nichenetr

NicheNet is an R implementation designed to study intercellular communication (CCC) by linking ligands to their downstream target genes. Unlike traditional approaches that only link ligands to receptors, NicheNet incorporates intracellular signaling and transcriptional regulation to predict how ligands influence gene expression in receiver cells.

Core capabilities include:

• Assessing how well sender-cell ligands predict gene expression changes in receiver cells.
• Prioritizing ligands based on their effect on gene expression.
• Inferring putative ligand-target links and potential signaling paths.
• Validating the prior ligand-target model.
• Constructing user-defined prior ligand-target models.

Input requirements: Users must provide cell type-annotated expression data (single-cell or sorted bulk) for human or mouse. The analysis extracts three key features: a list of potential ligands, a gene set capturing downstream effects, and a background set of genes.

NicheNet is designed to prioritize ligand-receptor interactions based on their observed effect on gene expression (target genes). Therefore, it requires some form of differential expression in the receiver cell population to function effectively. This differential expression can be:

1. Between two conditions: e.g., comparing a treated vs. control state.
2. Between two cell types: e.g., comparing a progenitor cell to a differentiated cell type where differentiation is influenced by the microenvironment.

If you only have steady-state (homeostatic) data and want to find ligand-receptor pairs, you can infer them based solely on expression levels (similar to CellphoneDB), but NicheNet's core strength—prioritizing based on downstream target gene enrichment—will be limited.

Warning: If you use all expressed genes in a cell type as your geneset_oi (gene set of interest) in steady-state data, you risk falsely linking ligands to 'cell-intrinsic' genes (genes that are part of the cell's identity rather than being regulated by the environment). This leads to noisy and irrelevant predictions.

NicheNet predicts how strongly a ligand may regulate a target gene by integrating biological knowledge from three main sources:

1. Ligand-receptor interactions
2. Signal transduction (e.g., protein-protein and kinase-substrate interactions)
3. Gene regulatory interactions (e.g., ChIP-seq and motifs)

The process involves two main steps:

• Network Integration: Individual data sources are integrated into a ligand-signaling network (pathways from ligands to transcriptional regulators) and a gene regulatory network (interactions between regulators and target genes).
• Network Propagation: Using algorithms like Personalized PageRank (PPR), the signal is propagated from the ligand, through receptors and signaling proteins, to transcriptional regulators, and finally to target genes. A high regulatory potential score indicates that the target gene's regulators are downstream of the ligand's signaling network.

You can prioritize the most important receptors for a given ligand by scoring interactions based on the expression strength of the receptor. This helps identify which receptors are most likely driving the observed ligand activity in a specific cell type.

Methodology: The score ligand_scaled_receptor_expression_fraction is calculated by averaging the normalized ranks of the receptor's expression and its expression fraction:

$$ ext{score} = 0.5 imes ext{normalized rank(expression)} + 0.5 imes ext{normalized rank(fraction)}$$

This approach ensures that ligands are prioritized based on receptors that are both highly expressed and present in a significant fraction of the receiver cells.

# Score ligand-receptor interactions based on receptor expression strength
exprs_sender_receiver = lr_network %>% 
  inner_join(exprs_tbl_ligand, by = c("ligand")) %>% 
  inner_join(exprs_tbl_receptor, by = c("receptor")) %>% inner_join(DE_sender_receiver %>% distinct(niche, sender, receiver))
  
ligand_scaled_receptor_expression_fraction_df = exprs_sender_receiver %>% 
  group_by(ligand, receiver) %>% 
  mutate(rank_receptor_expression = dense_rank(receptor_expression), 
         rank_receptor_fraction  = dense_rank(receptor_fraction)) %>% 
  mutate(ligand_scaled_receptor_expression_fraction = 0.5*( (rank_receptor_fraction / max(rank_receptor_fraction)) + ((rank_receptor_expression / max(rank_receptor_expression))) ) )  %>% 
  distinct(ligand, receptor, receiver, ligand_scaled_receptor_expression_fraction) %>% 
  distinct() %>% 
  ungroup() 

NicheNet's prioritization algorithm does not inherently distinguish between up- and downregulated genes. The regulatory potential score represents the evidence/probability that a ligand regulates a target gene, regardless of whether that regulation is inducing (upregulation) or repressive (downregulation).

Key implications:

• Prioritization Bias: If your gene set of interest contains more upregulated genes than downregulated genes, ligands that primarily upregulate genes will naturally receive higher priority scores.
• Input Strategy: To capture the full scope of signaling, it is recommended to include both upregulated and downregulated genes in your gene set of interest. However, you can choose to analyze only one direction if your research question specifically targets one (e.g., finding only inducing ligands).
• Verification: Because the model treats all retrieved interactions as 'active', you should manually verify the direction of effect (up vs. down) by visualizing the expression of target genes (e.g., using a heatmap).

The Differential NicheNet pipeline differs from the default NicheNet pipeline in how it prioritizes ligand-receptor (L-R) pairs. While default NicheNet prioritizes pairs based on ligand activity alone, Differential NicheNet incorporates information from the differential expression (DE) of L-R pairs compared to other niches or spatial locations.

Prioritization Patterns

• Differential NicheNet: Tends to find hits that are high-DE but low-activity. These may be biologically important ligands that have low predicted activity due to incomplete prior knowledge in NicheNet, or ligands that lack strong signaling effects (e.g., cell adhesion).
• Default NicheNet: Tends to find hits that are low-DE but high-activity. These may represent ligands regulated post-transcriptionally, or cases where high activity is an artifact of inaccurate prior knowledge.

Candidate Selection

When evaluating results, ligands that exhibit both high DE and high activity are considered the strongest candidates for regulating the biological process of interest and are recommended for experimental validation.