A physical sciences network characterization of nonmalignant and metastatic cells

hushedsnailInternet and Web Development

Nov 12, 2013 (3 years and 6 months ago)

88 views


1

A physical sciences network characterization

of

nonmalignant and metastatic cells

SUPPLEMENTARY MATERIAL

-
materials and methods

-
supplementary figures and captions


Authors:
The
Physical Sciences


Oncology

Centers

Network
1

PS
-
OC Network

1.

Principal Investigator: Paul C.W. Davies

Senior Co
-
Investigator: William M. Grady

Arizona State University Physical Sciences
-
Oncology Center

Arizona State University, Tempe, AZ

2.

Principal Investig
ator: Michael L. Shuler

Senior Co
-
Investigator: Barbara L. Hempstead

Cornell University Physical Sciences
-
Oncology Center

Cornell University, Ithaca, NY

3.

Principal Investigator: Franziska

Michor

Senior Co
-
Investigator: Eric C. Holland

Dana
-
Farber Cancer Ins
titute Physical Sciences
-
Oncology Center

Dana
-
Farber Cancer Institute, Boston, MA

4.

Principal Investigator: Robert A. Gatenby

Senior Co
-
Investigator: Robert J. Gillies

H. Lee Moffitt Cancer Center & Research Institute Physical Sciences
-
Oncology Center

H. Lee

Moffitt Cancer Center & Research Institute, Tampa, FL

5.

Principal Investigator: Denis Wirtz
2

Senior Co
-
Investigator: Gregg L. Semenza

Johns Hopkins University Physical Sciences
-
Oncology Center

Johns Hopkins University, Baltimore, MD

6.

Principal Investigator:
Alexander van Oudenaarden

Senior Co
-
Investigator: Tyler Jacks

Massachusetts Institute of Technology Physical Sciences
-
Oncology Center

Massachusetts Institute of Technology, Cambridge, MA

7.

Principal Investigator: Mauro Ferrari

Senior Co
-
Investigator: Steven
A. Curley

The Methodist Hospital Research Institute Physical Sciences
-
Oncology Center

The Methodist Hospital Research Institute, Houston, TX




1

Comprehensive list of authors

and affiliations appear at the end of the paper.

2

To whom correspondence should be addressed. E
-
mail:
wirtz@jhu.edu
.


2

8.

Principal Investigator: Thomas V. O’Halloran

Senior Co
-
Investigator: Jonathan D. Licht

Northwestern University
Physical Sciences
-
Oncology Center

Northwestern University, Chicago, IL

9.

Principal Investigator: Robert H. Austin

Senior Co
-
Investigator: Thea

Tlsty

Princeton University Physical Sciences
-
Oncology Center

Princeton University, Princeton, NJ

10.

Principal Investig
ator: Peter Kuhn

Senior Co
-
Investigator: Kelly J. Bethel

The Scripps Research Institute Physical Sciences
-
Oncology Center

The Scripps Research Institute, La Jolla, CA

11.

Principal Investigator: Jan Liphardt

Senior Co
-
Investigator: Valerie M. Weaver

University

of California
-
Berkeley Physical Sciences
-
Oncology Center

University of California
-
Berkeley, Berkeley, CA

12.

Principal Investigator: W. Daniel Hillis

Senior Co
-
Investigator: David B. Agus

University of Southern California Physical Sciences
-
Oncology Center

Uni
versity of Southern California, Los Angeles, CA






3

Supplementary Information:
Methods

I.

Morphology


A.

Differential Interference Contrast Microscopy

Breast cancer cell lines were grown on number 1.5 cover glass
(Carl Zeiss, Thornwood, NY)
, nuclei were
stained
with DAPI (Serotec
, Oxford, UK
), cells were washed with
PBS

and then mounted on microscope
slides using Fluoromount G (Southern Biotech
, Birmingham, AL
). Cells were imaged using Köhler
illuminated Nomarski differential interference contrast (DIC) optics wi
th a Zeiss 63X oil immersion 1.40
NA plan apochromat lens on a Zeiss Axiovert 200M microscope (Carl Zeiss, Thornwood, NY). The
aspect ratio
s

of the cells were

measured

using IMAGE J software (National Institutes of Health, Bethesda,
MD).

The

Jarque
-
Bera te
st was used to establish the normality of the aspect ratio data for each of the cell
types.


A 1
-
way ANOVA on the aspect ratio data
was performed
;

a
p
-
value of
0
.001

was obtained
.


B.

Single
-
cell Tomographic Imaging and 3D Morphometry

3D imaging of cells was
accomplished by optical CT using the Cell
-
CT
TM

instrument (VisionGate, Inc.,
Phoenix, AZ) that provides isotropic spatial resolution of 350 nm.


Cells were fixed with CytolLyt (Cytyc,
Malborough, MA
) and stained with 6.25% (w/w) hematoxylin and 1% (w/w) e
o
sin (Electron Microscopy
Sciences, Hatfield, PA), dehydrated with ethanol, and embedded into a carrier gel (Smart Gel, Nye
Lubricants, Fairhaven, MA). Cells were imaged sequentially by flowing the carrier gel through a rotating
glass capillary. For each ce
ll, a volumetric image was generated by acquiring 500 projections taken at
angular intervals of 0.72 degrees around the cell and subjecting the
m

to mathemati
cal reconstruction
algorithms
1
. Prior to reconstruction, we removed background noise and aligned the projections
to remove
pattern noise artifacts and to compensate for mechanical jitter and run
-
out of the capillary, respectively.
3D image processing algorithms
2

were used to quantify morphological parameters such as nuclear
sphericity. One hundred volumetric images
of each cell type were analyzed and s
tatistical analysis was
performed by two
-
tailed unpaired t
-
test.


C.

Partial Wave S
pectroscopy

PWS instrumentation and measurement technique
s have

previously been described

3
.
For

cell culture
,
25,000 cells
were plated in each chamber of two
-
well sterile glass chamber slides (Lab
-
Tek
Chamber
Slide System, NUNC, Rochester, NY) and incubated for 6 h. Cells were fixed with 70% cold ethanol
(v/v) overnight prior to PWS measurements. All experiments were carried out for early passages 2
-
5.
Nuc
lear disorder was quantified in

approximately 60

cells for each cell line. All p

values

were calculated
using standard

t
-
tests. The effect
-
size between two groups was calculated from the average disorder
strength,
)
(
g
d
L
and its standard deviation,
)
(
g

:



4

2
2
2
1
2
1








size
Effect

where
1


and
1


can be thought as the
)
(
g
d
L
and
)
(
g

for the MCF
-
10A cell population. Similarly,
2

and
2


can be assumed
to be

)
(
g
d
L

and
)
(
g

for the MDA
-
MB
-
231 cells.
This treatment

take
s

into account the
statistical significance of the average disorder strength difference
(
i.e.,
d
L

, for MCF
-
10A and MDA
-
MB
-
231
)
.

T
o minim
ize sample variability, the MDA
-
MB
-
231 cell L
d

values were normalized with
respect to MCF
-
10A values. For example, if the L
d

of MCF
-
10A is 1, the L
d

of MDA
-
MB
-
231 cells will
be 1.80.

D.

Matrix S
tiffness

Cells were trypsinized and plated on
ECM
-
crosslinked
polyacrylamide gels (2
,
000 cells/18

m
l

gel). MCF
-
10A cells were grown in 3D in media containing 2% (v/v) recombinant basement membrane (rBM)
as
described
4

and MDA
-
MB
-
231 cells were grown in 3D in
Dulbecco's Modified Eagle Medium

with
high
glucose, 10%
fetal bovine serum

and
2%

rBM. ECM
-
crosslinked
polyacrylamide

gels were prepared as
previously described
5

and
6
. Cells were processed for immunofluorescence with the following antibodies
and reagents: polyclonal Ki67 (Ab
cam, Cambridge, MA), Alexa Fluor 488 conjugated phalloidin, and
Alexa Fluor 594 rabbit IgGs (Invitrogen, Carlsbad, CA).Cells were visualized using a scanning confocal
laser attached to a Zeiss LSM 510

fluorescence microscope and images were recorded at 63X

magnification
(
numerical aperture

1.4). Cell proliferation was measured by calculating the percentage of
cells with Ki67
-
labeled nuclei. On average at least 120 cells where counted per condition. The Ki67 cell
proliferation data were statistically analyze
d using a two
-
tailed unpaired t
-
test.


E.

CD44
and
l
ipid
r
aft distribution
assay
s

Cells were detached using 5
-
10 mL of enzyme
-
free cell dissociation buffer (Invitrogen, Carlsbad, CA) for
5 min (MDA
-
MB
-
231) or 15 min (MCF
-
10A) to preserve the cell surface anti
gen presentation. Complete
media was added once
the cells detached, cells were
washed twice in 1X Dulbecco’s phosp
hate
-
buffered
saline (DPBS) without Ca
2+

or Mg
2+

(Invitrogen
)
) and centrifuged at 1
,
000 RPM

at

4

°C for 5 min. Cells
were then incubated over
ice with mouse FITC
-
conjugated anti
-
CD44 antibody clone G44
-
26 (BD
Biosciences, San Diego, CA) and Alexa Fluor 555 Vybrant
®

Lipid
Raft Labeling Kits (Invitrogen
) as well
as appropriate isotype controls

for 45 min. Cells were washed three times with 1mL of
1X DPBS without

Ca
2+

or Mg
2+

to remove unbound antibody.
Th
e lipid raft kit
uses an antibody specific for cholera toxin
subunit B (CT
-
B) that binds to the pentasaccharide chain of
plasma
membrane ganglioside GM1, which
selectively partitions into lipid raf
ts.



5

Prior to TIRF imaging, cells were fixed in cold 4% (v/v) paraformaldehyde (Electron Microscopy
Sciences, Hatfield, PA) over ice

for 30 min.
TIRF
and epifluorescence
images were taken with a Nikon
Ti
-
E/B automated inverted microscope with Perfect Focus

and automated TIRF angle

and

equipped with
a CFI APO 100X/1.45

oil TIRF objective
lens
and an EMCCD Detector. A 488
-
nm Argon laser was used
to detect CD44 FITC
-
conjugated antibody. A 561
-
nm laser was used to detect
Alexa Fluor 555
Vybrant
®

Lipid Raft dye.

Areas of fluorescence in the TIRF images were analyzed using ImageJ (National Institutes
of Health) and plotted using Prism 5.0b for Microsoft (GraphPad Software, San Diego, CA). Two
-
tailed
unpaired t
-
tests were employed with a significance level of α=0.0
5 where applicable.

E
ach e
xperiment
was
conducted

in triplicate
.


I.

Motility and Mechanics

A.

1D, 2D and 3D M
igration

1.

Measurement of Maximum D
isplacement

Maximum displacement from the origin was taken as the distance between the starting position of a cell
and

the furthest point from the starting position that it reached at any time during the 16.5 h experiment.
One value was obtained for each cell measured, and these values were averaged to give the value reported
herein.


2.

1D Migration: Device P
reparation an
d
T
ime
-
lapse M
icroscopy

The 13 micron
-
wide silicon microchannels were pattern
ed and etched to a depth of 20

m using standard
photolithography techniques and a Deep RIE etcher (Unaxis 770, Plasmatherm, St. Petersburg, FL). The
device was cleaned under an oxygen plasma treatment for 1 min before a slab of cured
PDMS

about
500


m thick was placed over one
-
half of t
he device. While the device was still hydrophilic, the micro
-
channels were coated through capillary action with 50

g/ml fibronectin (Sigma, St.

Louis MO
) in
1X

PBS solution. The device was washed with 1X PBS and placed inside a petri dish containing fresh

growth medium. Cells were seeded on top of the device at a density of 100,000 cells/cm
2
.
The presence of
the PDMS slab prevented cells from settling down on top of the microchannels. After an incubation
period of 8
-
12 h, the slab was removed and the petri

dish placed inside a 5% CO
2

and 37

°C temperature
controlled mic
roscope incubator. A computer
-
controlled microscope recorded a picture of the array of
channels every 5 min for 8 h. After 8 h, a 7.5 nM dose of

paclitaxel
was added to the media
,

and

images
were
recorded for an additional 8 h. Resulting images were processed using ImageJ in order to highlight
the motile cells in each image. A particle tracking algorithm written in MATLAB (Mathworks, Natick,
MA) was used to reconstruct the motion of each cell
inside the microchannels. The instantaneous speed of

6

a cell was averaged over the duration of the experiment to obtain the average speed. The maximum radial
displacement over the course of the experiment wa
s used to compute the invasion

distance.


3.

2D
Migration: 2D C
olla
gen Substrate

2D cell culture plates were prepared by adding soluble rat tail type I collagen in acetic acid (BD
Biosciences, San Jose, CA) to achieve coverage of 32.48

g/cm
2

and incu
bated at room temperature for
2

h. Plates were washed

gently with 1x ph
osphate
-
buffered saline (PBS) and plated with a low density of
cells. Migration studies were conducted as previously described
7
.


4.

3D Migration: 3D Collagen M
atrix

Cell
-
impregnate
d 3D collagen matrices were prepared as previously described
7
.

A collagen concentration
of 2 mg/ml was used so that the average matrix pore size (<1

m) was significantly smaller than
the
cell
body and nucleus. Cell density was kept low (20,000 cells per gel) so as to ensure that only single cell
measureme
n
ts were taken. Mean and standard error of the mean (SEM) values
were calculated and
t
wo
-
tailed unpaired t
-
tests were conducted using Graphpad Prism.



B.

Traction Force Microscopy

Polyacrylamide gels of specific Young’s moduli (5 kPa) were synthesized as prev
iously described
8

with
a constant ratio of 0.175% (v/v) bis
-
acrylamide to 7.5% (v/v) acrylamide in the gel solution mixture a
nd
derivatized with 0.1, 10, or

50

g/mL of laminin from Engelbreth
-
Holm
-
Swarm murine sarcoma
basement membrane (Sigma, St. Louis, MO). Cells were seeded on
polyacrylamide

substrates embedded
with 0.5
-

m fluorescent beads (Invitrogen, Carlsbad, CA), allowed to a
ttach and spread overnight, an
d

imaged in a custom temperature, humidity, and CO
2
-
controlled envir
onment on a Zeiss Axio Observer
Z1m inverted phase microscope with a Hamamatsu ORCA
-
ER camera. Traction forces were determined
as previously described

8
. Data were analyzed with Tukey’s Honestly S
ignificant Difference test or t
-
test

after natural logarithm transformation to ensure assumptions of normality and equal variance. Sta
tistical
tests were performed using

JMP software (SAS Institute, Cary, NC
)
and significance was considered
p
<0.05.


C.

Shear Stresses on Rolling Cells on E
-
Selectin

1.

Cell P
reparation

Cells were trypsinized for 3 min (MDA
-
MB
-
231) or 15 min (MCF
-
10A) and then recovered in com
p
lete
media in the incubator for 2 h before experiment to
ensure normal surface receptor expression. Cells were
washed twice with 1X DPBS, centrifuged at 1,000 RPM

at 4


C and resuspended in the flow buffer at a

7

concentration of 10
6
cells/mL. The flow buffer consisted of PBS containing Mg
2+

(100 mg/mL) and
satura
ted with Ca
2+
. At least 90% viability of cells was confirmed using trypan blue stain.


2.

Preparation of Immobilized Protein S
urfaces

Recombinant human E
-
selectin
-
IgG chimeric protein (R&D,
Minneapolis, MN)

was dissolved in
1X

DPBS to a final concentration of 5

g/mL. The surface was first rinsed with 75% (v/v) ethanol and
then 1X DPBS. The surface was subsequently incubated with 10

g/mL protein
-
G (EMD Chemicals,

Gibbstown, NJ) solution for 1.5 h, followed by a 2 h incubation w
ith E
-
selectin chimera then blocked
with 5% (w/v) milk protein in 1X DPBS for 1 h. Control tubes were bloc
ked with 5% milk protein in
1X

DPBS for 1 h.


3.


Rolling E
xperiment

Micro
-
Renathane

microtubing (300

m ID; Braintree Scientific, Braintree
, MA) was
cut to a length of
50

cm, and secured to the stage of the Olympus IX81 motoriz
ed inverted research microscope

after
surface functionalization as described above. A CCD camera (model no: KP
-
M1AN, Hitachi, Tokyo,
Japan) and a DVD recorder were used to record

experiments for offline analysis. Flow of cell suspension
at a concentration of 10
6

cells/mL through the microtubes was produced by a syringe pump (KDS 230,
IITC Life Science, Woodland Hills, CA). For both surface conditions, cells were introduced into th
e
microtubes at a wall shear stress of 1.0 dyn/cm
2
.


4.

Data A
nalysis

Rolling

cells were defined as those observed to translate in the direction of flow with an average velocity
less than 50% of the calculated hydrodynamic free stream velocity. The rolling
velocity was calculated by
measuring the distance a rolling cell traveled over a 30
-
second interval. Videos of rolling cells were taken
at three randomly selected locations along the microtube. The quantity of cells rolling or adherent to the
surface was d
etermined by recording images at 30 randomly selected locations along the microtube. All
errors are reporte
d as SEM values

and two
-
tailed unpaired t
-
test
s were performed using GraphPad Prism
.


D.

Extracellular Matrix Isolation and I
mmunofluorescence

Cells wer
e grown in a 12 well plate (tissue culture treated; Nunc
,
Fisher Scien
tific) containing 1 12
-
mm
glass coverslip/well (Fisher Scientific, Pittsburgh PA). Matrix isolation was performed accor
ding to an
established protocol
9
. For scanning electr
on microscopy
, isolated ECM components were fixed
in a
glutaraldehyde/formaldehyde containing buffer (3% (v/v) formaldehy
de, 1.5% (v/v) glutaraldehyde,
0.1

M

sodium cacodylate, 5

mM

MgCl
2
, 2.3

M

sucrose, pH

7.4
) for 20 min and washed with 1x

PBS.

8

Samples were post
-
fixed with 1% (v/v) osmium tetroxide for
20 min (Sigma, Allentown, PA), followed
by dehydration in ethanol. Samples were critical point dried (Tou
simis 795) and coated with
2

nM

platinum using a sputter coater (Anatech Hummer 6.2 Sputter Coater, San Diego, CA). Samples
were visualized using a
Qua
nta 200 environmental scanning electron microscope (FEI,
Hillsboro,
Oregon
)
. For immunofluorescence staining, MCF
-
10A and MDA
-
MB
-
231 cells were cultured for 6, 9
and 12 d in respective growth media, fixed with 3.7% (v/v) paraformaldehyde for 30 min, washed

in
1X

PB
S and incubated with 50

g/ml FITC
-
tagged hyaluronan (Sigma) for 1 h. Coverslips were mounted
using 1 drop of fluorescent mounting media (Dako, Denmark).Hyaluronic acid receptors were visualized
using an Olympus BX60 microscope and images were cap
tured and acquired using constant exposure
time in MagnaFire
TM

(Melville, NY).


E.

Hyaluronic Acid
Micropatterning

Standard photolithographic techniques were utilized to fabricate silicon masters with 80

m x 80


m
square patterns. PDMS pre
-
polymer elastomer solution and curing agent (Sylgard 184) were mixed in a
10:1 wt ratio, cast onto the silicon masters and cured overnight at room temperature to form a
complementary elastomeric stamp. Glass substrates were micr
opatterned with HA using a previously
established protocol
10
. Cells were grown to 80% confluence, washed with 1X PBS and digested usin
g
trypsin
-
EDTA. Cell densities between 0.75 and 1 x10
5

cells were seeded onto each HA patterned
substrate.


After a culture period of 24 h, cells were fixed using 3.7% (v/v) paraformaldehyde for 20 min, washed
with 1X PBS, and incubated with 1% BSA in 1X
PBS for 30 min to prevent
non
specific binding. Cells
were washed with 1X PBS and incubated with anti
-
human CD44 for 1 h

(1:100

dilution
; Clone A3D8;
Sigma, Saint Louis, MO). After rinsing with 1X PBS, cells were incubated with anti
-
mouse IgG Cy3
conjugate
(1:50; Sigma) for 1 h. Cells were washed with 1X PBS, counterstained with

the DNA dye

DAPI (1:1000; Roche Diagnostics) for 10 min, rinsed with 1X PBS, and mounted with glass slides using
fluorescent mounting medium. All substrates were imaged using
an
Olym
pus BX60

fluorescent
microscope
. To quantify cell adhesion, fluorescently
-
labeled substrates were analyzed to determine the
percentage of cells (via cell number and cell area) that overl
apped HA patterned regions. The

total
number of cell nu
clei and

the nu
mber of cell nuclei located direc
tly on HA patterned regions were

counted
in the field of view

and used to determine cell adhesion (cell # on HA patterns/total cell number).
All image analyses were performed
i
n triplicate experiments with triplicate fields

of view
. A

second
quantification method

validated and confirmed these findings:

image analysis

software
(Image J)
was

9

used to measure cell area using the “freehand selection” tool to outline the respective cell areas. Unpaired
two
-
tailed t
-
tests were
performed where appropriate.


F.

Atomic force microscopy (AFM)

To prepare samples for AFM measurements, cells were plated onto 50
-
mm glass
-
bottom dishes (World
Precision Instruments, Sarasota, FL) as described in
11
. Cell
s were incubated with 5 µM SYTO

9 nucleic
acid stain (Invitrogen) for 30 min prior to i
nd
entation followed by incubati
on with 5 µM CellMask Deep
Red lipid membrane stain (Invitrogen) 5 min prior to indentation, or cells were incubated with 5 µM
Nuclear ID Red (Enzo Life Sciences, Plymouth Meeting, PA) and 5 µM Nucleolar ID Green (Enzo Life
Sci
ences) 30 min prior to indentation.


The combined AFM
-
Confocal Laser Scanning Microscopy (AFM
-
CLSM) setup consists of a sample
scanning AFM (MFP
-
3D Bio, Asylum Research, CA) and a single molecule sensitive CLSM (Microtime
200, PicoQuant, Germany). The opt
ical setup allows fluorescence lifetime imaging microscopy (FLIM)
with two
-
color excitation (470 nm and 640 nm). The combined AFM
-
CLSM setup is described in detail
in
12
.



Soft silicon nitride AFM probes with nominal spring constants k ≈ 10 pN/nm (MSNL, Veeco Instrument
s,
Plainview, NY) were used for the indentation experiments. The spring constant of each cantilever was
determined from the thermal noise spectrum
13
. After the experiment, the probes were imaged with a
scanning electron microscope

to determine the tip radius. The AFM tip and the confocal volume were
aligned using the pattern of back
-
scattered light
12
.


After align
ment, the AFM tip was fully retracted and the sample stage moved until a cell of interest was
under the tip in the crosshairs of the eyepiece. A two color FLIM image of the cell was acquired while the
tip was still retracted. In the FLIM image, points of i
nterest on the cell were selected. The AFM tip was
moved directly over the first point and force

distance curves were acquired with 2 µm/s approach and
retract speeds in the continuous force curve mode with a trigger force of 600 pN. Approximately 20 such
curves were taken at each point of interest on the cell. After indentation, a subsequent FLIM image was
taken of the cell with exactly the same settings as the preliminary image. The two images were super
-
im
posed in software (Image J) to determine the exte
nt to which the cell moved during the measurements.

If the cell moved more than ~1

µm, then data were excluded. The alignment of the tip was verified as
described above and another cell of interest was located and measured. The measured force distance
curv
es were analyzed with custom written software (Igor Pro, Wavemetrics) as described in
11
.


10


G.

Ballistic injection nanorheology (BIN)

Cells were seeded on 35
-
mm plastic dish and grown to ~80% confluence before ballistic injection of
fluorescent 100
-
nm carboxylated polystyrene particles (Invitrogen, Carlsbad,
CA) using a Biolistic PDS
-
1000HE particle delivery system (Bio
-
Rad, Hercules, CA) as previously described

14
. In brief,
nanoparticles were placed on macrocarriers and allowed to dry for 2 h. Rupture disks of 1550 psi
were
used along with hepta adaptor. After ballistic injection, cells were washed with 1X PBS and incubated for
1 h before re
-
seeding onto a poly
-
L
-
lysine (Concentration/time for PLL coating, Sigma, St. Louis, MO)
treated glass
-
bottom dish (MatTek, Ashland,

MA). Particle
-
injected cells were allowed to grow overnight
before observation on a microscope. A Nikon TE2000
-
E inverted microscope equipped with a 60X oil
-
immersion, NA 1.4 objective lens and a Cascade 1K camera (Roper Scientific, Tucson, AZ) were used
to
acquire the high
-
speed image of fluorescent particles movement with 30 frames/s for 20 s. Images were
analyzed by custom software in MATLAB (The MathWorks, Natick, MA) to obtain the movement of
particles.

The

time
-
averaged mean squared displacement

(MSD
) were calculated to determine t
he random
displacement of probe

nanospheres by following formula,





2
2
)
(
)
(
)
(
)
(
)
(
t
y
t
y
t
x
t
x
MSD










where
τ

is
the
time lag and
t

the
is elapsed time.


II.

Stress Response

and Survival

A.

Cell Number and Viability in Low

pH
,
Low O
xygen
, and Paclitaxel Treatment

For low pH treatment, complete cell culture growth medium was further supplemented with 25 mM each
of PIPES and HEPES and the pH adjusted to 7.4 or 6.7 with 1N HCL or NaOH. Cells were seeded in
neutral pH (7.4) comp
lete growth medium at 17% oxygen, 5% CO
2

and 37

°C, 24 h prior to low pH
treatment.

The following day, growth medium was maintained at neutral pH (7.4) or changed to low pH
(6.8) growth media for 72 h.


For low oxygen treatment, complete cell culture
growth medium was further supplemented with 25 mM
each of PIPES and HEPES and the pH adjusted to 7.4 with 1N NaOH.

Cells were seeded in pH 7.4
growth medium at 17% oxygen, 5% CO
2

and 37

°C, 24 h prior to low oxygen treatment.

The following
day, cells were
either cultured in pH 7.4 growth me
dium at 17% or 1% oxygen for 72

h.


Cell viability was determined every 24 h using the Invitrogen Live/Dead Viability and Cytoxicity Kit (L
-
3224).

Images were collected with an automated Zeiss Observer Z.1 inverted micros
cope through a 5X

11

/0.15 N.A. objective using
green fluorescent protein

and Rhodamine filters.

Tiled mosaic images were
captured using the AxioCam MRm3 CCD camera and Axiovision version 4.7 software suite (Carl Zeiss
Inc., Germany). Quantification of viable

cells was performed using fully automated metho
ds in Definiens
Developer XD.


For paclitaxel treatment, a
pproximately 4,000 cells/well were plated in a 96
-
well plate (Corning
Incorporated #29442
-
056, Lowell, MA) in PS
-
OC media conditions. The following
day, increasing
concentrations of paclitaxel (LC Laboratories #P
-
9600, Woburn, MA) were added to the respective wells
and allowed to incubate. After a specified p
eriod of time, as indicated in Fig. 3f
, the CellTiter 96
Aqueous Non
-
Radioactive Cell prolife
ration Assay reagent was added (following manufacturer’s
instructions, Promega #G5421, Madison, WI) to the individual wells. Cell viability was determined by
reading the plate at an absorbance wavelength of 490 nm. The results are averages of four wells
per
experimental sample.



B.

3D Alginate
-
based Tumor M
odels

RGD
-
modified

calcium alginate discs (200

m thick, 4 mm in diameter) were made by suspending cells
in RGD peptide
-
modified alginate (Protanal LF; FMC Biopolymer, Philadelphia, PA; dissolved, 4%
[w/v], in serum
-
free DMEM (for MDA
-
M
B
-
231) or DMEM/F12 (for MCF
-
10A
) at a concentration of
20×10
6

cells/mL, followed by casting in a machined plexiglass mold and cross
-
linking with 0.1 M CaCl
2
,
as previously published
15
. Cell
-
seeded discs were cultured in 24
-
well plates (
one disc per well) on an
orbital shaker. Culture was performed at 37
o
C, 5% CO
2
, and either hypoxic (1%) or ambient (17 ± 1%)
O
2
, in a controlled atmosphere incubator (Thermo Fisher Scientific Inc., Waltham, MA). Culture media
was changed daily (500

L/disc
), with harvest time points at days 1, 3, and 6.


1.

Cell V
iability in 3D

Culture

Viability of cells seeded in alginate discs was
monitored

using live
-
dead staining. Alginate discs were
submerged in a live/dead staining solution (5 μM

calcein
-
AM and 5 μM

ethidium homodimer
-
1;
Invitrogen, Carlsbad, CA). The submerged discs were incubated for 30 min prior to fluorescence
microscopy. Images were taken on an inverted microscope (Axio Observer; Carl Zeiss;
excitation/emission wavelengths used were 470/525 and 5
45/605 nm for green and red fluorescence).
ImageJ enabled the merging of green and red fluorescence images. Plots represent mean
±

SD
.

Statistical

significance was assessed with

ANOVA and a Tukey post
-
test, wherein

p < 0.05 (*) was considered
significant

u
sing

Graph
p
ad Prism.


12


2.

Measurement of O
2

Consumption in 3D C
ulture

Cellular O
2

consumption was measured as previously reported
15
. Twelve cell
-
seeded 200 μm
-
t
hick
alginate discs at day 6 of either 1% or 17% O
2

3D cultures were submerged in 2 mL media in a sealed
glass chamber (Agilent Technologies, Inc., Santa Clara, CA), and kept stirring at 37
o
C. Media was
initially equilibrated at ambient O
2

and 5% CO
2
, and
reduction in O
2

level due to cellular consumption
was measured with a dissolved oxygen meter (Innovative Instruments, Inc., Tampa, FL). O
2

drawdown
measurements were taken at two
-
minute intervals for 30 min, and draw
-
downs were run in duplicate.
Consumptio
n rate was calculated from a linear fit to the O
2
level versus time and normalized to total
sample DNA content for comparisons of different conditions.


C.

Carcinoembryonic Antigen Expression under Hypoxia

Cells were grown at < 80% confluence under normal (am
bient atmosphere supplemented with 5% CO
2
)
or hypoxic (1% O
2

with 5% CO
2
) conditions fo
r 72 h. To harvest cell singletons
, cells were mildly
trypsinized (≤ 5 min, 0.05% trypsin with EDTA) and resuspended to 10
6

cells/mL in cold 1X PBS with
0.1% (w/w)
bovin
e serum albumin
. Cells were stained on ice for single
-
color flow cytometry
(FACSCalibur, BD Biosciences, San Jose, CA) with mouse anti
-
human CEA primary
antibodies
(clone
COL
-
1, BD Bio
sciences, San Diego, CA) with phycoerythrin
-
conjugated anti
-
mouse IgG se
condary
antibodies (Vector Labs, Burlingame, CA), or proper isotype controls
16
.

Mean fluorescence intensities

are equal to the geometric mean of the fluorescence intensity of each collected cell (<2
,
500 cells per
sample). Dat
a are presented as the mean ± SE
M. of
n
≥ 3 for all e
xperiments. Two
-
tailed unpaired t
-
test
was used to determine statistical significance
(
p< 0.05)
.


D.

Wound healing assay

Cells were plated in a six
-
well plate at a density of 10,500 cells/cm
2
. At confluence, a sterile 200


L
pipette tip was used to
scratch 3 parallel lines into the monolayer. The media in each well was refreshed,
and the initial time point was imaged and then every 12 h until confluency in the wound was reached.



E
.

Single
-
cell respirometry

Single
-
cell oxygen consumption rates
(OCR) measurements were performed as described elsewhere
17
.
The technique is based on enclosing individual cell in hermetically sealed glass microwells that contain
an extracellular optical oxygen sensor. Oxygen concentration changes inside of the microwells are
measured as changes in sensor emission intensity.
Cells were loaded into microwells utilizing a custom

13

built high
-
precision pump that enables manipulation of sub
-
nanoliter volumes. After loading cells were
incubated for 18
-
24 hours in an incubator under 5% CO
2

at 37
o

C in “open” condition, i.e. the wells
were
fully open for nutrient and gas exchange with the outside medium. Immediately after incubation the
microwells were sealed be placing a “lid” with integrated oxygen sensor on top. The sensor emission
intensity was measured at time intervals of 5 second
s and oxygen concentration inside of the sealed
microwells was determined using a two
-
point calibration curve. The OCR values were calculated from
the slope of the oxygen concentration time course.



III.

Molecular Signatures for Morphology, Motility, and Survi
val

A.

Proteomics Experiments

1.

Stable Isotope Labeling with Amino Acids in Cell C
ulture

(SILAC)

Cells were grown in Advanced DMEM F
-
12 Flex media (Invitrogen, # MS10033) containing 5% or 10%
(MCF
-
10A or MDA
-
MB
-
231 cells, respectively) dialyzed FBS (10kDa MW cu
t
-
off and dialyzed against
0.15M NaCl) and either
13
C
-

lysine or unlabeled lysine for six passages according to the standard SILAC
protocol
18
. MCF
-
10A cells were also supplemented with 0.5

µg/ml hydrocortisone (Sigma; #H
-
0888),

20
ng/ml hEGF (Invitrogen; #PHG0311), 10 µg/ml insulin (Sigma; #I
-
0516), and 100 ng/ml cholera toxin
(Sigma; # C
-
8052). After the completion of si
x passages in the presence of SILAC media, experiments
were set
-
up and cells then harvested.


2.

Protein Sample Preparation for LC
-
MS/MS

Cells were washed with 1X PBS to remove cellular debris and residual FBS. Approximately 2

x

10
7

cells
were lysed in 0.75
ml of 1X PBS containing 1% w/v

octyl
-
glucoside and HALT protease inhibitor
cocktail (Pierce, Rockford, IL #78430) by scraping and needle treating using a 27G1/2 needle. The cell
lysate was centrifuged at 16,000
g
, the supernatant of which was passed through a 0.22
-
µm filter. Cell
lysates were mixed 1:1 using protein quantitation results determined by the
bicinchoninic acid

assay (i.e.
,

MCF
-
10A lysate from cells grown in light lysine media was mixed 1:1 with
MDA
-
MB
-
231

lysate from
cells grown in heavy lysine).


3
. C
ell T
reatments

For ROCK inhibitor experiments, heavy labeled cells (
13
C
-
lysine) from the MCF
-
10A and MDA
-
MB
-
231 cell lines were treated with 40

M Y
-
27632 for 24 h and the corresponding light labeled ce
lls (
12
C
-
lysine) were untreated. For the paclitaxel experiments, MCF
-
10A light labeled cells were treated with 15
nM paclitaxel (LC Laboratories, Woburn, MA; #P
-
9600) for 18 h and the heavy labeled cells were

14

untreated. For the MDA
-
MB
-
231 line, heavy label
ed cells were treated with 120 nM paclitaxel for 18 h
and the light labeled cells were untreated. The respective heavy and light lysates were mixed at a 1:1 ratio
following cell lysis and protein quantitation. Reciprocal experiments were also performed (da
ta not
shown).


4.
Protein Identification and Quantification by LC
-
MS/MS

Protein digestion and identification by LC
-
MS/MS was performed as described previously
19
. Acquired

data was automatically processed by the Computational Proteomics Analysis S
ystem (CPAS)
20
. The

tandem mass spectra were searched with X!Tandem against version 3.13 of the human IPI database. A
fixed modification of 6.020129 mass units wa
s added to lysine residues for database searching to account
for incorporation of the heavy lysine isotope. All identifications with a PeptideProphet probability greater
than 0.9 were submitted to ProteinProphet and the subsequent protein identifications w
ere filtered at a 1%
error rate
21
. Quantitative information was extracted and only peptides with a minimum of 0.90

PeptideProphet score were considered in the analysis. All normalized peptide ratios for a specific protein
were averaged
to compute an overall protein ratio.


5.
Cytoskeleton S
taining

Cells were cultured at 50,000 cells/cm
2

in Lab
-
Tek Chamber Slides (Sigma
-
Aldrich, St. Louis, MO) in
standard media conditions with or without ROCK Inhibitor (40µM Y
-
27632) for 2 h (EMD Biosciences,
Gibbstown, NJ; #688000). Cells were then fixed in 4% (w/v) paraformaldehyde and permeabilized usi
ng
0.1% (w/w) Triton
-
X. FITC
-
conjugated phalloidin (Invitrogen, Carlsbad, CA; #A12379) was used to
stain F
-
actin and cell nuclei were counterstained using a DAPI mounting medium (Vector Laboratories,
Burlingame, CA; #H
-
1500). Fluorescent images were acquir
ed using Perkin Elmer spinning disk confocal
microscope with a 63x/1.4 oil immersion lens.



B.

Computational Analysis

Starting with the microarray data, we performed network inference using an Inferelator
-
based inference
pipeline
22

, and visualized the r
esults using Cytoscape
23
, a program that enables

network visualization
and analysis. We also analyzed the microarray and SILAC data, yielding sets of differentially expressed
genes that we further analyzed with Sungear
24
. Sungear enables comparison and investigation of large
data sets.
An example is shown in Suppl. Fig. 4.
Sungear is connected to The Gagg
le Boss
1
, whic
h allows
seamless communication and data exchange with an ensemble of other analysis tools including
Cyotscape, FireGoose
25
, the Data Matrix Viewer, and the MultiExperiment Viewer
26
. This integrated

15

analysis platform allowed for the iterative selection

of sub
-
networks and genes in Cytoscape, followed by
analysis and visual overlays in Sungear, resulting in our final networks and sub
-
networks.


C.

Genomic Analysis

Microarray data were taken from publicly available data sets in th
e Gene Expression Omnibus (G
EO)

27
.
MCF
-
10A data was taken from several different experiment sets: GEO accession numbers GSE4917
28
,
GSE6784
29
, GSE8240
30
, GSE10070
31
, GSE12764
32
, GSE14987
33
, and
GSE20285
34
. MDA
-
MB
-
231
data were all from one experiment set: GSE2603
35
. These eight data sets were generated
on four different
Affymetrix mi
croarray platforms: GSE2603 and GSE4917 on Human Genome (HG) U133A arrays,
GSE6784 on HGU133A 2.0 arrays, GSE8240 on HT HGU133A arrays, and the remaining four on
HGU133 Plus 2.0 arrays. We used MAS5
36

(without quantile normalization) to normalize data in order to
account for the lab
-
to
-
lab and platform
-
to
-
platform variation. The original full dataset contained 12
,
000
genes, from these we selected those who
se standard deviation across experiments was in the top

25
th

percentile along with the 2
,
000 genes that were most differentially expression as determined by
Significance Analysis of Microarrays (SAM). This resulted in a final network of 4
,
619 genes, includ
ing
289 transcription factors (TFs).

1.

Network inference

We used our network inference pipeline to infer a regulatory network from microarray data characterizing
the signaling networks of MD
A
-
MB
-
231 and MCF
-
10A cells (non
PS
-
OC lines). Final networks edges
(
depicted as an arrow from regulatory TF to gene)
we
re determined by the voting of multiple methods, to
produce networks of higher quality. The inference pipeline we use
d

here is composed of two methods:
1)
time
-
lagged Context Likelihood of relatedness (tlC
LR)
37

and 2) a modified version of the

Inferelator
38
.
We combine the results of these multiple methods using a resampling approach. Time
-
lagged Context
Likelihood of Relatedness (tlCLR)

is based on Context Likelihood of Relatedness
(CLR)
39

and explicitly
uses the time
-
series data and mutual information (MI)
40

(a measure of

similarity) to assign confidences to
regulatory interactions. The Inferelator learns regulatory dynamics as well as topology by explicitly using
the time
-
series data to parameterize a linear o
rdinary differential
equation model using an

l
1

constrained
linear regression and model selection method
41
. The modified

version of the Inferelator used here uses an
ordinary differential equation model, as in
15
, but uses an
adaptive
l
1

constraint
41
, which allows the
inference procedure to incorporate previously

known

regulatory edges, derived here from publicly
available interaction databases.

2.

Problem set up


16

We denote the expression levels of the genes by
x

(
x
1
,
.
.
.
,
x
N
g
)
. We store the
C

observations of these
N
g

genes in an

N
g
x
C

matrix, where the columns correspond to the experimental observations. These
observations can be of two types: time
-
series data (
X
ts
), and steady state data (
X
ss
). Since we make
explicit use of the time series data in the d
escription of our inference procedure we denote by
t
1
,
t
2
,
.
.
.
,
t
K

t
he
K

time series observations (columns of
X
ts
). Our inference procedure produces a network in the form
of a ranked list of regulatory interactions, ranked according to

confidence. We refer to the final list of
confidences as an

N
g
x
N
p

matrix
Z
final
, where
N
p

denotes the possible predictors. Element
i
,
j

of
Z
final

represents our confidence in the existence of a regulatory interaction between
x
i

and
x
j
.

3.

Core Method 1: Time Lagged Context Likelihood of Relatedness

tlCLR is an MI based method that extends the original CLR algorithm to take
advantage of time
-
series
data

16
. The original formulation of CLR was unable to learn directionality of regulatory edges as MI is a
sym
metric measure. In the tlCLR algorithm we make explicit use of the time
-
series data to learn directed
regulatory edges. We describe, in brief, three main steps: 1) model the temporal changes in expression as
an ODE, 2) calculate the MI between every pair o
f genes, 3) apply a background correction (filtering)
step to remove least likely interactions. We refer the reader to
14

for a thorough description of this method.


We assume that the temporal changes in expression
of each gene,
x
i
, can be approximated by the linear
ODE:


d
x
i
(
t
)
d
t



i
x
i


i
,
j
x
j
(
t
)
j

1
N

,


i

1
,
.
.
.
,
N



(1)


where



i

0

is the first
-
order degradation rate of
x
i

and the
b
i
,
j
’s are a set of dynamical parameters to
be estimated. The value of
b
i
,
j

describes the extent and sign of the regulation of target gene
x
i

by
regulator
x
j
. We store the dynamical parameters in an
N

×
P

matrix,

, where
N

is the number of genes,
and
P

is the number of possible regulators. Note that


is typically sparse
(
i.e.
, most entries are 0
,

reflecting the sparsity of transcriptional regulatory networks). Later, we describe how to calculate the
values
b
i
,
j

by a modified version of Inferelator. Now we briefly describe how to use the time
-
series data
in the context of improving the calculation of MI values between a gene
x
i

and its potential regulator
x
j
.



17

We first apply a finite approximation to Eq. 1, for each
x
i
,
i
=1
,

N
g

and rewrite it as a response vector
y
i
, which captures the rate of change of expression in
x
i
. We pair the response
y
i

with a corresponding
explanatory variable
x
j
j
=1

N
p
. Note each
x
j

is time lagged with respect to the response
y
i
, i.e.
x
j
(
t
k
)

is
used to predict
y
i
(
t
k
+1
)
.

For more details of this transformation we refer the reader to

14
. As a measure of
confidence for a directed regulatory interaction between a pair of genes we use,
I
(
y
i
,
x
j
)
, where a pair that
shows a high MI score (relative to other pairs) is more likely to represent a true regulatory interaction.
Note that
I
(
y
i
,
x
j
)

I
(
y
j
,
x
i
)
, making one regulatory direction more likely than the other. We refer to the
MI calculated from
I
(
y
i
,
x
j
)

as dynamic
-
MI, as it takes advantage of the temporal information available
from time
-
series data (distinguishing time
-
series data from steady
-
state data). As described above, we
calculate
I
(
x
i
,
x
j
)

and
I
(
y
i
,
x
j
)

for every pair of genes and store the values in the form of two
N
g
x
N
p

matrices
M
stat

and
M
dyn
, respectively. Note that
M
stat

is symmetric while
M
dyn

is not. We now briefly
describe how tlCLR integrates both static
-

and dynamic
-
MI to produce a final confidence score for each
regulatory interaction. For a more detailed explanation we refer the reader to
14
.


For each regulatory interaction we compute two positive Z
-
scores (by setting all negative Z
-
scores to
zero): one for the regulation of
x
i

by
x
j

based on dynamic
-
MI (
i.e.

based on
M
dyn
),
z
1
(
x
i
,
x
j
)
=max








0
,

M
dyn
i
,
j
-



j
'

M
dyn
i
,
j
'
N
s
i

where
s
i

is the standard deviation of the entries in the
i
’th row
of
M
dyn
. And one for the regulation of
x
i

by
x
j

based on both static and dynamic
-
MI,
z
2
(
x
i
,
x
j
)
=max








0
,

M
dyn
i
,
j
-



i
'

M
stat
i
'
,
j
N
s
j
,

where
s
j

is the standard deviation of the entries in the
j
’th
column of
M
stat
. We combine the two scores into a final tlCLR score,
z
tlCLR
i
,
j
=
(
z
2
1
+
z
2
2
)
. Note that
some entries in
Z
tlCLR

are zero,
i.e.
Z
tlCLR

is somewhat sparse. For a more detailed description of tlCLR
we refer the reader to

14
.


4.

Core method 2: Inferelator


18

We use a modified version Inferelator to learn a sparse dynamical model of regulation for each gene
x
i
.
As potential regulators of
x
i

we consider only the
P

highest confidence (non
-
zero) regulators.
Accordingly, for each gene,
x
i
, we denote this subset of potential regulators as
x
i
. We then learn a sparse
dynamical model of regulation for each
x
i

as a function of
x
i
’s (using Inferelator). We assume that the
time evolution in

the
x
i
’s is governed by
d
x
i
(
t
)
d
t



i
x
i


i
,
j
x
j
(
t
)
j

1
N

,


i

1
,
.
.
.
,
N



which is exactly (eq.
1)

with our constraint on the number of regulators. Adaptive Least Angle Regression
20


is used to
efficiently implement an
l
1

constraint on

, which minimizes the following objective function, amounting
to a least
-
square estimate based on the ODE (eq. 1): under an adaptive
l
1
-
norm penalty on regression
coefficients,


|
j

1
P
i


i
,
j

i
,
j
|



j

1
P
i


i
,
j
(
o
l
s
)

(2)


where,

i
,
j

corre
sponds to a weight on the predictor
x
i
,
j

chosen in a data
-
dependent manner as suggested
in

20
.

Known interactions are encoded by setting



i
,
j

1

when it is previously known that
x
j

regulates
x
i
,
thus making it less likely that

i
,
j
is shrunk out of the model. Note that it is unlikely for any weight of


to rescue

i
,
j

if it is not at all supported by the idea (i.e. the
l
1

constraint takes precedence over the
adaptive weight

i
,
j
). After applying the adaptive elastic net, we have


an

N
g
x
N
p

matrix of dynamic
parameters for each regulatory interaction. We use the percentage of explained variance of each
parameter
b
i
,
j

as described in
[
22
]
, as confidence in these regulatory interactions. We store these
confidences in
Z
Inf
. We combine these confidences in a rank
-
based way, such that each method is
weighted equally, as described
in
14
, to generate
Z
tlCLR
-
Inf
, which represents the our confidence in each
regulatory interaction after

running our pipeline one time. We now describe how we resample our
network inference pipeline to generate an ensemble of predic
ted networks (i.e. lists of confidence for each
regulatory interaction).


5.

Combining genetic and dynamic information in a resampling approach

To further improve the quality of our ranked list we applied a resampling approach to the pipeline
described above
to generate an ensemble of putative regulatory networks. We denote the matrix of

19

response variables
y
i
,
i
=1

N
g
, by
Y
. Similarly we denote the matrix of predictor variables
x
j
,
j
=1

N
p

by
X
. We sample with replacement from the indices of the columns of
Y
, generating a permutation of the
indices,
c
s
*
. We use this permutation,
c
*
, to permute the columns of
Y

and
X
, generating
Y
*
, and
X
*
,
respectively. Note that the columns of
Y

match the columns of
X

in the sense that the time
-
lagged
relationship between the response on the predictors is preserved. We generated
Z
tlCLR
, and
Z
Inf
, and
Z
tlCLR
-
Inf

as d
escribed before, with the only difference being that we use the response and explanatory
vectors from
Y
*

and
X
*
, respectively, instead of
Y

and
X
. We repeat this procedure
B

times, with
B
=50.
This generated an ensemble of
B

predicted regulatory networks. The final weight we assign to each
regulatory interaction is the median weight assigned to that interaction from each of the
B

networks. Thus,
the final weight can be considered an "ensemble vote" of the confidence the ensem
ble of networks has in
that edge
z
final
i
,
j
=
median
(
z
tlCLR
-
Inf
i
,
j
(
1
)
,
z
tlCLR
-
Inf
i
,
j
(
2
)
,

,
z
tlCLR
-
Inf
i
,
j
(
B
)
)
.


6.

Visualization and analysis

We use Cytoscape

1
as our primary means of visualizing regulatory networks and overlaying asso
ciated
data (in this case, genes in proteomics and microarray experiments)

since it
display
s

attributes (
e.g.
,
coloring based on gene expression)
and
convey
s

multiple channels of information in a single interactive
network representation.


The Inferelator
output is a ranked list of regulatory interactions of the form "A regulates B". Each of these
interactions is assigned a confidence by the Inferelator, and the predicted interactions are sorted in order
of decreasing confidence. The usual process for visua
lizing networks generated by the Inferelator is to
convert the Inferelator output file to a Cytoscape .sif file, assign confidence and kinetic values to the
edges using edge attribute files, filter out low
-
confidence edges from the network, and display the

filtered
network using edge coloring that reflects the kinetic parameters.

However, here we wish to create a single
network visualization that will allow direct comparison of our two individual cell
-
line networks. We
approach this problem by defining two
new confidence metrics for each edge in the final network: the
overall confidence that a regulatory interaction exists in either network and the likelihood that the edge
exists on one network but not the other. Since the confidence values generated by the
Inferelator are
essentially Z
-
scores, we use Stouffer’s method to combine the scores for one edge in two networks into an
overall confidence metric for that edge. Likelihood for each edge uses the rank of that edge in the ordered
list of edges for each ind
ividual network, and is calculated as the log ratio of the two ranks. Likelihoods
of edges that exist only in one network or the other are set to the

largest or smallest defined values across

20

all likelihood scores calculated as above. Undefined values are
removed implicitly in the filtering stage,
which removes all edges with a combined confidence of zero. After defining overall confidence and
likelihood scores, processing for visualization proceeds as usual from the step of generating a .sif file,
using th
e overall confidence score instead of the single
-
network confidence scores. Edge coloring for
network comparisons usually reflects the edge likelihood score instead of the kinetic parameter.


Sungear

24

is a visualization tool that enables rapid exploration of large data sets. Here the data are sets of
genes
(
e.g.
, genes that are differentially expressed in microarray
experiments on one cell line as compared
to the other
)
. The main Sungear window is a generalization of a Venn diagram to arbitrarily many sets.
By displaying all possible discrete intersections of input sets, the features that are unique to one set or
shar
ed among combinations of sets are readily apparent. We prepared lists of genes from both microarray
and proteomics data for analysis using Sungear. We used Significance Analysis of
Microarrays (SAM)
42


to compare the set of 103 MCF
-
10A microarray experiments to the 121 MDA
-
MB
-
231 microarray
experiments, selecting genes that were most differentially expressed in e
ach cell line according to the
SAM t
-
statistic. This yielded one list of differentially expressed genes for each cell line. Control studies
have shown that changes greater than 1.5 are typically statistically significant. Consequently, we then
analyzed the

SILAC
proteomics
data for differential expression based on L/H ratio, looking for change in
either direction (
e.g.
, greater than 3/2 or less than 2/3), producing one further list for each cell line for
each SILAC experiment. Finally, we created a "present
" set for each pair of SILAC experiments where
the criteria for list inclusion was simply identification of the protein as present in either experiment, with
or without a usable L/H ratio between experiments. These lists were then made available for explor
ation
within Sungear.


The Gaggle is a framework that enables communication between different analysis tools via a lightweight
server called the Gaggle Boss. Programs can be connected to the Gaggle via a simple "goose" API,
effectively allowing these progr
ams to communicate directly with each other. We developed a new
Gaggled version of Sungear that worked with the Gaggled Cytoscape to enable both programs to send
lists of genes to the other, enabling rapid, iterative exploration of networks and genomic/pro
teomic data.
For example, Sungear can quickly isolate the sets of genes that are present or differentially expressed in
the cell line comparison SILAC experiments. These genes can then be broadcast to Cytoscape, assigned
attributes, and quickly highlighted

by any of Cytoscape’s visual attributes (
e.g.
, different colors on the
network visualization). Differentially
-
expressed genes that appear in key places in the network (
e.g.
,
transcription factors hypothesized to regulate many other genes or TFs) can th
en

be selected and sent back

21

to Sungear, where they can be examined for membership in other sets (
e.g.
, differential regulation under
taxol treatment)

(see Supplement Fig. 4)
.





22

References



1

Kak, A. C., Author, Slaney, M., Wang, G. &

Reviewer. Principles of Computerized
Tomographic Imaging.
Medical Physics

29
, 107
-
107 (2002).

2

Nandakumar, V., Kelbauskas, L., Johnson, R. & Meldrum, D. Quantitative
characterization of preneoplastic progression using single
-
cell computed tomography and
three
-
dimensional karyometry.
Cytometry. Part A : the journal of the International
Society for Analytical Cytology

79
, 25
-
34 (2011).

3

Damania, D.

et al.

Role of cytoskeleton in controlling the disorder strength of cellular
nanoscale architecture.
Biophysi
cal journal

99
, 989
-
996 (2010).

4

Debnath, J., Muthuswamy, S. K. & Brugge, J. S. Morphogenesis and oncogenesis of
MCF
-
10A mammary epithelial acini grown in three
-
dimensional basement membrane
cultures.
Methods

30
, 256
-
268 (2003).

5

Johnson, K. R., Leight,
J. L. & Weaver, V. M. Demystifying the effects of a three
-
dimensional microenvironment in tissue morphogenesis.
Methods in cell biology

83
, 547
-
583 (2007).

6

Weaver, V. M.

et al.

beta4 integrin
-
dependent formation of polarized three
-
dimensional
architectur
e confers resistance to apoptosis in normal and malignant mammary
epithelium.
Cancer cell

2
, 205
-
216 (2002).

7

Fraley, S. I.

et al.

A distinctive role for focal adhesion proteins in three
-
dimensional cell
motility.
Nature cell biology

12
, 598
-
604 (2010).

8

Kraning
-
Rush, C. M., Carey, S. P., Califano, J. P., Smith, B. N. & Reinhart
-
King, C. A.
The role of the cytoskeleton in cellular force generation in 2D and 3D environments.
Physical biology

8
, 015009 (2011).

9

Beacham, D. A., Amatangelo, M. D. & Cukierman
, E. in
Current Protocols in Cell
Biology

(John Wiley & Sons, Inc., 2001).

10

Dickinson, L. E., Ho, C. C., Wang, G. M., Stebe, K. J. & Gerecht, S. Functional surfaces
for high
-
resolution analysis of cancer cell interactions on exogenous hyaluronic acid
.
Biomaterials

31
, 5472
-
5478 (2010).

11

Fuhrmann, A.

et al.

AFM stiffness nanotomography of normal, metaplastic and
dysplastic human esophageal cells.
Physical biology

8
, 015007 (2011).

12

Schulz, O.

et al.

Simultaneous single molecule atomic force and fluorescence lifetime
imaging.
Proc Spie

7571

(2010).

13

Hutter, J. L. & Bechhoefer, J. Calibration of atomic
-
force microscope tips.
Review of
Scientific Instruments

64
, 1868
-
1873 (1993).

14

Wu, P. H., Arce, S
. H., Burney, P. R. & Tseng, Y. A novel approach to high accuracy of
video
-
based microrheology.
Biophysical journal

96
, 5103
-
5111 (2009).

15

Verbridge, S. S.

et al.

Oxygen
-
controlled three
-
dimensional cultures to analyze tumor
angiogenesis.
Tissue engineer
ing. Part A

16
, 2133
-
2141 (2010).

16

Hanley, W. D.

et al.

Variant isoforms of CD44 are P
-

and L
-
selectin ligands on colon
carcinoma cells.
FASEB journal : official publication of the Federation of American
Societies for Experimental Biology

20
, 337
-
339 (20
06).

17

Kelbauskas, L.

et al.

Method for physiologic phenotype characterization at the single
-
cell
level in non
-
interacting and interacting cells.
J Biomed Opt

17
, 037008 (2012).


23

18

Ong, S. E.

et al.

Stable isotope labeling by amino acids in cell culture,
SILAC, as a
simple and accurate approach to expression proteomics.
Molecular & cellular proteomics
: MCP

1
, 376
-
386 (2002).

19

Faca, V.

et al.

Quantitative analysis of acrylamide labeled serum proteins by LC
-
MS/MS.
Journal of proteome research

5
, 2009
-
2018

(2006).

20

Rauch, A.

et al.

Computational Proteomics Analysis System (CPAS): an extensible,
open
-
source analytic system for evaluating and publishing proteomic data and high
throughput biological experiments.
Journal of proteome research

5
, 112
-
121 (2006)
.

21

Keller, A., Nesvizhskii, A. I., Kolker, E. & Aebersold, R. Empirical statistical model to
estimate the accuracy of peptide identifications made by MS/MS and database search.
Analytical chemistry

74
, 5383
-
5392 (2002).

22

Greenfield, A., Madar, A., Ostr
er, H. & Bonneau, R. DREAM4: Combining genetic and
dynamic information to identify biological networks and dynamical models.
PloS one

5
,
e13397 (2010).

23

Shannon, P.

et al.

Cytoscape: a software environment for integrated models of
biomolecular interaction networks.
Genome research

13
, 2498
-
2504 (2003).

24

Poultney, C. S.

et al.

Sungear: interactive visualization and functional analysis of
genomic datasets.
Bioinformatics

23
, 259
-
261 (2007).

25

Bare, J. C., Shannon, P. T., Schmid, A. K. & Baliga, N. S. The Firegoose: two
-
way
integration of diverse data from different bioinformatics web resources with desktop
applications.
BMC bioinformatics

8
, 456 (2007).

26

Saeed, A. I.

et a
l.

TM4: a free, open
-
source system for microarray data management and
analysis.
BioTechniques

34
, 374
-
378 (2003).

27

Edgar, R., Domrachev, M. & Lash, A. E. Gene Expression Omnibus: NCBI gene
expression and hybridization array data repository.
Nucleic Acids

Res

30
, 207
-
210
(2002).

28

Wu, W., Zou, M., Brickley, D. R., Pew, T. & Conzen, S. D. Glucocorticoid receptor
activation signals through forkhead transcription factor 3a in breast cancer cells.
Mol
Endocrinol

20
, 2304
-
2314 (2006).

29

Amit, I.

et al.

A modu
le of negative feedback regulators defines growth factor signaling.
Nature genetics

39
, 503
-
512 (2007).

30

Andarawewa, K. L.

et al.

Ionizing radiation predisposes nonmalignant human mammary
epithelial cells to undergo transforming growth factor beta induce
d epithelial to
mesenchymal transition.
Cancer research

67
, 8662
-
8670 (2007).

31

Marshall, A. M., Pai, V. P., Sartor, M. A. & Horseman, N. D. In vitro multipotent
differentiation and barrier function of a human mammary epithelium.
Cell and tissue
research

335
, 383
-
395 (2009).

32

Hoeflich, K. P.

et al.

In vivo antitumor activity of MEK and phosphatidylinositol 3
-
kinase inhibitors in basal
-
like breast cancer models.
Clinical cancer research : an official
journal of the American Association for Cancer Research

15
, 4649
-
4664 (2009).

33

Wolfer, A.

et al.

MYC regulation of a "poor
-
prognosis" metastatic cancer cell state.
Proceedings of the National Academy of Sciences of the United States of America

107
,
3698
-
3
703

(2010).

34

Mazzone, M.

et al.

Dose
-
dependent induc
tion of distinct phenotypic responses to Notch
pathway activation in mammary epithelial cells.
Proceedings of the National Academy of
Sciences of the United States of America

107
, 5012
-
5017 (2010).


24

35

Minn, A. J.

et al.

Genes that mediate breast cancer metastasis to lung.
Nature

436
, 518
-
524 (2005).

36

Hubbell, E., Liu, W. M. & Mei, R. Robust estimators for expression analysis.
Bioinformatics

18
, 1585
-
1592 (2002).

37

Madar, A., Greenfield, A., Vanden
-
Eijnden, E. & Bonnea
u, R. DREAM3: network
inference using dynamic context likelihood of relatedness and the inferelator.
PloS one

5
,
e9803 (2010).

38

Bonneau, R.

et al.

The Inferelator: an algorithm for learning parsimonious regulatory
networks from systems
-
biology data sets
de novo.
Genome biology

7
, R36

(2006).

39

Faith, J. J.

et al.

Large
-
scale mapping and validation of Escherichia coli transcriptional
regulation from a compendium of expression profiles.
PLoS biology

5
, e8 (2007).

40

Shannon, C. E. A Mathematical Theory of
Communication.
At&T Tech J

27
, 379
-
423
(1948).

41

Zou, R. & Ghosh, A. Automated sensitivity analysis of stiff biochemical systems using a
fourth
-
order adaptive step size Rosenbrock integration method.
Systems biology

153
, 79
-
90 (2006).

42

Tusher, V. G., Ti
bshirani, R. & Chu, G. Significance analysis of microarrays applied to
the ionizing radiation response.
Proceedings of the National Academy of Sciences of the
United States of America

98
, 5116
-
5121

(2001).




25

SUPPLEMENTARY FIGURES

Supplementary Figure 1
.





26

Supplementary Figure 2.



27


Supplementary Figure 3.




28

Supplementary Figure 4.



29

Supplementary Figure 5.



30

SUPPLEMENTAL FIGURE LEGENDS

Supplemental Figure 1:
I
n
-
depth characterization of MCF
-
10A and MDA
-
MB
-
231 cell

line
s


(A)
Distribution of cell
lines to PSOC Network
. (B)
10× Phase micrographs

of cells in culture
. Scale bars:
100
µm
. (C)
Epifluorescence micrographs of cells stained with Hoechst 33258 (nuclear DNA) and Alexa
Fluor 488 phalloidin (cytoskeletal actin).
Top: MCF
-
10A. Bottom:
MDA
-
MB
-
231. Left: Untreated.
Right: 2 h after treatment with 40 µM Y
-
27632

ROCK inhibitor
. Scale bars:
20
µm.
(D) MCF
-
10A cells
at various time points after initial seeding and at two magnifications. Cell density shown at the 72 h time
point is optimal for

passaging the cells. Cells shown at 120 h are overgrown. (E) MDA
-
MB
-
231 cells at
various time points after initial seeding and at two magnifications. Cell density shown at the 96 h time
point is optimal for passaging the cells. Cells shown at 144 h are ov
ergrown. (F) MCF
-
10A cell cycle
analysis profile. (G) MDA
-
MB
-
231 cell cycle analysis profile. (H) MCF
-
10A karyotype. (I) MDA
-
MB
-
231 karyotype.


Supplemental Figure 2:

Stress Response

(A)

Fluorescence micrographs from viability assay of cells grown in 3D cu
lture (alginate discs)
.
(B)
Oxygen drawdown over time of MCF
-
10A and MDA
-
MB
-
231 cells grown in normoxic and hypoxic
conditions.
(C)

Raw fluorescence histograms of CEA expression of CEA in MCF
-
10A and MDA
-
MB
-
231 cells in normoxic and hypoxic conditions.
(D)

Mean fluorescence intensity minus isotype (MFI) of
CD44 in MCF
-
10A and MDA
-
MB
-
231 cells in normoxic and hypoxic conditions.
(E)

Schematic of 3D
alginate gel fabrication and experiment.


Supplemental Figure
3
:
Overall i
ntegrated regulatory network

The over
all network with the top 5000 edges ranked by combined confidence from the two cell line
inference runs. As in Figures 1A
-
1D, edge color denotes differential inferred regulation in MCF
-
10A
(yellow) or MDA
-
MB
-
231 (blue). Nodes are rendered semi
-
transparent
so that the distribution of cell
-
line
-
specific regulatory edges

(connecting lines)

can be clearly seen. The inset histogram also shows the
distribution of cell
-
line
-
specific edges: edges specific to MDA
-
MB
-
231 are more prevalent, particularly at
the extreme. Proteomics data from MCF
-
10A/MDA
-
MB
-
231 comparison are also shown using node
co
lors: differential expression in MCF
-
10A is shown in yellow, and MDA
-
MB
-
231 in blue. Genes
present but not differentially expressed are shown in darker gray.


Supplemental Figure
4
:
Sungear Diagram

Sungear diagram
showing
differential expression
of ROCK an
d Taxol
-
treated MDA
-
MB
-
231 and MCF
-
10A cells.
Shown is the main Sungear plot, which
displays

data set labels around the outside of the plot
and circles representing intersections between these sets in the interior. The data here are eight sets of
proteins
up
-

or down
-
regulated in different conditions: from the top, these are proteins up
-
regulated in
MCF
-
10A with ROCK treatment (Rock_10A+), down
-
regulated in MCF
-
10A/ROCK (Rock_10A
-
), up
-

and down
-
regulated in MDA
-
MB
-
231 with ROCK treatment (Rock_231+, Rock_2
31
-
), and up
-

and
down
-
regulated proteins in both cell lines with taxol treatment (Taxol_10A+, Taxol_10A
-
, Taxol_231+,
Taxol_231
-
). Circle size represents the number of proteins in a given intersection, and location is based
on the sets in the intersection
: the large circle closest to the Rock_10A
-

label represents the 57 proteins
that are unique to that set, while the highlighted
red
circle between Rock_10A
-

and Taxol_10A+
represents the six proteins common to both sets. Using the Gaggle, proteins up
-

and
down
-
regulated in
different treatments were sent to Cytoscape
.

We then used the Gaggle again to send the list of
differentially expressed proteins in

the ITGB4 1
-
hop network (Fig. 1d
) back to Sungear: of these six
genes, three were in the intersection high
lighted in pink.


31


Supplemental Figure
5
:
Wound healing assay

Wound healing assay. Top: Micrographs taken at the initial time point and at 12 hr intervals after
scratching. Bottom: Quantitation of scratched area filled over time.