The diversification of methods for studying cell–cell interactions and communication

Abstract

No cell lives in a vacuum, and the molecular interactions between cells define most phenotypes. Transcriptomics data provide rich information to infer cell–cell interactions and communication, thus accelerating the discovery of fundamental roles of cells within their communities. Such research relies heavily on algorithms to infer which cells are interacting and the ligands and receptors involved. Specific pressures on different research niches are driving the evolution of next-generation computational tools, enabling new conceptual opportunities and technological advances. More sophisticated algorithms now account for the heterogeneity and spatial organisation of cells, multiple ligand types and intracellular signalling events, and enable the use of larger and more complex datasets, including single-cell and spatial transcriptomics. Similarly, new high-throughput experimental methods are increasing the number and resolution of interactions that can be analysed simultaneously. Here, we explore recent progress in cell–cell interaction research and highlight the diversification of the next generation of tools, which have yielded a rich ecosystem of tools for different applications and result in invaluable discoveries.

Table of Contents (ToC) blurb:

In this Review, the authors summarise recent progress in cell-cell interaction (CCI) research. They describe recent evolution in computational tools that underpin CCI studies, discuss higher throughput led by improvements in experimental methods to analyse CCIs, and highlight future directions for the field.

Introduction

Cell–cell interactions (CCIs) are a cornerstone of multicellular life, allowing cells to live in communities and perform collective functions¹. Cells interact by producing diverse molecules and membrane structures that activate signalling pathways in other cells¹, coordinate their gene expression², and drive cellular functions³. These interaction mechanisms include structural and functional proteins, small compounds, extracellular matrix, membrane projections, and extracellular vesicles. However, ligands that bind to cognate receptors on other cells, or ligand–receptor interactions (LRIs), are a dominant mechanism used for CCIs. The cellular interactions mediated by LRIs, used specifically to relay signals, are also known as cell–cell communication (CCC), but for simplicity we refer to this specific subtype simply as CCIs.

Over the last decade, there has been an increasing interest in studying CCIs to understand molecular mechanisms that govern tissue physiology⁴, disease⁵ and development⁶. Transcriptomics, in particular, has been the basis of a plethora of computational methods that capture distinct aspects of CCIs, such as the participating cell types and the molecular mechanisms involved^7–15. To support these CCI analyses, it has been crucial to build high-confidence LRI databases⁷, including for protein subunits, activators, inhibitors, and/or competitors^16–18, or even downstream target genes to capture intracellular gene regulation^18–20. Moreover, cutting-edge experimental methods are enabling high-throughput analysis of CCIs^21–27, leading to new biological discoveries and helping to refine and validate computational tools.

A variety of computational tools have been developed^7–15, each of which infer CCIs within a sample using known LRIs. The core methods use the expression levels of ligands and receptors to elucidate how cells communicate, generating new biological hypotheses⁷. Briefly, as a generic example, CCIs can be predicted as follows: First, a gene expression matrix is filtered to include only ligands and receptors. Second, the expression level of each gene is aggregated across all single cells of a specific cell type (such as by averaging expression across single cells). Third, for each pair of cell types in a sample, each candidate LRI is evaluated by considering the ligand expression level in the sender cell type and the receptor expression level in the receiver cell type. Finally, a communication score [G] is computed for each LRI in each cell-type pair separately (for example, this could be the product, mean, or geometric mean between the ligand and receptor expression levels in a sender-receiver pair of cell types). To further identify hypothesis-driven CCIs (such as identifying cell-type specific LRIs), statistical analyses including permutation, parametric, and non-parametric tests can be further performed to identify significant interactions. As a result, computational tools enable the identification of important CCIs and the generation of biological hypotheses that can be experimentally evaluated.

Several core tools have helped unravel many important CCIs, and new tools are emerging to address yet more complex nuances of intercellular interactions. For example, signalling events vary from cell to cell²⁸, even within the same cell type. In addition, proper interactions highly depend on cell proximity^29–31 and on multiple types of molecules (such as proteins, metabolites, etc.). Broadly speaking, CCIs are shaped by the biological context or condition wherein they occur^32–34. Thus, next-generation tools are now modelling CCIs better by addressing some of these aspects (Fig. 1a). Indeed, these CCI tools have become: finer, by considering full single-cell resolution and heterogeneity of CCIs; more localised, spatially contextualising cells; deeper, by expanding the ligand types and evaluating intracellular events of CCIs; and/or broader, by scaling CCI analyses up to multiple biological conditions (such as patients, treatments, life stage, and genotypic background).

Figure 1. Methodological advancement of cell-cell interaction research.

(a) A wide range of computational tools have been developed to infer cell-cell interactions from gene expression. Recent innovations have expanded the capabilities of such analyses to account for finer resolution of single cell interactions, spatial information, larger datasets, and deeper information. (b) Experimental methods have expanded the power of conventional methods to increase the throughput for tracking cell-cell interactions.

Conventional experiments employed to validate CCI predictions often focus on co-localization of the mediating molecules, including techniques such as fluorescence in situ hybridization (FISH), fluorescence resonance energy transfer (FRET), and immunostaining^7,9. Although these methods allow hypothesis-driven study of specific interacting partners, cutting-edge high-throughput techniques can simultaneously track large numbers of interactions, enabling a hypothesis-free approach. These elucidate physical CCIs at resolutions spanning cell–cell interfaces to systemic CCI networks, revealing, for example, physical networks of CCIs, different biomolecules participating in specific cell–cell contacts, new LRIs and mechanisms of communication, and complex signalling activities^21–27. Thus, these techniques are key to scaling up CCI studies and validating next-generation computational tools.

Previous reviews have introduced the concepts for inferring CCIs, collecting and building LRI databases, implementing analysis workflows, and understanding the mathematical and statistical strategies that tools employ^7–15. Here, we review the different ways computational and experimental tools have recently evolved to improve CCI research (Fig. 1), offering here an expansive catalogue of tools. We first highlight the defining features of different next-generation computational tools, their limitations, example applications, and resulting discoveries. Next, we introduce cutting-edge methods with higher throughput to experimentally track CCIs, describing the evaluations that they enable, and how they help refine computational tools and make them more accurate. Finally, we discuss challenges and future directions of the field.

Next-generation computational tools

A variety of core tools^{16,17,35–43}, such as CellPhoneDB¹⁶ and CellChat¹⁷, established a set of methods used by the community to infer CCIs from transcriptomics. These ‘core tools’ are methods that that are based on the expression of ligands and receptors to infer CCIs through core scoring functions⁷ (such as expression mean, expression product, expression correlation, expression-based Hill function, and differential combinations) without being devised to specifically address any of the aspects in Fig. 1a. However, conceptual opportunities and technological advances are now driving further evolution of computational tools, which are adapting to specific pressures in different research niches (Fig. 2). Here we discuss features of this evolution and how they fulfil specific needs by adapting rule-based and data-driven strategies (Box 1) to answer additional biological questions.

Figure 2. Phylogenetic tree of computational tools for inferring cell-cell interactions.

There has been an evolution of computational tools to infer CCIs, derived from a ‘root’ of core tools. From this root, methods have become more specialised to address specific opportunities (grey arrows in the centre). Main branches growing from the centre represent the predominant features of tools (“core tools”, “finer”, “deeper”, “broader”, “more localised”, or “other improvements”). Coloured boxes indicate secondary features of each fate or sub-group of tools (coloured shades). A total of 105 tools are displayed here (see Supplementary Table 1 for further details, including summaries and repository availability). Tools are grouped as follows: (i) The “core tools” rely on core or similar scoring functions⁷ and are general frameworks for CCI analysis. (ii) Other branches capture tools with predominant features such as those associated with Fig. 1a (secondary features from different branches are further shown in Supplementary Table 1). (iii) The “other improvements” branch highlights tools whose predominant features are not in Fig. 1a (such as implementing interactive interfaces, enabling benchmarking, and focusing on multiple ligand-receptor interactions (LRIs) simultaneously to infer CCIs). (iv) Sub-branches and leaves are sorted by similarities in underlying algorithms, secondary features, or date of publication, defining the distinct sub-groups (coloured boxes and shades). A tool name is marked with an asterisk (*) if published as a preprint upon writing this review. Multiple versions of the same tool are treated separately if they include different features and were published in separate articles. DE, differentially expressed; LR, ligand–receptor; ST, spatial transcriptomics

Box 1. Rule-based and data-driven computational strategies.

CCI tools utilise diverse computational strategies, many of which can be categorised into rule-based and data-driven approaches (Supplementary Table 1). Rule-based tools (such as SoptSC¹⁹, NICHES⁴⁴, ICELLNET³⁹, and NATMI¹²⁵) incorporate assumptions or prior knowledge about CCI behaviour and model interactions using principles associated with ligand and receptor quantity (such as thresholding ligand and receptor expression, or using expression levels as inputs of continuous core functions describing the mode of interaction⁷), or by defining interaction rules (such as agent-based models). However, rule-based methods may struggle with accommodating higher CCI complexity, noisy data and unaccounted variables. Alternatively, data-driven tools primarily use statistical tests or machine learning (such as differential analysis, label permutations, regression models, factorization methods, and deep learning) to interpret gene expression. These methods can reveal unexpected correlations and hidden patterns within large datasets, even when the underlying mechanisms are poorly understood (such as those involving non-LRIs). For instance, DIALOGUE¹²¹, MISTy⁵⁷, MOFAcell¹²², scITD¹²⁰, and Tensor-cell2cell³³ employ distinct types of factorization methods to extract properties of CCIs, although they demand substantial amounts of data.

Rule-based tools typically yield consistent results due to their reliance on gene-expression-based formulas. In contrast, data-driven tools might generate varying outputs on identical datasets due to inherent randomness in statistical tests (such as permutations) and the initialization of machine-learning algorithms (such as gradient-based methods). The reproducibility and robustness of data-driven models can be improved by using, for example, similarity metrics to assess and stabilise results from separate runs of the same tool, regardless of the stochastic nature of the algorithms^33,204. Hybrid models exploit strengths of rule-based and data-driven strategies by implementing a combination of both. For instance, CellChat¹⁷ and CellPhoneDB¹⁶ infer CCIs first through an expression-based formula for consistency, then employ statistical tests to extract significant LRIs⁷.

Although differing in their outputs, both tool types facilitate various downstream analyses. Rule-based methods’ results enable direct comparisons between top LRIs, CCI overrepresentation analysis, cell type clustering, and evaluation of signalling functions^{17,44,45,103,105,106,135,137}. Meanwhile, data-driven outputs from deep-learning and factorization methods compress data into loadings [G] and/or embeddings [G] that can be further analysed. For example, principal components or factors representing MCPs or CCI patterns, can help to cluster and rank samples, LRIs, and cell pairs and to associate biological functions through enrichment analysis^{33,121,187,205}. Tool outputs can also be inputs for other machine-learning tools that classify samples or LRIs^{33,121,187,206,207}. Understanding each CCI tool type is crucial for recognizing pros and cons, output behaviours, and potential downstream analyses.

Finer: Gaining insights at full single-cell resolution

Computational tools using scRNA-seq are predominantly applied at a pseudo-bulk [G] level, wherein single cells are aggregated into clusters or cell types. This approach has effectively dealt with the typical sparsity of measured transcripts in each cell in scRNA-seq. However, recent methods can now handle these data at true single-cell resolution (Fig. 2, Supplementary Table 1), inferring the communication between pairs of individual cells.

Inferring CCIs at the resolution of single cells does not rely on the cluster-wise average expression of genes. In this regard, SoptSC¹⁹ accounts for single-cell CCIs; however, its main focus is to assess intracellular signalling activities triggered by LRIs. Other tools such as NICHES⁴⁴ and Scriabin⁴⁵ leverage the methods applied by core tools⁷ to compute LRIs directly from single-cell pairs in a label-free manner, and these further expand the applications, outputs, and visualization options for single-cell CCIs. SPRUCE⁴⁶ and DeepCOLOR⁴⁷ project the gene expression of individual cells into a latent space [G] and use this information to infer CCIs between single-cell pairs. Furthermore, they facilitate downstream analyses from low-dimensional data, without unwanted biases introduced by cell-type annotations. That way, these next-generation methods leverage biological insights previously missed when aggregating cells.

Analysing CCIs at single-cell resolution provides insights into interaction heterogeneity across single cells both between and within cell groups. For instance, NICHES evaluates the ligands and receptors used by individual cell pairs (Fig. 3a), thus revealing another layer of heterogeneity and better defining cell-type subclusters that use different molecular mechanisms⁴⁴. Similarly, Scriabin helped reveal the diversity of interactions between individual T cells and CD1C+ dendritic cells (DCs) in the tumour microenvironment⁴⁵. In this case, the exhaustion phenotype was found to not be exclusive to discrete cell-subtype clusters; rather, it is present across multiple clusters (Fig. 3b). Importantly, non-exhausted and exhausted T cells interact differently with CD1C+ DCs, with exhausted T cells interacting through CTLA4 and TIGIT, while reducing their expression of pro-inflammatory chemokines, such as CCL4 and CCL5. Thus, these methods can capture the single-cell nature and heterogeneity of biological phenomena that may be missed by typical agglomerative tools.

Figure 3. New features and analyses performed by next-generation computational tools.

(a) Analysing cell-cell interactions (CCIs) at single-cell resolution enables one to identify ligand-receptor interactions (LRIs) that define heterogeneity of cell pairs (markers) and helps obtain more precise annotations related to their communication mechanisms. (b) Similarly, analysis of interactions between single cells (circles) can reveal phenotypic heterogeneity (different colours) that is not associated with cell annotations. (c,d) Tools that account for spatial transcriptomics explore local neighbourhoods of cells so they can spatially visualise signalling activities (c), and infer directionality of signalling pathways given the spatial distribution of receptors and/or diffusion of ligands (d). (e) Strategies integrating metabolite- or small-compound-based LRIs depend on expression of enzymes catalysing their production or consumption. (f) The inclusion of gene-regulatory networks and receptor-downstream transcription factors (TFs) can be used to infer signalling feedback loops between two cells, helping to capture interconnected layers of intracellular activity. (g,h) Next-generation tools also enable comparisons of multiple conditions either pairwise to detect differential CCI changes (g) or simultaneously to identify trends or patterns of CCIs (such as factors from factorization methods) (h).

There are challenges when inferring interactions between single cells. These include how to handle dropouts or zero-counts and scaling the analysis to larger datasets. To deal with data sparsity, denoising algorithms can be used before inferring CCIs⁴⁵. Furthermore, Scriabin and NICHES use zero-preserving [G] communication scores to avoid changing data sparsity. To handle large numbers of cells in single-cell datasets, NICHES subsamples cells to reduce computational demand⁴⁴; however, this procedure reduces statistical power as valid data are omitted and noise from random sampling is introduced⁴⁸. Alternatively, Scriabin can prioritize cell-cell pairs using properties of interest, or it can summarize them into a network of overall potential of interaction⁴⁵, reducing the demand on computational memory. Similarly, DeepCOLOR prioritizes co-localized cells in spatial spots⁴⁷. In contrast to other tools that scale to hundreds of thousands of cells, SPRUCE is uniquely more scalable to >10 million cell pairs due to its low dimensional nature⁴⁶, enabling the generation of large CCI atlases.

More localised: Spatially contextualising cells

Cell location affects intercellular interactions and associated gene expression^{2,29–31,49–51}. Molecules mediating CCIs form concentration gradients as they diffuse from their producing cells and trigger different signalling programmes in receiver cells^52,53. For this reason, it is important to account for the spatial context of cells to understand how tissues function. For example, the spatial division of labour among liver cells can be observed when incorporating spatial information into single-cell transcriptomics⁵⁴. In consequence, new CCI tools that inspect the spatial context of each cell will more clearly decipher biologically meaningful communication in complex tissues.

The blossoming field of spatial transcriptomics has enabled the evolution of computational tools to include cell location when inferring CCIs^{49,50,55–82} (Fig. 2, Supplementary Table 1). One branch of these tools more generally facilitates the analysis of spatial data^55–58. For instance, Giotto⁵⁵ and Squidpy⁵⁶ help users visualize spatial data and perform analyses across cells or spots, such as clustering and neighbourhood detection, but these also provide information for CCI inference (Fig. 3c). Other frameworks such as MISTy⁵⁷ identify more specific properties of spatial data through multiple ‘views’, that is, models describing marker expression relationships and their sources of variability in a domain-specific manner. Each ‘view’ encompasses distinct ranges of intercellular distances given a distance threshold, which helps to identify niche-specific mechanisms of CCIs. Such an approach has helped identify features that separate breast cancer biopsies by tumour-grade and clinical subtypes, while also revealing key signalling pathways⁵⁷.

More specialised spatial CCI inference started with earlier tools such as SVCA⁵⁹ and SpaOTsc⁷⁶ accounting for individual-gene and single-cell interactions respectively. Available tools directly incorporate intercellular distance, either to constrain the analysis to proximal cells and spots^59–69,82 or to weigh communication scores representing interaction potential^70–74,83. For instance, SpaTalk⁶³ spatially constrains the analysis by using the Euclidean distance between cells to build a cell graph network based on their K-nearest neighbours. It then computes CCIs of connected cells (that is, proximal cells). Another method, COMMOT⁶⁷, limits distance to infer CCIs between proximal cells. COMMOT introduces a collective optimal transport algorithm⁸⁴ [G] that uses a specified distance limit to define neighbourhoods and infer proximal CCIs. This algorithm seeks the most efficient way to transport a set of resources (such as ligands) between locations, minimising the overall cost while considering the distances between the source and destination (for example sender and receiver cells). Mathematically, it optimises the potential of LRI-based CCIs (transport plan) given intercellular distances (cost), while also penalising ligands and receptors that remain non-transported or ‘unused’. Furthermore, this algorithm can handle multiple competing species of ligands and receptors simultaneously and can infer the spatial signalling directionality (Fig. 3d). Other tools, such as stMLnet⁷¹, weigh the interaction potential, instead of constraining the analysis. For example, stMLnet uses principles of ligand diffusion by assuming an inverse relationship between intercellular distance and ligand concentration by scaling communication scores with intercellular Euclidean distances as a denominator. These various methods highlight the diverse ways in which distance can be explicitly accounted for in CCI analysis.

Deep learning has also helped explore spatial properties of CCIs (Box 2) using gene expression matrices and spatial cell graphs that connect neighbouring cells^47,79,81,85. For example, DeepLinc⁸⁰ implements an autoencoder [G] to infer a CCI network after considering both expression and cell graph inputs. This model differs from others by revealing both proximal and distal interactions; the latter are often missed by methods focused on the neighbourhood of a cell. Another tool, spaCI⁸¹, uses an encoder to generate latent features [G]. By considering gene-gene co-expression, this tool can handle dropouts in spatial transcriptomics and infer both ligand-receptor interactions and interactions of any gene pairs, such as upstream transcription factors (TFs) and their target ligands or receptors. Although many deep-learning-based tools can use similar inputs, they differ on how their neural networks are built (Box 2). DeepLinc incorporates a variational graph autoencoder and an adversarial network for regularisation⁸⁰, enabling the inference of unobserved CCIs (such as distal CCIs). Meanwhile, spaCI uses two separate encoders for gene expression information and spatial data, respectively. A third encoder, using the other two encoder outputs, then deciphers gene-triplet relationships, which detects upstream TFs mediating LRIs⁸¹. Thus, deep-learning methods can unveil biologically meaningful signalling activities.

Box 2. Deep learning leverages the inference of CCIs.

Deep-learning methods are rapidly expanding in computational biology²⁰⁸. Such models mimic the brain’s neural architecture and use interconnected nodes (neurons) to learn from the data. Notable neural network architectures (see figure) include (a) feedforward neural networks (FNN), (b) convolutional neural networks (CNN), (c) recurrent neural networks (RNN), (d) autoencoders, (e) graph neural networks (GNN), (f) generative adversarial networks (GAN), and (g) transformers. Each can be tailored for different purposes. Autoencoders have helped in CCI inference by using gene expression to predict LRIs⁴⁶ and to integrate spatial information^{47,79–81,85}. Similarly, GNNs can infer CCIs directly from single-cell¹³⁸ and spatial transcriptomics^62,79. Other deep-learning applications in CCIs include FNNs for inferring CCIs from gene-gene interactions¹¹⁴, predicting receptor conversion rates¹⁵⁰, and inferring high-quality LRIs from data rather than relying on databases^147,148. Large language models²⁰⁹ (LLMs)–designed to process and generate text using generative pre-trained transformers (GPTs)–are proving invaluable to computational biology²¹⁰. Indeed, specialised transformer-based models are surpassing the performance of previous models for specific tasks in single-cell omics. For instance, DeepMAPS²¹¹ implements a graph transformer to infer biological networks from multiple omics simultaneously. LLMs such as scGPT²¹² and scFoundation²¹³, trained on tens of millions of single-cell transcriptomes, can extract patterns of cells and enhance downstream analyses, such as cell-type annotation, gene expression enhancement, and drug/perturbation response predictions. These models and their downstream analysis capabilities show considerable potential for CCI analysis at single-cell and spatial levels. Furthermore, there are opportunities to design LLMs explicitly for studying CCIs in different cellular contexts, by integrating LRIs, gene expression, and spatial data from inception. Both strategies present robust opportunities for more advanced CCI-focused LLMs, capable of handling larger amounts of data and achieving better performance.

Other studies demonstrated that the spatial organisation of cells can be deciphered from the expression of ligands and receptors. For example, Neighbor-seq⁷⁸ creates a network of physical CCIs by inferring cells forming multiplets [G] to find enriched cell-cell contacts from a scRNA-seq dataset. This strategy identified CCI architectures in the spleen, small intestine, lungs, in pancreatic and skin cancers⁷⁸. Similarly, CSOmap⁵⁰ uses every ligand-receptor pair to compute an affinity score for each pair of cells. Then, cell pairs are projected into a pseudo-physical space, revealing their relative position within a tissue. Hence, these tools leverage CCIs to spatially contextualise cellular functions without necessarily using spatial locations as inputs. CCIs have also been reconstructed at a whole-body level in C. elegans^49,86. By using cell2cell⁴⁹, a negative correlation between CCIs and intercellular distances was identified. This method uses a genetic algorithm [G] to prioritise LRIs that are informative of cellular functions in distinct body regions. Thus, these methods can also provide hypotheses about how LRIs encode spatial information driving the 3D organisation of cells. Leveraging these results, a new benchmarking approach evaluates whether highly scored LRIs are enriched in spatially-proximal cell pairs⁸⁷. This strategy may help tune tools to generate biologically meaningful predictions, but may obscure biological processes involving long-distance interactions^88,89 and signals propagated through the circulatory system⁹⁰.

Deeper: Peering into intracellular activities

CCIs involve multiple types of molecules, including ions, small compounds, peptides, and proteins^7,91. Such molecules are important extracellularly (such as for ligand-receptor interactions) and intracellularly (such as for signalling, regulation of gene expression, and ligand biosynthesis). However, transcriptomics-based tools are often limited to simply inferring protein ligands, whose abundances correlate more with gene expression than other ligand types⁹². Nevertheless, systems biology pathway analysis methods can deepen insights into intracellular processes to infer abundance and activity of other classes of ligands, thus broadening the capabilities of CCI tools (Fig. 2, Supplementary Table 1).

To estimate CCIs involving non-protein ligands^6,91,93,94 (Fig. 2), such as small molecules, next-generation tools can analyse the expression of enzymes producing or consuming metabolite ligands, thereby capturing their communication potential (Fig. 3e). MEBOCOST⁹¹ incorporates metabolite-receptor and metabolite-transport interactions, based on a curated database. This tool was used to elucidate the role of epsins in macrophage-mediated metabolic regulation, revealing that epsins promote lipid uptake in atherosclerotic macrophages through CD36⁹⁵. Specifically, decreased macrophage communication through cholesterol-CD36 was observed in epsin-knockout mice. NeuronChat⁹³ is a tool that specialises in molecules that mediate communication among neurons, incorporating additional information that accounts for vesicular transporters and enzymes producing small compounds such as neurotransmitters. This tool was used to identified context-specific CCIs of the anterior lateral motor cortex (ALM) and primary visual cortex (VISp) in mouse brain; the glutamatergic cell-subtype L4/5 IT CTX was inferred as a major communication node using glutamate-Grin3a in ALM but not in VISp⁹³. Although strategies including metabolites have yielded additional biological insights, major limitations are apparent. Naturally, metabolite abundance depends less on the gene expression level of enzymes and transporters and more on metabolic fluxes and other modes of regulation, following more complex dynamics than proteins⁹⁶. Further work is therefore needed to refine predictions by incorporating other processes involved in metabolite production and secretion to better represent their abundance and activity.

Intracellular signalling pathways, including TFs and their downstream targets, can be also considered^{6,19,20,97–106} (Fig. 2), as introduced by SoptSC¹⁹ and NicheNet²⁰. For instance, some tools weigh the impact of CCIs on signalling and transcriptional regulation in the receiver cells^20,97–99. LRLoop⁹⁹ leverages a network propagation [G] algorithm¹⁰⁷ with intracellular gene regulatory networks to prioritise ligand-receptor interactions that form intercellular feedback loops [G] between two cells. That is, it uses gene expression to find sets of two ligand-receptor pairs that are intracellularly connected through gene regulatory networks, forming a closed loop (Fig. 3f). This versatile method can be further combined with other existing tools to find biologically-meaningful intercellular signalling, thus helping improve the true positive rate⁹⁹ [G] of the tool. Other tools directly use gene expression of TFs and/or target genes to prioritise specific LRIs^100–102. For example, Domino¹⁰⁰ focuses on receptors and their downstream genes to understand the immune response to implanted biomaterials. This tool was used to identify the activation of anti-inflammatory macrophages through Il4ra and its downstream gene Esrra, and the activation of injury repair signalling by endothelial cells through Osmr and its downstream TF Sox17. A third subgroup of tools uses downstream genes to weigh communication scores assigned to each ligand-receptor pair^19,103–106. CellComm¹⁰⁵ incorporates these weights into a network of ligand-receptor-to-regulon interactions and uses them to maximise the signalling flow with the lowest cost for the intercellular communication. Another tool, exFINDER¹⁰⁶ further leverages these weights to study signals received from cells that are not directly present in the data. Thus, tools incorporating intracellular signalling can more comprehensively represent the mechanisms associated with CCIs.

Broader: Accounting for multiple conditions

CCIs depend heavily on the biological context wherein they occur^32,33. Biological contexts, such as different genetic, cellular, extracellular, or temporal states¹⁰⁸, affect the behaviour of single cells and whole tissues. These contexts vary across phenotypic conditions, making it difficult to study CCIs in single-cell atlases, especially with their vast number of samples. As the number of experimental variables increases across samples, the complexity of analysis exponentially increases¹⁰⁹. This has been particularly challenging in CCI research, in which one also wants to consider the phenotypic association of all combinations of cells and LRIs.

Multiple approaches can facilitate the comparison of CCIs across different samples or conditions^33,110–131 (Fig. 2, Supplementary Table 1). However, most tools are limited to simple pairwise comparisons by performing differential analysis either on the genes encoding ligands and receptors or on the final scores of the CCI inference (Fig. 3g). For example, Connectome¹¹⁰, CellChat^111,112, scDiffCom¹¹³, and scTenifoldXct¹¹⁴ rely on pairwise comparisons between samples to find differential CCIs. scTenifoldXct offers further inference of CCIs that is not limited to LRIs. By leveraging two-sample comparisons, these tools can reveal specific LRIs and communication pathways that are altered in one condition. For example, comparing wound healing in young and aged mice revealed that ageing affects CCIs through growth factor, chemokine, and cytokine pathways¹¹². Specifically, aged skin wounds present dysregulated fibroblast signalling via communication through BMP-2/4/7 and TGF-β1/2/3. Hence, this imbalance affects fibroblast proliferation and therefore wound healing.

Dimensionality reduction approaches now enable direct comparison of >2 samples simultaneously (Fig. 3h). Tools such as sciTD¹²⁰, DIALOGUE¹²¹, and MOFAcell¹²² use gene expression values to infer multicellular programmes (MCPs) through factorization methods [G]. In these cases, the resulting factors correspond to each MCP and are assigned to pertinent samples. Thus, changes in CCIs can be extracted by using MCPs wherein ligands and receptors are determinant. Another method, Tensor-cell2cell³³, applies tensor factorization¹³² [G] directly on CCI scores computed across samples for each combination of LRIs, sender cells, and receiver cells. Instead of MCPs, this tool finds context-driven programmes of CCIs, which can be correlated, for example, with COVID-19 severity³³ or neurodevelopmental stages¹³³. TraSig¹²⁷ and TimeTalk¹²⁸ focus on differences across multiple (pseudo)time points, similarly identifying changes in CCI dynamics rather than in gene expression dynamics. In contrast to dimensionality reduction approaches, MultiNicheNet¹²⁹ performs a differential expression analysis¹³⁴ that can handle multiple samples, enabling it to account for batch effects and covariates. These sorts of methods will be increasingly important as most single-cell and spatial omics studies are expanding beyond two samples.

Other features emerging in the evolution of tools

Computational tools have evolved in ways beyond the aforementioned categories, enabling improvements that range from data handling and visualisation to the construction of high-accuracy LRI databases^135–150 (Fig. 2, Supplementary Table 1). User-friendly interactive interfaces are one key aspect implemented in tools such as InterCellar¹³⁵, Cellinker¹³⁶, and TALKIEN¹³⁷. Although most tools require coding skills, these tools facilitate each step in CCI analysis by implementing simple graphical user interfaces for handling inputs and deploying visualisations.

A few tools are emerging that benchmark or leverage diverse tools. Benchmarking is particularly important for next-generation tools; LIANA¹⁴⁰ established a platform for implementing methods from multiple tools, enabling a comparative analysis that revealed low overlap between methods, mainly due to the distinct approaches to prioritise relevant CCIs. To help manage the heterogeneity of outputs from diverse tools, LIANA also includes a consensus prediction among the different options, providing more reliable results. ESICCC¹⁵¹ is a more recent systematic framework that can benchmark more tools than LIANA and report multiple performance metrics. Furthermore, a few tools aim to find significant results by simulating the universe of possible CCIs through, for example, Monte Carlo¹⁴² and agent-based models¹⁵⁰, which enable in silico experimentation to explore ‘what-if’ scenarios without the need for further experiments.

Other methods employ data-driven models (Box 1) to evaluate the behaviour of multiple LRIs simultaneously and infer CCIs. For example, Calligraphy¹⁴⁶ uses communication gene co-expression networks to identify modules of signalling genes. This approach assumes that intercellular groups of genes behaving similarly provide stronger and less noisy signals than methods focused on one ligand–receptor pair at a time. Thus, Calligraphy successfully identified co-expression of modules involving IL-33 between epithelial and immune cells that direct pancreatic tumorigenesis¹⁴⁶. Some approaches also focus on building lists of high-confidence LRIs^147,148, a crucial step of the CCI analysis workflow, before inferring CCIs. These are merely a few examples of other features that are appearing among computational tools, each of which aim to improve more specific aspects of CCI predictions.

Tracking cell–cell interactions

Validating CCIs at the single-cell level has always been most easily approached with experimental studies focused on single interactions, such as fluorescence in situ hybridization (FISH), fluorescence resonance energy transfer (FRET), and immunostaining^7,9. However, it is difficult to use such approaches to validate CCI predictions coming from large-scale scRNA-seq studies, as they are limited to examining only a few cells and LRIs in a given experiment. Next-generation experimental methods are emerging, increasing the throughput of CCI measurements. These methods may measure many LRIs and cell-cell pairs simultaneously (Table 1), and many also couple experimentally-measured interactions with downstream sequencing. For example, such approaches have helped characterise large networks of CCIs^152–154, reveal molecular mechanisms^155–158, and trace CCIs in vivo¹⁵⁹. Although these technologies have been reviewed in detail^21–26, here we highlight examples that are pertinent to gaining biological insights about CCIs, and could help refine the aforementioned computational tools and validate their predictions, improving the fidelity of CCI inference.

Table 1.

Illustrative next-generation methods for tracking cell–cell interactions.

Name	Category	Description	Throughput	Limitations	Refs.
Methods based on sequencing technologies
MAPseq	Barcode-based	Barcodes delivered via a recombinant Sindbis virus uniquely label single-neurons, enabling tracking of their projections from source regions to target regions, as well as downstream NGS analysis.	High	Low spatial resolution. Intermediary cells between regions are not tracked.	161
BRICseq	Barcode-based	Extension of MAPseq that enables probing of multiple source regions simultaneously.	High	Limitations in sensitivity (that is, sequencing depth) prevent genuine single-cell resolution analysis, resulting in long-distance region-to-region connectivity analysis instead.	162
RABID-seq	Barcode-based	Barcodes delivered via a glycoprotein G–deficient pseudorabies virus uniquely label single-cells that transgenically express viral glycoprotein G and the EnvA receptor (TVA), enabling cell-type specific tracking of interacting cells, as well as downstream NGS analysis.	High	Proof-of-concept has only been applied to interactions between two cell types.	154
PIC-seq	Droplet-based	FAC-sorted doublets (physically interacting cells) are selected, sequenced, and reads are computationally assigned as a linear combination of the contribution from each cell type.	High	Can only capture physically interacting cells. Deconvolution approach only assigns cell types and read contributions with some probability according to a maximum likelihood estimate.	153
ProximID	Droplet-based	Identifies multiplets in single-cell RNA-sequencing outputs and relies on random forest classifiers trained on virtual structures to identify participating cell types.	High	Can only capture physically interacting cells. Only identifies participating cell types with some probability.	152
SPEAC-seq	Droplet-based	Cells engineered to express reporters and/or genetic perturbations are encapsulated and co-cultured in droplets and subsequently sequenced, identifying differential short-range molecular mediators of interactions.	Medium	In vitro co-culturing may not fully represent in situ interactions. Requires engineering of reporters.	163
Methods based on proximity labelling
LIPSTIC	Contact-dependent	Genetic fusions of a ligand with SrtA and a receptor with 5 N-terminal glycines are generated, enabling labelling of the receptor with SrtA substrate upon a physical interaction. Labelled receiving cells can further be sequenced.	Medium	Only one ligand–receptor pair can be probed in a given experiment.	155
uLIPSTIC	Contact-dependent	Extends LIPSTIC to probe cell interactions in a ligand–receptor pair independent manner such that interactions can be rapidly captured by fusing the enzyme and acceptor pairs to the cell membrane directly. Interaction intensity measured by extent of receptor labelling can be a proxy for interaction strength. Interaction readouts can be coupled downstream with sequencing to identify cell types and molecular mediators.	High	When multiple source cell types are generically labelled, it is not possible to identify which sender cell type participated in the interaction that caused the labelling.	169
EXCELL	Contact-dependent	Similar to LIPSTIC; however, engineers SrtA to promiscuously label receiver cells without the need for an engineered receptor.	Medium	The sender cell type(s) that can be probed is limited by the ligand fusion.	156
FucoID	Contact-dependent	Sender cells are functionalized on the surface with a fucosyltransferase. Receiver cells are labelled with a guanosine diphosphate (GDP)–fucose (Fuc)–biotin (GF-Biotin) on the cell surface disaccharide LacNAc upon addition of substrate.	High	A sender cell of interest must be specified (cannot probe multiple simultaneously).	157
TransitID	Contact-independent	Uses two orthogonal proximity labelling enzymes to first label proteins in a source location and subsequently those in a target location. Dual-labelled proteins represent those that transit between two locations (including across interacting cells).	Medium	The number of cell types involved in an interaction is limited by the number of orthogonal enzymes that can be used for labelling (currently two).	171
PhoTag	Contact-independent	Ligand-specific antibodies conjugated to photoactivatable flavin-based cofactors enable labelling of the respective receptor upon irradiation.	Medium	A specific ligand must be targeted, limiting each experiment to a given ligand–receptor pair.	158
μMap	Contact-independent	Photo-labelling of the (high-resolution) spatial microenvironment (nearby proteins) of a surface receptor or ligand–receptor pair targeted by a photocatalyst-antibody conjugate.	Low	Sender cells of a given ligand cannot be identified.	166
Methods based on synthetic circuits
SynNotch	Synthetic Notch receptor	Replaces Notch receptor’s extracellular sensor module and the intracellular transcriptional module with heterologous protein domains to perform user-defined functional responses in transfected and interacting cells.	Low	Limited to probing in immortalised and primary cell lines. Requires substantial genetic engineering.	178
SyNPL	Synthetic Notch receptor	Engineers the sender cell to express membrane bound GFP (EGFP) and the receiver cell to express a synthetic Notch receptor that induces expression of mCherry upon engagement with EGFP, thus reporting on sender-receiver cell contacts, even in vivo.	Medium	Limited to probing specific cell pairs. Requires substantial genetic engineering.	180

EnvA, envelope protein of subgroup A; FAC, fluorescence-activated cell; LacNAc, N-Acetyllactosamine; NGS, next-generation sequencing; SrtA, sortase A.

Profiling cell–cell interactions through sequencing technologies

By taking advantage of sequencing technologies, several experimental tools have facilitated the study of cell-cell contacts^{152–154,160–164}. Nucleic acid barcodes can be introduced into sender cells and then transferred to receiver cells via CCIs (Fig. 4a). Alternatively, droplet-based methods can isolate multiplets of cells to capture physically interacting cells (Fig. 4b). Both methods report unique sender-receiver contacts at single-cell resolution upon sequencing, which can be used to test computational CCI predictions.

Figure 4. Major approaches in next-generation experimental methods for studying cell-cell interactions.

Sequencing technologies have enabled the measurement of single-cell transcriptomes while also (a) tracking barcodes passed between physically interacting cells (such as those delivered through transfection or by engineered viruses) or (b) directly isolating interacting cells by generating multiplets (such as by using fluorescence-activated cell sorting (FACS) or microfluidics). Thus, cell-cell contact networks can be built while inferring the ligand-receptor interactions (LRI) used by these cells. (c,d) Cell-cell interactions (CCIs) between proximal cells can be evaluated by labelling methods either relying on (c) enzymes catalysing the binding of a probe to an acceptor to tag the LRIs in a contact-dependent way or (d) catalysts inducing a highly-reactive state of a diffusible tag (with a radius that depends on its half-life) to label LRIs in a contact-independent way. (e) Synthetic receptors can be devised to induce a transcriptional response of interest within a receiver cell each time they interact with a specific ligand produced by a sender cell. For example, a membrane-bound green fluorescent protein (GFP) in a sender cell can be used with a synthetic receptor made from an anti-GFP (a-GFP) nanobody and a Notch receptor. Such synthetic receptors can help track sender-receiver interactions in vivo¹⁵⁹.

Barcode-based technologies have been invaluable in neuroscience. For instance, MAPseq¹⁶¹ uses viral particles to introduce RNA barcodes and track single-neuron projections from a source region. BRICseq¹⁶² and RABID-seq¹⁵⁴ both build on this concept to study CCIs across multiple tissue regions and at single-cell resolution, respectively. By accounting for the transcriptomes of interacting cells, BRICseq helped identify genes relevant to brain connectivity across brain regions¹⁶², whereas RABID-seq helped discover Sema4D/PlexinB1, Sema4D/PlexinB2, and Ephrin-B3/EphB3 as LRIs dictating pathological interactions between microglia and astrocytes in autoimmunity.

Droplet-based methods such as ProximID¹⁵² and PIC-seq¹⁵³ enable high-throughput determination of CCIs from transcriptomic analysis. Rather than using barcodes, these methods take advantage of the fact that physically interacting cells form multiplets (such as doublets or triplets) during cell-sorting. Such methods are particularly good for probing interactions between physically interacting cells at the moment of measurement. Additionally, they require probabilistic algorithms to deconvolve the gene expression profiles of multiple cell types sequenced together, which can be more challenging as the number of cell types increases in each multiplet. Nevertheless, their value has been proven in various applications. PIC-seq, for example, was used to identify enriched cell interactions and differentially regulated genes in the presence of specific CCIs, such as upregulation of glucocorticoid-responsive genes between premature alveolar macrophages and alveolar type I epithelial cells during lung development. SPEAC-seq^163,165 circumvents the requirement for cells to be physically interacting by instead encapsulating cell pairs in water-in-oil droplets. Beyond capturing additional mediators such as soluble secreted ligands, this approach can screen for genetic perturbations in the sender cell and quantify responses in the receiver cell. However, it requires engineering of reporters prior to encapsulation that decrease throughput and co-culturing after encapsulation that may not reflect in vivo conditions.

Proximity labelling of cell–cell interactions

CCIs can be labelled with chemical modifications that depend on the proximity of enzymes or highly reactive compounds. Such approaches can help researchers study the microenvironments and molecular presence at the cell-cell interface between interacting cells^{155–158,166–168}. These techniques can be classified as contact-dependent or contact-independent tagging methods²². Here we discuss both approaches, focusing on methods that helped detect intercellular relationships and uncover important aspects of CCIs.

In contact-dependent methods, an enzyme is localised to the cell surface of one interacting cell type, and this enzyme catalyses the binding of a probe to an acceptor on the surface of surrounding cells (Fig. 4c). Thus, these methods exhibit a small radius in which interacting cells are labelled due to the physical constraint imposed by the enzyme-acceptor interaction. For example, LIPSTIC¹⁵⁵ performs a Sortase A (SrtA)‐mediated ligation to enable the in vivo labelling of ligand–receptor pairs. Here, a peptide substrate containing the motif LPETG is transferred by SrtA onto an N-terminal pentaglycine (G5) acceptor. Thus, by fusing SrtA and G5 to one member of a LRI respectively, this method was applied to study murine immune cell synapses. Specifically, by focusing on the CD40-CD40L interaction, LIPSTIC helped show that T-cells initially interact with dendritic cells (DCs) in an antigen-specific manner, but upon activation they interact promiscuously¹⁵⁵.

Many contact-dependent methods require engineering of specific enzyme-acceptor pairs, and so are inherently limited in throughput and in the total number of cell types or LRIs that can be simultaneously tested. However, novel strategies such as EXCELL¹⁵⁶ bypass this limitation by engineering SrtA to promiscuously label N-terminal monoglycines of interacting cells. FucoID¹⁵⁷, a method that relies on a fucosyltransferase expressed on a bait-cell membrane and the addition of a biotinylated substrate to label a prey cell, increases throughput by using chemoenzymatic functionalization of enzymes rather than genetic engineering. Moreover, this method can distinguish strong CCIs from weaker ones. An extension of FucoID that uses antibody-based probes also enabled detection of CCIs without purification of cells of interest from their samples¹⁶⁸. A more recent version of LIPSTIC, universal LIPSTIC (uLIPSTIC)¹⁶⁹, overcomes the limitation of studying individual LRIs by discarding the direct fusion of SrtA and G5 to the interacting ligands and receptors. Instead, it fuses SrtA and G5 to ubiquitously expressed proteins on the cell membrane. When opposing membranes come into close proximity (<14nm) during functional interactions, the addition of a biotin-LPETG substrate facilitates SrtA mediated biotin-labelling of G5 to act as a universal reporter. By accompanying uLIPSTIC with functional perturbations of key markers and/or LRI partners, this method can report the importance of any LRIs of interest for the overall CCI.

Contact-independent methods rely on highly reactive compounds to label CCIs. In these cases, after a catalyst transforms small compounds into a highly reactive state, they diffuse and tag neighbouring molecules (Fig. 4d). Thus, their tagging radius depends on the half-life of their reactive state, meaning that distinct compounds can be employed to control the resolution for studying CCIs. Multiple methods use engineered ascorbate peroxidases (APEX) to make compounds reactive^27,170. TransitID¹⁷¹ leverages two orthogonal promiscuous proximity labelling enzymes (PLEs), TurboID¹⁷² and APEX2¹⁷³, to track protein trafficking from a source location to a destination, including subcellular trafficking and intercellular interactions. In particular, TransitID captures proteome-wide interactions at multiple time points and nanometer spatial resolution. Furthermore, it identifies CCIs beyond direct cell contacts, distinguishing between soluble protein ligands binding cell membrane receptors and exosome and nanotubule trafficking to the intracellular compartments of the receiver cell. However, without multiple orthogonal PLEs, it cannot probe more than two cell types at a time. Furthermore, a major limitation of APEX-based methods is that they rely on a toxic peroxide, thus limiting the experimental conditions to perform CCI studies.

Other contact-independent methods have overcome the issue of employing toxic compounds. BioID¹⁷⁴, TurboID¹⁶⁷, and Split-TurboID^172,175 use promiscuous biotin ligases fused to a protein of interest. These engineered proteins can adenylate biotin, which later diffuses and labels other proximal proteins, enabling the study of LRIs. Photocatalytic tagging approaches also avoid the use of toxic compounds. Cases such as PhoTag¹⁵⁸ and μMap¹⁶⁶ use light to control the highly-reactive state of the compounds, and therefore their tagging radius. Like FucoID, they also help to bypass the need for genetic engineering. In this regard, PhoTag helped identify T-cells that transiently interacted with Rajis via PD-1/PD-L1 and distinguish distinct response patterns across T-cell subpopulations. Although PhoTag is specific to a single LRI, μMap captures extremely high resolution (1nm) interactions of a ligand with all its local receptors. This resolution has high utility, capable of distinguishing between spatially distinct plasma membrane receptors on the same cell, although it does not identify the sender cell for a given ligand.

Synthetic circuits for tracking cell–cell interactions and intracellular activities

Synthetic receptors and pathway sensors can reveal CCIs and quantify the effect of a specific ligand-receptor interaction on receiver cells^{159,176–180}. Synthetic receptors allow customization of how cells sense and respond to different signals (Fig. 4e). These systems can be designed to report intercellular contacts after binding membrane-bound ligands, reveal cues in the cellular microenvironment after binding soluble ligands, and elucidate the presence of intracellular factors by binding ligands in the intracellular space²⁴. Synthetic receptors such as SynNotch¹⁷⁸ enabled the labelling of direct cell-cell contacts to assess temporal processes such as embryonic development¹⁵⁹. For example, it helped show that endothelial receiver cells that interact with cardiomyocytes migrate from the heart to the liver to form vasculature, and when interacting with tumours, they activate angiogenic, migratory, and inflammatory responses¹⁵⁹. SynNotch was also adapted to study CCIs in mammalian development in the SyNPL system, which tracks and manipulates sender-receiver interactions in vitro and in vivo¹⁸⁰.

Downstream intracellular activities triggered by the LRI, such as signal transduction, can be studied through synthetic circuits within intracellular pathways of interest²⁶. These circuits often rely on genetic constructs bearing pathway reporters. For example, the action of 21 different SARS-CoV-2 proteins in regulating 10 signalling pathways have been identified by coupling distinct pathway promoters with luciferase¹⁸¹. Although in this case the effect was assessed on cells overexpressing these proteins, this approach holds potential for evaluating the impact of signals coming from other cells. In another example, in vitro multiplex assays testing the effect of small molecule odorant ligands binding to olfactory receptors, the largest family of G protein-coupled receptors (GPCRs), identified tens of novel interactions¹⁸². Importantly, these new interactions revealed ligands that were unknown for 15 receptors. This approach was medium-throughput, as it involved engineering barcode reporters but could be scaled to 96-well plates¹⁸². Other previously mentioned techniques could be also combined with this approach to reach even higher throughput¹⁸³.

Challenges and opportunities

The aforementioned innovations have shed new light on CCIs with greater throughput and depth at the single-cell level. Although both computational and experimental methods have allowed new biological questions to be answered, various challenges and opportunities for further innovation remain (Fig. 5).

Figure 5. Challenges and opportunities for future enhancement of methods for cell-cell interaction research.

(a) Aligning single cells across multiple conditions is difficult because each sample inherently differs in cell number, which challenges cell-cell interaction (CCI) comparisons across conditions with single-cell resolution. Without correction, this case would lead to the pigeonhole principle²⁰³ (that is, not all cells can be assigned a one-to-one alignment from one condition to another). (b) Combinations of ligand-receptor interactions (LRIs) given their different protein variants¹⁹¹ also shape the expression of downstream genes, leading to different transcriptional responses for similar CCIs. By incorporating this information into CCI research, one may improve predictions and capture new biological insights of signalling activities. (c) Generating ground-truth data has been a major challenge for years; however, to address this, curated databases of real interactions are being developed, helping to benchmark and tune computational tools. Furthermore, community-organised efforts will be key for generating gold-standard data using next-generation experimental methods. (d) Including sub-cellular localization of ligands and receptors could refine predictions of interacting cells. (e) Most experimental methods are limited to in situ and in vitro deployment, with little work conducted in vivo. Further advances will be obtained with experimental methods that allow simultaneous evaluation of many LRIs and libraries of engineered ligands and receptors for tracking interactions.

Major discoveries rely on increased scale of CCI studies, either by introducing new methods that infer CCIs with single-cell resolution or that allow the simultaneous analysis of CCIs in multiple samples from diverse conditions. Hence, future work can further improve CCI tools by including both cases simultaneously; however, this requires alignment of single cells across conditions. Unfortunately, samples do not contain the same number of cells, and this impedes matching cells in one assay with cells in another (Fig. 5a). As such, differences in the cell number of samples can be incompatible with certain algorithms, meaning that there may be a need to include data imputation, especially at the single-cell resolution. For example, DURIAN¹⁸⁴ imputes gene expression in single-cell data and deconvolves bulk data into a single-cell level, holding potential to facilitate cross-condition comparisons with single-cell resolution of CCIs. Moreover, the development of comprehensive frameworks for inferring CCI, such as the most recent versions of CellChat (V2)¹⁸⁵, CellPhoneDB (V5)¹⁸⁶ and LIANA+¹⁸⁷, could help in this task since they are designed to include a variety of strategies and features covering most of the aspects of tool diversification (Fig. 1a). For example, LIANA+ is an all-in-one tool, even powered by multiple other tools, capable of integrating multimodal data, inferring CCI based on non-protein ligands, detecting CCI patterns, inspecting intracellular changes, and evaluating multiple conditions with both hypothesis-free and hypothesis-driven approaches¹⁸⁷.

Building LRI databases is critical for studying CCIs⁷. This step involves essential aspects such as focusing on high-confidence LRIs^{125,140,147,148,188}, including protein subunits, activators, inhibitors, and/or competitors to consider signalling activity of LRIs^16–18, and incorporating gene regulatory networks to capture intracellular processes^18–20. In this regard, AlphaFold¹⁸⁹ has shown great potential for revealing unidentified LRIs¹⁹⁰. Nevertheless, building databases that are biologically comprehensive remain to be fully addressed, limiting the performance that tools can achieve to infer CCIs. Specific combinations of ligands and their cognate receptors can trigger distinct signalling activities in different cell types^34,191. In particular, distinct ligand variants can promiscuously interact with different receptor variants. This results in a many-to-many relationship in which each ligand variant binds, with distinct strength, to multiple receptor variants and vice versa (Fig. 5b), meaning that ligands and receptors compete for interacting with their partners and this competition is dependent on genetic backgrounds. Hence, the signalling processes of a receiver cell would depend on context-specific expression profiles of its receptors. Therefore, novel databases and algorithms including ligand and receptor variants, their interaction affinity, and expression profiles could serve to better capture how intracellular events are activated, helping to increase the robustness of CCI predictions.

CCI tools also suffer from high false positives rates due to limited ground-truth data, which is essential for benchmarking analyses that evaluate and help improve the performance of these tools (Fig. 5c). This is especially critical when large atlases of inferred CCIs are starting to emerge¹⁹². Thus, efforts to generate gold-standard datasets will be crucial to evaluate tool performance through proper metrics (such as accuracy, precision, recall, F-score, and root-mean-square error). Attempts to build such reference data often involve manual curations of CCIs from literature, including diverse cell types and molecular mediators, as seen in, for example, citeDB¹⁹³ and Cellinker¹³⁶. Other cases are specialised in curated CCIs of cell- or tissue-specific datasets, such as a database of tumour–immune cell communication¹⁸⁸, and another of idiopathic pulmonary fibrosis¹⁹⁴. However, the existence of multiple tools leads into numerous benchmarking scenarios, making it challenging to conduct a comprehensive evaluation of tool performance. Nevertheless, a few studies have attempted to evaluate the ability of tools to capture CCIs in short-range distance⁸⁷, measure their robustness and consistency^140,149, and assess how tools perform in a specific benchmarking scenario¹⁹⁴ or in classifying samples by their phenotypes^33,187. Furthermore, the creation of frameworks such as scMultiSim¹⁴⁴ has been important to provide in silico ground truth, by simulating different modalities of single-cell behaviour, including gene regulatory network inference, RNA velocity estimation, batch effects, spatial organisation, and CCIs. However, none of these approaches cover in full the distinct aspects of CCIs to comprehensively help improve the performance of computational tools. Thus, further studies need to expand resources housing ground truth interactions by applying, for instance, the experimental methods described here (see Tracking cell–cell interactions). This could serve as the foundations of a bigger community-organised generation of gold-standard data across multiple benchmarking scenarios, as previously achieved for building a resource of surface proteins facilitating immune cell interactions¹⁹⁵.

Experimental methods have enabled the analysis of protein organisation at cell-cell interfaces²². However, current computational tools offer limited insight into membrane sub-localization of ligands and receptors (Fig. 5d). Novel algorithms that account for biophysical features (such as localization, strength, and cytoskeleton coupling) of membrane protein interactions could be developed to study the dynamics of the immune synapse and embryonic development¹⁹⁶. For example, the prediction of immunotherapy responses was improved by using the subcellular location of parent proteins of antigens¹⁹⁷, demonstrating that subcellular location of molecular mediators could be informative for CCI inference.

Recent experimental approaches have improved the throughput and accuracy of measuring CCIs. However, many challenges remain, such as constraints on measuring mostly direct cell contacts; difficulties in measuring many LRIs and cell pairs simultaneously; and the challenges associated with measuring interactions in situ (Fig. 5e). These limitations and the required specialised expertise for the experimental methods (Table 1) further extend the gap that exists in efforts to validate the outputs of computational tools. Thus, few experimental methods have been used in complement with computational tools^153,154. However, many approaches discussed here can be coupled to downstream RNA-seq and then used for constraining the computational inference of CCIs. We anticipate that additional technologies will emerge that report more validated CCIs and couple these to downstream sequencing and computational analysis. For example, uLIPSTIC can be combined with scRNA-seq to unravel complex CCIs¹⁶⁹. Furthermore, combining these methods with orthogonal approaches to evaluate molecule and cell colocalization will help identify further CCIs and to deal with obstacles for measuring interactions (such as identifying source cells in methods lacking such information). For instance, to help scale up both CCI and LRI networks obtained experimentally, experimental methods could be combined with multiplexed protein imaging methods¹⁹⁸ like CODEX¹⁹⁹, an approach using antibodies conjugated with DNA-barcodes to spatially visualise multiple protein markers within a tissue. Moreover, emerging methods labelling secreted signals^200,201 and mechanisms of long-range communication²⁰² will be essential to make these high-throughput experimental methods even more comprehensive when tracking CCIs. Further innovations will continue to address the aforementioned limitations, and such progress will be essential for computational inference, providing excellent sources of training and validation by direct measurement of many CCIs in a context-specific manner.

Conclusions

Tools for studying CCIs have experienced remarkable diversification in recent years. Computational tools have evolved from core gene expression-based methods to next-generation strategies that incorporate additional biological properties of cells. Similarly, more advanced experimental methods have increased throughput capabilities, allowing simultaneous analysis of multiple communication pathways; thus, providing more comprehensive and biologically meaningful insights into CCIs. Both computational and experimental methods hold great complementarity and synergy that can expand their potential for impactful applications in areas such as biomedicine and personalised medicine. Thus, addressing these new opportunities will be central to further deepen our understanding of CCIs.

Glossary

Communication score

Score computed from the gene expression of a ligand and its cognate receptor in a sender and a receiver cell, respectively. The communication score depends on a tool-specific mathematical function

Pseudo-bulk

Resolution resulting from aggregating the gene expression of single cells into a higher group of cells, such as cluster, cell type, sub-cluster, or sub-cell-type groups

Latent space

Term used in machine learning to refer to a lower-dimensional representation of complex data, which enables meaningful feature extraction and manipulation, and the identification of structures or patterns in data

Zero-preserving

A property of data transformations that maintain the number or proportion of zero values observed in the original input data

Optimal transport algorithm

A mathematical method for moving and transforming distributions from one state to another with minimum cost

Autoencoder

A neural network composed of an encoder and a decoder that is trained to reconstruct its inputs, typically used for dimensionality reduction or feature learning

Latent features

Unobservable variables inferred from observed data to capture underlying structures

Multiplets

Multiple cells inadvertently captured and sequenced together as one cell or barcode

Genetic algorithm

A search and/or optimization algorithm based on natural evolution in which individuals are selected by their optimal fitness to an objective function

Network propagation

A set of probabilistic processes that model the spread of information within a network across time

Intercellular feedback loop

A two-way communication in which one ligand-receptor interaction triggers the production of a ligand by one cell. The interaction of this second ligand and its cognate receptor induces the expression of the first ligand on the other cell

True positive rate

Metric used to evaluate the performance of a model. Specifically, it measures the proportion of true positives with respect to the total actual positives. Also known as sensitivity or recall

Factorization methods

Unsupervised techniques that extract low-dimensional structure from data (that is, data decomposition), preserving essential information while reducing complexity

Tensor factorization

Decomposition method designed to extract properties of a multidimensional data structure, also known as a tensor (a matrix is a tensor of two dimensions, whereas higher-order tensors have more dimensions)

Loadings

A representation of the contribution of variables to principal components or factors generated by dimensionality reduction methods, revealing their significance in each factor

Embeddings

Low-dimensional representations of data that capture essential features, enabling effective learning and similarity measurement by a given machine-learning technique