Acceleration of Novel Drug Design via Prediction of Drug Candidate Promiscuity

Tamas Nagy

Department of Chemistry Department of Computer Science

University of Kentucky Lexington, KY, USA 40508

March 26th, 2014

De novo drug discovery is difficult

•  Despite dramatic increases in expenditure, R&D productivity in the pharmaceutical industry is down

2 Ashburn, T. T.; Thor, K. B. Nat Rev Drug Discov 2004, 3, 673-683.

De novo drug discovery is difficult

•  It is a rare case in modern drug discovery that an unmodified natural product (e.g. taxol) becomes a drug.

•  Process is long and fraught with complications –  10-17 years from start to finish –  <10% overall probability of success

3 Ashburn, T. T.; Thor, K. B. Nat Rev Drug Discov 2004, 3, 673-683. Jorgensen, W. L. Science 2004, 303, 1813-1818.

Increasing success by understanding drug candidate polypharmacology in silico

•  Speed up process via protein-ligand binding studies that can elucidate the polypharmacology of drug candidates, i.e. their tendency to bind multiple targets. Eliminate those that may have off-target effects early.

–  E.g. Molecular docking studies •  Limited by crystal structure availability

•  Alternative: search for similarity between ligand and drug structure instead.

4 Jorgensen, W. L. Science 2004, 303, 1813-1818. Okimoto, N. et al. PLoS Comput Biol 2009, 5, e1000528. Hopkins, A. L. Nature 2009, 462, 167-168.

Determining Ligand Similarity

•  The Tanimoto coefficient relates the similarities of two sets A and B:

5

Tc =χA∩ χB

χA∪ χB

Krasowski, M. D. et al. BMC Emerg Med 2009, 9, 5. Willett, P. et al. J Chem Inf Comput Sci 1998, 38, 983-996.

Determining Ligand Similarity

6 Keiser, M. J. et al. Nat Biotechnol 2007, 25, 197-206.

•  Comparing the 216 ligands of Dihydrofolate reductase (DHFR) with:

–  Themselves •  4.7% of ligand pairs had Tc scores between

0.6-1.0 –  The 253 ligands of the similar functionality

TS antifolate enzyme •  1.6% of ligand pairs had Tc 0.6-1.0

–  The 1226 ligands of the unrelated protease thrombin.

•  0% of ligand pairs had Tc 0.6-1.0

Determining Ligand Similarity

7 Keiser, M. J. et al. Nat Biotechnol 2007, 25, 197-206.

8 Keiser, M. J. et al. Nature 2009, 462, 175-181.

Prediction of drug promiscuity via similarity ensemble approach (SEA)

3,665 drugs tested against 246 protein targets (~1,000,000 drug-target combinations)

Experimental confirmation of predicted drug promiscuity results

•  Radioligand competition binding assays for select drugs (30 in total) –  Confirm Prozac’s novel

interaction with β adrenergic receptors

–  Doralese shows higher affinity (Ki of 18nM) for the off target D4 receptor than its actual α1 adrenergic receptors

9 Keiser, M. J. et al. Nature 2009, 462, 175-181.

Drug binding across major protein boundaries

Whereas many of the predicted off-targets occur among aminergicGPCRs, a target class for which cross-activity is well-known (seelater)44, four of the drugs bound to targets unrelated by sequenceor structure to their canonical targets (Table 2). For instance, thereverse transcriptase (enzyme) inhibitor Rescriptor was predictedand shown to bind to the histamine H4 receptor, a GPCR. Thesetwo targets share no evolutionary history, functional role, or struc-tural similarity whatsoever. Intriguingly, although the Ki value of

Rescriptor for the H4 receptor is high at 5.3 mM (Table 2 andSupplementary Fig. 1), this is within its steady-state plasma concen-tration (minimum plasma concentration averages 15 mM) and isconsistent with the painful rashes associated with Rescriptor use45;likewise, H4 dysregulation has been associated with atopic der-matitis46. Similarly, the vesicular monoamine transporter (VMAT)inhibitor47 Xenazine binds two different GPCRs at sub-micromolarconcentrations (Table 2 and Supplementary Fig. 1). Despite its useover the last 50 years, Xenazine has not been reported to bind to any

Table 1 | Prediction and testing of new aminergic GPCR targets for drugs

Drug Pharmacological action E-value Predicted target Ki (nM)

N

N

O

O

F Sedalande Neuroleptic 8.3 3 102136 a1 adrenergic blocker* a1A, 1.2; a1B, 14;a1D, 7.0

1.7 3 10214 5-HT1D antagonist 140

N

N

O

O

Dimetholizine Antihistamine; antihypertensive 1.6 3 102129 a1 adrenergic blocker* a1A, 70; a1B, 240;a1D, 170

2.7 3 102113 5-HT1A antagonist 1107.4 3 10256 Dopamine D2 antagonist 180

O

O

OH

HO

HN

Kalgut Cardiotonic 3.1 3 10279 b3 adrenergic agonist 2.1 3 103

NN

Fabahistin Antihistamine 5.7 3 10257 5-HT5A antagonist 130

N+Prantal Anticholinergic; antispasmodic 5.5 3 10232 d-opioid agonist 1.4 3 104

N

HN N,N-dimethyltryptamine Serotonergic hallucinogen 3.1 3 10221 5-HT1B agonist 130

1.2 3 10213 5-HT2A agonist{ 1301.1 3 1027 5-HT5A antagonist 2.1 3 103

5.0 3 1026 5-HT7 modulator 210

NH

N

O

HN

Doralese Adrenergic a1 blocker; antihypertensive;antimigraine

2.8 3 10227 Dopamine D4 antagonist 18

HN

F

FF

O

Prozac 5-HT reuptake inhibitor; antidepressant 3.9 3 10215 b adrenergic blocker* b1, 4.4 3 103

NH

HN

N

N

N

Cl

O

O

Motilium Antiemetic; peristaltic stimulant 4.8 3 10211 a1 adrenergic blocker* a1A, 71; a1B, 530;a1D, 710

HN

F

O

O

O

Paxil 5-HT reuptake inhibitor; antidepressant 1.3 3 1027 b adrenergic blocker* b1, 1.0 3 104

Ki values are accurate 620% at two significant figures.* For the targets marked, the reference data set did not specify the receptor subtype, requiring a separate assay for each one. For instance, the MDDR contains an ‘a1 adrenergic blocker’ set, for whichit was necessary to test the a1A, a1B and a1D subtypes.{ 5-HT2A is a known target of DMT, but is shown here with its retrospective SEA E-value for comparison purposes.

ARTICLES NATURE | Vol 462 | 12 November 2009

178 Macmillan Publishers Limited. All rights reserved©2009

Novel off-target effects in common, over-the-counter drugs

10 Keiser, M. J. et al. Nature 2009, 462, 175-181.

Conclusions

•  Using the SEA method of ligand fingerprinting is an effective manner of predicting drug promiscuity and likely can be applied to ranking drug candidates.

–  Limits potential side effects that may not show up till human trials

•  It is not without its weaknesses

–  It compares drugs to ligand sets based on all shared chemical patterns instead of ones unique to specific binding sites (i.e. pharmacophores).

–  Method susceptible to false-positives (7 of 30 drugs were not active with predicted off-targets).

11

Digression

•  Last year’s Nobel Prize in Chemistry was the first to recognize the field of computational chemistry.

•  Martin Karplus, Michael Levitt, and Arieh Warshel shared the prize “for the development of multi-scale models for complex chemical systems.”

12 http://www.nobelprize.org/nobel_prizes/chemistry/laureates/2013/

Questions?

13