multinichenetr: the R package for differential cell-cell communication analysis from single-cell transcriptomics data with complex multi-sample, multi-condition designs. The goal of this toolbox is to study differences in intercellular communication between groups of samples of interest (eg patients of different disease states).
You can read all about MultiNicheNet in the following preprint: https://www.biorxiv.org/content/10.1101/2023.06.13.544751v1
The main goal of the MultiNicheNet package is to find which ligand-receptor interactions are differentially expressed and differentially active between conditions of interest, such as patient groups. Compared to the normal NicheNet workflow, MultiNicheNet 1) considers more information to prioritize interactions (differential and cell-type specific expression in addition to ligand activity); 2) is more suited to tackle complex experimental designs such as those with multiple samples and conditions, and multiple receiver cell types of interest; and 3) enables incorporation of complementary data types for prioritization of cell-cell interactions (e.g., proteomics).
In the MultiNicheNet approach, we allow the user to prioritize differential cell-cell communication events (ligand-receptor interactions and downstream signaling to target genes) based on the following criteria:
- Upregulation of the ligand in a sender cell type and/or upregulation of the receptor in a receiver cell type - in the condition of interest.
- Sufficiently high expression levels of ligand and receptor in many samples of the same group (to mitigate the influence of outlier samples).
- Cell-type and condition specific expression of the ligand in the sender cell type and receptor in the receiver cell type (to mitigate the influence of upregulated but still relatively weakly expressed ligands/receptors)
- High NicheNet ligand activity, to further prioritize ligand-receptor pairs based on the enrichment of their predicted target genes in the receiver cell type
MultiNicheNet combines all these criteria in a single prioritization score, which is also comparable between all sender-receiver pairs. This way, MultiNicheNet extends on the prioritization done by NicheNet, which is only based on the ligand activity score. Importantly, this prioritization framework is flexible and extendable: it is possible to design an additional criterion based on complementary data such as serum proteomics.
To summarize the main functionalities of multinichenetr:
- Prioritizing the most differential ligand-receptor interactions from different sender-receiver pairs between different sample groups, according to criteria such as condition specificity, cell-type specificity, ligand activity (= downstream signaling activity / target gene enrichement), and more.
- Predicting the specific downstream affected target genes of prioritized differential ligand-receptor links of interest.
- Predicting intercellular gene regulatory networks, connecting ligands to ligand- or receptor-encoding target genes in other cell types, enabling predictions concerning intercellular cascade and feedback mechanisms.
Note: at the basis of MultiNicheNet for defining differentially expressed ligands, receptors and target genes, is the the differential state analysis as discussed by muscat, which provides a framework for cell-level mixed models or methods based on aggregated “pseudobulk” data (https://doi.org/10.1038/s41467-020-19894-4, https://bioconductor.org/packages/release/bioc/html/muscat.html). We use this muscat framework to make inferences on the sample-level (as wanted in a multi-sample, multi-condition setting) and not the classic cell-level differential expression analysis of Seurat (Seurat::FindMarkers), because muscat allows us to overcome some of the limitations of cell-level analyses for differential state analyses. Some of these limitations include: a bias towards samples with more cells of cell type, a lack of flexibility to work with complex study designs, and a too optimistic estimation of the statistical power since the analysis is done at the cell-level and not at the sample level.
In the future, we might extend the differential expression analyses options to include other frameworks than muscat.
Since May 13 2024, we have put out multinichenetr 2.0.0. on the main branch. Please install this latest version of multinichenetr, which has been improved at several levels compared to the first version:
Here is a concise list of methodological novelties compared to the previous version. See the tutorials for more in-depth explanation on each of these aspects. These novelties will also be discussed in the revised version of the manuscript.
- It is now possible to add new prioritization criteria based on additional data modalities. This option can helps users prioritize interactions better if they have additional complementary data, such as proteomics data. This option can help overcome major limitations of cell-cell communication inference from transcriptomics data.
- An additional alternative workflow has been added to prioritize interactions involving condition-specific cell types. This option was added to not ignore biological signal from potentially relevant condition-specifc cell types.
- More emphasis on the usability of the “Intercellular regulatory network” to prune prioritized interactions.
- Integration with the Omnipath L-R database to check database source, quality and curation effort behind prioritized ligand-receptor interactions.
Here is a concise list of software changes compared to the previous version. See the tutorials for more in-depth explanation on each of these aspects.
- Gene filtering has been modified slightly, resulting in fewer and easier to interpret parameters and stronger consistency with the rest of the MultiNicheNet methodology
- Prioritization can now be done by only using the up-regulatory ligand activities instead of both up- and downregulatory activities. This option was added because of the difficulty of interpreting downregulatory activities.
- The option to flexibly change prioritization weights has been replaced by the option to select biological scenario’s. This reduces the number of parameters for the end users and limits unwanted tunability of end results.
- The standard interpretable bubble plot visualization has been extended and provides now information about cell-type specificity, fraction of expression, and curation effort of the ligand-receptor pairs according to Omnipath. As a result of this, this plot now summarizes all the criteria used for prioritization and gives an indication about the curation effort of the ligand-receptor prior knowledge. This can help users now to get better insights in their results and define better candidate interactions for follow-up validation.
multinichenetr is in ongoing development. We always appreciate it if you let us know how we can improve the software/documentation/algorithm/output visualizations further. The best way to do this is through the issues page: https://github.com/saeyslab/multinichenetr/issues
Installation typically takes a few minutes, depending on the number of dependencies that has already been installed on your pc.
You can install multinichenetr (and required dependencies) from github with:
# install.packages("devtools")
devtools::install_github("saeyslab/nichenetr")
devtools::install_github("saeyslab/multinichenetr")
It is possible that during installation the following warning is thrown:
“glmmTMB was built with TMB version 1.9.4” “Current TMB version is 1.9.5
This warning can be safely ignored since this does not affect multinichenetr’s installation and functionalities.
multinichenetr is tested via Github Actions version control on Windows, Linux (Ubuntu) and Mac (most recently tested R version: R 4.3.1.).
We provide several vignettes demonstrating the different types of analysis that can be performed with MultiNicheNet, and the several types of downstream visualizations that can be created.
We recommend users to start with the following vignette, which demonstrates the different steps in the analysis without too many details yet. This is the recommended vignette to learn the basics of MultiNicheNet.
- MultiNicheNet - comprehensive tutorial - Condition A vs Condition B vs Condition C | R Markdown version | HTML version
That vignette provides an example of a comparison between 3 groups. The following vignettes demonstrate how to analyze cell-cell communication differences in other settings. These vignettes are the best vignettes to learn how to apply MultiNicheNet to different datastes for addressing different questions. To reduce the length of these vignettes, the sections on downstream analysis has been reduced strongly. So we strongly recommend to read these vignettes to learn how to perform the analysis in other settings, but still perform all additional analyses and checks as demonstrated in the comprehensive tutorial vignette .
- Condition A vs Condition B - without repeated subjects | R Markdown version | HTML version
- Condition A vs Condition B - with repeated subjects: paired analysis with subject_id as covariate | R Markdown version | HTML version
- Condition A vs Condition B - from integrated atlas data: correcting for batch effects | R Markdown version | HTML version
- Multifactorial analysis: condition-differences in cell-cell communication dynamics: (Time2_ConditionA - Time1_ConditionA) vs (Time2_ConditionB - Time1_ConditionB) | R Markdown version | HTML version
In addition to these vignettes, we also provide 2 other vignettes showcasing the flexibility and extendibility of the MultiNicheNet framework: how to include additional complementary data modalities and assess condition-specific cell types:
- Incorporating additional prioritization criteria based on complementary data types: serum proteomics | R Markdown version | HTML version
- Altnernative workflow to assess interactions involving condition-specific cell types | R Markdown version | HTML version
When applying MultiNicheNet on datasets with many samples and cell types, it is often needed to run the analysis on HPC infrastructure. In those cases, we recommend to first set up your pipeline locally on a subset of the data (eg subset of 2/3 cell types). Then we recommend to run the core analysis on HPC infrastructure and save the output, and finally interpret this output locally. In the following scripts you can see an example of how we split up the analysis in two parts: 1) running MultiNicheNet (with qsub on gridengine cluster) and saving necessary output and plots; and 2) interpreting the results and generating visualizations. These scripts are illustrative and will not work directly when you would just run them.
- MultiNicheNet core analysis on gridengine cluster via HPC
- Interpreting MultiNicheNet results locally
To help users in interpreting parameter values and output figures, we provide the following two files:
- Parameter interpretation: provides an explanation of different parameter choices - can help users in deciding when the default parameter values would not be optimal for their own dataset
- Output figure interpretation: provides an explanation of different output figures - can help users in drawing hypotheses based on the output figures
- The input data needed for MultiNicheNet should be raw counts, and metadata of cells giving information about the sample, condition and cell type. In all vignettes, we assume that the data has been preprocessed adequately (proper cell filtering, doublet removal, ambient RNA correction,…).
- We strongly recommend having at least 4 samples in each of the groups/conditions you want to compare. With less samples, the benefits of performing a pseudobulk-based DE analysis are less clear and non-multi-sample tools for differential cell-cell communication might be better alternatives. If you want to perform differential cell-cell communication with a MultiNicheNet-like prioritization framework, you can have a look at this vignette: Differential cell-cell Communication for datasets with limited samples: “sample-agnostic/cell-level” MultiNicheNet. Just realize that the analysis is based on a limited number of samples, and it will be hard to draw strong conclusions. This may often be the best you can get out of your data, but it is not a practice we would recommend.
Browaeys, R. et al. MultiNicheNet: a flexible framework for differential cell-cell communication analysis from multi-sample multi-condition single-cell transcriptomics data. (preprint). https://www.biorxiv.org/content/10.1101/2023.06.13.544751v1
Browaeys, R., Saelens, W. & Saeys, Y. NicheNet: modeling intercellular communication by linking ligands to target genes. Nat Methods (2019) https://doi:10.1038/s41592-019-0667-5
Crowell, H.L., Soneson, C., Germain, PL. et al. muscat detects subpopulation-specific state transitions from multi-sample multi-condition single-cell transcriptomics data. Nat Commun 11, 6077 (2020). https://doi.org/10.1038/s41467-020-19894-4