Exploring Compound Combinations in High Throughput Settings: Going Beyond 1D Metrics

Exploring
Compound
Combina1ons
in

High
Throughput
Se9ngs

Going
Beyond
1D
Metrics

Rajarshi
Guha

NCATS

June
2014,
Novar:s,
Boston.

Background

•  Cheminforma:cs
methods

–  QSAR,
diversity
analysis,
virtual
screening,

fragments,
polypharmacology,
networks

•  More
recently

–  RNAi
screening,
high
content
imaging,

combina:on
screening

•  Extensive
use
of
machine
learning

•  All
:ed
together
with
soMware

development

–  User-‐facing
GUI
tools

–  Low
level
programma:c
libraries,
APIs,

databases

•  Believer
&
prac::oner
of
Open
Source

Outline

hUp://origin.arstechnica.com/news.media/pills-‐4.jpg

Why
combine?

Physical
infrastructure
&
workﬂow

Summarizing
and
exploring
the
data

Screening
for
Novel
Drug
Combina1ons

•  Increased
eﬃcacy

•  Delay
resistance

•  AUenuate
toxicity

•  Inform
signaling
pathway

connec:vity

•  Iden:fy
synthe:c
lethality

•  Highlight

polypharmacology

Transla5onal
Interest
Basic
Interest

How
to
Test
Combina1ons

•  Many
procedures
described
in
the
literature

– Fixed
dose
ra:o
(aka
ray)

– Ray
contour

– Checkerboard

– Gene:c
algorithm

C5,D5 C5
C4,D4 C4
C3,D3 C3
C2,D2 C2
C1,D5 C1,D4 C1,D3 C1,D2 C1,D1 C1
D5 D4 D3 D2 D1 0

Mechanism
Interroga1on
PlateE

•  Collec:on
of
~
2000
small
molecules
of
diverse

mechanism
of
ac:on.

•  745
approved
drugs

•  420
phase
I-‐III
inves:ga:onal
drugs

•  767
preclinical
molecules

•  Diverse
and
redundant
MOAs
represented

AMG-47a
Lck inhibitor
Preclinical
belinostat
HDAC inhibitor
Phase II
Eliprodil
NMDA antagonist
Phase III
JNJ-38877605
HGFR inhibitor
Phase I
JZL-184
MAGL inhibitor
Preclinical
GSK-1995010
FAS inhibitor
Preclinical

Development VEGF signaling and activation
Translation Non-genomic (rapid) action of Androgen Receptor
Transcription PPAR Pathway
Regulation of lipid metabolism RXR-dependent regulation of lipid metabolism via PPAR, RAR and VDR
Cytoskeleton remodeling TGF, WNT and cytoskeletal remodeling
Cell adhesion Chemokines and adhesion
Apoptosis and survival Anti-apoptotic action of Gastrin
Development VEGF signaling via VEGFR2 - generic cascades
Some pathways of EMT in cancer cells
Development EGFR signaling pathway
0 5 10 15
-log10(pValue)
Mechanism
Interroga1on
PlateE

Top
10
enriched
GeneGo
pathway
maps

Combina1on
Screening
Workﬂow

Run
single
agent
dose
responses

6x6
matrices
for

poten1al
synergies

10x10
for
conﬁrma1on

+
self-‐cross

Acoustic dispense, 15 min
for 1260 wells, 14 min for
1200 wells"

Where
Are
We
Now?

•  382
screens
in
total

– 65,960
combina:ons

– 3,024,224
wells

•  244
cell
lines

– Various
cancers

– Mainly
human

•  Combined
with
target

annota:ons
we
can
look

at
combina:on
behavior
as
a
func:on
of

various
factors

0
50
100
150
0 500 1000 1500 2000
Number of combinations
Numberofassays

Screening
Challenges

•  A
key
challenge
is
automated
quality
control

•  Plate
level
data
employs
standard
metrics

focusing
on
control
performance

•  Combina:on
level
is
more
challenging

– Single
agent
performance

is
one
approach

– MSR
across
all
combina:on

can
provide
a
high
level
view

– But
how
to
iden:fy
bad
blocks?

QC
Examples

•  Inves:ga:ng
an:-‐malarial
combina:ons

•  300
10x10
combina:ons
in
duplicate

•  15
compounds
included
more
than
ten
:mes

-4.0
-3.5
-3.0
-2.5
-2.0
-1.5
Artemether Artesunate Dihydro
artemisinin
Halofantrine Lumefantrine
logIC50(uM)

0 5 10 15 20
MSR
Compound
10
20
30
40
Freq
QC
Examples

•  Single
agents
with
very
high
MSR’s
could
be

used
to
ﬂag

combina:ons

containing
them

•  Doesn’t
help
for

compounds
with

only
one
or
two

replicates

QC
Score

A
heuris:c
score
that
can
be

used
to
focus
on
good
quality

combina:ons

Acceptable DMSO
response
Valid single agent
curve fit & IC50
Sufficient variance in
dose sub-matrix
Spatial autocorrelation
in dose sub-matrix
Acceptable single
agent efficacy
0
250
500
750
0 2 3 5 6 7 8 10 11 12 13 15 16
QC Score
Frequency
Strain
3D7
DD2
HB3

QC
Score

QCS
=
0

QCS
=
13
QCS
=
2

•  Depends
on
mul:ple

subjec:ve
thresholds

•  Passes
some
poor

quality
blocks

•  Quickly
ﬁlters
out
very

bad
combina:ons

Repor1ng
Combina1on
Results

Repor1ng
Combina1on
Results

•  These
web
pages
and
matrix
layouts
are
a

useful
ﬁrst
step

•  Does
not
scale
as
we
grow
MIPE

•  Need
beUer
ways
of
ranking
and
aggrega:ng

combina:on
responses
taking
into
account

– Response
matrix

– Compounds,
targets
and
pathways

– Clinical
status
and
other
external
informa:on

Network
Representa1ons

Combina:on
screens
lend
themselves
naturally

to
network
representa:ons

●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
∆ Bliss+
−4.3
−3.8
−3.3
−2.9
−2.4
−1.9
−1.4
−1.0
−0.5
0.0
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
∆ Bliss+
−3.4
−3.1
−2.7
−2.3
−1.9
−1.5
−1.2
−0.8
−0.4
0.0
immune system process
apoptotic process
transcription from RNA
polymerase II promoter
protein phosphorylation
cell communication
immune response

Network
Representa1ons

•  Things
get
more

interes:ng
when

we
have
n

m

screens

•  Can
be
simpliﬁed

using
a
variety
of

methods

– Neighborhoods

– Minimum
Spanning
Tree

●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
×

Comparing
Neighborhoods

Combina:ons
that
have
DBSumNeg
<
1st
quar:le
value
for

that
strain

3D7 DD2 HB3

Comparing
Neighborhoods

Alterna:vely,
consider
all
tested
combina:ons,

highligh:ng
distribu:on
of
synergis:c
and

antagonis:c
combina:ons

3D7 DD2 HB3

Iden1fying
the
Most
Synergis1c
Pairs

●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

When
are
Combina1ons
Similar?

•  Diﬀerences
and
their

aggregates
such
as
RMSD

can
lead
to
degeneracy

•  Instead
we’re
interested
in

the
shape
of
the
surface

•  How
to
characterize
shape?

– Parametrized
ﬁts

– Distribu:on
of
responses

0.000
0.005
0.010
0 25 50 75 100
0.00
0.02
0.04
0.06
0 25 50 75 100
0.00
0.05
0.10
0.15
0 50 100
D, p value

0
3
6
9
0.00 0.25 0.50 0.75 1.00
D
density
Similarity
via
the
KS
Test

•  Quan:fy
distance
between
response

distribu:ons
via
KS
test

– If
p-‐value
>
0.05,
we
assume

distance
is
0

•  But
ignores
the
spa1al

distribu:on
of
the
responses

on
the
concentra:on
grid

0.0
2.5
5.0
7.5
10.0
0.00 0.25 0.50 0.75
D
density
Similarity
via
the
Syrjala
Test

•  Syrjala
test
used
to
compare

popula:on
distribu:ons

over
a
spa:al
grid

– Invariant
to
grid
orienta:on

– Provides
an
empirical
p-‐value

•  Less
degenerate
than
just

considering
1D
distribu:ons

Syrjala,
S.E.,
“A
Sta:s:cal
Test
for
a
Diﬀerence
between
the
Spa:al
Distribu:ons
of
Two
Popula:ons”,
Ecology,
1996,
77(1),
75-‐80

Ibru1nib
Combina1ons
For
DLBCL

•  Primary
focus
is
on
inves:ga:ng
combina:ons

with
Ibru:nib
for
treatment

of
DLBCL

– Btk
inhibitor
in
Phase
II
trials

– Experiments
run
in
the
TMD8

cell
line,
tes:ng
for
cell
viability

Mathews-‐Griner,
Guha,
Shinn
et
al.
PNAS,
2014,
in
press

Viable
Cells
(% DMSO)
Ibrutinib* (nM)
MK-2206 (µM)
Ibrutinib
MK-2206
Ibrutinib* +
MK-2206

Clustering
Response
Surfaces
0.00.20.40.60.8
C1
(24)

C2(47)

C3(35)

C4(24)

response to stress
peptidyl-tyrosine phosphorylation
cell cycle checkpoint
interphase
peptidyl-amino acid modification
negative regulation of cell cycle
cellular process involved in reproduction
ubiquitin-dependent protein catabolic process
regulation of interferon-gamma-mediated signaling pathway
macromolecule catabolic process
0 1 2 3
-log10(Pvalue)
Cluster
C3

•  Vargatef,
vorinostat,

ﬂavopiridol,
…

•  Not
par:cularly

speciﬁc
given
the

range
of
primary

targets

0.000.050.100.150.200.250.30
302
281
128
174
285
153
177
210
144
35
60
457
180
39
111
272
288
166
231
104
106
417
319
44
218
279
219
121
119
34
102
286
230
178
179

Cluster
C4

•  Focus
on
sugar

metabolism

•  Ruboxistaurin,

cycloheximide,
2-‐
methoxyestradiol,
…

•  PI3K/Akt/mTOR

signalling
pathways
glycogen metabolic process
regulation of glycogen biosynthetic process
glucan biosynthetic process
glucan metabolic process
cellular polysaccharide metabolic process
regulation of generation of precursor metabolites and energy
peptidyl-serine phosphorylation
cellular macromolecule localization
regulation of polysaccharide biosynthetic process
cellular carbohydrate biosynthetic process
0 1 2 3
-log10(Pvalue)
0.000.020.040.060.08
361
254
215
164
143
82
125
327
241
194
145
116
139
371
163
165
384
339
322
217
184
150
52
136

Combina1ons
across
Cell
Lines

•  Cellular
background
affects
responses

•  Can
we
group
cell
lines
based
on
combina:on

response?

•  Or
find
“fingerprints”
that
characterize
cell
lines?

Working
in
Combina1on
Space

•  Each
cell
line
is
represented
as
a
vector
of

response
matrices

•  “Distance”
between
two

cell
lines
is
a
func:on
of
the

distance
between
component

response
matrices

•  F
can
be
min,
max,
mean,
…

L1
L2

=
d1

=
d2

=
d3

=
d4

=
d5

D L1, L2( )= F({d1,d2,…,dn})
,

,

,

,

,

Many
Choices
to
Make

01234
KMS-34
INA-6
L363
OPM-1
XG-2
FR4
AMO-1
XG-6
MOLP-8
ANBL-6
KMS-20
XG-7
OCI-MY1
XG-1
8226
EJM
U266
KMS-11LB
SKMM-1
MM-MM1
sum
0.00.10.20.30.40.50.6
L363
OPM-1
XG-2
KMS-20
XG-1
XG-7
ANBL-6
OCI-MY1
U266
XG-6
INA-6
MOLP-8
AMO-1
KMS-34
KMS-11LB
SKMM-1
MM-MM1
EJM
FR4
8226
max
0.000.050.100.150.200.25
INA-6
MM-MM1
8226
XG-1
U266
ANBL-6
SKMM-1
EJM
OPM-1
XG-2
OCI-MY1
KMS-20
L363
KMS-11LB
AMO-1
XG-6
FR4
KMS-34
MOLP-8
XG-7
min
0.00.20.40.60.81.01.2
L363
OPM-1
XG-2
KMS-34
INA-6
KMS-11LB
SKMM-1
EJM
U266
MM-MM1
FR4
AMO-1
XG-6
8226
MOLP-8
ANBL-6
OCI-MY1
XG-1
KMS-20
XG-7
euc

•  Vargatef
exhibited
anomalous
matrix

response
compared
to
other
VEGFR
inhibitors

Exploi1ng
Polypharmacology

Vargatef

Linifanib Axitinib Sorafenib Vatalanib
Motesanib Tivozanib Brivanib Telatinib
Cabozantinib Cediranib BMS-794833 Lenvatinib
OSI-632 Foretinib Regorafenib

Exploi1ng
Polypharmacology

•  PD-‐166285
is
a
SRC
&

FGFR
inhibitor

•  Lestaurnib
has

ac:vity
against
FLT3

Vargatef DCC-2036 PD-166285 GDC-0941
PI-103 GDC-0980 Bardoxolone methyl AT-7519AT7519
SNS-032 NCGC00188382-01 Lestaurtinib CNF-2024
ISOX Belinostat PF-477736 AZD-7762
Chk1 IC50 = 105 nM
VEGFR-1
VEGFR-2
VEGFR-3
FGFR-1
FGFR-2
FGFR-3
FGFR-4
PDGFRa
PDGFRb
Flt-3
Lck
Lyn
Src
0 200 400 600
Potency (nM)
Hilberg,
F.
et
al,
Cancer
Res.,
2008,
68,
4774-‐4782

Predic1ng
Synergies

•  Related
to
response
surface
methodologies

•  LiUle
work
on
predic:ng
drug
response
surfaces

– Peng
et
al,
PLoS
One,
2011

– Jin
et
al,
Bioinforma1cs,
2011

– Boik
&
Newman,
BMC
Pharmacology,
2008

– Lehar
et
al,
Mol
Syst
Bio,
2007
&

Yin
et
al,
PLoS
One,
2014

•  But
synergy
is
not
always
objec:ve
and
doesn’t

really
correlate
with
structure

Structural
Similarity
vs
Synergy

beta gamma
ssnum Win 3x3
0.1
0.2
0.3
0.4
0.1
0.2
0.3
0.4
0.1
0.2
0.3
0.4
0.1
0.2
0.3
0.4
0.85 0.90 0.95 1.00 1.05 1.10 1.15 0.75 0.85 0.95 1.05
0 5 10 15 20 25 -40 -30 -20 -10 0
Synergy measure
Similarity

Predic1on
Strategy

•  Don’t
directly
predict
synergy

•  Use
single
agent
data
to
generate
a
model

surface

•  Predict
combina:on
responses

•  Characterize
synergy
of
predicted
response

with
respect
to
model
surface

•  Reduced
to
a
mixture
predic:on
problem

•  Need
to
incorporate
target
connec:vity

Conclusions

•  Use
response
surfaces
as
ﬁrst
class
descriptors
of

drug
combina:ons

–  Surrogate
for
underlying
target
network
connec:vity
(?)

•  Response
surface
similarity
based
on
distribu:ons
is

(fundamentally)
non-‐parametric

•  Going
from
single
-‐
chemical
space
to
combina:on

space
opens
up
interes:ng
possibili:es

•  Manual
inspec:on
is
s:ll
a
vital
step

Acknowledgements

•  Lou
Staudt

•  Beverly
Mock,
John
Simmons

•  Lesley
Griner,
Craig
Thomas,
Marc
Ferrer,

Bryan
MoU,
Paul
Shinn,
Sam
Michaels

Exploring Compound Combinations in High Throughput Settings: Going Beyond 1D Metrics

More Related Content

What's hot

Similar to Exploring Compound Combinations in High Throughput Settings: Going Beyond 1D Metrics

More from Rajarshi Guha

Recently uploaded

Exploring Compound Combinations in High Throughput Settings: Going Beyond 1D Metrics