Supplementary Information

Predicting gene expression using morphological cell

responses to nanotopography

Cutiongco et al.

1
Supplementary Figure 1. Representative images of musculoskeletal cell types
cultured on four different nanotopographies. Cells were stained using antibodies
against the nucleus, actin cytoskeleton, focal adhesion kinase (FAK) and phosphorylated
(Y397) focal adhesion kinase (pFAK) and imaged after 2 days in culture on the
nanotopographies: FLAT, SQ, NSQ, and HEX. Scale bar = 50 µm.

2
3
 (caption for figure in previous page) Supplementary Figure 2. Conventional
measurements of cells and focal adhesions reveal cell type specific changes
induced by nanotopography. Histograms cell area, nucleus area, pFAK/FAK integrated
intensity ratio, actin integrated intensity and focal adhesion area all cell type and
topography combinations. Nanotopographies are color coded, with FLAT denoted in pink,
SQ denoted in purple, NSQ denoted in blue and HEX denoted in green.

Supplementary Note 1

Cells manifested qualitative differences after 2 days. Pre-myoblasts and myoblasts were
highly elongated on FLAT but were more irregularly shaped with highly discrete and large
focal adhesions on SQ. Pre-osteoblasts and osteoblasts had very prominent actin stress
fibers and large adhesions regardless of topography. Chondrocytes generally had low pFAK
expression, weak stress fiber formation, and stellate and irregular shapes, particularly on SQ
and NSQ. Focal adhesions were largest in osteoblastic cells on NSQ and SQ. Fibroblasts
had very diffuse focal adhesions and were small and elongated on FLAT and SQ, while
those on NSQ and HEX were larger and with more aligned actin stress fibers.

Statistically significant differences were also noted. NSQ significantly increased cell area
relative to FLAT. Changes in pFAK/FAK intensity ratios showed cell type dependence:
chondrocytes cultured on SQ, pre-osteoblasts cultured on HEX, and fibroblasts on HEX
showed the highest pFAK activation compared to the FLAT control. Generally, both pre-
myoblasts and myoblasts showed low pFAK activation compared to all cell types. Actin
intensity changed most markedly in chondrocytes cultured on HEX, showing higher variance
and higher intensity values compared with the FLAT control. Pre-osteoblasts and
osteoblasts on NSQ and HEX developed focal adhesions with the largest area compared to
all cell types.

4
 Pre-myoblasts Myoblasts
FLAT SQ NSQ HEX FLAT SQ NSQ HEX

Nucleus
TAZ
Actin
YAP

50 µm

Pre-osteoblasts Osteoblasts
FLAT SQ NSQ HEX FLAT SQ NSQ HEX
Nucleus
TAZ
Actin
YAP

50 µm

Chondrocytes Fibroblasts
FLAT SQ NSQ HEX FLAT SQ NSQ HEX
Nucleus
TAZ
Actin
YAP

50 µm

Supplementary Figure 3. Representative images of cellular YAP/TAZ expression

and localization in response to nanotopography. Cells were stained using antibodies
against the nucleus, actin cytoskeleton, and the mechanosensors YAP and TAZ then
imaged after 2 days in culture on the nanotopographies. Scale bar = 50 µm.

5
Supplementary Figure 4. Quantification of YAP/TAZ nuclear localization changed by nanotopography. The ratio of integrated intensity
in the nucleus normalized to total intensity in the cell of (A) YAP or (B) TAZ are shown here. Changes in the nuclear/cytoplasmic intensities of
YAP or TAZ indicate changes in mechanosignaling induced by nanotopographical cues. The effects of nanotopography on YAP or TAZ
nuclear translocation vary according to the cell type. Relative to its FLAT controls, NSQ induced higher whereas SQ showed lower nuclear
translocation of YAP and TAZ on myoblastic and osteoblastic cells, respectively. Boxplots show median and interquartile range, with minima
and maxima at the whiskers. Nanotopographies are color coded, with FLAT denoted in pink, SQ denoted in purple, NSQ denoted in blue and
HEX denoted in green.

6
Supplementary Table 1. Descriptors/features measured by various modules on CellProfiler. Descriptions from the CellProfiler
documentation1, Kumar et al.2, Vega et al.3, and Huang et al4.
Area and Shape
Area Total number of pixels in each object.
Perimeter Total number of pixels on the
boundary of each object.
Form Factor Ratio of the area to the perimeter of
each object. Ratio of 1 is for a perfect
circle.
Extent Given the smallest bounding box that
can enclose the object, the extent is
the ratio of the area of the object by
the area of the bounding box. Cells
that have fewer protrusions have
larger extent values.
Eccentricity Ratio of the distance between two
foci and the major axis of the ellipse
that bounds the cell. Ratio
approaching 1 indicates a more
elliptical shape.
Major Axis Length of the major axis of the ellipse
Length that bounds the cell given in pixels.
Minor axis Length of the minor axis of the ellipse
length that bounds the cell given in pixels.
Orientation Angle between -90 and 90 degrees
that indicates the orientation of the
major axis of the cell with respect to
the horizontal.

7
Maximum The greatest distance between any
Radius pixel inside the object to the nearest
pixel outside of the object.
Compactness Ratio of the variance in the radial
distance from the centroid and the
area of the object.

Maximum Feret The maximum distance between two

Diameter parallel lines tangent to opposite
sides of the object.
Minimum Feret The minimum distance between two
Diameter parallel lines tangent to opposite
sides of the object.

8
Solidity Measure of how many holes or
concave boundaries in each object.
Measured as area/convex hull area.
The convex hull area is fitted onto the
object by drawing a straight line
across straight or concave areas of
the cell but expanding outward along
object protrusions. Ratio of 1
indicates an object without any
concavities or indentations, while a
ratio less than 1 indicates an object
with holes or an irregular boundary.
Ratio of Area to convex hull area.
Zernike shape Set of polynomial coefficients used to
features describe cell shape with increasing
detail. The smallest circle that
encloses the object is used to
calculate Zernike features. Cells that
are more closely related to the
Zernike polynomial in question are
reflected in a higher value.

9
 Intensity
Integrated Total intensity values of pixels within
Intensity a specified object. (e.g.
measurement of total intensity of
focal adhesions within a single cell)

Integrated The sum of pixel intensities along the

Edge Intensity perimeter of the object. (e.g.
measurement of total intensity of
focal adhesions on the perimeter of
the object)

Fraction at The ratio of the stain in object to the

Distance total stain within a given radius
(FracAtD) (measured from the cell nucleus).

10
Mean The mean fractional intensity at a
Fractional specified radius (with the center
Intensity defined as the nucleus).
(MeanFrac)
Radial Coefficient of variation of intensity
coefficient of within a ring (with the center defined
variation as the nucleus).
(RadialCV)

Zernike Characterizes the distribution of

Magnitude intensity across an object from its
center described by Zernike
polynomials. Higher Zernike
polynomials describe the object with
higher detail. Values obtained are
correlation coefficients between the
intensity distribution within the object
and the given Zernike polynomial.
Zernike Phase Provides information about the
orientation of the Zernike magnitude
measurement.

11
 Haralick texture features
The Haralick texture features measure pixel frequency information in an object and this information is summarized in each Haralick
measure.
Angular Second Moment Measures object homogeneity in pixel gray levels. A high value indicates very similar pixels within the vicinity.
Based on Haralick’s measure, H1
Contrast Measures the local intensity variations between a reference pixel and its neighbors. A high values indicate
large differences in intensity. Based on Haralick’s measure, H2
Correlation Indicates the linearity of the object. A high value is obtained when the object contains linear structures. Based
on Haralick’s measure, H3
Variance (Sum of Squares) Measures heterogeneity of pixel gray levels in each object. Based on Haralick’s measure, H4
Inverse Difference Moment Measures local homogeneity. A high value indicates uniform local gray levels. Based on Haralick’s measure,
H5
Sum Variance Summation of variance measurements across each object, which gives the distribution of size in grains
measured by granularity. Based on Haralick’s measure, H7
Sum Entropy Summation of entropy measurements across each object, which is low for uniform objects. Based on
Haralick’s measure, H8
Entropy Measures the degree of randomness in the object. Higher entropy values occur when there is more
heterogeneity in neighboring pixels. Based on Haralick’s measure, H9

Difference Variance Difference in variance measurements across each object. Based on Haralick’s measure, H10

Difference Entropy Difference in entropy measurements across each object, which is high for uniform objects. Based on Haralick’s
measure, H11

Information Measure of Ratio of Entropy to inverse difference moment/local homogeneity of the object. Measure of the uncertainty in
Correlation 1 correlation measurements. Based on Haralick’s measure, H12
Information Measure of Value related to the square root of entropy. Measure of the uncertainty in correlation measurements. Based on
Correlation 2 Haralick’s measure, H13

12
Granularity Measures the grain size. High values indicate the coarseness of the texture.

Non-Haralick textures
Gabor factor Measures correlation between bands of intensities. A higher correlation of pattern of intensity results in a
higher Gabor feature measurement (e.g. parallel actin stress fibers lead to a higher Gabor factor value)

13
Supplementary Figure 5.
Measurement of
musculoskeletal marker
expression on all
combinations of cell
type and topography.
Quantitative polymerase
chain reaction (qPCR)
was used to determine
the changes in gene
expression across 14
different musculoskeletal
genes. Fold change in
expression of each gene
expression were
calculated with respect to
fibroblasts on Flat controls
(n>4 from 2 independent
experiments).
Nanotopographies are
color coded, with FLAT
denoted in pink, SQ
denoted in purple, NSQ
denoted in blue and HEX
denoted in green. Data
are presented as mean ±
standard deviation, with
individual data points
presented as open faced
circles.

14
Supplementary Figure 6. Comparison of gene expression between cells stimulated by nanotopography and biochemical inducers of
differentiation. (A-D) Control experiments were performed by differentiating pre-myoblasts, pre-osteoblasts, chondrocytes and fibroblasts
grown on standard tissue culture polystyrene with established biochemical inducers. Thereafter, gene expression from the control experiments
were measured at different timepoints. Day 0 indicates cells undifferentiated cells. (A) Myogenic, (B) osteogenic, (C) chondrogenic and (D)
fibrotic differentiation were measured across multiple timepoints (n=6, 2 independent experiments). Data are presented as mean ± standard
deviation, with individual data points presented as open faced circles. (E-H) Hierarchical clustering of average gene expression data.
Dendrogram shows relationship in gene expression between biochemically-differentiated controls and nanotopographically-stimulated cells.
Notable similarities in biochemically-driven differentiation (control) and nanotopography-induced gene expression are highlighted in blue. Gene
expression from cells on nanotopography were measured after 7 days of culture, and are originally presented in Figure 3 of the main text.

15
Supplementary Table 2. Exact p values denoting statistically significant
comparisons of gene expression. Gene expression measurements presented
in Figure 3 were compared using one-way ANOVA with Tukey’s post-hoc test for
pairwise comparison.
Gene Pairwise comparison p value
MYOD1 Pre-myoblast FLAT vs Pre-myoblast SQ 0.0078
MYOG Pre-myoblast FLAT vs Pre-myoblast SQ 0.0327
MYH7 Pre-myoblast FLAT vs Pre-myoblast SQ 0.030
RUNX2 Pre-osteoblast FLAT vs Pre-osteoblast SQ 0.0001
RUNX2 Pre-osteoblast FLAT vs Pre-osteoblast NSQ 0.0001
RUNX2 Pre-osteoblast FLAT vs Pre-osteoblast HEX 0.0001
RUNX2 Osteoblast FLAT vs Osteoblast SQ 0.0001
SP7 Pre-osteoblast FLAT vs Pre-osteoblast NSQ <0.0001
SP7 Osteoblast FLAT vs Osteoblast NSQ <0.0001
SP7 Osteoblast FLAT vs Osteoblast HEX <0.0001
BGLAP Pre-osteoblast FLAT vs Pre-osteoblast HEX <0.0001
BGLAP Osteoblast FLAT vs Osteoblast NSQ <0.0001
SPP1 Pre-osteoblast FLAT vs Pre-osteoblast SQ 0.0377
SPP1 Pre-osteoblast FLAT vs Pre-osteoblast NSQ <0.0001
SPP1 Pre-osteoblast FLAT vs Pre-osteoblast HEX <0.0001
SPP1 Osteoblast FLAT vs Osteoblast NSQ <0.0001
COL2A1 Chondrocyte FLAT vs Chondrocyte HEX 0.003
ACAN Chondrocyte FLAT vs Chondrocyte HEX 0.0001
TGFB1I1 Fibroblast FLAT vs Fibroblast NSQ 0.0037
COL3A1 Fibroblast FLAT vs Fibroblast SQ 0.0032
COL3A1 Fibroblast FLAT vs Fibroblast NSQ <0.0001
COL3A1 Fibroblast FLAT vs Fibroblast HEX 0.0006
ELN Fibroblast FLAT vs Fibroblast HEX 0.0003

16
(A)

(B)

Supplementary Figure 7. Hierarchical clustering of the morphome across nanotopographies. (A)

Mean measurements and (B) variance of the morphome across nanotopographies were clustered. For
visualization, each morphome feature was normalized to have a mean value of 0 and standard deviation of 1
across all datapoints. Features included were significantly varied across cell types (p < 0.05 using one-way
ANOVA). The complete list of morphome features used is found in Supplementary Data 3.

17
Supplementary Table 3. Accuracy of classifying the morphome into different cell types
using Bayesian logistic regression with a leave-one-out scheme. Bayesian logistic
regression models were trained by holding out from the entire dataset a cell type from one
independent experiment. The held-out dataset was used as test data to determine
classification accuracy. We present here classification accuracy from each held-out
experiment, and the mean ± standard deviation of classification accuracy from two
independent experiments. Essentially, we show that the morphome contains sufficient
information to predict cell type outside of the dataset.
Cell type Experiment 1 Experiment 2 Mean ± standard deviation
(%) (%) (%)
Pre-myoblast 56.0 45.2 50.6 ± 7.6
Myoblast 45.3 59.5 52.4 ± 10.1
Pre-osteoblast 60.1 66.7 63.4 ± 4.7
Osteoblast 60.1 70.1 65.1 ± 7.0
Chondrocyte 40.0 33.9 36.8 ± 4.3
Fibroblast 24.7 35.9 30.3 ± 7.8
Random - - 16.7

Supplementary Table 4. Accuracy of classifying the

morphome into different cell types using Bayesian
logistic regression. Bayesian logistic regression models
were trained using a 60% of the entire dataset. The
remaining 40% of the dataset was used to test accuracy of
cell type classification.
Cell type Classification accuracy (%)
Pre-myoblast 92.6
Myoblast 89.3
Pre-osteoblast 95.7
Osteoblast 95.2
Chondrocyte 97.2
Fibroblast 97.4
Random 16.7

18
Supplementary Note 2
The leave-one-out scheme used to measure classification accuracy in Supplementary Table
3 was used to prevent overfitting the data to experimental variation (e.g. in background
staining). For this new classification scheme, we iteratively removed from the entire dataset
a cell type from one independent experiment (the held-out dataset, e.g. Chondrocyte from
the first independent experiment). The remaining dataset was used to train a Bayesian
logistic regression model (details in Supplementary Methods) to classify a cell type using the
morphome. The left-out dataset was then used as test data to measure the accuracy of
classification using the morphome. That is, in the training set we do not see examples of the
cell types from a biological experiment, but we exploit the availability of other cell types to
avoid overfitting to any individual experimental setup. We averaged the classification
accuracy of each class from two independent experiments and present the average ±
standard deviation. In this manner, we obtained a measure of the predictive quality of the
morphome that does not conflate information about the specific cell type in question from
independent experiments. Essentially this leave-one-out scheme allows us to predict outside
of the dataset, an important task supporting the claim that the morphome contains sufficient
information to classify cell types.

Overall, we showed that classification of cells using the morphome still outperformed
random classification (Supplementary Table 3). This is similar to what we observed when we
use the conventional machine learning method that separates the dataset into train and test
sets without regard to experimental variation (Supplementary Table 4). Both these results
align with what we observed in Figure 5b, where we demonstrated that prediction of gene
expression outside of the training dataset suffers most with removal of the cell type related
to the gene expression of interest.

19
20
(caption for figure in previous page) Supplementary Figure 8. Morphome separated by
nanotopographies clustered characteristically across each cell type. (A-D) Hierarchical
clustering of the morphome separated by nanotopography. Cell type specific morphome
features are made up of 15 chromatin, 73 actin, 88 FAK and 56 pFAK features. Each
morphome feature was mean centered and normalized to standard deviation per cell type to
result in mean = 0 and standard deviation = 1. The total number of cells analysed per
nanotopography type are: (A) n=1251 for FLAT; (B) n=1215 for SQ; (C) n=1293 for NSQ; (D)
n=1180 for HEX obtained across 2 independent experiments. The colour and intensity of
each tile represents the average value of the feature for the specified cell type and
nanotopography.

Supplementary Note 3
On FLAT, pre-myoblasts and osteoblasts showed predominantly low average values for cell-
type specific features except for morphometric and texture measurements (Supplementary
Figure 8, cluster 4). This was reversed in fibroblasts, with particularly high average values of
FAK and actin radial distribution. Pre-myoblasts showed high average values for actin radial
distribution and intensity, nuclear morphometry, and chromatin textures on SQ (cluster 1).
Chondrocytes showed the opposite profile for the same features. Average values for FAK
measurements were high on pre-myoblasts and pre-osteoblasts and low on osteoblasts
(cluster 4). On NSQ, pre-osteoblasts and osteoblasts showed markedly high average values
of FAK radial distribution (cluster 5), relative to other cell types while pre-myoblasts and
myoblasts cells showed low average values of FAK radial distribution. Pre-myoblasts and
myoblasts cells showed the opposite for the same features. While pre-osteoblasts and
myoblasts demonstrated high average values of actin radial distribution and textures (cluster
6), osteoblasts and myoblasts showed low average values for the same features. Cell type-
specific features were mostly high on pre-myoblasts, myoblasts and osteoblasts on HEX. On
the other hand, fibroblasts and pre-osteoblasts showed low average values for the cell-type
specific features. Surprisingly, chondrocytes exhibited moderate values (close to the mean)
of the cell type-specific features. The cell type-specific hierarchical clusters show low
correlation (Supplementary Table 5).

21
Supplementary Table 5. Correlation coefficient of the morphome hierarchically
clustered by nanotopography.

All
FLAT SQ NSQ HEX
topographies

All
1
topographies

FLAT 0.356 1
SQ 0.389 0.140 1
NSQ 0.564 0.276 0.328 1
HEX 0.458 0.194 0.229 0.451 1

22
23
(caption for figure in previous page) Supplementary Figure 9. Spearman correlation
between the changes in morphome and changes in gene expression resulting from
nanotopography. Spearman correlation coefficients were used to measure the similarities
in changes between morphome and gene expression across nanotopographies. (A)
Hierarchical clustering of the correlation coefficients for each morphome and gene
expression pair. The colour and intensity of each tile represents the correlation coefficient
magnitude and direction of association between the morphome feature and gene
expression. Each feature was mean centered and normalized to standard deviation. (B-G)
Scatterplots showing changes in qPCR against individual morphome features. Data are
presented as mean (square) ± standard deviation (bars). Spearman rank correlation value
between gene expression from specified cell type and the morphome feature (y axis) are
presented in parenthesis.

24
Supplementary Table 6. Mean absolute
error (MAE) and coefficient of
correlation (R2) of the predicted gene
expression values using the
morphome.

Test set
Gene MAE R2
MYOD1 0.10 0.10
MYOG 0.10 0.10
MYH7 0.10 0.10
RUNX2 0.12 0.12
SP7 0.17 0.17
BGLAP 0.14 0.14
SPP1 0.18 0.18
SOX9 0.15 0.15
COL2A1 0.16 0.15
ACAN 0.12 0.16
COL10A 0.16 0.17
TGFB1I1 0.17 0.17
ELN 0.17 0.21
COL3A1 0.21 0.59
All Genes 0.11 0.72

25
26
(caption for figure in previous page) Supplementary Figure 10. Contribution of each
morphome feature to the Bayesian linear regression model. The parameters β from the
linear regression models (i.e. regression weight) weights the contribution of each morphome
feature in the Bayesian linear regression model for predicting gene expression. Higher
magnitudes of the parameter β indicate larger contribution of the morphome feature in
predicting the expression of the specified gene. Data shown are mean (dots) ± high density
interval (bars) of β for each morphome feature and gene. The morphome features were color
coded to denote chromatin (blue), actin (orange), FAK (yellow) and pFAK (green)
measurements. The average value of the model parameter β for each morphome feature
and gene is provided in Supplementary Data 5.

27
 Nucleus Actin FAK pFAK
FLAT
SQ
NSQ
HEX

100 µm

Supplementary Figure 11. Overview of co-culture of pre-osteoblasts and fibroblasts

on entire nanotopographies. Images of cells were captured at day 2 using 40X
magnification and stitched together to form a tiled image of the entire nanotopography,
which was then used for morphome extraction and subsequent analysis. A 2.12 mm x
2.12 mm grid was imaged for each nanotopography. Overview shows homogeneous cell
attachment across all nanotopographies. Scale bar = 100 µm.

28
Supplementary Figure 12. Histograms of predicted gene expression from pre-osteoblasts and
fibroblasts co-cultured on nanotopographies. The morphome was extracted from all cells on the
entire nanotopography. The morphome from the co-culture dataset was then centered to the mean
and normalized to the standard deviation of the dataset used to create the predictive linear models.
To predict the effect of nanotopographies on cell phenotype, the co-culture morphome was then used
as input into the Bayesian linear regression model to obtain qPCR values for 14 different genes.
Nanotopographies are color coded, with FLAT denoted in pink, SQ denoted in purple, NSQ denoted
in blue and HEX denoted in green.

29
 (A) Sum of osteogenic gene expression (B) Sum of fibrotic gene expression
2.0 * 2.5 ****
2.0
**
1.5
Fold change

Fold change
1.5
1.0
1.0
0.5
0.5

0.0 0.0
FLAT SQ NSQ HEX FLAT SQ NSQ HEX

(C) Immunohistochemistry at Day 28

FLAT SQ NSQ HEX
Alizarin red
Von Kossa

1 mm

Supplementary Figure 13. Functional response of co-cultured pre-osteoblasts and

fibroblasts to nanotopographies. Changes in (A) osteogenic and (B) fibrotic gene
expression induced by nanotopography after 7 days of culture. Gene expression was
obtained from 2 independent experiments measured twice (n=4). Data shown are mean ±
standard deviation, with symbols denoting mean value of each osteogenic (circle) or
fibrotic (square) gene. Statistical significance in fold change between FLAT and indicated
nanotopography FLAT was calculated using one-way ANOVA with Dunnet’s post-hoc test.
Significance levels were denoted by * (p<0.05), ** (p<0.01), *** (p<0.001), and ****
(p<0.0001). (C) Mineralization of cocultured pre-osteoblasts and fibroblasts after 28 days
of culture.

30
Supplementary Table 7. Sequence of primers used for quantitative measurement of gene
expression.
Gene Primer sequence Amplicon size
18s ribosomal RNA Fwd AAGTCCCTGCCCTTTGTACACA 100
(reference gene) Rev GATCCGAGGGCCTCACTAAAC
RUNX2 (Runx2) Fwd AAGTGCGGTGCAAACTTTCT 90
Rev TCTCGGTGGCTGCTAGTGA
BGLAP (Osteocalcin) Fwd CTGACCTCACAGATGCCAAG 98
Rev GTAGCGCCGGAGTCTGTTC
SPP1 (Osteopontin) Fwd TCAGGACAACAACGGAAAGGG 139
Rev GGAACTTGCTTGACTATCGATCAC
SP7 (Osterix) Fwd GGTCCAGGCAACACACCTAC 184
Rev GGTAGGGAGCTGGGTTAAGG
SOX9 (Runx2) Fwd GGAAGGGAGAGAGAGAGAGAAA 138
Rev CGGGATTTAAGGCTCAAGGT
COL2A1 (Collagen type 2 Fwd ACGAAGCGGCTGGCAACCTCA 73
alpha 1 chain) Rev CCCTCGGCCCTCATCTCTACATCA
COL10A1 (Collagen type Fwd TTCTCCTACCACGTGCATGTG 191
X alpha 1 chain) Rev AGGCCGTTTGATTCTGCATT
ACAN (Aggrecan) Fwd GTGAGGACCTGGTAGTGCGAGTGA 103
Rev GAGCCTGGGCGATAGTGGAATATA
MYOG (Myogenin) Fwd GAGACATCCCCCTATTTCTACCA 106
Rev GCTCAGTCCGCTCATAGCC
MYOD1 (Myogenic Fwd CCACTCCGGGACATAGACTTG 109
differentiation 1) Rev AAAAGCGCAGGTCTGGTGAG
MYH7 Fwd CTCAAGCTGCTCAGCAATCTATTT 153
(Myosin heavy chain 7) Rev GGAGCGCAAGTTTGTCATAAGT
COL3A1 (Collagen type III Fwd CTGTAACATGGAAACTGGGGAAA 144
alpha chain 1) Rev CCATAGCTGAACTGAAAACCACC
TGFB1I1 (Transforming Fwd ATGTCACGGTTAGGGGCTC 83
growth factor beta 1 Rev GGCTTGCATACTGTGCTGTATAG
induced transcript 1)
ELN (Elastin) Fwd TGTCCCACTGGGTTATCCCAT 92
Rev CAGCTACTCCATAGGGCAATTTC

31
(A)

(B)

Supplementary Figure 14. Effect of using different Pearson correlation cutoffs to

filter the dataset. Total variance from the (A) entire data set or (B) the variance of the
topography-specific dataset was measured at different Pearson correlation cutoffs. At a
correlation cutoff of 0.9, the data set considerably reduced in feature number without
compromising on the dataset variance.

32
Supplementary Methods

Conventional cell measurements

Images of cells stained against the chromatin, actin, focal adhesion kinase (FAK) and FAK
phosphorylated at the Tyrosine 397 site (pFAK) were used to measure cell area, nucleus
area, total pFAK/FAK integrated intensity ratio, actin integrated intensity and individual focal
adhesion area. The nuclei were segmented using the image of chromatin, the cell body was
segmented using the image of actin, while individual focal adhesions were segmented using
the image of pFAK. Measurements were taken from distinct cells across 1 independent
experiment.

Staining for mechanosensors YAP and TAZ

After 2 days, cells on nanotopographies were fixed and permeabilised as described in the
main text (see Materials and Methods). Cells were stained against YAP (Cell Signalling
Technologies 4912, 1:70) and TAZ (BD Pharmingen 560235, 1:200) overnight at 4°C then
incubated with Alexa Fluor-conjugated secondary antibodies (ThermoScientific, 1:500). Cells
were also stained with DAPI and fluorescently conjugated phalloidin (ThermoScientific) to
outline the nucleus and the cell body. All substrates were mounted on 0.17 µm thick glass
coverslips with ProLong mounting medium (ThermoScientific) and dried overnight at 4°C
before imaging.

Biochemical differentiation of musculoskeletal cells

C2C12, MC3T3, chondrocytes and NIH3T3 cells were stimulated biochemically to provide
gene expression landmarks of lineage commitment. C2C12 and MC3T3 were biochemically
stimulated with established inducers of mature myogenic and osteogenic phenotypes,
respectively. Mouse C2C12 myoblasts (ATCC) were cultured in DMEM with 20% FBS and
1% penicillin-streptomycin, and committed into mature myoblastic cells using DMEM
supplemented with 2% horse serum and 1% penicillin streptomycin5,6. Mouse chondrocytes
were cultured in minimum essential medium alpha (MEM) with nucleosides, ascorbic acid,
glutamate, sodium pyruvate supplemented with 10% FBS and 1% penicillin-streptomycin.
Mouse MC3T3 cells (ATCC) were cultured in MEM with nucleosides and L-glutamine
without ascorbic acid and supplemented with 10% FBS and 1% penicillin-streptomycin. To
commit MC3T3 into mature osteoblasts, MC3T3 media was supplemented with 10 nM
dexamethasone, 50 µg per ml ascorbic acid and 10 mM β-glycerophosphate7,8. Additionally,
we stimulated the commitment of primary chondrocytes and fibroblasts into the
chondrogenic and fibrotic lineages through similarly established methods. Chondrocytes

33
were stimulated by supplementing the growth media with 1% insulin-transferrin-selenium
(Gibco), 20 µg per ml ascorbic acid (Sigma Aldrich) and 50 ng per ml bone morphogenetic
protein-2 (animal free, PeproTech, Inc)9. NIH3T3 were stimulated into the fibrotic phenotype
by supplementing the growth media with 5 ng per ml mouse transforming growth factor beta
1 (CellSignaling Technologies)10-12. Cells on conventional tissue culture plastic were
biochemically stimulated for 16 days, with media change every 2 days. RNA was collected at
multiple timepoints and used to measure gene expression, as described in the main text
(see Methods section). Gene expression was measured twice from each independent
experiment.

Logistic regression
Logistic regression was used as a supervised classification algorithm to separate the
morphome according to cell type. A Bayesian logistic regression algorithm was utilised. We
consider probability of classification into each cell type y to be modelled from the multinomial
family13. All model parameters were assumed a priori to come from a normal distribution
parameterized by mean = 0 and standard deviation = 5. A logistic regression model was
trained by excluding from the dataset an entire cell type from one independent experiment
(e.g. Chondrocyte from first independent experiment). The held-out dataset was then used
as test data to determine classification accuracy. Model convergence was confirmed with the
potential scale reduction statistic split-Rhat ≥ 1, effective sample size smaller than total
sample size and low autocorrelation. To determine the suitability of the prior distribution, the
data regenerated from the prior predictive distribution (i.e. using only model parameters and
without seeing any data) closely matched the dataset. For each cell type, classification
accuracy was calculated twice (corresponding to two independent experiments). This leave-
one-out scheme is critical to measure the true predictive capability of the morphome for
classifying cell type without confounding factors from experimental variance14. Logistic
regression and confusion matrices were created using the brms10 and caret15 package,
respectively, on R.

Mineralization study
After 28 days, calcium phosphate deposition was studied using Alizarin red and Von Kossa
histological stains. Cells were fixed with 4% paraformaldehyde and stained against 4%
Alizarin red (pH 4.2, Sigma Aldrich), or von Kossa (Merck Millipore) and NuclearFast Red
(Sigma Aldrich). Samples were imaged at 10X magnification using a color CCD camera.

34
Supplementary References
1. Kamentsky, L. et al. Improved structure, function and compatibility for CellProfiler:
modular high-throughput image analysis software. Bioinformatics 27, 1179–1180
(2011). doi: 10.1093/bioinformatics/btr095
2. Kumar, R. M. & Sreekumar, K. A. survey on image feature descriptors. Int. J.
Comput. Sci. Inf. Technol. 5, 7668-7673 (2014).
3. Vega S.L. et al. Organizational metrics of interchromatin speckle factor domains:
integrative classifier for stem cell adhesion & lineage signalling. Integr. Biol. 7, 435–
446 (2015). doi:10.1039/c4ib00281d.
4. Huang K. and Murphy R.F. From quantitative microscopy to automated image
understanding. J. Biomed. Opt. 9, 893–912 (2004). doi:10.1117/1.1779233.
5. Quach, N. L., Biressi, S., Reichardt, L. F., Keller, C. & Rando, T. A. Focal Adhesion
Kinase Signaling Regulates the Expression of Caveolin 3 and β1 Integrin, Genes
Essential for Normal Myoblast Fusion. Mol. Biol. Cell 20, 3422–3435 (2009). doi:
10.1091/mbc.E09-02-0175
6. Clemente, C. F. M. Z., Corat, M. A. F., Saad, S. T. O. & Franchini, K. G.
Differentiation of C2C12 myoblasts is critically regulated by FAK signaling. Am. J.
Physiol. Regul. Integr. Comp. Physiol. 289, R862–R870 (2005).
7. Quarles, L. D., Yohay, D. A., Lever, L. W., Caton, R. & Wenstrup, R. J. Distinct
proliferative and differentiated stages of murine MC3T3-E1 cells in culture: an in vitro
model of osteoblast development. J. Bone Miner. Res. 7, 683–692 (1992).
8. Yan, X.Z. et al. Effects of continuous passaging on mineralization of MC3T3-E1 cells
with improved osteogenic culture protocol. Tissue Eng Part C Methods 20, 198–204
(2014).
9. Perrier-Groult, E., Pasdeloup, M., Malbouyres, M., Galéra, P. & Mallein-Gerin, F.
Control of collagen production in mouse chondrocytes by using a combination of
bone morphogenetic protein-2 and small interfering RNA targeting Col1a1 for
hydrogel-based tissue-engineered cartilage. Tissue Eng. Part C Methods 19, 652–
664 (2013).
10. Chen, Y.-L. et al. Sorafenib ameliorates bleomycin-induced pulmonary fibrosis:
potential roles in the inhibition of epithelial-mesenchymal transition and fibroblast
activation. Cell Death Dis. 4, e665 (2013). doi: 10.1038/cddis.2013.154.
11. Ji, Y.D. et al. BML-111 suppresses TGF-β1-induced lung fibroblast activation in vitro
and decreases experimental pulmonary fibrosis in vivo. Int. J. Mol. Med. 42, 3083–
3092 (2018).
12. Negmadjanov, U. et al. TGF-β1-mediated differentiation of fibroblasts is associated
with increased mitochondrial content and cellular respiration. PLoS ONE 10,
e0123046 (2015). https://doi.org/10.1371/journal.pone.0123046
13. Bürkner, P.C. brms: An R Package for Bayesian Multilevel Models Using Stan. J.
Stat. Softw. 80, 1–28 (2017). doi: 10.18637/jss.v080.i01
14. Shamir, L. Assessing the efficacy of low-level image content descriptors for
computer-based fluorescence microscopy image analysis. J. Microsc. 243, 284–292
(2011).
15. Kuhn, M. Building Predictive Models in R Using the caret Package. J. Stat. Softw. 28,
1–26 (2008). doi: 10.18637/jss.v028.i05

35