Relationship between nanotopographical alignment and stem cell fate with live imaging and shape analysis

The topography of a biomaterial regulates cellular interactions and determine stem cell fate. A complete understanding of how topographical properties affect cell behavior will allow the rational design of material surfaces that elicit specified biological functions once placed in the body. To this end, we fabricate substrates with aligned or randomly organized fibrous nanostructured topographies. Culturing adipose-derived stem cells (ASCs), we explore the dynamic relationship between the alignment of topography, cell shape and cell differentiation to osteogenic and myogenic lineages. We show aligned topographies differentiate cells towards a satellite cell muscle progenitor state - a distinct cell myogenic lineage responsible for postnatal growth and repair of muscle. We analyze cell shape between the different topographies, using fluorescent time-lapse imaging over 21 days. In contrast to previous work, this allows the direct measurement of cell shape at a given time rather than defining the morphology of the underlying topography and neglecting cell shape. We report quantitative metrics of the time-based morphological behaviors of cell shape in response to differing topographies. This analysis offers insights into the relationship between topography, cell shape and cell differentiation. Cells differentiating towards a myogenic fate on aligned topographies adopt a characteristic elongated shape as well as the alignment of cells.

shows the fabrication scheme of CNT coated coverslips with aligned drawable forest of CNTs. The procedure for both random, aligned and flat topographies is shown.
The areal density of random substrates was matched to that of aligned substrates ( Figure S2).
Absorbance through aligned and random substrate types, measured using a NanoDrop 1000 (Thermo Scientific), was found to be equivalent when a CNT solution of 31.3 µg mL -1 was used. This equates to an areal density of 2.7 µg cm -2 -a result comparable to previously reported values for sheets pulled from drawable aligned forests [2] . All subsequent random substrates were prepared as described above with a coating of CNT solution at 31.3 µg mL -1 . All coverslip control surfaces including uncoated and fibronectin coated glass were treated with vacuum oven, oxidation and pyrolysis treatments described above. Prior to experimentation all surfaces were immersed in 70% ethanol overnight. Figure S2: Absorbance of visible light through random coated substrates at varying concentration as compared to aligned coated substrates (left). Averaging results and plotting the absorbance at 350 nm against concentration of random topographies (or random carbon nanotubes (RCNT)) shows comparable areal density to aligned topographies (or aligned carbon nanotubes (ACNT)) coating at 31.3 µg mL -1 (right) The diameter of the CNTs on either substrate types was characterized to ensure mixing, dispersion, sonication, and coating processing had not altered the dimensionality of CNT populations between substrate types. Figure S3 shows histograms of the diameter of CNTs over aligned and random substrates where the mean diameter of 26.38 nm and 25.13 nm was comparable over aligned and random substrate types, respectively. Measurements of water contact angle is shown in Figure S4A and B with 2 µL water droplet over random and aligned topographies, respectively (measured using a contact angle goniometer, VCA2000 (AST Inc.)). Substrates were comparable ( Figure S4C) with mean contact angle of 89.143 (SD = 2.113, n = 8) for aligned topographies and 90.015 (SD = 4.017, n = 8) for random topographies; P = 0.58 (unpaired two sample Student's t-test). The sheet resistance of the substrates was measured using the 4-point probe / Keithley SCS-4200 ( Figure S4D). Mean sheet resistance of 2997 Ω sq -1 was measured for aligned topographies as compared to 3111 Ω sq -1 for random topographies; P = 0.84 (unpaired two sample Student's t-test) Figure S4: Measurements of water contact angle is shown in panels A and B with 2 µL water droplet over random and aligned topographies.
Water contact angle data is summarized in panel C comparing random and aligned topographies. Panel D compares the sheet resistance of the substrates measured using a 4-point probe.

Transduction with LifeAct
Cell shape parameters were obtained using fluorescent imaging of ASCs stably expressing LifeAct-GFP (an 17-amino-acide peptide that stains F-actin without interfering with actin dynamics -kind gift from R. Wedlich-Söldner, University of Münster, Germany). [3] Transduction was confirmed with fluorescent imaging and flow cytometry ( Figure S6). Subsequent to transduction, cells were sorted for double GFP on a BD Influx flow cytometer and then expanded following standard protocols prior to use in experiments. GFP-Lifeact was excited using the 488 nm laser and collected in the B430/30 channel; Data was analyzed with FlowJo software.

Circularity
Circularity of a cell as the scaled ratio of it's area and perimeter -equal to 4 Area -; equal to 1 for a perfectly circular object and decreases towards 0 for shapes with increasing perimeter for a given area.

Major Axis
The major axis of the cell is the longest axis of the smallest ellipse that completely encloses a cell.

Minor Axis
The minor axis of the cell is the shortest axis of the smallest ellipse that completely encloses a cell.

CellProfiler image segmentation process
To characterize cell shape in an unbiased fashion, time-lapse microscopy data was processed using CellProfiler analysis software. Image processing pipeline is shown below in Table S3. Calculates and corrects image for illumination artifacts, calculated for each image from the background using splines method with automatically calculated spline parameters RescaleIntensity Image intensity rescaled (divide by image maximum) for consistent thresholding and segmentation ImageMath Illumination function calculated above is subtracted from the image, values less than 0 are set to 0. IdentifyPrimaryObjects Segmentation process was completed using maximum correlation thresholding (MCT) with an adaptive window.
Specifically the following parameters were set: • Cells outside of the diameter range: 20 < diameter < 4000 are rejected • Cells touching the border are rejected • Adaptive MCT thresholding was used with automatic smoothing and a threshold correction value of 0.7 • An adaptive window size was set to 400 pixels • Objects were separated set by ther intensity • With a declumping smoothing filter size of 10 pixels • Declumping was suppressed if local maxima were less than 15 pixels • Image was not downsized from declumping

MeasureObjectSizeShape
Object shape parameters we measured SaveImages Grayscale images of objects (each object a different intensity) were saved for processing. ExportToSpreadsheet Object shape parameters we extracted into a spreadsheet. An example of an image before and post segmentation is shown in Figure S7. For all segmented objects in an image, the CellProfiler Module 'MeasureObjectSizeShape' was used to extract metrics on cell area and shape; more information on these metrics can be found a the CellProfiler website: http://www.cellprofiler.org/CPmanual/MeasureObjectSizeShape.html. Extracted shape metrics were then imported into MATLAB R2015b where data for all objects from the same substrate and time but different fields of view were combined (time series plots in main article Figure 5 and 6).
Using a hampel filter, the quality and focus of each image was assessed prior to further analysis. The hampel filters was set to compute the median and standard deviation of a window composed of 10 adjacent data point (5 on either side). Data from images was rejected if the number of cells differed from the median of more than 5 standard deviations. Consequently, gaps in the data can be observed were the microscope was unable to focus correctly (see Figure S8). To validate that the extraction of shape metrics using CellProfiler could be automated, we compared objects that were manually segmented to objects that were obtained from automated segmentation with CellProfiler. We tested that automated segmentation of objects was equivalent to that done manually.
To validate that image analysis could be automated we compared data which had been manually selected to that obtained from automated image analysis. Fitting the dynamic cell area to a power law model = 1 • 3 4 , as previously developed, [4] we determined the probability that the data sets were equal using a F-test. This suggested that the null hypothesis was false (p ~ 0); and automated selection of cell masks can not be used to replace selection done manually (F score can be found in Table S4). However, inspection of the two data sets suggests a high degree of correlation between . Further, data from automated analysis yielded consistently lower values relative to data manually selected. To correct for this difference, we calculated a constant the correction factor (CF) using least square method: 0 = ( :;<=;> − × C=DEF;DGH -DIFGJDIFG KLM DIFGJN . Scaling data obtained through automated selection by the correction factor constant observed that the two sets of data closely fit each other ( Figure S9A-Dii). Further reevaluation of the data with a F-test we demonstrated that the null hypothesis was true (p ~ 1); and both automated and manually selected data sets can be modelled with one equation. Using the same methods we completed analysis comparing the other metrics used herein. Table S4 lists the models, coefficients of trend line fits, F-scores and p-values obtained when comparing metrics between manually and automatically segmented images. Note that the model for circularity ( = 2 • ( 1 −2• 1 • 2 ) ) was developed through substitution and simplification of the area ( = 1 • 3 4 ) and perimeter ( = SGTIFGDGT • 3 U ) -developed elsewhere [4] -into the expression for circularity 4 • -.  To compare the shape metrics of different topographies throughout the culture period we used a F-test over segments of time shorter than the duration of the study. For these comparisons we assumed that the data could be segmented and the data could be fit to a linear model. Figure S8 below shows the comparisons for time segments over which different topographies were statistically the same.

Testing for similarity between topographies for a given cell shape parameter
Given that all cells begun with a similar non adherent circular shape, we examined the time taken for cell shape to become dissimilar between different topographies. Table 1 in the main manuscript summarizes this data. To obtain these values we truncated our data to cover short subsets of the time series beginning at t = 0 days, and ending at t = x days. As per the above analysis, we fit cell shape metrics to power law models or otherwise and used a F-test to determine the probability that the data could be from the same data set (P > 0.05). We then increased the time x until P < 0.05. The last time at which P > 0.05 is reported as the time at which cell shape is similar (as per Table 1).

FFT analysis -processing
To quantify the alignment of cells we compared two dimensional fast Fourier transforms (2D FFT) of images obtained from time-lapse microscopy. The power spectrum of the 2D FFT of an images can be used to compare the presence of features at a given orientation. [5] Summing the power spectrum of 2D FFT images at a given angle the prominence of cell features at given orientations is compared between the different substrate types. This method is advantages as it can provide orientation distributions for images of overlapping features. Further, its speed and robust nature make it the preferred calculation methods for quantification of alignment distributions in this large data set. Prior to calculation of 2DFFTs, the background of each image set was subtracted from each image. The background image was created from initially images prior to cell spreading and movement. Using MATLAB 2015b, a FFT transform of each frame of the fluorescent LifeAct-GFP channel of time-lapse microscopy was taken.
The magnitude spectrum was rearranged to move zero frequency components to the center of the FFT array. The angle of each pixel from the center was calculated and its saturation summed to give the relative strength of different angle bands within the FFT array. Points close to the center of the array were excluded as their close location to the center corresponds to low angular resolution and results in data skewed to certain angles. This gave the total power of the 2D FFT at a given angle and could be used to compare the presence of features at a given orientation. For each topography type, the power spectrums from each field of view at a given time point were combined and normalized. Following calculation of the power spectrums across the complete 21 days, spectrums for a given substrate were combined and graphed as a surface (as per Figure 6 in main article).