Advances in early human embryonic development and organogenesis from spatiotemporal multi-omics

In vitro embryo-like models serve as a new platform for studying early human development

To overcome the ethical and technological bottlenecks in human embryo research, scientists have developed a series of in vitro models based on hPSCs, collectively known as "embryo-like/embryo models". These models aim to simulate the structure and developmental process of real human embryos, providing unprecedented tools for studying early development^[29]. At the same time, they do not fully reproduce the biological complexity of in vivo human embryos, especially with respect to native tissue architecture, physical environment, and endogenous cell-cell interactions^[30]. Based on their cellular composition and developmental potential, these models are generally divided into two main categories: non-integrated embryo models and integrated embryo models.

Non-integrative embryonic models typically contain only intraembryonic cell types (such as epiblast cells) and contain little or no extraembryonic tissue. They do not have the potential to develop into a complete fetus and therefore face less ethical controversy, primarily being used to study specific tissues or developmental events. Their development can be traced back to 2000, when Itskovitz-Eldor et al. from Benvenisty's team at Rambam Medical Center discovered that suspended cultured hESCs spontaneously assembled into embryoid body (EB) structures, but these structures were highly disordered and randomly differentiated, quite different from the ordered developmental pattern of real embryos^[31]. In 2014, Warmflash et al. at Rockefeller University in the United States first constructed a 2D embryo (micro-patterned colony) structure spontaneously assembled from hESCs^[32]. Through BMP4-induced differentiation, they formed outer trophoblast-like cells, inner ectoblast-like cells, and intermediate mesoderm/endoderm-like cells, laying the foundation for simulating an early neurulation-like event^[32]. In the following years, embryonic models of amnion, organizer, and early neural development were successively established, and polarity was observed in in vitro embryos for the first time^[33-38]. In 2020, Moris et al. at the University of Cambridge, UK, treated hESCs with WNT agonists to obtain a 3D gastruloid structure, simulating the process from gastrulation to somitogenesis after implantation in human embryos^[39]. These systems are particularly useful for dissecting selected morphogenetic events under controlled conditions, but they still represent partial reconstructions rather than intact equivalents of in vivo embryos. In 2023, Yamanaka et al. at Kyoto University, Japan, further optimized in vitro culture conditions to induce hESCs into an axiotoid structure, breaking the anterior-posterior symmetry and simulating the head-tail morphology and molecular characteristics of somitogenesis in human embryos^[40].

Integrative embryonic models incorporate both intraembryonic and extraembryonic cell types, enabling a more complete simulation of the structure of a real embryo. For example, blastoids mimic the pre-implantation blastocyst structure, while post-implantation embryonic models simulate the post-implantation developmental stages. In 2021, Liu et al. and Yu et al. at Monash University in Australia reported methods for generating human blastoids using naive-state hPSC induction and somatic cell reprogramming, respectively^[41,42]. Liu's team's work is particularly noteworthy; they successfully constructed iBlastoids by reprogramming fibroblasts into induced pluripotent stem cells (iPSCs), providing an important platform for studying genetic diseases using patient-specific iPSCs^[41]. In 2023, Weatherbee et al. overexpressed key regulatory factors of trophoblast and primitive endoderm cells in hPSCs via transgenic methods, inducing trophoblast-like and primitive endoderm-like cells to spontaneously assemble into structures containing three lineage-like cell types, reproducing structures similar to amniotic sacs, extraembryonic mesoderm cells, and primordial germ cells (PGCs)^[43]. In the same year, Oldak et al. at the Weizmann Institute of Science in Israel systematically optimized the induction system of human hPSCs, constructing an embryo-like structure in vitro that simulates a real human embryo, known as a synthetic complete embryo model (SEM) based on stem cell synthesis. This model relatively completely simulates the key developmental events of a human embryo after implantation in the uterus until the onset of gastrulation^[44]. Even in these more integrated systems, increased structural resemblance does not mean full developmental equivalence to in vivo embryos, and careful benchmarking against real embryonic data remains necessary.

After the aforementioned technologies enabled embryo culture to reach the post-implantation stage and simulate the development of key primitive organs such as the neural tube, heart, and aorta-gonad-mesonephros region hematopoietic region^[45-47], recent embryo-like studies have expanded beyond embryo-like structures to include joint modeling of the maternal-fetal interface. In December 2025, Molè et al. from Peter J. Rugg-Gunn's team developed the Cell-Enhanced Receptivity Endometrial Scaffold (CREST) technology. This system successfully simulated the tissue receiving the implanted embryo (superficial endometrium) and cultured both real embryos and embryo-like structures to day 14^[48]. Following this, in January 2026, the laboratories of Li et al. not only constructed a small-scale microarray of the maternal-fetal interface system but also developed a relatively complete drug screening process based on it, enhancing the clinical trial value of this system^[49].

The main contributions of embryo-like technologies lie in their scalability, reproducibility, and operability. Compared to scarce human embryo samples, hPSCs (including hESCs and iPSCs) can be expanded indefinitely, enabling large-scale experiments. The introduction of iPSCs is particularly important because they can be obtained by reprogramming adult cells, avoiding the ethical controversies associated with using embryonic stem cells, while also enabling the creation of patient-specific disease models. Furthermore, these models are easily genetically manipulated, allowing researchers to precisely analyze the function of specific genes during development. However, these models have also raised complex ethical debates, particularly integrative models, which structurally resemble real embryos, making their developmental potential and ethical standing a focus of attention for the scientific community and the public^[50]. With the rapid accumulation of single-cell reference atlases, it has become increasingly common to compare in vitro embryo models with real embryonic datasets in order to identify both shared features and model-specific differences.

A key question in the aforementioned embryo-like studies is the extent of their similarity to real embryos. Can they accurately replicate the development of an in vivo embryo? To answer this question, conventional experimental methods require collecting real human embryos at the corresponding stage or embryos for immunofluorescence staining verification to ensure that the key expression characteristics of the cell types obtained from embryo-like studies are similar to those of real embryos. However, this verification requires precious in vivo embryo samples and has limited throughput. Bulk RNA-seq has been used to compare the transcriptomic differences between induced amniotic membranes and hPSCs as a control group to demonstrate that the induced cells possess the expression characteristics of the amniotic membrane^[38] [Figure 3].

Figure 3. Timeline of Embryo Model Technology Development. Publications utilizing single-cell and spatial omics technologies are highlighted. AM: Amnion; ChC: chorionic cavity; EB: Embryoid body; Ect: ectoderm; Endo: endoderm; Epi: epiblast; ES: embryonic stem cell; ExEM: extraembryonic mesoderm; hESC: human embryonic stem cell; hiPSC: human induced pluripotent stem cell; hnPSC: human naive pluripotent stem cell; hPSC: human pluripotent stem cell; Hypo: hypoblast; Mes: mesoderm; PL: placenta; PS: primitive streak; SEM: synthetic complete embryo model; Tb: trophoblast; TE: trophectoderm; xEM cells: extraembryonic mesoderm cells; YS: yolk sac.

scRNA-seq technology has opened the door to comprehensive mapping studies of embryonic development mechanisms

The rise of scRNA-seq substantially changed the study of early human embryonic development. For the first time, researchers could systematically resolve developmental heterogeneity at cellular resolution and begin to assemble transcriptomic atlases across successive embryonic stages. This shift was important not only because it expanded cell-type discovery, but also because it provided a practical reference framework for comparing embryo-like models with real embryos in vivo. At the same time, scRNA-seq still profiles development in a dissociated manner. It can describe cell states with increasing precision, but it cannot by itself explain how those states are positioned within intact embryonic structures or how spatially adjacent tissues shape developmental trajectories^[39]. Therefore, the development of embryo-like cells and the verification of their accuracy are inseparable from the development of sequencing technology, and even more so from the study of the mechanisms of real embryos in vivo (as prior knowledge for verifying the accuracy of embryo-like cells). In the context of preimplantation development, in 2013, Xue et al. first applied scRNA-seq to human and mouse embryos^[51]. By comparing gene expression profiles from the oocyte to morula, they not only identified specific expression patterns of maternal and paternal alleles (such as paternal activation of the CDCA2 gene), but also constructed stage-specific functional modules using weighted gene co-expression network analysis, elucidating the sequential activation process of cell cycle, transcriptional regulation, and metabolic pathways in embryonic development^[51]. This work not only laid the foundation for cross-species comparative studies of embryonic data but also highlighted the advantages of scRNA-seq for analyzing low-abundance transcripts and allele expression, directing research focus to deeper lineage-separation mechanisms.

Following the morula, researchers began to focus on the peri-implantation period. In 2016, Xue et al. used scRNA-seq to analyze human preimplantation embryos (3-7 days post-fertilization), revealing for the first time at the single-cell level the transcriptomic differences in TE, ICM, epiblast (EPI), and primitive endoderm (PE)^[51]. The study found that the expression of lineage-specific genes (such as GATA3, SOX2, and GATA4) begins to differentiate at the day 5 stage, and that X chromosome dose compensation is achieved through biallelic downregulation rather than complete inactivation, challenging the traditional view of X chromosome inactivation in mouse models^[51]. Furthermore, in 2021, Meistermann et al. reconstructed the developmental trajectories of human and mouse preimplantation embryos through integrated pseudotime analysis, finding that transcriptome segregation of human TE and ICM occurs during the blastocyst expansion stage (B2-B3), later than the mouse morula stage, indicating significant differences in developmental timing between species^[52]. In 2020, Xiang et al. developed a 3D culture system that successfully simulated human embryonic development from implantation to pre-gastrulation^[14]. scRNA-seq analysis showed that 3D cultured embryos could form structures such as the EPI, amnion, and yolk sac, and revealed the molecular pathways by which trophoblast cells differentiate into the cytotrophoblast, syncytiotrophoblast, and extravillous trophoblast^[14]. In the same year, Tyser et al. performed the first scRNA-seq analysis on CS7 gastrulating embryos and constructed a cellular atlas of germ layer differentiation and organ primordia formation, including cell groups from the ectoderm, mesoderm, and definitive endoderm, providing a valuable resource for the study of human gastrulation^[53]. With the accumulation of data, a reliable reference dataset is urgently needed for the identification of real embryonic stages and cells, or for the reliability assessment of embryonic-like systems. Therefore, recent research has focused on constructing a single-cell atlas of embryos across a longer developmental period. In 2023, Xu et al. published a single-cell transcriptome atlas of human embryos at weeks 4 to 6 (CS12-16), covering hundreds of cell clusters across 18 developmental lineages, revealing the axial patterning mechanisms of organs such as limb buds and the neural tube^[54]. The study also found that LIN28A plays a conserved regulatory role in the transition between developmental stages in vertebrate embryos, indicating the existence of punctuated stages in the developmental process. These findings were validated on human punctuated sections from CS13^[54]. In 2025, Zhao and colleagues integrated six independent datasets to develop an early embryogenesis prediction framework that provides harmonized reference resources and annotation benchmarking tools spanning human development from the blastocyst stage through gastrulation. This resource offers a practical foundation for assessing the fidelity of pluripotent stem cell-derived embryo models, including blastoids and gastruloids, and for improving cell-state assignment in spatially limited or low-coverage embryonic transcriptomic datasets^[55]. Yet even increasingly complete single-cell atlases do not restore the spatial coordinates, tissue neighborhoods, or anatomical continuity that are central to embryonic morphogenesis.

ST brings a new paradigm to embryo research

The second question is which cell states arise during early human embryogenesis and how transcriptomic data can resolve them. We address it by first introducing the major spatial profiling platforms and then showing how each has been applied to embryonic tissues. ST became important in this context because it began to reconnect molecular identity with physical location. This did not simply add spatial coordinates to existing datasets. More importantly, it allowed developmental time, tissue architecture, and cell-state information to be interpreted in relation to one another, making it possible to revisit questions that dissociation-based profiling could only approach indirectly. In practice, however, no single spatial strategy is equally suitable for all embryonic samples and research aims. Choices about assay design and downstream analysis depend on sample scarcity, developmental stage, spatial scale, transcriptome coverage, and the type of biological inference required. To make these considerations more explicit, we summarize below a comparative framework linking representative embryonic sample types and research aims to commonly used spatial assays and analytical strategies in current embryonic research [Table 2].

The following sections build on this overview and examine spatial profiling technologies and analytical strategies in greater detail, with emphasis on the embryological questions they help address [Box 1].

Bin selection and cell segmentation: a trade-off between sequencing depth and spatial precision

One of the first analytical challenges in high-resolution ST is that increased spatial precision often comes at the cost of increased data sparsity. In embryonic research, this trade-off is especially consequential because rare cell states and fine anatomical boundaries may be lost if analysis units are defined too coarsely, whereas excessively small units may lack sufficient transcript depth for robust interpretation. A common response to this problem is to aggregate neighboring capture spots into more information-rich analytical units. In practice, this is usually done in two ways: by merging neighboring spots into square bins or by using image- or transcript-informed segmentation to generate near-single-cell units, here referred to as Cell Bins^[97] in this paper.

Square bin remains a widely used method for embryonic atlas construction due to its computational efficiency and its ability to preserve transcriptomic information at the bin level. However, it inherently involves a trade-off between bin size and spatial resolution, and larger bins may reduce single-cell resolution and obscure rare cell populations. In the field of human embryos, the number of detected genes per unit is typically controlled between 400 and 1,300 [Table 2] to achieve a balance between capture quality and single-cell resolution. If the number of genes or the sequencing depth (total unique molecular identifier counts) is too low, the clustering algorithm will lack sufficient information; conversely, if spatial bin size is too large, it will lose its near-single-cell resolution, hindering the identification of valuable cell types. To address this issue, for example, 10X Genomics recommends starting with an 8 µm square bin to identify cell types through clustering when using data generated by Visium HD sequencing technology, and gradually increasing the square bin range to 16 µm to empirically determine an optimal balance between quality and resolution. Other technologies also employ similar strategies to select appropriate square bins; for example, in Stereo-seq with a spot spacing of 500nm, bins from 20 (10 µm) to 50 (25 µm) can be selected. In practice, Square Bins remain particularly useful when the priority is to preserve transcript depth and analytical stability across large embryonic sections, rather than to approximate single-cell boundaries as closely as possible [Table 4].

Ideally, each analytical unit should be a real, bounded cell rather than a bulk. This is crucial for preserving heterogeneity in dense and small tissues like the embryo. Cell segmentation delineates cell (nuclei) boundaries by identifying structures in DAPI- or antibody-stained images, and subsequently aggregates the spatial transcripts (spots) falling within each delineated region into a Cell Bin, which approximates a single cell. Cell Bin can greatly reduce erroneous transcript allocation caused by square bin. As sequencing depth and segmentation algorithms improve in ST platforms, Cell Bin segmentation is becoming increasingly feasible^[106]. The evolution of cell segmentation algorithms has progressed from image processing techniques to deep learning-based paradigms. Early methods, such as the Watershed algorithm, are mathematical morphology-based segmentation algorithms grounded in topological principles. They infer cell boundaries by iteratively analyzing pixel intensity gradients in the stained image, treating them analogously to a topographic landscape to identify catchment basins (corresponding to cells) separated by Watershed lines^[107]. However, this algorithm struggles to accurately delineate cell boundaries in high-density scenarios, often leading to over-segmentation or under-segmentation^[108], and is now primarily used as a baseline for benchmarking more advanced methods. Cell segmentation is a long-standing challenge in computer vision and image analysis, and deep learning-based algorithms are currently the mainstream approach. For instance, CellPose employs a U-Net-based neural network to predict spatial flow fields (gradients) for each pixel. By learning to predict these vector fields, the algorithm directs pixels toward the centroid of their respective cells, effectively grouping them to reconstruct cell morphology^[77]. Through subsequent iterations, CellPose 2.0 incorporated denoising modules^[75,76], significantly enhancing its robustness and generalizability when processing low-quality images. In contrast, StarDist utilizes a U-Net to predict parameters defining star-convex polygons^[109]. This representation, which models objects as convex when viewed from a central point, performs better at segmenting densely packed nuclei and accommodating non-rectangular cell morphologies. However, the above methods rely on high-quality images that match the transcripts. When high-quality staining images are lacking, methods are needed to parse cell boundaries from transcript data and spatial information. This technical approach explores segmentation strategies that do not rely on traditional imaging information, such as EM + BP, gOTSU, Moran's I, VI + BP, and Baysor methods^[63,78]. These methods rely solely on spatial representation data features and employ various mathematical methods to segment cell boundaries. Notably, Baysor's Markov random fields (MRFs) framework is not a transcript-based segmentation method; rather, it is an algorithm that directly identifies cell affiliation of transcripts through clustering without requiring segmentation. A conceptually similar approach is embodied in FICTURE. It learns a latent representation with a probabilistic generative model and, at each spatial location (pixel/spot), infers a probability distribution over “factors” or states. This enables direct derivation of cluster or annotation maps without performing explicit cell segmentation^[110].

Although cell segmentation technology has not yet been widely used in human embryonic research, it has been demonstrated in other mammalian embryology research to accurately cluster cells, identify rare cell types, and enable co-visualization with microscopic images, thereby enhancing the interpretability of spatial gene expression patterns. Xie et al. used Cellpose to process three mouse embryos from the E7.5-E8.0 stage and performed cell segmentation based on ssDNA images^[111]. Following integration with scRNA-seq-derived cell annotations, they achieved highly concordant results, reconstructed the spatial expression gradient of HOX genes along the anteroposterior (AP) axis, and delineated the precise molecular boundary between the embryonic and extraembryonic tissues^[111]. Similarly, Stereo-seq's Cell Bin was applied to mouse E9.5-E11.5 embryos^[112] and salamander embryos^[113] to reveal the asynchronous development of the left and right lungs and the neurodevelopmental patterns of brain regeneration. In zebrafish, research spanning gastrulation to organogenesis has employed weMERFISH-based Cell Bin segmentation to reveal dynamic changes in gene expression across multiple cell types. The weMERFISH technique enhances gene detection efficiency on three-dimensional tissue sections^[114].

In early human embryonic research, the choice between Square Bins and Cell Bins is not merely a preprocessing step, but a decision that directly affects the balance between anatomical precision and transcriptomic robustness. Due to staining and algorithmic errors, cell segmentation cannot identify cell boundaries with absolute precision, resulting in 40%-60% of transcripts being unable to be assigned to neighboring cells because they are far from the cell boundaries^[110]. Therefore, the Cell Bin approach is advisable primarily when sequencing depth is high, as this is a prerequisite for approaching true single-cell resolution and maximizing annotation accuracy. Conversely, Square Bins are recommended when sequencing depth is limited or when the tissue of interest contains multinucleated or anucleate cells (e.g., some blood cells), which are challenging for nucleus-based segmentation algorithms [Figure 4].

Figure 4. Selection of different analysis unit segmentation strategies based on sample condition and data quality. The square lattice area represents the sequencing chip, and the blue spots indicate the regions where RNA has been captured. The diamond-shaped boxes contain specific criteria for determining the segmentation strategy. The choice between these strategies determines the trade-off between transcriptomic depth and spatial resolution, directly affecting the detection of rare cell populations and fine anatomical structures.

Leveraging deconvolution and read-depth enhancement for embryonic studies

Deconvolution becomes important in embryonic ST when the spatial assay does not directly resolve individual cells, but the biological question still requires cell-type level interpretation. This situation is common in human embryonic studies, where low-resolution ST platforms, limited sample availability, and scarce matched sections often make direct cell-level annotation difficult. In such settings, deconvolution uses scRNA-seq-derived reference profiles to infer the likely cellular composition of each spatial spot or bin, thereby improving the interpretability of spatial maps.

The landscape of deconvolution algorithms encompasses a variety of approaches, including complex probabilistic models and deep learning-based methods^[115]. In practice, the main distinction is not simply algorithmic form, but whether the method relies on matched single-cell references, how well it tolerates platform and batch differences, and whether it is intended primarily for annotation transfer, composition inference, or expression enhancement. Commonly used deconvolution methods in embryonic research include RCTD^[116] and Cell2location^[117]. Both methods employ generalized linear models, utilizing Poisson (RCTD) and negative binomial (Cell2location) distributions, respectively, to infer the cellular composition (i.e., the proportion of each cell type) within each ST spot or bin from scRNA-seq-derived reference profiles. Although these methods were published several years ago, their performance remains robust and has been widely validated across diverse ST platforms^{[115,118,119]}. In 2021, Tangram introduced a deep learning approach to ST deconvolution. Building upon deconvolution, it further integrates single-cell and ST expression matrices to impute or enhance gene expression information in the spatial data, offering unique advantages for gene expression prediction. It continues to be actively used in embryonic research^[120]. To improve accuracy, Xiao Yang and Jinjin Xiang developed the Spatialsmooth software package, which integrates multiple deconvolution methods and uses a convolutional autoencoder to smooth different cell-type combinations in space^[121]. Beyond continuous improvements in accuracy, optimization of deconvolution algorithms primarily focuses on three directions: (1) enhancing computational scalability to handle ever-increasing cell numbers (e.g., jMF2D^[122], NLSDeconv^[123]); (2) integrating histological image information to improve reliability (e.g., SpaDecon^[124], GraphST^[125], FAST^[126]); and (3) enabling deconvolution without reliance on scRNA-seq reference data (e.g., STdeconvolve^[127], RETROFIT^[128], gwSPADE^[129], SpatialDeX^[130]). For embryonic applications, these three directions correspond to three recurring practical needs: scaling to increasingly large atlases, improving robustness in histologically complex tissues, and reducing dependence on perfectly matched reference datasets. Given the scarcity of early human embryonic samples and the practical challenge of obtaining multiple replicates, reference-free deconvolution methods (those not reliant on scRNA-seq data) hold particular promise for this field.

Deconvolution techniques have been successfully applied to low-resolution ST data to resolve the spatial localization of diverse cell types within human embryonic tissues. Xu et al. (2023) employed the RCTD algorithm to map 239 scRNA-seq-derived cell type annotations onto two CS13 human embryonic tissue sections profiled by 10X Visium (55 µm resolution)^[54]. This successfully localized both distinct and subtle anatomical structures visible in H&E staining, such as the optic vesicles, renal epithelium, head muscles, and pancreas^[54]. Similarly, Cell2location has been applied in limb development studies to precisely identify key mesenchymal cell subsets, thereby delineating the spatial expression patterns of genes implicated in limb malformations within specific niches. Beyond the probabilistic models discussed, the deep learning-based method Tangram has been applied to human embryonic brain sections at the postconceptional week (PCW) 4 stage. It generated annotations for brain structures, including the neural tube, that align accurately with established anatomical landmarks^[57]. Collectively, these studies show that deconvolution is most convincing when supported by anatomically interpretable landmarks and by reference data that are reasonably matched in developmental stage.

Current applications demonstrate that deconvolution excels at identifying spatially dispersed cell populations (e.g., immune cells) and is also highly effective for detecting rare cell types. To mitigate batch effects arising from differences in technology platforms, biological sample variability, and experimental protocols, it is ideal to perform deconvolution using matched, concurrently generated scRNA-seq data from the same embryonic stage as the reference. Therefore, this algorithm is widely used in human embryonic samples from later, relatively easily obtainable organ development stages. This is because the embryonic samples from a relatively later period can be obtained more readily and repeatedly.

Finding the anatomical structures of spatially adjacent tissues: spatial domain identification

Spatial domain identification becomes necessary when the goal is not only to annotate cell states, but also to recover anatomically coherent tissue regions from spatial data^[131]. This is particularly important in embryonic research, where neighboring tissues often differ sharply in developmental potential and signaling environment despite sharing partially overlapping transcriptional features. Traditional clustering methods based only on gene expression similarity therefore tend to miss one of the key features of embryo organization: the continuity of local structure in space. To fill this gap, a range of spatial domain identification methods have been developed to integrate gene expression with spatial topology and, in many cases, histological context.

The application of convolutional neural networks (CNNs) and graph neural networks (GNNs) represents a standard computational approach to address this challenge. In practice, these methods mainly differ in how they weight spatial continuity, histological information, computational cost and robustness to spatially dispersed cell populations. Since the recognition task often involves co-registered histological image data, it naturally extends into a multimodal deep learning problem. Statistical model-based methods, such as Hidden Markov Random Field models^[132], were the first to adapt to spatial multi-omics^[133]. These methods are based on the assumption that neighboring cells have similar gene expression (BayesSpace^[134]). Subsequently, deep learning-based methods have rapidly advanced in the field of spatial domain identification. Initial efforts focused on leveraging CNNs to extract features that capture spatial context. stLearn^[135] uses a pretrained CNN model, ResNet-50, to extract image features and integrates them with a smoothed representation matrix to obtain a new matrix. Based on the integrated matrix, traditional Louvain/Leiden clustering is performed, and further sub-clustering based on the DBSCAN algorithm (an algorithm based on pure spatial location information clustering) is performed on spatially discrete clusters to further distinguish spatially distant clusters of the same type. SpaGCN^[136] integrates gene expression, spatial information, and H&E image information using graph convolutional networks (GCNs). STAGATE significantly improves upon SpaGCN by introducing an adaptive graph attention autoencoder to dynamically learn the similarity between neighboring points^[137]. Following this, methods for integrating multimodal information in deep learning to achieve spatial domain dimensionality reduction have proliferated (e.g., DeepST^[138], SpaceFlow^[139], GraphST^[125], SpaSEG^[140]).

In the past two years, the introduction of the transformer architecture and large-scale pretrained models has substantially addressed a key limitation of previous methods: incorrectly classifying spatially non-clustered cell types, such as immune cells, into nearby anatomical hierarchical structures (TransformerST)^[141]. spaLLM utilizes the foundation model scGPT^[142] pretrained on a single-cell dataset to obtain the spatial group expression enhancement matrix, and then fuses multimodal information through graph neural networks and multi-view attention mechanisms, significantly outperforming existing methods on multiple datasets^[143]. Similarly, Novae adapts the SwAV visual self-supervised learning framework to ST, enabling accurate spatial domain identification and dimensionality reduction even for complex tissue sections^[144]. While the aforementioned neural network-based methods are gradually improving, they often come with substantial computational costs. In contrast, Banksy adopts a distinct strategy. It constructs a feature matrix for each cell by integrating its own gene expression profile with neighborhood information. The latter includes the average expression of neighboring cells and the spatial gradient (magnitude and direction) of expression changes. This composite representation is termed the Banksy matrix. This matrix encapsulates the intrinsic cell expression, the local microenvironment expression profile, and the spatial patterns of expression variation. The weight of the cell's own expression information is flexibly adjusted by the λ parameter. Standard clustering algorithms like Leiden or Louvain, when applied to the Banksy matrix, can identify spatially coherent domains while simultaneously preserving the identity of spatially dispersed cell populations. Compared to neural network methods, it greatly saves computation time while maintaining competitive and stable accuracy^[145]. The algorithms mentioned above all emphasize spatial domain identification on a single two-dimensional slice, while SpaDo emphasizes the continuity of three-dimensional spatial domains on continuous slices. This algorithm can accurately transfer scRNA-seq annotations to continuous slices, i.e., supervised identification of spatial domains, and can also accurately identify adjacent spatial domains on three-dimensional continuous slices without supervision^[146].

Using spatial domain identification algorithms alone struggles to balance the contradiction between the heterogeneity of single cells and the continuity of their neighborhoods. This means they may overlook small clusters or scattered cells in space and categorize them into neighboring large anatomical cell types. In embryonic analysis, spatial domain identification is most informative when the objective is to delineate contiguous tissue architecture or to locate developmental organizers within complex local neighborhoods. However, because these methods may obscure scattered or rare cell populations, they are best interpreted together with deconvolution, single-cell references, or conventional clustering results when higher annotation fidelity is required [Figure 5].

Figure 5. The workflow for cell annotation in human embryonic spatiotemporal omics studies, from raw data processing (bin selection and deconvolution) through spatial domain identification and cell-type assignment, enabling the transition from spatial measurements to biologically interpretable cell-type maps.

Spatial domain identification technology is mainly used in early embryonic research. Chen et al. applied Stereo-seq in mice and proposed spatially constrained clustering^[73]. Based on the union of spatial proximity maps and transcriptome similarity maps, they generated a combinatorial map and then used Leiden to perform clustering on the combinatorial map^[73]. This relatively simple and computationally efficient spatial domain identification strategy was subsequently extended to study human early embryonic development during the peri-gastrulation period^[2,3,56], also utilizing Stereo-seq technology. This application helped researchers locate key structures, such as the anterior visceral endoderm (AVE). This illustrates why spatial domain methods are especially valuable in early embryonic samples, where identifying small but spatially organized signaling centers can be more informative than transcript-only clustering.

Human early embryonic cell migration and cell fate decisions in spatiotemporal dimensions

In embryonic development, lineage progression and cell migration are inseparable from space: cells do not differentiate in isolation, but within changing tissue geometries, signaling gradients, and neighborhood relationships. For this reason, trajectory inference in ST is valuable not only because it estimates developmental order, but because it allows inferred fate transitions to be interpreted in their anatomical context. This is especially important during gastrulation and early organogenesis, when transient cell states and directional movements are tightly coupled to morphogenesis.

In practical analysis, mainstream differentiation methods can be divided into two main categories: graph-based or minimum spanning tree-based methods and velocity-based methods based on mRNA splicing. The main distinction in practice is whether the method infers developmental progression from transcriptional similarity alone or estimates directional change from RNA dynamics. Algorithms such as Monocle2^[147], Monocle3^[148], and Slingshot^[149] work by constructing a cell-to-cell graph within a low-dimensional embedding [e.g., derived from DDRTree or uniform manifold approximation and projection (UMAP)] of the gene expression matrix. This graph is then used to infer a pseudotemporal ordering of cells along putative differentiation trajectories. Trajectory inference methods that have developed alongside spatial omics technologies primarily rely on velocity-based approaches. Traditional velocity methods, such as the steady-state model Velocyto^[150], the dynamic model scVelo^[151], and deep learning-based CellDancer^[152] and DeepVelo^[153], work by comparing the abundance of unspliced and spliced mRNA in cells to infer the instantaneous direction of gene expression changes, thereby predicting the future differentiation state of cells. Although these methods do not directly utilize spatial information, the calculated pseudotime, latent time, or vector projection can be projected onto the spatial dimension (or onto a 3D point cloud if reconstruction is involved), allowing visual observation of cell differentiation in space. These methods typically require principal component analysis (PCA) dimensionality reduction, and high-quality spatial PCA results generated by SpatialPCA can be used as preliminary data to further improve performance in ST^[154].

The past three years have also seen the proposal of velocity algorithms specifically designed for ST. These methods are particularly attractive for embryonic studies because they attempt to couple dynamic inference with local spatial context, rather than treating developmental progression as independent of tissue position. For high-resolution ST based on NGS, loom files containing splicing information can be directly generated from the aligned BAM files. For instance, Stereopy^[97] can directly generate loom files from Stereo-seq data. For other platforms like 10X Visium HD, the BAM files first need to be merged and processed according to the chosen spatial bin size, after which tools like Velocyto^[150] can be used to generate the loom files. These upstream files can serve as the basis for various velocity algorithms. Spateo^[63] introduced the concept of vector fields from geoinformatics, combined with Dynamo^[155] to model the differentiation rate, and determined the spatial differentiation and migration direction of cells in mouse developmental data across time periods. TopoVelo^[156] employs a GNN to model interactions between spatially neighboring cells, thereby deriving velocity parameters that are intrinsically coupled to spatial context. It can be further integrated with lineage inference tools like scFate^[157] to estimate the direction and rate of cellular migration within the tissue space. spVelo^[158] is a generative model designed for ST. Building upon advancements in RNA velocity analysis, it specifically addresses the challenge of batch effect correction when integrating data from multiple developmental time points. For techniques like MERFISH based on fluorescent probes, the velocity calculation results from scRNA-seq need to be deconvolved spatially (Cell2Fate^[159], SIRV^[160]).

In actual differentiation analysis of embryonic development, stemness genes, the CytoTRACE stemness scoring algorithm^[161,162], and the researcher's prior knowledge are often used to comprehensively identify cells in the current differentiation lineage that are suspected to be at the differentiation origin. Next, traditional methods such as minimum spanning trees and the Velocity algorithm are used to determine the differentiation path in a combination of supervised and unsupervised methods. A representative application is the study of human CS7 embryos by Cui et al.^[56]. They projected pseudotime inferred by Monocle3^[148] onto both UMAP embeddings and the actual 3D embryo model. These pseudotime projections were then integrated with inferences from the dynamic model of scVelo^[151] and the partition-based graph abstraction (PAGA) method^[163]. These three complementary approaches were used to reconstruct the spatial differentiation trajectory of the mesoderm lineage. In the same study, analysis of PGC lineage differentiation utilized the scVelo dynamic model within the UMAP space. This suggested an epiblast origin for PGCs and pointed to a potential differentiation pathway from amniotic ectoderm towards PGCs. In a study of mouse E9.5 embryos^[102], the authors also used Monocle3 to infer pseudotimes of neuromesodermal progenitor cell (NMP) lineages to presomitic mesoderm (PSM), somites, and neural tube, etc., and projected the results into three-dimensional space. They found that the degree of transcriptome change was not a simple linear relationship with spatial distance, revealing a gene expression change pattern along the dorsoventral (DV) axis. The studies mentioned above employed high-resolution ST methods. When using platforms with relatively lower spatial resolution, such as in a study of human embryonic limb development^[58], researchers have successfully combined multiple computational approaches. These include Diffusion Maps^[164], force-directed graph layout^[165], and PAGA to infer differentiation trajectories of cells originating from the lateral plate mesoderm. This revealed a differentiation network that radiated from six core mesenchymal progenitor cells at PCW5, differentiating into various cell types such as cartilage, bone, and tendon.

Traditional methods like Monocle2 and Monocle3 often require extensive prior biological knowledge to determine reliable cell types at the origin of differentiation, which can introduce subjective bias. While RNA velocity provides an unsupervised framework, the accuracy and assumptions underlying its inferences have been subjects of ongoing discussion and refinement^[166]. Furthermore, standard ST data often represent transcript counts from spots or bins containing multiple cells, rather than pure single-cell measurements. Consequently, the observed spliced and unspliced mRNA counts are aggregated signals, which may not accurately reflect the kinetics within individual cells. These issues all challenge the accuracy of differentiation analysis with ST data. Therefore, until a completely reliable, unsupervised single differentiation method emerges, complex differentiation events like embryonic cell differentiation require the use of multiple differentiation methods to cross-validate and determine the direction and sequence of differentiation. In embryonic research, trajectory and velocity analyses are most informative when they are supported by recognizable anatomical landmarks, temporally ordered sampling, or independent lineage knowledge. They are especially useful for reconstructing directional processes such as gastrulation movements, germ-layer transitions, and early organ-specific patterning. However, because inferred trajectories remain sensitive to data sparsity, batch effects, and the assumptions of individual models, they should be interpreted as spatially constrained hypotheses rather than direct proof of lineage progression or migration.

Analyzing developmental signals: communication algorithms under spatial constraints

In embryonic development, signaling is not simply a property of cell type, but of spatially organized interactions among neighboring tissues. This is especially important during gastrulation and early organogenesis, when local signaling centers and short-range inductive interactions help establish body axes, tissue boundaries, and organizer function. Under these conditions, spatially constrained communication analysis becomes valuable not only because it reduces false positive interactions caused by tissue dissociation, but also because it helps reconstruct the signaling architecture of developing embryos. The CellChat v2 upgrade introduces spatial constraints by defining spatial neighborhoods, which identify each cell's neighbors based on calculating the Euclidean distance threshold between cells, thereby effectively filtering out false positive signals introduced by tissue dissociation^[167]. CellPhoneDB v5, a classic communication analysis tool designed for ST upgrades, introduces the concept of microenvironments. When inferring communication, it only considers the interactions between cell types belonging to the same microenvironment, thereby refining the communication network and making it more consistent with the actual structure of the tissue^[168]. Furthermore, new algorithms designed natively for ST data are continuously emerging. SpaTalk uses graph networks and knowledge graph methods to model and score the co-expression of ligand-receptor-target signaling networks between spatially adjacent cells^[169]. COMMOT, developed by the same team as CellChat, incorporates optimal transport theory to model ligand diffusion from secretory to receiving cells. This provides a more realistic simulation of signal propagation through tissue space^[170]. Signaling regulation operates across multiple hierarchical levels. Building upon COMMOT's output, FlowSig reveals more precisely how ligand-receptor interactions drive downstream signal output by regulating intracellular gene modules through a causal inference graph model that includes signal inflow, gene expression modules, and signal outflow^[171]. The limitation of the above methods is that they rely on known, human-summarized communication network databases such as CellChatDB, and cannot discover potential new communication control relationships. To address this limitation, OrgaCCC^[172] was developed. This method utilizes two independent variational graph auto-encoders to capture features from the cellular/spot domain and the gene domain, respectively. These two autoencoders are integrated by maximizing the similarity between their respective reconstructions of the gene expression matrix. The resulting intercellular relationship matrix is the inferred communication network. This method was tested on mouse placental ST data and identified receptor-ligand pairs that were not detected by other communication methods. Significant interactions between trophoblast cells and endometrial spiral artery endothelial cells were observed to anchor trophoblast cells within the vascular lumen, thereby ensuring blood perfusion and embryonic health. Real cells live in three-dimensional space and communicate with each other. Spateo^[63] developed a complete analytical framework. After rearranging cells using the concept of a spatial field, its communication analysis module integrates spatial location, gene expression, and morphological information to construct a spatially aware regression model to identify receptor-ligand interactions in three-dimensional space and reveal complete cascade signaling pathways.

Computational inference alone does not establish that cell-cell communication is active during embryonic development. We therefore distinguish four levels of evidence for inferred interactions. (1) A predicted interaction is inferred from ligand-receptor co-expression and serves only to formulate hypotheses. (2) A spatially supported interaction is additionally constrained by spatial proximity between ligand- and receptor-expressing cells, which raises confidence without providing functional validation. (3) An experimentally perturbed interaction is supported by perturbation experiments in embryo models. Such experiments provide direct functional evidence that the candidate interaction can influence downstream responses, although they do not by themselves establish a causal developmental mechanism. (4) A causally validated developmental mechanism is demonstrated through perturbation together with rescue and lineage tracing, a standard rarely achievable in human embryos. In the CS7-CS9 examples discussed below, we primarily report evidence at the first two levels, and these interactions should therefore be interpreted as spatially constrained hypotheses requiring experimental validation.

These analytical advances are particularly relevant to embryonic studies, where signaling centers are often small, spatially restricted, and developmentally transient. The evolution of communication algorithms, from single-level to multi-level regulation and from two-dimensional slices to three-dimensional structures, also influences research on embryonic development, particularly in the study of axial signal regulation and the identification and regulation mechanisms of Organizer cells. In the early CS7 stage of gastrulation, researchers, using CellChat v2, discovered that the AVE may serve as a major source of signaling molecules such as DKK1 and CER1, as well as the transcription factor OTX2. These secreted factors inhibit signaling activity associated with the anterior primitive streak, thereby helping to establish the embryo’s anteroposterior polarity^[56]. As gastrulation proceeds into the CS8 stage, the body axis undergoes further extension and regional specialization. Spatial communication analysis identified two key signaling centers: the notochord and the connecting stalk. These centers are predicted to jointly regulate the spatial gradients of signaling factors such as WNT, SHH, and BMP within the embryo^[2]. In the CS9 stage, the neural tube begins to close, and brain regions begin to specialize. Researchers identified the isthmic organizer at the midbrain-hindbrain boundary as a candidate signaling center. This organizer exhibits high expression of signaling factors like Fgf8 and Wnt1, which are crucial for patterning the anterior (midbrain) and posterior (hindbrain) regions^[3].

When selecting specific communication algorithms for embryonic data, it is important to note that not all spatial communication methods are currently compatible with Square Bin-based spatial sequencing units. Therefore, careful consideration should be given to cell segmentation methods when choosing a specific communication tool. Furthermore, it's necessary to consider whether using only known database interactions as a basis is sufficient for discovering new communication relationships. The development of spatial communication inference tools mainly focuses on the following directions: first, ensuring broad applicability across ST platforms with varying sequencing depths and segmentation methods; second, refining spatial distance constraints to reduce false positives while preserving genuine signals; and third, modeling multi-cascade signaling networks and long-range communication.

Embryo reconstruction in serial slices

Embryo reconstruction in serial slices becomes necessary when the biological question extends beyond isolated sections to embryo-scale organization. This is particularly important because body axes, signaling gradients, tissue continuity, and cell migration are inherently three-dimensional, whereas most spatial assays still generate data as two-dimensional slices. The close co-development of reconstruction algorithms and embryonic applications has therefore been driven by several intrinsic features of embryonic samples: their rarity, which makes maximal information recovery from each specimen especially valuable; their relatively small size, which makes serial sectioning technically feasible; and their highly ordered spatial organization, which is best interpreted in a continuous 3D framework. For this reason, embryonic datasets have become both a major application scenario and an important benchmark resource for the development of 3D reconstruction algorithms. For example, the MOSTA^[73] embryo database derived from the mouse embryo article published by BGI in 2022, besides being used for 3D reconstruction in that article, has also appeared as a test dataset in many subsequent 3D reconstruction algorithm articles (SLAT^[173], SPATEO^[63]). Similarly, the STitch3D algorithm was applied to integrate multi-slice spatial transcriptome and scRNA-seq data, achieving the first 3D spatial reconstruction of human heart development^[174]. In addition, sequencing technologies such as Magic-seq have directly performed continuous slices and 3D reconstruction of the developing mouse brain. These cases illustrate not only that embryonic ST data are key benchmark resource for reconstruction algorithms, but also that reconstruction is most informative when the aim is to recover spatial continuity across sections rather than to interpret each slice in isolation.

Three-dimensional reconstruction algorithms have been widely used in human embryonic research. ST data of the human CS8 embryo enabled researchers, for the first time, to accurately map gradients of morphogenetic signals and spatial distributions of cell types across a complete intact early human embryo^[2]. Specifically, researchers utilized the PASTE algorithm for initial alignment of sequential ST slices. The positions of key, spatially coherent cell populations (e.g., pre-annotated notochord cells) were then manually refined to generate a coherent 3D point cloud. Subsequently, Maya was used to model the outer shell of key tissue anatomy within the point cloud, and the modeling results were imported into Plotly (https://plot.ly) for dynamic visualization. Finally, a visualization webpage was built to publicly display the aforementioned three-dimensional virtual embryo model containing the expressed information [Figure 6].

Figure 6. The workflow for the 3D reconstruction of embryonic spatial transcriptomics data from serial sections. From left to right: serial sectioning and sequencing, algorithmic generation of the 3D point cloud, refinement of the point cloud using prior knowledge, and shell modeling.

This pioneering pipeline was subsequently referenced in several human embryology papers, including CS7^[56] and CS9^[3], which also constructed high-quality three-dimensional virtual embryos. Leveraging these reconstructions, researchers delineated the spatial distribution of signaling molecules along the AP and DV axes, characterized the 3D morphology of key tissue structures and signaling centers, and uncovered 3D cellular migration patterns during CS7 and CS8^[2]. At the CS9 stage, analysis revealed that NMPs are organized into a three-dimensional, bilayered structure along the DV axis. These findings show that reconstruction is particularly powerful when developmental interpretation depends on axial geometry, cross-section continuity, or spatially extended signaling architecture that cannot be inferred reliably from single slices alone.

From an analytical perspective, the evolution of 3D reconstruction algorithms has progressively addressed three recurring challenges in embryonic spatial data: cross-section alignment, non-rigid tissue deformation, and incomplete section continuity. This shift reflects progress toward more efficient, flexible 3D reconstruction of spatial transcriptomic data. Early attempts at 3D reconstruction often adapted traditional computer vision image registration algorithms, such as SIFT^[175]. These methods relied solely on pixel-level image features, entirely disregarding the biological information in gene expression similarity, and were incapable of handling the complex, nonlinear deformations inherent to embryonic development. Consequently, their application to biological data faced significant limitations^[176]. This was particularly problematic for embryonic samples, in which rapid morphological change and curved body plans make purely image-based alignment unreliable. A significant breakthrough came with PASTE in 2020, which achieved alignment of slices using ST data based on optimal transport theory^[96]. However, PASTE employed a rigid alignment strategy, which could not accommodate natural tissue deformations introduced during sectioning. Furthermore, its computational scalability was limited. Substantial progress commenced with the advent of high-resolution ST platforms and concurrent innovations in algorithmic design. The SLAT^[173] algorithm, proposed in 2023, breaks through the bottleneck of heterogeneous data integration. Its core innovation involves modeling spatial topology as a graph structure and employing a lightweight graph convolutional network trained with an adversarial strategy. This framework not only enhances compatibility and computational efficiency across diverse technological platforms but also enables cross-timepoint slice alignment, thereby facilitating the inference of differentiation and migration trajectories during mouse kidney development. Also released in 2023, the SPACEL algorithm leverages an adversarial graph convolutional neural network model and a global optimization strategy to perform well in cell type prediction, spatial domain recognition, and 3D alignment tasks^[177]. These methods improved cross-platform compatibility and made reconstruction more scalable in heterogeneous datasets. In 2024, the Spateo framework innovatively and effectively solved the problem of non-rigid alignment and morphological correction at the whole embryo scale through a probabilistic model of Gaussian processes and variational inference. Building upon this, it can automatically construct surface meshes from the aligned point clouds. These modules achieved excellent alignment results for both mouse and Drosophila embryos, accurately reconstructing complex morphological features such as the curled body plan of the mouse embryo. To address the common issue of non-consecutive or missing sections in ST datasets—which can disrupt 3D reconstruction—the SpatialZ algorithm was developed. It compensates for missing adjacent slices, ensuring the continuity of the reconstructed volume along the z-axis^[178]. Together, these advances moved the field closer to embryo-scale reconstruction under realistic conditions of deformation and incomplete sampling.

The value of 3D reconstruction in embryonic studies extends beyond visualization. Once embryo-scale spatial continuity is established, many downstream analyses become more biologically interpretable in three-dimensional context. For example, Spateo's spatial communication module allows researchers to study modular representations and cascaded communications in 3D space. Metrics such as stemness scores from CytoTRACE^[161,162], pseudotime from Monocle3^[148], and latent time^[151] from RNA velocity can be projected onto the 3D reconstruction, providing spatial context to understand the dynamics of cell differentiation and migration. Gene clustering algorithms developed specifically for ST, such as HotSpot^[179], can also more accurately identify co-expressed gene modules in three-dimensional space, helping researchers deepen their understanding of regional tissue functions.

Although 3D reconstruction algorithms have become increasingly sophisticated, assessing the accuracy of a reconstructed embryo remains a considerable challenge. In imaging-based reconstruction tasks, ground truth can sometimes be obtained from an independent modality. This is not possible for human embryonic ST: once a section has been sequenced, the same tissue cannot be re-sequenced for validation. Reconstruction accuracy must therefore be assessed through alternative strategies. Several complementary checks can help evaluate reconstruction accuracy. Cross-validation with histology provides a qualitative but essential check. Overlaying spatial transcriptomic spots on H&E or immunofluorescence images from adjacent sections can expose misalignments that transcriptomic data alone would not reveal. Alignment error can also be quantified, for example as the root mean square deviation (RMSD) between corresponding spots in adjacent sections. Independent anatomical landmarks provide a further check of global consistency. The relative positions of the notochord, the neural tube and the developing cardiac primordium can be compared with the expected anatomy. Such comparisons confirm that biologically meaningful spatial relationships are preserved in the reconstruction. To make 3D reconstructions transparent and interpretable, we recommend that future studies report: (1) the number and thickness of sections used, and the proportion of sections that were missing or discarded; (2) the alignment algorithm employed, any manual refinement steps, and the anatomical criteria used for curation; (3) quantitative alignment metrics, such as RMSD or correlation coefficients between neighboring sections; and (4) a sensitivity analysis relating variation in alignment parameters or bin size to downstream biological interpretation.

Despite these advances, current 3D reconstruction methodologies still face several limitations. First, most current algorithms utilize only sequencing-derived gene expression and spatial coordinate data, largely neglecting the complementary information available from high-resolution histology images. Future multimodal 3D reconstruction software that integrates transcriptomic and imaging data could not only improve alignment accuracy but also produce more anatomically realistic and visually informative models. Second, sequencing of continuous thin sections is too costly and unsuitable for larger samples such as late-stage embryos. A low-cost end-to-end thick-section ST-running sequencing technology needs to be developed to balance the cost of large-sample reconstruction^[89]. Finally, despite improvements in alignment accuracy, the process often still requires manual curation. Researchers must rely on prior biological knowledge to select tissue types with clear spatial continuity to guide and validate z-axis alignment. This places certain demands on the researchers' biological literacy and introduces some subjective bias into the alignment process. In embryonic research, serial-slice reconstruction is most informative when the goal is to interpret cell states, signaling gradients, and tissue morphology within a continuous three-dimensional framework. Its main limitation remains that alignment quality, section loss, and manual refinement can still influence the final biological interpretation [Box 3].