3. Mechanism Interrogation PlateE
• 1911 small molecules, with a primary focus on
oncology, but also addressing infectious
disease and stem cell biology
• Diverse and redundant MoA’s
• Employed in 1-vs-all & all-vs-all modes
AMG-47a
Lck inhibitor
Preclinical
belinostat
HDAC inhibitor
Phase II
GSK-1995010
FAS inhibitor
Preclinical
Approved
Phase III
Phase II
Phase I
Preclinical
Other
5. Where are we now?
• 81 projects, 773 screens
• 140,730 combinations
• 4.8M wells
• 320 cell lines
• Opportunities to look at
global trends in combination
behavior in the context of
physicochemical properties,
biological functionality, …
0
50
100
150
200
2011 2012 2013 2014 2015 2016
Year
NumberCombinationScreens
• Cancers
• Hodgkins lymphoma
• DLBCL
• Neuroblastoma
• Leukemia
• Malaria
• Transcriptional mechanics
Baranello, L et al, Cell, 2016
Jun, W et al, PNAS, 2016
Lewis, R et al, J. Cheminf, 2015
Bogen, D et al, Oncotarget, 2015
Mott BT et al, Sci Rep, 2015
Zhang, M et al, PNAS, 2015
Ceribelli, M et al, PNAS, 2014
Mathews, L et al, PNAS 2014
6. Digging into the data
• Lots of data across lots of cell lines for lots of (mostly
annotated) compounds
• How can we slice & dice?
• How do we characterize quality of combination response?
• Are there global trends in synergy based on target class, MoA,
chemical structure/property?
• What is the role of selectivity vs promiscuity?
• What is the relation between single & combination responses?
• Can we better prioritize large sets of combinations?
• Can we find interesting subsets of combinations?
• Are there alternatives to the table view?
• How does (can) the data inform us on polypharmacology?
• How do we prospectively predict combination responses
7. Quantifying combination quality
• A key challenge is automated quality control
• Control separation
– control performance ≠ combination performance
• Intra-plate or inter-plate pattern
– no room for lots of replicates and
– the assumption used in primary screen can’t be satisfied
• Data consistency
– IC50 not always available (we are searching for synergy!)
– Consistent single agent IC50 ≠ consistent synergy
Lu Chen (NCATS)
8. Deviation of block
control
mQC: Interpretable QC model
Feature name Importance Explanation
dmso.v 20.71 Normalized response of the negative control
smoothness.p 18.88 p-value for smoothness
moran.p 18.82 p-value for spatial autocorrelation (tested by Moran’s I)
mono.v 12.62 Likelihood of monotonic dose responses
sa.min 12.84 The smaller relative standard deviation of the single-agent dose response
sa.matrix 8.78 The relative standard deviation of the dose combination sub-matrix
sa.max 7.36 The larger relative standard deviation of the single-agent dose response
Smoothness Randomness Monotonicity Activity variance
Feature importance encoded by mQC is consistent with human intuition
Chen, L. et al, Sci. Rep., submitted https://matrix.ncats.nih.gov/mQC/ Lu Chen (NCATS)
9. Visualization & Ranking
3D7 DD2 HB3
Azalomycin−B
ABT−263 (Navitoclax)
Cabozantinib
AZD−2014
Selumetinib
Volasertib
Midostaurin
SB−415286
IC−87114
GDC−0941
Neratinib
NCGC00021305
LY2157299
GMX−1778
PCI−32765
Torin−2
BEZ−235
Ruxolitinib
INK−128
Tipifarnib
MK−2206
PD 0325901
Imatinib
G−Strophanthin
Ketotifen
Clomipramine
NCGC00014925
2−Fluoroadenosine
MK−0752
Rolipram
Alvespimycin hydrochloride
Ganetespib
NCGC00183656
Sulindac
Carfilzomib
Bardoxolone methyl
LLL−12
JQ1
Suberoylanilide hydroxamic acid
Panobinostat
Azalo
m
ycin
−B
ABT−263
(N
avitocla
x)
C
abozantin
ib
AZD
−2014
Selu
m
etin
ib
Vola
sertib
M
id
ostaurin
SB−415286
IC
−87114
G
D
C
−0941
N
eratin
ib
N
C
G
C
00021305
LY2157299
G
M
X−1778
PC
I−32765
Torin
−2
BEZ−235
R
uxolitin
ib
IN
K−128
Tip
ifarnib
M
K−2206
PD
0325901
Im
atin
ib
G
−Strophanthin
Ketotifen
C
lo
m
ip
ram
in
e
N
C
G
C
00014925
2−Flu
oroadenosin
e
M
K−0752
R
olipram
Alvespim
ycin
hydrochlo
rid
e
G
anetespib
N
C
G
C
00183656
Sulindac
C
arfilzom
ib
Bardoxolo
ne
m
ethyl
LLL−12JQ
1
Suberoyla
nilid
e
hydroxam
ic
acid
Panobin
ostat
DBSumNeg
(−7,−4]
(−4,−3]
(−3,−2]
(−2,−1]
(−1,0]
Azalomycin−B
ABT−263 (Navitoclax)
Cabozantinib
AZD−2014
Selumetinib
Volasertib
Midostaurin
SB−415286
IC−87114
GDC−0941
Neratinib
NCGC00021305
LY2157299
GMX−1778
PCI−32765
Torin−2
BEZ−235
Ruxolitinib
INK−128
Tipifarnib
MK−2206
PD 0325901
Imatinib
G−Strophanthin
Ketotifen
Clomipramine
NCGC00014925
2−Fluoroadenosine
MK−0752
Rolipram
Alvespimycin hydrochloride
Ganetespib
NCGC00183656
Sulindac
Carfilzomib
Bardoxolone methyl
LLL−12
JQ1
Suberoylanilide hydroxamic acid
Panobinostat
Azalo
m
ycin
−B
ABT−263
(N
avitocla
x)
C
abozantin
ib
AZD
−2014
Selu
m
etin
ib
Vola
sertib
M
id
ostaurin
SB−415286
IC
−87114
G
D
C
−0941
N
eratin
ib
N
C
G
C
00021305
LY2157299
G
M
X−1778
PC
I−32765
Torin
−2
BEZ−235
R
uxolitin
ib
IN
K−128
Tip
ifarnib
M
K−2206
PD
0325901
Im
atin
ib
G
−Strophanthin
Ketotifen
C
lo
m
ip
ram
in
e
N
C
G
C
00014925
2−Flu
oroadenosin
e
M
K−0752
R
olipram
Alvespim
ycin
hydrochlo
rid
e
G
anetespib
N
C
G
C
00183656
Sulindac
C
arfilzom
ib
Bardoxolo
ne
m
ethyl
LLL−12JQ
1
Suberoyla
nilid
e
hydroxam
ic
acid
Panobin
ostat
DBSumNeg
(−7,−4]
(−4,−3]
(−3,−2]
(−2,−1]
(−1,0]
0.00.20.40.60.8
10. LogP & Synergy?
• Yilancioglu et al (JCIM 2014) suggested that you can
predict synergicity using only logP
• Synergicity of a compound is the frequency of synergistic
pairs involving the compound
Synergy doesn’t correlate with logP
10
20
30
-4 0 4 8
logP
Numberofsynergisticcombinations
Synergicity may correlate with logP
http://blog.rguha.net/?p=1265
12. -3
-2
-1
0
0.0 0.1 0.2 0.3 0.4
Tanimoto Similarity
DBSumNeg
Structural similarity vs synergy?
• Do structurally
similar compounds
lead to synergistic
combinations?
• No reason they
should
• Synergy driven by
(off-)targets
16. Explicitly consider targets
Descriptors used for learning
Three classes of descriptors generated per combination
• StructuralFingerprint
• Morgan, 2,048 bits, radius 2 (RDKit).
• PredictedTargets
• 1,080 human target probabilities of affinity
(PIDGIN V1)
• Combined
• StructuralFingerprint and PredictedTargets.
Input data required:
• Compound structure for training and test data (names, SMILES)
• Combination data (which compounds, synergy score)
Output:
• New combinations predicted to be synergistic
• Probability of being synergistic (classifier model,
worked best for this project)
• Predicted synergy value (quantitative model,
did not work so well for this project)
Dan Mason, Andreas Bender (U. Cambridge)
18. Outlook
• Accurate predictions will enable virtual screening of
combinations
• Many aspects of the process are yet to be explored
• Differential analysis of combination response
• Are some pathways or mechanisms more amenable to
combination screening than others?
• Viability is easy to measure. What about other readouts?
• Is there a better way to characterize synergy?
• Tang, J. et al, Frontiers. Pharmacol., 2015
https://tripod.nih.gov/matrix-client
19. Acknowledgements
• Lu Chen
• Alexey Zakharov
• Kelli Wilson
• Mindy Davis
• Xiaohu Zhang
• Richard Eastman
• Bryan Mott
• Craig Thomas
• Marc Ferrer
• Paul Shinn
• Crystal McKnight
• Carleen Klumpp-
Thomas
• Anton Simeonov
• Dan Mason
• Rich Lewis
• Yasaman Kalantar
Motamedi
• Krishna Bulusu
• Andreas Bender