Read InterProScan Results: Domains & GO

InterProScan searches many signature libraries in one run

InterProScan is a search framework rather than one universal domain model. It runs member-database methods such as Pfam, PANTHER, PROSITE, PIRSR and others, each built for different kinds of evidence: profile hidden Markov models, families, conserved sites, sequence patterns or rules. Jones et al. describe the InterProScan 5 architecture in Bioinformatics 2014;30:1236–1240, doi:10.1093/bioinformatics/btu031.

Different methods can detect the same biological region at slightly different boundaries or levels of specificity. Agreement is useful support; duplicate-looking rows are not necessarily redundant computation. Conversely, one highly specific site match and one broad family match answer different questions even when they overlap.

SeqBench strips proper FASTA headers and nonletters other than *, joins parsed records, uppercases the result and submits that normalized sequence. A multi-record FASTA is therefore concatenated into one artificial protein rather than analyzed separately. Submit exactly one record, then confirm that its sequence and length are the intended construct.

A signature accession is not the same as an InterPro entry

The signature accession identifies the member database's model or pattern that matched. Its library tells you where that model came from. The InterPro accession, usually beginning IPR, identifies an integrated biological entry to which one or more member signatures have been assigned after curation.

A signature can have no InterPro entry. That does not make the match false; it means the member model has not been integrated into an InterPro entry in the result's release. Inspect the member accession and description rather than treating a blank InterPro column as a failed scan.

InterPro entry types include families, domains, repeats, homologous superfamilies and sites. A family says the whole protein or region belongs to an evolutionary group; a domain is a structural or functional unit; a site can be only a few residues. Do not draw all entry types as interchangeable domain boundaries.

Coordinates tell you which sequence claim the match covers

A location written 45–130 covers residues 45 through 130 inclusively, an 86-residue span. Compare that interval with the full protein length, known signal peptides, transmembrane segments, low-complexity regions and construct boundaries. A hit that ends at the last submitted residue may be clipped by the input rather than by a natural domain boundary.

Repeated domains should appear as separate match locations even when they use the same signature accession. SeqBench's “Domain architecture” list deduplicates InterPro accessions for a compact summary, while the signature table and domain cartoon retain each location. Use the location-level rows to count repeats.

Boundary disagreement of a few residues is common because models are trained on different alignments and score local evidence differently. Treat a consensus region as more robust than one exact terminal coordinate unless structural or experimental data define the boundary.

An E-value measures chance expectation under one search model

For methods that report it, an E-value estimates how many matches with that score or better would be expected by chance in the search context. Smaller values indicate stronger sequence evidence under that model: 1e-30 is stronger than 1e-3. It is not a probability that the biological annotation is true and not a percentage identity.

Raw E-values should not be compared blindly across member libraries. The models, databases, sequence lengths, calibration and reporting thresholds differ. A PROSITE rule may report no E-value at all, while a profile-HMM method does. A dash in the E-value column can therefore mean “this method does not provide that field,” not “zero evidence.”

Near-threshold matches deserve context. Ask whether another independent signature supports the same region, whether the alignment covers the model well, whether key residues are present and whether the proposed function is plausible for the protein's organism and architecture.

Overlapping rows can represent support, nesting or conflict

Do not resolve every overlap by keeping only the smallest E-value. That can discard a useful specific site beneath a broad family model. First group rows by location, then compare what type of claim each row makes.

A defensible summary might read “residues 45–130 are supported as a kinase-like catalytic region by two signatures, with an active-site match inside it,” rather than listing three domains or selecting one label without the evidence chain.

• Support — two member databases independently detect approximately the same domain or family.
• Nesting — a broad superfamily spans a narrower catalytic domain or conserved site.
• Architecture — one long family model overlaps several component domains because it models the protein as a whole.
• Repeats — the same signature occurs at several non-overlapping locations.
• Conflict — incompatible family assignments cover the same residues and require closer inspection of scores, coverage and sequence quality.

GO terms are transferred annotations, not assays on your protein

InterPro entries can carry Gene Ontology cross-references in three namespaces: molecular function, biological process and cellular component. InterProScan returns those terms when the matched entry supports the mapping. SeqBench deduplicates the terms across all matches.

The evidence path is therefore sequence → signature → InterPro entry → GO cross-reference. A molecular-function term may be a useful hypothesis, but it is not direct biochemical evidence that the submitted protein performs that activity. Biological-process and cellular-component terms often require organism, expression and interaction context that sequence alone cannot establish.

Broad ancestor terms and several related descendants can appear together because GO is hierarchical. Preserve the term IDs and inspect the contributing matches. A long list of related terms is not the same as many independent observations.

Worked interpretation: one region with three kinds of evidence

Consider a hypothetical 310 aa protein with a Pfam signature from residues 48–137 at E = 2e-24, a homologous-superfamily match from 43–151, and a PROSITE active-site pattern at residue 102 with no E-value. Suppose the first two integrate to one IPR domain entry and the site maps to a second IPR entry.

The careful reading is not “three domains.” It is one domain-scale region supported at two resolutions plus one short conserved site inside it. The Pfam E-value gives strong statistical sequence evidence for its model; the superfamily supplies broader evolutionary context; and the site pattern identifies a residue motif whose functional importance still needs alignment and biochemical context.

If a GO catalytic-activity term is attached to the integrated domain, report it as an InterPro-derived annotation. Before assigning that function to the protein, confirm the full catalytic residue set, expected cofactors, domain partners, cellular location and evidence from orthologs or experiments.

No match does not mean no domain or no function

Short proteins, rapidly evolving regions, intrinsically disordered proteins, taxon-specific families and incomplete fragments can fall outside current signature coverage. A novel domain has no model by definition. Low-complexity masking and unusual composition can also reduce detectable sequence evidence.

Check that the sequence is protein rather than untranslated DNA, that internal stop symbols and ambiguous residues are intentional, and that the expected region was not omitted. Search close homologs, examine predicted structure, conservation, disorder and transmembrane topology, and consider profile searches against appropriate databases.

The correct negative statement is “no reported member-database signature passed the scan's criteria for this sequence and release.” It is not “the protein has no function.”

Build an annotation from strongest claim to weakest

1. Validate the submitted sequence, length, organism and construct boundaries.
2. Group location-level matches into regions instead of reading the table only by E-value.
3. Separate family, domain, repeat, superfamily and site entry types.
4. Look for agreement across independent signatures and inspect unsupported or conflicting rows.
5. Use GO terms as transferred hypotheses and preserve the signature and InterPro accessions that generated them.
6. Cross-check architecture with signal peptides, transmembrane segments, disorder, structure and known homologs.
7. State what remains untested and validate function experimentally when the conclusion matters.

Know where the scan runs and what the page returns

SeqBench submits protein sequences to EMBL-EBI's InterProScan Job Dispatcher, then polls the job and parses the result. The sequence therefore leaves SeqBench and is processed on EBI infrastructure. Do not submit confidential sequences without reviewing the relevant data policy and your organization's requirements.

The page is intended for one protein up to 5,000 amino acids after normalization and normally waits a few minutes. The current normalizer does not reject multi-record FASTA, so supplying several records creates one concatenated sequence and must be avoided. The page displays deduplicated InterPro entries, deduplicated GO terms and every parsed signature location with library, coordinates and E-value where supplied; it does not return the complete raw report or every method-specific alignment field.

Database releases and integrations change. Record accessions and the date of analysis if the annotation will support a publication, regulatory record or long-lived construct dossier.

Related references

Related tools

Related guides

• What Does pLDDT Mean in an AlphaFold Structure? →
• Hydrophobicity Plots for Transmembrane Domains →
• Interpret GO Enrichment: FDR & Background →