Pharos – A Torch to Use in Your Journey In the Dark Genome
2 likes718 views
This document summarizes a webinar presentation about the Illuminating the Druggable Genome (IDG) Knowledge Management Center (KMC) and its Pharos interface. The presentation discusses how Pharos provides access to integrated data on over 20,000 protein targets to help characterize well and poorly studied targets. It highlights how Pharos analyzes target data using methods like knowledge availability scoring and similarity in knowledge space to help prioritize "dark" targets lacking extensive research that may have therapeutic potential. The long-term vision is for Pharos to generate customized, semi-natural language summaries of target data to act as a biological dashboard for users.
Pharos – A Torch to Use in Your Journey In the Dark Genome
- 4. Target Development Level • Protein classification schemes are based on structural and functional criteria. • For therapeutic development, it is useful to understand how much and what types of data are available for a given protein, thereby highlighting well-studied and understudied targets. T. Oprea et al., Nature Rev. Drug Discov. poster, Jan 2017
- 5. Target Development Level • Proteins annotated as drug targets are Tclin • Proteins for which potent small molecules are known are Tchem • Proteins for which biology is better understood are Tbio • Proteins that lack antibodies, publications or Gene RIFs are Tdark T. Oprea et al., Nature Rev. Drug Discov. poster, Jan 2017
- 7. Why Should ANYONE Fund Tdark? Data from Tudor Oprea & Christian Bologa Leptin SMO S1PR1 Orexin PCSK9 Ghrelin 1995 2000 2005 2010 2015 Median time to go from Tdark to bearing fruit is 17 years
- 8. The Causality Dilemma • Are Tdark proteins underfunded because there is no scientific interest in this category, or is the lack of knowledge perpetuated by lack of funding? • It is possible that the absence of high quality, well characterized molecular probes may be a root cause for this situation. • However, lack of tools leads to lack of interest, and lack of interest diminishes the probability of such tools being developed Tudor Oprea
- 9. There is a Knowledge Deficit • > 37% of proteins are poorly described (Tdark) • ~10% of the Proteome (Tclin & Tchem) can be targeted by small molecules • 10% of NIH R01’s (2011-2015) awarded to study 11 targets (out of 7,934 targets funded in total) • Dark genes need funding and patience https://academic.oup.com/nar/article-lookup/doi/10.1093/nar/gkw1072
- 11. Current Status191 facets 17.8 GB database 30 GB Lucene indexes 36K LoC (Java) 14K LoC (Scala) Image available Source code available 20,120 targets 15,094 diseases 2.3M publications 4,500 drugs
- 12. What’s Included? • Pharos presents data from a variety of sources, integrated by U. New Mexico • Primary focus is the protein target • Wherever possible, targets are linked to other entities (which are also interlinked) • Small molecules, Diseases, Publications • Target related data include • Identifiers, ontology terms, sequence, expression data, publications (curated & text mined), phenotypes, PPI
- 15. Target Audience Biologists & Clinical Researcher • Characterize & validate novel targets • Identify key small molecules or biologics Informatics Scientists • Data mining • Support target validation projects Program Staff • Explore the research landscape • New directions for research & funding
- 16. Different Ways to Use Pharos Random Access Direct Access Manual Interaction Programmatic Interaction Search Entity Info Precomputation converts analysis in to browsing
- 17. Supporting Both Types of Users • Efficient full text search, coupled to relevant auto- suggestion • Primary entry point when exploring and for hypothesis generation • Extensive list of facets • Supports easy construction of complex filtering rules • Extensive details for each target • Linked to external and internal resources
- 19. Sequence Search • Query is ABL,ARG (from LINCS KiNativ dataset), similarity > 0.7
- 20. Structure Search • Search by substructure or similarity • Identify targets enriched in a scaffold
- 21. Visualization • Key requirement for efficient exploration, summary • Increase information density in limited screen real estate, take context into account • Interactivity is desirable, high quality for easy inclusion in documents • Simple is better than fancy but pretty pictures have value, make for a better experience • Integrate and link to external visualization • TinX, Harmonizome
- 25. Entity Dossier • As you explore the knowledge base it’s useful keep track of data • Pharos implements a dossier function • Analogous to e-commerce shopping carts • Support for task-specific dossiers • Download a dossier as a ZIP file
- 28. Quantifying Knowledge About Targets • The Harmonizome represents the data available around a given target • Compute the number of associations for each gene in a data source and convert to ECDF • Precomputed in TCRD • Used by Harmonogram and radar chart viz • Define the Knowledge Availability Score (KAS) KAST = Ci i=1 n ∑
- 29. Knowledge Availability Score 0 50 100 150 0 20 40 60 Knowledge Availability Score Frequency
- 32. KAS vs. Other measures • Best correlation with Pubmed count • As expected, data for Tdark is noisier • Of interest are those targets with higher values of knowledge availability but small values of another metric • In particular the Jensen Pubmed Score seems to lead to such targets 1 100 10000 0 20 40 60 Knowledge Availability Score JensenPubmedScore Tbio Tchem Tclin Tdark
- 33. (Dis)similarity in Knowledge Space • There are 114 unique data sources via the Harmonizome • We represent each target as a 114-element vector • Where a source has no data for the target, we set it's value to 0 • Not necessarily the best choice, since it's really missing data • Uniform weighting may not be appropriate • Compute a pairwise Euclidean distance matrix or cosine similarity matrix for 1757 targets
- 35. Similarity in Knowledge Space • Consider Euclidean distance matrix • Of particular interest is to identify Tdark targets that have a knowledge profile that is most similar to targets that are not Tdark • 44 such targets • Within this set, 10 targets have a knowledge profile that is most similar to a Tchem or Tclin target
- 36. Target Similarity • Compute target similarity in “Harmonizome space” • Supports recommendations, prioritization • Currently extending to a generalized Target Knowledge Vector approach Tdark targets whose most similar target is not Tdark
- 37. What might this mean? • Publications are one way to prioritize targets • But we should also consider the extent of data around targets • Alternatively, all (or multiple types of) the data about a target is subsumed into a small set of publications • One paper might include RNAseq, CNV, pharmacology • Publications lag data • Tdark targets with a (relative) higher knowledge availability value but low publication based score could be rising stars?
- 38. Next Steps - Target Knowledge Vectors • Based on sparse vector representation of data availability, applied to 20K targets • A target is a document mixture of discrete and continuous variable descriptors • Set of facet values/terms and frequencies • Amino acid sequence length and individual AA residue profiles • Counts of related publications, ligands, Xtals, diseases, protein-protein interactions, etc. • Similar to TD-IDF, facet value frequencies are inversely weighted by popularity • The similarity is calculated as generalized Tanimoto
- 39. Outreach & Dissemination Activities User Feedback Deployment Webinars Documentation NER API for targets & diseases @idg_pharos Recent papers to Pharos links via Tweets
- 40. Pharos Usage • Usage statistics over the last one year are generally increasing • 89K pageviews • 14K sessions • 7.5K users
- 42. The Long Term Vision • Incorporate dependencies between data types to support inference and sophisticated filters • From presentation to summarization • Use explicit links & computational inference to generate (semi-) natural language summary using all known data • Influenced by the query • The result is a biological dashboard, customized for the user and the query Target X has been implicated in 3 diseases related to skeletal, urological and nervous systems. It has been investigated in 5 in vitro assay, 2 in vivo assays. There are 4 compounds active against this target, 3 of which are in clinical trials.
- 43. Feedback • Explore the UI, try it, break it, and let us know what works and what doesn’t • Are there data types and relations that would help you but are not available? • Nguyen & Mathias et al, Nucl. Acids Res., 2017 https://pharos.nih.gov https://spotlite.nih.gov/pharos https://hub.docker.com/r/ncats/pharos/ pharos@nih.gov @idg_pharos
- 44. Acknowledgements • Dac-Trung Nguyen, Kyle Brinacombe, Timothy Sheils, Geetha Mandava, Noel Southall, Ajit Jadhav • Steve Mathias, Oleg Ursu, Jeremy Yang, Christian Bologa, Daniel Canon, Tudor Oprea • Nicholas Fernandez, Andrew Rouillard, Avi Mayan • Finkbeiner lab, Tomita Lab • Ajay Pillai, Aaron Pawlyk, Christine Colvis