the Institute for Advanced Research Publication Grant of United International University, Ref. No.: IAR-2025-Pub-033. Authors' contributions: FR contributed to conceptualization, methodology, software development, visualization, writing – original draft, investigation, formal analysis, and validation. AAS was involved in conceptualization, methodology, software development, visualization, writing – original draft, investigation, formal analysis, and validation. RA was responsible for supervision, conceptualization, methodology, formal analysis, writing – review & editing, investigation, and validation. FA contributed to validation, formal analysis, writing – original draft, writing – review & editing, investigation, project administration, and funding acquisition. SAE contributed to writing – original draft and visualization. UAB contributed to writing – original draft and visualization. Acknowledgement: None Declared References Maynard KR, Collado-Torres L, Weber LM, Uytingco C, Barry BK, Williams SR, et al. Transcriptome-scale spatial gene expression in the human dorsolateral prefrontal cortex. Nat Neurosci. 2021;24:425–36. Wang Y, Liu B, Zhao G, Lee Y, Buzdin A, Mu X, et al. Spatial transcriptomics: Technologies, applications and experimental considerations. Genomics. 2023;115:110671. Shengquan C, Boheng Z, Xiaoyang C, Xuegong Z, Rui J. stPlus: a reference-based method for the accurate enhancement of spatial transcriptomics. Bioinformatics. 2021;37 Supplement_1:i299–307. Satija R, Farrell JA, Gennert D, Schier AF, Regev A. Spatial reconstruction of single-cell gene expression data. Nat Biotechnol. 2015;33:495–502. Abdelaal T, Mourragui S, Mahfouz A, Reinders MJT. SpaGE: Spatial Gene Enhancement using scRNA-seq. Nucleic Acids Res. 2020;48:e107. Liu J, Gao C, Sodicoff J, Kozareva V, Macosko EZ, Welch JD. Jointly defining cell types from multiple single-cell datasets using LIGER. Nat Protoc. 2020;15:3632–62. Teng H, Yuan Y, Bar-Joseph Z. Clustering spatial transcriptomics data. Bioinformatics. 2022;38:997–1004. Dong K, Zhang S. Deciphering spatial domains from spatially resolved transcriptomics with an adaptive graph attention auto-encoder. Nat Commun. 2022;13:1739. Wolf FA, Angerer P, Theis FJ. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 2018;19:15. Xu H, Fu H, Long Y, Ang KS, Sethi R, Chong K, et al. Unsupervised spatially embedded deep representation of spatial transcriptomics. Genome Med. 2024;16:12. Long Y, Ang KS, Li M, Chong KLK, Sethi R, Zhong C, et al. Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST. Nat Commun. 2023;14:1155. Zhao E, Stone MR, Ren X, Guenthoer J, Smythe KS, Pulliam T, et al. Spatial transcriptomics at subspot resolution with BayesSpace. Nat Biotechnol. 2021;39:1375–84. Hu J, Li X, Coleman K, Schroeder A, Ma N, Irwin DJ, et al. SpaGCN: Integrating gene expression, spatial location and histology to identify spatial domains and spatially variable genes by graph convolutional network. Nat Methods. 2021;18:1342–51. Xu C, Jin X, Wei S, Wang P, Luo M, Xu Z, et al. DeepST: identifying spatial domains in spatial transcriptomics by deep learning. Nucleic Acids Res. 2022;50:e131. Zong Y, Yu T, Wang X, Wang Y, Hu Z, Li Y. conST: an interpretable multi-modal contrastive learning framework for spatial transcriptomics. 2022. Cao Y, Lin Y, Patrick E, Yang P, Yang JYH. scFeatures: multi-view representations of single-cell and spatial data for disease outcome prediction. Bioinformatics. 2022;38:4745–53. Raredon MSB, Yang J, Kothapalli N, Lewis W, Kaminski N, Niklason LE, et al. Comprehensive visualization of cell-cell interactions in single-cell and spatial transcriptomics with NICHES. Bioinformatics. 2023;39:btac775. Zhang Z, Cui F, Su W, Dou L, Xu A, Cao C, et al. webSCST: an interactive web application for single-cell RNA-sequencing data and spatial transcriptomic data integration. Bioinformatics. 2022;38:3488–9. Saviano A, Henderson NC, Baumert TF. Single-cell genomics and spatial transcriptomics: Discovery of novel cell states and cellular interactions in liver physiology and disease biology. J Hepatol. 2020;73:1219–30. Hou W, Ji Z. Palo: spatially aware color palette optimization for single-cell and spatial data. Bioinformatics. 2022;38:3654–6. Pham D, Tan X, Balderson B, Xu J, Grice LF, Yoon S, et al. Robust mapping of spatiotemporal trajectories and cell-cell interactions in healthy and diseased tissues. Nat Commun. 2023;14:7739. Gui Y, Li C, Xu Y. Spatial domains identification in spatial transcriptomics using modality-aware and subspace-enhanced graph contrastive learning. Comput Struct Biotechnol J. 2024;23:3703–13. Wang T, Zhu H, Zhou Y, Ding W, Ding W, Han L, et al. Graph attention automatic encoder based on contrastive learning for domain recognition of spatial transcriptomics. Commun Biol. 2024;7:1351. Wang Y, Liu Z, Ma X. MNMST: topology of cell networks leverages identification of spatial domains from spatial transcriptomics data. Genome Biol. 2024;25:133. Scrucca L, Fop M, Murphy TB, Raftery AE. mclust 5: Clustering, Classification and Density Estimation Using Gaussian Finite Mixture Models. R J. 2016;8:289–317. Nagy C, Maitra M, Tanti A, Suderman M, Théroux J-F, Davoli MA, et al. Single-nucleus transcriptomics of the prefrontal cortex in major depressive disorder implicates oligodendrocyte precursor cells and excitatory neurons. Nat Neurosci. 2020;23:771–81. Chelebian E, Avenel C, Kartasalo K, Marklund M, Tanoglidi A, Mirtti T, et al. Morphological Features Extracted by AI Associated with Spatial Transcriptomics in Prostate Cancer. Cancers (Basel). 2021;13:4837. Bergenstråhle J, Larsson L, Lundeberg J. Seamless integration of image and molecular analysis for spatial transcriptomics workflows. BMC Genomics. 2020;21:482. Bae S, Choi H, Lee DS. Discovery of molecular features underlying the morphological landscape by integrating spatial transcriptomic data with deep features of tissue images. Nucleic Acids Res. 2021;49:e55. Kleshchevnikov V, Shmatko A, Dann E, Aivazidis A, King HW, Li T, et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat Biotechnol. 2022;40:661–71. Song Q, Su J. DSTG: deconvoluting spatial transcriptomics data through graph-based artificial intelligence. Brief Bioinform. 2021;22:bbaa414. Ratajczak F, Joblin M, Hildebrandt M, Ringsquandl M, Falter-Braun P, Heinig M. Speos: an ensemble graph representation learning framework to predict core gene candidates for complex diseases. Nat Commun. 2023;14:7206. Veličković P, Fedus W, Hamilton WL, Liò P, Bengio Y, Hjelm RD. Deep Graph Infomax. In: 7th International Conference on Learning Representations, ICLR 2019. New Orleans, Louisiana; 2018. Kipf TN, Welling M. Semi-Supervised Classification with Graph Convolutional Networks. In: 5th International Conference on Learning Representations, ICLR 2017 - Conference Track Proceedings. Toulon, France; 2016. Chen W-T, Lu A, Craessaerts K, Pavie B, Sala Frigerio C, Corthout N, et al. Spatial Transcriptomics and In Situ Sequencing to Study Alzheimer’s Disease. Cell. 2020;182:976–e99119. Kleshchevnikov V, Shmatko A, Dann E, Aivazidis A, King HW, Li T et al. Compr Mapp tissue cell Archit via Integr single cell Spat transcriptomics bioRxiv. 2020. Wang X, Almet AA, Nie Q. The promising application of cell-cell interaction analysis in cancer from single-cell and spatial transcriptomics. Semin Cancer Biol. 2023;95:42–51. Yuan Y, Bar-Joseph Z. GCNG: graph convolutional networks for inferring gene interaction from spatial transcriptomics data. Genome Biol. 2020;21:300. Song J, Lamstein J, Ramaswamy VG, Webb M, Zada G, Finkbeiner S, et al. Enhancing Spatial Transcriptomics Analysis by Integrating Image-Aware Deep Learning Methods. Biocomputing 2024. WORLD SCIENTIFIC; 2023. pp. 450–63. Moncada R, Barkley D, Wagner F, Chiodin M, Devlin JC, Baron M, et al. Integrating microarray-based spatial transcriptomics and single-cell RNA-seq reveals tissue architecture in pancreatic ductal adenocarcinomas. Nat Biotechnol. 2020;38:333–42. Li J, Chen S, Pan X, Yuan Y, Shen H-B. Cell clustering for spatial transcriptomics data with graph neural networks. Nat Comput Sci. 2022;2:399–408. McInnes L, Healy J, Saul N, Großberger L. UMAP: Uniform Manifold Approximation and Projection. J Open Source Softw. 2018;3:861. Lábár JL, DiffMap. A new free computer program to process scanned electron diffraction patterns. Resolut Discov. 2022;6:7–11. van Ooijen JW. DrawMap: a computer program for drawing genetic linkage maps. J Hered. 1994;85:66. Wu C, Tu T, Xie M, Wang Y, Yan B, Gong Y, et al. Spatially resolved transcriptome of the aging mouse brain. Aging Cell. 2024;23:e14109. Weger M, Alpern D, Cherix A, Ghosal S, Grosse J, Russeil J, et al. Mitochondrial gene signature in the prefrontal cortex for differential susceptibility to chronic stress. Sci Rep. 2020;10:18308. Nicholls C, Li H, Liu J-P. GAPDH: a common enzyme with uncommon functions. Clin Exp Pharmacol Physiol. 2012;39:674–9. Dai S, Wang H, Lin Z. ACTB Mutations Analysis and Genotype–Phenotype Correlation in Becker’s Nevus. Biomedicines. 2021;9:1879. Geng W, Ren J, Shi H, Qin F, Xu X, Xiao S, et al. RPL41 sensitizes retinoblastoma cells to chemotherapeutic drugs via ATF4 degradation. J Cell Physiol. 2021;236:2214–25. Herbet M, Piątkowska-Chmiel I, Motylska M, Gawrońska-Grzywacz M, Nieradko-Iwanicka B, Dudka J. Alteration in the Expression of Genes Involved in Cerebral Glucose Metabolism as a Process of Adaptation to Stressful Conditions. Brain Sci. 2022;12:498. Kang Y, Wu Z, Cai D, Lu B. Evaluation of reference genes for gene expression studies in mouse and N2a cell ischemic stroke models using quantitative real-time PCR. BMC Neurosci. 2018;19:3. Tian Y, Lu J, Hao X, Li H, Zhang G, Liu X, et al. FTH1 Inhibits Ferroptosis Through Ferritinophagy in the 6-OHDA Model of Parkinson’s Disease. Neurotherapeutics. 2020;17:1796–812. Christopher JR, Ponnaiyan D, Parthasarathy H, Tadepalli A. Association of CST3 Gene with Its Protein: Cystatin C in Health and Severe Periodontal Disease. Genet Test Mol Biomarkers. 2021;25:405–10. Zhang Y, Lv X, Chen L, Liu Y. The role and function of CLU in cancer biology and therapy. Clin Exp Med. 2022;23:1375–91. Liu T, Han X, Zheng S, Liu Q, Tuerxun A, Zhang Q, et al. CALM1 promotes progression and dampens chemosensitivity to EGFR inhibitor in esophageal squamous cell carcinoma. Cancer Cell Int. 2021;21:121. Wang C, He Y, You Z, Chen X. TMSB10 Promotes Progression of Clear Cell Renal Cell Carcinoma via JUN Transcription Regulation. Ann Clin Lab Sci. 2022;52:230–9. Additional Declarations No competing interests reported. Supplementary Files SupplementaryDocument.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6619020","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":457970546,"identity":"6141e5f1-c9ce-4ba7-a92e-1862035da4be","order_by":0,"name":"Fahad Rahman","email":"","orcid":"","institution":"United International University","correspondingAuthor":false,"prefix":"","firstName":"Fahad","middleName":"","lastName":"Rahman","suffix":""},{"id":457970547,"identity":"cb082f6b-b311-4de6-b2f8-01eccc4a54cf","order_by":1,"name":"Abdullah Al Sakib","email":"","orcid":"","institution":"United International University","correspondingAuthor":false,"prefix":"","firstName":"Abdullah","middleName":"Al","lastName":"Sakib","suffix":""},{"id":457970548,"identity":"e0b1ea8d-f19a-4dc5-9033-921536f6a73a","order_by":2,"name":"Riasat Azim","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIiWNgGAWjYDACZjBpA+EkgEkeBgnCWhLSSNECUXoYmUdAi3k78+MPP3+cl5dvb2CTeNjGIM/fwHvwBj4tMofZzCR7Em4bbjhzgE0isY3BcMYBvmQLfFokmBnMGHgSbicYSCSAtTBuYOAxw+swCWb2zx//JJxLkJ//AKzFnggtPAbSPAkHEhhuMIC1JBKjpUxaJi0Z6JfEZouEcxLJMw4T8gv/8c0f39jYAUPs8MGbP8psbPvbe/GHGBJgbAAZAUsPo2AUjIJRMAooAQBB0T3HJvV/ugAAAABJRU5ErkJggg==","orcid":"","institution":"United International University","correspondingAuthor":true,"prefix":"","firstName":"Riasat","middleName":"","lastName":"Azim","suffix":""},{"id":457970549,"identity":"dd5607e4-738d-482d-9fe2-b17de2eb3e21","order_by":3,"name":"Fahim Hafiz","email":"","orcid":"","institution":"United International University","correspondingAuthor":false,"prefix":"","firstName":"Fahim","middleName":"","lastName":"Hafiz","suffix":""},{"id":457970550,"identity":"99a29fa6-78c9-45c5-bad1-84fcce275d86","order_by":4,"name":"Shabrina Airen Esha","email":"","orcid":"","institution":"United International University","correspondingAuthor":false,"prefix":"","firstName":"Shabrina","middleName":"Airen","lastName":"Esha","suffix":""},{"id":457970554,"identity":"7a2b6976-8c67-4877-a0e0-9f914b12311c","order_by":5,"name":"Umme Aimon Bristy","email":"","orcid":"","institution":"United International University","correspondingAuthor":false,"prefix":"","firstName":"Umme","middleName":"Aimon","lastName":"Bristy","suffix":""}],"badges":[],"createdAt":"2025-05-08 09:23:35","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6619020/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6619020/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83018497,"identity":"76a935d0-5574-49c1-9694-b18f161b8ab6","added_by":"auto","created_at":"2025-05-19 06:54:16","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4121437,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of the SEDIST algorithm. (A) A feature vector for each spot tile was constructed using an autoencoder. (B) An initial graph is constructed from morphological locations and embedded PCAs generated from the Image matrix and spatial gene expressions are integrated with the initial graph to construct an enhanced modified graph. (C) Enhanced and accurate matrix representations containing the spatial gene expression are generated using two graph encoders where the weights are shared between the encoders. (D) Applying contrastive learning on original and modified matrix representation, a final matrix representation is generated to predict the clusters in the spatial locations.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/a1421efabbb084dc284cdd06.png"},{"id":83018471,"identity":"eb6d3122-5048-4238-8b66-32cd8c2537a8","added_by":"auto","created_at":"2025-05-19 06:54:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2771687,"visible":true,"origin":"","legend":"\u003cp\u003eSEDIST clustering improves the identification of tissue structures in the human dorsolateral prefrontal cortex (DLPFC) . SEDIST enhances the recognition of spatial domains within tissue. (A) The layers of the DLPFC were labeled by Maynard et al. The accurate positions of spots on slide 151673 were aligned with the ground truth, separating them into six cortical layers (L1–L6) and white matter (WM). (B) SEDIST’s ability to identify spatial domains was compared with other advanced algorithms GraphST, DeepST, conST, GRAS4T, GAAEST, MNMST, SpaGCN and SpaceFlow for slide 151673. SEDIST enhances the recognition of Spatial Organization and Genetic Dynamics. (C-L) Comparing the ARI of the SEDIST method with the eight state-of-the-art (SOTA) techniques: GraphST, DeepST, conST, GRAS4T, GAAEST, MNMST, SpaGCN and SpaceFlow focusing on slices 151507, 151508, 151509, 151510, 151669, 151670, 151671, 151672, 151673, 151674, 151675 and 151676.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/0e391a5a0419cd597caaf9f3.png"},{"id":83018489,"identity":"5c69be24-8fe0-4551-a8c2-304c243292c6","added_by":"auto","created_at":"2025-05-19 06:54:16","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":9444576,"visible":true,"origin":"","legend":"\u003cp\u003eDifferent visualization perspectives demonstrating SEDIST's accuracy in capturing spatial relationships (A) 3D visualizations were generated for slide 151673, displaying both the ground truth (left) and the embeddings derived from SEDIST (right). (B) UMAP visualizations were created for slide 151673, showing the embeddings derived from SEDIST (color-coded). (C) Diffmap projections in two diffusion component were crafted for slide 151673, exhibiting both the ground truth (left) and the embeddings derived from SEDIST (right). (D) Draw Graph visualizations in two functional regions were crafted for slide 151673, exhibiting both the ground truth (left) and the embedding derived from SEDIST (right). (E) Expression pattern of spatially variable genes in different regions of the tissue.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/7b0a54d2cfb2c35bd9fa18ca.png"},{"id":83018500,"identity":"3fcab2c6-1a40-4f27-bd7c-5638fe2adcd9","added_by":"auto","created_at":"2025-05-19 06:54:16","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":278109,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of the Top 10 Gene’s expression in different Cortical Layers of Slice 151673.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/81a61de2b6c651f4edb4a867.png"},{"id":83018494,"identity":"84f9b75d-af5b-4dca-8ca5-1485be3fcb23","added_by":"auto","created_at":"2025-05-19 06:54:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":3930841,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial clustering performance of SEDIST on the BRCA dataset, demonstrating 20 clusters. SEDIST achieved an ARI score of 0.6185when compared to the ground truth.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/41f2556c7f2209d2b8867e84.png"},{"id":94063975,"identity":"40416db0-a84f-4396-b0b8-4ff34cc25092","added_by":"auto","created_at":"2025-10-22 07:31:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":25510099,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/c8de8b11-f6ab-4c96-8b3b-bcd13f90fe19.pdf"},{"id":83018501,"identity":"42313314-2fce-40d4-92ff-231954d63281","added_by":"auto","created_at":"2025-05-19 06:54:17","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1013033,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryDocument.docx","url":"https://assets-eu.researchsquare.com/files/rs-6619020/v1/93023eaa53f6d27d266da7da.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"SEDIST: Spatially Enhanced Domain Identification through Spatial Transcriptomics","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eRecent technological advances in spatial transcriptomics(ST) have brought new opportunities for biomarker identification, spatial heterogeneity analysis and characterization of gene expression while taking into consideration the spatial location of gene expression [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The spatial domains and their associations are often responsible for the tissue microenvironments and their biological functions. The ways in which complex biological tissue functions are closely related to the spatial context of the cells.\u003c/p\u003e \u003cp\u003eUnderstanding gene functions is crucial to discovering activities in different organs, and understanding disease pathways. Traditional techniques such as bulk RNA sequencing (RNA-seq) and single-cell RNA sequencing (scRNA-seq) have significantly advanced our knowledge by identifying which genes are expressed. However, these methods lack spatial resolution, meaning that they do not provide information about where gene expression occurs in the tissue. This limitation is critical, as the spatial context of gene expression plays a key role in cellular function, organization, and communication. Spatial transcriptomics is an emerging technology that overcomes this limitation by enabling the simultaneous measurement of gene expression levels and their spatial locations within tissues. By integrating spatial data with gene expression profiles, this technique allows researchers to gain deeper insights into cell-to-cell interactions, tissue architecture and disease mechanisms at the cellular level. As a result, spatial transcriptomics offers a more comprehensive understanding of biological processes and has the potential to advance both basic research and clinical applications.\u003c/p\u003e \u003cp\u003eExperimental technology in spatial transcriptomics can be broadly categorized into two types: (i) imaging-based technologies and (ii) sequencing-based technologies. Imaging-based technologies use highly multiplexed adaptations of single-molecule fluorescence in situ hybridization (smFISH), such as sequential FISH (seqFISH) and MERFISH along with histology, molecular, and imaging tools without relying on sequencing methods [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].Molecular cartography, MERSCOPE, and 10x Xenium are examples of imaging-based methods that are adopted from sequential FISH (seqFISH), MERFISH, and 10x Genomics technology. These imaging-based methods provide single-cell or subcellular level resolution such as Molecular Cartography (275nm), MERSCOPE (100nm), and Xenium (50nm). However, they can only provide the profile of a few hundred to a few thousand genes due to various technical limitations [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Conversely, the sequencing-based methodology includes 10x Visium, GeoMx DSP and Stereo-seq.\u0026nbsp;These methods combine spatial barcoding with next-generation sequencing technology to preserve spatial context. Although these methods can profile the whole transcriptome, the spatial resolution varies across them such as 10x Visium (100 \u0026micro;m), GeoMx DSP (10 \u0026micro;m) and Stereo-seq (0.5 \u0026micro;m).\u003c/p\u003e \u003cp\u003eRegardless of the spatial transcriptomics methodology employed, the initial step in any spatial expression analysis workflow is to identify and define the cell types within the 2D tissue sample. The most essential task of spatial transcriptomics is the identification of domains and cell types. Further downstream analysis and understanding of the tissue microenvironment often rely on the identification of the cell types. Several methodologies have been constructed to facilitate such clustering for single-cell spatial transcriptomics data. These methods usually convert the high-dimensional data into low low-dimensional space and cluster the cells. Furthermore, clusters are categorized into known or novel cell types by analyzing the expression of a certain group of marker genes.\u003c/p\u003e \u003cp\u003estPlus proposed an autoencoder-based framework with a tailored loss function to enhance the spatial transcriptomics data by leveraging scRNA-seq data as a reference [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. This proposed method outperformed other methods, such as Seurat [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], SpaGE [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], and Liger [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] in predicting gene expression in spatial transcriptomics data. Furthermore, it enables better identification of cell populations and more accurate segmentation of tissue sections. This demonstrates that integrating additional data with spatial transcriptomics data enhances the accuracy of the analysis. FICT proposed a novel clustering method designed for spatial transcriptomics data by integrating both gene expression and spatial information to improve cell-type assignment. It demonstrated superior performance on both simulated and real datasets such as MERFISH, osmFISH, and seqFISH, than other methods that use only gene expression or partial spatial information [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. STAGATE analyzed spatial transcriptomics data by identifying spatial domains and tissue architectures, particularly in brain tissues [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. To better identify genes that are spatially expressed and reduce noise, STAGATE integrates spatial information with gene expression profiles using a graph attention method. It builds a spatial neighbor network and learns low-dimensional representations using gene expression and geographic data via a graph attention auto-encoder. The approach outperforms SOTA techniques such as SCANPY [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and SEDR [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] in uncovering tissue architecture and laminar organization across many datasets. GraphST used spatial location and spatial gene expression for spatial clustering, multi-sample integration, and integration of spatial transcriptomics and scRNA-seq [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. The authors proposed augmentation-based contrastive learning by creating a corrupted graph by random shuffling of gene expression across spots. This allows the self-supervised contrastive model to capture the local context surrounding the environment. Then GNN encoder is used to learn the latent representation of the corrupted and uncorrupted data. The local context in the latent representation helps to determine accurate gene expression and graphs that are reconstructed using a decoder. This graph self-supervised contrastive learning approach in clustering achieves 10% higher accuracy than other methods such as STAGATE [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], BayesSpace [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], SpaGCN [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], DeepST[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], GraphST[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] and ConST [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. scFeatures [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] published an R package designed to generate a multi-view molecular representation of samples by creating interpretable features from scRNA-seq and spatial transcriptomics data, to enhance the understanding of disease outcomes. The findings indicate that different feature types are useful for different datasets. Another study, NICHES [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], integrates the detailed gene expression profiles from scRNA-seq with the spatial context provided by spatial transcriptomics. Constructing matrices that capture ligand-receptor interactions using the integrated data and visualizing these interactions creates a detailed map of how cells communicate within their local environment, offering a deeper understanding of cell-cell interactions in tissues. Similar approaches of integrating spatial transcriptomics data with scRNA-seq data, interactive analysis, and visualization of the integrated data have been discussed in [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e][\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e][\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAnother approach to analyzing spatial transcriptomics data is utilizing histology information along with gene expression and spatial data. stLearn [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] integrates gene expression, spatial location, and histology data to cluster similar regions within tissue sections, providing a basic understanding of tissue organization. While it combines these data sources to identify spatial domains, it treats spatial domain identification and the detection of spatially variable genes (SVGs) as separate tasks. This limits stLearn's ability to link gene expression patterns with histology, making it difficult to understand how differences in cell types relate to the overall organization of the tissue. SpaGCN [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], on the other hand, addresses this issue by integrating spatial domain identification and SVG detection into a unified approach while utilizing scRNA-seq data, spatial location, and histology. In another study, DeepST [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] integrated the spatial location, gene expression, and tissue image to enhance the identification of spatial domains in samples. Image (H\u0026amp;E staining tiles) is segmented according to the coordinates of the spots, and consecutively, a pre-trained convolutional neural network (CNN) is used to transform each spot into 2048 latent variables, creating an image matrix. Additionally, graph representations of the spatial gene expression are used to determine the adjacency matrix between spots. These image matrices, the adjacency matrix, and gene expression are used to determine enhanced gene expression. A denoising autoencoder with a variable graph autoencoder is used to determine the final reconstructed spatial gene expression with higher accuracy compared to SpaGCN, stLearn, SEDR, BayesSpace, DeepST, ConSt, and GraphST. Furthermore, we also benchmarked some recent spatial transcriptomics methods such as: GRAS4T[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], GAAEST [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], and MNMST [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] with the SEDIST.\u003c/p\u003e \u003cp\u003eBy addressing limitations in feature extraction, spatial integration, and computational inefficiencies, we proposed this deep learning-based method named SEDIST to accurately identify spatial domains from spatial transcriptomics data. Initially, we extract features of tissue image tiles from the morphological image utilizing an autoencoder architecture. A feature vector was created for each spot of the tissue, and integration with the gene expressions and spatial locations was performed to create enhanced representations of the spatial domains. Furthermore, we converted these integrated representations into a graph where the graph undergoes a shuffling process along with randomness for enhanced robustness of the graph data. Moreover, this modified graph is passed through a graph autoencoder to learn an enhanced matrix representation of the graph. To achieve the final matrix representation, a self-contrastive mechanism is applied to project the gene expression data into a spatial space by emphasizing the spot similarities. After that, the final representation is used to cluster the spatial domains using Mclust. In order to ensure alignment with ground truth data, the number of clusters is decided by manually annotating tissue slices. Utilizing human dorsolateral prefrontal cortex (DLPFC) data from the LIBD repository and Human Breast Cancer (BRCA) dataset, SEDIST is tested extensively and provides better accuracy in identifying spatial domains compared to other SOTA methods. Another extensive feature of this method named SEDIST is capable of presenting the 2D embedding in a 3D spatial space. SEDIST effectively maintains consistency across multiple datasets, such as Visium 10x, standard Visium. This method\u0026rsquo;s consistent performance across different datasets demonstrates its flexibility and robustness. Additionally, balances clustering accuracy with computational efficiency, optimizing execution time and memory usage for large-scale datasets and resource-constrained environments. Furthermore, it showed more computational effectiveness compared to the other recent methods, such as GraphST, SpaGCN, DeepST, GAAEST, GRAS4T, and MNMST. In summary, the Findings we present in this paper underscore the robustness of SEDIST in accurately identifying spatial domains, while also highlighting its scalability for processing additional spatial omics data.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\n\u003ch3\u003e2.1 Overview of SEDIST\u003c/h3\u003e\n\u003cp\u003eThe proposed method SEDIST, consists of four levels of architecture, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Initially, an H\u0026amp;E-stained image undergoes segmentation, resulting in its division into multiple sections, which are known as spots. Subsequently, an autoencoder is applied to discern the underlying patterns, generating an image feature vector. By utilizing spatial locations, a graph is constructed using the distance measure. Subsequently, the feature vector, along with the image feature vector and corresponding gene expressions of specific spots undergoes feature extraction through a scalar function and principal component analysis (PCA). Here, log normalization is applied to the datasets as a scalar function. Such preprocessing step is crucial before applying PCA as it ensures that no single feature disproportionately influences the PCA due to its magnitude, thereby giving equal importance to all the features.\u003c/p\u003e \u003cp\u003eFrom the resultant combined features and undirected neighborhood graph, a modified graph is constructed which subsequently goes undergoes a shuffling process to introduce randomness and enhance robustness. Spatial information is then incorporated to spatially map the graph, aligning it with the original spatial distribution of spots. After that, a graph encoder is subsequently deployed to unveil latent features from graphs and transform them into two enhanced matrix representations. We implemented self-supervised contrastive learning (SSCL) to those matrix representations to increase the correctness of adjacent or neighborhoods and converted them into a final matrix representation. Finally, the modified matrix representation is utilized using the nonspatial assignment algorithm Mclust[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] for the clustering process, which is specialized in clustering spatial domains through expectation maximization .\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003e2.2 Data Collection and Preprocessing\u003c/h3\u003e\n\u003cp\u003eIn SEDIST, the human dorsolateral prefrontal cortex (DLPFC) and BRCA (human breast cancer) dataset is used which is collected from the LIBD repository [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] and SEDR [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. The DLPFC dataset comprises 12 tissue slices obtained through the 10x Visium platforms. Each tissue slice has a varying number of spots, ranging from 3460 to 4789, and captures information from 33,538 genes. Manual annotations were applied to categorize each slice into distinct regions, including DLPFC layers and white matter. DLPFC datasets consisted of cell type composition, spatial organization, and gene expression profiles within each spatial spot, utilizing deconvolution techniques in spatial transcriptomics (ST) using scRNA-seq reference data were employed. In one instance, we utilized slice #151673 from the human brain sample DLPFC dataset. The corresponding scRNA-seq data for the dorsolateral prefrontal cortex (BA9) tissue, obtained from postmortem archives, were generated using the 10x Genomics Chromium platform [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. This scRNA-seq dataset comprises 78,886 cells and captures gene expression from 30,062 genes. The BRCA dataset consists of 3799 cells having 20 layers. All data utilized in the experiment and validation process are presented in \u003cb\u003eSupplementary Table S2\u003c/b\u003e. To facilitate a comprehensive comparison, additional slices from the DLPFC dataset were utilized for assessing the performance of the proposed method against other existing state-of-the-art (SOTA) methods.\u003c/p\u003e\n\u003ch3\u003e2.3 Image Matrix generation\u003c/h3\u003e\n\u003cp\u003eUsing H \u0026amp; E staining, the morphological image is divided into smaller sections or images according to the spatial coordinates of each spot. Spot tiles are augmented and transformation method such as rotation, normalization, sharpness adjustment are applied [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Subsequently, an autoencoder is applied to extract each spot tile\u0026rsquo;s high-level features. The dense layer in the autoencoder began at 1024. Each dense layer of the autoencoder contains a number of fixed morphological features. Extracted a total of 14400 high-level morphological features from the autoencoder\u0026rsquo;s output in this instance [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e][\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Finally, the feature weights of morphological similarity \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{MS}_{ij}\\)\u003c/span\u003e\u003c/span\u003e between spot \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{S}_{i}\\)\u003c/span\u003e\u003c/span\u003e and adjacent \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{S}_{j}\\)\u003c/span\u003e\u003c/span\u003e were calculated using the cosine distance as follows Eq.\u0026nbsp;\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{MS}_{ij}=1-\\frac{{S}_{i}\\cdot\\:{S}_{j}}{{\\Vert\\:{S}_{i}\\Vert\\:}_{2}{\\Vert\\:{S}_{j}\\Vert\\:}_{2}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003e2.4 Graph Construction\u003c/h3\u003e\n\u003cp\u003eUsing a predetermined number of K\u0026thinsp;=\u0026thinsp;3 neighbors, we created an undirected neighborhood graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:G=\\left(V,\\:E\\right)\\:\\)\u003c/span\u003e\u003c/span\u003eto employ the spatial data fully. The set of edges that connect the neighborhood spots represented by E to V, the collection of spots, is represented in graph G. The adjacency matrix of graph G is defined as follows:\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{}\\text{(}A\\in\\:{\\mathbb{R}}^{{N}_{\\text{spot}}\\times\\:{N}_{\\text{spot}}}\\text{)},\\:\\)\u003c/span\u003e\u003c/span\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{N}_{\\text{spot}}\\)\u003c/span\u003e\u003c/span\u003e is the number of spots. If spot \u003cem\u003ei\u003c/em\u003e is the neighbor of spot \u003cem\u003ej\u003c/em\u003e, then \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{ij}=1\\)\u003c/span\u003e\u003c/span\u003e; otherwise, 0. As a result, the Euclidean distance, which is computed using the spatial position data, determines a spot\u0026rsquo;s neighbors based on how close it is to [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Finally, we select k spots as neighbors from the top-ranked nearest neighbors. In our experiments, the best performance across most tested datasets was achieved when \u003cem\u003ek\u0026thinsp;=\u0026thinsp;3\u003c/em\u003e. It also aligns with the parameter setting utilized by GraphST in its implementation.\u003c/p\u003e\n\u003ch3\u003e2.5 Data Augmentation\u003c/h3\u003e\n\u003cp\u003ePrimarily, another neighborhood graph is constructed from augmented data. More specifically, the combined features of the image matrix and arbitrarily rearranged gene expression vectors \u003cem\u003eX\u003c/em\u003e among the spots are added to create the modified graph while preserving the original graph\u0026rsquo;s topological structure [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{G}^{{\\prime\\:}}=\\left({V}^{{\\prime\\:}},\\:{E}^{{\\prime\\:}}\\right)\\:\\:\\)\u003c/span\u003e \u003c/span\u003esubsequently represents the modified graph, and \u003cem\u003eX\u0026rsquo;\u003c/em\u003e represents the shuffled gene expression.\u003c/p\u003e\n\u003ch3\u003e2.6 Latent Representation Encoding\u003c/h3\u003e\n\u003cp\u003eIn order to learn spot representations that incorporate the informative portions of the gene expression profiles and spatial locations we leveraged a GNN-based encoder [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. The normalized gene expression profiles \u003cem\u003eX\u003c/em\u003e and the neighborhood graph \u003cem\u003eG\u003c/em\u003e are fed into the encoder, which then delivers the modified gene expressions \u003cem\u003eH\u003c/em\u003e to the decoder. To be more precise, a graph convolutional network (GCN) is used as an encoder to combine the representations of a spot\u0026rsquo;s neighbors in order to learn a latent representation for it. In mathematical terms, the encoder\u0026rsquo;s \u003cem\u003el-th\u003c/em\u003e layer representations are defined by \u003cb\u003eEq.\u0026nbsp;2\u003c/b\u003e:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{Z}_{s}^{l}=\\sigma\\:\\left(\\stackrel{\\sim}{A}{Z}_{s}^{l-1}{W}_{e}^{l-1}+{b}_{e}^{l-1}\\right)\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{Z}_{s}\\:\\)\u003c/span\u003e \u003c/span\u003eis denoted as the final output of the encoder while \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Z}_{s}^{l}\\)\u003c/span\u003e\u003c/span\u003e denotes the \u003cem\u003el-th\u003c/em\u003e layer output representation .\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\stackrel{\\sim}{A}={D}^{-\\frac{1}{2}}A{D}^{-\\frac{1}{2}}\\)\u003c/span\u003e \u003c/span\u003e indicates the location of the diagonal elements of the diagonal matrix in the normalized neighboring matrix and denotes a trainable weight matrix \u003cem\u003eW\u003c/em\u003e and a bias vector \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:b\\)\u003c/span\u003e\u003c/span\u003e, respectively. The Rectified Linear Unit (ReLU) is applied as a nonlinear activation function which is defined by \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\sigma\\:\\)\u003c/span\u003e\u003c/span\u003e. Subsequently, a decoder receives the latent representations and reverses it, restoring it to the original gene expression space. Similar to the encoder, the decoder reconstructs the gene expression using a symmetric architecture. The decoder is specified in particular as follows in \u003cb\u003eEq.\u0026nbsp;3\u003c/b\u003e:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:{H}_{s}^{t}=\\sigma\\:(\\stackrel{\\sim}{A}{H}_{s}^{t-1}{W}_{d}^{t-1}+{b}_{d}^{t-1})\\:\\:\\:\\:\\:\\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{H}_{s}^{t}\\)\u003c/span\u003e\u003c/span\u003e is set as the encoder\u0026rsquo;s output representation at the \u003cem\u003et\u003c/em\u003e-th layer and\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:{H}_{s}^{t-1}\\)\u003c/span\u003e\u003c/span\u003e, represents the reconstructed gene expression profiles at the (\u003cem\u003et-1)\u003c/em\u003e th layer. All nodes in the network share the trainable weight matrix, denoted by\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:W}_{d}\\)\u003c/span\u003e\u003c/span\u003e, and the bias vector, denoted by\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:{b}_{d}\\)\u003c/span\u003e\u003c/span\u003e. The model is trained by minimizing the self-reconstruction loss of gene expressions, following a specified process to fully leverage the gene expression profiles. The loss function is calculated using \u003cb\u003eEq.\u0026nbsp;4\u003c/b\u003e:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:{\\text{L}}_{\\text{recon\\:}}=\\sum\\:_{i=1}^{{N}_{\\text{spot\\:}}}{\\Vert\\:{x}_{i}-{h}_{i}\\Vert\\:}_{F}^{2}\\:\\:\\:\\:\\left(4\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eFor spot \u003cem\u003ei\u003c/em\u003e, the original normalized gene expression is represented by\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:x}_{i}\\)\u003c/span\u003e\u003c/span\u003e, and the reconstructed gene expression is represented by\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:h}_{i}\\)\u003c/span\u003e\u003c/span\u003e. The \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\Vert\\:.\\Vert\\:}_{F}\\)\u003c/span\u003e\u003c/span\u003e refers to the Frobenius norm, which is used for measuring the magnitude between two matrices.\u003c/p\u003e\n\u003ch3\u003e2.7 Self-supervised contrastive learning\u003c/h3\u003e\n\u003cp\u003eSEDIST employs a Self-Supervised Contrastive Learning (SSCL) technique to make sure the model captures the local spatial context of spots, hence improving the representation\u0026rsquo;s discriminative and informative qualities. In particular, using the original and corrupted graph G and G\u0026rsquo; as input, the GNN-based encoder initially creates two distinguished representation matrices for the specific spots. Inspired by deep graph informax (DGI), which is an unsupervised learning framework designed to learn node representations in graph-structured data, we combine the representations of the neighbors to describe a spot\u0026rsquo;s local context or its neighborhood microenvironment [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn contrast to DGI's use of global neighbors, the readout function here is redefined by applying a sigmoid to the mean of the immediate neighbors' representations. This approach assumes that each point in the spatial data is characterized by its local environment, which typically reflects the cell types and gene expression profiles of its local environment. For each spot \u003cem\u003ei\u003c/em\u003e in the graph, the local context vector and its representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{P}_{i}\\)\u003c/span\u003e\u003c/span\u003e create a positive pair, and the local context vector g and its equivalent representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{P}_{i}^{{\\prime\\:}}\\)\u003c/span\u003e\u003c/span\u003e from the corrupted graph form a negative pair. Maximizing positive pairs\u0026rsquo; mutual information while limiting negative pairs\u0026rsquo; mutual information is the fundamental principle of SSCL [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Physically adjacent spots will exhibit similar representations through the application of contrastive learning, whereas non-adjacent spots will display dissimilar representations. Subsequently, SSCL is modeled by utilizing binary cross-entropy (BCE). The \u003cb\u003econtrastive loss\u003c/b\u003e has the following formal definition as shown in \u003cb\u003eEq.\u0026nbsp;5\u003c/b\u003e:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:{\\mathcal{A}}_{1}=\\sum\\:_{i=1}^{{N}_{\\text{spot\\:}}}\\left({\\mathbb{E}}_{\\left(X,A\\right)}\\left[\\text{log}{\\Phi\\:}\\left({p}_{i},{q}_{i}\\right)\\right]\\right)\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:{\\mathcal{B}}_{1}={\\mathbb{E}}_{\\left({X}^{{\\prime\\:}},{A}^{{\\prime\\:}}\\right)}\\left[\\text{log}\\left(1-{\\Phi\\:}\\left({p}_{i}^{{\\prime\\:}},{q}_{i}\\right)\\right)\\right]\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\:{\\mathcal{L}}_{SSCL}=-\\frac{1}{2{N}_{\\text{spot\\:}}}\\left({\\mathcal{A}}_{1}+{\\mathcal{B}}_{1}\\right)\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(5\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\varPhi\\:\\left(\\cdot\\:\\right)\\:\\)\u003c/span\u003e\u003c/span\u003edual neural network discriminator \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{D}:{\\mathbb{R}}^{d}\\times\\:{\\mathbb{R}}^{d}\\to\\:\\mathbb{R}\\mathbb{\\:}\\)\u003c/span\u003e\u003c/span\u003ethat separates the positive pairs from the negative pairs. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\varPhi\\:({p}_{i},{q}_{i})\\)\u003c/span\u003e\u003c/span\u003e denotes the probability score assigned to the positive pair \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:({p}_{i},{q}_{i})\\)\u003c/span\u003e\u003c/span\u003e. Since the original graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:G\\:\\)\u003c/span\u003e\u003c/span\u003eand the modified graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{G}^{{\\prime\\:}}\\)\u003c/span\u003e\u003c/span\u003e share the same topological structure, a symmetric contrastive loss is employed, denoted as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{B}\\)\u003c/span\u003e\u003c/span\u003e. This loss, defined in Eq.\u0026nbsp;5, is applied to the corrupted graph used in the modified loss\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:{L}_{SSCL}\\)\u003c/span\u003e\u003c/span\u003e to improve the model\u0026rsquo;s stability and balance. The detailed explanation of the contrastive loss of the original and modified graphs is provided in \u003cb\u003esupplementary equation S1\u003c/b\u003e and \u003cb\u003esupplementary equation S2\u003c/b\u003e.\u003c/p\u003e\n\u003ch3\u003e2.8 Loss Function for SSCL\u003c/h3\u003e\n\u003cp\u003eThe self-reconstruction loss and contrastive loss are minimized to train the ST data representation-learning module. In a nutshell, this module\u0026rsquo;s total training loss is defined as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{L}\\)\u003c/span\u003e\u003c/span\u003e in \u003cb\u003eEq.\u0026nbsp;6\u003c/b\u003e [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]:\u003cdiv id=\"Equg\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equg\" name=\"EquationSource\"\u003e\n$$\\:\\mathcal{L}={\\lambda\\:}_{1}{\\mathcal{L}}_{\\text{recon\\:}}+{\\lambda\\:}_{2}\\left({\\mathcal{L}}_{SSCL}+{\\mathcal{L}}_{SSC{L}_{corrupt}}\\right)\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(6\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe weight parameters that balance the effects of the contrastive loss and the reconstructive loss are \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\lambda\\:}_{1}\\)\u003c/span\u003e\u003c/span\u003e and\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:\\lambda\\:}_{2}\\)\u003c/span\u003e\u003c/span\u003e. We predefined \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\lambda\\:}_{1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\lambda\\:}_{2}\\)\u003c/span\u003e\u003c/span\u003e as 10 and 1 empirically. This module is trained without the use of ST data, and the optimization is achieved using the Adam optimizer. For spatial clustering, the training epoch and learning rate are set to 600 and 0.0025, respectively.\u003c/p\u003e\n\u003ch3\u003e2.9 Location-specific gene expression analysis\u003c/h3\u003e\n\u003cp\u003eSpatial transcriptomics combines gene expression analysis with spatial information, enabling researchers to map where specific genes are active within tissues, offering deeper insights into cellular organization and function [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e][\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This behavior can be mapped as explained in \u003cb\u003eEq.\u0026nbsp;7\u003c/b\u003e:\u003cdiv id=\"Equh\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equh\" name=\"EquationSource\"\u003e\n$$\\:S\\left(x,y\\right)=\\sum\\:_{i=1}^{N}{G}_{i}\\cdot\\:\\partial\\:\\left(x-{x}_{i}\\right)\\cdot\\:\\partial\\:\\left(y-{y}_{i}\\right)\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(7\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:S\\left(x,y\\right)\\)\u003c/span\u003e\u003c/span\u003e shows how genes (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{G}_{i}\\)\u003c/span\u003e\u003c/span\u003e) are expressed in specific positions in the cell, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{(x}_{i},{y}_{i})\\)\u003c/span\u003e\u003c/span\u003e is the coordinates of each cell. The equation ensures precise measurement of gene activity at specific cellular locations, facilitating an understanding of how genes interact across different cellular regions and providing insights into cellular architecture [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Incorporating spatial location allows for the observation of gene interactions at specific sites, offering a more detailed view of cellular organization and function. The spatial transcriptomics method increases our understanding of cell architecture and opens up new possibilities for studying cellular interactions in various biological processes [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e][\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003e2.10 Evaluation Matrix\u003c/h3\u003e\n\u003cp\u003eThe Adjusted Rand Index (ARI) works as a pivotal metric within the research framework, providing a robust assessment of the clustering algorithm\u0026rsquo;s efficacy in capturing meaningful cellular and molecular distinctions [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. ARI evaluates the agreement between true and predicted cell or gene cluster assignments, offering a comprehensive understanding of the spatial and transcriptional landscape within tissues [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. Here, the ARI is used to evaluate the performance of various clustering techniques including SEDIST on datasets containing spot-type labels. The ARI is calculated using the following \u003cb\u003eEq.\u0026nbsp;8\u003c/b\u003e:\u003cdiv id=\"Equi\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equi\" name=\"EquationSource\"\u003e\n$$\\:ARI=\\frac{\\sum\\:_{ij}\\left(\\begin{array}{c}{n}_{ij}\\\\\\:2\\end{array}\\right)-\\left[\\sum\\:_{i}\\left(\\begin{array}{c}{a}_{i}\\\\\\:2\\end{array}\\right)\\sum\\:_{j}\\left(\\begin{array}{c}{b}_{j}\\\\\\:2\\end{array}\\right)\\right]/\\left(\\begin{array}{c}n\\\\\\:2\\end{array}\\right)}{\\frac{1}{2}[\\sum\\:_{i}\\left(\\begin{array}{c}{a}_{i}\\\\\\:2\\end{array}\\right)+\\sum\\:_{j}\\left(\\begin{array}{c}{b}_{j}\\\\\\:2\\end{array}\\right)]-\\left[\\sum\\:_{i}\\left(\\begin{array}{c}{a}_{i}\\\\\\:2\\end{array}\\right)\\sum\\:_{j}\\left(\\begin{array}{c}{b}_{j}\\\\\\:2\\end{array}\\right)\\right]/\\left(\\begin{array}{c}n\\\\\\:2\\end{array}\\right)}\\:\\:\\:\\:\\:\\:\\:\\:\\left(8\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{n}_{ij}\\)\u003c/span\u003e\u003c/span\u003e represents the number of data points shared between clusters \u003cem\u003ei\u003c/em\u003e and \u003cem\u003ej\u003c/em\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{i}\\)\u003c/span\u003e\u003c/span\u003e denotes the total number of data points in cell or gene cluster \u003cem\u003ei\u003c/em\u003e. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{b}_{j}\\)\u003c/span\u003e\u003c/span\u003e represents the total number of data points in cluster \u003cem\u003ej\u003c/em\u003e, and \u003cem\u003eN\u003c/em\u003e is the total number of data points. The resulting normalized index ranges from 0 to 1, with a higher ARI indicating a more accurate agreement between true and predicted cluster assignments [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e].\u003c/p\u003e"},{"header":"3 Result Analysis","content":"\u003cp\u003eIn this section, we discuss SEDIST's overall performance. The proposed method showed enhanced capability of identifying spatial domains in most of the tissue slices, while achieving around a maximum of 0.6466 ARI. Additionally, we discuss the exploratory analysis capability of this method such as 3D embedding, diffusion map, draw map and identification of top genes within each cell.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Performance Comparison with State-of-the-Art (SOTA) Methods on DLPFC Datasets\u003c/h2\u003e \u003cp\u003eThe DLPFC benchmarking dataset, annotated by Maynard et al. for cortical layers and white matter using gene markers and cytoarchitecture (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA), was used to compare the performance of SEDIST with other state-of-the-art (SOTA) spatial clustering methods DeepST, GraphST, conST, GRAS4T, GAAEST, MNMST, SpaGCN and SpaceFlow. All models were run under a standardized development environment with PyTorch version 2.3.0 and the hardware configuration specified in Section 3.5 to ensure fair and consistent evaluation.\u003c/p\u003e \u003cp\u003eOur results demonstrate that SEDIST consistently aligns well with manual annotations of the DLPFC dataset and outperforms several existing methods across most of the 12 dataset slices. Our results show that SEDIST achieved the highest Adjusted Rand Index (ARI) score of 0.6466 on slice 151673 by outperforming GraphST (0.6123), DeepST (0.5855), and MNMST (0.6012). These findings highlight the effectiveness of SEDIST in accurately identifying spatial domains.\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB, the visual comparison further emphasizes SEDIST's strong ARI of 0.6466 on slide 151673 for layer boundary delineation. It is evident that SEDIST outperforms DeepST, GraphST, conST, GRAS4T, GAAEST, MNMST, SpaGCN and SpaceFlow on this particular slice. Among these, only SEDIST and GraphST were successful in clearly distinguishing between cortical layers L1 and L2.\u003c/p\u003e \u003cp\u003eAcross the 12 slices listed in \u003cb\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e, SEDIST outperformed all other methods in 8 slices, specifically 151508, 151510, 151670, 151671, 151672, 151673, 151674, and 151676, demonstrating a consistent trend of strong performance. For example, on slice 151670 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eH), SEDIST achieved an ARI of 0.6244 which is higher than GraphST (0.5004) and DeepST (0.3048). Similarly on slice 151672 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eJ), SEDIST obtained an ARI of 0.6273 by outperforming GraphST (0.5733), DeepST (0.4512) and MNMST (0.5738). Another example is slice 151674 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eL), where SEDIST reached an ARI of 0.5993, again surpassing GraphST (0.5561) and MNMST (0.5361). These examples highlight the SEDIST of ability to consistently identify spatial domains more effectively than other state-of-the-art approaches across diverse samples.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eHowever, SEDIST showed moderate performance in some other slices. For example, in slice 151507 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC) SEDIST scored an ARI of 0.4361, which was lower than DeepST (0.5529) and MNMST (0.5203). Similar trends were observed in slice 151509 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eE), where DeepST slightly outperformed SEDIST.\u003c/p\u003e \u003cp\u003eThese variations highlight the importance of evaluating models across diverse biological conditions. While SEDIST excels in most cases, its performance can vary depending on the specific structural and biological features of the tissue slices. Figures\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC through \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eN provide detailed visualizations of the spatial clustering outcomes across all slices and demonstrate the relative strength of SEDIST in real-world spatial domain identification tasks.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Exploring Spatial Organization and Gene Expression from SEDIST Visualizations and Gene Profiling\u003c/h2\u003e \u003cp\u003eIn this proposed method, we also employed a 3D embedding technique for clustering gene expression data, enabling the integration of spatial information with gene expression profiles, and facilitating a more comprehensive understanding of transcriptomic patterns as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA. The 3D embedding helps us visualize how genes are expressed in different tissue areas due to their projection in 3D space. We additionally observe from Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA that the ground truth of the tissue slides closely matches with the predicted 3D embedding by our method, especially Layer 1, 2, 3, 5 \u0026amp; 6 as well as white matter. Furthermore, UMAP[\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] and partition-based graph abstraction in SEDIST revealed a clear, structured arrangement of cortical layers from L1 to L6 and white matter (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Unlike the less-defined layer separation in 3D embedding (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA), the UMAP results show the delineation of each layer. This demonstrates how UMAP outperforms other methods in providing a more accurate and detailed organization of cortical structures. We also generated a diffusion map to understand the cell-cell connections among different tissue layers in terms of physical space and gene expression as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC. In Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC, we can observe the ground truth and the predicted diffusion map shows a similar pattern. We observe the continuum of cellular states in both maps as we move from one cortical layer to another. The subtle transitions between layers 1\u0026ndash;6 and WM as well as their distributions are also visible in this figure. For example, the cells are distributed within a similar range along DC1 (Diffusion Component 1), indicating these layers are within the same tissue sections. However, we observe in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC that WM and Layer 6 exhibit a broad spread along DC2 (Diffusion Component 2), indicating significant heterogeneity within these layers. In contrast, layers 1\u0026ndash;5 are narrowly clustered and overlapped along DC2, indicating a higher degree of homogeneity in their secondary characteristics. This pattern highlights distinct gene expression profiles and transitions across different cortical layers. Furthermore, we analyzed to facilitate the visualization of cells clustering in two-dimensional space FR1 (Functional Regions 1) \u0026amp; FR2 (Functional Regions 2) based on the gene expression patterns as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD. Unlike diffusion maps, which cluster cells based on their gene expression and spatial location, focus only on gene expression. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD demonstrates that in both ground truth and predictions of SEDIST, layer 1 is distinctly clustered (blue color in ground truth and brown color in SEDIST), suggesting that layer 1 is well-defined in terms of gene expression and our method can distinguish it accurately. Additionally, the WM is separated in both visualizations, indicating that SEDIST can identify the transition between cortical layers and white matter. However, the predicted drawgraph has some discrepancies in layers 2\u0026ndash;6 compared to the ground truth. While these layers overlap in the ground truth, SEDIST predicts more distinct clusters (green, orange, and blue), indicating deviations in the clustering. The combined integration of UMAP, Diffmap[\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], and Drawmap [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e] in SEDIST results enhanced the interpretability of our proposed method and provided a comprehensive view of the spatial and transcriptional layout within cortical layers.\u003c/p\u003e \u003cp\u003eFurthermore, SEDIST can figure out spatially variable gene influence across the tissue slice to understand the intricate cell behavior as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eE. In Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eE, we demonstrate these genes\u0026rsquo; (MBP, PLP1, NEFL, COX6C, IGKC, HPCAL1 and PCP4) spatial distribution across the overall tissue slices.\u003c/p\u003e \u003cp\u003eWe observed varying levels of expression across the tissue for these genes, with a few spots of higher expression (yellow color), indicating specific cellular processes influenced by these genes. The consistent low expression (black color) of these genes may provide insights into their role in maintaining cellular homeostasis. We can understand how gene expression varies across tissue by visualizing such a spatial distribution graph for particular genes, which is crucial for understanding complex biological processes, cell behavior, cell-cell communications, and advancing therapeutic interventions. Additionally, we discuss how SEDIST can perform a comprehensive analysis of individual cell gene expression to identify the main genetic drivers influencing cellular behavior within spatial transcriptomics. Figure S3 shows a box plot of the top 20 genes by percentage of total counts in the dataset identified by SEDIST.\u003c/p\u003e \u003cp\u003eHere, we calculated each gene expression level across all the spots in the tissue and showed the top genes that showed the highest mean expression level across all the cells. In this result, we observe that a significant portion of the top-expressed genes are Mitochondrial (e.g., MT-CO2, MT-ND4, MT-CYB), indicating cells undergoing metabolic activity since mitochondrial genes are involved in energy production. Mitochondrial genes usually perform different metabolic functions in tissues. However, several MT genes exhibit outliers, indicating that certain cells have unusually high metabolic activity or are in a low-energy state. These variations might reflect differences in cellular states, such as heightened metabolic activity in certain regions or phases of the tissue. A recent study showed that mitochondrial genes were down-regulated in the tissue of aging mice whereas some mitochondrial genes were up-regulated for specific biological roles [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. Another study discussed how mitochondrial gene expression levels were up-regulated in the prefrontal cortex due to chronic stress [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Such studies indicate that the overall mean expression level of mitochondrial genes across cortical layers can be up-regulated which we also demonstrated in \u003cb\u003eSupplementary Figure S3\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eIt is also observed that GAPDH[\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e], ACTB[\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e], and RPL41[\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e] genes (a housekeeping gene) among the top 20 genes which are expected because these genes are often highly expressed in many cell types. studies suggest that certain GAPDH genes can have high levels of expression due to stressful conditions and other biological reasons [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e][\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. For example, a study asserted that housekeeping genes are usually expressed stably across tissues and used as a reference. Genes like FTH1 [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e], CST3 [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e], and CLU[\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e] in the top genes suggest that some cells in the tissue slice may be under stress or actively maintaining homeostasis. This could be due to the tissue type, experimental conditions, or cell state (e.g., proliferation, differentiation, or apoptosis). The expression of CALM1[\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e] and TMSB10[\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e] suggests active calcium signaling and cytoskeletal dynamics, which are important for many cellular processes. However, under various conditions, they can also show high expression levels[\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis analysis helps to identify the genes that are most actively expressed in cells and gives different biological insights, such as the metabolic state of cells, possible differences between cell types, and gene expression roles in governing cellular functions and interactions within the tissue microenvironment.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Insights from SEDIST: Exploratory Analysis of Gene Expression and Spatial Architecture\u003c/h2\u003e \u003cp\u003eWe also examined the complexities of feature representation in PCA through graphical visualization (See \u003cb\u003eSupplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e). We selected the top 20 principal components (PCA20) to retain the most significant variance in the dataset while reducing dimensionality and minimizing redundancy in the feature space. This allowed our analysis to focus on the most informative features (essential gene expression patterns) while avoiding data overload. To interpret the distinctions between normalized and non-normalized gene expression, SEDIST utilized variance versus mean expression plots (See \u003cb\u003eSupplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e), which was essential for ensuring the accuracy and reliability of our gene expression analyses, as normalization significantly affects the interpretation of gene expression data. Moreover, the top 10 genes in each layer were identified within slice 151673 as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. These genes, such as MBP, PLP1, and CNP dominated in the cortical layer 1, contributing to processes like myelination. Other layers showed distinct gene profiles such as KRT17 and DIRAS2 in layer 2, and NEFL and NRGN in the third, each demonstrating valuable insights into the spatial organization and unique biological functions within the cortical layers.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Performance of SEDIST on Visium Dataset\u003c/h2\u003e \u003cp\u003eIn this study, we also applied the SEDIST to spatial transcriptomics data from 10x Genomics Visium for human breast cancer, a cancer known for its high levels of variability within and between tumors [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. To check the robustness of SEDIST we performed spatial transcriptomics analysis on Visium dataset along with Visium 10x. This was performed to ensure the robustness. We tested it on the Human Breast Cancer Block A (Section 1) dataset. This BRCA-1 H\u0026amp;E slice has a total number of 20 different regions based on the pathological features, with 4 major morphotypes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eOur results, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, highlight SEDIST\u0026rsquo;s ability to identify spatial domains and provide different visual perspectives of its clustering performance. SEDIST achieved an Adjusted Rand Index (ARI) score of 0.6185, which indicates its accuracy in dividing tissue layers. Most of the existing literature has manually separated 20 different regions of the tissue slice, while SEDIST successfully identified all 20 clusters with an ARI score of 0.6185.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Development Environment and Computational Benchmarking\u003c/h2\u003e \u003cp\u003eSEDIST was developed and tested on a MacBook Pro with the following hardware configuration: Model: MacBook Pro M2 Pro 14-Inch, Processor: 10-Core CPU, Graphics: 16- Core GPU, Memory: 16GB RAM, Storage: 512GB SSD. The powerful combination of a 10-core CPU and a 16-core GPU provides significant computational capabilities for both the training and testing phases. The 16GB RAM ensures efficient handling of data and model parameters, and the 512GB SSD offers ample storage for datasets and model checkpoints. This hardware configuration offers a robust and efficient environment for developing, training, and evaluating SEDIST.\u003c/p\u003e \u003cp\u003eIn addition to evaluating the clustering performance of SEDIST against state-of-the-art (SOTA) methods based on Adjusted Rand Index (ARI), we conducted a comparative analysis of their computational efficiency in terms of time and memory usage. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the execution time (in seconds) and memory consumption (in megabytes) for SEDIST, GraphST, DeepST, SpaGCN, GRAS4T, GAAEST, and MNMST. SEDIST demonstrated competitive time efficiency, completing the analysis in 121.45 seconds. This was faster than DeepST (261.45 seconds), SpaGCN (161.45 seconds), GRAS4T (189.07 seconds), and MNMST (776.2 seconds). However, SEDIST required slightly more time than GraphST (90.45 seconds) and GAAEST (109.3 seconds).\u003c/p\u003e \u003cp\u003eAmong the evaluated methods, MNMST was the slowest, requiring 776.2 seconds, which is more than six times the execution time of SEDIST. Conversely, GraphST and GAAEST were the fastest methods, with execution times of 90.45 and 109.3 seconds, respectively. Despite these differences, SEDIST\u0026rsquo;s execution time strikes a balance between computational speed and its superior clustering accuracy, making it a viable choice for applications where both precision and efficiency are critical. In terms of memory usage, SEDIST required 2195.3 MB, demonstrating a favorable balance between computational demands and performance. It used significantly less memory than GAAEST (5646.06 MB) and DeepST (3295.56 MB), both of which exhibited the highest memory consumption among the methods evaluated. SpaGCN and GraphST also consumed more memory than SEDIST, requiring 2469.46 MB and 2288.67 MB, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTime and Memory benchmark for different State-of-the-Art methods against SEDIST.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSEDIST\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGraphST\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeepST\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpaGCN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGRAS4T\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eGAAEST\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMNMST\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime (sec)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e121.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e90.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e261.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e161.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e189.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e109.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e776.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMemory (mb)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2195.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2288.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3295.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2469.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1663.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5646.06 MB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e1761.28\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eGRAS4T and MNMST, in contrast, exhibited lower memory consumption than SEDIST, with GRAS4T requiring the least memory at 1663.48 MB, followed by MNMST at 1761.28 MB. While these methods demonstrated efficiency in memory usage, their clustering accuracy was generally lower compared to SEDIST, as discussed earlier. The results presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e highlight SEDIST\u0026rsquo;s balanced computational profile. While slightly slower than GraphST and GAAEST, SEDIST outperformed these methods in clustering accuracy, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Moreover, SEDIST's memory usage is significantly more efficient than that of GAAEST and DeepST, making it a more practical choice for large-scale datasets or resource-constrained environments. This analysis underscores the trade-offs between computational efficiency and clustering performance among the methods. SEDIST\u0026rsquo;s ability to deliver superior clustering accuracy, coupled with reasonable time and memory requirements, further establishes its robustness and applicability in spatial clustering tasks.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eDeciphering the activity within a complex cellular microenvironment is crucial. Also, the identification of spatial domains is significant for understanding these interactions inside a multicellular organism. In this study, we introduce SEDIST, a novel deep learning-based approach designed to identify spatial domains by integrating gene expression profiles with spatial information.\u003c/p\u003e \u003cp\u003eInitially, the morphological image is segmented into spots based on H\u0026amp;E staining, from which high-level morphological features are extracted using an autoencoder. By incorporating spatial locations along with gene expression profiles and morphological features, we constructed and enhanced an undirected neighborhood graph, then applied a graph encoder to unveil latent features. Finally, we utilized self-supervised contrastive learning to refine these features and employed the MCLUST method to cluster the spatial domains based on their morphological and spatial characteristics.\u003c/p\u003e \u003cp\u003eTo validate the performance of SEDIST, a comprehensive comparative analysis is conducted with state-of-the-art method such as GraphST, DeepST, ConST, SpaceFlow, SpaGCN. The performance of identifying the spatial domain is evaluated using Adjusted Rand Index (ARI) and SEDIST provides more stable and comparatively better performances in most of the cases. For slice 151673, SEDIST achieves an ARI of 0.6466, outperforming stLearn (0.3031), SpaGCN (0.4557), DeepST (0.5855), and GraphST (0.6123). Furthermore, the proposed method demonstrates superior performance most of the cases across multiple slices, including 151507, 151508, 151509, 151669, 151670, 151671, 151672, 151674, 151675, and 151676, particularly within the Dorsolateral Prefrontal Cortex (DLPFC). This consistent improvement highlights the robustness and effectiveness of SEDIST in comparison to existing approaches.\u003c/p\u003e \u003cp\u003eAnother key strength of SEDIST is that it employs a 3D embedding technique for clustering gene expression data, where this approach integrates spatial information with gene expression profiles, allowing for a more comprehensive understanding of transcriptomics patterns and improving the accuracy of downstream analysis. By leveraging the spatial context, the method enhances the interpretability and precision of gene expression clustering. The integration of diffusion-based visualization approaches significantly improved the interpretability, also offering a holistic view on the spatial and transcriptional organization within cortical layers.\u003c/p\u003e \u003cp\u003eTo understand tissue organization and biological functions, it is important to accurately identify spatial domains and extract spatially expressed genes. SEDIST allows for the identification and visualization of highly expressed genes within spatial domains. Moreover, the top 10 highly expressed genes were shown within slice 151673. Several genes were identified as playing critical roles in key biological processes, including cell proliferation, axon recognition, axon-glia signaling, differentiation, membrane component transport, myelin compaction, and node formation. Distinct gene expression patterns revealed functional specializations and characteristics, providing valuable insights into the organization and biological activity of cortical regions\u003c/p\u003e \u003cp\u003eIn terms of model stability, SEDIST consistently identifies spatial domains that align with ground truth data, outperforming existing methods. However, some discrepancies were observed between our reproduced results and those reported in the original literature, likely due to differences in hardware and the convergence challenges inherent to unsupervised methods. To address this limitation, all comparative methods were evaluated under the same hardware and software conditions, ensuring a fair and consistent comparison.\u003c/p\u003e \u003cp\u003eIn summary, SEDIST is a novel deep learning-based approach that modifies feature representation for each spot to identify spatial domains. As the generation of spatial transcriptomics (ST) data continues to grow, it is anticipated that SEDIST will contribute to revealing new insights into cellular organization within a spatial context. Therefore, optimizing memory efficiency through mini-batch processing and parallel computing in distributed learning systems can be an important enhancement to spatial transcriptomics (ST) studies. Additionally, integrating spatial omics and scRNA-seq data can further enhance the resolution of spatial transcriptomics results and facilitate automatic annotation of spatial domains.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eClinical trial number:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and material:\u0026nbsp;\u003c/strong\u003eAll the data and codes used in the methods are available at https://github.com/SEDIST-014/SEDIST.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u0026nbsp;\u003c/strong\u003eNone declared.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eFunded by the Institute for Advanced Research Publication Grant of United International University, Ref. No.: IAR-2025-Pub-033.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions: FR\u003c/strong\u003e contributed to conceptualization, methodology, software development, visualization, writing \u0026ndash; original draft, investigation, formal analysis, and validation. \u003cstrong\u003eAAS\u003c/strong\u003e was involved in conceptualization, methodology, software development, visualization, writing \u0026ndash; original draft, investigation, formal analysis, and validation. \u003cstrong\u003eRA\u003c/strong\u003e was responsible for supervision, conceptualization, methodology, formal analysis, writing \u0026ndash; review \u0026amp; editing, investigation, and validation. \u003cstrong\u003eFA\u003c/strong\u003e contributed to validation, formal analysis, writing \u0026ndash; original draft, writing \u0026ndash; review \u0026amp; editing, investigation, project administration, and funding acquisition. \u003cstrong\u003eSAE\u003c/strong\u003e contributed to writing \u0026ndash; original draft and visualization. \u003cstrong\u003eUAB\u003c/strong\u003e contributed to writing \u0026ndash; original draft and visualization.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgement:\u0026nbsp;\u003c/strong\u003eNone Declared\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMaynard KR, Collado-Torres L, Weber LM, Uytingco C, Barry BK, Williams SR, et al. Transcriptome-scale spatial gene expression in the human dorsolateral prefrontal cortex. Nat Neurosci. 2021;24:425\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Liu B, Zhao G, Lee Y, Buzdin A, Mu X, et al. Spatial transcriptomics: Technologies, applications and experimental considerations. Genomics. 2023;115:110671.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShengquan C, Boheng Z, Xiaoyang C, Xuegong Z, Rui J. stPlus: a reference-based method for the accurate enhancement of spatial transcriptomics. Bioinformatics. 2021;37 Supplement_1:i299\u0026ndash;307.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSatija R, Farrell JA, Gennert D, Schier AF, Regev A. Spatial reconstruction of single-cell gene expression data. Nat Biotechnol. 2015;33:495\u0026ndash;502.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdelaal T, Mourragui S, Mahfouz A, Reinders MJT. SpaGE: Spatial Gene Enhancement using scRNA-seq. Nucleic Acids Res. 2020;48:e107.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Gao C, Sodicoff J, Kozareva V, Macosko EZ, Welch JD. Jointly defining cell types from multiple single-cell datasets using LIGER. Nat Protoc. 2020;15:3632\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeng H, Yuan Y, Bar-Joseph Z. Clustering spatial transcriptomics data. Bioinformatics. 2022;38:997\u0026ndash;1004.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDong K, Zhang S. Deciphering spatial domains from spatially resolved transcriptomics with an adaptive graph attention auto-encoder. Nat Commun. 2022;13:1739.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolf FA, Angerer P, Theis FJ. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 2018;19:15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu H, Fu H, Long Y, Ang KS, Sethi R, Chong K, et al. Unsupervised spatially embedded deep representation of spatial transcriptomics. Genome Med. 2024;16:12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLong Y, Ang KS, Li M, Chong KLK, Sethi R, Zhong C, et al. Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST. Nat Commun. 2023;14:1155.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao E, Stone MR, Ren X, Guenthoer J, Smythe KS, Pulliam T, et al. Spatial transcriptomics at subspot resolution with BayesSpace. Nat Biotechnol. 2021;39:1375\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu J, Li X, Coleman K, Schroeder A, Ma N, Irwin DJ, et al. SpaGCN: Integrating gene expression, spatial location and histology to identify spatial domains and spatially variable genes by graph convolutional network. Nat Methods. 2021;18:1342\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu C, Jin X, Wei S, Wang P, Luo M, Xu Z, et al. DeepST: identifying spatial domains in spatial transcriptomics by deep learning. Nucleic Acids Res. 2022;50:e131.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZong Y, Yu T, Wang X, Wang Y, Hu Z, Li Y. conST: an interpretable multi-modal contrastive learning framework for spatial transcriptomics. 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao Y, Lin Y, Patrick E, Yang P, Yang JYH. scFeatures: multi-view representations of single-cell and spatial data for disease outcome prediction. Bioinformatics. 2022;38:4745\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaredon MSB, Yang J, Kothapalli N, Lewis W, Kaminski N, Niklason LE, et al. Comprehensive visualization of cell-cell interactions in single-cell and spatial transcriptomics with NICHES. Bioinformatics. 2023;39:btac775.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Z, Cui F, Su W, Dou L, Xu A, Cao C, et al. webSCST: an interactive web application for single-cell RNA-sequencing data and spatial transcriptomic data integration. Bioinformatics. 2022;38:3488\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaviano A, Henderson NC, Baumert TF. Single-cell genomics and spatial transcriptomics: Discovery of novel cell states and cellular interactions in liver physiology and disease biology. J Hepatol. 2020;73:1219\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHou W, Ji Z. Palo: spatially aware color palette optimization for single-cell and spatial data. Bioinformatics. 2022;38:3654\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePham D, Tan X, Balderson B, Xu J, Grice LF, Yoon S, et al. Robust mapping of spatiotemporal trajectories and cell-cell interactions in healthy and diseased tissues. Nat Commun. 2023;14:7739.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGui Y, Li C, Xu Y. Spatial domains identification in spatial transcriptomics using modality-aware and subspace-enhanced graph contrastive learning. Comput Struct Biotechnol J. 2024;23:3703\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang T, Zhu H, Zhou Y, Ding W, Ding W, Han L, et al. Graph attention automatic encoder based on contrastive learning for domain recognition of spatial transcriptomics. Commun Biol. 2024;7:1351.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Liu Z, Ma X. MNMST: topology of cell networks leverages identification of spatial domains from spatial transcriptomics data. Genome Biol. 2024;25:133.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScrucca L, Fop M, Murphy TB, Raftery AE. mclust 5: Clustering, Classification and Density Estimation Using Gaussian Finite Mixture Models. R J. 2016;8:289\u0026ndash;317.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagy C, Maitra M, Tanti A, Suderman M, Th\u0026eacute;roux J-F, Davoli MA, et al. Single-nucleus transcriptomics of the prefrontal cortex in major depressive disorder implicates oligodendrocyte precursor cells and excitatory neurons. Nat Neurosci. 2020;23:771\u0026ndash;81.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChelebian E, Avenel C, Kartasalo K, Marklund M, Tanoglidi A, Mirtti T, et al. Morphological Features Extracted by AI Associated with Spatial Transcriptomics in Prostate Cancer. Cancers (Basel). 2021;13:4837.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBergenstr\u0026aring;hle J, Larsson L, Lundeberg J. Seamless integration of image and molecular analysis for spatial transcriptomics workflows. BMC Genomics. 2020;21:482.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBae S, Choi H, Lee DS. Discovery of molecular features underlying the morphological landscape by integrating spatial transcriptomic data with deep features of tissue images. Nucleic Acids Res. 2021;49:e55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKleshchevnikov V, Shmatko A, Dann E, Aivazidis A, King HW, Li T, et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat Biotechnol. 2022;40:661\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSong Q, Su J. DSTG: deconvoluting spatial transcriptomics data through graph-based artificial intelligence. Brief Bioinform. 2021;22:bbaa414.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRatajczak F, Joblin M, Hildebrandt M, Ringsquandl M, Falter-Braun P, Heinig M. Speos: an ensemble graph representation learning framework to predict core gene candidates for complex diseases. Nat Commun. 2023;14:7206.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVeličković P, Fedus W, Hamilton WL, Li\u0026ograve; P, Bengio Y, Hjelm RD. Deep Graph Infomax. In: 7th International Conference on Learning Representations, ICLR 2019. New Orleans, Louisiana; 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKipf TN, Welling M. Semi-Supervised Classification with Graph Convolutional Networks. In: 5th International Conference on Learning Representations, ICLR 2017 - Conference Track Proceedings. Toulon, France; 2016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen W-T, Lu A, Craessaerts K, Pavie B, Sala Frigerio C, Corthout N, et al. Spatial Transcriptomics and In Situ Sequencing to Study Alzheimer\u0026rsquo;s Disease. Cell. 2020;182:976\u0026ndash;e99119.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKleshchevnikov V, Shmatko A, Dann E, Aivazidis A, King HW, Li T et al. Compr Mapp tissue cell Archit via Integr single cell Spat transcriptomics bioRxiv. 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang X, Almet AA, Nie Q. The promising application of cell-cell interaction analysis in cancer from single-cell and spatial transcriptomics. Semin Cancer Biol. 2023;95:42\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan Y, Bar-Joseph Z. GCNG: graph convolutional networks for inferring gene interaction from spatial transcriptomics data. Genome Biol. 2020;21:300.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSong J, Lamstein J, Ramaswamy VG, Webb M, Zada G, Finkbeiner S, et al. Enhancing Spatial Transcriptomics Analysis by Integrating Image-Aware Deep Learning Methods. Biocomputing 2024. WORLD SCIENTIFIC; 2023. pp. 450\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoncada R, Barkley D, Wagner F, Chiodin M, Devlin JC, Baron M, et al. Integrating microarray-based spatial transcriptomics and single-cell RNA-seq reveals tissue architecture in pancreatic ductal adenocarcinomas. Nat Biotechnol. 2020;38:333\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, Chen S, Pan X, Yuan Y, Shen H-B. Cell clustering for spatial transcriptomics data with graph neural networks. Nat Comput Sci. 2022;2:399\u0026ndash;408.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcInnes L, Healy J, Saul N, Gro\u0026szlig;berger L. UMAP: Uniform Manifold Approximation and Projection. J Open Source Softw. 2018;3:861.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL\u0026aacute;b\u0026aacute;r JL, DiffMap. A new free computer program to process scanned electron diffraction patterns. Resolut Discov. 2022;6:7\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Ooijen JW. DrawMap: a computer program for drawing genetic linkage maps. J Hered. 1994;85:66.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu C, Tu T, Xie M, Wang Y, Yan B, Gong Y, et al. Spatially resolved transcriptome of the aging mouse brain. Aging Cell. 2024;23:e14109.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeger M, Alpern D, Cherix A, Ghosal S, Grosse J, Russeil J, et al. Mitochondrial gene signature in the prefrontal cortex for differential susceptibility to chronic stress. Sci Rep. 2020;10:18308.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicholls C, Li H, Liu J-P. GAPDH: a common enzyme with uncommon functions. Clin Exp Pharmacol Physiol. 2012;39:674\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDai S, Wang H, Lin Z. ACTB Mutations Analysis and Genotype\u0026ndash;Phenotype Correlation in Becker\u0026rsquo;s Nevus. Biomedicines. 2021;9:1879.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGeng W, Ren J, Shi H, Qin F, Xu X, Xiao S, et al. RPL41 sensitizes retinoblastoma cells to chemotherapeutic drugs via ATF4 degradation. J Cell Physiol. 2021;236:2214\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHerbet M, Piątkowska-Chmiel I, Motylska M, Gawrońska-Grzywacz M, Nieradko-Iwanicka B, Dudka J. Alteration in the Expression of Genes Involved in Cerebral Glucose Metabolism as a Process of Adaptation to Stressful Conditions. Brain Sci. 2022;12:498.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKang Y, Wu Z, Cai D, Lu B. Evaluation of reference genes for gene expression studies in mouse and N2a cell ischemic stroke models using quantitative real-time PCR. BMC Neurosci. 2018;19:3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTian Y, Lu J, Hao X, Li H, Zhang G, Liu X, et al. FTH1 Inhibits Ferroptosis Through Ferritinophagy in the 6-OHDA Model of Parkinson\u0026rsquo;s Disease. Neurotherapeutics. 2020;17:1796\u0026ndash;812.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChristopher JR, Ponnaiyan D, Parthasarathy H, Tadepalli A. Association of CST3 Gene with Its Protein: Cystatin C in Health and Severe Periodontal Disease. Genet Test Mol Biomarkers. 2021;25:405\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Lv X, Chen L, Liu Y. The role and function of CLU in cancer biology and therapy. Clin Exp Med. 2022;23:1375\u0026ndash;91.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu T, Han X, Zheng S, Liu Q, Tuerxun A, Zhang Q, et al. CALM1 promotes progression and dampens chemosensitivity to EGFR inhibitor in esophageal squamous cell carcinoma. Cancer Cell Int. 2021;21:121.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang C, He Y, You Z, Chen X. TMSB10 Promotes Progression of Clear Cell Renal Cell Carcinoma via JUN Transcription Regulation. Ann Clin Lab Sci. 2022;52:230\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6619020/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6619020/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eMotivation:\u003c/h2\u003e \u003cp\u003eThe recent advancement of spatially resolved transcriptomics provides a way to investigate cellular heterogeneity and tissue micro-environment using gene expression profiles with spatial context, often accommodated with histological data. However, precise spatial domain identification, multisample integration, and cellular decomposition still remain challenging.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eHere we introduce SEDIST, a novel method that integrates autoencoders and graph neural networks with self-supervised contrastive learning to utilize spatial transcriptomics data effectively. This approach enhances the model\u0026rsquo;s ability to extract informative and discriminative spot representations by minimizing the embedding distance between spatially adjacent spots through self-supervised contrastive learning. Highly expressed genes in different domain layers are also identified and 3D embedded plot of clusters along with diffusion map are also graphically represented as downstream analysis. This comprehensive integration of spatial context leads to superior performance compared to existing state-of-the-art (SOTA) methods in the human brain dorsolateral prefrontal cortex (DLPFC) and the BRCA (human breast cancer) 10X Visium dataset, as evidenced by achieving an Adjusted Rand Index (ARI) of 0.6466 and 0.6185, respectively.\u003c/p\u003e","manuscriptTitle":"SEDIST: Spatially Enhanced Domain Identification through Spatial Transcriptomics","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-19 06:53:36","doi":"10.21203/rs.3.rs-6619020/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9e5c6803-a0fc-4db2-8ecd-193ffe5b79bf","owner":[],"postedDate":"May 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-22T07:23:27+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-19 06:53:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6619020","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6619020","identity":"rs-6619020","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}