Exploring the chemical space of screening results Edmund Champness, Matthew Segall, Chris Leeding, James Chisholm, Iskander Yusof, Nick Foster, Hector Martinez ACS Spring 2013, 7 th April 2013 Optibrium, StarDrop, Auto-Modeller and Glowing Molecule are trademarks of Optibrium Ltd.
Overview Introduction Part 1: Chemical Space Part 2: Balancing Properties in Drug Discovery Part 3: Exploring the Space Around Us Conclusions 2
Part 1: Chemical Space 3
Chemical Space: What is it? Chemical space is the space spanned by all possible (i.e. energetically stable) molecules and chemical compounds [1] estimated to exceed 10 60 an amount so vast when compared to the number of such molecules we have made, or indeed could ever hope to make, that it might as well be infinite how we should best direct our efforts towards regions of chemical space that are most likely to contain molecules with useful biological activity? [2] [1] Wikipedia [2] Kirkpatrick, P.; C. Ellis (2004). "Chemical space, Nature - Vol 432 No 7019 (Insight) pp823-865 4
Chemical Space: How do we think about it? Outer space Moon Landing 240,000 miles Solar system exploration 3.5 billion miles Galaxy mapping 30,000 light years Universe? Quite large! Chemical space Preclinical several compounds Lead optimisation Tens/hundreds of compounds Hit-finding Many thousands of compounds All theoretical molecules Too many! 5
Methodology: Chemical Space Chemical information Fingerprints Descriptors Similarity measure Tanimoto Euclidean Dimension reduction Similar compounds close together Expectation maximisation PCA [3] tsne (t-distributed Stochastic Neighbour Embedding) [4] [3] Sam Roweis (1998) Neural Information Processing Systems 10 (NIPS'97) pp.626-632 [4] van der Maaten, L., & Hinton, G. (2008). Visualizing High-Dimensional Data Using t-sne. Journal of Machine Learning Research, 9, 2579-2605. 6
Two Approaches to Visualisation PCA Well known Can be optimised for large sets Linear transformation results in information loss Biased towards keeping dissimilar compounds far apart rather than similar compounds close together tsne (t-distributed Stochastic Neighbour Embedding) Computationally expensive Optimised for visual representation at expense of some intracompound relationships 7
How They Compare...NK2 Compounds 8
How They Compare...Dopamine Actives 9
Clustering? Clustering - Tanimoto level of 0.7 using path based fingerprints [5] [5] Butina, D. (1999). Unsupervised Data Base Clustering Based on Daylight's fingerprint and Tanimoto Similarity: A fast and automated way to cluster small and large data set. J. Chem. Inf. Comput. Sci, 39, 747-750. 10
Comparing Chemical Families 11
Part 1: 5HT1A Hit Space pk i = 7 pk i = 9.5 Looking promising... 12
Part 2: Balancing Properties in Drug Discovery 13
Property 2 Property 2 The Objectives of Drug Discovery Multi-parameter optimisation Identify chemistries with an optimal balance of properties Hit X Solubility Potency X Drug Metabolic stability Safety Absorption Quickly identify situations when such a balance is not possible Fail fast, fail cheap Only when confident Solubility No good drug Property 1 Potency Property 1 Safety Metabolic stability Absorption 14
An Approach to MPO Probabilistic Scoring [6] [6] Segall et al. (2009) Chem. & Biodiv. 6 p. 2144 15
Score Probabilistic Scoring Property data Experimental or predicted Criteria for success Relative importance Uncertainties in data Data do not separate these as error bars overlap Experimental or statistical Score (Likelihood of Success) Confidence in score Error bars show confidence in overall score Bottom 50% may be rejected with confidence Best Worst
Part 1: 5HT1A Hit Space pk i = 7 pk i = 9.5 Looking promising... 17
Part 2: Potential Property Space Score = 0 Score = 0.4 Still promising? 18
Part 2: Property + Hit Space Score = 0 Score = 0.4 Still promising? 19
Part 3: Exploring the Surrounding Space Generating New Ideas... 20
Generating Compound Ideas Applying Med. Chem. Transformation Rules Compounds generated must make sense from a medicinal chemistry perspective Apply transformation rules, derived from medicinal chemistry experience, to initial compound [7][8] Library of >200 transformations Not only functional group replacement but also framework transformations Functional group addition: e.g. sulfonamide addition [7] Drug Guru: A computer software program for drug design using medicinal chemistry rules K.D. Stewart et. al. Bioorg. Med. Chem. 14 (2006) p. 7011 [8] Applying Medicinal Chemistry Transformations and Multiparameter Optimization to Guide the Search for High-Quality Leads and Candidates Matthew Segall, Ed Champness, Chris Leeding, Ryan Lilien, Ramgopal Mettu and Brian Stevens J. Chem. Inf. Model. (2011) 51(11) pp. 2967-2976 21
Generating Compound Ideas Applying Med. Chem. Transformation Rules Compounds generated must make sense from a medicinal chemistry perspective Apply transformation rules, derived from medicinal chemistry experience, to initial compound [7][8] Library of >200 transformations Not only functional group replacement but also framework transformations Linker modification: e.g. ester to amide [7] Drug Guru: A computer software program for drug design using medicinal chemistry rules K.D. Stewart et. al. Bioorg. Med. Chem. 14 (2006) p. 7011 [8] Applying Medicinal Chemistry Transformations and Multiparameter Optimization to Guide the Search for High-Quality Leads and Candidates Matthew Segall, Ed Champness, Chris Leeding, Ryan Lilien, Ramgopal Mettu and Brian Stevens J. Chem. Inf. Model. (2011) 51(11) pp. 2967-2976 22
Generating Compound Ideas Applying Med. Chem. Transformation Rules Compounds generated must make sense from a medicinal chemistry perspective Apply transformation rules, derived from medicinal chemistry experience, to initial compound [7][8] Library of >200 transformations Not only functional group replacement but also framework transformations Ring addition: e.g. benzene to indole [7] Drug Guru: A computer software program for drug design using medicinal chemistry rules K.D. Stewart et. al. Bioorg. Med. Chem. 14 (2006) p. 7011 [8] Applying Medicinal Chemistry Transformations and Multiparameter Optimization to Guide the Search for High-Quality Leads and Candidates Matthew Segall, Ed Champness, Chris Leeding, Ryan Lilien, Ramgopal Mettu and Brian Stevens J. Chem. Inf. Model. (2011) 51(11) pp. 2967-2976 23
Iterative Application Exponential Growth! 24
Part 1: 5HT1A Hit Space pk i = 7 pk i = 9.5 Looking promising... 25
Part 2: Potential Property Space Score = 0 Score = 0.4 Still promising? 26
Part 3: Exploring the Space Score = 0 Score = 0.4 Interesting? 27
Part 3: Exploring the Space Score = 0 Score = 0.4 Interesting... 28
Conclusions We can use chemical space visualisation to help identify potentially interesting areas of chemistry Using an appropriate approach to MPO we can make confident decisions despite the uncertain nature of drug discovery data We can explore a library to find chemistries with the best chance of having a good balance of properties while avoiding missed opportunities due to uncertainty By automatically generating new chemistry ideas we can determine the potential of different areas of the chemistry space around which to expand upon our hit data 29
Acknowledgements StarDrop group past and present, including: Matthew Segall Alan Beresford Chris Leeding Dawn Yates Iskander Yusof Dan Hawksley James Chisholm Joelle Gola Nick Foster Brett Saunders Hector Martinez Simon Lister Olga Obrezanova Mike Tarbit 30
...and if you would like to try it yourself... StarDrop Hands-on Workshop Tuesday 9 th, 12:00 2:30p.m. Hall B2-C, Exhibitor Workshop Room 1 Lunch provided... Register at Optibrium booth #708 31