The biology in your ligand-receptor analysis, and how to read it

Ligand-receptor analysis has become a routine step after clustering, and it is easy to run: choose your cell groups, choose a method, read off a ranked table of interactions. That table is where most analyses stop. It is better treated as a list of hypotheses than as a result, because each score is a plausibility estimate built from gene expression, not a measurement of signalling.

Whether that table becomes a chord diagram or a set of testable leads depends on a few decisions: which score you trust, how well your cells are labelled, and what you do with the top hits. Named tools appear throughout, because the value here is in the method, not in any one product.

What a ligand-receptor score is claiming

Every method starts from the same idea. If a ligand gene is expressed in one group of cells and its receptor in another, those groups might communicate through that pair. One group is the sender, the other the receiver, and direction matters because signalling is directional. Autocrine loops, where sender and receiver are the same population, fit the same scheme.

Two caveats travel with that idea. Signalling happens at the protein level, and mRNA is only a proxy for protein abundance. And co-expression of a ligand and receptor does not show that the two proteins ever meet, bind, or trigger anything. A score ranks plausibility from expression; it does not measure signalling, and reading it as if it does is where most over-interpretation starts.

Magnitude and specificity measure different things

LIANA formalised a distinction worth applying whatever tool you use: the magnitude of an interaction versus its specificity.

Sorting by magnitude alone returns the loudest interactions, which are often the least informative. Broadly expressed ligands sit near the top of a magnitude ranking in almost any dataset while saying little about what makes one niche different from another. Specificity is what pulls out the interactions particular to your system, the ones that separate a tumour-associated macrophage from a resting one, or a disease niche from its healthy counterpart. The interactions worth an experiment usually rank well on both.

No single method is the right answer

Different methods disagree on the same data, because each encodes its own assumptions about what makes an interaction strong or specific. CellPhoneDB, CellChat and the other established scoring functions can rank the same matrix differently. A recent review counts more than a hundred inference tools and around fifty ligand-receptor databases, which leaves a lot of room for disagreement.

LIANA's answer, and the one we find most defensible, is to aggregate rather than pick a favourite. Described in its founding paper and extended in LIANA+, it re-implements several established methods, including CellPhoneDB and CellChat, runs them on the same data, and combines their outputs into a consensus rank. An interaction that ranks high across methods is a stronger lead than one that only appears under the single method that happens to favour it.

One database detail is worth knowing. Many ligands and receptors work as complexes, and requiring every subunit to be expressed before an interaction counts reduces false positives. CellPhoneDB and CellChat both handle complexes; a method that scores a heteromeric receptor as present from a single detected subunit will report interactions that cannot physically occur.

Your interactions are only as good as your labels

Ligand-receptor inference runs on your cell groups and inherits everything about them, mistakes included.

Run it on raw Leiden clusters and the output reads "Cluster 4 to Cluster 9," which is bookkeeping, not biology. Annotate the same clusters and it reads "cytotoxic T cell to tumour-associated macrophage," which is something a biologist can act on. The numbers are identical; only the interpretability changes, and interpretability is the whole value of the step. We cover annotation decisions in more detail in our scRNA-seq cluster annotation guide.

Resolution matters as much as accuracy:

• Coarse labels average away the signalling that separates the states inside them.
• Finer, more homogeneous groups reveal interactions that were diluted in a mixed cluster.
• Merge two distinct states under one name and an interaction driven by one of them gets attributed to both.

So annotation is part of the analysis, not a formality before it. Getting the labels right, at a resolution that matches the biology you are after, often decides whether a result is real or an artefact of how you clustered.

Where the deeper biology is

The top of the table is a starting point. The results that hold up come from interrogating it rather than screenshotting it.

• Directionality. The result is a directed network. Which populations are the dominant senders of a pathway, and which are wired to receive it, is often the actual question. Reading direction turns a symmetric-looking diagram into a description of how the tissue is organised.
• Per-cluster programmes. Instead of the single strongest interaction, look at the set a population is engaged in. A cell type defined by a coherent set of incoming and outgoing signals is a stronger object than any one pair. In a tumour microenvironment, this is how you move from a list of ligands to an account of how stroma, immune and tumour cells hold each other in a state.
• Condition contrasts. An interaction present in both disease and control is context. One that appears, strengthens or drops between them is a candidate mechanism. Differential communication is harder to infer than steady-state, because it has to account for variation between samples rather than pooling cells, but it is where the biology you can act on usually sits.
• Downstream corroboration. Co-expression is a hypothesis; evidence that the receiver shows the expression changes the pathway should produce is corroboration. NicheNet approaches this directly, linking ligands to the target genes they are expected to regulate in the receiving cell. Pairing a specific interaction with a receiver-side response, and ideally a known druggable node, is how many of these findings become testable therapeutic hypotheses.

What a top interaction is, and is not

A high-ranking pair from scRNA-seq is evidence that two populations could communicate through that axis, from transcript co-expression, aggregated across methods, given your annotation. Being clear about what it is not is what makes that worth trusting:

• It is not proof the interaction occurs.
• It carries no spatial information. Dissociated data cannot tell whether sender and receiver were ever neighbours, and many interactions only work over short distances. Two populations can share a perfect ligand-receptor match and sit at opposite ends of a tissue. For when spatial context matters, see our guide to integrating spatial data with scRNA-seq.
• It is shaped by the database. Well-studied pathways are richly annotated and recur; novel or poorly characterised interactions are underrepresented because no one has curated them yet. A ranking reflects the database as much as your cells.

That makes it a hypothesis generator, which is what it should be. The interactions that survive, specific, consistent across methods, coherent with a downstream signature, resting on defensible labels, are strong leads.

Running this in ScarfWeb

ScarfWeb's ligand-receptor analysis runs LIANA's consensus scoring in the browser, with no code and the interpretation steps built into the workflow.

Figure 1: The Ligand-Receptor Analysis modal in ScarfWeb, where you set inference parameters before running LIANA consensus scoring.

Start with annotation, since everything above depends on it. Label clusters with auto-annotation or CyteType first, so the results read in cell-type terms. From the Explore page, open a run, pick the grouping to infer over, and set a few parameters:

• Minimum cells per group. Groups below the threshold are excluded. Lower it for small datasets so rare populations are not dropped.
• Expression proportion. The fraction of cells in a group that must express a ligand or receptor for it to count.
• Cell selection. Run over all cells, a categorical subset such as disease against control, or a saved custom cell set.
• Species and gene symbols. Human or mouse, using official gene symbols.

Results come back as a table of source and target populations with their ligand and receptor complexes, scored for both magnitude and specificity. You can filter and search the full table, export it, or open a dotplot to look closely at chosen senders, receivers and interactions. Because the scoring is a consensus rather than one method, the top of the table stays closer to what holds up.

We built it this way to make the careful version of the analysis the default one, so what you take to the bench is a set of leads you can defend.

Frequently asked questions

What does a ligand-receptor score actually measure?

A ligand-receptor score ranks plausibility from transcript co-expression between sender and receiver cell groups. It does not measure protein binding, signalling, or whether the two cell types were ever neighbours in tissue. Treat ranked interactions as hypotheses, not confirmed communication events. See What a ligand-receptor score is claiming above for the full context.

What is the difference between magnitude and specificity in ligand-receptor analysis?

Magnitude reflects the strength of ligand and receptor expression in the sender and receiver groups. Specificity reflects how unique an interaction is to that sender-receiver pair within the dataset. Broadly expressed ligands often rank high on magnitude alone; interactions worth following up usually score well on both. The comparison table in Magnitude and specificity measure different things walks through both.

Can you trust a top-ranked interaction from scRNA-seq?

A top-ranked pair is a lead, not proof. It carries no spatial information, is shaped by the ligand-receptor database used, and inherits any mistakes in your cell labels. Interactions that are specific, consistent across methods, coherent with downstream receiver signatures, and grounded in defensible annotation are the strongest candidates. See What a top interaction is, and is not and our cluster annotation guide.

How does ScarfWeb run ligand-receptor analysis?

ScarfWeb runs LIANA consensus scoring in the browser. After annotating clusters with auto-annotation or CyteType, you pick a grouping, set parameters such as minimum cells per group and expression proportion, and receive a table scored for both magnitude and specificity. Results can be filtered, exported, or explored in a dotplot. See Running this in ScarfWeb above for the workflow.