title: "MultiNicheNet analysis: SCC paired analysis - wrapper function"

author: "Robin Browaeys"

package: "`r BiocStyle::pkg_ver('multinichenetr')`"

output:

BiocStyle::html_document

output_dir: "/Users/robinb/Work/multinichenetr/vignettes"

vignette: >

<style type="text/css">

.smaller {

font-size: 10px

}

</style>

```{r setup, include = FALSE}

In this vignette, you can learn how to perform a MultiNicheNet analysis to compare cell-cell communication between two conditions of interest (one-vs-one comparison) in a __paired design__ setting. A MultiNicheNet analysis can be performed if you have multi-sample, multi-condition/group single-cell data. 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. For those datasets, you can check and run our alternative workflow that makes use of cell-level sample-agnostic differential expression tools.

As input you need a SingleCellExperiment object containing at least the raw count matrix and metadata providing the following information for each cell: the **group**, **sample** and **cell type**.

As example expression data of interacting cells, we will here use scRNAseq data of patients with squamous cell carcinoma from this paper of Ji et al.: [Multimodal Analysis of Composition and Spatial Architecture in Human Squamous Cell Carcinoma](https://www.sciencedirect.com/science/article/pii/S0092867420306723). We have data of **normal and tumor tissue for each patient**, and thus a **paired design**.

We will use MultiNicheNet to explore tumor microenvironment interactions that are different between tumor and normal tissue. In this vignette, we will prepare the data and analysis parameters, and then perform the MultiNicheNet analysis. In contrast to a classic pairwise analysis between conditions/groups, we will here demonstrate how you can include in the DE model that tumor and healthy tissue come from the same patient.

We will first prepare the MultiNicheNet core analysis, then run the several steps in the MultiNicheNet core analysis, and finally interpret the output.

```{r load-libs, message = FALSE, warning = FALSE}

MultiNicheNet builds upon the NicheNet framework and uses the same prior knowledge networks (ligand-receptor network and ligand-target matrix, currently v2 version).

The Nichenet v2 networks and matrices for both mouse and human can be downloaded from Zenodo [](https://doi.org/10.5281/zenodo.7074291).

We will read these object in for human because our expression data is of human patients.

Gene names are here made syntactically valid via `make.names()` to avoid the loss of genes (eg H2-M3) in downstream visualizations.

```{r, results='hide'}

In this vignette, we will load in a subset of the scRNAseq data of the Ji el al Squamous Cell Carcinoma data. For the sake of demonstration, this subset only contains 4 cell types.

If you start from a Seurat object, you can convert it easily to a SingleCellExperiment object via `sce = Seurat::as.SingleCellExperiment(seurat_obj, assay = "RNA")`.

Because the NicheNet 2.0. networks are in the most recent version of the official gene symbols, we will make sure that the gene symbols used in the expression data are also updated (= converted from their "aliases" to official gene symbols). Afterwards, we will make them again syntactically valid.

```{r, results='hide'}

In this step, we will formalize our research question into MultiNicheNet input arguments.

In this case study, we want to study differences in cell-cell communication patterns between cells in tumor tissue and healthy tissue. The meta data columns that indicate this tissue status is `tum.norm` (values: Tumor and Normal).

Cell type annotations are indicated in the `celltype_alt` column, and the sample is indicated by the `sample_id` column.

If your cells are annotated in multiple hierarchical levels, we recommend using a relatively high level in the hierarchy. This for 2 reasons: 1) MultiNicheNet focuses on differential expression and not differential abundance, and 2) there should be sufficient cells per sample-celltype combination (see later).

__Important__: It is required that each sample-id is uniquely assigned to only one condition/group of interest. See the vignettes about paired and multifactorial analysis to see how to define your analysis input when you have multiple samples (and conditions) per patient.

If you would have batch effects or covariates you can correct for, you can define this here as well. For this dataset, we can set the patient ID (given by the `patient` column) as covariate. Important: for a MultiNicheNet analysis there is a difference between a covariate and batch in the following sense: covariates will just be included in the DE GLM model, whereas batches will be included in the DE GLM model AND normalized pseudobulk expression values will be corrected for the batch effects. In this dataset, we want to take into account the patient effect, but not correct the expression values for the patient effect. Therefore we add patient as covariate and not as batch.

__Important__: for categorical covariates and batches, there should be at least one sample for every group-batch combination. If one of your groups/conditions lacks a certain level of your batch, you won't be able to correct for the batch effect because the model is then not able to distinguish batch from group/condition effects.

__Important__: The column names of group, sample, cell type, batches and covariates should be syntactically valid (`make.names`)

__Important__: All group, sample, cell type, batch and covariate names should be syntactically valid as well (`make.names`) (eg through `SummarizedExperiment::colData(sce)$ShortID = SummarizedExperiment::colData(sce)$ShortID %>% make.names()`)

For this analysis, we want to compare Tumor tissue and normal tissue. To do this comparison, we need to set the following contrasts:

__Very Important__ Note the format to indicate the contrasts! This formatting should be adhered to very strictly, and white spaces are not allowed! Check `?get_DE_info` for explanation about how to define this well. The most important points are that:

*each contrast is surrounded by single quotation marks

*contrasts are separated by a comma without any white space

*all contrasts together are surrounded by double quotation marks.

If you compare against two groups, you should divide by 2 (as demonstrated in other vignettes), if you compare against three groups, you should divide by 3 and so on.

For downstream visualizations and linking contrasts to their main condition, we also need to run the following:

This is necessary because we will also calculate cell-type+condition specificity of ligands and receptors.

Other vignettes will demonstrate how to formalize different types of research questions.

If you want to focus the analysis on specific cell types (e.g. because you know which cell types reside in the same microenvironments based on spatial data), you can define this here. If you have sufficient computational resources and no specific idea of cell-type colocalzations, we recommend to consider all cell types as potential senders and receivers. Later on during analysis of the output it is still possible to zoom in on the cell types that interest you most, but your analysis is not biased to them.

Here we will consider all cell types in the data:

In case you would have samples in your data that do not belong to one of the groups/conditions of interest, we recommend removing them and only keeping conditions of interest:

Now we will run the core of a MultiNicheNet analysis. This analysis consists of the following steps:

* 1. Cell-type filtering: determine which cell types are sufficiently present

* 2. Gene filtering: determine which genes are sufficiently expressed in each present cell type

* 3. Pseudobulk expression calculation: determine and normalize per-sample pseudobulk expression levels for each expressed gene in each present cell type

* 4. Differential expression (DE) analysis: determine which genes are differentially expressed

* 5. Ligand activity prediction: use the DE analysis output to predict the activity of ligands in receiver cell types and infer their potential target genes

* 6. Prioritization: rank cell-cell communication patterns through multi-criteria prioritization

Following these steps, one can optionally

* 7. Calculate the across-samples expression correlation between ligand-receptor pairs and target genes

* 8. Prioritize communication patterns involving condition-specific cell types through an alternative prioritization scheme

After these steps, the output can be further explored as we will demonstrate in the "Downstream analysis of the MultiNicheNet output" section.

In this vignette, we will demonstrate these steps one-by-one, which offers the most flexibility to the user to assess intermediary results. Other vignettes will demonstrate the use of the `multi_nichenet_analysis` wrapper function.

In this step we will calculate and visualize cell type abundances. This will give an indication about which cell types will be retained in the analysis, and which cell types will be filtered out.

Since MultiNicheNet will infer group differences at the sample level for each cell type (currently via Muscat - pseudobulking + EdgeR), we need to have sufficient cells per sample of a cell type, and this for all groups. In the following analysis we will set this minimum number of cells per cell type per sample at 10. Samples that have less than `min_cells` cells will be excluded from the analysis for that specific cell type.

We recommend using `min_cells = 10`, except for datasets with several lowly abundant cell types of interest. For those datasets, we recommend using `min_cells = 5`.

First, we will check the cell type abundance diagnostic plots.

The first plot visualizes the number of cells per celltype-sample combination, and indicates which combinations are removed during the DE analysis because there are less than `min_cells` in the celltype-sample combination.

The red dotted line indicates the required minimum of cells as defined above in `min_cells`. We can see here that some sample-celltype combinations are left out. For the DE analysis in the next step, only cell types will be considered if there are at least two samples per group with a sufficient number of cells. But as we can see here: all cell types will be considered for the analysis and there are no condition-specific cell types.

__Important__: Based on the cell type abundance diagnostics, we recommend users to change their analysis settings if required (eg changing cell type annotation level, batches, ...), before proceeding with the rest of the analysis. If too many celltype-sample combinations don't pass this threshold, we recommend to define your cell types in a more general way (use one level higher of the cell type ontology hierarchy) (eg TH17 CD4T cells --> CD4T cells) or use `min_cells = 5` if this would not be possible.

You can always explore this plot for a more lenient or stringent setting of `min_cells` in case of doubt. In this case study, it may be useful to be more lenient. Why? Because we added the Patient ID as covariate, we need to have sufficient cells of one patient for BOTH tumor and normal tissue to include that patient in the DE analysis. Since we seem to have many "borderline" cases here and could up with a low nr of included patients, we will drop our stringency level.

Running the following block of code can help you determine which cell types are condition-specific and which cell types are absent.

Absent cell types will be filtered out, condition-specific cell types can be filtered out if you as a user do not want to run the alternative workflow for condition-specific cell types in the optional step 8 of the core MultiNicheNet analysis.

For this dataset, there are no condition-specific or absent cell types, so this does not really matter.

Before running the DE analysis, we will determine which genes are not sufficiently expressed and should be filtered out.

We will perform gene filtering based on a similar procedure as used in `edgeR::filterByExpr`. However, we adapted this procedure to be more interpretable for single-cell datasets.

For each cell type, we will consider genes expressed if they are expressed in at least a `min_sample_prop` fraction of samples in the condition with the lowest number of samples. By default, we set `min_sample_prop = 0.50`, which means that genes should be expressed in at least 2 samples if the group with lowest nr. of samples has 4 samples like this dataset.

But how do we define which genes are expressed in a sample? For this we will consider genes as expressed if they have non-zero expression values in a `fraction_cutoff` fraction of cells of that cell type in that sample. By default, we set `fraction_cutoff = 0.05`, which means that genes should show non-zero expression values in at least 5% of cells in a sample.

We recommend using these default values unless there is specific interest in prioritizing (very) weakly expressed interactions. In that case, you could lower the value of `fraction_cutoff`. We explicitly recommend against using `fraction_cutoff > 0.10`.

Now we will calculate the information required for gene filtering with the following command:

Now only keep genes that are expressed by at least one cell type:

After filtering out absent cell types and genes, we will continue the analysis by calculating the different prioritization criteria that we will use to prioritize cell-cell communication patterns.

First, we will determine and normalize per-sample pseudobulk expression levels for each expressed gene in each present cell type. The function `process_abundance_expression_info` will link this expression information for ligands of the sender cell types to the corresponding receptors of the receiver cell types. This will later on allow us to define the cell-type specicificy criteria for ligands and receptors.

Normalized pseudobulk expression values per gene/celltype/sample can be inspected by:

An average of these sample-level expression values per condition/group can be inspected by:

Inspecting these values for ligand-receptor interactions can be done by:

In this step, we will perform genome-wide differential expression analysis of receiver and sender cell types to define DE genes between the conditions of interest (as formalized by the `contrasts_oi`). Based on this analysis, we later can define the levels of differential expression of ligands in senders and receptors in receivers, and define the set of affected target genes in the receiver cell types (which will be used for the ligand activity analysis).

We will apply pseudobulking followed by EdgeR to perform multi-condition multi-sample differential expression (DE) analysis (also called 'differential state' analysis by the developers of Muscat).

Check DE output information in table with logFC and p-values for each gene-celltype-contrast:

Evaluate the distributions of p-values:

These distributions look fine (uniform distribution, except peak at p-value <= 0.05), so we will continue using these regular p-values. In case these p-value distributions look irregular, you can estimate empirical p-values as we will demonstrate in another vignette.

To end this step, we will combine the DE information of senders and receivers by linking their ligands and receptors together based on the prior knowledge ligand-receptor network.

In this step, we will predict NicheNet ligand activities and NicheNet ligand-target links based on these differential expression results. We do this to prioritize interactions based on their predicted effect on a receiver cell type. We will assume that the most important group-specific interactions are those that lead to group-specific gene expression changes in a receiver cell type.

Similarly to base NicheNet (https://github.com/saeyslab/nichenetr), we use the DE output to create a "geneset of interest": here we assume that DE genes within a cell type may be DE because of differential cell-cell communication processes. In the ligand activity prediction, we will assess the enrichment of target genes of ligands within this geneset of interest. In case high-probabiliy target genes of a ligand are enriched in this set compared to the background of expressed genes, we predict that this ligand may have a high activity.

Because the ligand activity analysis is an enrichment procedure, it is important that this geneset of interest should contain a sufficient but not too large number of genes. The ratio geneset_oi/background should ideally be between 1/200 and 1/10 (or close to these ratios).

To determine the genesets of interest based on DE output, we need to define some logFC and/or p-value thresholds per cell type/contrast combination. In general, we recommend inspecting the nr. of DE genes for all cell types based on the default thresholds and adapting accordingly. By default, we will apply the p-value cutoff on the normal p-values, and not on the p-values corrected for multiple testing. This choice was made because most multi-sample single-cell transcriptomics datasets have just a few samples per group and we might have a lack of statistical power due to pseudobulking. But, if the smallest group >= 20 samples, we typically recommend using p_val_adj = TRUE. When the biological difference between the conditions is very large, we typically recommend increasing the logFC_threshold and/or using p_val_adj = TRUE.

We will first inspect the geneset_oi-vs-background ratios for the default tresholds:

We can see here that for all cell type / contrast combinations, all geneset/background ratio's are within the recommended range (`in_range_up` and `in_range_down` columns). When these geneset/background ratio's would not be within the recommended ranges, we should interpret ligand activity results for these cell types with more caution, or use different thresholds (for these or all cell types).

Because we are only out of range for the `KC_other` cell type, we will explore these ratio's in case we would increase the stringency of the logFC cutoff a little bit.

Now we are in range for all cell types, and we will therefore proceed with these tresholds for the ligand activity analysis.

After the ligand activity prediction, we will also infer the predicted target genes of these ligands in each contrast. For this ligand-target inference procedure, we also need to select which top n of the predicted target genes will be considered (here: top 250 targets per ligand). This parameter will not affect the ligand activity predictions. It will only affect ligand-target visualizations and construction of the intercellular regulatory network during the downstream analysis. We recommend users to test other settings in case they would be interested in exploring fewer, but more confident target genes, or vice versa.

The NicheNet ligand activity analysis can be run in parallel for each receiver cell type, by changing the number of cores as defined here. Using more cores will speed up the analysis at the cost of needing more memory. This is only recommended if you have many receiver cell types of interest.

Running the ligand activity prediction will take some time (the more cell types and contrasts, the more time)

You can check the output of the ligand activity and ligand-target inference here:

In the previous steps, we calculated expression, differential expression and NicheNet ligand activity. In the final step, we will now combine all calculated information to rank all sender-ligand---receiver-receptor pairs according to group/condition specificity. We will use the following criteria to prioritize ligand-receptor interactions:

* 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.

* Cell-type specific expression of the ligand in the sender cell type and receptor in the receiver cell type in the condition of interest (to mitigate the influence of upregulated but still relatively weakly expressed ligands/receptors).

* Sufficiently high expression levels of ligand and receptor in many samples of the same group.

* High NicheNet ligand activity, to further prioritize ligand-receptor pairs based on their predicted effect of the ligand-receptor interaction on the gene expression in the receiver cell type.

We will combine these prioritization criteria in a single aggregated prioritization score. In the default setting, we will weigh each of these criteria equally (`scenario = "regular"`). This setting is strongly recommended. However, we also provide some additional setting to accomodate different biological scenarios. The setting `scenario = "lower_DE"` halves the weight for DE criteria and doubles the weight for ligand activity. This is recommended in case your hypothesis is that the differential CCC patterns in your data are less likely to be driven by DE (eg in cases of differential migration into a niche). The setting `scenario = "no_frac_LR_expr"` ignores the criterion "Sufficiently high expression levels of ligand and receptor in many samples of the same group". This may be interesting for users that have data with a limited number of samples and don’t want to penalize interactions if they are not sufficiently expressed in some samples.

Finally, we still need to make one choice. For NicheNet ligand activity we can choose to prioritize ligands that only induce upregulation of target genes (`ligand_activity_down = FALSE`) or can lead potentially lead to both up- and downregulation (`ligand_activity_down = TRUE`). The benefit of `ligand_activity_down = FALSE` is ease of interpretability: prioritized ligand-receptor pairs will be upregulated in the condition of interest, just like their target genes. `ligand_activity_down = TRUE` can be harder to interpret because target genes of some interactions may be upregulated in the other conditions compared to the condition of interest. This is harder to interpret, but may help to pick up interactions that can also repress gene expression.

Here we will choose for setting `ligand_activity_down = FALSE` and focus specifically on upregulating ligands.

Check the output tables

First: group-based summary table

This table gives the final prioritization score of each interaction, and the values of the individual prioritization criteria.

With this step, all required steps are finished. Now, we can optionally still run the following steps

* Calculate the across-samples expression correlation between ligand-receptor pairs and target genes

* Prioritize communication patterns involving condition-specific cell types through an alternative prioritization scheme

Here we will only focus on the expression correlation step:

In multi-sample datasets, we have the opportunity to look whether expression of ligand-receptor across all samples is correlated with the expression of their by NicheNet predicted target genes. This is what we will do with the following line of code:

To avoid needing to redo the analysis later, we will here to save an output object that contains all information to perform all downstream analyses.

In a first instance, we will look at the broad overview of prioritized interactions via condition-specific Chordiagram circos plots.

We will look here at the top 50 predictions across all contrasts, senders, and receivers of interest.

Whereas these ChordDiagrams show the most specific interactions per group, they don't give insights into the data behind these predictions. Therefore we will now look at visualizations that indicate the different prioritization criteria used in MultiNicheNet.

In the next type of plots, we will 1) visualize the per-sample scaled product of normalized ligand and receptor pseudobulk expression, 2) visualize the scaled ligand activities, 3) cell-type specificity.

We will now check the top 50 interactions specific for the Tumor-tissue

```{r, fig.height=13, fig.width=16}

Samples that were left out of the DE analysis are indicated with a smaller dot (this helps to indicate the samples that did not contribute to the calculation of the logFC, and thus not contributed to the final prioritization)

As a further help for further prioritization, we can assess the level of curation of these LR pairs as defined by the Intercellular Communication part of the Omnipath database

Now we add this to the bubble plot visualization:

```{r, fig.height=13, fig.width=16}

Further note: Typically, there are way more than 50 differentially expressed and active ligand-receptor pairs per group across all sender-receiver combinations. Therefore it might be useful to zoom in on specific cell types as senders/receivers:

Eg CLEC9A DCs as receiver:

```{r, fig.height=13, fig.width=16}

Eg CLEC9A as sender:

```{r, fig.height=13, fig.width=16}

You can make these plots also for the other groups, like we will illustrate now for the S group

```{r, fig.height=13, fig.width=18}

__Note__: Use `make_sample_lr_prod_activity_batch_plots` if you have batches and want to visualize them on this plot!

In another type of plot, we can visualize the ligand activities for a group-receiver combination, and show the predicted ligand-target links, and also the expression of the predicted target genes across samples.

For this, we now need to define a receiver cell type of interest. As example, we will take `CLEC9A` cells as receiver, and look at the top 10 senderLigand-receiverReceptor pairs with these cells as receiver.

```{r, fig.width=40, fig.height=7}

What if there is a specific ligand you are interested in?

```{r, fig.width=20, fig.height=7}

In the previous plots, target genes were shown that are predicted as target gene of ligands based on prior knowledge. However, we can use the multi-sample nature of this data to filter target genes based on expression correlation between the upstream ligand-receptor pair and the downstream target gene. We will filter out correlated ligand-receptor --> target links that both show high expression correlation (spearman or pearson correlation > 0.50 in this example) and have some prior knowledge to support their link. Note that you can only make these visualization if you ran step 7 of the core MultiNicheNet analysis.

Now we will visualize the top correlated target genes for the LR pairs that are also in the top 50 LR pairs discriminating the groups from each other:

```{r, fig.width=50, fig.height=16}

You can also visualize the expression correlation in the following way for a selected LR pair and their targets:

```{r, fig.width=30, fig.height=6}

In the plots before, we demonstrated that some DE genes have both expression correlation and prior knowledge support to be downstream of ligand-receptor pairs. Interestingly, some target genes can be ligands or receptors themselves. This illustrates that cells can send signals to other cells, who as a response to these signals produce signals themselves to feedback to the original sender cells, or who will effect other cell types.

As last plot, we can generate a 'systems' view of these intercellular feedback and cascade processes than can be occuring between the different cell populations involved. In this plot, we will draw links between ligands of sender cell types their ligand/receptor-annotated target genes in receiver cell types. So links are ligand-target links (= gene regulatory links) and not ligand-receptor protein-protein interactions! We will infer this intercellular regulatory network here for the top50 interactions. You can increase this to include more hits of course (recommended).