CellContrast: Reconstructing spatial relationships in single-cell RNA sequencing data via deep contrastive learning

Summary

A vast amount of single-cell RNA sequencing (SC) data have been accumulated via various studies and consortiums, but the lack of spatial information limits its analysis of complex biological activities. To bridge this gap, we introduce CellContrast, a computational method for reconstructing spatial relationships among SC cells from spatial transcriptomics (ST) reference. By adopting a contrastive learning framework and training with ST data, CellContrast projects gene expressions into a hidden space where proximate cells share similar representation values. We performed extensive benchmarking on diverse platforms, including SeqFISH, Stereo-seq, 10X Visium, and MERSCOPE, on mouse embryo and human breast cells. The results reveal that CellContrast substantially outperforms other related methods, facilitating accurate spatial reconstruction of SC. We further demonstrate CellContrast’s utility by applying it to cell-type co-localization and cell-cell communication analysis with real-world SC samples, proving the recovered cell locations empower more discoveries and mitigate potential false positives.

Highlights

• •A new tool for reconstructing spatial relationships using gene expression of cells
• •Better performance in local spatial context reconstruction by contrastive learning
• •Robust performance across diverse tissues and platforms
• •Recovered cell spatial relationships enhance discoveries and mitigate false positives

The bigger picture

Spatial transcriptomics (ST) could revolutionize our understanding of biological tissues and processes. However, its usage could be associated with high costs and technical challenges. Meanwhile, single-cell RNA sequencing (SC) technologies have matured and there is currently a large amount of data available. These rich SC datasets hold great potential for biological discoveries, but they lack spatial relationship information between cells. CellContrast is a computational method that uses gene expression data and a model trained with ST data to reconstruct the spatial relationships between cells. CellContrast revealed that spatial information can still be transferred to the SC data even if the SC and ST information do not align perfectly. Mining of existing SC datasets coupled with limited ST information can maximize the value of these two resources and contribute to further understanding of cell systems.

Explore the unseen spatial relationships in single-cell sequencing data with CellContrast. This tool, trained on spatial transcriptomics data by contrastive learning, can project gene expressions into a latent space where spatially proximate cells share similar representations. CellContrast demonstrated a better performance in reconstructing local spatial context and maintained robustness when a certain mismatch exists across tissues and platforms. Utilizing the spatial relationships inferred by CellContrast can facilitate discoveries with single-cell sequencing data and mitigate potential false positives.

Introduction

Spatial-resolved transcriptomic profiling has emerged as a crucial tool for investigating both cellular gene expression and its corresponding positional context. This approach is vital for unraveling complex tissue organizations and intercellular communication that underlies cooperative biological functions. However, the current landscape of spatial transcriptomic (ST) technologies, spanning imaging- and sequencing-based methods,¹ faces several challenges that prevent comprehensive problem resolution. Imaging-based methods, such as SeqFISH+,² STARmap,³ and MERFISH,⁴ have achieved single-cell resolution; however, their gene-profiling capabilities are limited to a few hundred predefined genes. Conversely, sequencing-based methods such as 10X Visium, Slide-seq,⁵ and Stereo-Seq⁶ provide the complete transcriptome, albeit at spatial resolutions larger than a single cell. In recent years, single-cell RNA sequencing (SC) has become a widely used technology, and large data resources have been built based on it. Although SC data lack spatial information, they yield more comprehensive gene profiles than imaging-based ST methods while also affording higher resolutions than sequencing-based ST approaches. Therefore, researchers have started to explore combining ST and SC to overcome their individual limitations. Various methods have been developed to enhance ST data with SC as a reference, including spot deconvolution^7,8 and gene imputation.^9,10 While these methods address the limitations of ST to some extent, the potential of existing vast SC resources remains untapped.

An alternative paradigm has arisen from an inverse perspective, whose aim is to reconstruct the spatial locations of SC data using ST as a spatial reference. Several methods have been proposed, including NovoSpaRc,¹¹ which assigns SC cells to predefined locations based on the assumption that physically proximate cells share similar gene expression profiles. Its location assignment is performed based on optimal transport. Although NovoSpaRc does not necessarily require an ST reference, the locations and marker genes derived from ST can be utilized to improve its performance. Subsequently, Tangram¹² and CytoSpace¹³ were introduced to establish mapping relationships from SC to ST. Although both methods minimize the correlation-based cost function inferred from gene expressions, their assumptions and outcomes are different. Tangram considers cell density and gene expression, generating mapping probabilities from SC to ST. CytoSpace generates distinct SC-to-ST predictions based on inferred cell counts per ST spot and corresponding gene expression. However, Tangram can generate low probabilities for matched SC-ST pairs, and CytoSpace might filter out some SC cells to ensure optimal subsets of samples, potentially leading to incomplete reconstructions. Recently, supervised methods have emerged, such as SageNet¹⁴ and CeLEry.¹⁵ SageNet employs a graph neural network to model the gene interaction network from the ST data and produces an estimated cell-cell distance matrix for the query cells. It has demonstrated accurate and robust performance across distinct samples. However, the preprocessing step of making spatial partitions is dataset specific and intricate for non-experts. The memory consumption of SageNet also explodes when the number of genes increases. CeLEry represents another supervised deep neural network approach for recovering spatial information of SC data. By ingesting the gene expression inputs, CeLEry was trained by minimizing the loss between the predicted domain layer or spatial coordinates and the truth. It has shown reliable performance in global spatial reconstruction across various species and technologies. However, it does not fully consider the performance of local cell environment reconstruction, which is vital for many downstream tasks.

In this study, we introduced CellContrast, a deep-learning method that employs a contrastive learning framework for spatial relationship reconstruction of SC data. Our fundamental assumption is that gene expression profiles can be projected into a latent space, where physically proximate cells demonstrate higher similarities. To achieve this, we employed a contrastive framework of an encoder-projector.¹⁶ We built training batches containing positive and negative contrastive pairs based on their truth spatial relationship. We optimized the model using infoNCE loss,¹⁷ which pulls positive pairs closer and pushes negative pairs apart. During inference, we discarded the projector and used the output of the encoder for spatial relationship reconstruction, based on the principle that higher cosine similarity indicates shorter spatial distance. We benchmarked the performance in multiple settings, including matched samples, independent samples, and independent platforms between training and testing with mouse embryos, human breast, and mouse brain data. Our experiments found that CellContrast outperformed related methods by a large margin and that, specifically, local heterogeneity can be well captured. We also demonstrated the power of spatial reconstructed SC data in two downstream tasks: cell-type spatial co-localization and cell-cell communication (CCC). The results showed that, with the help of CellContrast, cell-type co-localization patterns were well detected and potential false-positive CCC signals were reduced.

Results

Overview of CellContrast

The aim of CellContrast is to train an encoder that extracts spatially preserved representations based on the gene expression profiles of ST datasets. The fundamental assumption is that gene expressions can be projected into a feature space where spatially proximate cells exhibit similarity. This trained encoder is then applied to SC samples, enabling estimation of distances between SC samples and between SC and ST samples. We adapted contrastive learning for this purpose (see section “experimental procedures”). As illustrated in Figure 1A, our method employs a deep neural network comprising two main components: an encoder for spatial semantic representation extraction and a projection head for loss calculation. Both components were implemented as multi-layer perceptron (MLP). Constructing the training batches containing contrastive pairs is crucial. Given batch size n, we first randomly selected n/2 cells. For each sample X, we selected one of its k nearest neighbors as the positive pair and the rest of n − 2 cells (the 2 being the X and its random selected positive pair) served as the negative pairs. The projection head generated vectors that were used for infoNCE loss calculation, which encouraged positive pair similarity and negative pair dissimilarity. Model optimization was achieved via stochastic gradient descent (SGD).

Figure 1.

Overview of our method

(A) The computational framework. Our method utilized a framework with two key components: an encoder and a projection head. Both components were implemented as MLP. During training, we harnessed spatial transcriptomics (ST) samples as the training dataset. For each sample X, we selected a positive pair from its nearest neighbors and chose n − 2 random samples as negative pairs. The model’s neural network takes log-normalized gene expressions as input and is optimized through the infoNCE loss, which minimizes the distance between positive pair projections and maximizes the separation of negative pairs.

(B) Spatial reconstruction workflow. The gene profiles of SC samples are processed by the trained encoder to generate representations. These representations enable the identification of cell neighbors and global cell-cell relationships by comparing pairwise SC-SC representation similarities. Moreover, the SC-ST representation similarities was included to locate the most matched ST sample’s coordinates, effectively predicting the SC sample’s spatial location based on the reference dimension.

During inference, the projection head was removed and the trained encoder received SC gene expressions and generated spatially related representations (Figure 1B). Depending on the downstream tasks, the representations can be used in two ways: (1) obtaining SC-SC similarities to identify cell neighbors and cell-cell distances of the query data, and (2) aligning SC samples to ST samples based on a weighted similarity metric and generating coordinates in the reference space. For each of the assignment between SC and ST, we also provide a mapping confidence score (see section “experimental procedures”).

Evaluation involved two steps. First, we focused on accurately capturing the local cell environment. This was crucial for downstream tasks, such as CCC and cell-type co-localization. We used two metrics: the average neighbor hit within k nearest neighbors, which measures the identification of true individual neighbors, and the Jessen-Shannon distance (JSD) to compare predicted and true neighbor cell-type distributions, considering downstream tasks often rely on detecting the interaction among cell-type groups. Second, we assessed global spatial relationships using the Spearman’s rank correlation coefficient, which compares predicted and true distances among all cells (see section “experimental procedures”).

Benchmarking spatial reconstruction of mouse gastrulation cells using single-cell ST reference

To evaluate the performance of the cell spatial reconstructions of CellContrast, we employed single-cell-resolution mouse gastrulation data generated from SeqFISH.¹⁸ The dataset contains expression profiles of 351 genes collected from three mouse embryos that are on embryonic day (E)8.5 (marked embryos1, 2, and 3). Each mouse embryo has two sections from distinct tissue layers $12\mu \mathrm{m}$ apart (marked L1 and 2). We trained the model using embryo 1 L1 with 10,150 cells, and assessed the performance in two scenarios: (1) testing different layers from the same mouse embryo (embryo 1 L2), and (2) testing independent samples (embryo 2 L2 and embryo 3 L2) (Figure 2A).

Figure 2.

Spatial reconstruction of mouse gastrulation cells using a single-cell-resolution ST reference

(A) Strategy for evaluating reconstruction performance. The training dataset consistently featured SeqFISH1 L1 data. We investigated two distinct scenarios: (1) evaluation using embryo 1 L2 data, sourced from the same embryo as the training data but originating from different samples; and (2) assessment using embryo 2 L2 and embryo 3 L2, derived from independent mouse embryos not utilized in the training phase.

(B) Evaluation of cell neighbor reconstruction. We quantified the performance of different methods by analyzing the average hit number within various k nearest neighbors across the three distinct mouse test datasets.

(C) Evaluation of neighborhood cell-type reconstruction. We computed the average Jessen-Shannon distance (JSD) for each cell type across the three mouse test datasets.

We conducted a comprehensive comparison of five related methods (Data S1, S2, and S3). Our evaluation encompassed the detection of true cell neighbors within the predicted k nearest neighbors surrounding individual cells (Figure 2B). Notably, our method outperformed all others, displaying the highest average hit numbers across all tested datasets. Our method and SageNet managed to identify twice the number of true neighbors compared with alternative approaches (for example, the median average hit ratios of 47.32% by ours, 36.22% by SageNet, and 14.27% by CeLEry). Furthermore, our method outperformed SageNet by 10 additional neighbors when $k\ge 120$ (about a 22% gain). Even at lower k values, such as 20, our method exhibited an improvement, identifying nearly 1.5 times the number of true neighbors compared to SageNet (4.73 vs. 2.69, 3.71 vs. 2.01, and 2.66 vs. 1.43 for L2 of embryo 1, 2, and 3, respectively). The distributions of the hit number within nearest 20 neighbors are shown in Figure S1A, indicating a consistent conclusion. It is worth noting that all methods yielded a higher average hit number in scenario 1 than in scenario 2. This observation underscores the importance of spatial consistency between training and testing cells, as a stronger alignment enhances the accuracy of predicted neighbors.

We further explored the extent of reconstruction of local cell-type heterogeneity by comparing cell types between the true and predicted nearest 20 neighbors and quantifying their divergence through JSD. A lower JSD value signifies a closer resemblance in cell-type distribution between predicted and true cell neighbors. The results, presented in Table 1, underscore the notable performance of our method, exhibiting the lowest average JSD of 0.237 for embryo 1 L2. SageNet was next, attaining a JSD of 0.289, while CytoSpace had a JSD of 0.324. In contrast, other methods displayed average JSDs large than 0.4. Of particular significance, our method maintained consistent JSD performance across independent testing datasets for embryo 2 and 3, illustrating its robustness against over-fitting to new contexts. When comparing the JSD of all cells with the null distribution—wherein locations for the testing sample were randomly shuffled—it was observed that our method shows a consistent lower JSD pattern than null distribution (Figure S1B). Delving deeper into the JSD averages for individual cell types, as depicted in Figure 2C, our method’s advantage is evident across almost all cell types: for example, gut tube in all three testing samples. The only exception was presomitic mesoderm cells in embryo 2 L2, partly due to the larger local cell-type heterogeneity of embryo 2 compared to the training embryo 1 L1. Another observation is that our method and SageNet exhibited better capability in identifying neighbors for rare cell types present in the training samples (e.g., allantois, erythroid, sclerotome in embryo 3 L2; Figure S2). This could potentially be originated from their approach of comparing each cell with all other cells in the SC dataset based on the learnt spatial patterns instead of assigning the cells to referenced locations. Worth noting is the discernible trend of increased JSD across all methods as the inherent truth local heterogeneity of cell types escalated.

Table 1.

Average JSD (k = 20) and Spearman’s correlation coefficient on SeqFISH1 L1 reference of mouse gastrulation

Method	Embryo 1 L2		Embryo 2 L2		Embryo 3 L2
	JSD	Spearman’s ρ	JSD	Spearman’s ρ	JSD	Spearman’s ρ
Ours	0.237∗	0.874∗	0.225∗	0.574	0.343∗	0.477
SageNet	0.289	0.543	0.273	0.514	0.391	0.391
CeLEry	0.435	0.769	0.44	0.59∗	0.562	0.485∗
CytoSpace	0.324	0.623	0.356	0.331	0.433	0.296
NovoSpaRc	0.417	0.47	0.432	0.155	0.522	0.197
Tangram	0.525	0.304	0.538	0.073	0.613	0.095

Spearman’s ρ denotes Spearman’s correlation coefficient.

JSD assesses the consistency of local cell-type proportions within k nearest neighbor. Spearman’s ρ evaluates the correlation of global spatial distances, especially including the non-neighbors as the majority.

Asterisks denote the best performed method for each metric.

The global spatial distributions were reconstructed based on SC-ST mapping (Figure S3). It was observed that blood progenitors and endothelium cells showed low confidence scores across all three samples, potentially caused by their diverse distribution (Figures S4 and S5). We further assessed the global relationships by Spearman’s rank correlation between true and predicted pairwise distances for each cell. Our method performs the best in scenario 1, while it achieved comparable performance with CeLEry in scenario 2 (Table 1). Our method, along with SageNet and CeLEry, maintained relatively stable performance across independent mouse embryos, while Tangram, NovoSpaRc, and CytoSpace exhibited significant declines. This divergence in performance may be attributed to the fact that those methods are not focused primarily on mapping SC samples to ST. Although they can perform this task, their main design focuses on ST spot deconvolution based on gene expression, lacking precise spatial arrangement.

Robustness of spatial location inference across platforms and samples

After validating the efficiency of spatial reconstruction with single-cell reference, we further explored CellContrast in more complicated settings. First, we benchmarked the reconstruction of the cell locations of SC samples with the assistance of a spot-based ST reference. In this experiment, the performance was evaluated in a setting where reference and query cells were (1) collected from different mouse embryos at slightly mismatched developmental stages, and (2) obtained by different techniques (with StereoSeq and SeqFISH, respectively). The ST reference was spot based. We took one sample of mouse organogenesis spatiotemporal transcriptomics atlas (MOSTA), which was collected from a mouse embryo on E9.5, obtained by Stereo-seq.⁶ The single-cell-resolution embryos1, 2, and 3 L2 were included for testing (Figure 3A; Data S4, S5, and S6). There were 5,266 spots with 348 overlapped genes for training the model. As shown in Figures 3B and S6, our method achieved the highest mean average hit number of the three test datasets. Specifically, when the nearest neighbor was k = 200, CellContrast detected 55.388 true cell neighbors. Second-place SageNet got an average hit number of 43.7 and detected 11 fewer than ours. The results suggest the good transferability of our method by capturing complex spatial-dependent features. The reconstructed local cell-type heterogeneity was assessed by JSD for the nearest 20 neighbors. As shown in Table 2, cell neighbors reconstructed by our method identified the most similar cell neighbors with cell types with the truth with the lowest mean average JSD (0.305 by ours vs. 0.359 by second-best SageNet). The JSDs for each cell type are shown in Figures 3C and S7. Our method and SageNet’s kept robust JSDs because they can estimate the query cell-cell distance directly. However, methods that can only project the cells into the spatial coordinates based on the ST reference decreased drastically. This decrease is observed even for cell types that exhibit low local heterogeneity with respect to their author-annotated cell-type labels, such as forebrain/midbrain/hindbrain, gut tube, and cardiomyocytes. Similar patterns were observed for global reconstruction with the Spearman’s rank correlation coefficient (Table 2; 2D distribution shown in Figure S8).

Figure 3.

Spatial location inference across platforms

(A) Evaluation of spatial reconstruction for the mouse gastrulation cells using array-based ST reference. The training dataset consistently utilized the Stereo-seq E2S1 sample. The benchmarking was against three independent single-cell-resolution ST datasets derived from E8.5 mouse embryos.

(B) Evaluation of cell neighbor identification. We evaluated the efficacy of different methods by quantifying the average hit number within varying k nearest neighbors across the three distinct mouse test datasets; the mean average hit numbers of three testing sets are presented.

(C) Evaluation of local neighborhood cell-type heterogeneity reconstruction. We computed the average JSD for each cell type across the three mouse test datasets and showed the mean values of the three testing sets.

(D) Evaluation of spatial reconstruction for 10X Visium human breast cells. The training utilized the normal sample S06, consisting of 3,037 spots, while the testing involved sample S08, comprising 2,801 spots.

(E) Assessment of cell neighbor identification. We quantified performance by analyzing the average hit number across varying nearest k neighbors of the test human breast cells.

(F) Evaluation of neighborhood cell-type reconstruction of human breast cells using JSD.

Table 2.

Average JSD (k = 20) and Spearman’s correlation coefficient of across multiple platform benchmarking experiments

Method	Mean of three mouse embryos		Human breast
	JSD	Spearman’s ρ	JSD	Spearman’s ρ
Ours	0.305∗	0.283∗	0.306∗	0.103
SageNet	0.359	0.198	0.369	0.172∗
CeLEry	0.577	0.156	0.465	0.077
CytoSpace	0.568	0.099	0.444	0.005
NovoSpaRc	0.579	0.066	0.393	0.044
Tangram	0.614	0.053	0.485	0.007

Spearman’s ρ denotes Spearman’s correlation coefficient.

Asterisks denote the best performed method for each metric.

Based on the successful application of our method in spatial reconstruction in mouse embryos with single-cell and spot-based ST reference, we extended our investigation to assess its performance within a recently released spatially resolved genomic atlas of the adult human breast, generated using 10X Visium.¹⁹ We employed two distinct normal samples—S06 (3,037 spots) and S08 (2,081 spots)—as a training and testing dataset, respectively (Figure 3D; Data S7). A total of 36,503 genes were used as input. As SageNet could not handle such a large pool of genes with 1T memory computing server, we used 1,024 highly variable genes as input for it. As a result, our method detected the most true cell neighbors, identifying 58.556 truth neighbors when k = 200, about 1.8 times that of the next-best method (Figure 3E). For reconstructing neighborhood cell-type heterogeneity, our method achieved the most similar neighborhood cell-type distribution of an average of JSD 0.306, followed by SageNet’s 0.369 and NovoSpac’s 0.393, and ours was the most robust for most of the cell types (Figures 3F and S9). When we limited the input to the 1,024 HVGs, our method experienced a slight decrease in performance. However, it maintained the same trend as other methods, indicating that comprehensive input genes are beneficial for mining spatially related features. Furthermore, in terms of capturing global spatial relationships among cells (2D distribution shown in Figure S10), the Spearman’s ρ among SageNet, ours, and CeLEry are comparable.

To further examine CellContrast’s transferability to different techniques, we incorporated the single-cell-resolution mouse brain dataset generated by MERSCOPE into our assessment.²⁰ Specifically, we utilized cells from the left hemisphere for training and those from the right hemisphere for testing. Similar to our observations with other datasets, our methods demonstrated the highest performance in both local and global spatial reconstruction (Figure S11; Table S1; Data S8).

In addition, we investigated scenarios where there exists a large divergence between donors' phenotypes for training and testing data. Specifically, we applied CellContrast on a dataset derived from human lung tissues obtained via 10X Visium.²¹ Notably, the training reference originated from a non-smoking male donor, whereas the testing data were sourced from a current female smoker with a 30-year smoking history. Our findings reveal that our method exhibited robust performance in local context reconstruction, surpassing other related methods (Figure S12; Table S2).

Unveiling cell-type co-localizations in dissociated mouse gastrulation cells with CellContrast

With the solid foundation laid out in the previous benchmarking sections, we showed the practical efficacy of our method for enhancing biological data analysis. We applied CellContrast to a real-world mouse atlas with embryos at the same developmental stage (E8.5).²² Leveraging the reconstructed spatial coordinates of cells, we showcased the precise detection of cell-type co-localization patterns.

Initially, we built a reference pattern for cell-type co-localizations based on the six SeqFISH datasets with truth cell spatial distributions. To identify cell-type spatial associations, we employed Moran’s R from spatialDM, adopting a significance threshold of p value 0.05. Figure 4A presents Moran’s R and p values for all cell-type co-localizations obtained from SeqFISH embryo 1 L1. Notably, dataset-specific co-localizations exhibited lower Moran’s R values. "Reference support" for a cell-type co-localization was defined as the number of SeqFISH datasets producing a significant signal for the co-localization pattern. A reference support of 6 denotes the utmost confidence in co-localization, whereas 0 implies the absence of evidence in SeqFISH datasets. Reference supports for each cell pair are shown in Figure 4B. Notably, 88 cell-type co-localizations were identified in the SeqFISH dataset, each with at least one reference support. Intriguingly, a spatial association was detected among blood progenitors, endothelium, and hematoendothelial progenitors, which were closely linked to blood generation.²³

Figure 4.

Cell-type co-localization analysis of spatially reconstructed SC data of the mouse embryo at E8.5

(A) Relationships between Moran’s R and the p values of all cell-type pairs derived from SeqFISH embryo 1 L1.

(B) Reference cell-type co-localization patterns. We incorporated six SeqFISH datasets; the value of reference support indicates the number of datasets demonstrating significant signals ($p\le 0.05$).

(C) Reconstructed spatial locations of the SC sample and its ST reference (SeqFISH embryo 1 L1).

(D) Detected p values of cell-type co-localizations based on a spatially reconstructed SC sample.

(E) Detected co-localization proportions in the SC sample vs. reference patterns.

We further applied our method to sample (S27) of the mouse SC atlas. In total, the spatial information of 6,460 cells was reconstructed. The model was trained on SeqFISH embryo 1 L1, featuring 350 overlapping genes. The spatial locations of SC cells were achieved by aligning them to the training ST samples. Figure 4C shows the reconstructed spatial locations of SC and reference ST. Employing Moran’s R, cell-type co-localizations were detected, and the corresponding p values are presented in Figure 4D. Importantly, the reference’s major co-localization patterns were captured by spatially reconstructed SC data. We extended our analysis by applying two related spatial reconstruction methods, SageNet and CeLEry, to the same SC sample, and subsequently conducted similar analyses (see section “experimental procedures”). We explored detected cell-type pairs with reference support exceeding one. Notably, both our method and CeLEry identified 42 cell-type pairs, while SageNet detected 38 pairs, with 27 co-localizations consistently detected across all three methods. We then employed Fisher’s exact tests for detected and reference co-localizations with support >1. The p values for our method, CeLEry, and SageNet were 2.793e−9, 1.121e−6, and 7.409e−6, respectively (Figure 4E; Tables S3, S4, and S5). We also performed the same analysis for an additional SC sample, which suggested the consistent conclusion (Figure S13). This indicates the effectiveness of spatial reconstruction methods in capturing cell-type co-localization patterns.

Improved CCC inference in mouse embryos through reconstructed spatial locations of SC data

CCC analysis benefits from spatial transcriptomic technologies, which provide spatial context and reduce false positives. As shown in Figure 5A, the traditional approach identifies ligand-receptor interactions (LRIs) based on gene expressions among different cell groups.²⁴ CellChat, a popular tool, uses statistical models to generate the LRIs and the interacting cell groups. With ST, gene expression and spatial locations enable more accurate CCC detection using cell-cell distance constraints. SpatialDM employs Moran’s R to identify co-expression of LRIs and detect interacting local spots or cells. In this study, we demonstrated that our spatial reconstruction method enhances CCC signal accuracy of SC data compared to gene expression alone.

Figure 5.

Analysis of CCC in spatially reconstructed SC data of the mouse embryo at E9.5

(A) Demonstration of the CCC analysis with SC and ST data. We employed CellChat to decipher CCC among different cell groups to compute the co-signals of ligand and receptors based on their gene expression. With the spatial distance constraint incorporated by ST data, SpatialDM detected the global LRIs and the local CCCs among individual cells.

(B) Reconstructed locations of E1. The spatially reconstructed locations of E1 were accomplished by leveraging the predicted cell-cell dissimilarities into a 2D space through the MDS algorithm.

(C) Comparison of global ligand-receptor pairs. We compared the identification of global ligand-receptor pairs using CellChat and spatialDM. Notably, CellChat exclusively employs E1’s gene expression data, while spatialDM integrates both gene expression and predicted spatial cell locations.

(D) Reference support for global ligand-receptor pairs.

(E) Reference support for local CCC.

(F) Spatial demonstration of the detected ligand-receptor interaction of LGF2-LGF2R.

We reconstructed the spatial locations of mouse embryo SC sample S1 (scRNA-S1), derived from GEO: GSE87038. The sample was collected on E9.5 with cells from forebrain, hindbrain, heart, liver, intestine, and somite tissues.²⁵ The model was trained with one Stereo-seq sample, using 444 genes that are highly variable and overlapped with the scRNA-S1. The pairwise cell-cell similarities were derived from the learnt representations of the cells and can be leveraged to 2D space using the multidimensional scaling (MDS) algorithm (Figure 5B). Notably, forebrain and hindbrain cells are clearly grouped into different clusters and near each other, suggesting cell-cell spatial relationships were well characterized. The median distances between cell-type pairs in the spatially reconstructed SC sample and the five ST references are shown in Figure S14. It demonstrated a consistent relative spatial relationship between ST and spatially reconstructed SC samples.

Subsequently, we conducted two distinct analyses for CCC detection. Initially, we employed CellChat to detect LRIs based solely on gene expression across various tissues. Then we employed spatialDM to identify LRIs and associated interacting cells by incorporating the reconstructed cell positions derived from the scRNA-S1 (see section “experimental procedures”). First, we examined global LRIs, by ignoring specific interacting tissues or cells. Both methods identified a total of 349 global ligand-receptor pairs, with CellChat detecting more unique LRIs (Figure 5C). To assess the reliability of these LRIs, we established reference CCCs through the application of spatialDM on five Stereo-Seq samples possessing ground-truth cell locations. These samples originated from two mouse embryos at developmental stage E9.5, aligning temporally with the scRNA-S1. The reference support for an LRI was defined as the number of Stereo-seq datasets that produced the same signals. A higher reference support indicated enhanced reliability in the LRI. As shown in Figure 5D, where spatialDM consistently identified a greater portion of LRIs with reference support >1 (for example, 0.28 VS 0.16 for reference support 5), accompanied by a reduction in LRIs with no support (0.12 vs. 0.30). In contrast, for a large proportion of CellChat-identified LRIs, no evidence has been found in the reference datasets. This result underscores the enhanced precision of LRI detection facilitated by the incorporation of reconstructed cell spatial information.

In addition to global LRIs, we further explored local CCCs, which are vital for understanding the role of the cells in the local context. We focused on tissues that overlapped the reference datasets—forebrain, hindbrain, heart, and liver—for validation. We evaluated local CCCs using "local spot evidence" to denote matched spots inferred from reference datasets, while "number of reference support" denotes the number of reference datasets that have at least one local spot evidence for the corresponding CCC (see section “experimental procedures”). We obtained similar conclusions to those of global LRIs (Figure 5E): the reconstructed cell locations notably reduced local unsupported CCCs, minimizing potential false positives.

As an example, we examined the established insulin-like growth factor 2 (IGF2) and its high-affinity receptor IGF2R, which is well known for trophoblast morphogenesis regulation.²⁶ Forty-five cells within the scRNA-S1, primarily from heart and liver tissues, exhibited significant IGF2-IGF2R signals using spatialDM. The normalized expressions and reconstructed spatial coordinates of these cells are shown in Figure 5F, effectively supporting CCC within a short range of reconstructed physical distances.

Discussion

In this study, we introduced a method, named CellContrast, designed to reconstruct spatial relationships among SC based on gene expression profiles, a task of paramount importance in deciphering complex biological activities. CellContrast leverages a deep contrastive learning framework with an encoder-projector structure. This design enables the projection of gene expressions into a latent space, where physically proximate cells exhibit similar representation values. This approach differs from that of NovoSpaRc, which assumes neighboring cells share similar gene expressions. With the utilization of learnt latent space, CellContrast allows capture neighbor-dependent gene expression patterns identified by a local CCC study.²⁷ Achieving this involves constructing training batches containing contrastive pairs, derived from ST data with truth locations and optimizing them using the infoNCE loss. The resulting cell-cell distances are estimated through cosine similarities of learnt representations and therefore identifying the potential cell neighbors. The spatially related representations of SC cells can be further mapped to location coordinates using ST reference or MDS algorithm. MDS only relies on SC-SC estimated distances, thus it is more robust when there is a large gap between SC and ST reference, but it loses the pre-knowledge of reference’s cell distribution. SC-ST mapping preserve the information but requires a high-quality and matched reference ST sample.

To evaluate CellContrast, we conducted extensive benchmarking experiments on diverse datasets, including mouse embryo, human breast, and mouse brain samples. These datasets included various platforms, ranging from spot-based (Stereo-Seq and 10X Visium) to single-cell-based resolution STs (SeqFISH, MERSCOPE). Our method’s performance and robustness were assessed in different scenarios: (1) training and testing on distinct layers of the same mouse embryo with single-cell resolution; (2) training and testing on different mouse embryos, both at single-cell resolution; (3) training on spot-based resolution data and testing on single-cell-resolution data from different mouse embryos; (4) both training and testing on spot-based resolution data with independent human breast samples; and (5) training and testing on single-cell-resolution MERSCOPE data with mouse brain cells. In all benchmarking scenarios, CellContrast consistently outperformed existing methods in local spatial reconstruction, such as SageNet,¹⁴ CeLEry,¹⁵ CytoSpace,¹³ NovoSpaRc,¹¹ and Tangram,¹² demonstrating its effectiveness and transferability in capturing local spatial relationships among neighboring cells. Our contrastive model primarily focuses on learning features related to neighboring cells, treating all distant cells equally. To address the challenge of estimating global spatial distance relationships, we project the SC samples onto reference ST locations based on the similarities of the learnt representations. This projection employs a weighted similarity metric based on both SC neighbors and ST samples. Notably, when evaluating global reconstruction performance using Spearman’s rank correlation coefficient, we observed improved results by incorporating more predicted nearest SC neighbors (as shown in Table S6). However, the approach of SC-ST mapping also introduces a potential "cell clumping" issue, where multiple neighbors are projected into the same reference location. For example, several spatially close SC cells may share the same ST sample with the highest weighted similarity score. Consequently, these cells are assigned to the same ST location, causing overlap and loss of predicted SC-SC distances. This issue was also discussed in CellTrek,²⁸ which addresses it by the point repulsion algorithm. The problem originated from the sparsity of the ST reference and the sample-wise greedy mapping approach. While this can be mitigated by combining local and global estimated distances into the assignment loss, such as the optimal transport method used by NovoSpaRc, it can be a double-edged sword. Particularly when there is a large gap between ST and SC, forcing the SC assignment to resemble the ST could introduce more bias. In our method, we can adjust the trade-off between accurately ranking global distances and addressing cell clumping can be adjusted by the number of SC neighbors used for SC-ST alignment. Moreover, in SC-ST mapping, global reconstruction performance heavily influenced by the spatial distribution similarity between ST and SC samples. We observed that optimal performance is achieved with matched pairs, but large gaps between them can lead to a performance drop in global reconstruction across all methods. Despite these challenges, our method maintains robust performance in identifying cell neighbors by directly estimating SC-SC distances using learned patterns. In such scenarios, we recommend using predicted neighbors based on SC-SC distances for downstream tasks. This may result in a loss of similarity to the reference spatial distributions, but it ensures more accurate cell neighbor identification. On the other hand, we believe that, in the task of spatial reconstruction, more emphasis should be placed on the local environment. This emphasis is rooted in the fact that neighboring cells are often actively engaged in mutual interactions, such as paracrine and juxtacrine signaling. Conversely, non-neighboring cells function within a more complex system, often requiring the involvement of mediators, such as blood plasma,²⁴ where the importance of spatial reconstruction diminishes. Additionally, contemporary downstream biological algorithms typically consider the interactions among neighboring cells in their calculations.

Reconstructing the spatial locations of SC data unlocked the potential for two crucial downstream analyses, exemplified with real-world SC samples in this study. The first is cell-type co-localizations analysis: using CellContrast, we reconstructed spatial locations for cells from a mouse embryo on E8.5 and detected cell-type co-localization patterns with Moran’s R. To validate our findings, we built reference patterns from six ST datasets with truth cell locations. Remarkably, the detected co-localizations aligned well with the reference patterns, exhibiting substantial agreement across the three spatial reconstruction methods (CellContrast, SageNet, and CeLEry). This analysis demonstrated that spatially reconstructed SC data can effectively capture cell-type co-localization patterns, thus enhancing biological insights. The second is CCC analysis, which is crucial for comprehending cell differentiation and tissue organization. While many methods infer CCCs from SC data, the lack of spatial constraints might introduce false positives. We employed CellContrast for CCC analysis on E9.5 mouse embryo SC data and compared CCCs inferred solely from gene expression (with CellChat) to those incorporating spatial information (with SpatialDM). Validated against five ST datasets, our results exhibited reduced false positives in CCCs when spatial information was considered. We also noted that, because the reference CCCs were identified through the same approach with the spatial reconstructed SC data, the results could be advanced to it to some extent. However, we consider the overall effect of spatial reconstruction to remain the same.

In contrast to ongoing efforts to enhance ST data resolution and impute missing genes through SC reference, our study adopts an inverse approach. Leveraging the wealth of SC data available, which covers a broader range of samples and species than available ST data, we present a solution for spatial relationship inference. This approach allows us to capitalize on the spatial patterns embedded in ST data to reconstruct the spatial relationships of SC samples when they were collected from the same region of the same species. A certain level of mismatch may exist between the query SC data and the ST reference due to limited ST resources (e.g., the developmental stages of embryos, less captured cells/spots in ST data, different cell-type distribution), but our method shows robustness, thanks to its primary aim of learning from a spatially preserved representation instead of explicit translation of spatial location. In addition to a single ST reference, we explored performance using multiple ST samples as training data. While the performance of matched samples experienced a slight decrease compared to a single ST reference, others demonstrated improved reconstruction performance (Figure S15; Table S7). It was concluded that optimized performance is achieved using a single ST sample closely aligned with the SC sample. When a matched sample is unavailable, multiple references are beneficial for performance optimization. An additional observation is that tissues with a high level of repeating elements could pose a challenge for the spatial reconstruction of SC cells. One potential solution could be preprocessing the data to decrease the presence of repeating elements.

In addition, we investigated the way of mining the most relevant genes for spatial reconstruction through the analysis of gene contribution to the latent representation (see section “experimental procedures”). Notably, we successfully detected spatial gene patterns, such as En1 and Hoxb1 (Figure S16). These findings indicated the latent representations of CellContrast as a potential resource for studying spatial biology. Looking ahead, CellContrast holds significant promise for comprehensive utilization of SC data in diverse contexts. Future directions include refining the method’s performance with improved ST data and exploring new applications of spatial reconstruction in the realm of single-cell analysis.

In conclusion, our proposed method, CellContrast, offers a powerful solution for accurate spatial relationship reconstruction in SC data. By harnessing the intricate spatial patterns inherent in ST data, CellContrast provides a tool for enhancing various downstream analyses and facilitating critical insights into cellular behavior. It provides the way for deeper exploration of spatial dynamics in single-cell research and offers opportunities for more discoveries reliant on spatial information.

Experimental procedures

Resource availability

Lead contact

For further details, questions, or requests, direct correspondence to Yuanhua Huang at yuanhua@hku.hk.

Materials availability

No physical materials were utilized in the conduct of this research.

Data and code availability

• •This study used publicly available data and the original data can be obtained at the following links: (1) seqFISH data on mouse organogenesis (https://crukci.shinyapps.io/SpatialMouseAtlas/); (2) Stereo-Seq data on mouse embryo (https://db.cngb.org/stomics/mosta/download/, E9.5_E2S1); (3) 10X Visium human breast data (https://cellxgene.cziscience.com/collections/4195ab4c-20bd-4cd3-8b3d-65601277e731, S06, S08); (4) MERSCOPE mouse brain data (https://info.vizgen.com/mouse-brain-data, s1r1); (5) SC data of mouse embryo on E9.5 obtained from the GEO database (GEO: GSE87038); and (6) SC data of mouse embryo on E8.5 (https://content.cruk.cam.ac.uk/jmlab/atlas_data.tar.gz, s29).
• •The source code is accessible at https://github.com/HKU-BAL/CellContrast and has been archived at Zenodo.²⁹ The trained models were accessible at http://www.bio8.cs.hku.hk/CellContrast/ and have been archived at Zenodo.³⁰

Implementation of CellContrast

Training batch construction

Constructing positive and negative cell pairs within training batches is a crucial aspect of contrastive learning. The primary goal of the CellContrast is to project cells into a latent space where spatially proximate cell pairs are similar. This requires careful formation of training batches, where neighboring cells are considered as positive pairs and non-neighboring cells act as negative pairs.

Given a batch comprising n cells, initially, n/2 cells are randomly chosen from the training dataset. For each cell X from the n/2 pool, one of the k nearest spatial located neighbors to X is considered to be a positive pair, which in total is also n/2 samples. Therefore, for each of cells in the batch, it has one positive pair and n − 2 negative pairs. The neighboring size k is determined based on the specific dataset: for the SeqFISH, MERSCOPE, Stereo-seq, and Visium 10X training data, values of 80, 80, 20, and 20, respectively, were used. This adjustment takes into account the decreased spatial resolution associated with the above four Spatial ST technologies. Besides the positive cell pairs in the training batch, the remaining cell pairs are treated as negative pairs. The batch size n was set as 64 in this study.

Model structure and training

CellContrast is composed of an encoder and a projection head, both implemented as MLP. In this study, there were two layers for both the encoder and the projection head. The node numbers were (1,024, 512) for the encoder and (256, 128) for the projection head. For the encoder, BatchNorm was added for each layer and leaky rectified linear unit (ReLU) was used as the activation function. For the projection head, ReLU was added to the first layer of MLP.

In the training phase, CellContrast accepted log-normalized gene expressions, which were calculated by scran.³¹ The projections were employed to calculate the infoNCE loss (Equation 1).

$$L_{\mathrm{q}}=-\log \frac{\exp \left(\frac{q·k_{+}}{\tau }\right)}{\sum _{i=1}^{N-2}\exp \left(\frac{q·k_i}{\tau }\right)}$$ (Equation 1)

Here, L_q denotes the loss of sample q. $\tau$ means the temperature parameter, which was set as 0.05 in this study. k₊ denotes the positive pair of sample q, and k_i denotes the i^th the negative sample. N denotes the batch size, and N − 2 is the number of negative samples.

The loss of each training sample was obtained by the above, and the batch mean loss value was used for subsequent optimization. The optimizer of SGD (momentum = 0.9, weight_decay = 5e−4) was adopted. The learning rate of the initial value 0.1 was scheduled with cosine decay. The model was trained 3,000 epochs. The runtime and memory consumption of fitting the model are available on (Table S8). Additionally, we investigated the impact of training epochs on spatial reconstruction performance (Figure S17) and found that sufficient training epochs (>1,000) are beneficial for contrastive learning.

Spatial location inference of SC data

During the inference phase, the projection head was discarded because the encoder is better on learning positive pairs’ spatial invariance.^16,32 and log-normalized gene expressions were fed into the previously trained encoder to generate representations. The cosine similarities among representations were utilized to predict spatial distances. Two different approaches were employed. First, to identify the neighbors among the query SC samples, SC-SC similarities were calculated. The top k samples with the highest cosine similarities were regarded as the nearest neighbors. Additionally, the coordinates of the SCs could be projected onto a pseudo space using the MDS algorithm based on the pairwise SC-SC similarities. The second approach is aligning SC to ST samples. This alignment is achieved through a weighted similarity metric that takes both SC-SC and SC-ST similarities into account. For a given query SC sample, we aim to estimate its spatial similarity to all available ST samples. This spatial similarity is represented as S:

$$\mathbf{S}=\left(s_{i,0},s_{i,1},\dots ,s_{i,n}\right)$$ (Equation 2)

where $s_{i,j}$ denotes the spatial similarity between the i^th SC sample and j^th ST sample, with N being the total number of reference ST samples, and $s_{i,j}$ is calculated based on the following:

$$s_{i,j}=\cos \_sim\left({\mathbf{R}}_i,{\mathbf{R}}_j\right)+\frac{1}{K}\left(\sum _{k=1}^K{e}^{-{\left(1-\cos \_\text{sim}\left({\mathbf{R}}_i,{\mathbf{R}}_k\right)\right)}^2}·\cos \_\text{sim}\left({\mathbf{R}}_k,{\mathbf{R}}_j\right)\right)$$ (Equation 3)

where cos_sim represents the cosine similarity, ${\mathbf{R}}_i$ is the learnt representations of i^th SC sample, and ${\mathbf{R}}_k$ is the representation of predicted k^th nearest SC neighbor of i^th SC cell/spot. ${\mathbf{R}}_j$ denotes the learnt representation of j^th ST cell/spot. K denotes the nearest m SC neighbors inferred from SC-SC similarities. This procedure is equivalent to smoothing the original SC-ST distance by the inferred SC-SC neighborhood graph. By default, K is set as 1 (no SC neighbor was included) in the benchmarking experiments.

Subsequently, we align each SC sample with the ST sample that demonstrates the maximum weighted similarity. The coordinates of this matched ST sample are considered as the predicted location for the SC sample.

We have adapted a confidence score, derived from Shannon entropy, to represent the confidence level of assigning SC data to ST locations. The rationale behind this is that, if there are multiple distant locations with high and similar estimated SC-ST spatial similarities, the confidence level should be low, as it indicates that the SC data are challenging to map to a unique predefined location. To achieve this, we first identified the top n ST locations that meet the following criteria: (1) they are not neighboring each other by m neighbors. (2) They have the highest SC-ST similarities that satisfy condition (1). N is 20 and m is 80 as the default.

For the $j^{th}$ SC cell, we first convert the spatial similarities into probabilities using the softmax function:

$$p_{i,j}=\frac{e^{s_{i,j}}}{\sum _0^n{e}^{s_{k,j}}}$$ (Equation 4)

where $p_{i,j}$ denotes the probability of assign $j^{th}$ cell to the $i^{th}$ ST location. $S_{i,j}$ indicates the spatial similarity between the $j^{th}$ cell and $i^{th}$ ST sample (Equation 3).

We then calculate the Shannon entropy:

$$H_i=-\sum _k{p}_{i,k}\log \left(p_{i,k}\right)$$ (Equation 5)

Where $H_i$ is the Shannon entropy for location assignment of $j^{th}$ SC cell.

Finally, the confidence score is obtained by:

$$C_i=1-\frac{H_i}{\mathit{max}\left(\mathbf{H}\right)}$$ (Equation 6)

H denotes all of the entropies of the SC cells.

For downstream tasks that require cell coordinates as input, such as cell-type co-localization and CCC, we utilized the first approach for the reference of spot-level resolution and the second approach for the reference of single-cell resolution.

Identification of spatially related genes

To investigate which genes are most relevant to the spatial reconstruction, we defined the gene contribution by the rate of representation change as follows:

$$C_{i,j}=\frac{{\parallel R_i-R_i^j\parallel }_2}{{\parallel X_i-X_i^j\parallel }_2}$$ (Equation 7)

$$X_i^j=\{\begin{array}{l}{x}_k,k\ne j\\ x_k+\epsilon ,k=j\end{array}$$ (Equation 8)

$C_{i,j}$ is the ratio of representation change for the $j^{\text{th}}$ gene to the $i^{\text{th}}$ cell. A larger $C_{i,j}$ denotes a larger contribution of $j^{\text{th}}$ gene to the $i^{\text{th}}$ cell’s learnt representation. $X_i$ is the original gene expression of the $i^{\text{th}}$ cell, and $X_i^j$ is the gene expression of $i^{\text{th}}$ cell by adding a small constant $\epsilon$ to the $j^{\text{th}}$ gene’s expression (Equation 8). In this study, $\epsilon$ was set as 0.01. $R_i$ and $R_i^j$ are the representation of $X_i$ and $X_i^j$, respectively, which are the outputs of the encoder.

Benchmarking analysis with real-world datasets

Related methods

To fully assess CellContrast, we performed benchmarking experiments with several related methods, including sageNet,¹⁴ CytoSpace,¹³ Tangram,¹² NovoSpaRc,¹¹ and CeLEry.¹⁵ The running and evaluation scripts of the benchmarking analysis can be found at: https://github.com/HKU-BAL/CellContrast and in Zenodo.²⁹

SageNet

SageNet employed a graph neural network to reconstruct the latent spatial locations by modeling the interactions between cells and genes. The model generated predicted pairwise cell-cell distance for the query samples. For the seqFISH training data, we partitioned the data and set the regularization parameters of the graphical lasso (GLASSO) to be the same as those reported by the authors for the same dataset. In the case of training with the Stereo-Seq and 10X Visium samples, which are both array-based ST technologies, we created three distinct partitions by assigning the spots into 10 × 10, 15 × 15, and 20 × 20 square grids and set the regularization parameters to 0.5 and 0.75, respectively, in line with the authors' approach for spot-based spatial datasets. All models were trained for 20 epochs using the default settings. For the benchmarking experiment of the human breast, 1,024 highly variable genes were used for training, as our 1TB memory computing server was not capable of handling all available genes in the dataset. All genes were used in other experiments.

CytoSpace

CytoSpace assigns SC samples to spatial locations by minimizing a correlation-based cost function for gene expressions between query and reference data. Thus, it returns the most closely matched reference spatial coordinates for the query SC samples. In our benchmarking experiments, the neighbors and cell-cell distances of each query sample were determined using the Euclidean distance of the predicted coordinates. It should be noted that CytoSpace may not be capable of allocating positions to all samples, and the assigned positions could not be unique, so the evaluation of CytoSpace was limited to successfully assigned samples, and we adopted the first appearance of predictions. In benchmarking experiments that use single-cell-resolution seqFISH samples as the spatial reference, we configured CytoSpace with the option "-single-cell" and supplied the author-annotated cell types for both the reference and query datasets. The parameter "-noss" was set at 10,000. For experiments that use spot-based Stereo-Seq and 10X Visium samples as the spatial reference, we ran CytoSpace with its default parameters.

Tangram

Tangram is similar to CytoSpace in that it seeks to learn mapping functions from SC cells to ST samples through deep learning. At the cell level, Tangram generates the probability of SC samples being present in all available reference ST positions. In our benchmarking experiments, we selected the reference position with the maximum probability as the predicted spatial location. We ran Tangram with all genes and default parameters for all benchmarking settings.

NovoSpaRc

NovoSpaRc is a computational method that assigns cells to tissue locations by assuming that cells with similar gene expressions are closer in spatial locations. Like Tangram, NovoSpaRc produces mapping probabilities for SC samples to locations of the reference. Therefore, we selected the positions with the maximum probability as the prediction in our benchmarking experiments. The predefined locations were constructed based on the reference ST dataset, which were as indicated in Figures 2A, 3A, and 3D. Following the NovoSpaRc analysis pipeline, we first identified marker genes from the reference dataset and then performed the mapping process based on these markers. As per the authors' reported practice, we set num_neighbors_s and num_neighbors_t as 5, alpha_linear as 0.8, and epsilon as 5e−3.

CeLEry

CeLEry is a supervised deep-learning model designed to recover the spatial domain and locations of SC datasets. It achieves this by reconstructing the individual cell’s spatial location and producing spatial coordinates. We trained CeLEry with the default parameters for all benchmarking experiments and inferred the predicted cell neighbors and cell-cell distances by calculating the Euclidean distance of the predicted coordinates.

Cell-type co-localization analysis

We employed SpatialDM’s Moran’s R³³ to investigate the cell-type co-localization patterns of ST data and spatially reconstructed SC data. To utilize it, we encoded a binary variable (0 or 1) for each location to signify whether or not it is occupied by a specific cell type. The significant spatial associations between different cell types were determined based on $p-\text{value}\le 0.05$.

To construct reference patterns, we analyzed six SeqFISH datasets from three mouse embryos. We set n_neighbors to 20 and the radial basis kernel (l) to 0.1 to identify the reference patterns. For the spatial reconstructed SC sample of CellContrast, we set the l as 0.1. And for CeLEry’s, we set the l as 0.1 and n as 20. While SageNet does not directly output the coordinates of SC samples, we used MDS to project its produced cell-cell distance as pseudo space and set the l as 1e−4, and n_neighbors as 20, because the scale of predicted distance is very small. The parameters of spatialDM were tuned based on the visualization of the weight_matrix according to spatialDM guidance (Figure S18).

CCC analysis

In the CCC analysis, we employed two methods. We used CellChat to detect LRI among different tissues of the SC mouse embryo data using gene expression only.³⁴ We used the recommended analysis pipeline and set the parameter "type" as "truncatedMean.” To detect CCCs in the spatially reconstructed SC data and reference ST data, we used spatialDM. We set the parameter n_neighbors as 20, and the l as 0.1 for all. We set the other parameters to default. We kept only the results with p value ≤0.05.

To validate the detected CCCs, we compared the detected global LRI and local CCCs with the six reference ST datasets. For global LRIs, we defined reference support as the number of datasets that detected at least one same LRI, regardless of where it was detected. We defined local CCCs as a combination of an LRI and where it was detected. The local spot evidence was defined as follows: for identified CCCs of CellChat, which are a combination of a specific LRI (X) and the involved tissue groups (A and B), local spot evidence refers to at least one spot with significant X in A and B detected in the individual reference dataset. For identified CCCs of spatialDM, which were a combination of LRI X and the specific cell C, the local spot evidence was defined if the same X was detected in the spots in the tissues where C was located in the individual reference ST dataset. The reference support denotes the number of datasets that have local spot evidence. Only results found in tissues that overlapped with the reference were included in the analysis of local CCCs.

Evaluation metrics

We concluded that the performance of spatial reconstruction of SC samples could be evaluated by the following criteria: (1) how many truth neighbors can be found; (2) how well the local cell-type heterogeneity can be preserved, which is fundamental for understanding many biological processes; and (3) whether global cell-cell relationships can be captured. To achieve these goals, we used three evaluation metrics: nearest neighbor hit, JSD, and the Spearman’s rank correlation.

The nearest neighbor hit refers to the number of successfully detected true neighbors within the k nearest neighbors. This is the strictest metric that requires individual-level spatial reconstruction correctness. For each test sample, we can obtain its corresponding nearest neighbor hit, and the average nearest neighbor hit denotes the mean values for all testing samples. In this study, we tested the performance by varying k values of 20, 40, …, 200.

To evaluate local cell-type heterogeneity, we adopted the JSD to compare the cell-type distributions of predicted k nearest neighbor cells and the true k nearest neighbors. K was set as 20. The ratios for each cell type of the predicted neighbors and truth neighbors were obtained and then JSD was calculated using the jensenshannon function in the Scipy package. Further, the average JSD was calculated for all test samples as well as by cell-type groups.

To evaluate global spatial reconstruction performance, we employed the Spearman’s rank correlation coefficient. Given a test sample, we obtained its coordinates by SC-ST mapping approach and the pairwise Euclidean distances between it and all other test samples. Then, the truth distances can be obtained through true locations. We incorporated the Spearman’s rank correlation to assess the monotonic relationship between the predicted and true distances. We calculated the average Spearman’s rank correlation coefficient across all query cells as a measure of the global spatial reconstruction performance.

Published: July 9, 2024