A unified spatial transcriptome profiling of ten mouse organs

Abstract

Spatial transcriptomics has enabled numerous deep learning models in this area, and training them requires large amounts of high-quality data, especially expression matrices paired with histological images. Here, we present a unified spatial transcriptomic dataset generated using the Stereo-seq platform, covering 10 mouse organs—including brain, kidney, lung, thymus, large intestine, skin, spleen, ovary, testis, and uterus—encompassing 23 tissue sections generated from 21 chips, each with matched ssDNA or H&E staining images. The dataset comprises single-cell-resolution (cell-bin) or square bin-50 (25 µm × 25 µm) expression matrices for each sample, accompanied by corresponding cell type annotations. Annotation robustness was further supported by concordance across different sections of the same tissue and corroboration with canonical marker gene expression patterns. Finally, we compared the characteristics of the cell-bin and bin-50 expression matrices and demonstrated the advantages of cell-bin resolution for cell annotation. This dataset provides a standardized resource for spatial transcriptomics method development, benchmarking, and multimodal analysis.

Background & Summary

Spatial transcriptomics (ST) has rapidly emerged as a transformative technology, offering critical insights into tissue context and function by combining spatial location information with transcriptome-wide gene expression^1,2. Among various ST platforms, Stereo-seq stands out due to its subcellular resolution (500 nm) and exceptionally broad transcriptome coverage, as well as its capacity to provide corresponding stained histological images that enable cell segmentation and in-depth study of tissue architecture^3,4.

Parallel to this, the surge in spatial transcriptomics (ST) technologies has spurred the development of diverse deep learning models and algorithms, particularly large-scale architectures, tailored to decipher these complex multimodal data. They opened new avenues for tasks such as automated cell annotation^5–7, the decoding of complex cellular interactions^8–10, and joint analysis of histological images and gene expression matrices^11–13. Despite these methodological advances, the performance of such models heavily depends on large-scale, high-quality training data¹⁴. However, the relatively high experimental cost of ST has limited the accumulation of large-scale public datasets—particularly those with unified and standardized profiles^15,16. Substantial variability across experimental platforms, laboratory protocols, and data processing strategies has led to pronounced technical heterogeneity, reduced data consistency, and limited cross-dataset comparability. These challenges collectively hinder the development of robust and generalizable deep learning models, highlighting the urgent need for large-scale, unified ST datasets encompassing diverse tissue types^17,18.

In this study, we present a dataset comprising 10 mouse organs—including brain, kidney, lung, thymus, among others—profiled across 23 sections and 21 chips using Stereo-seq technology. Among these, two ovary chips each contained two tissue sections. Each chip is accompanied by corresponding stained histological images (ssDNA or H&E). We systematically assessed image quality to determine whether individual cells could be reliably segmented. For samples with high-quality images enabling accurate cell segmentation, we provide single-cell resolution expression matrices, called “cell-bin”; for samples in which reliable image-based cell segmentation was not feasible, we generated square bins by aggregating spots within a 50 × 50 pixel square region into a single spatial unit (hereafter referred to as “bin-50”). Using three independent single-cell reference databases, we performed cell type annotation on both the cell-bin and bin-50 matrices. The annotation results demonstrated consistent spatial distribution within the same tissue and section, as well as concordance with marker gene expression patterns, collectively validating the reliability of the dataset. This dataset provides a valuable resource for deep learning models in spatial transcriptomics. Furthermore, we compared annotation outcomes derived from bin-50 data with those from single-cell resolution data, highlighting the advantages of cell-level expression matrices for refined cell type identification.

Methods

Tissue collection for Stereo-seq experiment

Mouse tissue samples were collected from 13 6-week-old C57BL/6 J mice provided by the Guangzhou Institutes of Biomedicine and Health, Chinese Academy of Sciences, through collaboration with the BGI laboratory. Mice were maintained under standard laboratory conditions, with food and water provided ad libitum before tissue collection. All experimental protocols for generating the dataset adhered to ethical regulations regarding animal research and were approved by the Institutional Review Board of BGI (Approval No. BGI-IRB A21001-T1). All methods were carried out in accordance with relevant guidelines and regulations, and this study is reported in accordance with the ARRIVE guidelines (https://arriveguidelines.org). Mice were anesthetized with isoflurane inhalation and euthanized by cervical dislocation by trained personnel, followed by rapid tissue collection. Fresh-frozen samples were prepared and processed according to the corresponding STOmics Stereo-seq Transcriptomics Set user manuals, including the STOmics Stereo-seq Transcriptomics Set User Manual for V1.2 samples (Manual Version C, August 2023; https://enfile.stomics.tech/STOmics%20Stereo-seq%20Transcriptomics%20Set%20User%20Manual_C%20August%202023.pdf.pdf) and the STOmics Stereo-seq Transcriptomics Set V1.3 for Chip-on-a-slide User Manual for V1.3 samples (STUM-TT001; https://enfile.stomics.tech/STUM-TT001%20Stereo-seq%20Transcriptomics%20Set%20V1.3%20for%20Chip-on-a-slide%20User%20Manual.pdf). Briefly, tissues were rapidly dissected, embedded in optimal cutting temperature (OCT) compound, and snap-frozen in liquid nitrogen. Cryosections of 10 μm thickness were prepared using a cryostat and mounted directly onto the surface of Stereo-seq chips, followed by tissue fixation with methanol.

Image acquisition and staining

During the imaging stage, an automated microscope equipped with a 10 × objective was used for image acquisition. For H&E-stained tissue sections, the epi-brightfield (color camera) mode was selected, whereas for ssDNA-stained sections, the epi-fluorescence mode was used. For skin slice 1, fluorescence images were acquired using a STOmics Microscope Go Optical system (Scanner Version 1.2.2, Ximea MC124 camera). For the remaining samples, images were acquired using a Motic PA53FS6 microscope (PA53 Scanner 1.0.0.14, PA53FS6 SCAN 5LITE RGB module). Images were saved in TIFF format and subjected to image quality control using the imageQC module in StereoMap. Because the dataset was assembled from samples generated during the development of Stereo-seq technology and related data resource construction, H&E and ssDNA staining were not selected according to predefined fixed criteria for each organ or slice. Instead, the staining modality was mainly determined by data availability and the overall quality of the imaging and spatial transcriptomic data. All images were captured following the imaging requirements specified in the user manual (https://enfile.stomics.tech/STUM-PE002-Stereo-seq-Imaging-Requirements-and-Guidelines.pdf).

Stereo-seq library preparation and sequencing

Tissue sections placed on the chip were permeabilized with 0.1% pepsin (Sigma, P7000) in 0.01 M HCl at 37 °C for 10 min. Released mRNA molecules were captured by spatially barcoded probes on the chip surface and reverse transcribed in situ to generate cDNA. The amplified cDNA was then used for DNA nanoball (DNB) generation and library construction. Libraries were sequenced together with coordinate identity (CID) information on the MGI DNBSEQ-Tx platform using paired-end reads. Read 1 contained the coordinate identity (CID) and unique molecular identifier (UMI) sequences, with CID occupying bases 1–25 and UMI occupying bases 26–35, whereas Read 2 consisted of the captured cDNA sequence.

Both V1.2 and V1.3 workflows were included because the samples were generated over a relatively long period during different stages of Stereo-seq technology development and related data resource construction. For samples processed using Stereo-seq V1.2, tissue removal was performed following reverse transcription prior to cDNA release. The released cDNA was subsequently purified and subjected to PCR amplification and purification. Detailed procedures are provided in the STOmics Stereo-seq Transcriptomics Set User Manual (Manual Version C, August 2023). For samples processed using Stereo-seq V1.3, the downstream workflow was optimized to simplify the experimental procedure and shorten the overall processing time. The separate tissue removal and standalone cDNA purification steps were omitted. Instead, cDNA release and denaturation were carried out using KOH and Elute Additive, followed by neutralization with Neutralization Solution. The resulting cDNA was then directly subjected to partitioned PCR amplification and purification. Detailed procedures are provided in the STOmics Stereo-seq Transcriptomics Set V1.3 for Chip-on-a-slide User Manual (STUM-TT001).

Stereo-seq raw data processing

Stereo-seq V1.2 and V1.3 raw FASTQ data were processed using the SAW (v8.1.0) pipeline¹⁹. Read 1 contains the coordinate identity (CID) and unique molecular identifier (UMI) sequences. CID sequences were aligned to the Stereo-seq capture chip coordinates with up to one mismatch allowed. The corresponding Read 2 sequences were mapped using STAR to the SAW-provided mouse reference genome. The reference genome was specified as GRCm38.p6, with genome-build accession NCBI:GCA_000001635.8. Only alignments with MAPQ (mapping quality) = 255 were retained. UMIs sharing the same CID and gene were collapsed into a single count, allowing one mismatch to account for PCR or sequencing errors. Gene expression counts were then aggregated into a spatial profile matrix based on spatial coordinates.

Cell-bin and bin-50 expression matrices generation

Spatial expression matrices were output as “GEF” files by SAW. For tissue staining images, image processing was performed using CellBin package²⁰. CellBin was first used to align the tissue staining image with the gene expression matrix based on chip track lines, followed by identification of the tissue-covered region and removal of background regions. For tissue sections with image quality sufficient for reliable cell segmentation, CellBin was further used to perform cell segmentation, and the resulting CellBin outputs were subjected to manual inspection and correction. For samples in which reliable image-based cell segmentation was not feasible, a square bin strategy was adopted to generate a bin-50 expression matrix. The bin-50 resolution was selected because it is a commonly used spatial aggregation scale in Stereo-seq data analysis^3,21. Specifically, thymus slice 1, thymus slice 2, thymus slice 3, and spleen slice 2 contained local regions with highly dense cellular structures, where the boundaries between adjacent cells were unclear and individual cell contours and cell numbers could not be reliably distinguished by visual inspection. For spleen slice 1, in addition to local regions with highly dense cellular structures and indistinct cell boundaries, overexposure was observed at the tissue edge, further reducing the reliability of cell segmentation. Therefore, these five samples were considered unsuitable for reliable image-based cell segmentation and were processed using the square bin strategy to generate bin-50 expression matrices.

Downstream processing and cell annotation

The cell-bin and bin-50 “GEF” file were transformed into Anndata format by Stereopy^22,23. Before cell type annotation, basic quality filtering was performed using Scanpy to remove cells/bins and genes with no or nearly no detected expression. Specifically, only cells or bins with total counts ≥ 1 and at least one detected gene were retained, and only genes with total counts ≥ 1 and detected in at least one cell/bin were retained. The retained cell-bin and bin-50 matrices were then used for cell type annotation. Cell type annotation was performed using the cell2location implementation provided in the STOmics SDAS repository (https://github.com/STOmics/SDAS). Reference single-cell transcriptomic datasets were curated from multiple sources according to tissue type, including the Mouse Cell Atlas²⁴ (MCA 3.0; adult mouse), the Tabula Muris Senis²⁵ (three-month-old mouse), and the original cell2location publication²⁶. Cell type annotations were first assessed by comparing cell type distributions across different tissue sections of the same organ to ensure consistency. Resulting annotations were further validated by examining the spatial distribution of canonical marker genes to confirm biological accuracy and spatial coherence.

Data Records

The raw data have been deposited into the Spatial Transcript Omics DataBase (STOmics DB, https://db.cngb.org/stomics/project) under accession ID STT0000184²⁷. The dataset includes raw tissue images in.tif format and matrix registered images, raw gene expression matrices in “GEF” format, tissue restricted gene expression matrices in.tissue.gef format, as well as expression matrices at bin-50 or cell-bin resolution, together with the corresponding annotation files. The dataset has also been submitted to the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) under accession number GSE333479²⁸. In addition, the raw tissue images in.tif format and the raw gene expression matrices in “GEF” format have been deposited in Zenodo (https://zenodo.org/) under https://doi.org/10.5281/zenodo.18463527²⁹.

Technical Validation

The spatial transcriptomic dataset presented in this study was generated from freshly dissected mouse organs collected from 13 mice, including brain, large intestine, lung, thymus, kidney, spleen, testis, skin, ovary, and uterus—comprising a total of 23 tissue sections from 21 chips (Fig. 1a,b). Among these, two ovary chips each contained two tissue sections. Variations in chip size, kit type, and staining method were present across samples (Fig. 1c,d,e, and Table 1). Corresponding ssDNA and H&E-stained images were acquired for each chip. For samples exhibiting clear, single-cell-resolvable morphology, cell-bin expression matrices were generated; for samples in which reliable image-based cell segmentation was not feasible, bin-50 expression matrices were generated (Table 1).

Fig. 1.

Generation and overview of spatial transcriptomes across 10 mouse organs and 23 tissue sections. (a) Workflow of Stereo-seq procedure and generation of cell-bin and bin-50 expression matrices. (b) Staining types and section numbers for the corresponding tissues. (c) Proportion of different chip sizes used in the dataset. (d) Sequencing kit types used in the dataset. (e) Overview of staining types.

Table 1.

Overview of 10 mouse organ dataset.

Tissue	Staining type	Slice	Chip size	Kit type	Matrix type for annotation	Mouse ID
brain	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_1
		slice2	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_1
		slice3	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_1
kidney	H&E	slice1	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_2
		slice2	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_3
lung	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_4
thymus	H&E	slice1	1*1	Stereo-seq Transcriptomics V1.2 Kit	bin-50	mouse_5
		slice2	1*1	Stereo-seq Transcriptomics V1.2 Kit	bin-50	mouse_5
		slice3	1*1	Stereo-seq Transcriptomics V1.2 Kit	bin-50	mouse_5
large intestine	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_6
		slice2	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_6
skin	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_7
	H&E	slice2	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_8
spleen	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	bin-50	mouse_9
	H&E	slice2	0.5*0.5	Stereo-seq Transcriptomics V1.3 Kit	bin-50	mouse_10
ovary	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_11
		slice2	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_11
testis	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_12
		slice2	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_12
uterus	ssDNA	slice1	1*1	Stereo-seq Transcriptomics V1.3 Kit	cell-bin	mouse_13
		slice2	1*1	Stereo-seq Transcriptomics V1.2 Kit	cell-bin	mouse_13

The table summarizes tissue type, staining type, chip size, kit type, matrix type used for cell type annotation, and the corresponding mouse ID for each chip.

The high quality of the resulting expression matrices was validated by their robust transcriptional profiles. At the bin-50 resolution, individual spots across all tissue types consistently contained more than 700 genes and exhibited an average UMI count greater than 1000. Furthermore, at the bin-200 (100 µm × 100 µm) resolution, each spot encompassed over 4500 distinct genes with an average UMI count exceeding 16,000 (Table 2). On the other hand, over 76% of the tissue sections exhibited sufficient image clarity for accurate cell segmentation, enabling the generation of cell-bin matrices (Table 1). For these cell-bin expression matrices, the resulting matrices demonstrated well-defined transcriptional profiles, with an average of over 120 genes detected per single cell and an average UMI count exceeding 170 per cell (Table 3).

Table 2.

Statistics of genes and UMIs captured at different squarebin sizes.

Slice	bin-20				bin-50				bin-200
	Gene Types		UMI Counts		Gene Types		UMI Counts		Gene Types		UMI Counts
	Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median
brain-1	137	132	203	183	716	734	1,264	1,240	5,070	5,537	19,599	20,619
brain-2	138	131	222	198	715	726	1,383	1,353	5,085	5,543	21,451	22,922
brain-3	158	155	244	229	806	826	1,520	1,509	5,397	5,815	23,595	24,365
kidney-1	154	147	219	205	760	751	1,360	1,282	4,997	5,257	20,626	18,828
kidney-2	145	158	197	212	720	823	1,197	1,339	4,393	5,557	16,461	19,868
lung-1	498	476	992	922	2,258	2,273	6,103	5,852	9,921	10,991	90,498	94,652
thymus-1	152	156	192	197	775	829	1,187	1,233	4,878	5,778	17,745	19,101
thymus-2	237	245	315	327	1,161	1,258	1,944	2,029	6,183	7,357	29,041	31,030
thymus-3	198	209	248	258	986	1,082	1,523	1,634	5,609	6,490	22,493	25,688
large intestine-1	749	791	1,513	1,519	2,812	3,122	9,227	9,203	8,983	10,690	130,338	132,681
large intestine-2	299	318	560	591	1,291	1,431	3,369	3,700	6,445	7,428	46,991	53,000
skin-1	145	122	216	166	703	621	1,315	1035	4,840	5,122	18,588	16,268
skin-2	145	103	237	153	721	564	1,445	962	4,688	4,910	20,528	14,806
spleen-1	402	327	1,007	778	1,766	1,541	6,209	4,828	8,297	8,665	92,982	77,435
spleen-2	233	218	311	291	1,109	1,082	1,906	1,808	6,013	6,771	27,632	28,174
ovary-1	668	598	1,232	997	2,685	2,697	7,505	6,162	9,336	11,038	106,801	83,634
ovary-2	214	215	342	328	974	1,032	2,040	1,999	4,915	5,721	26,044	22,322
testis-1	540	575	740	794	2,284	2,475	4,577	5,003	9,488	10,445	69,357	77,974
testis-2	310	311	437	403	1,427	1,494	2,706	2,509	7,379	8,049	41,134	41,106
uterus-1	248	126	442	219	1,103	624	2,655	1,310	5,610	4,729	35,443	17,155
uterus-2	174	127	435	309	790	600	2224	1867	4645	4337	34,701	24,823

Squarebin Size indicates the number of DNBs combined into a single analysis unit. Mean Gene Types and Median Gene Types represent the average and median number of gene types captured per bin, respectively. Mean UMI Counts and Median UMI Counts denote the average and median unique molecular identifier (UMI) counts per bin, reflecting the levels of gene expression.

Table 3.

Statistics of cell bin properties across mouse organs.

Tissue	Slice	Cell Count	Mean Cell Area	Median Cell Area	Mean Gene Types	Median Gene Types	Mean UMI Counts	Median UMI Counts
brain	slice1	113,170	185.25	171.25	290	248	476	385
	slice2	106,238	177.5	163.5	281	236	504	403
	slice3	108,434	169.5	154.25	295	248	505	403
kidney	slice1	156,908	76.5	66.25	125	95	178	125
	slice2	66,975	76.25	68.25	129	115	175	151
lung	slice1	202,860	84.75	75.25	497	446	972	826
large intestine	slice1	28,685	92.5	80.5	860	815	1786	1571
	slice2	29,310	86.25	75.5	283	261	529	466
skin	slice1	24,885	152.25	134	249	179	406	258
	slice2	23,805	119.5	100.5	246	174	424	263
ovary	slice1	51,698	76.75	65.25	694	610	1277	307
	slice2	46,599	76.75	65.25	209	186	331	307
testis	slice1	148,070	74.5	66.5	450	421	605	298
	slice2	140,228	83.25	72	273	243	382	333
uterus	slice1	12,941	102.5	85.75	647	584	1208	1039
	slice2	13,325	109.5	87.75	329	281	855	659

Cell Count refers to the total number of cells captured in the cell-bin expression matrices. Mean Cell Area and Median Cell Area represent the average and median cell areas, respectively, measured in square micrometers (µm²). Mean Gene Types and Median Gene Types indicate the average and median number of unique genes detected per cell. Mean UMI Counts and Median UMI Counts denote the average and median Unique Molecular Identifier (UMI) counts per cell, reflecting the overall transcript abundance.

After basic quality filtering, we performed cell annotation on those tissue sections. The resulting cellular annotations accurately delineated the histological architecture, with well-demarcated structural layers consistent with the organ’s anatomy (Fig. 2). For example, in the mouse brain sections, our Cell2location-based annotation clearly resolved distinct neuroanatomical structures, such as the anterior commissure, olfactory limb (aco) and the genu of corpus callosum (ccg) (Fig. 2b). These spatial patterns are highly concordant with the Allen Mouse Brain Atlas and align with the refined regional annotations reported by Yao et al.³⁰. Specifically, our results successfully captured the high-density enrichment of mature oligodendrocytes within these white matter fiber tracts, consistent with the high-resolution MERFISH data from the aforementioned study. Furthermore, in the mouse testis sections, our annotation successfully reconstructed the centripetal distribution of germ cells: spermatogonia were precisely localized at the basement membrane, while spermatids were significantly enriched in the center of the seminiferous tubules (Fig. 2f). Additionally, the mapping results showed that Leydig cells distributed as clusters precisely within the interstitial spaces. These spatial features are highly consistent with the testicular spatial atlas reported by Chen et al.³¹.

Fig. 2.

Spatial cell type annotations across 10 mouse organs. Panels containing multiple slices are ordered from left to right as listed. (a–h) Spatial distribution of cell type annotations based on cell-bin expression matrices: (a) large intestine slice 1 and slice 2; (b) brain slice 1, slice 2, and slice 3. Astro, astrocyte; Endo, endothelial cell; Ext, excitatory neuron; Inh, inhibitory neuron; Micro, microglia; Nb, neuroblast; OPC, oligodendrocyte precursor cell; Oligo, oligodendrocyte; AMY, amygdala; CTX, cortex; HPC, hippocampus; HYPO, hypothalamus; STR, striatum; THAL, thalamus; WM, white matter; ClauPyr, claustrum; Pir, piriform area; hab, lat and med denote habenular, lateral and medial, respectively; CA1, CA2 and CA3 denote cornu ammonis areas 1–3, and DG denotes dentate gyrus. (c) uterus slice 1 and slice 2; (d) lung slice 1; (e) kidney slice 2 and slice 1; (f) testis slice 2 and slice 1; (g) ovary annotation results, where the first two panels correspond to slice 1 and the last two panels correspond to slice 2; (h) skin slice 2 and slice 1. (i,j) Spatial distribution of cell type annotations based on bin-50 expression matrices: (i) thymus slice 1, slice 2, and slice 3; (j) spleen slice 2 and slice 1.

As a validation of data robustness, we first assessed the concordance of cell annotations between different slices of the same tissue, and the high cross-slice consistency confirmed the reliability of our cellular annotations as well as the accuracy of both the cell-bin and bin-50 expression matrices across 10 mouse organs (Fig. 2). We then compared the proportional distribution of annotated cell counts and found that cell type compositions were highly consistent among biological replicates of the same organ (Fig. 3a,d). Notably, separate visualization of the most abundant cell type in each organ revealed a consistent spatial distribution of these cells across distinct slices from the same tissue (Fig. 3b,e). To further validate this observation, we examined the expression patterns of canonical markers for each dominant cell type and observed that the marker signals co-localized precisely with the annotated cell positions (Fig. 3c,f,i,l). Specifically, we confirmed the following marker–cell type pairs: Meis2 for Int_Meis2 cells in the mouse brain, Rag1 for immature T cells in the mouse thymus³², Spp2 for S1 proximal tubule cells in the mouse kidney³³, and Slc4a1 for erythroblast cells in the mouse spleen³⁴. Importantly, despite morphological variations among different sections of the same tissue type, the spatial distribution of the most abundant cell type and its marker genes remained consistent (Fig. 3g–l).

Fig. 3.

Validation of cell type annotation across representative organs. (a–c) Brain sections (ssDNA, cell-bin): (a) composition of the top 10 annotated cell types. Inh, inhibitory neuron; Ext, excitatory neuron; Oligo, oligodendrocyte; Endo, endothelial cell; Micro, microglia; ClauPyr, claustrum. (b) spatial distribution of Inh_Meis2, and (c) spatial expression of its marker Meis2. (d–f) Thymus sections (H&E, bin-50): (d) composition of annotated cell types, (e) spatial distribution of immature T cell, and (f) expression of its marker Rag1. (g–i) Kidney sections (H&E, cell-bin): (g) composition of the top 10 annotated cell types, (h) spatial distribution of S1 proximal tubule cell, and (i) expression of its marker Spp2. (j–l) Spleen sections (slice 1: ssDNA; slice 2: H&E; bin-50): (j) composition of the top 10 annotated cell types, (k) spatial distribution of erythroblast, and (l) expression of its marker Slc4a1.

We additionally compared the performance of the cell-bin and bin-50 matrices, motivated by their distinct characteristics: the cell-bin matrix provides superior single-cell UMI assignment accuracy, whereas the bin-50 matrix yields higher UMI counts per spot but may contain transcripts derived from multiple cells and potentially multiple cell types within a single bin (Table 2 and Table 3). The spatial distribution of cells, as visualized by histological staining, was consistent with that reflected in the cell-bin expression matrix (Fig. 4a,b,e,f, and Fig. 4i–j). Furthermore, we observed a high degree of concordance between the cell-bin and bin-50 annotation results at the global level (Fig. 4c,d, Fig. 4g,h, and Fig. 4k,l). Notably, at the local architectural level, the cell-bin matrix preserved intact hollow structures within the tissue and enabled the identification of additional cell populations, uncovering richer molecular characteristics (Fig. 4c,g,k). In the mouse testis, the cell-bin expression matrix successfully annotated macrophages, whereas the bin-50 strategy failed to identify this population within the same localized region (Fig. 4m). Similarly, in the mouse ovary, the cell-bin matrix enabled the precise annotation of endothelial cells and epithelial cells, while the bin-50 strategy failed to capture these signals (Fig. 4n). In localized areas of the mouse kidney, compared to the bin-50 matrix which only identified Cryab high epithelial cells and ureteric epithelium, the cell-bin matrix demonstrated finer resolving power, capturing a more diverse range of cell types within these regions (Fig. 4o).

Fig. 4.

Comparison of spatial expression representation strategies for cell type annotation. (a–l) Cell segmentation and annotation across diverse tissue slices, including (a–d) testis slice 1, (e–h) ovary slice 1, and (i–l) kidney slice 1. For each tissue, panels display the full tissue image with a magnified view of a selected region of interest (ROI) (a, e, i), the corresponding cell segmentation mask (b, f, j), and cell type annotations within the ROI generated using the CellBin (c, g, k) and bin-50 (d, h, l) strategies. (m–o) Quantitative comparison of annotation performance. Bubble plots summarize the cell counts for each cell type within the corresponding magnified ROIs for (m) testis, (n) ovary, and (o) kidney. Bubble size represents the total count of annotated units identified by each strategy. Light grey bubbles indicate cell types not captured under the bin-50–based annotation. Total count indicates the number of annotated units within the selected ROI and provides a descriptive summary of the local annotation results for each strategy.

Author contributions

X.R., Q.K., and D.W. designed the research. X.R., T.L., N.L., C.S., J.F., Q.K., and D.W. have obtained and analyzed the data. X.R., and T.L., drafted the manuscript. X.R., T.L., N.L., N.Z., N.Z., Q.K., and D.W. revised and edited the manuscript. X.R., Q.K., and D.W. supervised the study. All authors have made a substantial contribution to the manuscript.

Data availability

The datasets generated in this study are available in STOmics DB under accession number STT0000184 and can be accessed at: https://db.cngb.org/stomics/project/STT0000184. The datasets have also been submitted to the Gene Expression Omnibus (GEO) under accession number GSE333479 and can be accessed at: https://identifiers.org/geo:GSE333479 (2026). In addition, raw tissue images and raw gene expression matrices have been deposited in Zenodo and are available at: https://doi.org/10.5281/zenodo.18463527.

Code availability

All data preprocessing and analysis workflows were performed using CellBin v2 (https://github.com/STOmics/cellbin2), Scanpy (v1.9.8, https://github.com/scverse/scanpy), and SDAS (https://github.com/STOmics/SDAS).

Competing interests

The authors declare no competing interests.