Automated Analysis of Cell-Matrix Adhesions in 2D and 3D Environments

Cell-matrix adhesions are of great interest because of their contribution to numerous biological processes, including cell migration, differentiation, proliferation, survival, tissue morphogenesis, wound healing, and tumorigenesis. Adhesions are dynamic structures that are classically defined on two-dimensional (2D) substrates, though the need to analyze adhesions in more physiologic three-dimensional (3D) environments is being increasingly recognized. However, progress has been greatly hampered by the lack of available tools to analyze adhesions in 3D environments. To address this need, we have developed a platform for the automated analysis, segmentation, and tracking of adhesions (PAASTA) based on an open source MATLAB framework, CellAnimation. PAASTA enables the rapid analysis of adhesion dynamics and many other adhesion characteristics, such as lifetime, size, and location, in 3D environments and on traditional 2D substrates. We manually validate PAASTA and utilize it to quantify rate constants for adhesion assembly and disassembly as well as adhesion lifetime and size in 3D matrices. PAASTA will be a valuable tool for characterizing adhesions and for deciphering the molecular mechanisms that regulate adhesion dynamics in 3D environments.

dimensional (1D) patterned fibril-like structures, which were used as a model system for oriented 3D fibrillar matrices 27,29 . However, little mechanistic data for the regulation of adhesion assembly, maturation, and disassembly in 3D matrices is currently available. Progress in this rapidly emerging field has been greatly hampered by the lack of available tools to analyze adhesion dynamics in 3D environments.
To address this need, we have created an automated platform, PAASTA, for analyzing adhesion dynamics in cells migrating on both 2D substrates and in 3D environments, that is based on an open source MATLAB framework, CellAnimation 30 . We manually validate our platform using an established adhesion analysis method 9 and use PAASTA to quantify adhesion dynamics in 3D matrices.

Results
An automated platform, PAASTA, for adhesion analysis. In order to perform automated detection and quantification of adhesions over time, we begin with raw images, which are acquired with time-lapse microscopy, of cells with fluorescently-labeled adhesions (Fig. 1). Initially, a Gaussian smoothing module is applied to the raw images to reduce noise and then corrected for uneven background illumination by dividing each smoothed image with a low pass filtered version of itself. To detect adhesions, we employ a local thresholding module that compares the intensity of each pixel with the mean value of the local neighborhood of the pixel. If the value of the pixel is higher than the local average, it is classified as an adhesion This platform subjects raw images to a series of steps to identify adhesions. A flow chart detailing each step with a corresponding output image is shown. A segmentation pipeline (red lines), where images are processed in order to segment adhesions, is followed by a tracking pipeline (blue lines) that identifies and tracks adhesions through a series of time-lapse images. Upon completion of the track assignment process for a particular frame, an overlay image is generated which displays the ID numbers for each adhesion in the frame. These images may be used for visual inspection of the automated tracking assignments. Intensity values are calculated for all tracked adhesions at every time point using the background-corrected intensity image. When the tracking assignment is completed for all the frames, the final adhesion ID numbers along with centroid locations and intensity values are exported to comma-separated text files for further analysis. pixel; otherwise, it is assigned to the background pixel class. Objects less than 1 mm 2 are excluded to ensure that background noise is eliminated from the analysis. Cell outlines are detected by thresholding the background-corrected images using a global intensity threshold module. Adhesions are selected by combining the binary mask of the cell with the binary image of the adhesions. Individual adhesions, which are assigned identification (ID) numbers, are tracked over time using a nearest neighbor algorithm. Adhesion ID numbers, individual adhesion integrated intensities, and area information at every time point are exported to comma-separated text files for further analysis. Sets of images showing the detected adhesion outlines, with or without ID numbers, are overlaid on the original images for manual validation of the automated quantification (Fig. 2).
Manual validation of PAASTA. We manually validated the automated tracking data received from PAASTA from time-series of cells expressing the fluorescently-tagged adhesion proteins, paxillin or vinculin. In these experiments, GFP-paxillin transfected HT1080 cells were plated on glass bottom dishes, which were coated with the ECM protein fibronectin (2D substrate), and imaged using total internal reflection fluorescence (TIRF) microscopy (Supplementary Movie 1). From these images, we manually tracked individual adhesions, measured fluorescence intensities in these adhesions, and quantified the kinetics of adhesion assembly and disassembly as previously described 9 . We next compared this manual adhesion data to the output generated by PAASTA using the same raw images. The relative changes in fluorescence intensities obtained from PAASTA for assembling adhesions was similar to that measured manually ( Fig. 3A and B). To calculate apparent rate constants for adhesion assembly, we generated semilogarithmic plots of fluorescence intensities of individual adhesions as a function of time. The slopes of these graphs, which correspond to the apparent rate constant for adhesion assembly, were similar for the manually generated data and the data from PAASTA (Fig. 3C). In addition, for disassembling adhesions, the fluorescence intensity profiles attained manually and from PAASTA were similar ( Fig. 3D and E), and the rate constants for adhesion disassembly were comparable for data obtained manually and with PAASTA (Fig. 3F). Indeed, the rate constants for adhesion assembly and disassembly, which we express as t 1/2 values, that were obtained from manually tracking GFP-paxillin adhesions were very similar to those attained with PAASTA (Fig. 3G). An individual value plot shows the range of t 1/2 values for adhesion assembly and disassembly ( Supplementary Fig. S1). Moreover, in HT1080 cells expressing GFP-vinculin, another adhesion protein, the t 1/2 values for adhesion assembly and disassembly were comparable for manual adhesion tracking and PAASTA (Fig. 3H). When we extended these observations to U2OS cells, we obtained very similar results for adhesion assembly ( Applications of PAASTA to adhesion dynamics. To further demonstrate the capabilities of PAASTA, we compared adhesions in HT1080 and U2OS cells (Fig. 5A). In these time-series, PAASTA   tracked a total of 46 adhesions in the HT1080 cell and 50 adhesions in the U2OS cell. The average adhesion lifetime, defined as the total time an adhesion was observed during the time course, was 6.7 6 0.7 min and 9.9 6 1.0 min for the HT1080 and U2OS cell, respectively. The U2OS cell had more adhesions with a lifetime of greater than 19 min compared to the HT1080 cell (Fig. 5B). The average adhesion size for the HT1080 and U2OS cell was 5.4 6 0.4 mm 2 and 5.8 6 0.6 mm 2 , respectively. Interestingly, even though the average size was comparable between the two cells, the U2OS cell had more small and large adhesions, while the HT1080 cell  had a majority of moderately sized adhesions (Fig. 5C). Furthermore, the two cells showed very similar trends when comparing adhesion lifetime to adhesion size (Fig. 5D).
Analysis of adhesion dynamics in a 3D environment with PAASTA. An attractive feature of PAASTA is that it is designed to analyze adhesion dynamics in 3D environments. Kubow et al. 20 recently showed that very low expression of GFP-tagged adhesion proteins under the control of a truncated CMV promoter is ideal for imaging adhesions in 3D matrices. Hence, we employed this approach to generate time-lapse images for analysis of adhesions with PAASTA. In initial experiments, we expressed GFP-paxillin and GFP-vinculin cDNAs with the truncated CMV promoter (Specpaxillin and Spec-vinculin) in both HT1080 and U2OS cells. We subsequently imaged these cells using time-lapse microscopy and quantified the t 1/2 values for adhesion assembly and disassembly both manually and with PAASTA. The average t 1/2 values obtained with these truncation constructs were quite similar to those attained in HT1080 cells with GFP-paxillin and GFPvinculin with the full-length CMV promoter (FL-paxillin and FLvinculin) (Fig. 3G and H). Moreover, we observed comparable t 1/2 values with these truncation constructs in U2OS cells (Fig. 4G). We therefore proceeded to use these constructs to analyze adhesion assembly and disassembly in U2OS cells embedded in 3D type I collagen matrices.
We generated multidimensional time-lapse images (with a zinterval of 0.5 mm) for U2OS cells expressing either Spec-paxillin or Spec-vinculin (Supplementary Movie 3). Only cells that were at least 100 mm from the coverslips were imaged to ensure they were embedded in the 3D matrices. Adhesions were identified in each zplane and tracked as a function of time through the z-stack using the nearest neighbor algorithm. This approach allows adhesions that are moving through different focal planes to be tracked over time. Adhesion ID numbers were exported along with the average integrated fluorescence intensities from the z-planes in which the adhesions were present. A profile of the fluorescent intensities obtained from PAASTA showed an adhesion assembling and disassembling in 3D type I collagen matrices (Supplementary Fig. S2A). Changes in fluorescence intensities for an assembling (Fig. 6A and B) and disassembling ( Fig. 6E and F) adhesion are also shown along with semilogarithmic plots of fluorescence intensities over time (Fig. 6C and G). The average R 2 values for the adhesion assembly and disassembly plots are 0.89 6 0.01 (S.E.M. from 43 adhesions) and 0.88 6 0.01 (S.E.M. from 48 adhesions), respectively. From these plots, t 1/2 values were calculated for assembly and disassembly of adhesions tracked through PAASTA (Fig. 6I). An individual value plot shows the range of t 1/2 values for adhesion assembly and disassembly ( Supplementary Fig. S2B). Distribution plots revealed that most paxillin and vinculin-containing adhesions have t 1/2 values of less than 10 min for assembly and disassembly in 3D type I collagen matrices   Fig. 6D and H). Moreover, the average adhesion lifetime was 13.4 6 1.0 min, and the average adhesion size was 6.1 6 0.4 mm 2 for U2OS cells in 3D type I collagen matrices ( Supplementary Fig. S3A and B). However, these plots show some variability with some adhesions having t 1/2 values of greater than 20 min, lifetimes of 30 min, and average sizes larger than 10 mm 2 . Others have similarly reported variability in adhesion parameters, including their size, distribution, shape, and location [31][32][33] . Interestingly, the average adhesion size correlated with the adhesion lifetime ( Supplementary Fig. S3C), suggesting that smaller adhesions have a shorter lifetime than the larger adhesions. Taken together, our data demonstrate that PAASTA is a useful platform for rapidly analyzing multiple adhesion parameters in 3D environments over time.

Discussion
Since adhesions were first shown to be direct regions of contact between a cell and the substratum using interference reflection microscopy, they have been extensively studied, characterized, and analyzed on flat 2D substrates [34][35][36] . These studies have been extremely beneficial in understanding adhesion organization, regulation, and structure and have laid the foundation for the identification of adhesions in more complex, physiologic 3D environments 17,26,37,38 . Adhesions in 3D environments differ from adhesions on 2D substrates in some respects and appear to more closely resemble adhesions in vivo 17,21,22 . These observations warrant a more thorough analysis of adhesion organization, regulation, and dynamics in 3D environments. However, current analysis methods and systems for quantifying adhesion parameters, such as assembly, disassembly, and size, in cells plated on 2D substrates 9,39,40 have not been shown to have the capability to analyze adhesion dynamics in 3D environments. This lack of available tools to analyze adhesions in 3D environments has hindered progress toward understanding adhesions in 3D. Consequently, we have developed a reliable, powerful platform (PAASTA) for the large-scale, rapid analysis of adhesions in 3D environments. PAASTA uses multidimensional images to identify and track adhesions through z-planes over time, permitting adhesion dynamics to be quantified in 3D. Therefore, PAASTA should prove to be a useful tool for investigating adhesions in 3D environments and for deciphering the molecular mechanisms that regulate adhesion dynamics in 3D.
Using PAASTA, we examined adhesion assembly and disassembly in U2OS cells embedded in 3D type I collagen matrices. The t 1/2 values for adhesion assembly and disassembly were approximately 7 min (Fig. 6). Based on data obtained on 2D substrates 41,42 , these results suggest that adhesions in cells in 3D type I collagen matrices are relatively stable. Small adhesions in cells on 2D substrates have been reported to turn over in a few minutes (Fig. 3) [41][42][43] . However, adhesion turnover on 2D substrates could differ among cell types, ECM proteins, and ECM concentration. Furthermore, adhesion in 3D is most likely more complex than adhesion on 2D substrates and influenced by factors, such as matrix composition, pliability, pore size, fiber alignment, as well as the immediate microenvironment of each adhesion 21 . Indeed, fiber orientation was shown to modulate adhesion size and maturation in 3D type I collagen matrices 28 . Therefore, future studies are needed to understand adhesion dynamics in 3D environments and how they compare to 2D substrates.
Investigating adhesions in 3D environments is attractive because they more closely resemble adhesions in vivo compared to adhesions on 2D substrates 17,22 . Thus, studies characterizing adhesions in 3D matrices will provide a wealth of information on the behavior of adhesions in more physiologic environments. These studies could also serve as a foundation for examining adhesions in vivo, which is currently difficult with available technologies. As innovative methodologies emerge, new analytical tools will be needed to characterize adhesions in 3D environments as well as in vivo.
We used PAASTA to analyze adhesion turnover, lifetime, and size in 3D matrices; however, PAASTA is a versatile, automated platform for analyzing many different adhesion characteristics in 3D environments and in vivo. For example, future incarnations of this platform could include features that would assess adhesion shape, distribution, and distance from the cell edge. Because PAASTA is built on the CellAnimation MATLAB platform, other modules can easily be added to the workflow. As technology advances, and adhesions can be more readily visualized in vivo, PAASTA should also provide a valuable platform for analyzing these structures.
Imaging adhesions on 2D substrates. Cells transfected with either GFP-paxillin, GFP-vinculin, Spec-paxillin, or Spec-vinculin were plated on glass-bottomed dishes, which were precoated with 2.5 mg/mL fibronectin (Sigma, St. Louis, MO), and permitted to adhere for 1 h at 37 o C. While imaging, cells were maintained in SFM4MAb TM media (Hyclone) supplemented with 2% FBS, pH 7.4. Cells were imaged on an inverted Olympus IX71 microscope (Melville, NY), which was equipped with a Retiga EXi CCD camera (QImaging, Surrey, BC) and an Olympus PlanApo 603 OTIRFM objective (NA 1.45), using MetaMorph software (Molecular Devices, Sunnyvale, CA). TIRF images were acquired by exciting with a 488 nm laser line from an Argon-Ion laser (Prairie Technologies, Middleton, WI). For TIRF imaging, a z488/543 rpc filter was used (Chroma, Brattleboro, VT). GFP-vinculin was a kind gift from Susan Craig (Johns Hopkins University, Baltimore, MD). Specpaxillin and Spec-vinculin were generously provided by Rick Horwitz (University of Virginia, Charlottesville, VA).
Imaging adhesions in 3D matrices. Rat-tail type I collagen (BD Biosciences, Bedford, MA) was mixed with sterile 103 DMEM (Invitrogen), sterile dH 2 O, FBS, and 1N NaOH to a final concentration of 2 mg/mL type I collagen, 10% FBS and 13 DMEM. NaOH was used for neutralization at 0.023 mL 3 the volume of type I collagen solution. U2OS cells transfected with either Spec-paxillin or Spec-vinculin were seeded (,1.5 3 10 5 cells) into 300 mL of type I collagen solution and pipetted into the bottom of glass-bottomed dishes. The type I collagen solution with embedded cells was allowed to polymerize for at least 30 min at 37uC in a cell culture incubator with 5% CO 2 . Subsequently, 2 mL of culture medium was gently added to each dish, and cells were incubated for approximately 18 h at 37uC in a cell culture incubator with 5% CO 2 . Prior to imaging, the culture medium was replaced with SFM4MAb TM medium supplemented with 10% FBS, pH 7.4. Z-series were acquired using a Quorum WaveFX-X1 spinning disk confocal system with a Yokogawa CSU-X1 spinning disk (Yokogawa Electric Corporation, Newnan, GA) modified with a Borealis upgrade (Guelph, Canada) and a Nikon Eclipse Ti microscope that was equipped with an EM-CCD camera (Hamamatsu, Hamamatsu City, Japan) and a Plan Fluor 403 objective (N.A. 1.3). Z-series were collected using MetaMorph software at time intervals of 45 sec -1 min with a z-interval of 0.5 mm. GFP was excited with a 491 nm laser line and imaged with a 525/50 emission filter (Semrock, Rochester, NY). Only cells that were completely embedded within the 3D collagen matrix (at least 100 mm from the coverslips) were imaged.
Manual adhesion analysis. All manual image analysis was performed using MetaMorph software. Individual adhesions were identified, and a region was created with the trace region tool that completely outlined the adhesion at the timepoint in which the adhesion had the greatest area. The integrated intensity for the fluorescently-tagged adhesion marker (paxillin or vinculin) in this region was recorded over time. An exact duplicate region was positioned within an area adjacent to the tracked adhesion, which was inside the cell and did not contain an adhesion at any timepoint. The integrated fluroescence intensity of this region was then used as background and was subtracted from each timepoint from the region containing the adhesion. These background-corrected data were then used in further processing steps to calculate adhesion kinetics.
Adhesion analysis with PAASTA. For adhesion analysis, a Gaussian smoothing module was applied to the raw images to reduce noise and then corrected for uneven background illumination by dividing each smoothed image with a low pass filtered version of itself. A local thresholding module was used to compare the intensity of each pixel with the mean value of the local neighborhood of the pixel in order to detect adhesions. Cell outlines were detected by thresholding the background-corrected images using a global intensity threshold module. Adhesions were selected by combining the binary mask of the cell with the binary image of the adhesions. Kernel sizes and thresholds were user designated because images can vary due to experimental conditions. Individual adhesions, which were assigned ID numbers, were tracked over time using a nearest neighbor algorithm. Adhesions that split into www.nature.com/scientificreports SCIENTIFIC REPORTS | 5 : 8124 | DOI: 10.1038/srep08124 two adhesions during imaging were tracked as separate adhesions; the adhesion that remained closest to the previous image was tracked with the same ID number whereas the other adhesion was assigned a new ID number. Adhesion ID numbers, individual adhesion integrated intensities, and area information at every time point were exported to comma-separated text files for further analysis. A set of images showing the detected adhesion outlines, with or without ID numbers, were overlaid on the original images for manual validation of the automated quantification. Adhesions were imaged with a temporal resolution (#1 min) that was sufficient to obtain numerous data points for adhesion analysis. Furthermore, because the temporal resolution was high, the number of adhesions did not vary greatly between subsequent images, which allowed the user to adjust any ID numbers due to splitting and merging adhesions. The vast majority of adhesions in cells were correctly identified and tracked using PAASTA, indicating that the adhesion density was amenable to tracking with the nearest neighbor algorithm. A modified nearest neighbor algorithm or binary integer programming can be added to PAASTA if the adhesion density increases dramatically and the nearest neighbor algorithm is no longer sufficient for adhesion tracking.
Calculating rate constants for adhesion assembly and disassembly. The background-corrected integrated fluorescence intensities in individual adhesions were determined manually or with PAASTA. Semilogarithmic plots of the background-subtracted fluorescence intensities over time were then generated as follows: ln [I 0 /I] vs. time for adhesion disassembly and ln [I/I 0 ] vs. time for adhesion assembly, where [I] is the intensity of the adhesion at a given timepoint and [I 0 ] is the initial intensity of the adhesion. Data were then fitted with a linear trendline, and rate constants were determined from the slopes. Rate constants were used to calculate half-life values for adhesion assembly and disassembly using the equation: t 1/2 5 ln(2)/k, where k is the rate constant.