Cresset: 25 year of Fields

25 Year of Fields: What Have we Learned?

Mark Mackey

Cresset

Bioisosteres Bioisosteric groups

Biologically relevant method for comparing molecules

How did we get there?

A glorious tale of

intrigue

deception

skullduggery

sex

How did we get there?

A glorious tale of

phosphodiesterases

enrichment graphs

unbelievably expensive graphics hardware

molecular electrostatics

Fortran 77

almost no sex at all

How did it all start?

“Some Italians in „73 or „74 did 2D plots of ESP”

Harel Weinstein (1982ish) 2D vectors on 5-HT

DHFR work at Wellcome mid-80s

SK&F

> COSMIC modelling package

> Modelling PDE III inhibitors (Davis, Warrington, Vinter, JCAMD

1987, 1(2), 97)

Promotion at SK&F

1988

All science ceased as Andy was promoted to

head of IT

1989

All science started again as Andy was fired

as head of IT

Lesson 1

Not all brilliant scientists make brilliant managers

Cambridge and Consulting

1990 – Jeremy Sanders and Chris Hunter

This led to the development of a full force field

along the same lines (Vinter, JCAMD 1994, 8, 653-

668)

Lesson 2

To get good answers using fields, you need good

fields

Publication at last!

“Multiconformational

composite molecular potential

fields in the analysis of drug

action. I. Methodology and

first evaluation using 5-HT

and histamine action as

examples”

J. G. Vinter and K. I. Trollope,

JCAMD 9 (1995) 297-307

The critics‟ verdict?

“Incomprehensible”

“Multiconformational composite molecular potential fields in the analysis of drug action. II” has yet to appear.

Lesson 3

If you write papers that people can‟t read, they

don‟t read them

“Molecular Field Extrema as Descriptors of Biological Activity: Definitions and Validation” T.

Cheeseright, M. Mackey, S. Rose and A. Vinter, JCIM 2006, 46, 655-676

Critics‟ verdict: “Mostly incomprehensible”.

James Black Foundation and Napp

> Field analysis now gave good(ish) qualitative

results

> Quantitation was a problem

Original idea

> Align and score purely on the position and size of the field points

> Define a „pseudo-Coulombic‟ potential between field points:

offsetdist

fpsizefpsizeE fpfp

)2()1(21

Original idea

> Align and score purely on the position and size of the field points

> Define a „pseudo-Coulombic‟ potential between field points:

offsetdist

fpsizefpsizeE fpfp

)2()1(21

Problems: Different well widths

Problems: Different well widths

> Not really soluble with a field point

representation– this is some of the information

we „throw away‟ going to a field minimum-based

representation

> Unfortunately, this leads to less-than-optimal

results

> Tried ellipsoidal field points etc but it didn‟t help

much

New idea – field sampling

> For a given field point in molecule A, instead of

estimating what the field would be at the

corresponding point in B from the positions of its

field points, why not calculate directly?

A B

New idea – field sampling

A B

Afp

ABABA fppositionFfpsizeE ))(()(

New idea – field sampling

A B

2

ABBAAB

EEE

BBAA

ABAB

EE

ES

2

Afp

ABABA fppositionFfpsizeE ))(()(

Advantages

> The entire „true‟ field is used in the calculation

> Potential well widths implicitly included

> Fast to calculate

> Only a few field values need to be calculated

> Samples fields at biologically-relevant points

> Gauge-invariant

Lesson 4

Field Points aren‟t enough

You need the field as well

More development

> Changed the vdW field

> Used to be scaled by visible surface area, calculated 13C

NMR constants and other stuff

> Added the hydrophobic field

> Improved methods for generating initial alignments

> Field permutations

> Monte Carlo

> Grid-sampled Monte Carlo

> Greedy clique matching

Cresset!

> Cresset founded in November 2001

> Business plan:

1. Condense field points into fingerprints

2. Stuff in Oracle

3. $$$$$

FieldPrints

Initial testing showed brilliant results

0

10

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60 70 80 90 100

% Database Retrieved

% H

its R

etr

ieved

Actual

Perfect

Random

FieldPrints

Later testing showed insipid results

Lesson 5

If the experiment works, never repeat it

Ok, not really

FieldPrints

Why did it look OK earlier?

Actives

• Large

• Positively charged

Decoys

• Small

• Neutral

Surprise! FieldPrints can tell the difference!

Lesson 6

Testing virtual screening methods is hard.

Really hard.

Even when you know how hard it‟s going to be, it‟s

harder than that.

See

“Benchmarking Sets for Molecular Docking”, Huang et al. J. Med. Chem., 2006, 49(23), 6789-6801

“What do we know and when do we know it?”, Nicholls, JCAMD, 2008, 22(3) 239-255

“FieldScreen: Virtual Screening using Molecular Fields”, Cheeseright et al. JCIM, 2008 48(11) 2108-2117

“Better than Random? The Chemotype Enrichment Problem”, Mackey and Melville, JCIM, 2009 49(5), 1154-62

and more

So where did we end up?

> FieldPrints didn‟t work very well

> But the full field similarity algorithm did (T. Cheeseright, M. Mackey, J. Melville, J. G. Vinter. (2008) 'FieldScreen: Virtual Screening Using Molecular Fields.

Application to the DUD Data Set' J. Chem. Inf. Model. 48, 2108)

> Used on ~100 virtual screening projects so far

> ~80% success rate

Lesson 7

See Lesson 4*

Sometimes you have to learn lessons twice

*“Field points aren’t enough: you need the field as well”

Other uses for field similarity

> FieldAlign

> Small-scale alignments and similarity scoring

> Useful for SAR

> FieldStere - Use field

similarity to score

bioisosteric replacements

> Avoids fragment scoring

limitations

> Allows for electronic influence

of replacing a moiety on the

rest of the molecule and vice

versa

> Allows for neighbouring group

effects

Other uses for field similarity

Other uses for field similarity

O

N+

NH

N

HO

O

O

OHO

H

N

N+

N

O

N

N

O

H

F

F F

F

F

O

HN N+

H

H

N N

NH

Use Fields to cross compare the actives Understand the pharmacophore - a detailed Field map of activity Employ the template in FieldAlign, FieldScreen, FieldStere

3 CCR5 actives

FieldTemplater

Other uses for field similarity

> Field-based QSAR

RMSE 0.19, PRESS 0.51, RMSEpred 0.64

0

1

2

3

4

5

6

7

8

9

4.5

5

5.5

6

6.5

7

7.5

8

8.5

9

4.5 5 5.5 6 6.5 7 7.5 8 8.5 9

Pre

dic

ted

Act

ivit

y

Activity

Training Set (1)

Test Set (1)

Residuals (Train)

Residuals (Test)

Electrostatics

Sterics

And more research

> Current field similarity algorithm works well

> But could do better

> Improved force field (XED FF3)

> Formal charges

> Dielectric/solvent attenuation

> Clipping

> Up/downweighting different regions of the fields

> Use the protein to determine which parts of the field

are relevant

Lesson 8

Even when it‟s good, it could be better.

There‟s always more research to do

Lesson 9

If you didn‟t want to listen to me waffle on, you

should never have let me begin

Acknowledgements

> Andy (of course)

> Tim Cheeseright

> James Melville

> Rob Scoffin

> Brian Warrington

> Lots of other people

25 Year of Fields: What Have we Learned?

Mark Mackey