Curation rationale

ImmPrint exists to score immune processes in transcriptomic data cleanly, and in a way you can defend. This page sets out the curatorial decisions that shape the collection.

Why not just use Hallmark, Reactome or GO?

General-purpose collections are invaluable, but they are built for breadth rather than for clean immune signatures. Used directly for module scoring, they tend to share a few recurring problems:

  • Non-specific drivers. Stress and activation-response genes (MAPK, ATF, JUN, FOS, HSP) recur across many sets, dampening the specificity of module scores.
  • Ligand contamination. Signalling sets frequently include the very ligand that triggers the pathway, either for completeness or to account for autocrine signalling in bulk RNA-seq data. However, single-cell RNA-seq allows ligand-producing and responding cells to be distinguished. In this context, a gene set that carries its own ligand pollutes the score of the cell you actually care about.
  • No internal structure. A single set may fold a receptor, its entire transducer cascade, generic stress machinery and dozens of loosely associated targets into one undifferentiated list.
  • Unsourced membership. Membership is rarely traceable to a specific paper linking that gene to that pathway.

Principle 1 - Mechanistic specificity

Each set aims to capture a single, mechanistically coherent process. Ubiquitous machinery that carries no pathway-specific information (ribosomal proteins, proteasome subunits, ubiquitination machinery, core housekeeping genes) is excluded from every set. The test for each candidate gene is not “is it expressed when this pathway is active?” but “does its expression specifically mark this process?”

Principle 2 - Inducer and responder separation

The most important and most context-dependent rule. A pathway’s own ligand is excluded from its own signalling set, because the ligand reports on the sending cell rather than the responding one. The same molecule is kept wherever it is a genuine output of the process being measured.

So interferon-gamma is treated differently by context: IFNG is not a member of IMMPRINT_IFNG_SIGNALLING, but it is a legitimate effector of IMMPRINT_TYPE1_RESPONSE. The same logic removes IL2 from IMMPRINT_IL2_SIGNALLING and TNF from IMMPRINT_TNF_SIGNALLING, while leaving receptors, transducers and downstream effectors in place.

Principle 3 - Focus on the signal-receiving cell

A generalisation of Principle 2: a set describes the cell receiving a signal or enacting a process rather than the cell sending it. The checkpoint sets make this concrete. The inhibitory-receptor sets contain the receptor and its intrinsic signalling apparatus rather than the ligands presented by the opposing cell (the PD-L1/PD-L2 ligands are deliberately absent from IMMPRINT_PD1_SIGNALLING). A corollary, for the coinhibitory programmes: these sets contain only genes that are induced or active in the programme and exclude genes the pathway represses.

Principle 4 - Cell type specificity

The same ligand drives different programmes in different cells. TGF-β, for instance, does very different things in T cells, fibroblasts and epithelium. ImmPrint layers cell type-specific effector refinements onto each pathway’s shared pan core, so a cell is scored against the version that matches what it is. These refinements are added only where the divergence is real and well-documented.

Principle 5 - Citations you can follow

Every gene set membership carries at least one PubMed ID. The citation should be primary literature that links that gene to that named pathway. A generic review is not enough. Where a gene appears in more than one set, each appearance is meant to be supported by a paper appropriate to that context.

Note

This is an alpha release. Citations are checked against PubMed where possible, but some may still need correcting. If you find a citation that does not support its membership, please flag it on GitHub so it can be fixed in the next version.

Principle 6 - Corroboration by MSigDB

Where a gene’s membership is corroborated by an assigned MSigDB reference gene set (from Hallmark, Reactome or GO:BP), that set is recorded in reference_gene_set. A blank entry means either the gene is not covered by any MSigDB counterparts or that there is no valid counterpart.

Principle 7 - Explicit family exclusions

A few gene families are excluded from all sets by standing policy, because they are general stress or activation responders rather than pathway-specific markers: the MAPK families (MAPK*, MAP2K*, MAP3K*), and the ATF, JUN, FOS and HSP families. Excluding them keeps sets from collapsing toward a generic “this cell is activated” signal.

The signalling hierarchy annotation

Separately from the principles above, every gene is annotated with a position in a three-tier signalling hierarchy (pathway_level):

  • Receptor (1): the sensor or input, meaning surface or intracellular receptors. For receptor-less metabolic or structural sets, the entry tier is the substrate-import transporters or the top driving transcription factor.
  • Transducer (2): kinases, adaptors and transcription factors acting at the signalling node (STATs, IRFs, NF-κB, SMADs, master lineage TFs), plus core processing enzymes. Initiator and inflammatory caspases sit here. The tier marks where a gene product acts, so negative regulators that operate at this node (e.g. SOCS proteins and the inhibitory SMADs) sit here too. Their direction is recorded separately by regulatory_role.
  • Effector (3): the output, meaning secreted cytokines and chemokines, cytotoxic and antiviral/antimicrobial molecules, presented MHC, terminal metabolic complexes and induced target-gene products. Executioner caspases sit here.

The tiers are interpretive metadata rather than a scoring axis. You can subset a pathway by tier, but the effector tier is often small, so per-tier subsetting is best treated as optional. By design, antigen-receptor and inhibitory-checkpoint sets (TCR, BCR, CTLA4, PD1) carry no effector tier, since they are detection or inhibition modules that terminate at the transducer level.

The regulatory role annotation

The tier locates a gene in the cascade; it does not say whether the gene drives the pathway or holds it back. That direction is recorded separately, per membership, in regulatory_role:

  • none: a forward component or a faithful output of the pathway. This is most genes, including receptors, transducers and effector products.
  • inhibitory: a negative regulator of the pathway, such as the SOCS and CISH negative-feedback proteins, the inhibitory SMADs, the Wnt destruction complex, decoy receptors and the anti-apoptotic BCL2 guards.
  • activating: a genuine positive-feedback amplifier, reserved for the rare gene that feeds back to boost its own pathway rather than merely transmitting it.
  • context_dependent: the sign genuinely varies, as for PTPN11 in cytokine signalling.

The sign is judged relative to the gene’s own pathway, which is why it lives on the membership rather than the gene. IL1RN is a faithful output of IL-10 even though it antagonises IL-1, so in IMMPRINT_IL10_SIGNALLING it is none, not a negative regulator. BCL2 opposes intrinsic apoptosis but is a survival output of IL-7 and IL-15, so it is inhibitory in one set and none in the others. PTPN11 enacts the inhibitory signal in PD1 and CTLA4 but modulates JAK/STAT either way elsewhere.

Versioning

Each collection is an immutable, versioned snapshot. The gene sets behind a published analysis can be traced back to the exact definitions used, the way an MSigDB release is pinned. New curation ships as a new version; released versions are never altered. See the changelog.

Intended use

ImmPrint is built for module scoring of immune processes, reading out the activity of specific, mechanistically defined programmes across cell types, tissues and disease states. The sets are plain, unweighted gene lists, so they work with any rank-based scoring method (UCell, AUCell, singscore, ssGSEA) and with average-based methods like Seurat::AddModuleScore. See Access to get started.