Multiscale cartilage biomechanics: technical...
Transcript of Multiscale cartilage biomechanics: technical...
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
rsfs.royalsocietypublishing.org
ResearchCite this article: Erdemir A, Bennetts C, Davis
S, Reddy A, Sibole S. 2015 Multiscale cartilage
biomechanics: technical challenges in realizing
a high-throughput modelling and simulation
workflow. Interface Focus 5: 20140081.
http://dx.doi.org/10.1098/rsfs.2014.0081
One contribution of 11 to a theme issue
‘Multiscale modelling in biomechanics:
theoretical, computational and translational
challenges’.
Subject Areas:biomechanics, computational biology,
biophysics
Keywords:cartilage, extracellular, pericellular,
chondrocyte, finite-element analysis,
post-processing
Author for correspondence:Ahmet Erdemir
e-mail: [email protected]
& 2015 The Author(s) Published by the Royal Society. All rights reserved.
Multiscale cartilage biomechanics:technical challenges in realizing ahigh-throughput modelling andsimulation workflow
Ahmet Erdemir1,2, Craig Bennetts1,2, Sean Davis1,2,3, Akhil Reddy1,2,4
and Scott Sibole1,2,5
1Computational Biomodeling (CoBi) Core, and 2Department of Biomedical Engineering, Lerner ResearchInstitute, Cleveland Clinic, Cleveland, OH 44195, USA3Department of Mechanical Engineering, University of Akron, Akron, OH 44325, USA4Weill Cornell Medical College, New York, NY 10065, USA5Human Performance Laboratory, Faculty of Kinesiology, University of Calgary, Calgary, Alberta, Canada T2N 1N4
Understanding the mechanical environment of articular cartilage and chon-
drocytes is of the utmost importance in evaluating tissue damage which is
often related to failure of the fibre architecture and mechanical injury to the
cells. This knowledge also has significant implications for understanding
the mechanobiological response in healthy and diseased cartilage and can
drive the development of intervention strategies, ranging from the design of
tissue-engineered constructs to the establishment of rehabilitation protocols.
Spanning multiple spatial scales, a wide range of biomechanical factors dictate
this mechanical environment. Computational modelling and simulation pro-
vide descriptive and predictive tools to identify multiscale interactions, and
can lead towards a greater comprehension of healthy and diseased cartilage
function, possibly in an individualized manner. Cartilage and chondrocyte
mechanics can be examined in silico, through post-processing or feed-forward
approaches. First, joint–tissue level simulations, typically using the finite-
element method, solve boundary value problems representing the joint
articulation and underlying tissue, which can differentiate the role of compart-
mental joint loading in cartilage contact mechanics and macroscale cartilage
field mechanics. Subsequently, tissue–cell scale simulations, driven by the
macroscale cartilage mechanical field information, can predict chondrocyte
deformation metrics along with the mechanics of the surrounding pericellular
and extracellular matrices. A high-throughput modelling and simulation
framework is necessary to develop models representative of regional and
population-wide variations in cartilage and chondrocyte anatomy and mech-
anical properties, and to conduct large-scale analysis accommodating a
multitude of loading scenarios. However, realization of such a framework is
a daunting task, with technical difficulties hindering the processes of model
development, scale coupling, simulation and interpretation of the results.
This study aims to summarize various strategies to address the technical chal-
lenges of post-processing-based simulations of cartilage and chondrocyte
mechanics with the ultimate goal of establishing the foundations of a high-
throughput multiscale analysis framework. At the joint–tissue scale, rapid
development of regional models of articular contact is possible by automating
the process of generating parametric representations of cartilage boundaries
and depth-dependent zonal delineation with associated constitutive relation-
ships. At the tissue–cell scale, models descriptive of multicellular and
fibrillar architecture of cartilage zones can also be generated in an automated
fashion. Through post-processing, scripts can extract biphasic mechanical
metrics at a desired point in the cartilage to assign loading and boundary con-
ditions to models at the lower spatial scale of cells. Cell deformation metrics
can be extracted from simulation results to provide a simplified description
of individual chondrocyte responses. Simulations at the tissue–cell scale can
rsfs.royalsocietypublishing.orgInterfa
2
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
be parallelized owing to the loosely coupled nature of the
feed-forward approach. Verification studies illustrated the
necessity of a second-order data passing scheme between
scales and evaluated the role that the microscale represen-
tative volume size plays in appropriately predicting the
mechanical response of the chondrocytes. The tools sum-
marized in this study collectively provide a framework
for high-throughput exploration of cartilage biomecha-
nics, which includes minimally supervised model
generation, and prediction of multiscale biomechanical
metrics across a range of spatial scales, from joint regions
and cartilage zones, down to that of the chondrocytes.
ceFocus
5:20140081
1. Background and motivationThe mechanics of cartilage and chondrocytes is a function of
complicated load sharing between musculoskeletal joints of
the body, individual tissue structures within the joint, and
the organization and mechanical properties of cells and micro-
structure within the tissue (figure 1). The external loads applied
to the body are translated to musculoskeletal joints, which may
also endure large applied muscular forces. To achieve equili-
brium, the passive structures of the joints often must sustain
high stresses. Contact is the modality through which cartilage
sustains load, where different regions of the musculoskeletal
joint may exhibit varying contact force transmission. As one
moves from the intermediate scale of joint compartments to
the lower spatial scale resolution of the tissue, the load trans-
mission at the articular surface is now resolved as a contact
stress field on the cartilage surface. This load distribution is
not only dependent on the total contact force, but it is also
likely to be influenced by the cartilage geometry and properties,
i.e. its thickness, radius of curvature and macroscopic material
properties. Different zones along the depth of the cartilage
(superficial, transitional and deep) exhibit varying stresses,
strains, fluid pressures, etc., depending on the zonal thickness
and underlying material properties, which are dictated by vari-
ations in collagen matrix organization and proteoglycan
content [2]. These zonal loads influence the deformation and
loading of chondrocytes, which are dependent on the relative
mechanical properties of chondrocytes, and pericellular and
extracellular matrices [3]. Chondrocyte shape, size and possibly
their organization affect the loads transmitted to these cells and
to their nuclei, which may cause cell damage and death [4] or
may trigger mechanobiological responses for the regulation of
extracellular matrix integrity [5].
A comprehensive understanding of the mechanical environ-
ment of cartilage and chondrocytes potentially has significant
implications to evaluate the function of this tissue in health
and disease. Such an understanding is necessary as cartilage
bears loads exceeding multiple body weights during the
activities of daily life and the tissue’s resulting mechanical
environment can lead to regulation of its structural integrity.
Mechanical insults not only result in matrix damage, but may
also decrease chondrocyte viability [6]. Chondrocytes also
respond to their mechanical environment, where mechanical
transduction results in biological responses to maintain the
structure of the extracellular environment [5]. Changes in
the multiscale load sharing pathway alter mechanical signals
transmitted to the cells, possibly compromising the mechanobio-
logical function of chondrocytes. When quantified for healthy
populations, the joint–tissue and tissue–cell scale mechanical
environments of cartilage can better elucidate the forcing func-
tions of chondrocyte mechanobiology, i.e. cellular tractions
and osmotic pressure gradients [7]. Such information can be
used as a design specification for tissue engineering of cartilage,
i.e. how construct loading protocols should be implemented to
promote matrix growth and adaptation in cell-seeded con-
structs. Mechanics also has a significant role in chronic and
acute disorders of the cartilage. For example, the mechanical
capacity and function of the cartilage and chondrocytes are
affected during the onset and progression of osteoarthritis [8].
This debilitating musculoskeletal disease is the most common
form of arthritis. Osteoarthritis influences more than 30% of
adults above the age of 65 [9] and can manifest itself in any syno-
vial joint. The knowledge of cartilage and chondrocyte loading
can also indicate when and how trauma or tears may occur
and it may offer an explanation for the formation of chondral
defects. In any of these diseased or injured populations, quanti-
fication of cartilage and chondrocyte mechanics may have
additional value for mechanical risk assessment and for diag-
nostic purposes, and may assist in the evaluation of disease
progression or efficacy of prevention and intervention strategies.
Computational modelling and simulation provides a means
of quantifying the mechanical environment of the cartilage and
chondrocytes [1]. In vivo measurement of cartilage and chon-
drocyte mechanics in human subjects, where the joints are
exposed to loadings of daily activities, is not necessarily
within the reach of the current state of the art. Recent exper-
imentation illustrated the possibility to quantify cartilage and
chondrocyte deformations of fresh cadaveric samples in situ[10] and of animals in vivo [11], through invasive approaches.
Modelling and simulation provide a means for in-depth
exploration of such data through descriptive analysis [12]. In
addition, through predictive simulations, it may be possible
to establish multiscale mechanical interactions and to evaluate
the influence of disease- or intervention-related perturbations
on the mechanics of cartilage and chondrocytes [1].
Musculoskeletal joints exhibit large variations, anatomically
and mechanically, at multiple scales, within an individual
and across populations. Regional cartilage anatomy (curvature,
thickness), stiffness, zonal properties, i.e. delineation of super-
ficial, transitional and deep layers, and fibrillar architecture
[13–16], and chondrocyte organization, shape, size [17,18]
and mechanical properties [19] all vary largely among species
and between individuals and can be altered by injury or
disease. A high-throughput modelling and simulation frame-
work seeks to provide the ability to represent such variations
and conduct individualized analysis to assess this multiscale
mechanical system under various loading conditions, which
can be used for understanding trauma risk, for evaluating the
protective performance of interventions, and for diagnosis
and prognosis through the use of biomechanical markers.
Likewise, parametric analysis supported by large-scale unsuper-
vised simulations can allow probabilistic studies (for population
analysis or for uncertainty estimation) [20], inverse analysis
(for estimation of patient-specific mechanical properties) [21]
and virtual prototyping (to design new therapeutic, surgical or
rehabilitative management strategies) [22]. While the use of
modelling and simulation for explorations of cartilage and chon-
drocyte mechanics is appealing, routine investigations are
hampered by many technical challenges associated with the
need to accommodate the nonlinear, multiscale and multiphasic
biomechanical nature of this tissue [1].
body~1 m
jointloads
jointloads
contact
externalloads
bonebone
cartilage
cartilage
cartilageloads
meniscus
ligament
joint~0.1 m
tissue~0.001 m
cell/fibre~0.00001 m
fibrearchitecture
chondrocytemuscleforces
Figure 1. The mechanics of articular cartilage and the chondrocytes are dictated by a multiscale load sharing pathway (illustration based on Halloran et al. [1]).External forces and inertial loads of the body are reflected upon musculoskeletal joints. The passive structures of the joint, e.g. cartilage, ligaments, menisci, bearlarge forces and deformations, which are distributed as a function of the joint’s anatomical architecture and the properties of individual tissues. For the articularcartilage, these loads are mainly in the form of contact, which induce deformations of the superficial, transitional and deep layers along the tissue’s depth andeventually the mechanics of chondrocytes, the cell type found in cartilage. (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
3
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
The goal of this paper is to summarize various strategies to
address the technical challenges of realizing a modelling and
simulation workflow to characterize the mechanical environ-
ment of cartilage and chondrocytes. The tissue of interest is
articular cartilage, but the outlined strategies are equally appli-
cable to fibrocartilage and other tissues. The modelling and
simulation strategy is based on a post-processing-type analysis
approach (from higher spatial scales to lower spatial scales),
where mechanical load sharing across joint–tissue and
tissue–cell scales can be differentiated in a feed-forward
manner. Individual components of the workflow are aimed to
address difficulties in model development and its automation,
scale coupling, computational burden and extraction of rel-
evant biomechanical metrics from simulation results.
Immediate utilization of a streamlined in silico framework
for understanding the regional and population-wide mechan-
ical responses of cartilage and chondrocytes may still not be
computationally feasible, particularly when the physical
phenomena need to be represented at a higher fidelity. None-
theless, approaches outlined here lay the foundation of a
high-throughput framework for large-scale analysis of cartilage
mechanics at multiple scales.
2. High-throughput frameworkThe modelling and simulation work presented herein provides
a feed-forward workflow where body-level musculoskeletal
loading is differentiated to tissue-level and cell-scale defor-
mations using multiple models in sequence (figure 2). If one
is interested in a certain contact region of the joint, e.g.
cartilage-to-cartilage contact area of the knee’s medial com-
partment, a model at the joint–tissue scale (spatial lengths of
0.001–0.01 m) can be constructed for prediction of detailed con-
tact mechanics and, more importantly, the internal mechanics of
the cartilage. Such models should incorporate the articulation
between the cartilage surfaces for evaluation of the distribution
of load transmission through the contact interface. These models
will ideally include zonal thickness and material properties of
the cartilage as well to appropriately predict the internal mech-
anical response of the tissue. Driven by the regional contact force
(or articulation kinematics), simulations can predict contact
pressure distributions and the detailed mechanics of the
superficial, transitional and deep zones. In the case of solid
phase analysis, tissue-level mechanical metrics include strains
and stresses. With the addition of the fluid phase, as in biphasic
simulations, they also include fluid flux and pressure. Such
information can be used to drive cell scale models to predict
chondrocyte deformation metrics. For example, one can pick
up any point in the joint–tissue scale model, e.g. midpoint of
the transitional zone, extract the deformation gradient, fluid
pressure and fluid pressure gradient from the solution of the
joint–tissue scale model, and apply these as loading and bound-
ary conditions to a tissue–cell scale model (spatial lengths
of 1026 to 0.0001 m) representative of the cartilage and chondro-
cyte characteristics of that specific region [23]. This process is
equivalent to using a one term Taylor series to approximate
the field of displacements and fluid pressures in the neighbour-
hood of the point of interest. From the results of the cell scale
models, chondrocyte mechanical metrics can be calculated
including the cell aspect ratio change, average effective (von
Mises) stress and strain, and mass exchange rate. Through
appropriate parametrization, streamlined data flow in between
scales, automation of post-processing and effective comput-
ing strategies, a high-throughput modelling and simulation
framework can be realized for prediction of cartilage and
chondrocyte mechanics.
At the spatial of scale of the musculoskeletal joint, simu-
lations using a finite-element (FE) representation of the
complete joint anatomy [24], e.g. the tibiofemoral joint, includ-
ing its substructures such as the ligaments and meniscus,
can provide the necessary information to define loading and
boundary conditions of the aforementioned joint–tissue scale
model. It should be noted that a comprehensive model,
where all structures of the joint are represented, may be used
to bypass the use of the compartmental joint–tissue scale
model [25]. Yet, a detailed representation of zonal anatomy,
requiring a finer level of discretization to accommodate zonal
thicknesses, may be computationally prohibitive to include in
the representation of the entire musculoskeletal joint. Contact
information, e.g. force or articulation kinematics, to drive the
joint–tissue scale model may be estimated from experimen-
tation. Such measurements are possible, either directly on
cadaver specimens by using contact pressure measurement
systems [26], or indirectly on live subjects through the use of
biplanar fluoroscopy by estimating proximity of articulating
regionaljoint loading
regionaljoint–tissue scale
model
zonaltissue–cell scale
model
zonaltissue
mechanicschondrocytemechanics
finite-elementanalysis
finite-elementanalysis
Figure 2. Sequential simulations using models at multiple spatial levels can predict cartilage and chondroycte loads and deformations. Regional joint – tissue scalemodels can provide representations of joint articulation and cartilage at the spatial order of 0.001 – 0.1 m. Zonal tissue – cell scale models allow modelling of cartilageand chondrocytes at the spatial order of 1026 to 0.0001 m. Given joint-level contact loads (or contact kinematics), the joint – tissue scale model of interfacing articularcartilage regions can be solved to obtain mechanical field information (deformation gradients, fluid pressure, etc.) for various layers of the cartilage. For a given point ofinterest in this macroscale tissue representation, the mechanics information can be mapped to a zone-specific tissue – cell scale model as boundary conditions.Subsequently, FE analysis can predict chondrocyte mechanics. The volume of the joint – tissue scale model can be adjusted to include a desired contact region ofthe joint such as cartilage-to-cartilage interaction area within the medial compartment of the knee. Detailed anatomical features of the cartilage’s zonal delineationcan also be modelled (illustrated as layers in different colours). The tissue – cell scale model represents chondrocyte distribution (seven cells are explicitly shown)and anatomical features including the representation of the pericellular matrix (transparent layer around cells). (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
4
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
surfaces [27]. In other words, an elaborate model of the joint
can be foregone, and a regional model can be employed
instead, with boundary conditions and loads taken directly
from regional measurements in the experiment.
Extensibility is paramount for a modelling and simulation
framework as the desired utility of the in silico workflow in
future may necessitate replacing models and analysis types
with higher fidelity versions. The feed-forward approach
lends itself to expansion towards higher and lower spatial
scales. At the higher spatial scale, body-level musculoskeletal
models can be used to determine muscular joint loading [28],
which can be processed to drive joint-level FE analysis, e.g.
for muscle force-driven patellofemoral joint contact analysis
[29]. At the lower spatial scale, cell deformation metrics
can be used to drive cellular–subcellular scale models,
i.e. incorporating the cytoskeleton (actin and tubulin) [30], to
understand the mechanical load transmission to intracellular
components, in particular to the nucleus. Similar submodelling
can be performed for the extracellular matrix, for example, to
evaluate microstructural-level loading of collagen fibres,
which may provide the relationship between macroscopic
tissue damage accumulation and microscale fibre failure [31].
At all spatial scales, it is also possible to increase the fidelity
of mechanical representation; in terms of physics (nonlinear
elasticity, multiphasic response) [32], with additional anatom-
ical components, and by using more advanced constitutive
relationships [24,33]. Utilization of open source and freely
accessible software, e.g. PYTHON (for scripting to automate
tasks; https://www.python.org/), SALOME (for geometry and
mesh generation; http://www.salome-platform.org/) and
FEBIO (for FE analysis; http://www.febio.org/) [34], facilitates
future development by any interested parties.
3. Tackling challengesThis section summarizes our recent published and unpub-
lished work to overcome various technical challenges
associated with high-throughput modelling and simulation
in multiscale biomechanics of the cartilage and chondrocytes.
The term throughput is used in a general sense, referring to
the rate at which something, in this case modelling and simu-
lation, can be conducted. The challenges (and the proposed
solutions) span the life cycle of the modelling and simulation
workflow. To enable high-throughput modelling and simu-
lation, the ‘challenge of automated model development’
should be met. Both joint–tissue and tissue–cell scale models
need to be developed on-the-fly and in an unsupervised
manner to facilitate large-scale analysis. Irrespective of being
high-throughput or not, the ‘challenge of appropriate scale
coupling’ exists. If the information exchange between scales
is erroneous, multiscale analysis of cartilage may become a use-
less effort. High-throughput multiscale simulations of cartilage
and chondrocytes also require expedited solutions of large
number of models, i.e. the ‘challenge of computing’, for dif-
ferent joint regions, tissue zones and loading conditions.
Interactive post-processing of simulation results, in order to
understand biomechanics of cartilage and chondrocytes, is
not suitable for high-throughput analysis. Therefore, the ‘chal-
lenge of streamlined extraction of information’ needs to be
addressed. The ‘challenge of accessible modelling and simu-
lation framework’ is readily addressed by using free and
open source software and publicly sharing models and tools
developed as part of the summarized studies (see §5).
3.1. Automation of modelling3.1.1. Joint – tissue scale model developmentFor a given contact region at the joint–tissue scale, the articulat-
ing cartilage tissue exhibits variations in its gross geometry,
e.g. thickness [13]. These parameters dictate how contact press-
ures are distributed during load transmission between
interfacing cartilage, e.g. for the hip [35]. Within the tissue,
superficial, transitional and deep zones of the cartilage may
have varying thicknesses and their mechanical responses
depend on the fibrous network of collagen (primarily type
joint region of interest
parametric modelling
anatomicalF
bone
bone
wxyz
contact
cartilage
rb
rd
rtrs
region size (w × h × l)zonal surface curvature (r)zonal thickness (t)
discretizationt(x, y, z)
E(x, y, z)v(x, y, z)k(x, y, z)
mechanical
examples
elastic modulus (E)Poisson’s ratio (v)permeability (k)fibre properties (x, b, –)contact force (F)lateral wall boundary conditions
x(x, y, z) b(x, y, z) –(x, y, z)
mesh density
Figure 3. Parametrization can simplify the modelling process at the spatialscale incorporating joint- and tissue-level features. Articular cartilage exhibitslarge variations at this spatial level. The curvature of contact surfaces andboundaries of the superficial, transitional and deep zones can be differentbased on the individual, the musculoskeletal joint and the region of interestwithin the joint. Cartilage also has depth-dependent mechanical properties,dictated by collagen and proteoglycan content, and collagen fibre architec-ture. Through an automated pipeline, the generation of numerous modelsand the execution of simulations under different loading conditions maybe conducted with ease. (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
5
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
II), water content (75–80% wet weight), dissolved ions and
aggregating proteoglycans [2]. The material properties are
depth-dependent [36], and as a result of varying fibre align-
ment in each zone, they also exhibit anisotropy [37]. All these
influence how the mechanical deformations and loads
distribute within the tissue.
A streamlined workflow for model generation at the
joint–tissue scale should exploit the parametrization of ana-
tomical and mechanical properties such that any desirable
variations can be represented expeditiously in a minimally
supervised manner, e.g. within an input text file. Our team
has been working on the development of an automated
model generation procedure to represent regional anatomy
and depth-dependent material properties for articulating
regions of cartilage [38,39] (figure 3). PYTHON scripting was
used along with SALOME to automate geometry and mesh
generation and to prepare a model ready to be analysed
using FEBIO, including definitions of material properties
and prescription of loading and boundary conditions. The
user can specify anatomical parameters in a human-readable
configuration file, which includes overall dimensions of the
articulating region of interest, cartilage boundaries and
zonal delineation with ellipsoidal surface approximations
(for all zones, essentially allowing representation of non-
uniform thickness), and mesh density. The execution of
PYTHON scripts generates the model geometry and discretizes
the whole region (into a hexahedral mesh) with appropriate
element, node and surface sets. The user can also prescribe
depth-dependent material properties, e.g. elastic modulus,
Poisson’s ratio, hydraulic permeability, fibre properties [37],
all as a function of FE centroid location. The regional contact
force (to drive the model) and boundary conditions for the
margins of the model, e.g. confined or unconfined, are other
user-specified inputs. In an additional step, execution of
PYTHON scripts allows the generation of a human-readable
model input file that can be used with FEBIO to conduct FE
analysis. Results of such analyses quantify the mechanical
response of the cartilage at a given anatomical and mechanical
state and for preferred loading and boundary conditions, which
can also be fed-forward to tissue–cell scale models for further
simulations. This scripted environment allows generation of a
plethora of models with different characteristics (figure 3),
with minimal user interaction through the automation of
geometry, mesh and model definition processes.
3.1.2. Tissue – cell scale model developmentModels at the tissue–cell scales, when and if generated in an
expedited fashion, can allow large-scale estimation of chondro-
cyte mechanics at any desired location within a given zone of
the cartilage. Diverse anatomical variations exist for chon-
drocyte organization, shape and size across the layers of
cartilage and sometimes within the layers [17]. A pericellular
matrix surrounds the chondrocyte(s) to form a unit called a
chondron. In the superficial zone, the chondrocytes exhibit
flatter, ellipsoidal shapes, with relatively denser cellular distri-
bution [17]. Depending on the body region, e.g. for the ankle
[40], several cells may cluster to form a multicellular chondron.
In the transitional zone, chondrocyte distribution is sparser
[17]. Chondrocytes are nearly spherical and typically form
chondrons with a single cell. In the deep zone, stacking of chon-
drocytes into columnar chondrons is observed [17]. Shape, size
and organization (and variations in each) can influence the way
mechanical loading is transmitted to individual cells, and
therefore the cell’s mechanotransduction. The relative material
properties of chondrocytes, their pericellular environment and
the extracellular matrix are also important to resolve the mech-
anical load sharing problem. The pericellular matrix has a
significant role in the regulation of chondrocytes’ mechanical
environment and their mechanobiological response [41]. The
region is primarily of type VI collagen, which is exclusive
to this zone [41]. The extracellular matrix includes a zone-
dependent fibrous network of collagen of primarily type II
[2] (as described above).
A robust workflow for model development at the tissue–cell
scale can include parametrization of anatomical and mechanical
properties such that rapid and appropriate representations
specific to specimen, and tissue location and depth can be
constructed for prediction of chondrocyte mechanics in a
superficial
transitional
deep
cartilage
bone
PCM
chondrocyte
d
Nc
rcrp
Nv
pericellular matrix (PCM) material (E, v, k; x, b, –) extracellular matrix (EM) material (E, v, k; x, b, –)
chondrocyte material (E, v, k)wall boundary conditions
z
x
anatomical
discretization
mechanical
mesh density
region size (w × h × l)layer thicknesseschondrocyte density (Nv)chondrocyte locationchodrocyte orientationchondrocytes/chondron (Nc)chondrocyte radii (rc)
chondron spacing (d)PCM radii (rp)
w
EM
tissue zone of interest
parametric modelling
examples
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
6
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
high-throughput manner. In a recent study, our team developed
an automated model generation procedure to represent cellular
organization, shape and size for zone-specific tissue–cell scale
models [42] (figure 4). PYTHON scripting was used to interact
with SALOME to automate geometry and mesh generation for a
rectangular volume with a distribution of spherical/ellipsoidal
inclusions. This geometrical model is a representative volume
for cellular-level description of the tissue. In a human-readable
file, the user can specify the tissue arrangement (a variable
rectangular volume, with single or multiple stacked layers),
cell shape and size (spherical or ellipsoidal with variable
radii), pericellular layer shape and size (spherical or ellipsoidal
with variable radii), cell arrangement (location and orientation,
clustering of multiple cells within a chondron) and mesh den-
sity. The geometrical properties can be specified explicitly, or
by providing cellular geometry and distribution statistics,
e.g. mean and standard deviations for cell radii, number of
cells per unit volume, and pericellular thickness. The execution
of PYTHON scripts provides discretized (tetrahedral) model
geometry in a human-readable form, including appropriate
element, node and surface sets to subsequently define loading
and boundary conditions. Through additional configuration
parameters, users can assign material properties for the extra-
cellular matrix in a similar fashion as was done for the
cartilage layer properties described for the joint–tissue scale
models. In addition, direction-dependent material properties
of the pericellular matrix can be prescribed, i.e. the region’s
fibre organization aligned tangentially to the chondrocyte sur-
face [43] as a function of element centroid location relative to
the chondrocyte. Finally, the loading and boundary conditions
can be prescribed using deformation gradient and fluid
pressure information (for biphasic case), e.g. from the results
of joint–tissue scale simulations. PYTHON scripting manages
the assembly of this heterogeneous information to generate
a human-readable model input file that can be used with
FEBIO to conduct FE analysis. Simulations can quantify the
mechanical response of individual or clusters of chondrocytes
for a given anatomical and mechanical property set and for
preferred loading and boundary conditions. Similar to the
pipeline provided for the joint–tissue scale modelling, this
scripted environment allows generation of a plethora of
models with varying characteristics individualized for the car-
tilage layer of interest (figure 4), with minimal user interaction
and through the automation of the geometry, mesh and model
definition processes.
Figure 4. Based on statistical descriptors of cellular anatomy and organization,model generation can be automated at the spatial scale of chondrocytes. Zonalcartilage features and chondrocyte properties influence the mechanics of cells.Chondrocyte size, shape and distribution vary largely between individual layersof the cartilage. These properties are also dependent on the location of the car-tilage within the joint and across joints. Cell size, shape and density can beprescribed in a statistical manner to generate a representative volume for adesired zone of interest. Definitions of parameters representing pericellularmatrix anatomy and material properties for chondrocytes, and pericellular andextracellular matrices complete the computational representation of the cartilageat the tissue – cell scale. Tissue-level mechanics information can be mapped onthe boundaries of the representative volume as loading and boundary conditionsto carry out simulations. This streamlined model generation approach provides theopportunity to build models in an expedited fashion. (Online version in colour.)
3.2. Scale couplingMultiscale analysis requires the transfer of mechanical infor-
mation from joint–tissue scale simulations (where cartilage
mechanics is quantified) to tissue–cell scale models (where
chondrocyte mechanics can be predicted). This data exchange
requires a multitude of assumptions. A fundamental assump-
tion is the required consistency of the tissue mechanical
behaviour between the joint–tissue and tissue–cell scales. At
the higher spatial scale, the cartilage may be represented by
homogenized material properties [24]. This constitutive
response should be equivalent to the average response of the
tissue–cell scale model, which represents the cartilage in more
detail, including the delineation of chondrocytes, pericellular
matrix and the extracellular matrix. Multiple strategies exist to
enforce this constraint. For example, the homogenized tissue
properties (of the joint–tissue scale model) and the properties
1.0
0.1 s 100 s0
2 × 10–5 2 × 10–3
0.4
100 × 100 × 100 µm
t = 5.0 s
chondrons
deformation time history
effective strain
fluid flux magnitude (mm s–1)
control casefirst-order
approximationsecond-order
approximation
control case
joint–tissue scale model
tissue–cell scale modelfirst-order
approximationsecond-order
approximation
deformation
cartilage
impermeablerigid wall
point ofinterest
Figure 5. Predictions of chondrocyte mechanics can be influenced by the assumptions of translating mechanical information obtained from joint – tissue scalemodels to drive models with cellular inclusions. A biphasic cartilage model with homogeneous material properties was used to simulate mechanics of thetissue under non-uniform loading and asymmetric permeability conditions [23]. Information on solid and fluid phase responses at the point of interest wasused to drive a tissue – cell scale model including 11 spherical chondrocytes with each surrounded by a pericellular matrix. A first-order approximation used defor-mation gradient, and fluid pressure and its gradient to prescribe boundary conditions. A second-order approximation used additional terms, including thedeformation and fluid pressure Laplacians. A control case was also simulated to obtain a reference solution. For this purpose, a region with the same cellularinclusions was explicitly represented in the joint – tissue scale model and macroscopic and cell scale solutions were carried out simultaneously. Reduced orderof data passing resulted in the loss of predictive capacity, both for mechanical metrics descriptive of the solid phase (effective strain) and the fluid phase(fluid flux magnitude). Additional quantitative details on the impact of data passing assumption during the prediction of chondrocyte mechanics can befound in Sibole et al. [23]. (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
7
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
of the extracellular matrix (of the tissue–cell scale model) can be
treated as the same. When the volume fraction of the extracellu-
lar matrix is large, the contribution of the pericellular matrix and
chondrocytes to the tissue scale response can be neglected. We
have used this assumption in our previous studies where
94.2% of the tissue–cell scale model was the extracellular
matrix (for an 11 cell model of the transitional zone of cartilage)
[23,25]. Alternatively, homogenization based on spherical
inclusions [44] can be adapted to estimate joint–tissue scale
macroscopic properties of the cartilage as an analytical function
of tissue–cell scale constituent properties and their volume frac-
tion. Finally, computational homogenization strategies can be
employed [45], where average mechanical response of a repre-
sentative volume element (in this case, the tissue–cell scale
model) can be calculated from simulations of uniform loading
cases, and a surrogate macroscopic constitutive law can be
fitted to determine homogenized material properties for the
joint–tissue scale model. This approach may be beneficial for
complicated and dense cellular arrangements [17] and for
explicit fibre-based modelling of the tissue matrix microstructu-
rally [31]. The gross impact of tissue heterogeneity and the
contributions of cells and fibres on tissue behaviour can be
incorporated within the continuum mechanics framework.
Another assumption when transferring information from
the higher spatial scale to the lower one is related to the level
of approximation of loading and deformation metrics that
need to be provided as boundary conditions for the tissue–
cell scale model. One option is to overlay the boundary of the
tissue–cell scale model within the joint–tissue scale model
geometry and linearly interpolate the solution of the macro-
scopic scale as displacement (and fluid pressure if necessary)
boundary conditions [46,47]. While this approach does not
require any further assumptions for the required detail of
data exchange, it may be cumbersome to do so, particularly
for large numbers of cellular-level analyses. Alternatively, one
can use loading and deformation information from a given
point within the cartilage of the joint–tissue scale model to
quantify chondrocyte deformations [25]. If the size of the
tissue–cell scale model is negligible when compared with
the characteristic size of the corresponding geometry in the
joint–tissue scale model, a first-order approximation may be
sufficient, i.e. using the deformation gradient at a given macro-
scale point to set boundary conditions of the cell scale model.
Unfortunately, for cartilage and chondrocytes, decoupling of
length scales can be problematic, potentially requiring a
second-order approximation for data exchange [45], i.e. using
the deformation gradient, which is a measure of local defor-
mations (shape change) at a given point, and deformation
Laplacian, which provides the spatial derivatives of the
deformation gradient. In a recent study, our team evaluated
the role of data passing assumptions in the characterization of
biphasic chondrocyte mechanics [23]. A 1 � 1 � 1 mm homo-
geneous and isotropic cartilage tissue underwent non-
uniform compression (imitating a curved indenter, 0–10%
nominal strain across the boundary; figure 5). Along the loading
axis, asymmetric permeability conditions were prescribed
(permeable at the indentation site, impermeable at the fixed
bottom); lateral walls of the model were confined and imper-
meable. Full loading was applied in 0.1 s with a relaxation
duration up to 100 s. The biphasic response of the model was
obtained using FEBIO. Loading and deformation informa-
tion at the centre point of this model was extracted to drive a
1.22
1.2238
1.24
number of cells
chondrocyte shape factor at steady state
2 1536
1189
47
10
5 14
131
12increasing representative volume
1 5 10 15 19
Figure 6. The tissue – cell scale model’s representative volume size may havean influence on the predictions of chondrocyte mechanics. Chondrocytes ofknee cartilage’s transitional zone were modelled by varying the representativevolume size. The cell density was kept constant at reported levels. The timehistory of cartilage mechanics was mapped on the representative volumesfrom the solution of a homogeneous biphasic cartilage model subjected tonon-uniform loading and asymmetric permeability conditions. A second-order data passing scheme was used. Chondrocyte shape factors wereobtained from biphasic simulations at steady state. The results indicatedthe role of representative volume size for consistent prediction of chondro-cytes’ mechanical environment. The solution for the 11-cell model ishighlighted. The edge length for this representative volume corresponds tothe commonly used 100 mm. While the illustrated differences in cellshape factor may not have significance in relation to other errors associatedwith modelling assumptions, the use of other cell scale biomechanical metricsand the exploration of different zones of cartilage or different types of tissueswarrant execution of similar analyses. (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
8
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
tissue–cell scale model. The lower scale model was a 100 mm �100 mm � 100 mm representative volume with spherical cellu-
lar inclusions representative of the transitional layer of the
cartilage [48] (11 cells within the volume; chondrocyte radius
of 5 mm and pericellular thickness of 2.5 mm; figure 5). Chon-
drocyte, pericellular matrix and extracellular matrix material
properties were homogeneous, isotropic and biphasic. Two
types of data passing schemes were employed to test the
required fidelity when prescribing tissue–cell scale boundary
conditions. First, a first-order Taylor series expansion was
used to map macroscale deformation gradient, and fluid
pressure and its gradient to the boundaries of the tissue–cell
scale model. Then, a second-order Taylor series expansion,
with additional terms containing the deformation and fluid
pressure Laplacians, was used. Results of both these cases
were compared against the predictions of a control model,
where the joint–tissue scale model had a region at its centre
explicitly incorporating the chondrocyte and pericellular
matrix representation exactly as in the separate tissue–cell
scale model. While the results were not necessarily compared
against the behaviour of actual cells, this investigation revealed
that data passing assumptions may lead to additional misre-
presentation of the chondrocytes’ mechanical environment, at
least for the case presented here (figure 5). It is advised to use
a higher-order data passing scheme or one that relies on interp-
olation (albeit with the increased difficulty for high-throughput
analysis). It should also be noted that for each additional term
included in a Taylor series approach, the order of the required
tensor increases by one. In three dimensions, this will contribute
3nþ 1 additional coefficients to be calculated for both the solid
and fluid phases, where n is the order of the series.
Another challenge for cell scale analysis has been the selec-
tion of an appropriate volume for the tissue–cell scale model.
Ideally, the representative volume should be small to allow
the scale separation assumption in that the characteristic
lengths of higher and lower scales differ by orders of magni-
tude. The representative volume should also be large enough
to include characteristic geometrical features of the region’s cel-
lular organization. Traditionally, computational simulations of
chondrocytes used models with an edge length of 100 mm [49],
originated from the study by Guilak & Mow [46]. This model
and its successors placed a single chondron at the centre of
the modelling volume, commonly overlooking the actual cell
density of the cartilage zone. One has multiple options when
building models which keep the volume ratio of the chondro-
cytes and extracellular matrix the same. Cartilage can be
represented at the tissue–cell scale using a single cell model
with a small overall model volume or by placing multiple
cells in larger volumes. Ideally, all these models should result
in similar chondrocyte mechanics predictions. Yet, with the
increased size of the representative volume, mechanical
interactions between the cells can be investigated and the influ-
ence of a cell’s location within the neighbourhood of the point
of interest can be explored. Our ongoing work explored the
influence of representative volume size on the prediction of
chondrocyte loading and deformation [50]. Tissue–cell scale
models representative of chondrocyte density of the cartilage’s
transitional zone were developed, starting with representative
volumes containing a single cell, through 20 cells (figure 6).
The model edge length varied between 30 and 120 mm to
keep cell density constant. The cells were randomly distributed
within the representative volume maintaining a prescribed
minimum distance between cells and the volume boundaries.
For each level of cellular representation, multiple models
were generated and solved providing mechanics information
for at least 130 chondrocytes in total. Chondrocytes were
spherical, each with a radius of 5 mm and a pericellular thick-
ness of 2.5 mm. Biphasic simulations were conducted with
loading and boundary conditions reflecting the deformation
gradient and fluid pressure (and their derivatives based on a
second-order data passing scheme) obtained from macro-
level simulations. The macroscale joint–tissue model was
loaded by a non-uniform surface displacement with asym-
metric fluid behaviour (described above) [23]. The average
predicted cell metrics (and their variability) became more con-
sistent as the representative volume size included larger
number of cells, e.g. for chondrocyte shape factor (defined in
[51]; figure 6). Differences can be attributed to the variability
of cell location within the representative volume and/or cell-
to-cell mechanical interactions. Simulations were completed
approximately 10 times faster at smaller representative
volumes with lower cell numbers. However, for this specific
illustrative case, distribution of cell mechanical metrics was
better captured as the volume of interest increased—possibly
providing location-dependent variability of cell mechanics
and cell-to-cell mechanical interactions.
3.3. ComputingSimulations both at the joint–tissue scale and the tissue–cell
scale can be inherently computationally intensive. Geometri-
cal and material nonlinearity, consideration of multiple
joint–tissue scaleeffective strain
0.2
0
parametric analysis at joint–tissue scale
M conditions
N conditions
M × N interactions
P points in tissue space
effective strain (cartilage)
effe
ctiv
e st
rain
(ch
ondr
ocyt
e)
0.4
00 0.2
+ + +...
+ + +...
+ + +...
+ + +...
+ + +...
parametric analysis at tissue–cell scale
parametric analysis at both scales
feed-forward multiscale analysis
Figure 7. Feed-forward modelling approach decouples joint – tissue andtissue – cell scale simulations and facilitates adaptation of various computingstrategies in order to expedite solutions. Probabilistic analysis, sensitivity studiesand uncertainty quantification can be conducted through the parallelization ofparametric simulations at individual scales or both (top three panels). Simu-lations can also be performed in parallel when one post-processes joint –tissue scale results to obtain chondrocyte response at a large number ofpoints within the cartilage (bottom panel). For example, FE analysis of theknee joint quantified femoral and tibial cartilage deformations under a compres-sive load of one body weight [25]. For each element centroid, a nonlinearlyelastic model of a tissue – cell scale model was solved. Utilization of parallel pro-cessing in a high-performance computing resource reduced the solution time foralmost 8000 models to less than a day, which would otherwise take more than70 days when using a single computer. (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
9
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
material phases, and large mesh sizes may all contribute to
the added cost of conducting simulations at the single spatial
scale. Multiscale analysis further complicates this issue as
multiples of models need to be solved for quantification
of cartilage and chondrocyte mechanics. Nonetheless, the
post-processing-based feed-forward approach to multiscale
simulations provides various opportunities to tackle the
computing burden. From a user’s perspective, multiple simu-
lation strategies may exist (figure 7): parametric studies at the
joint–tissue scale, parametric studies at the tissue–cell scale,
a combination of both, and lastly large-scale simulations at
the tissue–cell scale for characterization of chondrocyte
mechanical environment for a given solution of a joint–
tissue scale model. In the first use case, one may be interested
in the cartilage mechanics for a set of loading and boundary
conditions or in the sensitivity of the cartilage mechanical
environment to variations in anatomical and mechanical
properties. Similarly, as an example of the second use case,
one may be interested in chondrocyte mechanics as a function
of macroscale deformations or owing to variations in ana-
tomical and mechanical properties of the chondrocytes,
pericellular matrix and extracellular matrix. These analyses
can be conducted in conjunction to evaluate the changes in
multiscale load sharing path as model parameters at all
scales are perturbed simultaneously (third use case). All
these use cases can lead into probabilistic analysis [20],
i.e. for assessment of population variations, sensitivity studies
and uncertainty quantification. The final use case illustrates
the multiscale feed-forward analysis approach. First, a
joint–tissue scale model is solved. At each point of interest
in that model, e.g. all locations across all zones of the carti-
lage, tissue–cell scale models are simulated to map
chondrocyte mechanical metrics on the topology of the
macroscale tissue anatomy. This may require the solution
of hundreds to thousands of tissue–cell scale models, all
driven by the results of a single joint–tissue scale model. In
all the aforementioned cases, simulations are loosely coupled,
i.e. minimal or no communication is required between
models of the same scale. Massively scalable parallel comput-
ing strategies can be employed by distributing simulation
workload to independent computing nodes in order to expe-
dite the whole solution process. Albeit various computational
strategies, case-specific model simplifications may still be
necessary to balance the requirements of a high-throughput
simulation workflow and scientific accuracy.
In a previous study, an approach to tackle large-scale simu-
lations through the use of loosely coupled computing strategies
was presented for the prediction of chondrocyte deformations
in tibial and femoral cartilage [25]. First, a tibiofemoral joint
model was solved to obtain tissue-level deformation metrics
for the cartilage during the application of a compressive load
with a magnitude of one body weight. This model incorpor-
ated the anatomical detail of the joint articulation and the
passive tissue structures, e.g. ligaments, and a homogeni-
zed representation of tissue material properties, i.e. apparent
properties at the macroscale. Nonlinearly elastic FE analysis
provided deformation gradients at each element centroid
within the transitional layers of femoral and tibial cartilage.
For each element centroid, a tissue–cell scale model was
solved to obtain a joint-level distribution of chondrocyte
deformations within femoral and tibial cartilage. The tissue–
cell scale model was a 100 � 100 � 100 mm representative
volume with spherical cellular inclusions and homogeneous
and isotropic elastic properties for chondrocytes, pericellular
matrix and extracellular matrix. The cellular organization rep-
resented the transitional zone [48]: 11 cells within the volume;
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
10
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
chondrocyte radius of 5 mm and pericellular thickness of
2.5 mm. The boundary conditions of the model reflected defor-
mations based on a first-order data passing scheme. This
analysis required the solution of 7822 tissue–cell scale
models (same model simulated with different boundary con-
ditions extracted from 7822 locations in tibial and femoral
cartilage). Exploiting the parallel nature of feed-forward
analysis, these models were solved in a high-performance com-
puting environment (multiple compute nodes using 101
threads on Glenn Cluster of the Ohio Supercomputer Centre),
using FEBIO for FE analysis. At the time, Glenn Cluster pro-
vided up to 9572 compute cores, ranging in frequencies of
2.4–2.6 GHz, offering a peak performance of more than 75 ter-
aflops. Through parallel processing, solution of 7822 models
resulted in a wall-clock time of 19 h. If using a single CPU,
estimated wall-clock time would be 1735.1 h (72.3 days), indi-
cating a prohibitive computing burden. A study of scalability
was not performed, but because each tissue-scale model is
entirely independent, the limiting factor should only be the
initial distribution of problems to computer nodes, and like-
wise, the pooling of results after simulations have completed.
Some network overhead is expected but it is ideally low, and
thus scaling to many thousands of nodes is highly feasible.
The feed-forward analysis provided the possibility for this
loosely coupled parallel processing. In turn, timely prediction
of chondrocyte mechanics for large geometrical regions of
femoral and tibial cartilage was attainable (figure 7).
For cartilage and chondrocyte mechanics, nonlinear simu-
lations incorporating the biphasic or multiphasic nature of
the tissue and cell behaviour may also increase the compu-
tational cost. It should be noted that FE analysis software
provides capabilities for multithreading (on a single comput-
ing node) [52–54] and parallel processing (on a number of
distributed computing nodes) [53,54], during the generation
of system stiffness matrices and/or the solution of linearized
system equations. The use of these built-in computing strat-
egies may expedite single simulation of a given joint–tissue
or tissue–cell scale model. FEBIO, the FE analysis package
used in our modelling and simulation workflow, provides
multithreading on shared memory computing architectures
[52]. By relying on PARDISO, a shared-memory multiproces-
sing parallel direct sparse solver for linear systems of
equations [55], solution times can be reduced on a given
compute node equipped with multiple processors.
3.4. Post-processingFE analyses, both at the joint–tissue and tissue–cell scales,
result in a large set of variables to evaluate the mechanical
environment of the cartilage and chondrocytes. However, it
may be challenging to interpret the raw simulation results
without post-processing, where desirable biomechanical
metrics within the context of research question or clinical pur-
pose can be calculated. Similarly, movement of information
from joint–tissue simulations down to tissue–cell scale
models requires preparation of the raw simulation results of
the higher spatial scale in a form amenable to use in the
models of the lower spatial scale. In this section, post-proces-
sing solutions relying on PYTHON scripting are provided for
simulations conducted using freely accessible FEBIO software.
At the spatial scale of joint–tissue, interpretation of FE
analysis results, e.g. for individual zones of the cartilage, is
possible by using principal stress and strain (solid phase)
and fluid flux and pressure (fluid phase) distributions,
which are common for the evaluation of biphasic mechanics.
For multiscale analysis, more detailed information about the
mechanical loading and deformation is needed. One may
want to solve a tissue–cell scale model associated with a
point in the macroscale model, e.g. at an element centroid.
The complete mechanical state of the point of interest needs
to be mapped to the tissue–cell scale model as boundary con-
ditions. For biphasic simulations for example, the time
histories of the deformation gradient (and possibly the defor-
mation Laplacian) and the fluid pressure (and its derivatives)
will be necessary for data passing (also see discussion on
scale coupling above). Such information is not readily avail-
able in FE analysis results and it needs to be derived from
nodal displacement and fluid pressure information. To facili-
tate this procedure, PYTHON scripts were developed to parse
through FEBIO output files to extract raw nodal and element
results for desired sets, e.g. elements for which multiscale
analysis would be performed [23]. Using the FE isopara-
metric formulation, the nodal displacements and fluid
pressures can be used to compute time histories of defor-
mation gradient, fluid pressure and their spatial derivatives
for desired points of the joint–tissue scale models. This infor-
mation can subsequently be provided to tissue–cell scale
models in a streamlined human-readable form. For illus-
tration of joint–tissue scale post-processing, the solution of
a 1 � 1 � 1 mm homogeneous and isotropic cartilage tissue
was explored (figure 8). A summary of the model and simu-
lation conditions is provided here, and further details are
available elsewhere [23]. The cartilage model underwent
non-uniform compression, essentially representing the influ-
ence of a curved indenter (0–10% nominal strain across the
boundary). The biphasic response was sought, with asym-
metric permeability conditions along the loading axis
(permeable at the indentation site, impermeable at the fixed
bottom), and with confined and impermeable walls. Full
loading was applied in 0.1 s with a relaxation duration up
to 100 s. The time histories of compressive strain and fluid
pressure for the centroid of the element at the centre of the
tissue are shown in figure 8.
At the spatial scale of tissue–cell, interpretation of FE
analysis results, e.g. for the chondrocytes, pericellular
matrix and extracellular matrix, is also possible through the
evaluation of principal stress and strain (solid phase) and
fluid flux and pressure (fluid phase) distributions. This infor-
mation can provide an understanding of nearly focal stresses
and strains on chondrocyte surfaces and within, which may
indicate potential failure and mechanical signalling mechan-
isms. Yet, an understanding of multi-modal loading and
overall deformations of individual cells and/or populations
of cells may be necessary. Metrics such as cell height and
width, aspect ratio and volume (undeformed and deformed)
have commonly been used for describing experimental data
[11] and simulation results [56], with the intention to associ-
ate cell mechanics with cell damage and mechanical stimuli.
Such descriptors, however, are not readily available in raw
simulation results, and automated procedures are necessary
if these variables need to be evaluated for hundreds of cells
and for a large number of simulation conditions. Motivated
by this need, PYTHON scripts were developed to parse FEBIO
output files to extract raw results for nodes, elements and sur-
faces representing each chondrocyte [23,51]. The analysis
scripts are capable of calculating the volume of each cell,
100
0
1.8
chondrons
00 20 40 60 80 100
volu
me
aver
aged
eff
ectiv
ech
ondr
ocyt
e st
rain
chon
droc
yte
mas
s ex
chan
gera
te (
pgs–1
)
time (s)
0 20 40 60 80 100time (s)
100 × 100 × 100 µm
tissue–cell scale model
Figure 9. Post-processing of tissue – cell scale simulations can provide descrip-tors of chondrocytes’ multiphase mechanical behaviour in a usable form. Suchinformation is not readily available in raw results. In this case, time histories ofeffective chondrocyte strain and mass exchange rate are shown for a total of11 cells. The tissue – cell scale model is representative of the transitionalzone of the knee cartilage. Cartilage mechanics was mapped on this represen-tative volume from the solution of a homogeneous biphasic cartilage modelsubjected to non-uniform loading and asymmetric permeability conditions.A second-order data passing scheme was used. (Online version in colour.)
1.0
0.1 s 100 s
12
0
deformation time historydeformation
cartilage
impermeablerigid wall
0
–0.14
point ofinterest
joint–tissue scale model
time (s)0 20 40 60 80 100
time (s)0 20 40 60 80 100
stra
in a
long
load
ing
dire
ctio
nfl
uid
pres
sure
(M
Pa)
Figure 8. If one is interested in multiphasic mechanics of cartilage at thejoint – tissue scale, post-processing of simulation results can providecommon descriptors of the tissue’s mechanical environment at any desiredlocation. In this case, time histories of compressive strain and fluid pressurewere extracted for a point at the centre of a cartilage region (1 � 1 �1 mm). A non-uniform displacement profile was prescribed at the top.Impermeable walls confined the tissue. (Online version in colour.)
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
11
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
and using volume-weighted averaging, the effective stress
and strain, and maximum shear strain. The net mass
exchange rate for each chondrocyte is also calculated by the
surface integral of fluid flux. This metric, obtained from
biphasic analysis, can be an indicator of cell-to-cell signalling.
More common cellular deformation metrics are calculated by
first finding the mass moment inertia of each cell and then
finding the eigenvalues of this tensor, which represent the
principal moments of inertia [57]. These eigenvalues are pro-
cessed to calculate axis lengths of the inertial ellipsoid, which
in turn provide a shape factor from their ratio [51]. Further
processing provides chondrocyte height, width and depth,
and therefore aspect ratio [51]. From the results of biphasic
FE analysis, time histories of all these cell scale metrics are
available in a human-readable form. Post-processing capabili-
ties at the tissue–cell scale were illustrated using the results of
a 100 � 100 � 100 mm representative volume with spherical
cellular inclusions and homogeneous and isotropic biphasic
properties for chondrocytes, pericellular matrices and
extracellular matrix (figure 9). The cellular organization
represented the transitional zone [48]: 11 cells within the
volume; chondrocyte radius of 5 mm and pericellular thickness
of 2.5 mm. The boundary conditions of the model reflected
deformation gradient and fluid pressure (and their derivatives
based on a second-order data passing scheme) at the centre of a
joint–tissue scale model described previously in this section.
Extracellular matrix properties were kept the same as the
macroscopic properties of the joint–tissue scale. Time histories
of various biomechanical loading and deformation metrics for
each chondrocyte are provided in figure 9.
4. Concluding remarksThe vision and initial realization of a high-throughput mod-
elling and simulation workflow for large-scale analysis of
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
12
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
cartilage and chondrocyte mechanics was provided in this
document. Various strategies to realize individual com-
ponents of the modelling framework were implemented
with a specific focus on the automation of developing
models at the joint–tissue and tissue–cell scales that are
capable of incorporating subject-specific and regional vari-
ations in anatomical and mechanical parameters. Some of
the technical challenges associated with multiscale analysis
were addressed. In separate studies, the requirement for
higher-order data passing from the joint–tissue scale down
to the tissue–cell scale was emphasized, and the appropriate-
ness of the selection of the tissue–cell scale model size was
confirmed. From a usability perspective, the need for stream-
lined extraction of important biomechanical metrics, for the
evaluation of cartilage and chondrocyte mechanics and for
feed-forward movement of the data through the spatial
scales was met. Computing strategies, in particular the use
of parallel processing, were presented as a means to exploit
the loosely coupled nature of post-processing-based multi-
scale analysis to conduct thousands of simulations in an
expedited fashion.
The outlined modelling and simulation approach focused
on a feed-forward analysis strategy relying on the sequential
solution and post-processing of computational models at the
joint–tissue scale and then at the tissue–cell scale. This
approach was found to be suitable for solving the load shar-
ing problem starting from the higher spatial scales, and
determining the loads and deformations at various cartilage
layers and at the level of individual chondrocytes. A signifi-
cant advantage of this feed-forward approach is related to
the reduced demand on computing as macro- and microscale
models need to be solved only once for a desired loading con-
dition at preferred locations. This advantage fundamentally
relies on the assumption that tissue representations at multiple
scales are mechanically consistent, i.e. they provide the same
overall constitutive behaviour. It may not provide the struc-
ture–function relationships, which can be accomplished
through a feedback loop, i.e. by computational homogeniz-
ation that relies on iterative solutions of lower scale models
during the solution process of the model at the higher spatial
scale [45]. It should be noted that such a feedback loop is not
necessary to obtain the mechanical environment of cartilage
and chondrocytes for a given tissue state. It will be necessary
when constitutive properties change as a result of tissue
growth and damage [31]. Nonetheless, the tools presented
in this study can be adapted for such purposes. It is worth
mentioning that this study did not provide an understanding
of error propagation when one uses the proposed multiscale
modelling workflow. Owing to the feed-forward nature of
the simulations, it is anticipated that the errors will accumu-
late in an additive manner as one moves from the higher
spatial scale of the joint down to the scale of chondrocytes.
The components of the modelling and simulation workflow
were prototyped using PYTHON scripting, for pre-processing
of the models using user-specified parameters and for post-
processing of the simulation results. The integrated and
scriptable computer-aided design package SALOME was used
in an unsupervised manner for generation of models. Likewise,
FEBIO, FE analysis software for biomechanics [34], was used to
conduct simulations. These readily available, open source and
freely accessible pieces of software provided the means for
quick prototyping. Nonetheless, it is possible to integrate
model development, simulation and post-processing tools
into a turn-key modelling and simulation platform. Some desir-
able features for the expansion of this framework are noted as
follows. Geometrical model generation procedures can use
explicitly defined individual anatomy, e.g. from image segmen-
tation, for tissue representation at the macroscale, and for cell
representation at the microscale. For models at the joint–
tissue scale, the representation of additional components, e.g.
meniscus, will be necessary when the desired contact region
of interest becomes wider. For models at the tissue–cell scale,
incorporation of cell membrane may be needed to investigate
the influence this structure’s permeability on the mechanical
behaviour of the chondrocyte. Robust physics formulations pro-
vided in FEBIO [32] can also allow the modelling and simulation
workflow to include multiphasic behaviour of the cartilage and
chondrocytes beyond biphasic analysis, to incorporate the mech-
anical effects of charged particles and proteoglycans, namely
osmotic pressure gradients and swelling.
The capacity to explore cartilage mechanics for any joint
of interest, at any region of contact within the joint, for any
zone within the tissue and under different loading condi-
tions when supported by high-performance computing can
empower large-scale analysis of cartilage and chondrocyte
function in health and disease. Population responses can be
quantified based on probabilistic analysis and in-depth sensi-
tivity studies can be performed to understand the role of
biomechanical markers in cartilage function. It is anticipated
that this framework will be used to understand cartilage
and chondrocyte mechanics during ageing. Ageing mani-
fests anatomical and mechanical changes at multiple scales
[13,16,19,58], which in turn perturb the mechanical envi-
ronment and therefore associated biological responses to
mechanical stimuli. Similarly, this modelling and simulation
workflow will likely be a significant toolset to evaluate the
complicated mechanical state of the tissue and cells during
the onset and progression of osteoarthritis.
5. Access to data, models and scriptsVarious scripts, models, data and simulation results can be
found at the project site, https://simtk.org/home/j2c, in
the downloads section.
Acknowledgements. This study was conducted at the Cleveland Clinicand was previously presented at the Seventh World Congress of Bio-mechanics, 6–11 July 2014, Boston, MA, USA [59]. The documentprovides a synthesis of our group’s published and unpublishedwork on in silico mechanics of cartilage and chondrocytes. Theauthors acknowledge Simbios, NIH National Center for BiomedicalComputing at Stanford University, http://simbios.stanford.edu/(project hosting); Ohio Super Computer Center, https://www.osc.edu/ and Extreme Science and Engineering Discovery Environment,https://www.xsede.org/ (high-performance computing); FEBIO,http://www.febio.org/ (FE analysis software) [34]; SALOME, http://www.salome-platform.org/ (integrated software for computer-aided design); PYTHON, https://www.python.org/ (scripting). Theauthors are also grateful for in-depth discussions with JasonHalloran and Snehal Chokhandre (Cleveland Clinic), Cees Oomensand Rene van Donkelaar (Eindhoven University of Technology),Jeff Weiss and Steve Maas (University of Utah), and Farshid Guilak(Duke University).
Funding statement. The research activities were supported by theNational Institutes of Health, in particular by the National Instituteof Biomedical Imaging and Bioengineering (R01EB009643; PrincipalInvestigator: A.E., https://simtk.org/home/j2c) and partially bythe National Institute of General Medical Sciences (R01GM104139;Principal Investigator: A.E., https://simtk.org/home/openknee).
13
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
References
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
1. Halloran JP, Sibole S, van Donkelaar CC, vanTurnhout MC, Oomens CWJ, Weiss JA, Guilak F,Erdemir A. 2012 Multiscale mechanics of articularcartilage: potentials and challenges of couplingmusculoskeletal, joint, and microscalecomputational models. Ann. Biomed. Eng. 40,2456 – 2474. (doi:10.1007/s10439-012-0598-0)
2. Heinegard D, Oldberg A. 1989 Structure and biologyof cartilage and bone matrix noncollagenousmacromolecules. FASEB J. 3, 2042 – 2051.
3. Michalek AJ, Iatridis JC. 2007 A numerical study todetermine pericellular matrix modulus and evaluateits effects on the micromechanical environment ofchondrocytes. J. Biomech. 40, 1405 – 1409. (doi:10.1016/j.jbiomech.2006.05.025)
4. Torzilli PA, Deng X-H, Ramcharan M. 2006 Effect ofcompressive strain on cell viability in staticallyloaded articular cartilage. Biomech. Model.Mechanobiol. 5, 123 – 132. (doi:10.1007/s10237-006-0030-5)
5. Grodzinsky AJ, Levenston ME, Jin M, Frank EH.2000 Cartilage tissue remodeling in response tomechanical forces. Annu. Rev. Biomed. Eng. 2,691 – 713. (doi:10.1146/annurev.bioeng.2.1.691)
6. Kurz B, Lemke AK, Fay J, Pufe T, Grodzinsky AJ,Schunke M. 2005 Pathomechanisms of cartilagedestruction by mechanical injury. Ann. Anat. 187,473 – 485. (doi:10.1016/j.aanat.2005.07.003)
7. O’Conor CJ, Case N, Guilak F. 2013 Mechanicalregulation of chondrogenesis. Stem Cell Res. Ther. 4,61. (doi:10.1186/scrt210)
8. Martin JA, Brown TD, Heiner AD, Buckwalter JA.2004 Chondrocyte senescence, joint loading andosteoarthritis. Clin. Orthop. S96 – S103. (doi:10.1097/01.blo.0000143818.74887.b1)
9. Lawrence RC et al. 2008 Estimates of the prevalenceof arthritis and other rheumatic conditions in theUnited States. II. Arthritis Rheum. 58, 26 – 35.(doi:10.1002/art.23176)
10. Han S-K, Colarusso P, Herzog W. 2009 Confocalmicroscopy indentation system for studying in situchondrocyte mechanics. Med. Eng. Phys. 31, 1038 –1042. (doi:10.1016/j.medengphy.2009.05.013)
11. Abusara Z, Seerattan R, Leumann A, Thompson R,Herzog W. 2011 A novel method for determiningarticular cartilage chondrocyte mechanics in vivo.J. Biomech. 44, 930 – 934. (doi:10.1016/j.jbiomech.2010.11.031)
12. Moo EK, Han SK, Federico S, Sibole SC, Jinha A, AbuOsman NA, Pingguan-Murphy B, Herzog W. 2014Extracellular matrix integrity affects the mechanicalbehaviour of in situ chondrocytes undercompression. J. Biomech. 47, 1004 – 1013. (doi:10.1016/j.jbiomech.2014.01.003)
13. Hudelmaier M, Glaser C, Hohe J, Englmeier KH,Reiser M, Putz R, Eckstein F. 2001 Age-relatedchanges in the morphology and deformationalbehavior of knee joint cartilage. Arthritis Rheum.44, 2556 – 2561. (doi:10.1002/1529-0131(200111)44:11,2556::AID-ART436.3.0.CO;2-U)
14. McLeod MA, Wilusz RE, Guilak F. 2013 Depth-dependent anisotropy of the micromechanicalproperties of the extracellular and pericellularmatrices of articular cartilage evaluated via atomicforce microscopy. J. Biomech. 46, 586 – 592. (doi:10.1016/j.jbiomech.2012.09.003)
15. Deneweth JM, Newman KE, Sylvia SM, McLean SG,Arruda EM. 2013 Heterogeneity of tibial plateaucartilage in response to a physiological compressivestrain rate. J. Orthop. Res. 31, 370 – 375. (doi:10.1002/jor.22226)
16. Temple MM, Bae WC, Chen MQ, Lotz M, Amiel D,Coutts RD, Sah RL. 2007 Age- and site-associatedbiomechanical weakening of human articularcartilage of the femoral condyle. OsteoarthritisCartilage 15, 1042 – 1052. (doi:10.1016/j.joca.2007.03.005)
17. Quinn TM, Hunziker EB, Hauselmann H-J. 2005Variation of cell and matrix morphologies inarticular cartilage among locations in the adulthuman knee. Osteoarthritis Cartilage 13, 672 – 678.(doi:10.1016/j.joca.2005.04.011)
18. Quinn TM, Hauselmann H-J, Shintani N, Hunziker EB.2013 Cell and matrix morphology in articular cartilagefrom adult human knee and ankle joints suggestsdepth-associated adaptations to biomechanical andanatomical roles. Osteoarthritis Cartilage 21,1904 – 1912. (doi:10.1016/j.joca.2013.09.011)
19. Steklov N, Srivastava A, Sung KLP, Chen PC, LotzMK, D’Lima DD. 2009 Aging-related differences inchondrocyte viscoelastic properties. Mol. Cell.Biomech. 6, 113 – 119. (doi:10.3970/mcb.2009.006.113)
20. Laz PJ, Browne M. 2010 A review of probabilisticanalysis in orthopaedic biomechanics. Proc. Inst.Mech. Eng. H 224, 927 – 943. (doi:10.1243/09544119JEIM739)
21. Halloran JP, Erdemir A. 2011 Adaptive surrogatemodeling for expedited estimation of nonlineartissue properties through inverse finite elementanalysis. Ann. Biomed. Eng. 39, 2388 – 2397.(doi:10.1007/s10439-011-0317-2)
22. Causin P, Sacco R, Verri M. 2013 A multiscaleapproach in the computational modeling of thebiophysical environment in artificial cartilage tissueregeneration. Biomech. Model. Mechanobiol. 12,763 – 780. (doi:10.1007/s10237-012-0440-5)
23. Sibole SC, Maas S, Halloran JP, Weiss JA, Erdemir A.2013 Evaluation of a post-processing approach formultiscale analysis of biphasic mechanics ofchondrocytes. Comput. Methods Biomech. Biomed.Eng. 16, 1112 – 1126. (doi:10.1080/10255842.2013.809711)
24. Henak CR, Anderson AE, Weiss JA. 2013 Subject-specific analysis of joint contact mechanics:application to the study of osteoarthritis andsurgical planning. J. Biomech. Eng. 135, 021003.(doi:10.1115/1.4023386)
25. Sibole SC, Erdemir A. 2012 Chondrocytedeformations as a function of tibiofemoral joint
loading predicted by a generalized high-throughputpipeline of multi-scale simulations. PLoS ONE 7,e37538. (doi:10.1371/journal.pone.0037538)
26. Rudert MJ, Ellis BJ, Henak CR, Stroud NJ, PedersonDR, Weiss JA, Brown TD. 2014 A new sensor formeasurement of dynamic contact stress in the hip.J. Biomech. Eng. 136, 035001. (doi:10.1115/1.4026103)
27. Liu F, Kozanek M, Hosseini A, Van de Velde SK, GillTJ, Rubash HE, Li G. 2010 In vivo tibiofemoralcartilage deformation during the stance phase ofgait. J. Biomech. 43, 658 – 665. (doi:10.1016/j.jbiomech.2009.10.028)
28. Erdemir A, McLean S, Herzog W, van den Bogert AJ.2007 Model-based estimation of muscle forcesexerted during movements. Clin. Biomech. 22,131 – 154. (doi:10.1016/j.clinbiomech.2006.09.005)
29. Besier TF, Gold GE, Beaupre GS, Delp SL. 2005 Amodeling framework to estimate patellofemoraljoint cartilage stress in vivo. Med. Sci. Sports Exerc.37, 1924 – 1930. (doi:10.1249/01.mss.0000176686.18683.64)
30. Dowling EP, Ronan W, McGarry JP. 2013Computational investigation of in situ chondrocytedeformation and actin cytoskeleton remodellingunder physiological loading. Acta Biomater. 9,5943 – 5955. (doi:10.1016/j.actbio.2012.12.021)
31. Hadi MF, Sander EA, Barocas VH. 2012 Multiscalemodel predicts tissue-level failure from collagenfiber-level damage. J. Biomech. Eng. 134, 091005.(doi:10.1115/1.4007097)
32. Ateshian GA, Maas S, Weiss JA. 2013 Multiphasicfinite element framework for modeling hydratedmixtures with multiple neutral and charged solutes.J. Biomech. Eng. 135, 111001. (doi:10.1115/1.4024823)
33. Wilson W, van Donkelaar CC, van Rietbergen R,Huiskes R. 2005 The role of computational modelsin the search for the mechanical behavior anddamage mechanisms of articular cartilage. Med.Eng. Phys. 27, 810 – 826. (doi:10.1016/j.medengphy.2005.03.004)
34. Maas SA, Ellis BJ, Ateshian GA, Weiss JA. 2012FEBio: finite elements for biomechanics. J. Biomech.Eng. 134, 011005. (doi:10.1115/1.4005694)
35. Li J, Stewart TD, Jin Z, Wilcox RK, Fisher J. 2013 Theinfluence of size, clearance, cartilage properties,thickness and hemiarthroplasty on the contactmechanics of the hip joint with biphasic layers.J. Biomech. 46, 1641 – 1647. (doi:10.1016/j.jbiomech.2013.04.009)
36. Chen SS, Falcovitz YH, Schneiderman R, MaroudasA, Sah RL. 2001 Depth-dependent compressiveproperties of normal aged human femoral headarticular cartilage: relationship to fixed chargedensity. Osteoarthritis Cartilage 9, 561 – 569.(doi:10.1053/joca.2001.0424)
37. Ateshian GA, Rajan V, Chahine NO, Canal CE, HungCT. 2009 Modeling the matrix of articular cartilageusing a continuous fiber angular distribution
rsfs.royalsocietypublishing.orgInterface
Focus5:20140081
14
on May 13, 2018http://rsfs.royalsocietypublishing.org/Downloaded from
predicts many observed phenomena. J. Biomech.Eng. 131, 061003. (doi:10.1115/1.3118773)
38. Reddy A, Erdemir A. 2014 A modeling andsimulation pipeline for the prediction of localizedcartilage mechanics: towards a multiscaleanalysis framework for joint, tissue, and cellscales. In 7th World Congress of Biomechanics,Boston, MA, USA, 6 – 11 July 2014.
39. Erdemir A, Reddy A. 2014 Parametric modeling forin silico explorations of cartilage and chondrocytemechanics. In 12th Int. Symp. on Computer Methodsin Biomechanics and Biomedical Engineering,Amsterdam, The Netherlands, 13 – 15 October 2014.
40. Schumacher BL, Su J-L, Lindley KM, Kuettner KE,Cole AA. 2002 Horizontally oriented clusters ofmultiple chondrons in the superficial zone of ankle,but not knee articular cartilage. Anat. Rec. 266,241 – 248. (doi:10.1002/ar.10063)
41. Wilusz RE, Sanchez-Adams J, Guilak F. 2014 Thestructure and function of the pericellular matrix ofarticular cartilage. Matrix Biol. J. 39C, 25 – 32.(doi:10.1016/j.matbio.2014.08.009)
42. Bennetts CJ, Sibole S, Erdemir A. 2015 Automatedgeneration of tissue-specific three-dimensionalfinite element meshes containing ellipsoidal cellularinclusions. Comput. Methods Biomech. Biomed. Eng. 18,1293– 1304. (doi:10.1080/10255842.2014.900545)
43. Julkunen P, Wilson W, Jurvelin JS, Korhonen RK.2009 Composition of the pericellular matrixmodulates the deformation behaviour ofchondrocytes in articular cartilage under staticloading. Med. Biol. Eng. Comput. 47, 1281 – 1290.(doi:10.1007/s11517-009-0547-8)
44. Wu JZ, Herzog W. 2002 Elastic anisotropy ofarticular cartilage is associated with the
microstructures of collagen fibers and chondrocytes.J. Biomech. 35, 931 – 942. (doi:10.1016/S0021-9290(02)00050-7)
45. Kouznetsova VG, Geers MGD, Brekelmans WAM.2004 Multi-scale second-order computationalhomogenization of multi-phase materials: a nestedfinite element solution strategy. Comput. MethodsAppl. Mech. Eng. 193, 5525 – 5550. (doi:10.1016/j.cma.2003.12.073)
46. Guilak F, Mow VC. 2000 The mechanicalenvironment of the chondrocyte: a biphasic finiteelement model of cell – matrix interactions inarticular cartilage. J. Biomech. 33, 1663 – 1673.(doi:10.1016/S0021-9290(00)00105-6)
47. Moo EK, Herzog W, Han SK, Abu Osman NA,Pingguan-Murphy B, Federico S. 2012 Mechanicalbehaviour of in situ chondrocytes subjected todifferent loading rates: a finite element study.Biomech. Model. Mechanobiol. 11, 983 – 993.(doi:10.1007/s10237-011-0367-2)
48. Hunziker EB, Quinn TM, Hauselmann H-J. 2002Quantitative structural organization of normal adulthuman articular cartilage. Osteoarthritis Cartilage10, 564 – 572. (doi:10.1053/joca.2002.0814)
49. Chahine NO, Hung CT, Ateshian GA. 2007 In situmeasurements of chondrocyte deformation undertransient loading. Eur. Cell Mater. 13, 100 – 111;discussion 111.
50. Davis S, Erdemir A. 2014 Convergence ofrepresentative volume elements for prediction ofchondrocyte mechanics in multiscale simulations ofcartilage. In 7th World Congress of Biomechanics,Boston, MA, USA, 6 – 11 July 2014.
51. Sibole SC, Maas S, Halloran JP, Weiss JA, Erdemir A.In press. Evaluation of a post-processing approach
for multiscale analysis of biphasic mechanics ofchondrocytes, doi:10.1080/10255842.2013.809711.Comput. Methods Biomech. Biomed. Eng. (doi:10.1080/10255842.2013.869043)
52. Maas S, Rawlins D, Weiss J, Ateshian G. 2013 FEBio,finite elements for biomechanics, user’s manual,v. 1.8. Musculoskeletal Research Laboratories,University of Utah.
53. COMSOL. 2013 Introduction to COMSOLMultiphysics, v. 4.4. Burlington, MA.
54. Dassault Systemes Simulia. 2013 Abaqus AnalysisUser’s Guide, v. 6.13. Johnston, RI.
55. Schenk O, Gartner K. 2004 Solving unsymmetricsparse systems of linear equations with PARDISO.Future Gen. Comput. Syst. 20, 475 – 487. (doi:10.1016/j.future.2003.07.011)
56. Korhonen RK, Julkunen P, Wilson W, Herzog W.2008 Importance of collagen orientation and depth-dependent fixed charge densities of cartilage onmechanical behavior of chondrocytes. J. Biomech.Eng. 130, 021003. (doi:10.1115/1.2898725)
57. Peeters EAG, Bouten CVC, Oomens CWJ, Bader DL,Snoeckx LHEH, Baaijens FPT. 2004 Anisotropic, three-dimensional deformation of single attached cellsunder compression. Ann. Biomed. Eng. 32, 1443 –1452. (doi:10.1114/B:ABME.0000042231.59230.72)
58. Horton WE, Bennion P, Yang L. 2006 Cellular,molecular, and matrix changes in cartilage duringaging and osteoarthritis. J. Musculoskelet. NeuronalInteract. 6, 379 – 381.
59. Erdemir A. 2014 Multiscale cartilage biomechanics:technical challenges in realizing a high-throughputmodeling and simulation workflow. In 7th WorldCongress of Biomechanics, Boston, MA, USA, 6 – 11July 2014.