Differential cellular communication inference framework for large-scale single-cell RNA-sequencing data

Abstract

Single-cell transcriptomics data have been widely used to characterize biological systems, particularly in studying cell–cell communication, which plays a significant role in many biological processes. Despite the availability of various computational tools for inferring cellular communication, quantifying variations across different experimental conditions at both intercellular and intracellular levels remains challenging. Moreover, available methods are in general limited in terms of flexibility in analyzing different experimental designs and the ability to visualize results in an easily interpretable way. Here, we present a generalizable computational framework designed to infer and support differential cellular communication analysis across two experimental conditions from large-scale single-cell transcriptomics data. The scSeqCommDiff tool employs a statistical and network-based computational approach for characterizing altered cellular cross-talk in a fast and memory-efficient way. The framework is complemented with CClens, a user-friendly Shiny app to facilitate interactive analysis of inferred cell–cell communication. Validation through spatial transcriptomics data, comparison with other tools, and application to large-scale datasets (including a cell atlas) confirms the reliability, scalability, and efficiency of the framework. Moreover, the application to a single-nucleus transcriptomics dataset shows the validity and ability of the proposed workflow to support and unravel alterations in cell–cell interactions among patients with amyotrophic lateral sclerosis and healthy subjects.

Graphical Abstract

Graphical Abstract.

Introduction

In multicellular organisms, cells constantly send and receive signals to coordinate and regulate vital biological processes, ranging from organ maintenance to cell activity. Communication among cells, called here intercellular signaling, is typically mediated by the interaction of two signaling molecules, i.e. a ligand and a receptor. Moreover, the ligand–receptor binding activates, inside the receiver cell, an intracellular signaling cascade which, in turn, can trigger the cellular response, such as the transcriptional regulation of downstream target genes.

Dysfunctionalities or alterations in cellular communication may be implicated in diverse diseases and phenotypes, such as neurodegenerative diseases [1, 2], cancer progression [3, 4], and drug resistance [5, 6]. Deciphering cellular communication and its impairment (i.e. differential cellular communication) is of paramount importance in comprehending the molecular basis of the physiopathology of a living organism. In this endeavor, single-cell RNA-sequencing (scRNA-seq) technology has emerged as a powerful and invaluable tool to reveal cellular heterogeneity in gene expression patterns in a high-throughput manner and at the individual cell level, with a resolution never possible before. With the continuous advancement of technology, scRNA-seq has become more cost-effective, widely available, and easily accessible. Furthermore, extensive large-scale studies and cell atlases have grown in number and complexity, encompassing a broader range of conditions, samples, and even time series studies [7–9].

In recent years, there has been increasing development of computational tools aimed at systematically inferring intercellular interactions from scRNA-seq data [10–12]. Some of these tools have even delved into exploring the intracellular downstream response within receiver cells. These methods rely on (i) prior knowledge of cellular communication (e.g. curated ligand–receptor pair resources, protein–protein interactions, or signaling pathway databases) and (ii) computational strategies to predict “ongoing” interactions between pairs of cell clusters (annotated as cell types).

Most methods are designed to decipher signaling within individual biological conditions, but they often lack a specific focus on exploring variations in cell–cell communication across two conditions. For example, certain methods independently infer interactions within each biological condition and then identify common and/or distinct signaling events, failing to consider the intensity or strength of these interactions, and overlooking variations across conditions [13–15]. Among the few methods able to consider two experimental conditions, most solely infer differential intercellular signaling between conditions. These tools differ in the computational methodology proposed, such as differential expression (DE) analysis of ligands or receptors [16–18], statistical testing of differential predicted ligand–receptor strength [19, 20], and network-based analysis [21–23]. Conversely, only a few tools propose methodologies to investigate both differential intercellular signaling and differential intracellular signaling across experimental conditions; for example, CytoTalk [24], Domino [15], LIANA+ [25], CellCallExt [26], and MultiNicheNet [27] specifically focus on variations in the signal transduction mechanism. In particular, only the latter two offer quantitative measures of intracellular signaling: MultiNicheNet achieves this by linking ligands to sets of differentially expressed genes, whereas CellCallExt quantifies alterations in transcription factor (TF) activity. In contrast, the remaining methods [15, 24, 25] primarily provide qualitative assessments of signaling pathways, identifying altered networks of intracellular genes without quantifying their activity. Therefore, there is a significant limitation in the availability of tools that can both quantify and prioritize interactions, considering both dysregulated intercellular signaling and the subsequent intracellular responses across different experimental conditions. A brief description of all the mentioned tools is provided in Supplementary Section S1 and Supplementary Table S1.

Furthermore, current single-cell experiments often involve collecting cells from various samples, such as different subjects, under different biological conditions. The predominant approach among cell–cell communication inference tools involves aggregating all cells from individual samples within a biological condition, referred to as the “conditions-aware scenario” [16–19, 21]. This pooling procedure aids in identifying consistent trends or differences between conditions but falls short in capturing intra-condition variability. Only a minority of tools, such as ICELLNET, LIANA+,and MultiNicheNet, take into account individual samples within a condition in their computational strategies, which is referred to as the “multi-sample conditions-aware scenario”, particularly valuable when dealing with studies from large sample cohorts or where sample heterogeneity is of interest. However, a critical limitation is that none of the proposed methods generalizes its framework to work seamlessly with both approaches, i.e conditions-aware and multi-sample conditions-aware scenarios (see Supplementary Table S1). This lack of generalizability limits the application of these tools across diverse experimental designs and datasets.

In addition to inference capabilities, existing methods include robust visualization features, but they typically require (advanced) programming skills and a prior selection of the data for visualization. To our knowledge, only a few tools, such as CellChat and InterCellar [28], include an interactive platform for exploring cell–cell communication at the intercellular level. However, these tools often have limited capacity to represent multiple dimensions (e.g. inter- and intracellular signaling) and complex scenarios, such as differential cell–cell communication, limiting their effectiveness in interpreting and deriving meaningful insights.

Another crucial consideration is the increasing availability of large-scale datasets, which poses a substantial challenge in terms of memory management and computational performance [29, 30]. Despite this, there is currently a lack of emphasis on developing fast and memory-efficient tools for inferring cell–cell communication from large-scale scRNA-seq data.

In summary, there is a critical need for a unified, adaptable, and scalable workflow that not only conducts, but also facilitates, the interpretation of differential cellular communication analysis from large-scale scRNA-seq data. To address this issue, we have developed a two-step bioinformatics framework based on a pair of tools: scSeqCommDiff and CClens (see Fig. 1). The scSeqCommDiff tool introduces a comprehensive and scalable framework for analyzing cell–cell communication across two experimental conditions. Its key innovations include: (i) the ability to characterize differential cell–cell communication across diverse experimental designs (i.e. both conditions-aware and multi-sample conditions-aware scenarios); (ii) a novel inference strategy for differential cellular cross-talk at both inter- and intracellular levels using robust statistics and network-based methods; and (iii) an optimized computational pipeline enabling fast, parallel, and memory-efficient analysis of complex and multi-gigabyte datasets. Additionally, our solution provides support for the interpretation and exploration of complex and high-dimensional cell–cell communication data through an interactive and user-friendly interface implemented in the CClens package.

Figure 1.

Schematic overview of the proposed two-step framework for analyzing differential cellular communication. The framework requires as input both a priori biological knowledge of cell–cell communication (built-in or user-specified resources) and a normalized scRNA-seq gene expression matrix (in a conditions-aware or multi-sample conditions-aware setting). As a first step, the scSeqCommDiff tool identifies and quantifies alterations in cellular communication across experimental conditions (A versus B) at both intercellular and intracellular levels. Differential intercellular communication is assessed by examining the strength of ligand–receptor interaction through different built-in scoring schemes and using statistical methods to determine significance. For intracellular signaling, scSeqCommDiff incorporates network analysis and statistical tests to identify differential transcriptional regulation within downstream signaling pathways. The implementation is optimized for fast and memory-efficient analysis using shared memory or sparse matrix data representations and parallel computations in order to handle large-scale datasets efficiently. As a second step, the CClens visualization tool supports the interpretation and visualization of scSeqCommDiff outputs, as well as output from other cell–cell communication inference tools. CClens offers flexible data filtering criteria for inspection and exploration, as well as providing rich visualizations to enhance understanding and analysis of differential cellular communication. Created in BioRender. Cesaro, G. (2025) https://BioRender.com/7ji4vvb.

To validate the scSeqCommDiff tool, predictions of differential cell–cell communication inferred from scRNA-seq data in human myocardial infarction were corroborated using coupled spatial transcriptomics (ST) data. Additionally, inspired by existing benchmarking frameworks for single-condition cell–cell communication inference [31–33], we propose an initial attempt to develop a quantitative benchmarking framework for differential cell–cell communication inference tools, enabling comparative assessments using matched scRNA-seq and ST data. Our results demonstrate that scSeqCommDiff performs comparably with or outperforms existing methods, highlighting its accuracy and robustness. In addition, we demonstrate the scalability and memory efficiency of the scSeqCommDiff approach in analyzing large-scale datasets by conducting the entire analysis on a cell atlas (∼900 000 cells) using a standard workstation. Finally, to better appreciate the different facets of our framework in a real-case analysis, we applied it to a large-scale single-nucleus RNA-sequencing dataset of the human primary motor cortex (∼250 000 nuclei) from a patient cohort with amyotrophic lateral sclerosis (ALS) disease and a control group. We demonstrated the efficacy of scSeqCommDiff in identifying biologically known interactions in ALS and unveiling alterations in the cell–cell communication process, and we also showcased the effectiveness of CClens in supporting and facilitating the interpretation of these results.

Materials and methods

Mathematical notation

Let be the set of genes and the set of cells. Consider two disjoint sets of cells, and , describing two distinct experimental or biological conditions, where and , and . Let and represent, respectively, the partitions (or clusters) of cells within and . Partition corresponds to a set of cells assigned to cell cluster (), typically reflecting a cell type, under experimental condition , while represents the corresponding assignment under experimental condition . Let and be partitions of and , respectively, representing sample/patient identifiers associated with their respective experimental conditions, if present.

In conditions-aware scenarios, we solely consider the experimental condition information and without accounting for individual sample data, i.e. pooling all cells across samples within a biological condition. Conversely, in the multi-sample conditions-aware scenario, we include sample-specific variables and alongside the experimental conditions.

Let be the set of ligands and the set of receptors, with and . Let be the set of ligand–receptor pairs defined as the tuples where and , defined in the ligand–receptor database.

Potential cell–cell interactions are defined as all possible ligand–receptor pairs () between each pair of cell types. Thus, we define as the cell–cell interaction (CCI) mediated by the -th ligand–receptor pair, i.e. , involving as sender cell type (e.g. the cluster expressing the ligand ) and as receiver cell type (i.e. the cluster expressing the receptor ) in the experimental condition . Similarly, represents the interaction under condition .

A common approach in investigating cell–cell communication involves assigning a score to each cell–cell interaction. Let us denote as the intercellular score quantifying the strength or intensity of the given interaction within the experimental condition , and similarly for experimental condition . While many existing tools focus on intercellular signaling, in a previous work we introduced the use of an intracellular scoring scheme, denoted as , specifically designed to quantify the evidence of observing a cellular response downstream of a [34]. For detailed descriptions of the possible intercellular scores and intracellular scoring scheme implemented in scSeqCommDiff, refer to Supplementary Section S2.

Differential cellular communication analysis using scSeqCommDiff

We designed a computational method (named scSeqCommDiff), available within the R package scSeqComm (at https://gitlab.com/sysbiobig/scseqcomm), to infer both intercellular and intracellular signaling that significantly change between two experimental conditions. First, each pair of cell types is assigned an intercellular score, independently in each condition or sample/patient. Subsequently, statistical evaluation is applied to assess if the intercellular score of each CCI undergoes significant changes between the two conditions of interest. Additionally, our method quantifies the degree of intracellular alterations by investigating differential intracellular signaling activation in the receiver cell. Our approach implements different analysis strategies to work with conditions-aware or multi-sample conditions-aware data, as discussed in the following sections.

Input. The method requires as input a gene expression matrix from single-cell transcriptomics experiments along with additional metadata on sequenced cells (i.e. cell type annotation, experimental condition labels, and patient/sample IDs, if available). We assume that scRNA-seq data are already pre-processed and normalized.

Supported data structures for the gene expression matrix are R naïve matrix object, R sparseMatrix object (dgCMatrix) from the Matrix R package, and big.matrix object from the bigmemory R package [35] with genes on rows and cells on columns. The latter two data structures are recommended for handling large-scale datasets.

Metadata, which include cell partitions by cell type and by experimental condition on which the differential analysis will be performed (i.e. conditions-aware scenario), along with (optional) additional potential sample/patient information (i.e. multi-sample conditions-aware scenario), must be provided as a data.frame object.

Finally, as input, the user can choose among several public databases of ligand–receptor pairs (16 human and 11 mouse resources), transcriptional regulatory network (4 human and 3 mouse resources), and signaling pathway (2 human and 2 mouse databases) included in our package or can provide her/his own resource. While the tool includes resources for human and mouse, it can be applied to other species, as long as users supply the necessary species-specific prior knowledge.

Given the diversity of mathematical strategies proposed to compute the intercellular score, users have the flexibility to choose the intercellular scoring scheme to best suit the intended cellular communication analysis among the six available in our tool. These scoring schemes include scSeqComm [34], SingleCellSignalR [36], a simplified version of CellChat [23], product-expression and specificity NATMI weights [13], and scDiffCom [19] scores.

Algorithm for the conditions-aware scenario.Our approach relies on pooling all cells across samples within the same biological condition and on a permutation test to assess whether a CCI at the intercellular level is differentially ongoing between two distinct experimental conditions A and B. Denoting the intercellular scores of a specific cell–cell communication in experimental condition A and B as  and , respectively, the difference in the intercellular score between condition A and B is defined as:

(1)

The distribution of under the null hypothesis, which assumes no difference in the intercellular score values, denoted as , is derived by permuting the experimental condition labels independently within each cell cluster. Each permutation results in two random experimental groups and , and the difference in intercellular score under the null hypothesis is computed as:

(2)

The procedure is iterated times to create a distribution of values under the null hypothesis for each pair of cell types (default equal to 1000). The observed value is compared with the null distribution to compute the empirical P-value (denoted as ), assessing greater or lesser extreme values, as well as the two-sided P-value. The P-values are then corrected for multiple testing using the p.adjust R function (Bonferroni as default).

Along with statistical significance, the method provides additional metrics to prioritize results based on user needs, such as the magnitude of the intercellular score difference between and , i.e. (from Equation 1) and the log fold change

(3)

as well as the statistical significance of DE analysis applied independently to ligands and receptors, using method such as MAST [37] or the Wilcoxon Mann–Whitney test.

Leveraging our previous work [34] to identify and quantify differential intracellular signaling, we have adapted the scSeqComm intracellular scoring scheme to measure the evidence of ongoing differential transcriptional regulation of downstream target genes by TFs between the two experimental conditions. Specifically, the evidence of a differential TF regulation relies on DE analysis, computed by comparing target gene expression levels in the receptor-expressing cell type across experimental conditions. Common single-cell DE methods, such as the MAST test or, non-parametric statistical tests, such as the Wilcoxon Mann–Whitney test, are employed to identify up- and down-regulated target genes across conditions and . Subsequently, the evidence of TF differential regulation is weighted based on the topology of the intracellular gene network using the Personalized PageRank algorithm [38]. The resulting score , ranging from 0 to 1, quantifies the strength of the evidence for observing a different intracellular signaling in condition compared with  downstream of the given interaction . A higher score means greater evidence of observing an altered transcriptomics response. To complete the information provided as output, the list of up- and down-regulated target genes is also reported.

Additionally, we have integrated a Gene Ontology (GO) enrichment analysis step to perform functional characterization on the up-/down-regulated target genes and identify enriched biological processes associated with the detected differential intracellular signaling, thereby providing insights into the functional implications of the observed transcriptional changes.

Algorithm for the multi-sample conditions-aware scenario. To statistically compare the intercellular scores between the experimental condition and in a multi-sample conditions-aware scenario, we apply the intercellular scoring scheme at the sample level. This involves computing for every sample/patient  , and for every . Subsequently, we obtain distributions of || and || intercellular scores, where || represents the number of samples for the respective experimental condition. Next, we compare intercellular communication scores for each CCI between conditions using a non-parametric statistical test (i.e. Wilcoxon Mann–Whitney test), corrected for multiple testing (Benjamini–Hochberg as default). Similarly to the conditions-aware scenario, additional metrics, such as the standardized effect size measure Cohen's d [39] and ligand/receptor DE P-values, are provided to complement information from the statistical significance level.

Furthermore, our approach for computing differential intracellular signaling integrates contemporary pseudo-bulk methodology [40]. This process involves the aggregation of gene expression levels at sample level for each cell cluster (by average or sum). The resulting pseudo-bulk data represent the cell type-level expression profile across samples in both experimental conditions. Statistical tests, such as the Wilcoxon Mann–Whitney test, are then applied to identify differentially regulated target genes across conditions and . Similar to the conditions-aware scenario, we incorporate this approach within the scSeqComm intracellular scoring scheme to detect differential transcriptional regulation, yielding a 0-to-1 intracellular score. The output also includes the list of up- and down-regulated target genes, which can be used to functionally characterize the detected differential cellular response through the integrated GO.

Output. The main output of scSeqCommDiff is a table containing the results of differential cellular communication analysis (see Table 1). Each entry corresponds to a potential CCI, which is a ligand–receptor pair for each interacting pair of cell types. Details on differential intercellular signaling include (i) an independent intercellular score in each experimental condition, and ; (ii) adjusted P-values , indicating the significance of changes in intercellular scores across conditions; and (iii) the magnitude of the difference in intercellular signaling, such as (Equation 3). Furthermore, a differential intracellular score is assigned for each CCI, providing evidence of observing an altered intracellular response in the receiver cell type. This score is computed for each signaling pathway involved, allowing for a pathway-specific analysis of the intracellular changes. Furthermore, is associated with up- and down-regulated downstream target genes within each pathway, providing a more detailed understanding of the intracellular changes driven by altered cell–cell interactions.

Table 1.

Example of the main output provided by scSeqCommDiff in tabular format

L–R pair	Sender–Receiver						DE genes
NRG1–ERBB3	Oligodendrocyte-Fibroblast	0.066	1	0.008	-3.921	0.079	LAMP1, BCAS1, …
SLIT3–ROBO1	T cell–Microglia	0.976	0.192	0.065	2.346	0.739	EPHA4, CBLB, FOXO3, …
TGFB2–ACVR1	Microglia–Fibroblast	0.232	0.299	0.775	-0.366	0.042	UGT8
…	…	…	…	…	…	…	…

Memory-efficient and parallel implementation

Working with large-scale datasets poses a significant challenge in terms of efficient memory management and performance. Consequently, substantial effort has been dedicated to code design and optimization.

Loading and managing large single-cell datasets within R’s memory using a dense representation (i.e. R naïve matrix) can be challenging both for R’s intrinsic memory limit for elements to be indexed (i.e. R’s 32 bit indexing of matrices) and for a large memory footprint. To achieve efficient and flexible memory representation of multi-gigabyte gene expression matrices, our method supports alternative data structures. Given the inherent sparsity of single-cell gene expression data, an effective approach to represent data is using the R sparseMatrix (typically, dgCMatrix) data structure from the Matrix R package. By only storing the non-zero elements, this R object enables efficient memory usage and faster computations, and it is widely adopted by several R packages for storing and processing single-cell data (e.g. Seurat, SingleCellExperiment).

Additionally, we support representing the gene expression matrix using a big.matrix object [35]. This data structure allows memory to be saved by storing numeric values using single-precision floating-point numbers, in contrast to the R intrinsic double-precision representation. On the other hand, it requires the conversion of data from commonly used data structures to a big.matrix object, a step that involves additional computation. However, this initial effort is offset by the improved computational performance in the following analysis.

To further enhance performance, our method relies on the use of the Rcpp R package [41] to achieve fast operations on a massive gene expression matrix, integrating C++ code into our R package. Specifically, we implement de novo C++ computations (i.e. summary statistics on matrices and efficient Wilcoxon Mann–Whitney statistical test) that operate efficiently on sparseMatrix and big.matrix objects. The C++ code in our method achieves fast and memory-efficient computations compared with conventional R code, which significantly enhances computational performance, particularly when handling large-scale gene expression datasets.

Additionally, faster computation is achieved using the doParallel and doMC R packages which implement different parallel strategies tailored to the operation system (e.g. forking system for Unix-like operating systems). While parallelism with naïve R matrix and sparseMatrix objects is handled by R’s memory management, which may lack the same level of efficiency across different backends and operating systems, big.memory supports shared-memory parallelism, significantly reducing the memory footprint.

Validation: assessment and benchmarking framework

To validate the predictions, we integrated ST data with scRNA-seq data. The assumption is that cell types with differential cell–cell interactions should also exhibit spatial co-localization [27, 31, 42]. Specifically, if two cell types interact more in one condition than in another, they should be more spatially co-localized in that condition.

We quantified the expected spatial distance and observed the interaction strength for each cell type pair using ST and scRNA-seq data, respectively. The expected spatial distance was determined by assessing co-localization between cell type pairs. Meanwhile, the observed interaction strength was measured by applying differential cell–cell communication inference tools to identify differentially regulated interactions and counting the number of detected interactions for each cell type pair. Note that, since communication strength differs depending on the directionality of cell type pairs (i.e. the number of interactions from cell type to cell type is distinct from that of cell type to ), the total number of interactions was summed across both directions for each cell type pair. Interactions were considered condition specific if they were stronger in a particular condition.

Assessment of consistency between cellular communication prioritization inference and spatial co-localization

To assess the agreement of prioritized cell–cell interactions of scSeqCommDiff with spatial co-localization data, we adopted the same rationale proposed in MultiNicheNet. The framework was applied on single-nucleus RNA-sequencing data of human myocardial infarction [43] between the control (CTRL) and ischemic (IZ) condition (see Supplementary Section S3).

The expected spatial distance of cell type pairs in each experimental condition was defined by calculating the maximum co-localization across samples in both directions of communication. To quantify the condition-specific interaction strength, we first classified differential interactions as control specific when , and vice versa for ischemic-specific interactions. For each condition, we selected the top n ligand–receptor pairs exhibiting the greatest differences in intercellular scores (based on from Equation 3), with n equal to 250, 500, 750, or 1000 regardless of the cell type pair in which they were analyzed. Within each condition-specific top-ranked set, we identified the distinct cell type pairs represented in the selected interactions and, for each cell type pair, computed the number of associated differential interactions.

We plotted the control-specific and ischemic-specific differential interaction strengths for each pair of these cell types against their spatial co-localization in both ischemic and control samples. To quantitatively assess this relationship, we computed the Pearson correlation coefficient. As in MultiNicheNet, we expected (i) a positive correlation between control-specific interactions and co-localization in control samples, as well as between ischemic-specific interactions and co-localization in ischemic samples; and (ii) a negative correlation between control-specific interactions and co-localization in ischemic samples, and between ischemic-specific interactions and co-localization in control samples.

We repeated all the analyses described above considering the top-ranked differential interactions that also exhibited a high level of evidence of differential intracellular signaling. Since our differential intracellular score is pathway specific, we used the maximum observed differential intracellular response as evidence of a downstream cellular response, considering ligand–receptor pairs where maximum > 0.5. Our hypothesis was that the differential intracellular signaling score would strengthen the inference of differential communication.

We included an additional analysis: we identified cell type pairs that exhibited statistically differential co-localization across samples in ischemic compared with control samples. We expect that cells in close spatial proximity are more likely to signal to each other [42, 44]. A one-tailed Wilcoxon Mann–Whitney test P-value < 0.05 was used to distinguish higher co-localization in the ischemic versus control samples. We then assessed whether pairs of cell types inferred to have significant differential interactions exhibit stronger communication signals in the condition in which they were more strongly co-localized and, vice versa, lower communication signals in the condition in which they were less strongly co-localized. To calculate the communication signal, first, ligand–receptor interactions with statistically differential intercellular signaling ( < 0.05 after Benjamini–Hochberg correction) were considered. Second, we conducted a one-tailed Wilcoxon signed-rank test to confirm that intercellular score values in one condition (e.g. ischemic) were higher than their paired values in the other condition (i.e. control). Third, we evaluated whether the maximum differential intracellular score values, , associated with predicted differential intercellular interactions were higher than those associated with non-differential interactions. The rationale behind the latter analysis is based on the expectation that we are more likely to observe a differential cellular response downstream of predicted differential ligand–receptor binding.

Quantitative benchmarking framework for differential cell–cell communication methods

We developed a benchmarking framework for evaluating differential cell–cell interactions tools (see Supplementary Fig. S1), inspired by the metrics established by Liu et al. [32] to benchmark single-condition cell–cell communication methods.

Expected spatial distance. We categorized cell type pairs based on their relative co-localization patterns into two groups: those showing greater co-localization in condition A (denoted as ) and those more co-localized in condition B (denoted as ). To align with the underlying assumptions of cell–cell communication inference tools, we applied two distinct methodologies depending on the scenario.

• Conditions-aware scenario: co-localization for each cell type pair was computed at the condition level by averaging co-localization values across all samples within each experimental condition. Cell type pairs were then ranked based on the difference in co-localization between condition A and condition B.

• Multi-sample conditions-aware scenario: we conducted statistical tests comparing co-localization patterns for each cell type pair across multiple samples between conditions. Cell type pairs were ranked based on statistical significance.

We selected the top 10, 20, 30, and 40% of cell type pairs, with each pair assigned to either or based on its relative strength of co-localization, allowing for a robust evaluation of differential co-localization patterns at varying thresholds.

Observed interaction strength. We quantified condition-specific interaction strength using predictions from differential cell–cell communication inference tools, ensuring consistency by applying the same ligand–receptor database (ConnectomeDB2020 [13]) and using each tool’s default parameters (see Supplementary Section S4). For each differential inference tool, interaction strength for each cell type pair was computed for conditions A (i.e. ) and B (), and sorted in ascending order to obtain the ranked interaction list.

Evaluation metric. The distance enrichment score (DES) metric, adapted from Liu et al., used to evaluate cell–cell communication methods, quantifies the consistency between expected spatial distances (co-localization sets and ) and observed interaction strength (ranked interaction lists and ). Following the enrichment score methodology used in gene set enrichment analysis (GSEA), the DES was computed by systematically walking down the ranked interaction list , increasing the score when a cell type pair was present in the expected co-localization set at different thresholds (top 10, 20, 30, and 40%) and decreasing it when absent. The same procedure was applied for experimental condition B. A higher DES indicates that cell type pairs with greater condition-specific spatial co-localization also exhibit stronger condition-specific differential interactions. All DES computations were performed using the fgsea R package to ensure statistical robustness and reproducibility.

Tested tools and dataset. We benchmarked scSeqCommDiff against widely used differential cell–cell communication inference tools, including CellChat, scDiffCom, MultiNicheNet, and LIANA+. We applied the proposed benchmarking on two datasets (human myocardial infarction by Kuppe et al. [43] and multiple sclerosis by Lerma-Martin et al. [45]), each comprising two experimental conditions (case and control).

Assessment of computational burden

We assessed the computational burden of scSeqCommDiff by measuring the runtime and memory footprint of the tool when applied to real datasets. A large-scale single-nucleus RNA-sequencing dataset [46] (the Pineda et al. dataset), comprising ∼250 000 cells and 15 000 genes from 40 patients (81% of sparsity), and a single-cell lung cancer atlas [47] (LucA) with ∼900 000 cells and 13 000 genes, from a cohort of 318 patients (92% of sparsity), were used as test datasets. The scSeqCommDiff tool was run for all cell types, using the default parameters (e.g. scSeqComm as default intercellular scoring scheme), in both conditions-aware (with default 1000 permutations) and multi-sample conditions-aware scenarios, while varying the number of cores to test scalability. Since the Pineda et al. dataset was derived from human primary motor cortex samples, the communication analysis was performed using a neuronal-specific database (NeuronChatDB [48]), included in our package. In contrast, a general ligand–receptor pair database (i.e. Connectome DB [13]) was used for the analysis of the cancer-related LuCA dataset. The intracellular analysis was conducted using the combined resources of the Kyoto Encyclopedia of Genes and Genomes (KEGG) [49] and Reactome [50] signaling pathway.

Tests were performed on a PC equipped with Intel® Core™ i7-13700K (using up to eight performance cores), 64 GB of RAM, and Debian 12. To ensure reproducibility, R version 4.2.2 and the required dependencies were packed in a Singularity container. Execution time and memory usage were monitored using both OS built-in utilities (i.e. Linux time and free commands) and a dedicated R package (tictoc package).

Interactive and effective visualization tool with CClens

Upon inferring differential cell–cell communication, we implemented an interactive R/Shiny platform, namely CClens, to assist users in the interpretation and exploration of data, structured as an R package available at https://gitlab.com/sysbiobig/cclens.

To ensure portability, reproducibility, and usability of the app, a Docker file is also provided to run the app without the need to install R and package dependencies. Additionally, a user-friendly Windows-specific desktop application is distributed for easy utilization of the app.

The implementation of CClens follows the reactive programming paradigm and involves the use of several R packages, such as shiny, shinyWidgets, shinydashboard, and shinyalert. The app consists of a navbar page, organized into two-tab panels, one dedicated to uploading data (Upload page, see Supplementary Fig. S2) and the other to the customized analysis of cellular communication, tailored for interactive filtering and visualization of CCI results (Analysis page). The latter is organized into four tabs, each enhancing different aspects of the data, encompassing Tabular visualization, Overall Communication, Cell Clusters, and Ligands and Receptors (see Supplementary Figs S3-S6).

Input. Input data consist of a pre-computed dataset of predicted interactions resulting from either one condition or differential cellular communication inference. The platform accepts the results of scSeqCommDiff or other cell–cell communication inference methods, as well as custom analyses performed by the user, provided in csv, tsv, or RData file formats. The pre-computed table must represent predicted CCI in terms of (i) involved signaling molecules (“ligand” and “receptor” columns); (ii) interacting cell types (“sender” and “receiver” columns); and (iii) the intercellular score in the two experimental conditions (“<ScoreName>_<Condition>” columns). Additional information is supported, such as differential intracellular score and intercellular statistical significance. The application also validates and provides detailed feedback if the input format is not supported.

Filtering and customization. The application offers various filtering options for dynamically and interactively inspecting and exploring cell–cell communication data. Users can: (i) select ligands and/or receptors of interest; (ii) select a subset of cell types participating in the communication; (iii) filter results based on a custom intercellular and/or intracellular score range; and (iv) prioritize results based on additional criteria, such as statistical significance from differential intercellular analysis. The user settings automatically subset the data and update downstream statistics and visualizations accordingly.

Exploration and visualizations. Due to the high complexity and dimensionality of cellular communication datasets, we have equipped the interactive app with diverse and detailed graphical representations of data, including chord diagrams, heatmaps, network plots, and dot plots. These visualizations support both quantitative and qualitative inspection of cell–cell communication patterns across distinct conditions, facilitating exploration from a global perspective down to the molecular level. The graphical outputs are generated using custom R functions primarily based on ggplot2, plotly, chorddiag, and visNetwork.

A comprehensive and easily understandable tutorial of the app is available on the vignette of the CClens package.

Results

Agreement of prioritized differential interactions with spatial co-localization data

To assess scSeqCommDiff validity in prioritizing differential cell–cell interactions, we used a comprehensive multimodal dataset integrating both single-cell expression and ST profiling from patients with myocardial infarction and controls [43]. As explained in the Materials and methods, we applied our approach to identify differential cell–cell interactions between ischemic and control populations, and verified the agreement of prioritized differential interactions with spatial co-localization data.

As previous methods have hypothesized [27, 31], the highest ranked differential interactions across cell types are expected to be associated with spatial co-localization in both experimental conditions. Our analysis revealed that, when considering intercellular signaling only, the number of top prioritized differential interactions across different pairs of cell types tends to be positively associated with the spatial proximity for control-specific interactions and their co-localization in control samples, as well as for ischemic-specific interactions and their co-localization in ischemic samples (Fig. 2A, ). Conversely, we observed a negative correlation for control-specific interactions and co-localization in ischemic samples, and for ischemic-specific interactions and co-localization in control samples. This pattern is particularly evident for ischemic-specific interactions across all values of (Supplementary Fig. S7).

Figure 2.

Results of agreement with spatial co-localization data. (A and B) Spatial co-localization analysis of the top prioritized differential cellular communication analysis, considering only intercellular signaling (A) or both intercellular and intracellular signaling (B). The number of prioritized differential ligand–receptor interactions (y-axis) between cell type pairs within the top n= 250 condition-specific ligand–receptor interactions is computed for control and ischemic conditions. Spatial co-localization (x-axis) of each cell type pair is computed in both control and ischemic samples. Each data point represents a distinct cell type pair involved in at least one interaction among the top-ranked differential interactions predicted by scSeqCommDiff (with the number of points indicated in the upper corner of each panel). Pearson correlation between the number of prioritized differential interactions and co-localization is shown, along with blue lines indicating linear model fits and gray bands representing 95% confidence level intervals. (C) Boxplot showing the distribution of intercellular scores () of predicted differential interactions for cell type pairs statistically co-localized in ischaemic samples compared with controls. One-sided P-value from the Wilcoxon signed-rank test is shown. (D) Boxplot representing the distribution of differential intracellular scores () associated with predicted differential and non-differential interaction for cell type pairs statistically co-localized in ischemic samples. The P-value from the one-sided Wilcoxon signed-rank test is shown.

Additionally, Fig. 2B and Supplementary Fig. S8 demonstrate how incorporating evidence of altered intracellular signaling improved our findings. We observed a stronger positive association of condition-specific interactions with spatial proximity of their corresponding samples, as evidenced by higher Pearson correlation coefficients ( and for control- and ischemic-specific interactions, respectively). To evaluate the statistical significance of this improvement, we compare the Pearson correlation values across an increasing number of n (). The analysis revealed a significance increase in correlation (Wilcoxon signed-rank test, P-value < 0.05) in both control and ischemic conditions when differential intercellular signaling was combined with differential intracellular signaling compared with differential intercellular signaling alone (see Supplementary Fig. S9). Conversely, a more pronounced negative correlation was observed between condition-specific interactions and co-localization of the counterpart condition ( and ). This highlights the role of intracellular signaling analysis in strengthening the evidence of observing a differential cell–cell communication across two conditions.

To further validate our approach, we hypothesized that cell type pairs statistically significantly co-localized under one condition are more likely to communicate with each other compared with the other condition, i.e. we would expect that inferred differential interactions have higher intercellular signals in the co-localized samples with respect to the other [42, 44]. Our results confirmed that cell type pairs significantly co-localizing in ischemic samples (i.e. cardiomyocyte–pericyte pairs with one-sided Wilcoxon Mann–Whitney test P-value < 0.05) exhibit statistically higher values (P-value < 2.2e-16) within the ischemic population compared with controls (Fig. 2C). Additionally, the differential intracellular score associated with predicted differential interactions was significantly higher than that of non-differential interactions, as shown in Fig. 2D (P-value < 0.05). These observations suggest that inferred differential communication within the ischemic (co-localized) condition has stronger intercellular intensity compared with the control and presents more evidence of intracellular signaling dysregulation compared with non-differential communication.

These findings demonstrate the ability of scSeqCommDiff to prioritize condition-associated interactions between cell types that are spatially co-localized in their respective condition.

Benchmarking using distance enrichment score

To compare the performance of inferred differential cell–cell interactions across different tools, we used the quantitative benchmarking framework for differential cell–cell communication analysis described in the Materials and methods, which integrates scRNA-seq and ST data. Our underlying hypothesis was that the condition-specific observed interaction strength (i.e. the cardinality of detected interactions for each cell type pair) should align with expected spatial co-localization patterns [27, 31, 42]. We evaluated the performance of widely used differential cell–cell communication inference tools using the DES, originally proposed by Liu et al. [32] for benchmarking single-condition cell–cell communication inference tools. Higher DES values indicate that a given method successfully captures condition-specific CCI patterns that align with spatial co-localization trends.

Specifically, we assessed CellChat, scDiffCom, and scSeqCommDiff in the conditions-aware scenario and MultiNicheNet, LIANA+, and scSeqCommDiff in the multi-sample conditions-aware scenario. Our analysis leveraged two sets of paired and patient-matched ST and scRNA-seq data, ensuring a robust comparison of cell–cell interaction and spatial patterns across different experimental conditions. Figure 3 illustrates the distribution of DES values for both the myocardial infarction (Fig. 3A) and multiple sclerosis (Fig. 3B) datasets. Our results show that scSeqCommDiff consistently achieved high-ranked DES values across both benchmarking scenarios. In the conditions-aware setting, scSeqCommDiff outperformed competing methods, demonstrating its ability to capture condition-specific interaction patterns effectively. In the multi-sample conditions-aware scenario, LIANA+ and scSeqCommDiff performed comparably, both surpassing MultiNicheNet in accuracy.

Figure 3.

Results of benchmarking using DES. The box plot illustrates the DES distributions for various tools applied to the human myocardial infarction dataset (A) and the multiple sclerosis dataset (B). Additionally, tools are categorized based on their methodological approach, distinguishing between conditions-aware (left) and multi-sample conditions-aware (right) scenarios. The tools are ranked by their median DES values.

Notably, our framework is the only one capable of supporting both conditions-aware and multi-sample conditions-aware analyses, providing unmatched flexibility in differential cell–cell communication inference. In cases like this study, where a small sample cohort limits the statistical power of differential cell–cell communication inference, users can prioritize the conditions-aware approach for more robust results. As sequencing costs continue to decrease and larger sample cohorts become more feasible in future studies, the flexibility of our framework to handle multi-sample conditions-aware analyses will become an increasingly valuable resource in bioinformatics.

scSeqCommDiff is efficient and scalable for large-scale scRNA-seq data

One potential concern of performing computational analysis on large-scale scRNA-seq data is the considerable computational burden it imposes. We demonstrated that our implementation has been designed for optimal memory usage and fast computing by applying scSeqCommDiff to two real datasets, i.e. that of Pineda et al. [46] and LuCA atlas [47], with different numbers of cells and genes and different experimental designs (see “Assessment of computational burden”).

As shown in Table 2, scSeqCommDiff optimized memory usage using sparseMatrix and big.matrix object. Notably, this optimization was crucial for large datasets like the LuCA atlas, which could not be represented with a standard R naïve matrix object (i.e. dense representation).

Table 2.

Characteristics of the Pineda et al. and LuCA datasets and RAM memory usage required for loading gene expression matrix across scSeqCommDiff-supported data structures

	Characteristics			RAM memory usage
Dataset	No. of genes	No. of cells	Sparsity (%)	R naïve matrix (GB)	sparseMatrix (GB)	big.matrix (GB)
Pineda et al.	∼15 000	∼250 000	81	28.7	8.1	14.5
LuCA atlas	∼13 000	∼900 000	92	Not possible due to R limitation	14.9	46

We demonstrated that the entire analysis of differential cellular communication (inter- and intracellular) was performed with relatively low execution times and memory footprints using both data structures (see Fig. 4). scSeqCommDiff achieved an impressive performance when analyzing cellular communication in multi-sample conditions-aware settings: the runtime of a complete analysis in serial mode ranged from few seconds to analyze the Pineda et al. dataset with 40 samples and up to few minutes to analyze the ∼900 000 cells in the LuCA atlas across 318 samples (Fig. 4A).

Figure 4.

Evaluation of the computational performance of scSeqCommDiff across two different datasets, i.e. the Pineda et al. dataset and LuCA atlas. (A) Multi-sample conditions-aware scenario. Maximum observed memory overhead and running time for different data structures (big.matrix and sparseMatrix objects) when applying scSeqCommDiff in a multi-sample conditions-aware setting using serial mode. (B) Conditions-aware scenario. Maximum observed memory overhead and running time over an increasing number of cores for different data structures (big.matrix and sparseMatrix objects) when applying scSeqCommDiff in a conditions-aware scenario.

When inferring differential cell–cell communication in the conditions-aware scenario, where the computational burden is intensified by the heightened complexity of the analysis (such as the permutation test), scSeqCommDiff execution time scaled well with the number of cores, while maintaining a low memory footprint (Fig. 4B). Remarkably, this efficiency was achieved on a standard workstation without the need for a high-performance computing infrastructure, as reflected by the maximum footprint observed in the worst-case scenario.

Notably, the use of sparseMatrix or big.matrix object resulted in distinct efficiency advantages in terms of memory usage and running time (Table 2; Fig. 4). The sparseMatrix object minimized memory usage, especially for massive datasets like a cell atlas, but its operations might not fully exploit shared-memory parallelism (even in the context of a forking-system). In contrast, big.matrix objects supported shared-memory parallelism with lower relative memory footprints, but required more RAM for large datasets. Therefore, the choice of the appropriate data structure should consider the dimensionality and inherent characteristics of the dataset (i.e. factors such as sparsity, sample size, and number of cells), to balance RAM usage and parallel computing benefits.

On the other hand, this analysis highlights the efficiency and scalability of our computational approach for processing and manipulating large-scale datasets without the need for specialized and powerful resources.

Inferring differential cellular communication analysis in neurodegenerative disease

To demonstrate the validity and various functionalities of our framework and better appreciate the CClens app, we applied scSeqCommDiff in a multi-sample conditions-aware scenario to a publicly available large-scale single-nucleus RNA-seq dataset (∼250 000 nuclei) of 17 control patients (pathologically normal, PN) and 23 ALS patients [46]. The analysis aimed to identify dysregulated interactions associated with neurodegenerative processes, leveraging a neural-specific database [48], included in our package. To showcase the flexibility of the proposed framework, we used the SingleCellSignalR score for intercellular quantification, and the scSeqComm score for intracellular analysis. The predicted differential interactions were subsequently analyzed through the CClens Shiny app.

After uploading the predicted differential interactions, multiple filtering options allow users to customize downstream analyses by automatically sub-setting data and dynamically updating tabular visualizations. In this study, results were filtered to focus on overcommunicating signals in ALS versus PN, i.e. differential intercellular signaling ( lower than 0.05 after Benjamini–Hochberg correction computed per cell type pair) exhibiting high interaction intensity in ALS but not in PN ( and ). Moreover, we further filtered for those where there was evidence of altered intracellular signaling (maximum > 0.5), as shown in Fig. 5A, which represents a key aspect of our tool.

Figure 5.

Differential cellular communication analysis of ALS versus PN in the multi-sample conditions-aware scenario with CClens. (A) Screenshot of CClens displaying the filtering options. Results are filtered to analyze interactions that are statistically overcommunicating in ALS, with evidence of an altered intracellular response. (B) Interactive chord diagram displaying the number of interactions across cell type pairs. (C) Dot plot showing the top-ranked outgoing interactions (x-axis) across different cell type pairs (y-axis) involving astrocytes as the sender cell type. The intercellular score for each interaction is mapped as dot size, the differential intracellular score as a color fill, and the experimental condition is encoded by the colored stroke. (D) Dot plot displaying interactions involving GAD enzymes and GABA receptors (x-axis), astrocytes as the sender cell type, and microglia as the receiver cell type (y-axis). The values of intercellular score for each interaction are mapped as dot size, the differential intracellular score as colored fill, and the experimental condition is encoded by the colored stroke.

For initial high-level data exploration, CClens summarizes differential cellular communication through the cardinality of ligand–receptor interactions mediating communication between cells. This summary facilitates the identification of the cell types playing a prominent role as sender or receiver (or both) cell types. Specifically, we observed that astrocytes and microglia play significant roles in ALS, with astrocytes exhibiting the highest amount of dysregulated outgoing communication and microglial cells showing the highest amount of altered incoming communication (see Fig. 5B). The findings were aligned with a previous study on their pivotal role in ALS pathogenesis [51]. Specifically, astrocytes are fundamental in maintaining the homeostasis of extrasynaptic glutamate; impaired glutamate uptake in astrocytes has been proposed as a mechanism for motor neuron death in ALS through glutamate-mediated excitotoxicity [52].

CClens facilitates in-depth analysis of interactions characterizing specific cell types, allowing inspection of outgoing and incoming interactions. Thus, we focused on outgoing interactions involving astrocytes as the sender cell type. The top ligand–receptor interactions with the highest differences in intercellular scores between conditions are visually represented in dot plots (see Fig. 5C). Interestingly, these communications are primarily mediated by the glutamic acid decarboxylases (encoded by Gad1 and Gad2), which synthesize the inhibitory neurotransmitter gamma-aminobutyric acid (GABA) from glutamate. Notably, the role of these enzymes and of GABA metabolism in ALS and other neurodegenerative diseases has been studied extensively, but still remains partly unclear [53, 54]. There is increasing evidence of a disproportion between excitatory and inhibitory neurotransmitter levels in the brain [55]. Moreover, oxidative stress is also known to be crucial in neurodegenerative and neuromuscular diseases. One of the most sensitive and reliable markers of cellular oxidative stress is the antioxidant HMOX enzyme, that converts cellular heme into biliverdin, carbon monoxide (CO), and free iron [56]. Interestingly, isoform 2 of HMOX with the GUCY1A2 receptors was found to be overcommunicating in several cell types in ALS, including astrocytes and fibroblasts (Fig. 5C), suggesting high levels of oxidative stress, consistent with the literature [57].

A deeper investigation of the interactions can be performed at the molecular level. Users can visually inspect a subset of ligand–receptor interactions mediating communication between a user-specified sender–receiver pair or a single ligand–receptor across all pairs of cell types. To further investigate their significant role, we focused on differential communication occurring from astrocytes to microglia. Microglia were found to be receiving GABAergic communication, through several GABA receptors, namely GABBR1, GABBR2, GABRA2, and GABRA4 (see Fig. 5D). The relevance of this finding derives from the recently discovered role of this cell type: GABA-receptive microglia cells are actively involved in inhibitory remodeling of synapses [58].

Leveraging the key features of our tool, which provide insights into intracellular signaling alterations, we functionally characterized the differential cellular response by analyzing the target genes identified as DE downstream of these ligand–receptor pairs. The GO enrichment analysis (see Supplementary Table S2) on these target genes demonstrated biological processes related to response to oxidative stress (GO:0006979) and response to hydroperoxide (GO:0033194) [57]. Additionally, key processes identified include amyloid mechanism (GO:1902004, GO:1902993, GO:1902003, GO:1902991, GO:0034205, GO:0050435, and GO:0042987), which have been proposed as ALS signatures in the most recent studies [59, 60].

Next, we ran the analysis in the conditions-aware setting and analyzed predicted differential communication using CClens (see Supplementary Fig. S10). The prominent role of microglia as the most significantly altered receiver cell type has been confirmed (Supplementary Fig. S10B). Consistent with the multi-sample conditions-aware analysis, Gad1 and Gad2 were found to be differentially overcommunicating in ALS with respect to PN (see Supplementary Fig. S10C). Additionally, another enzyme involved in the same metabolic pathway was found to be differentially communicating in ALS versus PN, i.e. the glutaminase encoded by the Gls gene. This enzyme, which synthesizes the excitatory neurotransmitter glutamate from glutamine, exhibited significantly differential communication. Specifically, it was found to significantly overcommunicate in ALS using the GRIN2D receptor between neurons inhibitory/excitatory and oligodendrocytes (see Supplementary Fig. S10C). These findings were consistent with the current understanding of the role of Gls in ALS and neuroinflammation processes, as outlined in the literature [61], further supporting the biological validity of scSeqCommDiff results. Moreover, the vesicular transporter of glutamate, SLC17A6, also known as VGLUT2, whose activity is altered in ALS, was found to be overcommunicating in the conditions-aware scenario results, further suggesting alterations of glutamatergic activity (see Supplementary Fig. S10D) [62]. In more detail, SLC17A6, GAD1, and GAD2 resulted in altered communication from microglia to oligodendrocyte progenitor cells, suggesting a significant involvement of these cell types in the dysregulation of the glutamatergic and GABAergic activities in the ALS condition.

Discussion

In this work, we introduced a novel computational framework specifically designated for the analysis and interpretation of differential cell–cell communication analysis within complex and large-scale scRNA-seq data. The framework incorporates two tools: scSeqCommDiff, which identifies and characterizes alterations in cell–cell communication in a fast and memory-efficient manner, and CClens, which facilitates interpretation and exploration through an interactive R/Shiny interface.

Our framework offers several distinct advantages over current cell–cell communication tools for inferring differential cell–cell communication.

First, the framework is tailored to work with different experimental settings, accommodating both conditions-aware (aiming to identify population-level differences) and multi-sample conditions-aware (capturing within-condition heterogeneity) scenarios. This versatility enables the identification of dissimilarities across a wide spectrum of experimental designs. Furthermore, applying scSeqCommDiff in both conditions-aware and multi-sample conditions-aware scenarios to a neurodegenerative disease-related dataset demonstrated the agreement and consistency of inferred results with the disease literature. While both scenarios provided consistent findings, emphasizing the robustness of our results, each also revealed different aspects of altered communication, highlighting the complementary nature of these approaches. However, in datasets with limited sample sizes, the statistical power in the multi-sample conditions-aware scenario may be compromised, leading to fewer predicted differential interactions. In these cases, we recommend leveraging the statistical approach of conditions-aware analysis or adapting the multiple-testing correction.

Second, the generalizability and flexibility of the tool are ensured by its ability to use both build-in and user-specified apriori cell–cell communication resources (i.e. ligand–receptor databases and signaling pathway databases), as well as by the integration of several intercellular scoring schemes. The use of non-parametric statistical tests, such as the Wilcoxon test and permutation test, for assessing differential signaling avoids the need for any assumptions about the type and distributions of intercellular scores. This allows users to choose among several built-in intercellular scoring schemes. However, it is important to note that the choice of resource (e.g. ligand–receptor database) and scoring scheme may affect the results, as highlighted by recent reviews [31].

Third, a distinctive feature is the implementation of a computational tool to infer and quantify differential communication at both the intercellular and intracellular levels. Remarkably, the ability to quantitatively measure changes in downstream cellular responses helps in prioritizing the most reliable differential intercellular interactions. Indeed, the integration of differential intracellular evidence strengthens the reliability of predicted results, as demonstrated by our validation though ST data (see the Results). This contribution, coupled with the functional characterization of these alterations through GO enrichment analysis, facilitates a comprehensive understanding of the underlying transcriptomic alterations and their functional implications on the biological system, an aspect that is not currently addressed by existing tools.

As the throughput and accessibility of scRNA-seq workflows continue to increase, it becomes crucial that single-cell computational methods are scalable and capable of handling large-scale datasets. Indeed, the strategic use of efficient data structures in scSeqCommDiff, with the development of custom C++ code, ensures efficient memory utilization, as well as enabling effective parallelism to improve speed and reduce the computational demands of cellular communication analysis. Our test showed that scSeqCommDiff can successfully perform complete intercellular and intracellular analyses of large datasets, including cell atlases, on a standard workstation (i.e. using < 64 GB of RAM) (see the Results). However, we showed that different data structures yield distinct efficiency advantages in terms of memory overhead and running time, depending on dataset characteristics (such as sample size and sparsity). These factors should be considered when selecting the appropriate data structure.

Finally, the user-friendly and effective interface of CClens facilitates easy access, exploration, and interpretation of differential cell–cell communication data, whether resulting from scSeqCommDiff analysis or custom analysis performed by the user. Through flexible data inspection and rich visualizations, we showcased the ability of CClens to assist in the exploration of altered interactions among ALS and control patients, thereby unveiling valuable biological insights. Notably, since CClens is the only available interactive web app that allows filtering and visualizing of differential intercellular and intracellular signaling data, a direct comparison with existing interactive tools is currently not possible.

Like other existing tools, our method may suffer from inherent limitations of scRNA-seq data, such as sparsity and noise, and the use of gene expression as a proxy of protein expression. Another common constraint is the lack of a systematic and comprehensive assessment of cell–cell communication predictions due to the absence of a complete ground truth, as highlighted by several studies [11, 63]. In this work, we demonstrate the validity of scSeqCommDiff using two typical approaches in validating cell–cell communication tools: (i) applying it to real-case datasets, where the predictions can be compared with known biological knowledge and literature, and (ii) using data with additional information, i.e. cell co-localization from ST data, to be exploited during results analysis (see the Results). In both cases, the results supported the validity of the proposed method, suggesting its robustness and applicability in the analysis of scRNA-seq data.

Furthermore, as for single-condition cell–cell communication inference, comparing results across different differential cell–cell communication methods is non-trivial and often not feasible, since each method evaluates different aspects of cellular communication (e.g. presence of signaling events, signaling specificity, or signaling intensity) and focuses on distinct mechanisms (such as intercellular communication alone or both inter- and intracellular interactions), as reported by the recent literature [12]. These challenges are further amplified when benchmarking differential cell–cell communication tools. First, there is a limited availability of orthogonal or complementary datasets for scRNA-seq, particularly in large cohort studies with multiple experimental conditions. Second, benchmarking differential cell–cell communication analysis remains an underexplored area: existing benchmarking studies and frameworks [31–33] only focus on single-condition cell–cell communication tools, and no prior method or review has systematically addressed this challenge. To the best of our knowledge, none of the published differential cell–cell communication methods has been quantitatively assessed and compared with other differential tools. Building on the benchmarking framework proposed in Liu et al., we develop the first quantitative framework for assessing differential cell–cell communication tools. While we acknowledge that the current validation is limited to a small number of datasets, this framework represents a critical first step toward systematic and standardized benchmarking in this emerging area. Despite this only being a first attempt, and not the main focus of our work, the proposed framework allows for the first quantitative benchmarking of some of the most used differential cell–cell communication tools, including CellChat, scDiffCom, MultiNicheNet, and LIANA+, and compares their performance with that of scSeqCommDiff. Our results demonstrate that our tool achieved consistently high-ranking scores, outperforming the majority of differential inference tools in both condition-aware and multi-sample condition-aware scenarios. However, the development of more advanced benchmarking frameworks for differential cell–cell communication tools is far from being completed. Future efforts will need to focus on establishing more comprehensive and standardized benchmarking strategies, ideally incorporating a broader range of biological systems, experimental designs, and complementary data modalities, such as spatial transcriptomics and proteomics.

The benchmarking results show that intracellular-aware tools do not consistently outperform intercellular-only tools, which may seem counterintuitive. This probably reflects the inherent complexity of intracellular signaling and the fact that different modeling choices, such as focusing on direct protein–protein interactions versus signaling pathways, can lead to substantially different outcomes. Incorporating intracellular signaling adds a layer of complexity due to the highly intricate and interconnected nature of signaling pathways. Making effective use of this information requires not only high-quality curated databases, but also algorithmic approaches that can manage this complexity without introducing excessive noise. While such integration can enhance CCI inference by reducing false positives and providing deeper functional insights, it may also compromise performance if not carefully implemented [12, 10, 64]. This challenge is highlighted in the ESICCC benchmark [33], which identifies the inclusion of intracellular signaling as a critical factor influencing tool performance in single-condition cell–cell communication analysis, though not always beneficial. Notably, the intracellular score used in scSeqComm, which underlies the intracellular component of our scSeqCommDiff framework, achieved strong performance in ESICCC benchmark, even when other intracellular-aware tools performed poorly. This suggests that the integration of intracellular signaling can be beneficial when appropriately modeled, as in scSeqComm and scSeqCommDiff. Accordingly, our results show that incorporating intracellular signaling in scSeqCommDiff improves both predictive accuracy and the biological interpretability of differential cell–cell communication analysis.

One major constraint of our framework is its current limitation to pairwise differential communication comparisons: future developments will extend the framework to handle more than two conditions. Recent advances in spatially resolved transcriptomic techniques offer a valuable opportunity to explore the spatial organizations of cells. Future developments will include integrating ST data into our framework to account for spatial context in cell–cell interactions.

Despite these limitations, we believe our framework is a powerful and versatile resource for generating hypotheses and unraveling the cell–cell communication mechanisms that govern diverse experimental conditions.

Supplementary data

Supplementary data are available at NAR Genomics & Bioinformatics online.

Conflict of interest

None declared.

Funding

The MUR PNRR “National Center for HPC, BIG DATA AND QUANTUM COMPUTING” [CN00000013 CN1, spoke 8]; the “National Center for Gene Therapy and Drugs Based on RNA Technology”, in the framework of the National Recovery and Resilience Plan [PNRR-CN3, Mission 4, Component 2, Investment 1.4, Project CN00000041, CUP B93D21010860004, CN3, Spoke n.2, Cancer]; the ‘BRinging Artificial INTelligencE home for a better cAre of amyotrophic lateral sclerosis and multiple SclERosis’—‘BRAINTEASER’ project, funded from the European Union's Horizon 2020 research and innovation programme [grant agreement No 101017598]. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the EU. Neither the EU nor the granting authority can be held responsible for them. G.C. was funded by the Department of Information Engineering, University of Padova, ‘Research Grant type B junior'.

Data availability

No new data were generated for this study. All data used in this study are publicly available. The single-nucleus RNA-seq dataset from the study of Pineda et al. is available from the NCBI Gene Expression Omnibus (GEO) repository under accession number GSE174332. Pre-processed single-nucleus RNA-seq data from the study of Kuppe et al. are available at the cellxgene platform (https://cellxgene.cziscience.com/collections/8191c283-0816-424b-9b61-c3e1d6258a77). The (core atlas) single-cell Lung Cancer Atlas (LuCA) from Salcher et al. is available at the cellxgene platform (https://cellxgene.cziscience.com/collections/edb893ee-4066-4128-9aec-5eb2b03f8287). Processed snRNA-seq and ST data from the multiple sclerosis study of Lerma-Martin et al. are available at the UCSC cell browser: https://cells-test.gi.ucsc.edu/?ds=ms-subcortical-lesions. scSeqCommDiff and CClens tools are freely available at https://gitlab.com/sysbiobig/scseqcomm and https://gitlab.com/sysbiobig/cclens. The submitted software version is available in Zenodo at https://doi.org/10.5281/zenodo.12790607.