Calling on a million minds for community annotation in WikiProteins

Abstract

WikiProteins enables community annotation in a Wiki-based system. Extracts of major data sources have been fused into an editable environment that links out to the original sources. Data from community edits create automatic copies of the original data. Semantic technology captures concepts co-occurring in one sentence and thus potential factual statements. In addition, indirect associations between concepts have been calculated. We call on a 'million minds' to annotate a 'million concepts' and to collect facts from the literature with the reward of collaborative knowledge discovery. The system is available for beta testing at http://www.wikiprofessional.org.

Figure 1

PubMed grew beyond 14,000,000 abstracts in 2006 (by the end of 2007 the 17,000,000 mark was passed). In 2006, UMLS contained well over 1,300,000 concepts. Only 185,262 concepts from UMLS were actually mentioned in PubMed (2006 version) and, therefore, the concept space of the entire PubMed corpus could be captured in just over 185,000 Knowlets.

Figure 2

Any concept in the biomedical literature - for instance, a protein or a disease - can be treated as a source concept (depicted as a blue ball throughout the picture and the system). There may be curated information in authoritative databases such as UMLS or UniProtKB/Swiss-Prot concerning the concept and its factual relationships with other concepts. This information is captured and all concepts that have a 'factual' relationship with the source concept in any of the participating databases are thus included in the Knowlet of that concept. These 'factually associated concepts' are depicted in the Knowlet visualisation as solid green balls. In addition, the source concept may be mentioned with other concepts in one and the same sentence in the literature. In that case, especially when there are multiple sentences in which the two concepts co-occur, there is a high chance for a meaningful, sometimes causal, relationship between the two concepts. Most concepts that have a factual relationship are likely to be mentioned in one or more sentences in the literature at large, but as we have mined only PubMed so far, there might be many other factual associations that are not easy to recover from PubMed abstracts alone. For instance, many protein-protein interactions described in UniProtKB/Swiss-Prot cannot be found as co-occurrences in PubMed. Target concepts that co-occur minimally once in the same sentence as the source concept are depicted as green rings in the visualisation of the Knowlet. The last category of concepts is formed by those that have no co-occurrence per sentence in the indexed resources but have sufficient concepts in common with the source concepts in their own Knowlet to be of potential interest. These concepts are depicted as yellow rings and could represent implicit associations. Over one million Knowlets have been created so far. Each source concept has a relationship of varying strength with other (target) concepts and each of these distances has been assigned with a value for factual (F), co-occurrence (C) and associative (A) parameters. All Knowlets are dynamically coupled into the concept space. The semantic association between each concept pair is computed based on these values. In the near future additional data will be added, such as co-expression statistics between genes.

Figure 3

The WikiProteins Concept page of the CLB2 gene and its known formal synonyms (data from UniProtKB/Swiss-Prot as the authoritative database). Highlights are concepts recognized on the fly in the page that are linked to the corresponding Concept pages in the Wiki, to PubMed records, and to the concept space. Multiple terms selected in the page will create an AND query in external sources such as PubMed or a composite Knowlet with the selected concepts as source concepts (Figures 5-9). New co-occurrences on a given Wiki page due to edits by the community will be reported. Terms that represent concepts but are not recognized by the indexer can be added to the terminology system by selecting the terms in the text, starting a new Wiki page and defining the concept.

Figure 4

The basics of community annotation and semantic support. Once Knowlets have been created from authoritative sources and the indexed literature, a regular re-computing of the concept space with all changed semantic associations is performed. In case new co-occurrences, stated facts or significant associations emerge from the computational process, all experts that have expressed an interest in that part of the concept space will be alerted. Pre-constructed Knowlets for over one million authors have been created who currently have publications in PubMed. When they comment in the Wiki, their contributions will automatically be indexed and processed, forming an additional source for Knowlet enrichment alongside the classic literature and databases. UniProtKB/Swiss-Prot, GO Annotation, IntAct and UMLS have indicated that they wish to use the system as a source for accelerated annotation in their respective information resources.

Figure 5

A total of 26 concepts are recognized by the Peregrine tagger (2007) in abstract 15920482 from PubMed (see first column). The associated concepts in the composite Knowlet of these concepts include those that are expected, as discussed in the main text.

Figure 6

The Knowlet-based connections of four yeast proteins. (a) The composite Knowlet of the four yeast proteins as indicated in the text. (b) When the (factually connected) concept 'mitosis' is selected for explanation in the Knowlet, the factual association appears to be based on GO annotations. (c) Multiple co-occurrences are also found with more than one source concept including S. cerevisiae and CDC28. (d) In addition, there are multiple concepts that indirectly connect the source concepts with cell division. This means that the original example sentence used for this case study would have been repeatedly retrieved as relevant in the 'explain' window, supporting by co-occurrence the semantic association between the proteins involved.

Figure 7

Screenshot of the broad concept space of the concept 'antimalarials' and 46 actual antimalarial drugs. The closest, non-co-occurring drug is tegafur, which was explored further (see main text).

Figure 8

The Knowlet created with the source concept 'tegafur' with TS as a closely associated concept.

Figure 9

An artist's impression of the all-to-all semantic association matching of selected concepts followed by two-dimensional visualization. Several techniques for visualization of the concept space by standard techniques such as multi-dimensional scaling are currently under development.