Abstract
Background
Glioblastoma is the most aggressive malignant brain tumor with poor
survival due to its invasive nature driven by cell migration, with
unclear linkage to transcriptomic information. The aim of this study
was to develop a physics-based framework connecting to transcriptomics
to predict patient-specific glioblastoma cell migration.
Methods and Results
We applied a physics-based motor-clutch model, a cell migration
simulator (CMS), to parameterize the migration of glioblastoma cells
and define physical biomarkers on a patient-by-patient basis. We
reduced the 11-dimensional parameter space of the CMS into 3 principal
physical parameters that govern cell migration: motor number—describing
myosin II activity, clutch number—describing adhesion level, and
F-actin polymerization rate. Experimentally, we found that glioblastoma
patient-derived (xenograft) cell lines across mesenchymal (MES),
proneural, and classical subtypes and 2 institutions (N = 13 patients)
had optimal motility and traction force on stiffnesses around 9.3 kPa,
with otherwise heterogeneous and uncorrelated motility, traction, and
F-actin flow. By contrast, with the CMS parameterization, we found that
glioblastoma cells consistently had balanced motor/clutch ratios to
enable effective migration and that MES cells had higher actin
polymerization rates resulting in higher motility. The CMS also
predicted differential sensitivity to cytoskeletal drugs between
patients. Finally, we identified 18 genes that correlated with the
physical parameters, suggesting transcriptomic data alone could
potentially predict the mechanics and speed of glioblastoma cell
migration.
Conclusions
We describe a general physics-based framework for parameterizing
individual glioblastoma patients and connecting to clinical
transcriptomic data that can potentially be used to develop
patient-specific anti-migratory therapeutic strategies.
Keywords: cell migration, biophysical modeling, glioblastoma subtypes,
motor-clutch model, patient-derived cell lines
__________________________________________________________________
Key Points.
* Glioblastoma cells had balanced motor/clutch ratios to enable
effective migration.
* Mesenchymal cells had enhanced actin polymerization resulting in
higher motility.
* Eighteen genes were correlated with the biophysical parameters to
predict cell migration.
Importance of the Study.
Successful precision medicine requires biomarkers to define patient
states and identify personalized treatments. While biomarkers are
generally based on expression levels of protein and/or RNA, we
ultimately seek to alter fundamental cell behaviors such as cell
migration, which drives tumor invasion and metastasis. Our study
defines a new approach for using biophysics-based models to define
mechanical biomarkers that can be used to identify patient-specific
anti-migratory therapeutic strategies.
Glioblastoma is the most common malignant brain tumor with a median
survival of only 15 months and less than 5% 5-year survival
rate.^[35]1,[36]2 Complete surgical resection is difficult because the
tumor is highly invasive, and tumor cell infiltration into the
surrounding brain tissue drives disease progression and
recurrence.^[37]3,[38]4 Therefore, understanding the mechanics of
cancer cell migration can potentially be used to predict patterns of
invasion and guide efforts to disrupt migration and slow disease
progression.^[39]5 To target cell migration, cilengitide inhibits
adhesion proteins such as αvβ3 and αvβ5 integrin but failed in a phase
3 trial,^[40]6 suggesting that other adhesion proteins, such as CD44,
serve as major adhesive molecules in glioblastoma.^[41]7–9 Inhibiting
nonmuscle myosin II resulted in blocking glioma cell
migration,^[42]10,[43]11 and clinically safe derivatives of a myosin
inhibitor are under development.^[44]12 Fluvoxamine, an antidepressant,
can potently inhibit actin polymerization and block glioma cell
migration.^[45]13 However, in these studies, the connection of drug
potency to fundamental glioma cell migration mechanics as a function of
the transcriptomicly defined glioblastoma subtypes of proneural (PN),
classical (CL), and mesenchymal (MES)^[46]14–16 was still unclear. In
addition, it is not always feasible clinically to conduct in vitro
migration assays on patient cells, and different harvesting and
culturing methods may significantly alter the migration
behavior.^[47]17 Therefore, to effectively target cancer cell
migration, it is critical to understand the fundamental mechanics of
glioblastoma cell migration and its potential link to transcriptomic
information to predict tumor cell invasion based on patient-specific
omic analysis.
In the classic cell migration cycle, the first step is the extension of
a cell protrusion at the leading edge driven by actin polymerization
into self-assembled actin filaments (F-actin). F-actin undergoes
retrograde flow driven by myosin II (motor)-mediated contraction,
leading to protrusion retraction. At the same time, cell adhesion
molecule binding to the extracellular environment and subsequent
stretching of the actin-adhesion adaptor proteins constitute a
molecular “clutch” that resists myosin forces and biases the protrusion
toward net extension. The adhesion proteins can form focal adhesions
that allow the cell to transmit traction forces onto compliant
substrates. This system is known as the motor-clutch mechanism and is
widely used to describe cell migration.^[48]18–20 Stochastic
simulations of the motor-clutch model^[49]21 have been developed and
successfully predict the cell traction force, cell morphology, and
F-actin flow on various substrate conditions.^[50]21–23 Beyond single
protrusions, the cell nucleates multiple protrusions via F-actin
polymerization, each of which can be modeled as a motor-clutch system,
with traction forces balancing across the different protrusions.
Stochastic perturbations to the force balance due to adhesion bond
rupture enable larger-scale cell movements and can define the front and
the rear of the cell.^[51]24,[52]25 By imposing a force balance between
the protrusions, Bangasser et al.^[53]26 developed a whole-cell
motor-clutch model, which we refer to here as the cell migration
simulator (CMS, [54]Figure 1A). The CMS has successfully captured the
unique cell migration features on substrates with various
stiffnesses,^[55]26 various focal adhesion sizes,^[56]27 different
viscoelastic properties,^[57]28 different stiffness gradients,^[58]29
within 1D channels,^[59]30 and in brain tissue ex
vivo.^[60]5,[61]7,[62]8,[63]31 Therefore, the CMS provides a consistent
mechanical framework that can potentially be used to interpret and
synthesize cell migration and force measurements of glioblastoma
patient-derived (PD) cells across subtypes to predict cell migration.
Figure 1.
[64]Figure 1.
[65]Open in a new tab
Three cell migration simulator (CMS) physical parameters dictate cell
migration and traction force. (A) CMS schematic. (B) Parameter
sensitivities of the CMS were analyzed to predict maximum cell
migration speed (RMC), minimum F-actin flow, and maximum traction force
across substrate stiffnesses. With hierarchical clustering, 3 parameter
groups were identified: clutch group (
[MATH: nc :MATH]
,
[MATH: Fb :MATH]
,
[MATH: kon :MATH]
,
[MATH: koff :MATH]
), motor group (
[MATH: nm :MATH]
,
[MATH: Fm :MATH]
,
[MATH: vm :MATH]
) and actin group (
[MATH: vpoly :MATH]
,
[MATH: kmod :MATH]
,
[MATH: kcap :MATH]
). The 3 parameters (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) were chosen as fundamental physical expressions of the CMS. (C)
Maximum RMC, maximum traction, and minimum F-actin flow across
substrate stiffnesses as a function of the 3 biophysical expressions (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) predicted by the CMS were plotted. Conditions I–IV represent distinct
cell migratory behaviors with different (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
). (D) Condition I represents a typical migrating cell with base
parameter values. In Condition II, a higher motor number resulted in
lower traction, shorter protrusion length, and slower cell migration.
In Condition III, a higher clutch number resulted in near-maximal
constant traction, limited dynamic protrusions, and poor cell
migration. In Condition IV, a higher actin polymerization rate resulted
in more dynamic protrusions, longer protrusion length, highly
fluctuating traction, and faster cell migration.
In this study, we applied migration assays to glioblastoma PD xenograft
(PDX) and PD cells (collectively referred to here as “PD(X)”). We
explored the ability of the CMS to serve as a physics-based framework
for glioblastoma subtypes and PD(X) systems. We used the CMS parameters
representing myosin II motors, adhesion protein “clutches,” and F-actin
polymerization to predict cell migration generally, and then
mechanically parameterized glioblastoma cells obtained from a cohort of
11 glioblastoma patients across all 3 subtypes and 2 different culture
procedures. Using single-cell migration and force generation data
obtained on compliant 2D surfaces, we found distinct parameter sets for
glioblastoma patients across subtypes and culture conditions. In
addition, the CMS-predicted differential cell migration sensitivities
to cytoskeletal drugs between subtypes. Finally, we established
correlative links between the CMS parameter values and patient cell
transcriptomes. Our results suggest it is feasible to estimate cell
migration speeds using mRNA expression, similar to how migration speed
can be estimated via machine learning-based detection of features in
clinical MRI images.^[66]31 Overall, we describe a consistent
physics-based framework for parameterizing individual glioblastoma
patients, connected to clinical transcriptomic data, that can
potentially be used to develop subtype and patient-specific
anti-migratory therapeutic strategies for glioblastoma.
Materials and Methods
Cell Migration Simulator
The detailed governing equations and algorithms of the CMS were
described in Bangasser et al.^[67]26 and in [68]Supplementary Methods.
Briefly, the CMS comprises multiple protrusions or modules that were
nucleated randomly based on the rate
[MATH: kmod :MATH]
. Protrusions were elongated based on the polymerization rate
[MATH: vpoly :MATH]
. Protrusions were capped randomly at the rate
[MATH: kcap :MATH]
eliminating further polymerization and removed if the protrusion length
was shorter than the minimum length
[MATH: lmin :MATH]
. The cell position was determined by the force balance between
protrusion forces for modules and the cell body force. Clutches bound
and unbound to F-actin based on the clutch binding rate
[MATH: kon :MATH]
and unbinding rate
[MATH: koff :MATH]
([69]Table 1).
Table 1.
Parameters for the Cellular Level CMS
Parameter Symbol Value Ref.
Total number of myosin motors
[MATH: nm :MATH]
1000 Adjusted
Total number of clutches
[MATH: nc :MATH]
750 Adjusted
Maximum total actin length
[MATH: Atot :MATH]
100 µm [70]^*
Maximum actin polymerization rate
[MATH: vpoly :MATH]
200 nm/s Adjusted
Maximum module nucleation rate
[MATH: kmod :MATH]
1 s^–1 [71]^*
Module capping rate
[MATH: kcap :MATH]
0.001 s^–1 [72]^*
Initial module length
[MATH: lin :MATH]
5 µm [73]^*
Minimum module length
[MATH: lmin :MATH]
0.1 µm [74]^*
Cell spring constant
[MATH: kcell :MATH]
10 000 pN/nm [75]^*
Number of cell body clutches
[MATH: nc,cell :MATH]
10 [76]^*
Substrate spring constant
[MATH: ks :MATH]
0.3‒300 pN/nm Adjusted
Maximum number of module motors
[MATH: nm :MATH]
1000 [77]^*
Myosin motor stall force
[MATH: Fm :MATH]
2 pN [78]^*
Unloaded actin flow rate
[MATH: vm :MATH]
120 nm/s [79]^*
Maximum number of module clutches
[MATH: nc :MATH]
750 [80]^*
Clutch on-rate
[MATH: kon :MATH]
1 s^–1 [81]^*
Unloaded clutch off-rate
[MATH: koff :MATH]
0.1 s^–1 [82]^*
Clutch spring constant
[MATH: kc :MATH]
0.8 pN/nm [83]^*
Characteristic clutch rupture force
[MATH: Fb :MATH]
2 pN [84]^*
[85]Open in a new tab
^*Adebowale et al.^[86]28
Monte Carlo simulations were conducted using a direct Gillespie
Stochastic Simulation Algorithm, the event was executed based on
accumulated event rates, including
[MATH: kon :MATH]
,
[MATH: koff :MATH]
,
[MATH: kmod :MATH]
, and
[MATH: kcap :MATH]
, and the next time step
[MATH: tstep :MATH]
was determined based on the total event rates
[MATH: ∑ki :MATH]
. The C++ version of the CMS^[87]27 was used to conduct the simulations
on the Mesabi computer cluster at the Minnesota Supercomputing
Institute.
Grouped Clutch
Here we used the grouped-clutch algorithm to significantly enhance the
computational efficiency by grouping clutches together to have a
smaller number of clutches to represent all clutches, which produced
equivalent results but with much faster simulations ([88]Supplementary
Figure 2). The detailed governing equations and analysis were described
in [89]Supplementary Methods.
Parameter Sensitivity and Clustering
The cell migration predictions, including the maximum random motility
coefficient (RMC), the maximum traction force, and the minimum F-actin
flow over different stiffnesses were generated by the CMS with the
changes of the base parameter values ([90]Table 1), plotted in
[91]Supplementary Figure 1. The linear regression between the CMS
migration predictions Y and the logarithm of parameter ratios was
plotted in [92]Supplementary Figure 1. Parameter sensitivity values
were determined by the slope of the linear regression normalized by the
base prediction values (Y[0]) (slope (linear regression)/Y[0]). We
applied the agglomerative hierarchical clustering to the CMS parameter
sensitivities using the linkage function in MATLAB with an average
method to identify the main clusters for the CMS parameter
sensitivities ([93]Figure 1B).
Glioblastoma Patient Cell Lines and Cell Culture
Mayo PDX cell lines were developed and maintained by the Sarkaria lab
at Mayo Clinic (Rochester, MN).^[94]32 Cell lines were established by
implanting patient tumors into mouse flanks, and cells were derived in
short-term explant cultures with serum-containing medium. We used MES
(Mayo 16, 46, 59), PN (Mayo 64, 80, 85), and CL (Mayo 6, 38, 76, 91,
195) cells. Cells were shipped in fetal bovine serum (FBS) media
(Dulbecco's modified Eagle's medium [DMEM] + 10% serum) and grown
adherently until confluent, and then frozen in 10% DMSO 90% FBS media.
Cells prepared for experiments were thawed into a flask coated with 10%
Matrigel (Corning 354263) in Neural Stem Cell (NSC) Media (DMEM/F12
(Gibco 11320033) + 1X B-27 Supplement (Gibco 12587010) + 1X Pen/Strep
(Corning 45000-650) + 1ng/mL epidermal growth factor/fibroblast growth
factor (Peprotech AF10015/Peprotech AF10018B) (added every 2–3 days)).
Cells were allowed to recover for several days prior to imaging.
UCSD PD lines^[95]33 were developed and maintained by Clark Chen’s
Laboratory at the University of Minnesota (formerly at the University
of California San Diego). Cell lines were derived and established from
MES and PN glioblastoma patients and cultured as neurospheres.^[96]33
UCSD cells prepared for experiments were propagated in ultra-low
adhesion flasks (Corning 3814) with NSC Media and were allowed to
recover thawing for several days prior to imaging. For adherent culture
conditions, UCSD cells were grown on a Matrigel-coated T-flask in an
NSC medium for a minimum of 1 week prior to imaging.
Cell Migration, Traction Force, and F-Actin Flow on PA Gels
Polyacrylamide (PA) gels with different stiffnesses (0.7, 4.6, 9.3, and
19.5 kPa) were synthesized following the previous protocols^[97]26 and
described in [98]Supplementary Methods. Before imaging, Mayo cells were
plated on laminin with media containing 2% serum to promote adhesion.
UCSD cells could not adhere to laminin, collagen, or fibronectin,
except for Matrigel with no serum. Time-lapse light microscopic images
were taken using established protocols^[99]26 and described in
[100]Supplementary Methods. Time-lapse phase-contrast images were taken
for 10 h to track cell migration using a Nikon Eclipse TE200 microscope
with a Plan Fluor 10×/0.30NA objective. Time-lapse phase-contrast
images were taken for 3 min to measure F-actin flow. Phase-contrast and
epifluorescence images were taken before and after glioblastoma cells
detached by treatment with 0.05% trypsin to determine the cell
traction.
Cell motility (RMCs), cell area, aspect ratio, actin retrograde flow,
and traction strain energy were determined based on established
protocols^[101]26 and described in [102]Supplementary Methods. Cell
motility, area, and aspect ratio were measured for Mayo MES (Mayo
16(patient number), 46, 59), PN (Mayo 64, 80, 85), and CL(Mayo 6, 38,
76, 91, 195) cells and UCSD MES, PN cells. Actin flow was measured for
Mayo MES (Mayo 16, 46, 59), PN (Mayo 64, 80, 85), and CL(Mayo 6, 38,
76, 91, 195) cells and UCSD MES, PN cells. Strain energy was measured
for Mayo MES (Mayo 16, 46, 59), PN (Mayo 64, 80, 85), and CL (Mayo 6,
38, 76, 91, 195) cells and UCSD MES, PN cells.
Parameterizing Glioblastoma Cell Lines With (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) Values
CMS predictions (
[MATH: vF~-
actin :MATH]
(min),
[MATH: Fmodule :MATH]
(max), Rmc (max)) were linearly interpolated based on a 3-dimensional
parameter space defined by
[MATH: nm :MATH]
,
[MATH: nc :MATH]
, and
[MATH: vpoly :MATH]
. Cell traction force was estimated using the linear relation between
module force and experimental strain energy (
[MATH: Fmodule :MATH]
= 175*
[MATH: Estrain :MATH]
) based on the U251 maximum strain energy value and the assumption that
U251 cells had 7500 clutches in the model.^[103]26 The unique (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) values were found for each patient to limit the relative errors
between CMS and experimental results to within 10% (eg, (
[MATH: vF~-
actin :MATH]
(min, model) −
[MATH: vF~-
actin :MATH]
(min, patient))/
[MATH: vF~-
actin :MATH]
(min, patient) < 10%).
mRNA Expression Analysis
Detailed mRNA expression analysis was described in [104]Supplementary
Methods. In short, we derived RNAseq reads per kilobase million (RPKM)
expressions of Mayo PDX cells from Vaubel et al.^[105]32 (19 552 genes,
[106]www.cbioportal.org). We filtered out genes with geometric means
smaller than 1 and with counts in less than 80% of patients to achieve
a normal gene expression distribution compared with the original
distribution ([107]Supplementary Figure 6A and [108]B, 11 752 genes
left). We applied the 2-sample t-test to the RNAseq-derived mRNA
expression levels of Mayo MES and PN lines and derived 1 177
differential genes with P < .05 and FC > 2 in the volcano plot
([109]Supplementary Figure 6C). We applied pathway enrichment analysis
to these differential genes based on the KEGG database^[110]34 with the
false discovery rate (FDR)-adjusted P < .05 and derived 29 enriched
pathways ([111]Supplementary Figure 7). We derived the actin-motor gene
list (34 genes) based on the “Regulation of actin cytoskeleton”
pathway, and the clutch gene list (42 genes) based on the “Focal
adhesion” and “ECM-receptor interaction” pathways. We applied linear
correlation analysis between the mRNA expression ratios of the
actin-motor ([112]Figure 5A and [113]B) and clutch ([114]Figure 5C)
genes in the 10 Mayo lines used in the present study and their CMS
parameter values (
[MATH: vpoly :MATH]
,
[MATH: nm :MATH]
,
[MATH: nc :MATH]
) ([115]Figure 5A–C), respectively. We derived RNAseq normalized
transcripts per kilobase million (TPM) values by RSEM algorithm from
The Cancer Genome Atlas (TCGA) Glioblastoma Project
([116]www.cbioportal.org, 151 patients, 20 531 genes). We applied Cox
regression analysis between the mRNA expression ratios of the
actin-motor ([117]Supplementary Figure 10A) and clutch
([118]Supplementary Figure 10B) genes in a cohort of 66 Mayo patients
and their overall survival, and their hazard ratios with 95% confidence
interval were sorted and plotted in[119] Supplementary Figure 10, with
the significant hazard ratios in red.
Figure 5.
[120]Figure 5.
[121]Open in a new tab
CMS-based transcriptomic biomarkers lead to targeting strategies for
glioblastoma cell migration. We applied the pathway enrichment analysis
to the 1177 differential genes between Mayo MES and PN cell lines
(method: mRNA expression analysis) using the KEGG database^[122]34 with
the false discovery rate (FDR)-adjusted P < .05, and derived 29
enriched pathways ([123]Supplementary Figure 7. We derived the
actin-motor gene list (34 genes) based on the “Regulation of actin
cytoskeleton” pathway, and the clutch gene list (42 genes) based on the
“Focal adhesion” and “ECM-receptor interaction” pathways. Correlation
analysis was applied to the mRNA expression ratios of the actin-motor
(A, B) and clutch (C) genes in the 10 Mayo lines used in the present
study (no mRNA data was available for the Mayo line 16) and their CMS
parameter values,
[MATH: vpoly :MATH]
(A),
[MATH: nm :MATH]
(B),
[MATH: nc :MATH]
(C), respectively, along with their correlation coefficients (R) were
sorted and plotted in (A–C) (significant correlation is marked with *).
Highly correlated genes in (A) represent actin genes (RRAS, CXCR4,
TMSB4X, RRAS2, ARPC1B). Significantly correlated genes in (B) represent
motor genes (MSN, ACTN1, MLY12A; MYH9 was added). Significantly
correlated genes in (C) represent clutch genes (VCL, CD44, EMP1, ITGB1,
CAPN2, SHC1). Physical parameter ratios (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) were estimated by averaging the mRNA expression ratios of these
actin, motor, and clutch genes for each Mayo PDX line (66 lines) and
plotted in the 3D CMS parameter space (D) with 2D projections (E, F).
Actin polymerization rate is slightly linearly correlated with
motor-clutch number in (D) (actin = motor*0.3 (P = .01) + clutch*0.3
(P = .07) + 0.01). The clutch number is significantly correlated with
the motor number in (E) (clutch = motor (P < .001) + 0.01). Actin
polymerization rate is significantly correlated with motor number in
(F) (actin = motor*0.7 (P < .001) + 0.01).
Statistics
* Denotes P < .05, ** P < .01, and *** P < .001 derived from the
Kruskal-Wallis test with Dunn–Sidak post hoc analysis.
Code and Data Availability
All codes and data will be made available on reasonable request from
the corresponding author.
Results
Three CMS Physical Parameters Dictate Cell Migration and Traction Force
To understand the relationship between the CMS parameters and its
predictions, we computed the optimal cell motility (RMC), traction
force, and F-actin retrograde flow in the range of substrate
stiffnesses across a wide range of parameter values ([124]Supplementary
Figure 1). Similar to the analysis in Bangasser et al.,^[125]35 the
sensitivities of the fold changes in the CMS predictions to the fold
changes in parameter values, referred to as the parameter
sensitivities, were plotted in [126]Figure 1B. Results showed that
clutch-related parameters (
[MATH: nc :MATH]
,
[MATH: Fb :MATH]
,
[MATH: kon :MATH]
,
[MATH: koff :MATH]
,
[MATH: kc :MATH]
, [127]Table 1) increased the cell traction force and reduced the
F-actin flow, motor-related parameters (
[MATH: nm :MATH]
,
[MATH: Fm :MATH]
,
[MATH: vm :MATH]
, [128]Table 1) increased the F-actin flow and motility, and
actin-related parameters (
[MATH: kmod :MATH]
,
[MATH: vpoly :MATH]
,
[MATH: kcap :MATH]
, [129]Table 1) increased the motility ([130]Figure 1B). We applied
unsupervised hierarchical clustering to the parameter sensitivities,
and the motor, clutch, and actin-related parameters naturally clustered
into motor, clutch, and actin groups, respectively ([131]Figure 1B).
Therefore, the CMS parameters can be categorized broadly into 3 groups,
and each has its own unique influence on predicted cell migration,
which allows us to reduce the 11-dimensional parameter space (see
[132]Table 1) to 3 fundamental dimensions of motor, clutch, and actin
parameters.
Because of the natural clustering into 3 distinct groups, we chose 1
parameter from each clustered group as the fundamental physical
parameters: motor number (
[MATH: nm :MATH]
) representing myosin II motor activity, clutch number (
[MATH: nc :MATH]
) representing functional adhesion protein level, and F-actin
polymerization rate (
[MATH: vpoly :MATH]
) representing actin filament polymerization activity. These 3
parameters are the key components in the cell migration
process,^[133]18–20 they exhibit different values in different cell
types,^[134]24,[135]25 and they are easily manipulated by
drugs.^[136]26 Therefore, we used these 3 parameters (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) as a fundamental basis set to predict glioblastoma cell migration
across subtypes and culture conditions in the absence and presence of
drugs.
To illustrate the cell migration governed by the 3 fundamental physical
parameters, we plotted the optimal motility (RMC), traction force, and
F-actin flow as a function of the 3 physical parameters (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) ([137]Figure 1C). Here we used the grouped-clutch algorithm to
significantly enhance the computational efficiency ([138]Supplementary
Figure2). There are 4 different scenarios observed in the parameter
space ([139]Figure 1C, Conditions I–IV), and the time-dependent cell
protrusion dynamics and traction force fluctuations in these conditions
were plotted in [140]Figure 1D. In Condition I, a typical simulated
migrating cell with a balanced motor and clutch number showed the
dynamic protrusions with sequential phases of nucleation, elongation,
retraction, and elimination, and fluctuating traction force to produce
fast cell migration ([141]Figure 1C and [142]D, Condition I). In
Condition II with the higher motor number in the cells, the
motor-clutch mechanism became “free-flowing”^[143]35 with faster
F-actin flow, lower traction force, shorter protrusion length, and
hence slower cell migration ([144]Figure 1C and [145]D, Condition II).
In Condition III with the higher clutch number, the motor-clutch system
became “stalled”^[146]35 with near zero F-actin flow, near-maximal
constant traction force, limited dynamic protrusions, and hence poor
cell migration ([147]Figure 1C and [148]D, Condition III). In Condition
IV with the higher actin polymerization rate, the protrusion dynamics
became more significant, with longer protrusion length and highly
fluctuating traction forces, to produce faster cell migration
([149]Figure 1C and [150]D, Condition IV). Overall, these simulations
indicate that the CMS fundamental parameters (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) can uniquely describe motor-clutch-mediated cell migration.
Heterogeneity in Migration Phenotypes of Glioblastoma Patient Cells
To understand the migration phenotypes of glioblastoma patient cells,
we measured the cell migration of Mayo glioblastoma PDX lines of MES
(MM, 3 lines), PN (MP, 3 lines), CL (MC, 5 lines) subtypes with
adherent culture^[151]32 and UCSD glioblastoma PD lines of MES (UM, 1
line), PN (UP, 1 line) subtypes cultured as neurospheres^[152]33
(N = 13, [153]Figure 2C). We measured the migration of glioblastoma
cells on PA gels with different stiffnesses, coated with laminin for
the Mayo cells and Matrigel for the UCSD cells, to reach adequate cell
adhesion ([154]Figure 2A and [155]B). We found heterogeneity in cell
migration of glioblastoma cells with different subtypes and sources,
and their mean ± SEM of cell motility (RMC), traction strain energy,
and F-actin retrograde flow rate were all highly variable ([156]Figure
2D). Despite this heterogeneity, we found that glioblastoma cells
tended to have maximal motility on stiffnesses ranging from 4.6 to 19.5
kPa ([157]Figure 2D), which is comparable to brain tissue stiffnesses
(1–6 kPa^[158]36). Mayo MES and PN cells exhibited optimal traction
strain energy with the stiffness of 9.3 kPa, whereas the other cell
lines had low cell traction strain energy with no evidence of
optimality ([159]Figure 2D). All cell lines had no clear optimal
F-actin retrograde flow ([160]Figure 2D), cell area, and aspect ratio
([161]Supplementary Figure 3B) as a function of substrate stiffness.
The differences in cell migration between subtypes and sources were
similar at all substrate stiffnesses (ie Mayo MES had higher motility
than Mayo PN cells at all stiffnesses) ([162]Figure 2D), and therefore,
in subsequent analysis, we combined the cell migration data for a given
PD(X) line across all substrate stiffnesses.
Figure 2.
[163]Figure 2.
[164]Open in a new tab
Heterogeneity in migration phenotypes of glioblastoma patient cells.
(A) Mayo patient-derived xenograft (PDX) cell lines were established as
previously described by implanting patient tumors into mouse
flanks^[165]32 and then cultured in adherent conditions with FBS media.
(B) UCSD patient-derived (PD) cell lines were established as described
previously^[166]33 and cultured in neurosphere conditions with NSC
Media. Mayo and UCSD cells were plated on polyacrylamide gels with
Young’s modulus of 0.7, 4.6, 9.3, and 19.5 kPa, coated with laminin and
Matrigel, respectively. Cell motility (random motility coefficient,
RMC), traction strain energy, and F-actin retrograde flow were measured
using established protocols.^[167]26 (C) There were 3 Mayo MES lines
(MM), 3 Mayo PN lines (MP), 5 Mayo CL lines (MC), 1 UCSD MES line (UM),
and 1 UCSD PN line (UP). (D) The mean ± SEM values of RMC, strain
energy, F-actin flow of Mayo MES, PN, CL cells, and UCSD MES, PN cells
on PA gels with different stiffnesses were highly variable across cell
lines. (E) The mean values of RMC, strain energy, and F-actin flow
combining all substrate stiffnesses for each cell line were plotted in
the 3D space, along with their 2D projections. RMC, tractions strain
energy, and F-actin flow were all highly variable with no obvious
correlations with each other.
The mean cell motility (RMC), traction strain energy, and F-actin flow
rate for each PD(X) line were plotted in the 3D experimental
measurement space with 2D projections ([168]Figure 2E). The cell
migration data of all glioblastoma cells across all subtypes and
sources were plotted in [169]Supplementary Figure 3C with statistical
analysis. Mayo MES cells had significantly higher motility, F-actin
flow, cell area, and aspect ratio compared with Mayo PN cells
([170]Figures 2E and [171]Supplementary Figure 3C). Mayo CL cells had
intermediate values in motility, F-actin flow, and morphology, except
for the lower traction strain energy, compared with Mayo MES, PN cells
([172]Figures 2E and [173]Supplementary Figure 3C). UCSD MES cells had
higher motility and cell area compared with UCSD PN cells ([174]Figures
2E and [175]S3C). All Mayo cells had higher traction strain energy,
F-actin flow, and cell area with lower motility and aspect ratio
compared with all UCSD cells ([176]Figures 2E, and [177]Supplementary
Figure 3C). Overall, these results show significant heterogeneity in
cell migration mechanics of glioblastoma cells across different
subtypes and sources, and a general lack of correlation between
motility, traction force, and F-actin flow rate.
CMS-derived parameters of glioblastoma patient cells exhibit balanced
motors and clutches with F-actin assembly correlating with cell
migration motility.
To transform these empirical measurements of glioblastoma patient cell
migration mechanics into fundamental mechanistic interpretation, we
used the CMS to parameterize the cell migration of PD(X) lines by
fitting their motility, traction force, and F-actin retrograde flow
experimental data to simulations with adjusted physical parameters (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
). For each PD(X) line, this effectively mapped the 3D empirical
observation space into a 3D theoretical physical space. These physical
parameter values were then plotted in the 3D CMS parameter space of (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) for each patient with 2D projections in [178]Figure 3A, also plotted
as bar graphs in [179]Supplementary Figure 4. We found that despite the
apparent heterogeneity of physical parameters, glioblastoma PD(X) lines
consistently showed approximately balanced motor number and clutch
number (
[MATH: nc/
nm∼0.
75 :MATH]
) ([180]Figure 3A), which enabled robust cell migration and traction
forces that increased with the motor-clutch level ([181]Figure 1C).
Therefore, we can also plot the CMS parameter values on the heat map of
various values for (
[MATH: nm :MATH]
,
[MATH: v
poly :MATH]
) with a constant ratio
[MATH: nc/
nm=0.
75 :MATH]
in [182]Figure 3B. We found that Mayo MES cells had higher motor (
[MATH: nm :MATH]
) and clutch (
[MATH: nc :MATH]
) numbers, and higher F-actin polymerization rate (
[MATH: vpoly :MATH]
) compared with PN cells ([183]Figures 3A and [184]B, and
[185]Supplementary Figure 4), resulting in higher motility and F-actin
flow. UCSD MES cells had a higher F-actin polymerization rate (
[MATH: vpoly :MATH]
) compared with UCSD PN cells, resulting in higher motility. Mayo cells
had higher motor and clutch numbers and lower F-actin polymerization to
produce higher traction force and lower motility compared with UCSD
cells ([186]Figure 3A and [187]B). These results suggested that myosin
motors and adhesion clutches were well balanced in all glioblastoma
cells, fast-moving cells had higher F-actin polymerization with either
high or low cell traction, and adherently cultured cells had higher
myosin motors and adhesion clutches and lower F-actin polymerization
compared with neurosphere cultured cells. Overall, the 3D CMS parameter
space (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: v
poly :MATH]
) provides a more fundamental and revealing framework for describing
glioblastoma PD(X) migration than does the empirical 3D space defined
by the measured experimental quantities (cell motility (RMC), traction
strain energy, F-actin retrograde flow).
Figure 3.
[188]Figure 3.
[189]Open in a new tab
Cell migration simulator (CMS)-derived physical parameters of
glioblastoma patient cell migration. (A) Physical parameter values of
glioblastoma cells derived by fitting the cell migration with the CMS
predictions and plotted in the 3D CMS parameter space of (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) with 2D projections. (B) Physical parameter values of glioblastoma
cells were plotted in the heat map of the CMS prediction with various
values for (
[MATH: nm :MATH]
,
[MATH: v
poly :MATH]
) with a constant ratio between
[MATH: nm :MATH]
and
[MATH: nc :MATH]
(
[MATH: nc/
nm=0.
75 :MATH]
). Given the balanced motor-clutch ratio, cell speed is largely
determined by
[MATH: vpoly :MATH]
.
The CMS Predicts PD(X) Glioblastoma Cell Migration Upon Drug Perturbations
Knowing the CMS 3D parameter set for each PD(X) line not only allows us
to predict the cell migration using the CMS, but also to predict
different migration behaviors with the change of parameter values due
to hypothetical drug treatments. The CMS predictions with various
values of (
[MATH: nm :MATH]
,
[MATH: v
poly :MATH]
) and a constant clutch number (
[MATH: nc=22
5 :MATH]
) were plotted as the heat map in [190]Figure 4A, along with the UCSD
MES cells, which had higher F-actin polymerization rate (
[MATH: vpoly :MATH]
) and lower motor number (
[MATH: nm :MATH]
) ([191]Figure 4A orange) but higher motor/clutch ratio ([192]Figure
3A), resulting in higher cell motility and higher F-actin flow
([193]Figure 4A and [194]B, orange) compared with UCSD PN cells
([195]Figure 4A and [196]B, blue).
Figure 4.
[197]Figure 4.
[198]Open in a new tab
The cell migration simulator (CMS) predicts differential cell migration
and F-actin flow sensitivity upon actomyosin drug perturbation. (A) The
heat map of the CMS prediction, including motility (RMC) and F-actin
retrograde flow, with various values of (
[MATH: nm :MATH]
,
[MATH: v
poly :MATH]
) with a constant value of
[MATH: nc=22
5 :MATH]
. UCSD MES cells (UM) had a higher F-actin polymerization rate and
lower motor number compared with UCSD PN cells (UP), which produced
higher motility and higher sensitivity of F-actin flow while reducing
the motor number. (B, C) The mean ± SEM values of RMC and F-actin flow
of UCSD mesenchymal (MES) and PN cells with the decreasing motor number
(B) (solid-red arrow) and decreasing F-actin polymerization rate (C)
(dotted-red arrow) predicted by the CMS were plotted. (D, E) The
mean ± SEM values of RMC and F-actin flow of UCSD MES and PN cells with
different concentrations of myosin II inhibitor, blebbistatin (D), and
F-actin polymerization inhibitor, Latrunculin A (E), were plotted. The
RMC (F) and F-actin flow (G) sensitivities to decreasing parameter
values predicted by the CMS (x-axis in (F, G), also slopes in (B, C)
and [199]Supplementary Figure 5B and C) were correlated with the
sensitivities to cytoskeletal drugs (y-axis in (F, G), also slopes in
(D, E) and [200]Supplementary Figure 5D and E) with P = .15 and
P = .05, respectively.
When reducing the motor number, the motility and F-actin flow of UCSD
MES cells would be reduced to a greater extent than UCSD PN cells as
indicated by the solid-red arrows in [201]Figure 4A, which was
confirmed by the CMS predictions with the reducing motor number (
[MATH: Δnm
:MATH]
) in [202]Figure 4B. Consistent with the CMS predictions, when UCSD
cells were treated with blebbistatin to inhibit their myosin II
contractility, the cell motility and F-actin flow of UCSD MES cells on
PA gels decreased more significantly compared with UCSD PN cells
([203]Figure 4D). This test of the model was not only consistent with
the CMS predictions ([204]Figure 4B), but also confirmed the motor
number difference between UCSD MES and PN cells ([205]Figures 3A and
[206]4A).
When reducing the F-actin polymerization rate, we predicted that the
UCSD PN cells would reduce their motility to a greater extent than UCSD
MES cells, with the insensitivity of F-actin flow, as indicated by the
dotted arrow in [207]Figure 4A, also confirmed by the CMS predictions
with the reducing F-actin polymerization rate,
[MATH: Δvpoly :MATH]
, in [208]Figure 4C. We then treated the UCSD cells with Latrunculin A
to inhibit their actin polymerization by binding G-actin monomers and
found the motility of UCSD PN cells indeed decreased more significantly
compared with UCSD MES cells ([209]Figure 4E), which again was
consistent with the CMS predictions ([210]Figure 4C).
To test whether the culture conditions affect the migration phenotype,
we cultured the neurosphere UCSD MES cells adherently for 1 week to
create UCSD MES-AD cells before conducting migration assays, and we
found the migration phenotypes of UCSD MES-AD cells became closer to
the phenotypes of UCSD PN cells, with higher motor number and lower
F-actin polymerization rate resulting in lower motility and higher
F-actin flow compared with UCSD MES cells ([211]Supplementary Figure 5A
and [212]B). We also treated the UCSD MES-AD cells with cytoskeleton
drugs and found MES-AD cells had lower sensitivity to blebbistatin and
higher sensitivity to Latrunculin A in motility and F-actin flow
compared with UCSD MES cells ([213]Supplementary Figure 5D and [214]E),
which again was consistent with the CMS predictions ([215]Supplementary
Figure 5B and [216]C), and confirmed that adherent culture can increase
the functional motor number and decrease the F-actin polymerization
rate of neurosphere cultured cells ([217]Figure S5A).
Overall, we found that CMS-predicted sensitivities of cell motility
([218]Figure 4F) and F-actin flow ([219]Figure 4G) to decreasing
parameter values were highly correlated with the measured sensitivities
to cytoskeletal drugs. Not only can we use the CMS physical parameter
values of glioblastoma PD(X) lines to describe migration phenotypes
([220]Figure 3), but also to predict the differential migration changes
between patients due to the changes in the parameter values either by
cytoskeletal drugs or by the culture conditions ([221]Figures 4 and
[222]Supplementary Figure 5).
Identification of CMS-Based Transcriptomic Biomarkers
In principle, the CMS parameters should depend on transcriptional
levels of key genes controlling motor, clutch, and F-actin
polymerization activities. Thus, we sought to identify correlates of
the CMS parameters in previously collected transcription-level data. In
[223]Figures 3 and [224]Supplementary Figure 4, there are consistent
differential estimates of the CMS parameters between MES and PN for
both Mayo and UCSD patient cells. Moreover, PN and MES subtypes became
more frequent in patients with tumor recurrences.^[225]15 Therefore, we
decided to analyze the differential genes between MES and PN and
correlate them with the CMS parameters. We first derived RNAseq RPKM
expression of Mayo PDX cells from Vaubel et al.^[226]32 (19 552 genes
in 20 MES patients and 16 PN patients). We filtered out genes with
small mean values and low patient portions to achieve a normal gene
expression distribution ([227]Supplementary Figure 6A and B, 11 752
genes left). We derived 1177 differential genes between Mayo MES and PN
cells using the 2-sample t-test (P < .05, FC > 2,[228] Supplementary
Figure 6C). We applied the pathway enrichment analysis to these
differential genes using the KEGG database^[229]34 with the
FDR-adjusted P < .05 and derived 29 enriched pathways
([230]Supplementary Figure 7). We derived the actin-motor gene list (34
genes) based on the “Regulation of actin cytoskeleton” pathway, and the
clutch gene list (42 genes) based on the “Focal adhesion” and
“ECM-receptor interaction” pathways. This resulted in a final reduced
list of actin-motor and clutch genes that could potentially be
correlated with the CMS physical parameters.
We applied correlation analysis between the mRNA expression of the
actin-motor and clutch genes in the 10 Mayo PDX lines used in the
present study for which mRNA expression level was available, and their
CMS parameter values were (
[MATH: vpoly :MATH]
,
[MATH: nm :MATH]
,
[MATH: nc :MATH]
), respectively. In the correlation of actin-motor genes with the actin
polymerization rate, RRAS had the highest positive (although not
significant) correlation coefficient (R) and FGFR1 had the lowest (and
significant) R ([231]Figure 5A). In the correlation of actin-motor
genes with the motor number, VCL, ARPC1B, RRAS2, MSN, ACTN1, ITGB1,
RRAS, CXCR4, and MYL12A had statistically significant and positive R
(P < .05) ([232]Figure 5B). In the correlation of clutch genes with the
clutch number, VCL, CD44, EMP1, ACTN1, ITGB1, CAPN2, SHC1, and MYL12A
had a positive R that was statistically significant (P < .05)
([233]Figure 5C). In [234]Figure 5A and [235]B, altogether, 18 of 76
genes were significantly correlated (P < 10^−7 for a Poisson
distribution with mean = 0.05 × 76 = 4 and observed 18 events) and 17
of 18 genes were positively correlated (P < 10^−5 for binomial
distribution with 17 successes in 18 trials with P = .05 success in 1
trial), which is much higher than significance by chance (95%
confidence interval, 5% of the 76 total genes = ~4 genes). This result
validates the actual correlation of the CMS parameters to the gene
expression in cell migration.
With the genes highly correlating with the CMS parameters, we were able
to estimate the physical parameters of Mayo PDX lines (N = 66) based on
their mRNA expressions. RRAS, CXCR4, TMSB4X, RRAS2, and ARPC1B were
highly correlated with actin polymerization rate ([236]Figure 5A) and
were chosen to represent actin genes. MSN, ACTN1, MLY12A; MYH9 were
significantly correlated with a motor number ([237]Figure 5B) and
represent motor genes (MYH9 was added as a key component in myosin
assembly with P = .067 slightly above our cutoff of P = .05). VCL,
CD44, EMP1, ITGB1, CAPN2, and SHC1 were significantly correlated with a
clutch number ([238]Figure 5C) and represent clutch genes. By averaging
the mRNA ratios of these correlated genes (5 actin genes, 4 motor
genes, and 6 clutch genes), we estimated the physical parameter ratios
for each Mayo PDX line, plotted in the 3D CMS parameter space of (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) ([239]Figure 5D) with 2D projections ([240]Figure 5E and [241]F). MES
lines (red dots in [242]Figure 5) had higher (
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
) compared with PN lines (blue dots in [243]Figure 5), with CL lines
having intermediate values (green dots in [244]Figure 5). The trends of
the highly balanced motor and clutch number with correlated F-actin
polymerization ([245]Figure 5D–F) do not change when a smaller focused
subset of 3 genes was chosen for each parameter ([246]Supplementary
Figure 8). Even a single gene can largely reflect this same trend
([247]Supplementary Figure 9).
We also acquired RNAseq normalized TPM values from TCGA (151 patients,
20 531 genes) and applied Cox regression analysis on migration gene
expression and patient survival for both Mayo and TCGA patients, and we
found 17 genes were significantly correlated with patient survival, and
TCGA patients have more correlated genes than Mayo patients
([248]Supplementary Figure 10). Within these genes, MYL12A and ITGB1
can be regarded as representative motor and clutch genes, which is the
possible mechanism driving the correlation with patient survival. There
are fewer correlated genes in MES ([249]Supplementary Figure 11) and PN
([250]Supplementary Figure 12) patients, with larger patient variation,
due to the limited number of patients. In the end, our approach is an
integration of mechanistic modeling and data science approaches.
Discussion
In this study, we used the physics-based motor-clutch model, termed
here the CMS^[251]26 ([252]Figure 1A), to mechanistically parameterize
and predict glioblastoma cell migration mechanics and speed. We reduced
the 11-dimensional parameter space of the CMS (see [253]Table 1) into 3
dimensions based on their parameter sensitivities ([254]Figure 1B) and
identified 3 fundamental physical parameters: motor number, clutch
number, and actin polymerization rate that can uniquely govern cell
migration ([255]Figure 1C and [256]D). We found significant
heterogeneity in glioblastoma patient cell migration across subtypes
and sources ([257]Figure 2) and derived the physical parameter values
for each cell line by fitting their cell migration with the CMS
predictions ([258]Figure 3). Despite their heterogeneity, glioblastoma
cells had balanced motor/clutch ratios (
[MATH: nc/
nm∼0.
75 :MATH]
) to produce robust cell migration and traction force ([259]Figures 1C
and [260]3A). In addition, we found consistent trends by molecular
subtype, with Mayo MES cells having higher motor-clutch number and
F-actin polymerization rate relative to Mayo PN cells ([261]Figure 3),
resulting in higher motility and F-actin flow ([262]Figure 2E).
Similarly, UCSD MES cells had a higher F-actin polymerization rate
relative to the UCSD PN cells ([263]Figure 3) resulting in faster
migration ([264]Figure 2E). Moreover, the CMS accurately predicted the
differential sensitivities between MES and PN cells to cytoskeletal
drug perturbations ([265]Figure 4). Finally, we derived a list of
motor-clutch-associated genes in the Mayo cells having mRNA expression
correlating with the physical parameters of Mayo cells, which can be
used to predict CMS cell migration parameters and speeds. Overall, we
describe a simplified 3D physics-based framework for mechanically
parameterizing individual glioblastoma patients and connecting
biomechanics to clinical transcriptomic data, which can potentially be
used to predict cell migration and drug responses in glioblastoma
cells.
Our present study used 2D measurements with a range of stiffnesses that
have been reported for brain tissue (1–10 kPa).^[266]36 In addition,
our recent studies find that 2D measurements are predictive of 1D
confined migration in vitro^[267]37 and of migration in 3D brain tissue
ex vivo.^[268]5,[269]8 Our previous studies showed that the
motor-clutch model is relevant to glioblastoma cell migration in brain
tissue.^[270]5,[271]7,[272]8,[273]31 We also note that the 2D in vitro
migration speeds are predictive of clinical MRI features.^[274]31
Therefore, we have evidence that the 2D biomechanical measurements we
are making here will be directly relevant to 3D migration in brain. We
found that glioblastoma MES cells have faster migration than PN and CL
cells on 2D compliant PA gels, which is consistent with the 3D invasion
of glioblastoma spheroids in Munthe et al..^[275]38 Piao et al.^[276]39
similarly found the glioblastoma cell lines similar to the MES subtype
had higher invasive capacity and motility compared with other subtypes.
To provide the biophysical mechanisms of the differential cell
motility, we parameterized the cell migration of glioblastoma cells
with the CMS and found the higher F-actin polymerization rate best
explained the faster migration of the MES cells relative to the PN
cells ([277]Figures 3 and [278]Supplementary Figure 4). In the CMS
simulation, a higher F-actin polymerization rate promotes cell
protrusion dynamics, with longer protrusion length and highly
fluctuating traction force, causing highly unbalanced protrusion
forces, highly polarized cell morphology, and hence faster migration
([279]Figure 1C and [280]D). Adebowale et al.^[281]28 applied the CMS
simulation coupled with viscoelastic substrate and also confirmed that
the significant filopodia dynamics with longer filopodia length and
lifetime resulted in faster cell migration of fibrosarcoma cells on
fast-relaxing viscoelastic gels, demonstrating the protrusion dynamics
promoting cell motility in vitro. Therefore, F-actin polymerization
becomes a potential target to alter glioblastoma cell motility, and the
CMS can potentially predict the patient-specific cell motility and
treatment responses.
The subtype definitions/classifiers have shifted over time since the
original classification by Phillips et al.^[282]14 (PN, Proliferative,
and MES) and we are using the published classifications for these PDX
collections, following Wang et al.^[283]16 (PN, CL, and MES), which was
itself an update from earlier work by Verhaak et al.^[284]15 (PN,
Neural, CL, and MES). Wang et al.^[285]16 concluded that the Neural
subtype was likely due to contamination from adjacent brain tissue. The
recent single-cell RNAseq work (Neftel et al.^[286]40) has supported
the view of 4 subtypes, rather than 3, along with both intratumoral
heterogeneity and plasticity. Even so, Neftel et al.^[287]40 described
how their classification can be mapped onto the previous Wang-Verhaak
classification,^[288]16 with
oligodendrocyte-progenitor/neural-progenitor (OPC/NPC) being associated
with PN, astrocyte being associated with the CL subtype, and the same
MES subtype. Overall, for the past 15 years, the field has had a PN
(OPC/NPC) to MES axis, with the CL (AC) subtype being a robust
intermediate substate. Hence the PN/CL/MES classification used in our
study is very much in line with the current standard in the field, and
the subtypes for the Mayo PDX and UCSD PD lines have been
published.^[289]32,[290]33 In the present study, we account for the
variability on a patient-by-patient basis. Heterogeneity and plasticity
are clearly potential confounders of simple categorizations. However,
if these were strong effects, then the subtype classification of
patients based on a single tissue sample would not yield consistent
patterns with mechanistic measurements such as ours, that is, a similar
pattern of the model-based parameterization of
[MATH: nm :MATH]
,
[MATH: nc :MATH]
,
[MATH: vpoly :MATH]
throughout patient cell lines. In particular, we found that MES cells
had higher estimated
[MATH: vpoly :MATH]
, which enables faster migration, and higher motor-clutch levels
([291]Figures 3B and [292]5D). Our results suggest that
[MATH: vpoly :MATH]
and motor-clutch ratio, and their associated genes, may be better
predictors of functional subtype than the traditional PN/CL/MES
classification. However, we believe that this will require prospective
tests in more in vivo–like settings in the future.
By correlating with the CMS parameters, we found the representative
actin, motor, and clutch genes in [293]Figure 5A–C. Based on the
predicted cell migration in [294]Figure 1C, we can identify
patient-specific strategies to target the cell migration of
glioblastoma. For example, for the patients with higher motor-clutch
ratios ([295]Figure 5E, orange dotted line), we can inhibit adhesion
clutches resulting in free-flowing conditions ([296]Figure 1C and
[297]D, Condition II). For patients with lower motor-clutch ratios
([298]Figure 5E, brown dotted line), we inhibit myosin motors resulting
in stalled conditions ([299]Figure 1C and [300]D, Condition III). For
patients with a higher actin polymerization rate ([301]Figure 5F, green
dotted line, mostly MES lines), we then inhibit the actin
polymerization to block the migration. We also applied the Cox
regression analysis on the migration gene expression and patient
survival in the Mayo and TCGA cohorts. Among the genes highly
correlated with patient survival ([302]Supplementary Figure 10), ITGB1,
ITGA5, ITGA3 showed significant association with the patient survival
in Malric et al.,^[303]41 consistent with our results. The negative
correlation between CD44 levels and PN patient survival aligned with
Klank et al.’s.^[304]7 Collagen (COL9A2, COL6A2, COL4A1, COL4A2)
enrichment is associated with poor prognosis of glioblastoma
patients.^[305]42–44 Overall, these genes are significantly hazardous
to patient survival, and some of them (ITGB1, CD44) are also closely
associated with the CMS parameters, which makes them potential markers
and targets to enhance patient survival based on patient-specific
transcriptomic information.
Our study has some limitations and points to future work. While we
would expect that the faster migration speed in MES cells may
contribute to the lower survival found in MES patients vs. CL or
PN,^[306]15,[307]16 we suspect that other confounding variables need to
be included in the further analysis, such as age^[308]2 and immune
response^[309]15,[310]16 in order to better predict patient survival.
While we had hoped to identify genes that are associated with the
F-actin polymerization rate, we found only 1 statistically significant
correlation. It is likely that we will need a larger cohort of patient
cells with greater sequencing depth in actin, motor, and clutch genes.
In order to be of clinical utility, future work will need to
prospectively test transcriptomic predictions in terms of cell
migration and drug sensitivities, rather than the retrospective
relationships identified in the present study. Even so, the overall
results provide a proof of concept that CMS-based mechanical biomarkers
can be used to describe cell migration dynamics, predict differential
drug sensitivities, and identify correlations with mechanistically
relevant mRNA transcript levels. This platform allows us to predict on
a patient-by-patient basis the most effective intervention (ie
intervention with the greatest predicted sensitivity) to suppress
cancer cell migration, similar to how migration speed can be estimated
via machine learning–based detection of features in clinical MRI
images.^[311]31
Supplementary Material
vdae184_suppl_Supplementary_Figures_S1-S12
[312]vdae184_suppl_supplementary_figures_s1-s12.zip^ (11.2MB, zip)
Contributor Information
Jay Hou, Department of Neurosurgery, Rhode Island Hospital—Brown
University Health, Providence, Rhode Island, USA.
Mariah McMahon, Department of Biomedical Engineering, University of
Minnesota—Twin Cities, Minneapolis, Minnesota, USA.
Tyler Jubenville, Department of Pediatrics, University of
Minnesota—Twin Cities, Minneapolis, Minnesota, USA.
Jann N Sarkaria, Department of Radiation Oncology, Mayo Clinic
Rochester, Rochester, Minnesota, USA.
Clark C Chen, Department of Neurosurgery, Brown University, Providence,
Rhode Island, USA.
David J Odde, Department of Biomedical Engineering, University of
Minnesota—Twin Cities, Minneapolis, Minnesota, USA.
Funding
This work was supported by the National Institutes of Health (grant
numbers U54CA210190, P01CA254849 to D.J.O.).
Conflict of interest statement
None declared.
Authorship statement
Conceptualization: J.H., M.M., D.J.O. Funding acquisition: D.J.O. Cell
line development: J.N.S., C.C.C. Investigation, methodology, data
analysis, visualization: J.H., M.M. Simulation and modeling: J.H. mRNA
expression analysis: J.H., T.J. Supervision: D.J.O. Writing (original
draft): J.H., M.M., D.J.O. Writing (review and editing): all authors.
References