Data Review and Clean-Up Using Crowdsourced Input via the US EPA CompTox Dashboard
Download as PPTX, PDF0 likes188 views
The document discusses the EPA's CompTox Dashboard, which provides access to extensive chemical data including experimental and predicted properties for approximately 762,000 chemicals. It emphasizes the importance of data quality and the role of crowdsourced user feedback in curating and improving toxicological data. Additionally, it highlights efforts to enhance data accuracy and usability through collaboration and systematic review processes.
Data Review and Clean-Up Using Crowdsourced Input via the US EPA CompTox Dashboard
- 1. Data Review and Clean-Up Using Crowdsourced Input via the US EPA CompTox Dashboard Antony Williams1, Katie Paul-Friedman1, Jason Brown2, Chris Grulke1, Emma L. Schymanski3 and Jeff Edwards1 1) National Center for Computational Toxicology, U.S. Environmental Protection Agency, RTP, NC 2) ORAU 3) Luxembourg Centre for Systems Biomedicine (LCSB), University of Luxembourg, Campus Belval, Luxembourg August 2018 ACS Fall Meeting, Boston http://www.orcid.org/0000-0002-2668-4821 The views expressed in this presentation are those of the author and do not necessarily reflect the views or policies of the U.S. EPA
- 2. Overview • An introduction to the EPA’s CompTox Dashboard • Our Data: experimental and predicted property data, high-throughput screening assay data and hazard and environmental exposure data, and lots more • Data quality: high-quality data needed but challenging to produce – Millions of individual data points and annotations – 10’0s of 1000’s of chemicals • What is the role of user feedback? • Our efforts to curate our ToxCast bioassay data 1
- 4. The CompTox Chemistry Dashboard • A publicly accessible website delivering access: – New entry portal for all NCCT dashboards – ~762,000 chemicals with related property data – Searchable by chemical, product use, gene and assay (ToxCast) – Experimental and predicted physicochemical property data – “Bioactivity data” for the ToxCast/Tox21 project – Generalized Read-Across (GenRA) module – Links to other agency websites and public data resources – “Literature” searches for chemicals using public resources – “Batch searching” for thousands of chemicals – DOWNLOADABLE Open Data for reuse and repurposing 3
- 7. CompTox Dashboard Products and Use Categories 6
- 8. CompTox Dashboard Assays and Genes 7
- 11. Access to Chemical Hazard Data 10
- 12. Hazard Data from “ToxVal_DB” • ToxVal Database contains following data: –30,050 chemicals –772,721 toxicity values –29 sources of data –21,507 sub-sources –4585 journals cited –69,833 literature citations 11
- 13. In Vitro Bioassay Screening ToxCast and Tox21 12
- 14. In Vitro Bioassay Screening ToxCast and Tox21 13
- 15. How can be curate our data? • Crowdsourcing is well proven nowadays • Comments can be added at a record level • Submitted comments are reviewed by administrators and responded to 14
- 17. Reviewer comments are public 16
- 18. 17 MassBank/CompTox Curation of External Data o A “nice” example: 4-4'-Bis(2-sulfostyryl)biphenyl Purchased: CAS: 27344-41-8 DTXSID6036467 Registered: CAS: 38775-22-3 (UFZ) DTXSID7047017
- 19. Comments to date • The majority of comments to date: – Structure and names/CASRN do not match – Add additional synonyms – Request to add specific property data – Structure layout/depiction needs improving 18
- 20. Crowdsourcing Comments Single Cell Commenting added • Highlight an alphanumeric text string 19
- 22. Bioactivity Data • 100s of thousands of bioactivity curves to review • Impossible to review every one manually • Now accepting public Crowdsourced Comments • Public crowdsourcing will not suffice!!! 21
- 23. Internal Review of 25,000 curves 22 Screenshot of entry page for Beta R Shiny Application for NCCT users Brown & Paul-Friedman
- 24. Internal Review of 25,000 curves A “good fit” bioactivity curve 23
- 25. Internal Review of 25,000 curves 24 …and gain-loss fit with a cell viability assay (makes little sense) Single-Point in middle of concentration range drives ACTIVE Hit Call
- 26. Internal Review of 25,000 curves Abnormally High-Noise 25
- 27. Internal Curve Review Results • Internal curve review has resulted in: – Instances of correction of fitting procedures in the ToxCast Pipeline – Identification of issues with source data – Identification of additional flags or filters that could be used, depending on the application of ToxCast data – a beta implementation of quality assurance for HTS data – Brown & Paul-Friedman, Uncertainty in ToxCast Curve-Fitting: Quantitative and Qualitative Descriptors Inform a Model to Predict Reproducible Fits (in preparation) 26
- 28. The CompTox Dashboard for Structure Identification by MS 27
- 29. Collaborative Data Curation • Mapping between our data (and websites) has resulted in collaborative data curation • Collaboration with Emma Schymanski re. the NORMAN Suspects Exchange https://www.norman-network.com/?q=node/236 • Our process for mapping data is iterative 28
- 31. Example: NORMAN Priority List 30
- 32. Mapping on Two Identifiers 31
- 33. Mapping on Two Identifiers 32
- 34. Mapping Quality Control (I) 33
- 35. Mapping Quality Control (II) 34
- 36. Example: NORMAN Priority List 35
- 37. Example: NORMAN Priority List 36
- 38. >23 NORMAN Lists Available 37
- 39. Progressive Curation on #Lists 38(many) more registrations… Cleaning up lists to remove errors Undefined mixtures (UVCBs)
- 40. Conclusion • The CompTox Dashboard provides access to data for ~760,000 chemicals • Crowdsourced comments from users may help clean data • Record level data curation - any alphanumeric string • Improved data quality provides better data for modeling to underpin our prediction algorithms • Collaborative data curation benefits extended communities of users • Data curation is never complete – our data expands daily… 39
- 41. Contact Antony Williams US EPA Office of Research and Development National Center for Computational Toxicology (NCCT) Williams.Antony@epa.gov ORCID: https://orcid.org/0000-0002-2668-4821 40
- 42. 41 NORMAN Suspect Exchange o http://www.norman-network.com/?q=node/236
Editor's Notes
- #28: Comments from Katie: In terms of verbal comments, you might note that we are simply reviewing the uncertainty related to the curve-fitting itself. Sometime people conflate uncertainty in response, e.g., assay interference due to autofluorescence or cytotoxicity, with curve-fitting uncertainty. These are two distinct sources of uncertainty. An autofluorescent compound should produce a beautiful curve in an assay that measures fluorescence as an indirect marker of some bioactivity. These curves will get an “active” score. When filtering the data for use – say you only want to know the chemicals that definitively activate some receptor – then you might want to layer in not only the reproducibility of the curve fitting, but also other information such as the cytotoxicity burst, information on autofluorescence or nonspecific protein inhibition, etc. This is probably elementary to you, but it is a point I like to make because so many people think that we should auto-filter data for these different sources of assay interference. I am okay with pre-set options to select for filtering, but different applications will require different levels of stringency in terms of filtering.