Virtual Screening: Methods and Applications€¦ · • actives & 3D conformations as templates, growing size of compound libraries, limited computers, etc. • In practice, many

22/01/2013

1

EBI is an Outstation of the European Molecular Biology Laboratory.

Virtual Screening:

Methods and Applications

Dr Pedro J Ballester

MRC Methodology Research Fellow

EMBL-EBI, Cambridge, United Kingdom

Talk outline

Virtual Screening: Methods and Applications UCL MSc in Drug Design,

Jan 2013

2

1. Introduction

2. Ligand-based Virtual Screening: Methods

3. Ligand-based Virtual Screening: Applications

4. Structure-based Virtual Screening: Methods

5. Structure-based Virtual Screening: Applications

22/01/2013

2


Jan 2013

3

The Drug Discovery Process

• Developing new drug = average US$4 billion and 15 years http://www.forbes.com/sites/matthewherper/2012/02/10/the-truly-staggering-cost-of-inventing-new-drugs/

• While (pre)clinical trials are the most expensive stages,

the research determining approval at early stages:

• Finding a target linked to the disease and a molecule modulating

the function of target without trigering harmful side effects.

• New targets, but hard, less funding than traditional, etc.

Payne et al. (2007) Nat Rev. Drug Disc. 6:29

Payne et al. (2007) Nat Rev. Drug Disc. 6:29


Jan 2013

4

Virtual Screening: Why?

• HTS: Main strategy for identifying active molecules (hits)

by wet-lab testing a library of molecules against a target.

• Computational methods (Virtual Screening) are needed:

• HTS is slow: HTS of corporate collections many months

• HTS is expensive: Average cost US$1M per screen.Payne et al. 2007

• Growing # of research targets no HTS until target validation

• Limited diversity:

HTS 106 cpds...

but 1060 small molecules!

(Dobson 2004 Nature)

• Target really undruggable?

22/01/2013

3

Virtual Screening (VS): when and which?

VS can complement HTS by

enriching libraries w/ likely ligands


Jan 2013

5

HTS possible?

Can afford expense and

time?

Diversity essential?

HTS VS

VS

VS

Y

Y

N

N

Y

N

Ligand/s or structure for

target?

Target structure

available?

Ligand- > target-

based VS

Ligand-based VS

only

VS not possible

Y

Y

N

N

HTS or Virtual Screening (VS)? Type of Virtual Screening (VS)?

VS to predict target affinity (hit identification)

• Search for molecules that modulate

the function of a therapeutic target.

• Hypothesis: target fn modulation

cures/alleviates associated disease


Jan 2013

6

• In silico predictions must be validated in vitro (IC50, Ki, Kd)

• Important requirements:

• Potent threshold depends on target and [drug]plasma

• Different chemical structure new IP, avoid previous problems

• Work well in practice vs Looking good on paper

22/01/2013

4

VS to predict selectivity (lead identification)

• Drugs must selectively bind to their intended target, as

binding to other proteins may cause harmful side-effects

1. The more chemically diverse the found hits are, the

more likely directly having selective hits will be.


Jan 2013

7

2. Structure-based design: e.g. identify

hits that occupy a subpocket that is in

the target but not in related proteins.

3. Ligand-based design: 3D superpose

diverse hits and use activity against

related proteins to formulate hypothesis.

VS to predict whole-cell Activity (lead identif.)

• Drug molecules must also inhibit the

target in the cell environment.

• Whole-cell assay: e.g. cancer cell

growth inhibition (GI50) measured

in the presence of the molecule.


Jan 2013

8

• Phenotypic screening is attractive ∵ many molecules

binding to the target do not have whole-cell activity.

• However, very few in silico methods for predicting which

molecules are likely to have whole-cell activity.

22/01/2013

5

Talk outline


Jan 2013

9

1. Introduction





Ligand-based VS: classes by available data

• Using a single active molecule:

• 2D: similarity search based on connectivity, fingerprints,…

(e.g. circular fingerprints, MACCS keys, etc. etc.)

• 3D: similarity search based on shape, HBs, hydrophobicity…

(e.g. USR, ROCS, Shape Signatures, etc. etc.)

• Using a few active molecules:

• Pharmacophore models: 3D alignment, pharmacophore

elucidation and pharmacophore search (e.g. Pharmer).

• Using many active molecules:

• (2D)QSAR: chemical structures + regression/classification model

• 3D-QSAR: 3D conformers + regression/classification model


Jan 2013

10

22/01/2013

6

Ligand-based VS: Motivation and Challenges

• At least a known ligand, but ≠ chemical structures than

such template/s may have better properties:

• More selectivity to target

• Less toxicity

• More potency

• Less IP concerns, …

• Challenges (some specific to the type of method):

• Scaffold hopping: better with 3D techniques and larger dbs

• Efficiency: so that very large molecular dbs can be searched

• Prospective efficacy: high rates at suitable potency threshold –

depends on previously known ligands for the target


Jan 2013

11


Jan 2013

12

Ligand-based VS: Shape Similarity

• Principle:

Similarly shaped molecules

will fit the same binding pockets

likely similar bioactivity.

• Shape-based VS: Search a molecular database for small

molecules that are similar in 3D shape to a known active.

• Prospective success in the literature (alone/combined).

• Additional advantage: chemical structure is not specified

and hence novel chemical scaffolds may be found.

2C4W-GAJ 2C4W-GAJ

22/01/2013

7


Jan 2013

13

Shape-based VS: Challenges

Two challenges limiting the potential of shape similarity:

• Effectiveness

• fast pre-alignment of molecules is prone to error, shape

description might not be sufficiently accurate, etc.

• Efficiency

• actives & 3D conformations as templates, growing

size of compound libraries, limited computers, etc.

• In practice, many active molecules are not found

because these are not even screened in silico!


Jan 2013

14

Ultrafast Shape Recognition (USR)

22/01/2013

8


Jan 2013

15

USR descriptors 12 real-valued

USR descriptors

Molecular

Shape

Relative

position

of atoms

4 reference

points

interatomic

distances

1st, 2nd and

3rd sample

moments of

each dstn


Jan 2013

16

USR similarity

Shape similarity as the inverse of a distance between

vectors of USR descriptors Score (0,1]

Ranking molecules in terms of shape similarity to a

query/template molecule (e.g. a known inhibitor)

22/01/2013

9


Jan 2013

17

USR: most similar shapes in 2.5M DB

1st 2nd 3rd 4th

Query

Rank

1

2

3 Template is found to be the most similar molecule.

Top ranked molecules

are very similar, despite the large size of the database


Jan 2013

18

USR vs ESshape3D: effectiveness

1st 2nd 3rd 4th

Query

Rank

2

2 USR

ESshape3D

USR

22/01/2013

10


Jan 2013

19

USR vs ESshape3D: efficiency

1𝑡𝑠𝑐𝑟𝑒𝑒𝑛

𝑡𝑈𝑆𝑅 𝑞, 𝑛 = 𝑛 (𝑡𝑟𝑒𝑎𝑑 + 𝑞 𝑡𝑠𝑐𝑟𝑒𝑒𝑛)

𝑡𝑟𝑒𝑎𝑑 ~ 10−5 𝑠𝑒𝑐𝑠/𝑐𝑜𝑛𝑓

𝑡𝑠𝑐𝑟𝑒𝑒𝑛 ~4 10−8 𝑠𝑒𝑐𝑠/𝑐𝑜𝑛𝑓

• Interactively and accurately shape searching a multi-million

molecular DB was not feasible until USR. Now:

• Tom Blundell’s group, University of Cambridge, UK

• Stefano Moro’s group, University of Padova, Italy

• David Wild’s group, Indiana University, USA

Molecular data mining with USR


Jan 2013

20

22/01/2013

11

UFSRAT: USR + more chemical properties

Malcolm Walkinshaw’s group,

University of Edinburgh, UK


Jan 2013

21

J. Adie 2010 PhD thesis

S. Shave 2010 PhD thesis

Talk outline


Jan 2013

22

1. Motivation





22/01/2013

12


Jan 2013

23

First prospective application of USR


Jan 2013

24

Problem statement

• Interested in inhibitors for hNAT1 (Arylamine N-

acetyltransferases; EC 2.3.1.5; target for breast cancer).

• mNat2 is functionally analogous to hNAT1 assays

• Previously: mNat2 manual screen of 5000 cherry-picked

cpds 0.1% confirmed hit rate (IC50 < 10M)

• No protein structure available Ligand-based approach:

most potent hit (IC50 = 1.1M) as template for the VS.

• Goal: find additional inhibitors with diverse chemical

structure from template to be considered as probes.

22/01/2013

13


Jan 2013

25

Virtual Screen protocol

5.3M cpds (690M 3D conformers)

Extremely fast screening

100 USR queries on this DB, i.e. 69 billion shape

comparisons in just 83 mins. using a single dual-core proc.

690M confs 34 partitions of ~ 20M confs each

13,700,000 confs/sec using two execution cores.

Top 23 cpds bought

Testing by

collaborators

USR-best_hit

Testing virtual screening hits in vitro

#Compounds active against target?

ZINC7 purchasable


Jan 2013

26

Results of in vitro confirmatory tests

• 9/23 tested cpds

had IC50 < 10M

hit rate = 39%

• Just £500 to buy

23 cpds to test

• Manual screen yield

hit rate = 0.1% on

the same target.

• Importantly, all nine hits were scaffold hops with respect to

the active template (most hops with SMACCS < 0.4).

Template Most potent USR hits

22/01/2013

14

Prospective application of UFSRAT

• J. Adie 2010 PhD thesis (University of Edinburgh, UK):

• UFSRAT w/ 4 known inhibitors of human 11-hydroxysteroid

dehydrogenase 1 (h11-HSD1), a target for type 2 diabetes.

• a database with four million commercially available compounds


Jan 2013

27

• 35 tested cpds:

• 5 showed potent

inhibition both in cells

and recombinant

protein assays

• 14% confirmed hit

rate; Ki values

ranged from 51.8 nM

to 11.3 M).

Talk outline


Jan 2013

28

1. Motivation





22/01/2013

15

Docking

• Docking = Pose generation + Scoring

• Pose generation: estimating the conformation and orientation of

the ligand as bound to the target.

• Scoring: predicting how strongly the ligand binds to the target.

• Many relatively accurate algorithms for pose generation,

but imperfections of scoring functions continue to be the

major limiting factor for the reliability of docking.


Jan 2013

29

• If X-ray structure of the target

is available Docking:

• predicting whether and how a

molecule binds to the target.

• Possible variables:

• translation and rotation of the ligand relative to the binding site

involves six degrees of freedom (3D position + 3D orientation)

• Torsional/conformational degrees of freedom of both the

ligand (on the fly/stored) and the protein (flexible docking).

Pose generation in Docking

• Goal: finding a pose as similar as possible to that of the ligand co-crystallised with the target.

• How: search algorithm generates several poses of each considered ligand (multimodal optimisation problem). • tradeoff between time and search space coverage

22/01/2013

16

• Force Field-based SFs (e.g. DOCK score)

• Empirical SFs (e.g. X-Score)

• Knowledge-based SFs (e.g. PMF)

• SFs are trained on pK data usually through MLR:

• FF (Aij, Bij), Emp(w0,…,w4) and sometimes KB ( )

Scoring Functions for Docking: functional forms


Jan 2013

31

Scoring Functions for Docking: limitations

• Two major sources of error affecting all SFs:

1. Limited description of protein flexibility.

2. Implicit treatment of solvent.

• This is necessary to make SFs sufficiently fast.

• 3rd source of error has received little attention so far:

• Conventional scoring functions assume a theory-inspired

predetermined functional form for the relationship between:

• the structure-based description of the p-l complex

• and its measured/predicted binding affinity

• Problem: difficulty of explicitly modelling the various

contributions of intermolecular interactions to binding affinity.

• Also, SFs use an additive functional form, but this has been

specificly shown to be suboptimal (Kinnings et al. 2011 JCIM).


Jan 2013

32

22/01/2013

17


Jan 2013

33

non-parametric machine learning can be used to implicitly

capture the functional form (data-driven, not knowledge-based)

A Machine Learning Approach

A machine learning approach

• Main idea: a priori assumptions about the functional

form introduces modelling error no asumptions!

• reconstruct the physics of the problem implicitly in an

entirely data-driven manner using non-parametric ML.

• Random Forest (Breiman, 2001) to learn how the

atomic-level description of the complex relates to pK:

• Random Forest (RF): a large ensemble of diverse DTs.

• Decision Tree (DT): recursive partition of descriptor space s.t.

training error is minimal within each terminal node.

• But how do we characterise a protein-ligand complex as

set of numerical descriptors (features)?


Jan 2013

34

22/01/2013

18

Characterising the protein-ligand complex


Jan 2013

35

pKd/i C.C … C.Cl … C.I N.C … I.I PDB ID

5.70 95 30 0 73 0 2p33

+1 binding affinity

features or

descriptors

PDBbind benchmark

• De facto standard for SFs benchmarking: Cheng, T., Li, X., Li, Y., Liu, Z. & Wang, R. (2009) JCIM 49, 1079-1093

• Refined set 1300 manually curated protein-ligand

complexes with measured binding affinity ( diverse):

• Benchmark: 16 state-of-the-art SFs test set error

• RF-Score vs 16 SFs on test set error, but:

• Other SFs have an undisclosed number of cmpxes in common!

• RF-Score & X-Score (best) non-overlapping training-test sets.


Jan 2013

36

22/01/2013

19

Training and testing machine learning SFs


Jan 2013

37

pKd/i C.C – C.I N.C – I.I PDB

0.49 1254 – 0 166 – 0 1w8l

– – – – – – – –

13.00 2324 – 0 919 – 0 2ada

pKd/i C.C – C.I N.C – I.I PDB

1.40 858 – 0 0 – 0 2hdq

– – – – – – – –

13.96 4476 – 0 283 – 0 7cpa

Random Forest training

(descriptor selection, model selection)

RF-Score

(description and training choices)

Training set (1105 complexes) Test set (195 complexes)

1105

195

Generation of descriptors (dcutoff, binning, interatomic types)

1w8l

pKi=0.49

1gu1

pKi=4.52

2ada

pKi=13

2hdq

pKi=1.4

1e66

pKi=9.89

7cpa

pKi=13.96

RF-Score‘s performance


Jan 2013

38

Rp=0.776

SD=1.58

22/01/2013

20


Jan 2013

39

Fair assessment of scoring functions (SFs)

How to find best regression techniques for this problem?

Ballester PJ & Mitchell JBOM (2011) Journal of Chemical Information and modelling 51, 1739–1741

Benchmarking SFs: cultural barriers

1. Test all SFs on the same data set previously selected

by 3rd party: increasingly the case, but not always.

2. Have all SFs trained on the same data set:

• Almost never the case, often training set is not even disclosed.

• SF_A > SF_B simply ∵ A’s training set contains data more

relevant to predict the test set, not ∵ more accurate modelling!

3. No complexes in common between training and test

• Almost never the case, ∵ SFs are not recalibrated.

• Largest bias! Most SFs in PDBbind benchmark are likely to

perform worse if recalibrated to exclude overlaps.


Jan 2013

40

22/01/2013

21


Jan 2013

41

Performance vs overlap in RF-Score

If we allow 65 cpxes overlap

Rp=0.827

No overlap (unlike other SFs

but X-Score) Rp=0.776

Talk outline


Jan 2013

42

1. Motivation





22/01/2013

22


Jan 2013

43

First prospective application of RF-Score

44

Type II dehydroquinase - DHQase (EC 4.2.1.10)

• 3rd enzyme of the shikimate pathway

• Ideal target for antimicrobials development, as pathway

is essential for bacteria, but not present in mammals.

• bacteria human pathogens such that M. tuberculosis

(Tuberculosis ) and H. pylori (duodenal/gastric ulcers)

Hierarchical Virtual Screening for the discovery of new molecular scaffolds in

antibacterial hit identification

LibMedia Drug Design -

Oxford, UK, September 2012

UCL MSc in Drug Design,

Jan 2013 Virtual Screening: Methods and Applications

22/01/2013

23

45

High-Throughput Screening (HTS)

Hierarchical Virtual Screening for the discovery of new molecular scaffolds in

antibacterial hit identification

LibMedia Drug Design -

Oxford, UK, September 2012

• Robinson et al. (2006) J Med Chem 49:1282

• HTS 150,000 cpds against Type II DHQase from H. Pylori

• Primary screening: 100 cpds with 50% inhibition at cpd

concentration of 20g/mL (~106M)

• Secondary screening: Only one confirmed active reported

• Ki of this HTS hit (AH9095):

• HP: 20M

• SC: 230M

• MT: 10% at 200M

• As a result, only three known scaffolds for DHQase, the

other two being rational derivatives of the substrate.


Jan 2013 Virtual Screening: Methods and Applications

46

Hierarchical Virtual Screening

USR-(CA2,GAJ,RP4)

9M

USR-CA2

ZINC drug-like

GOLD-2BT4

USR-GAJ

ZINC drug-like

USR-RP4

ZINC drug-like

GOLD-2C4W GOLD-2CJF

RF-Score 3.9K

Testing virtual screening hits in vitro

Compounds active against target

148

?

ZINC drug-like

USR-CA2

ZINC drug-like

USR-CA2

ZINC drug-like

Testing virtual screening hits in vitro (£5000)

Compounds active against target

3.9K


Jan 2013

22/01/2013

24

47

Shape similarity screen

• Screening library: ZINC8 subset of 8,784,580 drug-like

commercially available molecules.

• i.e. 59 times larger than that of the HTS (150,000 mols.)

• USR: three distinct co-crystallised ligands as templates to

account for partial shape complementarity to some extent

promotes diversity of hits

• From ~9M to ~4K molecules similarly shaped to co-

crystallised ligands ~complementary to binding site

more likely to bind than those in the initial library


Jan 2013

48

GOLD pose generation + RF-Score rescoring

2cjf-rp4 (Scl) 2bt4-ca2 (Scl) 2c4w-gaj (Hp)

• Docking pose generation by Martina Mangold & Jochen

Blumberger (Dept. Chemistry, University of Cambridge):

• GOLD: generate 3D conformation, position and orientation of a

small molecule relative to a putative binding site.

• calibration: RMSD of co-crystallised vs predicted poses.

• Purchase: top ranked poses according to RF-Score.


Jan 2013

Virtual Screening: Methods and Applications

22/01/2013

25

Experimental confirmation: wet-lab assays

• Experimental validation of hits by Nigel Howard and

Chris Abell (Dept. Chemistry, University of Cambridge).

• Outstanding hit rates of ~ 60% with Ki 250 M 100

new and structurally diverse actives (£5,000 cost).

49

Overall Performance Ki ≤ 100M Ki ≤ 250M (L1, L2, L3)[M]

Against Mtb DHQase 35 (23.6%) 89 (60.1%) (23, 24, 40)

Against Scl DHQase 40 (27.0%) 91 (61.5%) (4, 21, 29)


Jan 2013

50

New active scaffolds for Type II DHQase

M. Tuberculosis

Ki

Computational Drug Design School of Computing,

University of Kent, Nov 2012

22/01/2013

26

51


M. Tuberculosis

Ki

Computational Drug Design School of Computing,

University of Kent, Nov 2012


Jan 2013

52


S. coelicolor

Ki

22/01/2013

27

Summary

• Computational methods needed to reduce expenses

and timescales, but also to tackle challenging targets.

• Goal: finding potent, selective and innovative drug-like

molecules with activity in target and whole-cell assays.

• If known active molecules, shape and other 3D props.

(e.g.USR/UFSRAT) to search very large molecular DBs.

• If only structure of target known, docking molecules with

latest generation of machine learning scoring functions.

• Remember: VS methods can be complementary to HTS

or combined hierarchically. Need experimental data.


Jan 2013

53


Jan 2013

54

Thanks – Q&A

Documents

Virtual Screening: Methods and Applications€¦ · • actives & 3D conformations as templates, growing size of compound libraries, limited computers, etc. • In practice, many