UniProt: a worldwide hub of protein knowledge

UniProt: a worldwide hub of protein knowledge The UniProt Consortium

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

UniProt: a worldwide hub of protein knowledge 2018 Nucleic Acids Research 10.1093/nar/gky1049 dogan2018-uniprot-a-worldwide-hub-of-protein dogan2018-uniprot-a-worldwide-hub-of-protein.txt, the project's staged copy 729d469428e47dab11ea367b301c26c90f6e1177 cc-by genetics

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them the UniProt Knowledgebase itself - 120 million protein records, of which just over half a million carry annotations a human curator read out of the literature Quotation withheld because this record is not cleared for redistribution. how to keep a knowledge base current, provenance-tracked and correct while the sequence data entering it grows exponentially Quotation withheld because this record is not cleared for redistribution. a two-year update: expanded reference proteomes especially for viruses, new rule-based and association-rule annotation systems, a bibliography layer built from other databases and text mining, and a licence change to CC-BY 4.0 Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. the gap between what has been sequenced and what has been read - 120 million entries against half a million with curated annotation, and expert curation reaching 35-50% of the curatable literature Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. why one wrong curated record matters far beyond itself: functional data is computationally transferred from Swiss-Prot to sequence-related but less studied proteins in TrEMBL, so a correction upstream stops misleading data being added to orthologues on the strength of an over-predictive signature annotation propagation Quotation withheld because this record is not cleared for redistribution. the growth problem is not more biology but more of the same: 90% of proteins have over 90% identity because sequencing is driven by very similar strains of the same species redundancy from near-identical strains Quotation withheld because this record is not cleared for redistribution. the premise of the DAAC predictor - that function depends not only on WHICH domains a protein has but on their arrangement and order, so two proteins with the same domain set can do different things domain architecture Quotation withheld because this record is not cleared for redistribution. how a cluster-level GO term is earned: it is assigned only when every annotated member carries it, or when it is a common ancestor of at least one term from each - agreement across the cluster rather than a majority conservative propagation Quotation withheld because this record is not cleared for redistribution. expert biocuration reaches only 35-45% of the curatable literature for a species, 50% for humans - the coverage gap is measured and published rather than left implicit Quotation withheld because this record is not cleared for redistribution. curation is labour intensive and the authors defend it anyway as the most efficient way to get everything out of a paper - which is a statement about what automation has NOT replaced Quotation withheld because this record is not cleared for redistribution. keeping a record current is hard because knowledge of protein function evolves and new techniques produce results that may appear to contradict earlier observations Quotation withheld because this record is not cleared for redistribution. MECHANISMS TO IDENTIFY AND HIGHLIGHT CONTRADICTORY INFORMATION IN EXISTING ENTRIES ARE A FUTURE PLAN, NOT A CAPABILITY - a 120-million-record knowledge base with per-claim evidence still cannot systematically surface where its own entries disagree Quotation withheld because this record is not cleared for redistribution. redundancy removal is not applied to viruses at all, because their small genomes carry too many variants - a stated exemption with a stated reason Quotation withheld because this record is not cleared for redistribution. UPCLASS, the classifier applied to 32 million accession-PMID pairs, has a published recall of 0.62 - so roughly a third of the classifications it should make, it does not Quotation withheld because this record is not cleared for redistribution. the categorisation of computationally mapped bibliography by source database is described by the authors as not optimal for all literature, which is why the classifier was brought in Quotation withheld because this record is not cleared for redistribution. an evidence tag on every individual piece of captured knowledge, naming its source eliminates: a claim in the database being untraceable to where it came from - and it is not restricted to the curated half: automatic annotations carry which system produced them · outcome: held Quotation withheld because this record is not cleared for redistribution. a CAUTION comment recording that a previously reported activity is an unsafe observation, kept on the entry alongside the annotation eliminates: a superseded finding either persisting silently or disappearing without trace - the reader gets the claim, the doubt and the reason together · outcome: held Quotation withheld because this record is not cleared for redistribution. redundancy removal that moves ~38% of complete proteomes out of UniProtKB and into UniParc, which retains historical obsolete sequence eliminates: growth in near-identical strains swamping the knowledgebase - WITHOUT losing the sequences, which is the difference between a withdrawal and a deletion · outcome: held Quotation withheld because this record is not cleared for redistribution. assigning a GO term to a UniRef cluster only when EVERY annotated member has it, or when it is a common ancestor of at least one term from each eliminates: a single member's annotation propagating to a whole cluster - the rule requires unanimity or a shared ancestor, not a majority · outcome: held Quotation withheld because this record is not cleared for redistribution. publishing UPCLASS's precision, recall and F-score before applying it at scale eliminates: the classifier's error rate being invisible to users of the 32 million pairs it classified - PARTIAL, because knowing recall is 0.62 does not tell a user WHICH third is missing · outcome: partial Quotation withheld because this record is not cleared for redistribution. prioritising for curation exactly what computational prediction gets WRONG - proteins a signature would place in an enzyme family but which are non-functional through the loss of specific residues eliminates: human effort being spent where the automatic systems already succeed, and leaves the curators on the cases where a prediction would be confidently wrong · outcome: held Quotation withheld because this record is not cleared for redistribution. admitting metagenome-assembled genomes only when they show high completeness and low contamination, and removing segmented viruses missing one or more segments eliminates: partial or contaminated assemblies entering the knowledgebase as if they were whole proteomes · outcome: held Quotation withheld because this record is not cleared for redistribution. re-importing curated data from the domain-specific resources it was contributed to - IMEx, the Complex Portal, GO - rather than duplicating their maintenance eliminates: the same fact diverging between UniProt and the specialist database that owns it; each is stored and maintained where it belongs and flows back · outcome: held Quotation withheld because this record is not cleared for redistribution. UniProtKB UniProtKB Quotation withheld because this record is not cleared for redistribution. annotations annotations Quotation withheld because this record is not cleared for redistribution. evidences evidences Quotation withheld because this record is not cleared for redistribution. expert biocuration expert biocuration Quotation withheld because this record is not cleared for redistribution. CAUTION CAUTION Quotation withheld because this record is not cleared for redistribution. UniParc UniParc Quotation withheld because this record is not cleared for redistribution. redundancy removal redundancy removal Quotation withheld because this record is not cleared for redistribution. InterPro InterPro Quotation withheld because this record is not cleared for redistribution. GO annotation GO annotation Quotation withheld because this record is not cleared for redistribution. UPCLASS UPCLASS Quotation withheld because this record is not cleared for redistribution. UniFire UniFire Quotation withheld because this record is not cleared for redistribution. curatable literature curatable literature Quotation withheld because this record is not cleared for redistribution. METTL14 METTL14 Quotation withheld because this record is not cleared for redistribution. METTL3 METTL3 Quotation withheld because this record is not cleared for redistribution. MAGs MAGs Quotation withheld because this record is not cleared for redistribution. CAUTION → METTL14 A RECORD HOLDING A CLAIM AND A DOUBT ABOUT THAT CLAIM AT ONCE. The entry keeps the reported methyltransferase activity AND a comment saying the report is an unsafe observation. There is no edge type for 'this annotation is flagged as unreliable' - all eight assert that a relation holds, and the whole content of a CAUTION is that something asserted elsewhere should not be relied on. This is the project's quine1951 finding with a WORKING IMPLEMENTATION on the other side: UniProt can write it and we cannot Quotation withheld because this record is not cleared for redistribution. expert biocuration → curatable literature A COVERAGE RATIO, WHICH IS A RELATION BETWEEN AN ACTIVITY AND WHAT IT HAS NOT REACHED. Curation captures 35-45% of the curatable literature per species, 50% for humans. 'part-of' is wrong, 'causes' is wrong, and there is no type for 'X covers this fraction of Y' - which is the single most useful thing a knowledge base can say about itself and the thing our own corpus has never stated about its own coverage Quotation withheld because this record is not cleared for redistribution. reviewed versus unreviewed - NOT A QUALITY GRADE BUT A PROVENANCE CLASS, and the two are wildly different sizes. `Reviewed` means a curator assimilated multiple lines of evidence from a paper's text and figures; `unreviewed` means a rule fired. Half a million against 120 million. A user filtering to `reviewed` is not filtering out errors, they are filtering to the fraction a person has read Quotation withheld because this record is not cleared for redistribution. evidence, in UniProt - A SOURCE TAG ON EVERY INDIVIDUAL CLAIM, not a citation on the record. Each piece of knowledge carries where it came from, and automatic annotations carry which system produced them. This is the same commitment this corpus makes with its `evidence` field, made by a database three orders of magnitude larger Quotation withheld because this record is not cleared for redistribution. CAUTION - A COMMENT TYPE THAT SAYS A PUBLISHED FINDING IS UNSAFE. When structural work showed METTL14 has a degenerate active site, the earlier reported methyltransferase activity was not deleted and not silently overwritten: a CAUTION comment was added to tell users the initial report is `an unsafe observation`. THE RECORD CARRIES BOTH THE CLAIM AND THE DOUBT Quotation withheld because this record is not cleared for redistribution. removed, of a proteome - MOVED TO AN ARCHIVE, NOT DELETED. Redundancy removal has taken about 38% of complete proteomes out of UniProtKB and placed their sequences in UniParc, which holds a complete set of known sequences including historical obsolete ones. A withdrawal convention at the scale of 241 million proteins Quotation withheld because this record is not cleared for redistribution. prediction precision, of the UniRef GO propagation - MEASURED AGAINST ITS OWN EARLIER OUTPUT. The >97% figure comes from comparing sequences that newly acquired a GO annotation with ones the same system annotated in previous releases. That is self-consistency across releases, which is worth knowing and is not external validation against experiment Quotation withheld because this record is not cleared for redistribution. curatable literature - THE DENOMINATOR THEY MEASURE THEMSELVES AGAINST, and it is not the literature. Coverage of 35-45%, rising to 50% for humans, is a share of the papers judged curatable for a species, so a paper outside that judgement is not counted as missed Quotation withheld because this record is not cleared for redistribution. contradicts → the initial characterisation of METTL14 as a methyltransferase, based on reported activity and MT-A70 family membership note: 'contradicts' is right here and it was not for cruz-jentoft2018's mass criterion. The paper does not demote the earlier claim, it reports that structural work showed the active site is degenerate and cannot accommodate the substrates - the earlier finding is not narrowed, it is wrong, and UniProt flags it as unsafe rather than deleting it Quotation withheld because this record is not cleared for redistribution. extends → the UniProt Consortium's own previous database update, published in the same journal in 2017 Quotation withheld because this record is not cleared for redistribution. builds-on → InterPro and its member databases PROSITE, SMART and Pfam, whose predictive signatures classify sequences and predict domains for the unreviewed records Quotation withheld because this record is not cleared for redistribution. predicts → the authors of future papers, who are asked to write protein mentions as (UniProtKB: P68369) so curators and text-mining tools can connect their work note: an instruction to a literature that does not exist yet, and the paper defers to any journal that already has its own format - a standard proposed with its own exception attached Quotation withheld because this record is not cleared for redistribution. builds-on → the specialist resources this data is contributed to and re-imported from - the IMEx Consortium, the Complex Portal and the Gene Ontology Quotation withheld because this record is not cleared for redistribution. annotations —part-of→ UniProtKB annotations UniProtKB strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. evidences —part-of→ annotations evidences annotations strength of the relation: 0.95 · note: RETYPED IN THE AUDIT from 'annotations --operationalised-as--> evidences', where my own note conceded it was not a measurement relation. The quote says every piece of knowledge is ASSOCIATED WITH evidences, which is attachment - and attachment is what 'part-of' says Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span writes 'every piece of knowledge we capture is associated with evidences', not 'annotations'. Quotation withheld because this record is not cleared for redistribution. CAUTION —part-of→ annotations CAUTION annotations strength of the relation: 0.9 · note: a comment type carried on an entry beside its annotation - the mechanism by which a record holds a claim and a doubt about that claim at the same time Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span writes 'a CAUTION comment to the METTL14 entry', not 'annotations'. Quotation withheld because this record is not cleared for redistribution. METTL3 —distinct-from→ METTL14 METTL3 METTL14 strength of the relation: 1 · note: and the distinction is the whole point: they are both MT-A70 family members and both were reported to have methyltransferase activity. Only one has a working active site Quotation withheld because this record is not cleared for redistribution. redundancy removal —causes→ UniProtKB redundancy removal UniProtKB direction on the measure: reduces · strength of the relation: 0.9 · note: and the reduction is the point - the paper reports that it made growth scalable rather than merely smaller Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: 'this process' = the redundancy removal process (previous sentences). Quotation withheld because this record is not cleared for redistribution. 120 million protein entries, of which just over half a million carry annotations extracted from the literature by expert curators Quotation withheld because this record is not cleared for redistribution. expert biocuration reaches 35-45% of the curatable literature per species, 50% for humans Quotation withheld because this record is not cleared for redistribution. Redundancy removal has taken about 38% of all complete proteomes (about 241 million proteins) out of UniProtKB. Quotation withheld because this record is not cleared for redistribution. Virus reference proteomes stood at 428 at the previous publication, about an order of magnitude below the 4405 known viral species. Quotation withheld because this record is not cleared for redistribution. Structural studies showed only METTL3 is catalytic; METTL14 has a degenerate active site and a non-catalytic role. Quotation withheld because this record is not cleared for redistribution. the bibliography layer adds 902,956 unique papers over 363,591 entries, and 36,147,401 accession-PMID pairs Quotation withheld because this record is not cleared for redistribution. UniProt data moved to a CC-BY 4.0 licence to make reuse easier - stated in the paper itself Quotation withheld because this record is not cleared for redistribution. UniProtKB = reviewed Swiss-Prot entries curated by an expert team, plus unreviewed TrEMBL entries annotated by automated and rule-based systems UniProt Knowledgebase Quotation withheld because this record is not cleared for redistribution. evidences = the source tag attached to every individual piece of captured knowledge evidences Quotation withheld because this record is not cleared for redistribution. CAUTION = a comment telling users that an initially reported finding is an unsafe observation CAUTION comment Quotation withheld because this record is not cleared for redistribution. UniParc = the archive holding a complete set of known sequences including historical obsolete ones, where redundancy-removed proteomes go UniProt Archive Quotation withheld because this record is not cleared for redistribution. UPCLASS precision 0.80, recall 0.62, F-score 0.70 - published before being applied to 32 million accession-PMID pairs Quotation withheld because this record is not cleared for redistribution. UniRef GO propagation precision above 97% in all three GO domains - measured by comparing newly annotated sequences against the system's own earlier releases Quotation withheld because this record is not cleared for redistribution. 90% of proteins have over 90% identity to another, which is what makes the growth a redundancy problem rather than a knowledge one Quotation withheld because this record is not cleared for redistribution. the two rule systems: UniRule and the Statistical Automatic Annotation System rule-based prediction systems Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1093/nar/gky1049.