The Royal Society of Chemistry has hosted the ChemSpider database and associated platforms for over five years. Technologies made significant progress over that period but, more importantly, the community needs in terms of the variety of data types as well as search performance have increased. The preprocessing of chemicals for improved similarity searching and compound database navigation is seen as one crucial component of major development efforts to architect a new data repository. This component is engineered and implemented in collaboration with the group of Professor Oliver Kohlbacher at University of Tübingen. They have developed an approach for clustering large chemical libraries based on a fast, parallel, and purely CPU-based algorithm for 2D binary fingerprint similarity calculation. Using this method, the complete similarity network of our seed set with tens of millions of chemicals has been analyzed at a Tanimoto threshold of 0.6 and all similarity links were fed into our database. The latter is highly beneficial and will allow us to create more complex and enriching visualizations of similar compounds with associated bioactivity data and physicochemical properties for the RSC chemical repository users. This presentation will provide an overview of our experiences in applying clustering to our compound data and how it will be used to enrich data navigation on the RSC data repository.

1. Clustering the Royal Society of Chemistry
chemical repository to enable enhanced
navigation across millions of chemicals
Valery Tkachenko, Ken Karapetyan, Antony Williams,
Oliver Kohlbacher, Philipp Thiel, Colin Batchelor
ACS, 248th National Meeting
San Francisco, CA
August 14th
2014

2. Chemical space - 1060

3. Navigation in chemical space

4. Clustering

5. Science dimensions

6. • ~30 million chemicals and growing
• Data sourced from >500 different sources
• Crowdsourced curation and annotation
• Ongoing deposition of data from our
journals and our collaborators
• A structure centric hub for web-searching

7. ChemSpider

8. Properties

9. Classification

10. ChemSpider Data Slices

11. Tagging in ChemSpider

12. RSC Archive – since 1841

13. DERA -
Digitally Enabling RSC Archive

14. Twelve broad categories

15. Twelve broad categories
Largest
category is
30 times
the size of
the smallest

16. 200 subcategories

17. How does it work?
Latent Semantic Analysis to build feature sets
for (1) articles (2) categories.
Features: words, citations and pairs of words.
Domain experts (Journal Development staff)
build a category vector.
All articles with a cosine similarity greater than
an adjustable threshold go into the category.

18. RSC Data Repository

25. Structures similarity
Molecule Similarity
Similarity ?Similarity ?
Suitable in silico representation:
2D binary fingerprints
Suitable in silico representation:
2D binary fingerprints
0 1 0 1 0 1 1 0Y:
0 1 1 0 1 1 0 1X:
25
0 1 2 3 4 5 6 7

26. Structures similarity
Molecule Similarity
26
• Important fingerprint properties:
1. Length: length of the binary vector
2. Density: fraction of 1-bits
• Various fingerprint types exist
– Different atom typing and generation procedure
– Different properties (length, density, ...)
• Alternative representation: Feature list
– Store only index numbers of vector positions
– Memory-efficient storage
0 1 0 1 0 1 1 0 0 1 0 1 0 1 1 0
Length
0 1 0 0 0 1 0 0 0 0 0 0 0 0 1 0
Sparse fingerprint (sFP)
1 1 0 1 0 1 1 0 0 1 1 1 0 1 1 1
Dense fingerprint (dFP)
0 1 0 1 0 1 1 0
1,3,5,6

27. Structures similarity
27
2. Jaccard P., Bulletin del la Société Vaudoise des Sciences Naturelles (1901), 37, 547-579
3. Tanimoto T.T., IBM Internal Report (1957)
• Molecules as binary vectors
• Various chemoinformatics dis-/similiarity measures:
– Euclidean distance
– Cosine similarity (inner product)
• Most frequently used: Tanimoto Coefficient 2,3
– Corresponds to Jaccard index
– Metric
– [0.0, 1.0] (dissimilar  similar)
Molecule Similarity

28. Full Similarity Matrix Clustering
28
Results: Clustering the Available Chemspace
• ZINC all purchasable set: ~17x106
compounds (sFP)
• Tanimoto cutoff analysis: 0.76
• Opteron, 64 threads, 100 GB main memory
Total run-time: 64 hours
CCs decomposition: 12 hours
Total run-time: 64 hours
CCs decomposition: 12 hours

29. Federated linked system

30. Thank you
Email: tkachenkov@rsc.org
Slides: http://www.slideshare.net/valerytkachenko16