Desperately seeking DARCP
Download as PPTX, PDF0 likes74 views
This document discusses the extraction of key relationships (D-A-R-C-P) reported in biomedical literature where a bioactivity (A) and result (R) are reported for a chemical structure (C) that modulates a protein target (P). It analyzes the statistics of DARCP entity accumulation from three manually curated databases and compares it to PubChem. While public databases have captured around 18% of known human protein targets, commercial databases have captured around 4 times more DARCP relationships through greater curation resources. The future of DARCP extraction depends on increased natural language processing, open access policies, and databases facilitating the input of these relationships.
Desperately seeking DARCP
- 1. Desperately seeking curated D-A-R-C-P: Assessing the past to predict the future Introduction Bioscientists reading papers or patents on bioactive chemistry strive to discern the key relationships reported within a document “D“ (e.g. with a PubMed ID) where a bioactivity “A” with a quantitative result “R” (e.g. an IC50) is reported for chemical structure “C” that modulates (e.g. inhibits) a protein target “P” (e.g. a UniProt ID). D – A – R – C – P While it cannot encompass all mechanistic cases a useful shorthand for this connectivity thus becomes DARCP. Biocuration for extraction and structured capture of this relationship chain in databases has high value that can be explored both manually and computationally, viz; • “D”: clustering by relatedness, entity content, citation networks, connections via authors and institutions • “A”: classified by various assay ontologies • “R”: log transformations (e.g. pIC50 or pKi) for potency ranking and SAR, sorting by molecular mechanism of action (mmoa), (e.g. where A-R indicates C to be a potent inhibitor of P) • “C”: the full range of cheminformatic analysis including 2D/3D clustering, property prediction, substructures, analogue searching and chemical ontologies • “P” a full range of bioinformatic analysis including; target classes, Gene Ontology (GO) assignments, pathway annotation, structural homology, disease associations and genetic variation (e.g. for target validation). The problem the community faces is that we have spent millions burying DARCP in paywalled PDFs (a.k.a. “Hamburgerisation”) over many decades but must now spend millions more trying to get it back out. Assessing the past The table below shows the statistics of DARCP entity accumulation from three manually curated resources over approximately the last decade. In the table these are compared with PubChem wherein these four are integrated as submitting sources (GtoPdb = IUPHAR/BPS Guide to Pharmacology, PMID 31691834). Statistical comparisons between databases can be confounded by differences in their data models, publication selectivity, curatorial practice and activity thresholds. Nonetheless, discrete entity count can be informative for assessing relative extraction capture of documents, structures and proteins. The DCP counts are shown below for the three sources. PubMed IDs PubChem CIDs Swiss-Prot human IDs Christopher Southan, TW2Informatics, Göteborg, Sweden 41266 Interpreting entity count differences The capture of PMIDs shows a pattern of intersects and differences that is to some extent also reflected in chemistry and protein targets. Each source has some unique capture but ChEMBL and BindingDB overlap for ~25K papers (partially due to collaborative mirroring between them). The total from all four of is ~75K PMIDs. The chemistry (as PubChem identifiers) shows similar disproportionation with ChEMBL, as expected, dominating with unique content of ~1.2 million. While this is skewed by their BioAssay subsumation of ~0.5 million, most has been extracted from ~35K unique papers. In BindingDB unique structures are mainly from SAR curation of US Patents. In terms of interpreting difference we should also note that GtoPdb extract on average ~ 1 lead compound per-paper, ChEMBL ~14 per-paper and BindingDB ~ 40 per-patent. For the differences in target coverage (i.e. as “P” in DARCP) further work is needed to know what selectivity causes this divergence (e.g. journal choice) but some BindingDB unique proteins are patent-only. While exploring further causes of target divergence are outside the scope of this work, the total of 3745 human proteins (with A-R-C modulating chemistry) covered by these three, represents ~18% of the UniProt proteome of 20,365. So how much could be captured? While an upper limit is difficult to assess, commercial DARCP extraction sources such as Exelra GOSTAR and Reaxys Medicinal Chemistry, declare curated entity counts in the range of 6-8 million activity-mapped compounds from ~200-350,000 papers plus ~70-130,000 patents. They also count over 10,000 targets (but not all as protein identifiers). While there are caveats with comparisons (i.e. not counting the entities in exactly the same way and no disclosure of entities-in-common) the indication is that these two sources have captured (very roughly) 4-fold more DARCP than public efforts, largely due to the larger number of curators employed or contracted. However, in terms of upper limits for public capture, we must not overlook issues of data reproducibility arising from the increasingly patchy quality of PubMed (i.e. many papers from which DARCP should perhaps not be extracted). Predicting the future The future flow of DARCP into databases is constrained by the following factors; • The three resources that continue to capture the majority of open DARCP are to be congratulated and we hope their funding will be sustained. However, their capacity is limited by the number of biocurators in the face of increasing bioactivity publications (and which cheminformatics AI may accelerate). • Progress in entity recognition via Natural Language Processing now means that the extraction of discrete D,A,R,C, and P per se can be automated with reasonable specificity as well as indexed by resource look-ups in European PubMed Central (EPMC). However, this has not been achieved for D-A-R-C-P relationships that biocurators can discern and extract from documents in minutes. • The good news on the journal front is that we have J.Med.Chem. supplementary SMILES listings (occasionally even with activities), Nat. Chem.Biol pointing to PubChem entries and Brit J. Pharmacol. incorporating GtoPdb out-links and (via those) links to PubChem. The bad new is we will move into 2020 without even a single journal (from 1000s across the domains of medicinal chemistry, drug discovery, pharmacology and chemical biology) facilitating author-specified explicit DARCP automatically piped to databases (e.g. PubChem BioAssay). • The FAIR initiative (Findable, Accessible Interoperable, Reusable) is gaining momentum and should lead to at least discrete D,A,R,C,P annotations flowing into various repositories, However, the proportion of fully connected D-A-R-C-P may be low and it is unclear technically how this might flow through to major databases. For example, there is currently neither push nor pull for DARCP to flow from Figshare into PubChem BioAssay. • While Open Access and Plan S are also gaining momentum, paywalls still seriously impede extraction. The legacy problem is that only 14% of the ~62K papers extracted by ChEMBL (as indexed in EPMC) are free full text. • For the future we need publications to facilitate FAIR data extraction. Non- document surfacing (e.g. open Electronic Notebooks and Wikidata) also needs encouraging as an alternative to journals. Both trends should increase DARCP flow into open databases to enable big data mining and knowledge distillation. N.b. additional details from this work are given in a ChemRxiv preprint (10.6084/m9.figshare.11295323) that is under consideration by a journal. https://sites.google.com/view/tw2informatics/home