Unravelling tumour spatiotemporal heterogeneity using spatial multimodal data

1.2.1 Spatial multi-omics technologies

Spatial omics technologies allow the simultaneous measurement of diverse molecular features – such as the genome,^{15, 16} epigenome,^17-26 transcriptome,^27-52 proteome,^53-73 and metabolome^{74, 75} while maintaining their spatial information^{76, 77} (Figure 1 and Table S1). These technologies have significantly advanced our ability to explore molecular, cellular, and structural patterns in both healthy and diseased states. Highlighted by Nature in 2022⁷⁸ as a key technology, spatial multi-omics has developed from previous spatial mono-omics methods.^{70, 79-85} Innovations like MISAR-seq enable combined chromatin accessibility and transcriptome analysis,⁸⁵ while SPOTS supports simultaneous proteomics and transcriptomics profiling.⁸³

Spatially resolved transcriptomics (SRT) data has emerged as a popular method for analysing disease progression, particularly within tumours,^{86, 87} as it allows for quantifying gene expression while maintaining spatial information within tissues. SRT methods are divided into imaging-based and sequencing-based approaches. In imaging-based methods, single-molecule fluorescent in situ hybridisation (smFISH)⁸⁸ quantifies multiple mRNA transcripts at subcellular resolution, with subsequent methods like seqFISH,⁸⁹ seqFISH+,³⁰ and MERFISH⁹⁰ multiplexed capabilities. Recently, nanoString commercialised the CosMx™ SMI platform to provide spatial multi-omics with FF and FFPE tissue samples at cellular resolution, quantifying up to 6000 RNAs and 64 proteins.⁹¹ Sequencing-based techniques, such as ST,⁹² MASC-seq,⁹³ Slide-seq,⁹⁴ Slide-seqV2,⁹⁵ and HDST,³⁵ measure the expression across spatial spots. 10x Genomics's Visium platform⁹⁶ has improved its resolution, reducing spot diameter from 100 to 55 µm. Furthermore, the recently released Visium HD enables resolutions of 2, 8, and 16 µm, while the Xenium platform offers true single-cell subcellular resolution. BGI's Stereo-seq achieves $$500\ {\mathrm{nm}}$$ resolution over larger tissue areas.⁹⁷ The sequencing-based technologies yield multimodal data, including gene expression, spatial location, and histology, the integration of which helps to reveal complex tissue architecture.⁵

1.2.2 Computational strategies

Integrating diverse spatial multi-slice multi-omics data facilitates characterising dynamic behaviours of different types of molecules, intracellular molecular networks, and intercellular regulation in the spatiotemporal progression of tumours (Figure 2a). However, spatial omics technologies often face resolution and capture efficiency challenges. Enhancing sensitivity and specificity by integrating public single-cell omics data or known interactions is thus crucial in bioinformatics. In this section, we will discuss key integration challenges, strategies for effective integration, current limitations and future directions.

The tools are divided into three categories based on how reference cells/spots are selected (Figure 2b and c and Table S2):⁹⁸ specifically, (i) for identical omics types across slices (horizontal integration), shared features across these slices serve as reference points; (ii) for different types of omics data from the same tissue slice (vertical integration), such as when both gene expression and protein data are collected using DBiT-seq technology,⁷⁹ the individual cells serve as the reference; and (iii) when different types of omics data are obtained from different tissue slices (diagonal integration), no common reference exists because the data types do not share similar features. For instance, gene expression looks at how genes are activated, while genetic data measures mutation throughout the genome. This difference in data types presents an initial challenge for combining these multi-omics data. In the following sections, we will explain the methods and analysis strategies for overcoming these practical challenges when integrating spatial multi-omics data.

In the analysis of omics data from multiple slices, computational methods are typically divided into two categories: those that deal with slices from the same tissue and those from different tissues. Integrating multi-slices omics data presents several challenges: (i) variations in tissue composition that affect cell densities, structures, and the surrounding microenvironment; (ii) physical shifts or distortions that make it difficult to align slices correctly; (iii) batch effects due to differences in how the samples were prepared, which can mask true biological signals; (iv) inconsistent markers leading to information gaps; (v) differences in resolution and detecting methods; and (vi) the risk of amplifying poor-quality data or noise. When slices come from different tissues, an additional layer of biological variation must be considered. Addressing these challenges requires advanced alignment algorithms, batch correction techniques, and noise reduction methods to ensure accurate integration and interpretation. Here, we illustrate strategies for addressing these challenges, using the integration of multi-slice SRT data as a representative example. Specifically,

(1) For multiple sections from the same tissue, PASTE⁹⁹ applies the optimal transport (OT) method to map and analyse neighbouring slices. Building on this, STitch3D¹⁰⁰ and GraphST¹⁰¹ use PASTE (or iterative closet point algorithm) to create a unified graph with 3D spatial coordinates and then apply a graph model to learn embeddings for spatial clustering. However, linear alignment in PASTE and its derivatives has limitations in detecting distortions in complex structures within slices caused by diseases with high variability. Moreover, SPACEL¹⁰² leverages graph models to predict cell type proportions, identify spatial domains, and reconstruct 3D tissue structure.

(2) For multiple slices across tissues: SEDR¹⁰³ combines gene expression and spatial coordinates using an autoencoder and graph model for spatial clustering. PRECAST¹⁰⁴ performs dimension reduction and spatial clustering via projection-based alignment, and its latest version, FAST,¹⁰⁵ is designed for large-scale data across slices. STAligner¹⁰⁶ uses a graph model and spot triplets to identify shared and conditional clusters. SLAT¹⁰⁷ employs graph and adversarial learning algorithms to map slices across technologies/omics. SPIRAL¹⁰⁸ integrates graph and OT methods to remove batch effects, predict unseen samples, and align coordinates. STELLAR¹⁰⁹ uses graph geometric learning via cell representations to transfer annotations from one slice to another across regions, tissues, and donors. However, these methods do not fully leverage the intricate inter-spot relations within and across slices, limiting their ability to capture partial relations in heterogeneous slices.

While most methods focus on integrating omics data with spatial location, they often overlook the complementary insights from modalities like histological images and annotations. Effectively leveraging this additional information – while addressing challenges related to scale, diversity, multimodality, and high dimensionality (where each sample contains a large number of features) – remains a complex task. Another important challenge we would like to highlight is how to utilise the large populations of cells/spots in tissue slices and phenotypes in tissue slices to obtain phenotypically relevant biological findings. Recently, CytoCommunity,¹¹⁰ DeepSP,¹¹¹ and scPROTEIN¹¹² have been introduced to integrate spatial proteomics data, offering new approaches for handling spatial information across multimodalities. For other omics data – such as chromatin openness and metabolism^{113, 114} across multi-slices – these methods provide useful frameworks for inspiration.

In integrative analysis of spatial multi-omics data from the same slices, most methods address the challenges such as (i) differences in scale and measurement units across omics types; (ii) sparsity, low resolution, and noise in multi-omics data; and (iii) high dimensionality, which requires substantial computational power. These methods typically begin by mapping data into a shared or coordinated feature space to minimise differences across omics. While single-cell methods^115-123 like scGPT, GLUE, totalVI, MultiVI, and Seurat (see previous review for details¹²⁴) can be adapted for spatial data integration, they may not fully capture the spatial context information, which is important for elucidating tissue composition. Combining spatial coordinates with multi-omics data is a new research area, and few methods have been proposed so far. For instance, Cellcharter¹²⁵ and SLAT¹⁰⁷ use preprocessing tools like scVI¹²⁶ and GLUE,¹¹⁵ and then leverage graph models to learn cell/spot representations. MaxFuse¹²⁷ smooths input data using graphs and iteratively maps different omics after co-embedding. SpatialGlue¹²⁸ combines spatial information with feature graphs using graph and attention methods, while moscot¹²⁹ models cell mapping across time and space as an OT problem. PRESENT¹³⁰ uses contrastive learning to capture cross-modal representations in multi-omics data. Moreover, for slices from different tissues, additional challenges arise due to biological heterogeneity and variations in resolutions that capture cellular features at different scales, complicating accurate alignment for integrative analysis.

Beyond the methods described for identifying spatial domains or cellular niches from batch-corrected features through the integration of multi-slice multi-omics data, future research should focus on exploring potential conversion relationships between cellular niches and the intracellular and intercellular molecular interactions or regulations in driving disease progression. This can be achieved by integrating spatial locations with multi-omics data,¹³¹ histological images, and annotations, as well as public single-cell omics data and known molecular interactions.¹³² To improve the biological explanation and interpretation of spatial multi-omics data, computational solutions can be developed in several key directions: (i) Fine-tuning pre-trained models derived from large-scale single-cell reference data to adapt to spatial omics data. This approach can enhance sparse or low-resolution spatial data, enabling more precise cell type annotation and deeper function insights; (ii) leveraging known gene-gene interactions or public ATAC-seq profiles¹³³ to construct associations across different molecular data types, facilitating the creation of cross-modal relations in the context of disease proregression and mitigating the impact of low-quality of omics data; and (iii) inferring cause relationships between regulator elements and target genes involved in disease progression, aiding in the identification of potential targets for therapeutic intervention. Such comprehensive integration would provide deeper insights into cellular interactions and the spatiotemporal evolution of disease, effectively addressing the inherent complexity and multidimensionality nature of spatial omics data.

1.3.1 Spatial clustering

In the analysis of SRT data to identify spatial domains within tissues, various methodologies have emerged as indispensable tools. Key statistical methods in the field, such as BayesSpace,¹³⁴ Giotto,¹³⁵ and DR-SC,¹³⁶ leverage probabilistic models to identify spatial components by combining gene expression and spatial coordinates. In contrast, deep learning models like SpaGCN,¹³⁷ STAGATE,¹³⁸ stMVC,⁵ and stKeep¹³⁹ apply graph-based approaches to reveal spatial patterns within multimodal data. A recent review provides detailed benchmark comparisons of these methods on simulated and real data.¹⁴⁰

Despite recent progress, several challenges remain in the field. With the increasing availability of single-cell and subcellular resolution data, addressing issues such as sparsity, high-dimensionality, scalability, and interpretability is crucial for further progress. Overcoming these challenges is essential to fully unlock the potential of SRT data and to gain a deeper understanding of spatial tissue organisation and cellular composition.

1.3.2 Identification of SVG

Detection of SVGs is essential for elucidating tissue biology, which involves identifying genes with differential expression across tissue or specific domains. Researchers have developed many computational approaches to address this problem, each with its strengths. For example, trendsceek¹⁴¹ uses a permutation process to estimate spatial dependencies, while SpatialDE¹⁴² applies Gaussian process regression to assess spatial variance. SPARK¹⁴³ and SPARK-X¹⁴⁴ detect spatial patterns through statistical methods, SOMDE utilises self-organising maps¹⁴⁵ and Hotspot¹⁴⁶ employs graph models. SpaGCN¹³⁷ tests the hypothesis for each gene based on identified domains, and STAMarker¹⁴⁷ uses saliency maps to highlight important features. BSP¹⁴⁸ and its enhanced version scBSP¹⁴⁹ use a big-small patch algorithm to detect SVGs at different scales, while PROST¹⁵⁰ introduces an indicator (PI) to evaluate spatial expression variation. A comparison of these methods is available in a recent review.¹⁵¹

Despite these advancements, challenges remain. Handling the high-dimensional and spatially structured nature of the data, while ensuring robustness and scalability across diverse conditions, is difficult. Additionally, dealing with noise, tissue heterogeneity, and complex interactions within the tissue microenvironment, along with ensuring scalability to large or 3D data, is vital for detecting SVGs. Addressing these challenges will be key to dissecting gene expression patterns and functions of different spatial components in disease progression.

1.3.3 Inference of cell–cell communications

Multicellular organism complexity arises from communication between various cell types. Computational inference of CCC from local spatial components is important for understanding cellular heterogeneity and functions,² and can be divided into two categories: (1) cell-population-level: Giotto¹³⁵ calculates interactions between neighbouring cell types based on ligand–receptor pairs. stLearn¹⁵² tests significant enrichment of ligand–receptor pairs in neighbouring cells using co-expression analysis. CellPhoneDB v3¹⁵³ examines interactions from the same local domains. SVCA¹⁵⁴ and MISTy¹⁵⁵ leverage statistical and machine-learning models to infer spatially dependent cellular gene-gene interactions. deeplinc¹⁵⁶ constructs CCC maps from scratch based on ligand–receptor genes; and (2) single-cell-level: SpaTalk¹⁵⁷ constructs and quantifies the ligand–receptor-target signalling network between nearby cells through knowledge-based graph models. NCEM¹⁵⁸ calculates how the composition of the cellular environment affects gene expression using graph models. COMMOT¹⁵⁹ leverages collective OT to infer CCC by considering the competition between different ligands and receptors and their spatial arrangement. stKeep¹³⁹ adopts a heterogeneous graph (HG) to infer CCC, ensuring that learned CCC patterns are comparable across different cell states through contrastive learning. IGAN¹⁶⁰ infers gene programs influenced by CCC using spatial correlation. A comparative analyses of some of these methods are provided in the recent review.¹⁶¹

Despite these advances, current methods still face significant challenges. Many struggle to capture the dynamic and context-dependent nature of CCCs, especially in heterogeneous conditions and environments. Scalability and computational efficiency are also issues, partially when dealing with large-scale datasets and integrating multi-slice data. Advances in single-cell and imaging technologies will be crucial for providing detailed insights into CCCs at both the cellular and molecular levels, improving our understanding of how cells communicate. In the context of spot-level SRT data analysis (involving multiple cells), it is critical to dissect how various cells coordinately respond to dynamic conditions.

1.3.4 Prediction of gene-regulatory network

Cell identity is controlled by GRNs, and transcription factors (TFs) interact with enhancers and promoters to regulate gene expression. Inferring GRNs from omics data is key to identifying impaired gene functions and critical drivers of disease progression. Accurately inferring regulatory networks that characterise cell states while addressing challenges such as high dimensionality, sparsity, and high noise in omics data is very challenging, especially for the spot-level (multiple cells) data. As a result, there are currently few research methods available. For example, SCING¹⁶² utilises gradient boosting and mutual information methods to infer stable GRNs. CLARIFY¹⁶³ employs a graph model to construct cellular networks, supporting CCC inference and enhancing the accuracy of cell-specific GRNs. stKeep¹³⁹ leverages an attention-based multi-relation graph embedding method to aggregate information from cells and cell states while ensuring that co-related genes are co-embedded to learn gene embeddings. The embeddings can be used to identify cell-state-specific GRNs.

We want to highlight that it's important to be cautious when interpreting results from spot-level omics data. That is because the relations between two genes in these omics do not always indicate they are co-regulated or co-expressed within a cell. stKeep solves this by using known relations between genes from public databases to avoid false positives. However, some co-associated gene pairs within one cluster may still be missed. Moreover, it is important to infer the direction of gene regulation in a cell. Leveraging spatial multi-omics data, along with unpaired ATAC-seq or Chip-seq data, is essential for uncovering gene-gene relations and their directions.¹³¹

1.3.5 Pseudo-time-space analysis

Pseudo-time-space methods allow researchers to track cell state changes throughout tissue space and time, providing insights into homeostasis, repair, and responses to environmental signalling.¹⁶⁴ To explore temporal changes within SRT data, several methods have been developed. SpaceFlow¹⁶⁵ leverages graph model to learn cell/spot features through combining gene expression and spatial coordinates then calculates pseudo-spatiotemporal MAP (pSM) using a diffusion pseudo-time (DPT) algorithm.¹⁶⁶ stLearn,¹⁵² STAGATE, and stMVC construct a spatial PAGA graph using gene expression similarity between spatial domains, and infers the pseudo-temporal order among these domains using a minimum spanning tree algorithm. SIRV¹⁶⁷ estimates RNA velocities at single-cell resolution by incorporating spliced and un-spliced mRNA data from reference scRNA-seq into SRT. Paella¹⁶⁸ uses initial pseudo-temporal values and spatial coordinates to create a network that progressively identifies several sub-trajectories. STT¹⁶⁹ uses a dynamic model to describe multistability in space through mRNA splicing and SRT data. spaTrack¹⁷⁰ creates spatial pseudo-temporal sequences by addressing OT problems between two cell groups. Additionally, spatial RNA velocity provides a way to directly infer developmental trajectories by representing temporal changes in cells.¹⁷¹ Recently, CalicoST¹⁷² has enabled the simultaneous inference of allele-specific copy number aberrations while also reconstructing spatial tumour evolution from SRT data.

In the future, more efforts should focus on identifying key regulators and genes that drive the spatial and temporal transitions of 3D tissue space to better understand cell differentiation and disease progression. Moreover, enhanced methods to fill in missing temporal data will be important to study dynamic cellular behaviours and increase the data's usefulness.

1.3.6 Cell type deconvolution

High-throughput platforms such as Visium capture the full transcriptome but lack single-cell resolution. Moreover, tissue slice thickness can also cause overlapping RNA signals from multiple cells. Imaging-based SRT methods achieve subcellular resolution but are limited in gene numbers, restricting their broader use. Hence, accurate prediction of cell types in spot-level SRT data is crucial for identifying disease-associated cellular composition and structures.

Current integrative analysis of whole-transcriptome and scRNA-seq data falls into two categories: (1) deconvolution-based methods: CARD,¹⁷³ Cell2location,¹⁷⁴ RCTD,¹⁷⁵ and POLARIS,¹⁷⁶ which use statistical or probabilistic-based models to spatial map cell types. DestVI¹⁷⁷ utilises deep learning to capture gene expression differences among cells of identical type. GraphST¹⁰¹ and CellMirror¹⁷⁸ use contrastive learning to estimate cell type proportions. DSTG¹⁷⁹ and STdGCN¹⁸⁰ leverage graph models to predict cell type composition based on graphs created from both real and simulated SRT data, with the simulated data generated from the single-cell reference database. Redeconve¹⁸¹ estimates the single-cell composition of SRT spots through non-negative least regression method; (2) alignment-based methods: NovospaRc,¹⁸² Tangram,¹⁸³ Celltrek,¹⁸⁴ and CytoSPACE,¹⁸⁵ map single-cell locations to SRT data by analysing gene expression similarities; and (3) reference-free methods like Stdecon¹⁸⁶ and RETROFIT¹⁸⁷ handle challenges between scRNA-seq and SRT data, including batch-effects, uneven of cell type coverage, and variations in gene expression.

Most current methods provide the proportion of different cell types in each spot, but do not provide the precise localisation of each cell type within the spot, indicating a need for computationally improved resolution in the future. Integrating histological images as complementary information could help enhance this resolution. In addition, when using deconvolution results for further CCC inference, caution is needed, as many different cell types may express the same ligand.