Many of you will have noticed a new section on the ChEMBL interface, specifically at the Document Report Card page, called Related Documents. It consists of a table listing the links for up to 5 other ChEMBL documents (i.e. publications aka papers) that are scored to be the most similar to the one featured in the report card. Here's an example.
How does this work? There are examples of related documents sections online, e.g. in PubMed or in various journal publishers' websites. Document 'related-ness' or similarity can be assessed by comparing MeSH keywords or by clustering documents using TF-IDF weighted term vectors. Fortunately, ChEMBL puts a lot of effort in manually extracting and curating the compounds and biological targets from publications, so why not using these as descriptors to assess document similarity instead - as far as we know this is the first time this approach has been implemented?
So, here's how it works:
Firstly, for each document in ChEMBL, its list of references is retrieved using the excellent EuropePMC web services. By considering documents as nodes which are connected with an edge if one paper cites the other, a directed graph structure emerges. By doing this for all ~50K documents in ChEMBL, you get the massive graph illustrated above in Cytoscape. As a bonus, by measuring the in- and out- degree of the nodes, one could check which are the most cited papers in ChEMBL - but that's the topic of another blog post. This graph could be further annotated with protein target families, authors and institutions, as it has been elegantly done here.
Moving on, once a relationship between two documents is established, we need a way to quantify their similarity. As hinted above, we used the normalised overlap of compounds and targets reported in the two documents. This is done using the classic Tanimoto coefficient, so if doc A reports compounds (1,2,3) and doc B reports compounds (3,4,5), their compound Tanimoto similarity T is 1/5 or 0.2. Exactly the same applies for the target-based document similarity. The composite score we use to rank docs in the Related Documents section is simply the maximum of the two individual ones.
What does all that mean in practice? It means that 2 papers are listed as similar if they their reported compounds or biological targets overlap significantly (and one cites the other). For example, papers with follow-up experiments on the same candidate drug will be deemed similar, e.g. this one. The same will apply to two papers that involve kinase panel screening assays. A desirable side-effect is that by following the links, the tenacious user may traverse the whole graph displayed above!
George & Mark