This document discusses using larger datasets to build, validate, and share machine learning models for drug discovery. It notes that datasets were historically very small, but are now much larger, containing thousands to tens of thousands of compounds. Larger datasets allow modeling a broader range of properties and endpoints. The document examines examples of published models for various targets and properties built using large datasets. It emphasizes the importance of making these models openly accessible so they can be more widely used and built upon by others.

1. Mining 'Bigger' Datasets to Create, Validate and Share Machine
Learning Models
Sean Ekins1,2,3* and Alex M. Clark4
1 Collaborations Pharmaceuticals, Inc., 5616 Hilltop Needmore Road, Fuquay-Varina, NC 27526, USA
2 Collaborative Drug Discovery, 1633 Bayshore Highway, Suite 342, Burlingame, CA 94010, USA
3 Collaborations in Chemistry, 5616 Hilltop Needmore Road, Fuquay-Varina, NC 27526, USA
4 Molecular Materials Informatics, 1900 St. Jacques #302, Montreal H3J 2S1, Quebec, Canada
Disclosure: As well as employee of above funded by NIH, EC FP7, consultant for several
rare disease foundations, drug companies and consumer product companies etc.

2. Laboratories past and present
Lavoisier’s lab 18th C Edison’s lab 20th C
Author’s lab 21th C
+ Network of global collaborators

3. "Rub al Khali 002" by Nepenthes
The chemistry/ biology data desert outside of pharma circa early 2000’s
Limited ADME/Tox data
Paucity of Structure Activity Data
Small datasets for modeling
Drug companies – gate keepers of information for drug discovery

4. "Oasis in Libya" by Sfivat
The growing chemistry/biology data Oasis outside of pharma circa 2015

5. ADME/Tox models 15 yrs on: Then & Now
• Datasets very small < 100 cpds
• Heavy focus on P450
• Models rarely used
• Very limited number of
properties addressed
• Few tools / agorithms used
• Limited access to models
• Much bigger datasets > 1000s
cpds >10,000
• Broader range of models
• Models more widely used and
reported
• More accessible models
• Pharma making data available
 70 hERG models (Villoutreix and
Taboroureau 2015)
 19 protein binding models
(Lambrinidis et al 2015)
 40 BBB models upto 2009

6. Model resources for ADME/Tox

7. CYP 1A2 2C9 2C19
Substrate (mM) phenacetin (10) diclofenac (10) omeprazole (0.5)
Inhibitor naphthoflavone sulfaphenazole tranylcypromine
Compounds IC50 (mM) IC50 (mM) IC50 (mM)
JSF-2019 2.25 3.55 10.8
Retinal dehydrogenase 1
ADME SARfari predicts importance of CYP1A2, CYP2C9, CYP2C19
The Naïve Bayes model
was built with 142345
compounds (training and
validation) and features
135 learned classes.
Testing by
Dr. Joel Freundlich

9. Just a matter of scale?
Drug Discovery’s
definition of Big data
Everyone else’s definition of Big data

10. • Data Sources
• PubChem
• ChEMBL
• ToxCast over 1800 molecules tested against over 800 endpoints
Where can we get the datasets

11. Open source – but much smaller
400 diverse, drug-like molecules active against neglected diseases
400 cpds from around 20,000 hits
generated screening campaign ~ four million compounds from the
libraries of St. Jude Children's Research Hospital, TN, USA, Novartis
and GSK.
Many screens completed

12. Bigger datasets and model collections
• Profiling “big datasets” is going to
be the norm.
• A recent study mined PubChem
datasets for compounds that have
rat in vivo acute toxicity data
• This could be used in other big data
initiatives like ToxCast (> 1000
compounds x 800 assays) and
Tox21 etc.
• Kinase screening data (1000s mols
x 100s assays)
• GPCR datasets etc (1000s mols x
100s assays)
Zhang J, Hsieh JH, Zhu H (2014) Profiling Animal
Toxicants by Automatically Mining Public
Bioassay Data: A Big Data Approach for
Computational Toxicology. PLoS ONE 9(6):
e99863. doi:10.1371/journal.pone.0099863
http://127.0.0.1:8081/plosone/article?id=info:d
oi/10.1371/journal.pone.0099863

14. ‘Bigger’ and not ‘Big’

15. (220463)
(102633)
(23797)
(346893)
(2273)
(1783)
(1248)
(5304)
(218640)
(102634)
(23737)
(345011)
1771924
Are bigger models better for tuberculosis ?
Ekins et al., J Chem Inf Model
54: 2157-2165 (2014)

16. No relationship between internal or external ROC and the
number of molecules in the training set?
PCA of combined
data and ARRA(red)
Ekins et al., J Chem Inf Model
54: 2157-2165 (2014)
Internal and leave out 50%x100 ROC track each other
External ROC less correlation
Smaller models do just as well with external testing
~350,000

17. The Opportunity
•Get pharmas to use open source molecular descriptors and algorithms
•Benefit from initial work done by Pfizer/CDD
•Avoid repetition of open source tools vs commercial tools comparisons
•Change the mindset from real data to virtual data – confirm predictions
•ADME/Tox is precompetitive
•Expand the chemical space and predictivity of models
•Share models with collaborators – Companies could share data as models
Ekins and Williams, Lab On A Chip, 10: 13-22, 2010.
Gupta RR, et al., Drug Metab Dispos, 38: 2083-2090, 2010

18. Pfizer Open models and descriptors
Gupta RR, et al., Drug Metab Dispos, 38: 2083-2090, 2010
• What can be developed with very large training
and test sets?
• HLM training 50,000 testing 25,000 molecules
• training 194,000 and testing 39,000
• MDCK training 25,000 testing 25,000
• MDR training 25,000 testing 18,400
• Open molecular descriptors / models vs
commercial descriptors

19. • Examples – Metabolic Stability
Gupta RR, et al., Drug Metab Dispos, 38: 2083-2090, 2010
HLM Model with CDK and
SMARTS Keys:
HLM Model with MOE2D and
SMARTS Keys
# Descriptors: 578 Descriptors
# Training Set compounds:
193,650
Cross Validation Results: 38,730
compounds
Training R2: 0.79
20% Test Set R2: 0.69
Blind Data Set (2310
compounds):
R2 = 0.53
RMSE = 0.367
Continuous  Categorical:
κ = 0.40
Sensitivity = 0.16
Specificity = 0.99
PPV = 0.80
Time (sec/compound): 0.252
# Descriptors: 818 Descriptors
# Training Set compounds:
193,930
Cross Validation Results: 38,786
compounds
Training R2: 0.77
20% Test Set R2: 0.69
Blind Data Set (2310
compounds):
R2 = 0.53
RMSE = 0.367
Continuous  Categorical:
κ = 0.42
Sensitivity = 0.24
Specificity = 0.987
PPV = 0.823
Time (sec/compound): 0.303
PCA of training (red) and test (blue)
compounds
Overlap in Chemistry space

20. • Examples – P-gp
Gupta RR, et al., Drug Metab Dispos, 38: 2083-2090, 2010
Open source descriptors CDK and C5.0 algorithm
~60,000 molecules with P-gp efflux data from Pfizer
MDR <2.5 (low risk) (N = 14,175) MDR > 2.5 (high risk) (N = 10,820)
Test set MDR <2.5 (N = 10,441) > 2.5 (N = 7972)
Could facilitate model sharing?
CDK +fragment descriptors MOE 2D +fragment descriptors
Kappa 0.65 0.67
sensitivity 0.86 0.86
specificity 0.78 0.8
PPV 0.84 0.84

21. MoDELS RESIDE IN PAPERS
NOT ACCESSIBLE…THIS IS
UNDESIRABLE
How do we share them?
How do we use Them?

22. Open ExtendedConnectivity Fingerprints
ECFP_6 FCFP_6
• Collected,
deduplicated,
hashed
• Sparse integers
• Invented for Pipeline Pilot: public method, proprietary details
• Often used with Bayesian models: many published papers
• Built a new implementation: open source, Java, CDK
– stable: fingerprints don't change with each new toolkit release
– well defined: easy to document precise steps
– easy to port: already migrated to iOS (Objective-C) for TB Mobile app
• Provides core basis feature for CDD open source model service
Clark et al., J Cheminform 6:38 2014

23. Uses Bayesian algorithm and FCFP_6 fingerprints
Bayesian models
Clark et al., J Cheminform 6:38 2014

24. Exporting models from CDD
Clark et al., JCIM 55: 1231-1245 (2015)

25. Machine Learning – Different tools
• Models generated using : molecular
function class fingerprints of maximum
diameter 6 (FCFP_6), AlogP, molecular
weight, number of rotatable bonds,
number of rings, number of aromatic
rings, number of hydrogen bond
acceptors, number of hydrogen bond
donors, and molecular fractional polar
surface area.
• Models were validated using five-fold
cross validation (leave out 20% of the
database).
• Bayesian, Support Vector Machine and
Recursive Partitioning Forest and single
tree models built.
• RP Forest and RP Single Tree models
used the standard protocol in Discovery
Studio.
• 5-fold cross validation or leave out 50%
x 100 fold cross validation was used to
calculate the ROC for the models
generated
• *fingerprints only Ai et al., ADDR 86: 46-60, 2015
KCNQ1

26. Ames Bayesian model built with 6512 molecules (Hansen et al., 2009)
Features important for Ames actives. Features important for Ames inactives.

27. Ames Bayesian model built using CDD Models showing ROC
for 3 fold cross validation. Note only FCFP_6 descriptors were
used

28. FCFP6 fingerprint models in CDD
Clark et al., JCIM 55: 1231-1245 (2015)

29. ECFP6 fingerprint only models in MMDS
Clark et al., JCIM 55: 1231-1245 (2015)

30. Using AZ-ChEMBL data for CDD Models

31. • Human microsomal
intrinsic clearance
• Rat hepatocyte
intrinsic clearance

32. What if the models were already
built for you
• Instead of having to go into a database and find
data
• The models are already prebuilt
• Ready to use
• Shareable
• Create a repository of models

33. Previous work by others
• Using large datasets to predict targets with Bayesian algorithm
• Bayesian classifier - 698 target models (> 200,000 molecules, 561,000
measurements) Paolini et al 2006
• 246 targets (65,241 molecules) Similarity ensemble analysis Keiser et al 2007
• 2000 targets (167,000 molecules) target identification from zebrafish screen
Laggner et al 2012
• 70 targets (100,269 data points) Bender et al 2007
• Many others…..
• None of these enable you qualitatively or quantitatively predict activity for a
single target.

34. Recent Studies
• Bit folding – trade off between performance & efficacy
• Model cut-off selection for cross validation
• Scalability of ECFP6 and FCFP6 using ChEMBL 20 mid size
datasets
• CDK codebase on Github (http://github.com/cdk/cdk: look
for class org.open-science.cdk.fingerprint.model.Bayesian )
• Made the models accessible http://molsync.com/bayesian2
Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60

35. What do 2000 ChEMBL models
look like
Folding bit size
Average
ROC
http://molsync.com/bayesian2 Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60

36. ChEMBL 20
• Skipped targets with > 100,000 assays and sets with
< 100 measurements
• Converted data to –log
• Dealt with duplicates
• 2152 datasets
• Cutoff determination
• Balance active/ inactive ratio
• Favor structural diversity and activity distribution
Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60

37. Desirability score
• ROC integral for model using subset of molecules and
threshold for partitioning active / inactive (higher is
better)
• Second derivative of population interpolated from the
current threshold (lower is better)
• Ratio of actives to inactives if the collection partitioned
(actives+1) / (inactives+1) or reciprocal..whichever greater
Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60
Ekins et al Drug Metab Dispos 43(10):1642-5, 2015

38. Models from ChEMBL data
http://molsync.com/bayesian2 Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60

39. Results
• Bit folding – plateau at 4096, can use 1024 with little
degredation
• Cut off – works well
• Evaluated balanced training: test and diabolical were test
and training sets are structurally different
Easy ROC 0.83 ± 0.11 Hard ROC 0.39 ± 0.23
Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60

40. Models in mobile app
• Added atom
coloring using
ECFP6 fingerprints
• Red and green
high and low
probability of
activity,
respectively
Clark and Ekins, J Chem Inf Model. 2015 Jun 22;55(6):1246-60

41. Results for Bayesian model cross validation. 5-fold and Leave one out (LOO)
validation with Bayesian models generated with Discovery Studio and Open
Models implemented in the mobile app MMDS. * = previously published
Ekins et al Drug Metab Dispos 43(10):1642-5, 2015
Transporter models

42. Ekins et al Drug Metab Dispos 43(10):1642-5, 2015
Transporter models

43. ToxCast data
• Few studies use the ToxCast data for machine learning
• Recent reviews Sipes et al., Chem Res Toxicol. 2013 Jun 17; 26(6): 878–895.
• Liu et al., Chem Res Toxicol. 2015 Apr 20;28(4):738-51
• A set of 677 chemicals was represented by 711 in vitro bioactivity descriptors
• (from ToxCast assays), 4,376 chemical structure descriptors (from QikProp,
OpenBabel, PaDEL, and PubChem), and three hepatotoxicity categories
• six machine learning algorithms: linear discriminant analysis (LDA), Naïve Bayes
(NB), support vector machines (SVM), classification and regression trees (CART), k-
nearest neighbors (KNN), and an ensemble of these classifiers (ENSMB)from animal
studies)
• nuclear receptor activation and mitochondrial functions were frequently found in
highly predictive classifiers of hepatotoxicity
• CART, ENSMB, and SVM classifiers performed the best

44. CDD Models for human P450s (NVS data) from ToxCast (n=1787) <1uM cutoff
CYP1A1 CYP1A2 CYP2B6 CYP2C18
CYP2C19 CYP2C9 CYP3A4 CYP3A5

45. ToxCast models in a mobile app
IC50 1A2 = 2.25 uM
IC50 2C9 = 3.55 uM
IC50 2C19 = 10.8 uM
In vitro data
Courtesy Dr. Joel Freundlich

46. PolyPharma a new free app for drug discovery

47. Composite models - Binned Bayesians
Clark et al., Submitted 2015

48. Summary
• Shown that open source models/ descriptors comparable to
previously published models with commercial software
• Implemented Bayesian machine learning in CDD Vault
• Can be used on private or public data
• Can enable sharing of models in CDD Vault
• Enabled export of models – can use models in 3rd part mobile apps
or other tools
• Demonstrated various ADME/Tox models and transporters
• Make ToxCast data into models that can be used by anyone
• Provide more information on models and predictions
• Visualize training set molecules vs test compounds
• Use a model to predict compounds and then test them

49. Future ?
+ = Big Models
Thousands of Big Models
How do you validate 1000’s of models
How do algorithms hands 500K – 1M molecules
Need new algorithms, data visualization, mining approaches
Model sharing is here
Need for broad Biology & Chemistry knowledge – open minds, BIG thinkers

50. Acknowledgments
• Alex Clark Antony Williams
• Joel Freundlich Robert Reynolds
• Steven Wright
• Krishna Dole and all colleagues at CDD
• Award Number 9R44TR000942-02 “Biocomputation across distributed private datasets to enhance drug
discovery” from the NIH National Center for Advancing Translational Sciences.
• R41-AI108003-01 “Identification and validation of targets of phenotypic high throughput screening” from NIH
National Institute of Allergy and Infectious Diseases
• Bill and Melinda Gates Foundation (Grant#49852 “Collaborative drug discovery for TB through a novel
database of SAR data optimized to promote data archiving and sharing”).

51. Software on github
Models can be accessed at
• http://molsync.com/bayesian1
• http://molsync.com/bayesian2
• http://molsync.com/transporters