# UniProt Evidence Codes Explained A UniProt evidence code is a machine-readable tag, drawn from the Evidence and Conclusion Ontology (ECO), that records how a specific annotation on a UniProt Knowledgebase (UniProtKB) entry was derived, for example by direct experiment, by similarity to a characterized protein, or by an automated rule. Reading the code next to the annotation tells you whether a claim about a protein is backed by bench data or by prediction, which is the difference between citing an established fact and quietly inheriting an assumption. This matters because the two halves of UniProtKB differ enormously in how their content is produced. The reviewed section, historically called Swiss-Prot, contains manually curated entries, while the unreviewed section, historically called TrEMBL, contains computationally generated records from Ensembl and other source databases. The gap between the two has widened as sequencing has accelerated. When UniProt release 2020_01 reached 178 million records, automated annotation systems already covered roughly 30 percent of the knowledgebase, with 53 million proteins annotated by UniRule alone [1]. Most entries will never receive individual expert review, so the evidence code attached to each functional statement is often the only signal a researcher has about how much to trust it. ## What UniProt Evidence Codes Are UniProtKB is split into two sections. Reviewed entries (Swiss-Prot) are curated by human experts who read the primary literature, verify sequences, and attach annotations with defined provenance. Unreviewed entries (TrEMBL) are created automatically from coding sequences submitted to nucleotide databases and annotated largely by automated pipelines. Every individual annotation inside either type of entry, such as a function, a subcellular location, a catalytic activity, a Gene Ontology (GO) term, or a domain assignment, can carry its own evidence tag. Each evidence tag resolves to a term in the Evidence and Conclusion Ontology, a formal ontology shared across bioinformatics resources so that a claim of "experimental evidence" means the same thing in UniProt, GO, and other databases that adopt ECO. In UniProt, the code does three jobs at once. First, it states the type of support, such as a direct assay or sequence similarity. Second, it states whether the assertion was made manually by a curator or automatically by software. Third, it points to the source of the support, such as a PubMed identifier, a reviewed UniProtKB accession, an InterPro signature, or a named rule set. Three facts about scoping are worth fixing early because they cause most confusion: 1. Evidence codes attach to individual annotations, not to whole entries. A single entry can mix experimentally verified functions with predicted domain hits and an automatically assigned subcellular location. The entry itself does not have one global evidence code. 2. Evidence codes describe the method of support, not the correctness of the claim. A code tells you the annotation came from an experiment or a rule. It does not certify that the experiment was well designed or that the rule fired correctly. 3. Evidence codes are distinct from reviewed status. Reviewed status describes the entry as a whole and its curation process. Evidence codes describe individual statements inside the entry. A reviewed entry contains manual assertions, but an unreviewed entry can still contain some manually propagated information transferred from reviewed homologs. When you see the phrase "UniProt evidence codes explained interpretation" in a training context, the practical skill being taught is exactly this: read the annotation, read the tag, read the source behind the tag, and decide how much weight the claim deserves for your purpose. ## Quick Reference: The Main Evidence Tags The table below summarizes the tags you will encounter most often when you use UniProt. The "displayed as" column reflects how UniProt renders the evidence on the website. | Displayed as | Typical ECO code | What it means | Typical origin | |--|--|--|--| | Experimental evidence | ECO:0000269 | A manually asserted annotation supported by published experimental work | Curator reading primary literature | | Direct assay evidence | ECO:0000314 | A specific experimental assay directly demonstrated the annotation | Enzyme assays, localization microscopy, binding assays | | Physical interaction evidence | ECO:0000353 | A direct physical interaction supports the annotation | Two-hybrid, co-immunoprecipitation, affinity capture | | By similarity | ECO:0000250 | A curator transferred the annotation from a reviewed homolog | Curated similarity-based transfer | | Phylogenetic evidence | ECO:0000318 | Inference based on evolutionary relationships and ancestral reconstruction | Curator or tool-assisted phylogenetic analysis | | Combinatorial evidence | ECO:0000305 | Several independent evidence types combine to support the assertion | Curator synthesis | | Curator inference | ECO:0000303 | The curator's reasoned inference from biological context | Expert judgment | | Sequence model | ECO:0000255 | A curated family model, such as a PROSITE-ProRule, supports the assertion | Rule-based manual assertion | | Electronic annotation | ECO:0000256 | Software assigned the annotation automatically | UniRule, ARBA, InterPro scans, similarity pipelines | The ECO codes listed here are the ones that appear most frequently in UniProt files and API output. The ontology contains many more specific terms, and UniProt maps its internal evidence types onto ECO when publishing data, so a given entry may show a more specific descendant code in raw downloads. ## The Evidence Categories Explained ### Experimental evidence Experimental evidence is the strongest category. The tag ECO:0000269 marks a manual assertion supported by experimental work, and more specific codes narrow the kind of experiment. A direct assay, tagged ECO:0000314, means the assay itself demonstrated the property, for example a purified enzyme shown to convert a specific substrate in vitro. Physical interaction evidence, tagged ECO:0000353, means a direct interaction method such as a two-hybrid screen or affinity capture demonstrated that two proteins bind. Curators attach these tags only after reading the underlying paper, and the tag is paired with the source reference, typically a PubMed identifier, so you can pull the original study and judge it yourself. When a protein carries experimental evidence for a function, that function is usually established for the exact protein in the entry, in the exact organism, under the conditions of the published experiments. Extrapolation across conditions or strains still requires judgment, but the anchor point is solid. ### Sequence similarity, the "By similarity" tag The tag ECO:0000250 appears in entries as the phrase "By similarity." It means a curator examined a homologous protein in a related organism and transferred an annotation. For example, if a [protein kinase](/knowledge/molecular-biology/protein-kinase) activity was demonstrated experimentally for the human enzyme, a curator may annotate the mouse ortholog with the same activity and mark it "By similarity," citing the reviewed human entry as the source. This tag is honest and useful, but it carries a specific limitation you must respect. Similarity-based transfer assumes the function is conserved between the two proteins. The closer the organisms, the safer that assumption. Transferring a metabolic function between two mammalian orthologs is usually sound. Transferring a developmental function from a fly protein to a distant invertebrate ortholog based on sequence similarity alone can be wrong, because sequence conservation does not guarantee functional conservation across large evolutionary distances. Chains of similarity transfers compound this risk, so when an annotation says "By similarity" and cites another reviewed entry, check how distant the two organisms are before you treat the claim as established for your protein. ### Phylogenetic evidence Phylogenetic evidence, tagged with codes in the ECO phylogenetic family such as ECO:0000318, marks assertions derived from evolutionary analysis. A curator or a phylogenetic tool reconstructs the history of a gene family and infers, for example, that a function arose at a particular node and was likely retained or lost in specific lineages. This approach complements plain similarity. Two proteins can be similar overall yet differ in a lineage-specific expansion where one copy retained an ancestral role and a duplicated copy specialized. Phylogenetic reasoning can distinguish those cases in a way that a pairwise similarity score cannot. ### Combinatorial and curator inference Two related codes capture expert synthesis rather than a single source. Combinatorial evidence, ECO:0000305, means several independent lines of evidence together support the assertion, for example a conserved catalytic residue plus a known interaction partner plus a characterized pathway membership. Curator inference, ECO:0000303, marks a reasoned conclusion drawn by the curator from biological context, such as assigning a subcellular location because all characterized members of a small family localize the same way. These codes signal thoughtful manual work, and they are generally reliable, but they are not experiments. If your analysis depends on a specific claim, the difference between a combinatorial assertion and a direct assay can matter. ### Sequence model evidence The code ECO:0000255 marks assertions supported by a curated sequence model, most often a PROSITE-ProRule. These rules encode family signatures plus conditions, for example the requirement that a catalytic residue be present at a specific position for a rule to apply. Because the rule encodes expert knowledge in executable form, its output is a manual assertion rather than a simple pattern match. This category sits between curator work and fully automatic annotation, and in files it often appears with a rule identifier as the source. ### Electronic annotation ![Calculation dividing 53 million UniRule annotated proteins by 178 million records to get roughly 30 percent.](/article-images/illustrations/uniprot-evidence-codes-explained-unirule-share.webp) Electronic annotation, ECO:0000256, is the workhorse of the unreviewed section of UniProtKB. Software assigned the annotation with no human inspecting the individual protein. The volume is enormous. In release 2020_01, UniRule, UniProt's unified rule engine, provided annotation for 53 million proteins, roughly 30 percent of the 178 million records in the knowledgebase at that release [1]. UniRule integrates rule sets developed by UniProt consortium members, including RuleBase, HAMAP, PIRSR, and PIRNR, and selects candidate proteins using protein family signatures from InterPro combined with taxonomic and other constraints [1]. The rules fire only when a protein matches the conditions defined by expertly curated sets of reviewed proteins with experimentally supported functions [1]. More recently, UniProt has added automatic annotation approaches based on machine learning, such as automatic rules, which likewise publish ECO:0000256 tags with the generating system as the source. The practical consequence is that ECO:0000256 covers a range of quality. A UniRule assertion backed by a tightly constrained family rule is far more trustworthy than a loose name assigned by a simple similarity hit, yet both carry the same code. When the source field next to the tag names the system, read it. The source tells you which pipeline produced the claim, and that distinction is often more informative than the code alone. ## Protein Existence Levels Are Not Evidence Codes A frequent point of confusion in any UniProt database tutorial is the difference between evidence codes and Protein Existence (PE) levels. Both describe confidence, but they describe different things and appear in different places. PE levels summarize the state of evidence for the physical existence of the protein as a whole, on a scale of one through five. Level 1 means experimental evidence at the protein level, for example the protein has been identified by mass spectrometry or sequencing of purified protein. Level 2 means experimental evidence at the transcript level, so the mRNA is demonstrated but the protein itself has not been directly detected. Level 3 means the protein is inferred from homology to a characterized protein. Level 4 means the protein is purely predicted from [nucleotide sequence](/knowledge/molecular-biology/nucleotide-sequence). Level 5 means the existence of the protein is uncertain. Evidence codes, by contrast, attach to individual functional annotations and use the ECO framework described above. A level 4 predicted protein can carry a well-supported electronic annotation from a strict UniRule, and a level 1 protein can carry a predicted subcellular location. When you audit an entry, read both signals: the PE level for existence, the evidence tags for each functional claim. ## Where Evidence Codes Appear in Practice Knowing what the codes mean is only half the skill. You also need to find and use them efficiently across the UniProt interfaces you use daily. ### On the uniprot.org website Every annotation on an entry page is labeled with its evidence. Hovering over a term, such as a GO annotation or a subcellular location, reveals the ECO code, the source, and a link to the underlying reference or rule. The entry page also groups annotations visually, so manually asserted items and automatically assigned items are distinguishable at a glance. On search result pages, UniProt exposes an "Evidence" facet that lets you restrict results, for example to entries containing experimentally supported annotations, without writing a query by hand. ### In the text flat file In downloadable text format, evidence appears in braces at the end of the relevant annotation line, followed by the source. For a manually asserted function the pattern is the ECO code paired with a PubMed identifier for the supporting paper. For a similarity transfer it is the ECO code paired with the reviewed UniProtKB accession of the donor protein. For an automatic annotation it is the ECO code paired with the generating system, such as a UniRule identifier or an InterPro-derived pipeline. GO annotation lines carry the GO term, its aspect, a GO evidence abbreviation such as IDA for inferred from direct assay, and an ECO-based provenance with its source. Learning to scan flat files for these brace-delimited tags is a core skill for anyone building pipelines on UniProt data. ### In XML and REST output The XML format carries an evidence attribute on annotation elements, holding the ECO code, and separate provenance elements carry the source reference. The same information is available through the UniProt REST API, which returns JSON by default and accepts format parameters for other formats. ### Through search queries The UniProt search engine, used both on the website and in the REST API, supports a dedicated evidence field. A query of the form `ecoi:ECO_0000269` restricts results to entries carrying experimental manual assertions. Combining evidence queries with other fields is how most reproducible workflows pull curated subsets. ## Workflows: Using Evidence Codes in Real Analyses The following workflows cover the situations where evidence codes earn their keep. Each one is written as a repeatable procedure you can adapt. ### Workflow 1: Characterize an unknown protein using experimentally supported homologs Suppose you have a predicted protein from a newly sequenced organism and want the best available functional anchor. The goal is to find the closest reviewed entries whose annotations rest on experimental evidence, because those are the annotations worth transferring. 1. Prepare your [protein sequence](/blog/guides/protein-sequence) in FASTA format. 2. Search UniProtKB with the sequence using the BLAST service on uniprot.org, restricting the target to reviewed entries. Reviewing only reviewed hits removes the noise of unreviewed entries that merely repeat predictions. 3. For the top hits, open each entry and check the evidence tags on the FUNCTION and catalytic activity annotations. Prefer hits where the matching function carries ECO:0000269 or ECO:0000314 over hits where the same function carries ECO:0000256. 4. Record the taxonomic distance between your organism and the hit organism. A function transferred between closely related species is more defensible than one transferred across phyla. 5. In your notes or manuscript, cite both the reviewed entry and the evidence code, and quote the underlying reference when the code is experimental. ### Workflow 2: Download a curated, experimentally supported dataset For training sets, benchmarking, or pathway reconstruction, you often need every reviewed protein whose function of interest is experimentally supported. The REST API performs this in one call: ``` curl "https://rest.uniprot.org/uniprotkb/search?query=(reviewed:true)+AND+(ecoi:ECO_0000269)&format=tsv&fields=accession,protein_name,organism_name" ``` This returns a tab-separated table of reviewed entries carrying experimental manual assertions. To narrow to a specific annotation type or protein family, add terms to the query, for example a GO term, a keyword, or a gene name, joined with AND. To include only direct assay evidence, substitute ECO_0000314 in the evidence field. If your downstream tool needs full entries rather than a table, change the format parameter to `fasta` for sequences or `json` for complete structured records. Saving the query URL alongside your data makes the dataset reproducible, because anyone can rerun the same call against the current release and see what changed. A caution on dataset size: excluding electronic annotation shrinks results dramatically, since automatic systems annotate tens of millions of proteins [1]. For a small curated benchmark that shrinkage is the point. For a broad survey across poorly studied taxa, insisting on experimental evidence everywhere can leave you with no data at all, and a graded approach is usually better. ### Workflow 3: Annotate proteins from a newly sequenced genome Genome projects produce thousands of predicted proteins that need functional labels, and evidence codes give you a principled policy for which labels to adopt. 1. Predict protein-coding genes from your assembly and export the proteome as FASTA. 2. Build or download a local BLAST database of reviewed UniProtKB entries. A standard command for a local search is: ``` blastp -query proteins.fasta -db uniprot_sprot -outfmt "6 qseqid sseqid pident length evalue" -evalue 1e-5 ``` 3. For each predicted protein, examine the best reviewed hits and the evidence codes on their annotations. Accept an annotation only when the hit is a clear ortholog and the annotation carries experimental, phylogenetic, or well-constrained rule-based evidence. Record which code justified each transfer. 4. Where no reviewed match exists, run the unreviewed search as a secondary source, but tag those predictions in your own results as lower confidence so downstream users see the difference. 5. Assess the quality of the whole predicted proteome before you trust any of it. OMArk, a tool that compares a query proteome against precomputed gene families across the tree of life, evaluates both the completeness and the consistency of a gene repertoire relative to close relatives and reports likely contamination [2]. This step matters because quality problems upstream contaminate every annotation you transfer downstream. The published spider genome of Trichonephila antipodiana illustrates the scale and standard of this workflow. The project assembled a 2.29 Gb chromosome-level genome, predicted 19,001 protein-coding genes, found that 96.78 percent were supported by transcriptome evidence, and reported that 96.32 percent of the predicted proteins matched records in UniProt [3]. Those UniProt match rates, reported alongside transcript support, serve as a quality signal for the gene prediction itself, and they were only meaningful because the matching records carry interpretable provenance. ### Workflow 4: Audit a protein list before you cite or publish it ![Decision path asking whether the code is experimental, similarity based or electronic, and what wording each supports.](/article-images/illustrations/uniprot-evidence-codes-explained-which-code-to-cite.webp) Before a manuscript, a grant, or a database submission claims that a set of proteins performs particular functions, walk the evidence chain. 1. For each claim, open the UniProt entry and locate the exact annotation line, not just the entry as a whole. 2. Note the ECO code, the assertion type (manual or automatic), and the source. 3. For experimental codes, follow the PubMed reference and confirm the experiment matches the claim you intend to make. 4. For "By similarity" codes, confirm the cited donor entry itself carries acceptable evidence and that the organism distance is defensible. 5. For electronic codes, note the generating system and any constraints it applied. 6. Reword your own claims to match the evidence. An annotation supported only by inference should be described as inferred, not established. ## How Evidence Codes Get Assigned Understanding the assignment pipeline helps you predict where codes are likely to be strong or weak. Manual curation is the origin of the manual assertion codes. UniProt curators read published papers, verify sequence identity between the paper and the entry, and attach annotations with the specific ECO code matching the experimental method. This process scales with human effort, which is why the reviewed section covers a small fraction of the knowledgebase while the unreviewed section covers most of it. Similarity propagation extends curated knowledge. When a reviewed entry carries an experimentally demonstrated function, curators transfer that function to reviewed orthologs in related organisms with the ECO:0000250 tag, creating the "By similarity" system described earlier. Each transfer names its donor, so a reader can trace the chain back to the original experiment. Rule-based systems automate at scale. UniRule encodes curated decision logic that examines protein sequences against InterPro family signatures, applies taxonomic constraints, and, when conditions are met, propagates annotation from sets of reviewed proteins with experimentally supported functions to unreviewed records [1]. Because the rules are public, researchers can inspect the logic behind any UniRule-derived assertion rather than treating the annotation as a black box [1]. Alongside UniRule, other automatic pipelines, including similarity searches and machine-learned automatic rules, generate the ECO:0000256 bulk annotations that make unreviewed entries usable. ## Interpreting Evidence Codes in the Wild Interpretation is where training pays off, and it starts with a simple mental hierarchy. Experimental evidence for the specific protein outranks curated inference, which outranks constrained rule-based assertion, which outranks loose automatic annotation. But the hierarchy is a starting point, not a verdict, because several factors shift the weight. The first factor is specificity. ECO:0000314, a direct assay, is more specific and usually stronger than the umbrella ECO:0000269, which covers experimental support generally. When both appear, read the more specific one as the operative evidence. The second factor is the source. Two ECO:0000256 annotations can differ enormously. One may come from a UniRule with strict active-site requirements and taxonomic limits, the other from a broad similarity pass. The code alone cannot distinguish them, so the source field does the discriminating work. The third factor is the biology of transfer. Similarity-based and rule-based annotations assume functional conservation. That assumption weakens as taxonomic distance grows and as protein families expand, because duplicated genes routinely specialize after duplication. An assertion labeled "By similarity" in a protein from a lineage rich in recent duplications deserves more skepticism than the same assertion in a single-copy conserved enzyme. The fourth factor is the quality of the reference proteomes that feed the annotation ecosystem. Automated annotation propagates from reviewed sources outward, so errors in influential reference proteomes can spread through derived annotations. A genome-scale audit with OMArk across 1,805 UniProt eukaryotic reference proteomes found strong evidence of contamination in 73 of them, and identified error propagation in avian gene annotation traceable to a fragmented zebra finch proteome used as a reference [2]. That finding is a concrete demonstration of why evidence interpretation must extend beyond the individual entry to the provenance of the resources the entry was built from. When your annotation chain passes through a reference proteome of uncertain quality, the apparent strength of the evidence code overstates the real confidence. ## Common Mistakes and Limitations Avoid these specific, recurrent errors when working with evidence codes. - Confusing PE levels with evidence codes. PE level 4 (predicted) and ECO:0000256 (electronic annotation) often co-occur, but a PE level describes existence of the protein while evidence codes describe individual annotations. Filter on the right one for the question you are asking. - Treating reviewed status as a guarantee. Reviewed entries contain a mixture of evidence strengths. An experimental function plus a predicted location can sit in the same entry. Read the tags on each line you intend to use. - Reading only the code and ignoring the source. The code tells you the category of support. The source tells you which paper, rule, or donor protein backs it. Skipping the source loses most of the interpretive value. - Overextending "By similarity." Sequence similarity does not guarantee functional conservation, especially across large evolutionary distances or within expanded gene families. State the assumption when you rely on it. - Assuming all electronic annotation is equal. UniRule assertions, which fire only for proteins meeting curated family and taxonomic conditions, are not comparable to loose name assignments, yet both carry ECO:0000256 [1]. Check the generating system. - Filtering so strictly that coverage collapses. Requiring experimental evidence everywhere can leave poorly studied organisms with almost nothing. Match the stringency to the purpose: strict for benchmarks and mechanistic claims, graded for broad surveys. - Building on contaminated or inconsistent proteomes. Quality problems in source proteomes propagate into downstream annotation, as demonstrated for avian references [2]. Run a proteome-level quality check such as OMArk before transferring annotations at scale [2]. - Citing a stale snapshot. UniProt releases update continuously, and evidence can change when papers are reevaluated or rules are revised. Record the release date or the query URL in your methods so the result is reproducible. - Forgetting that interpretation requires domain judgment. Evidence codes inform but do not replace expertise in the relevant biology, and specific research or diagnostic decisions should be reviewed by a subject-matter expert before they are acted on. ## Quick Review - UniProt evidence codes are ECO ontology terms attached to individual annotations, not to whole entries. - The code records the method of support, the assertion type (manual or automatic), and a pointer to the source. - Experimental codes (ECO:0000269, ECO:0000314, ECO:0000353) are the strongest. "By similarity" (ECO:0000250) marks curated transfer from a reviewed homolog. ECO:0000256 marks automatic annotation and dominates the unreviewed section. - UniRule and related systems generate a large share of automatic annotation, covering 53 million proteins at release 2020_01 [1]. - PE levels (1 through 5) describe protein existence and are separate from evidence codes. - Evidence tags appear on the website, in brace-delimited flat file annotations, in XML attributes, and are searchable through the `ecoi` field in the REST API. - Quality of the upstream reference proteome bounds the value of any evidence code, which is why proteome-level audits such as OMArk belong in serious annotation workflows [2]. ## Frequently Asked Questions ### What does an evidence code in UniProt mean? It is a standardized tag from the Evidence and Conclusion Ontology that records how a specific annotation on a UniProt entry was derived, such as by direct experiment, by similarity to a characterized homolog, or by an automated rule. Each tag also points to its source, for example a PubMed reference or a reviewed donor entry. ### Are reviewed and unreviewed UniProt entries the same thing as evidence codes? No. Reviewed status describes the curation process for the entry as a whole, while evidence codes describe individual annotations inside any entry. Both reviewed and unreviewed entries can carry a mix of manually asserted and automatically assigned annotations. ### What is the difference between PE levels and evidence codes? PE levels, numbered 1 through 5, summarize the evidence for the physical existence of the protein, from experimentally demonstrated protein to uncertain existence. Evidence codes attach to individual functional annotations and use the ECO framework. You need both signals to interpret an entry fully. ### How do I download only experimentally supported entries from UniProt? Use the REST API with an evidence query, for example `query=(reviewed:true)+AND+(ecoi:ECO_0000269)`, and choose your output format and fields. On the website, the Evidence facet on search results applies the same restriction without writing a query. ### Can I trust automatic annotations marked ECO:0000256? It depends on the generating system, which is listed next to the code. UniRule assertions come from curated rules that require InterPro family signatures plus taxonomic constraints, so they are more constrained than loose similarity-based assignments, but neither carries the weight of direct experimental evidence [1]. ### How do evidence codes matter for annotating a new genome? They define which predicted annotations you should accept and how you should label them. Match predicted proteins against reviewed UniProt entries, transfer annotations only from experimentally or rule-supported homologs, record the evidence code for each transfer, and run a proteome quality check such as OMArk, because problems in the predicted proteome or its references propagate into every downstream claim [2][3]. ## Sources 1. [UniRule: a unified rule resource for automatic annotation in the UniProt Knowledgebase.](https://pubmed.ncbi.nlm.nih.gov/32399560/) 2. [Quality assessment of gene repertoire annotations with OMArk.](https://pubmed.ncbi.nlm.nih.gov/38383603/) 3. [A chromosome-level genome of the spider Trichonephila antipodiana reveals the genetic basis of its polyphagy and evidence of an ancient whole-genome duplication event.](https://pubmed.ncbi.nlm.nih.gov/33739402/)