A Miniaturized Screening Platform to Identify Novel Regulators of Extracellular Matrix Alignment

Extracellular matrix (ECM) alignment contributes to metastasis in a number of cancers and is a known prognostic stromal factor; however, the mechanisms controlling matrix organization remain unclear. Cancer-associated fibroblasts (CAF) play a critical role in this process, particularly via matrix production and modulation of key signaling pathways controlling cell adhesion and contractility. Stroma normalization, as opposed to elimination, is a highly sought strategy, and screening for drugs that effectively alter ECM alignment is a practical way to identify novel CAF-normalizing targets that modulate ECM organization. To meet this need, we developed a novel high-throughput screening platform in which fibroblast-derived matrices were produced in 384-well plates, imaged with automated confocal microscopy, and analyzed using a customized MATLAB script. This platform is a technical advance because it miniaturizes the assay, eliminates costly and time-consuming experimental steps, and streamlines data acquisition and analysis to enable high-throughput screening applications. As a proof of concept, this platform was used to screen a kinase inhibitor library to identify modulators of matrix alignment. A number of novel potential regulators were identified, including several receptor tyrosine kinases [c-MET, tropomyosin receptor kinase 1 (NTRK1), HER2/ERBB2] and the serine/threonine kinases protein kinase A, C, and G. The expression of these regulators was analyzed in publicly available patient datasets to examine the association between stromal gene expression and patient outcomes. Significance: ECM fiber organization and alignment contribute to metastasis in a number of cancers and are a known prognostic stromal factor; however, the mechanisms controlling matrix organization remain unclear. Here, a high-throughput assay was developed to enable discovery-based screening for an in vitro ECM fiber alignment assay. As proof of concept, this platform was used to screen a kinase inhibitor library and identified several novel modulators of matrix alignment.


Introduction
The extracellular matrix (ECM) provides a structural and biochemical niche that regulates cell function and supports tissue homeostasis. During tumor progression, the normal ECM architecture undergoes dramatic remodeling. In the disease-free survival in patients with breast cancer (8,9). In a recent report, increased collagen alignment also correlated with severity of breast cancer diagnosis (10). Similarly in pancreatic cancer, highly aligned ECM is a negative prognostic factor at time of pancreatic ductal adenocarcinoma (PDAC) resection (11). In colon, ovarian cancer, and other epithelial cancers, collagen organization and alignment are also associated with malignancy (12)(13)(14).
While compelling evidence demonstrates a correlation between tumor stroma fiber alignment and patient outcomes, analysis of tissue fiber alignment is not currently part of standard clinical diagnoses. ECM alignment has also been directly implicated in modulating cancer cell behavior, such as cancer cell invasion and intravasation (2,(4)(5)(6). Matrix alignment in vivo has been shown to facilitate premetastatic breast cancer cell invasion (2,15,16). In vitro, fibroblastic-generated matrices effectively mimic in vivo ECM (17). Changes in ECM topography can modulate cancer cell speed (18,19) and directionality (4,(18)(19)(20)(21) during invasive spread within these natural substrates. Furthermore, matrix alignment can promote cell migration even under counterintuitive conditions, including increased chemokine gradient and/or concentrations or matrix stiffness (4,22). ECM derived from CAFs in vitro mimics the ECM alignment observed in the tumor microenvironment (1,23), which effectively simulates the ECM's protumoral stimulation via regulation of cellular function (ref. 24; i.e., increased migration, invasion, and intravasation). However, disorganized, isotropic ECM derived from normal (or normalized) fibroblasts represses these tumorigenic traits (13,18,19).
Because of the compelling evidence supporting the role of ECM alignment in human disease, considerable effort has been devoted to uncovering the genes and pathways that regulate matrix organization and strategies for therapeutically targeting alignment. CAFs, fibroblasts in the tumor stroma that have been activated to a myofibroblast phenotype, are the principal regulators of the ECM, both by producing ECM molecules and controlling ECM organization and alignment. Therefore, CAFs are an obvious therapeutic target. However, directly eliminating the CAF population has also been shown to reduce the total amount of ECM and can promote tumor progression. For example, ablating CAFs in murine models of PDAC by targeting the myofibroblast marker α-smooth muscle actin or the sonic hedgehog signaling pathway, promoted tumor aggressiveness and decreased survival (25,26). In addition, clinical trials therapeutically targeting CAFs were halted because of concerns for patient safety (27)(28)(29). While ablation of the ECM or CAF populations can promote tumor growth and invasion, restoration of a tumor suppressive, disorganized stroma while maintaining the fibroblast population will be a powerful strategy to inhibit cancer cell invasion and metastasis (30,31).
Collagen crosslinking enzymes, lysyl oxidases (LOX), have also been found to be critical regulators of ECM alignment. LOX expression correlates with aligned collagen at the invasive front of ductal breast carcinomas (32). Inhibition of LOX-like 2 (LOXL2) decreased matrix alignment and tumor volume in a murine breast cancer model (6). Furthermore, inhibition of LOX with β-aminopropionitrile (BAPN) in a murine model reduced tissue stiffness and ECM alignment surrounding mammary tumors (33). However, lack of a complete crystal structure of LOX has hampered development of a specific pharmacologic inhibitor (34). In addition, in a clinical trial in which the LOX inhibitor, BAPN, was applied topically to reduce hypertrophic fibrotic scarring, the trial was halted because of toxicity issues (34).
In addition to fibroblast activation and matrix remodeling enzymes, engagement of the cell-matrix adhesion and mechanotransduction machinery in CAFs is critical for ECM organization and alignment. Several integrins, the transmembrane receptors that mediate cell-matrix adhesion, have been found to play important roles in ECM alignment, including integrins α 2 β 1 , α v β 3 , and α v β 5 (35)(36)(37)(38) . The Rho-Rho kinase (ROCK) signaling pathway, which is activated by cell-matrix adhesion, and actomyosin contractility are key regulators of ECM remodeling and alignment (13,(39)(40)(41). Indeed, cellular traction forces generated by actomyosin contractility direct FN matrix assembly (42), which is needed for collagen fibrillogenesis. However, therapeutic inhibition of the mechanotransduction machinery has remained challenging. Both FAK and integrin inhibitors have entered early clinical trials; however, FAK inhibition has demonstrated high rates of toxicity, and integrin inhibitor trials have frequently been terminated because of lack of efficacy (43,44). ROCK is a promising target, yet despite two decades of research there is no clinically approved inhibitor (45)(46)(47)(48). Targeting actomyosin contractility directly is challenging, as this pathway is vital for cardiovascular function, and side effects are dose limiting.
While several molecular mechanisms regulating matrix alignment have been identified and have shown remarkable success in preclinical disease models, supporting the significant role of matrix alignment in cancer, they have not translated to clinical use. This is likely due to the ubiquitous nature of several of these pathways in normal cellular function, the difficulty in developing specific pharmacologic inhibitors, and that some of these strategies ablated the fibroblastic or ECM populations, promoting tumor progression (25,26). As an alternative approach, identifying mechanisms that normalize functional ECM topography while maintaining the fibroblast population will likely yield more tractable therapeutic targets (30,31,49).
There is an unmet need to identify new mechanisms and therapeutic strategies that promote ECM normalization as opposed to its elimination (30,31). To therapeutically restore normal ECM organization with the goal of limiting cancer cell invasion and metastasis [including metastatic reseeding (50,51)], it will be necessary to elucidate the molecular mechanisms governing ECM remodeling. To investigate ECM remodeling and fiber alignment in vitro, cells can be embedded in a three-dimensional collagen matrix, and collagen fiber organization visualized using confocal reflectance microscopy (21,52,53). This method has been used to examine how interstitial flow induces fibroblast remodeling of collagen fibers (52) and to investigate how breast cancer cells reorganize and align the ECM during invasion (21). As an alternative to providing a purified ECM, fibroblasts can be cultured to produce their own ECM in vitro, termed cell-derived matrices (CDM; refs. 1,13,20,18). Fibroblastic CDM are the gold standard model for investigating mechanisms regulating ECM alignment in vitro (1,13,20,18). CDM are produced by fibroblasts cultured in the presence of ascorbic acid to stabilize collagen incorporation into the ECM (54). Using this in vitro system, it has been shown that CAFs produce significantly more aligned ECM than normal fibroblasts, reproducing the characteristics of the in vivo TACS-3 and tumor microenvironment (1,13,20,18). Using CDM to study fiber alignment significantly reduces the experimental cost and complexity as compared with three-dimensional collagen culture by eliminating the need for the exogenous collagen matrix; however, there is less control over the properties of that matrix (composition, density, mechanical properties) and how those properties may feedback to regulate cell behavior.
While CDM have been an essential tool for studying matrix organization in a controlled and reproducible manner in vitro, current methods for generating and analyzing CDM are not compatible with high-throughput screening (HTS) approaches, limiting their ability to conduct unbiased discovery-oriented screens to identify new molecular mechanisms or novel therapeutics. The assay is experimentally labor-intensive, expensive, and cumbersome for data analysis.
Previous studies have miniaturized the production of CDM in a 96-well format to investigate the effects of a given type of CDM on cancer cell responses to assorted chemotherapies (55); however, analysis of matrix alignment has not been adapted for HTS applications. To address this need, we adapted a CDM alignment assay to a robust and reproducible HTS platform. The key advancements that enabled this platform include the development of a scaled-down experimental workflow for use in 384-well plates, streamlining of ECM visualization, and the development of an automated fiber alignment scoring system. The assay was credentialed for HTS, and a kinase inhibitor library was screened to demonstrate the use of the platform. This preliminary kinase inhibitor screen identified small molecules that target several known regulators of matrix alignment in addition to identifying compounds that target novel matrix alignment modifiers.

Fluorescent Labeling of Fibronectin
Fibronectin was labeled in-house with NHS-Fluorescein according to previously published protocols (60) with minor modifications. Briefly, 10 mg of bovine fibronectin (MilliporeSigma, catalog no. F4759) was suspended in PBS at 1 mg/mL. Fibronectin was dialyzed in PBS overnight using 8 kDa molecular weight cut-off dialysis tubing, then incubated with 125 μL of 1 mol/L sodium bicarbonate (Sigma-Aldrich, catalog no. S8875) in water, pH 9, and 125 μL of 1 mg/mL NHS-Fluorescein (Thermo Fisher Scientific, catalog no. PI-46409) in dimethyl sulfoxide (DMSO) for 2 hours at room temperature. Labeled fibronectin was separated using PD-10 desalting columns (GE Life Sciences, catalog no. 17085101) and the absorbance was measured at 280 and 494 nm for 50 μL of solution in a 96-well plate. An equivalent volume of PBS was used to correct for background absorbance. The values were corrected for a path length of 1 cm and the fibronectin concentration was calculated using the following formula (60): where CF is a correction factor adjusting for the amount of absorbance at 280 nm caused by the dye, in this case 0.3. ε is the molar extinction coefficient for fibronectin, in this case 677,800 M −1 cm −1 .
The degree of labeling of the fibronectin was calculated as follows (60): where εʹ is the molar extinction coefficient of fluorescein, in this case 68,000 For these experiments, the labeled fibronectin stock was at 0.3 mg/mL and was found to have a degree of labeling of approximately 6.

AACRJournals.org
Cancer Res Commun; 2(11) November 2022 Thermo Fisher Scientific, catalog no. P36930). CDM imaging and fiber orientation analyses were performed as described previously (61). To improve data visualization, all images in this study were pseudocolored by orientation relative to the mode angle of the fibers in that image using Adobe Photoshop (RRID:SCR_014199) to shift the color scale of each individual image such that all mode angles are displayed as the same color, cyan (59).

384-well Fibroblast-derived Matrix Production and Drug Screening
On day 0, Pten −/− MMF were seeded in a 384-well gelatin-coated plate at 50,000 cells/cm 2 (1,500 cells/well) using a multichannel pipet and 40 μL of medium per well. The plates were rocked gently on the lowest setting of an orbital shaker (Belly dancer, IBI Scientific) for 10 minutes at room temperature prior to transfer to the incubator to promote even seeding. On days 1 and 3, the media was replaced, and cells were treated with DMSO control or 1 or 10 μmol/L of a kinase inhibitor library (Cayman Chemical #10505, batch no. 0550547), 50 μg/mL ascorbic acid, and 6 μg/mL fluorescently labeled fibronectin (labeling performed as described above). The drug library was diluted 1:10 in DMSO, and subsequent dilutions were performed in cell culture medium such that the final medium contained 0.1% (v/v) DMSO regardless of drug concentration. 0.1% DMSO was used as a negative control and 10 μmol/L Y-27632 (a ROCK inhibitor in 0.1% DMSO) was used as a positive control. On day 5, the matrices were fixed by incubation with 40 μL/well of 4% (w/v) paraformaldehyde/5% (w/v) sucrose solution for 30 minutes. The fixative was removed, and nuclei were stained using 1:1,000 Hoechst 33342 in PBS for 1 hour, prior to washing the matrices three times in PBS. After the final wash, 40 μL of PBS was added to each well, and the plate was stored in the dark at 4°C until imaging.

Automated Imaging of Fibroblast-derived Matrices
Matrices were imaged at room temperature using a Nikon A1R confocal microscope with a 20× air objective (N.A. 0.75) and a 4× optical zoom. The NIS Elements software was used to automate image acquisition. Z-stacks were taken through an 8 μm depth at 1 μm intervals using the microscope's perfect focus system to set the focal point of the matrix at the center of the Z-stack. Four locations were randomly generated within a 0.8 mm restricted radius and imaged in each well.

Automated Image Analysis
Images were exported as individual Tag Image File Format (TIFFs), and a custom MATLAB (R2018a, RRID:SCR_001622) code was written to automate image analysis for quality control, fiber alignment, and nuclei number, with Im-ageJ (v1.50e, RRID:SCR_003070) running under the command of MATLAB using the Miji plugin (v1.3.6, downloaded from http://bigwww.epfl.ch/sage/ soft/mij/; ref. 62). Briefly, images were imported into MATLAB, passed to Im-ageJ, and z-projected using the maximum intensity of each pixel. As a quality control for insufficient fibrillogenesis or imaging errors in each field of view (FOV), the percentage of pixels in each z-projection that fell below the threshold of the background signal [100 arbitrary units (a.u.) of fluorescence intensity] was quantified. This background level was chosen for this study based on images that had little to no matrix present and may need to be adapted for other imaging setups. Any FOV with >10% of pixels at or below the background was removed from further analysis. This cutoff was chosen by analyzing the distribution of background pixel percentages across all images in the study. In addition, wells that had two or more (of four total) FOVs flagged for exclusion were excluded from further analysis. Finally, drug treatments that resulted in the exclusion of two or more replicate wells were excluded from the final hit list. To quantify the number of nuclei per image, the Hoechst channel images were z-projected, and a Gaussian blur was performed with σ = 2. The images were thresholded and watershed segmentation was performed to separate overlapping nuclei. The number of nuclei in each image was counted and compared with the control. Wells with <40% of the nuclei of the control were considered to have high toxicity which could affect matrix production and were excluded from subsequent analysis.
To quantify fiber alignment, each stack of the fibronectin channel images was z-projected and made into a 32-bit image, and the plugin OrientationJ was used to create a histogram of the fiber alignment, which was imported back to MAT-LAB and stored. The histogram was centered at its mode, and the fraction of fibers within 20 degrees of the mode was calculated as a metric of the fiber alignment. The average of the four locations within each well was taken as the final alignment value for that well.

Determination of Hits
The robust Z-score was calculated for each screen according to the following formula: whereỸ N is the median value and MAD N is the median of absolute deviation from the negative control (63). A composite Z-score was calculated by averaging the values from each experiment (n = 3), and hits were ranked according to this value.

Stromal Gene Expression and Patient Outcomes Analysis
The correlation between gene expression in ductal breast carcinoma stroma and patient 5-year survival data for the Boersma and colleagues cohort (64) was downloaded from Oncomine (oncomine.org). To confirm the prognostic significance of the selected genes in ductal breast carcinoma samples or specifically in ductal breast carcinoma stroma, overall survival and gene expression data from The Cancer Genome Atlas (TCGA) Breast Invasive Carcinoma and Boersma ductal breast carcinoma stroma cohorts were collected from the cBio Cancer Genomics Portal (65,66) and the Gene Expression Omnibus GSE5847 (GEO, RRID:SCR_005012). The respective cohorts were stratified into high (upper 50%) and low (lower 50%) expression groups for each gene, and Kaplan-Meier curves were generated. Differences in survival were determined by log-rank tests.

Statistical Analysis
Statistical analyses were performed using GraphPad Prism v8 (RRID:SCR_002798). For comparisons between multiple groups, one-way ANOVA was performed, followed by Tukey multiple comparisons posttest.
To determine whether the drug library used was skewed toward alignment inhibitors, a Shapiro-Wilk normality test was used on the composite robust Z-scores. Differences in patient survival were assessed using unpaired twotailed t tests. Differences in patient survival curves were assessed using the Mantel-Cox log-rank test. Differences were considered statistically significant at P < 0.05.

Data Availability
Robust Z-scores for the entire inhibitor screen are presented in Supplementary Table S1. The MATLAB code for data analysis is available at GitHub: https://github.com/jtsharick/HighThroughput_FDM_Screen/ blob/main/HT_Matrix_Analysis_Paper_FINAL.m. Gene expression data were collected from Oncomine and GEO (GSE5847). Additional data generated in this study are available from the corresponding author upon request.

Matrix Alignment Assay is Suitable for Adaption to HTS
CDM faithfully recapitulate the organizational characteristics of the in vivo stroma from which the fibroblasts originated (13,35,67). Specifically, the CDM produced by CAFs are highly organized and aligned parallel (i.e., anisotropic), whereas the CDM produced by fibroblasts isolated from normal tissue are disorganized (i.e., isotropic). However, the inherent heterogeneity of primary cell cultures and the limited time in culture are significant limitations for HTS design and execution. Adaptation of the CDM assay to work with well-established immortalized cell lines is a key step toward the development of an HT CDM platform. To this end, we selected the mouse embryonic fibroblast cell line NIH/3T3 and tested its compatibility with the CDM assay, as it is one of the most widely used fibroblast cell lines in biological studies and amenable to expansion for genetic and compound HTS campaigns. Using the low-throughput 24-well CDM assay as an initial step to examine the suitability of the assay for HTS, NIH/3T3 cells were stimulated with 0.1 to 10 ng/mL of TGFβ1 to induce CDM alignment to a similar degree as that observed in diseased tissue or with CAFs (refs. 1, 13, 20, 18; Fig. 1). TGFβ is a critical factor responsible for fibroblast activation in cancer (68,69) and matrix alignment (70). Furthermore, in previous reports, TGFβ1-treated fibroblasts significantly increased collagen alignment in vivo in a murine head and neck squamous cell carcinoma xenograft model, whereas control fibroblasts had no effect on alignment (70). Consistent with these studies, NIH/3T3s treated with 1-10 ng/mL TGFβ1 produced highly aligned matrices, as indicated by the increased fraction of fibers with an alignment angle within 20 degrees of the mode angle, as compared with the control cells (Fig. 1). To examine the suitability of the CDM for HTS, the Z'-factor, which indicates the degree of separation of the positive and negative signals within the assay, and the coefficient of variance were determined. For the control and 1 ng/mL TGFβ1-treated matrices, the Z'-factor was greater than 0.5, and the coefficient of variation was 3%-4%, indicating sufficient separation of the positive and negative signals, low variance, and feasibility for adaptation to HTS (71).

Miniaturization and Streamlining of the Matrix Organization Assay for a HTS Platform
To develop a HTS screening platform for the CDM alignment assay, the assay was miniaturized to a 384-well format. Black film bottom microplates were used to provide a flat, thin surface necessary for high content imaging (Fig. 2).
The plates were coated with gelatin prior to cell seeding to promote CDM adhesion to the well and prevent detachment during media changes and processing steps. After addition of the cells to the plate, gentle agitation at the slowest speed on an orbital shaker for ten minutes at room temperature ensured even cell seeding within the wells. The fiber organization of CDM is typically visualized through immunofluorescent labeling of fibronectin or collagen. To eliminate these costly and time-consuming immunostaining steps and to enable visualization of the matrix for the screening platform, cells were cultured with exogenous fluorescein-labeled fibronectin. Previous studies have demonstrated the incorporation of fluorescent fibronectin into CDMs (72,73), and the addition of exogenous fluorescent fibronectin has previously been used to study fibronectin fibrillogenesis (60,74). Furthermore, fibronectin is oriented in an orientation (Fig. 3). The extent of CDM alignment is indicated by the fraction of fibers distributed around the mode angle, as illustrated by the highlighted region in the histogram (Fig. 3). However, further data processing is needed to center the histogram and produce a quantitative fraction of the total fibers aligned within a certain degree (typically 10 or 20 degrees) of the mode angle.
To automate image and data analysis for quality control, fiber alignment, and nuclei number, a MATLAB code was written (Fig. 3), with ImageJ running under the command of MATLAB using the Miji plugin (Supplementary Data S1; ref. 62). Confocal microscopic images were exported as individual TIFFs and the image files were imported into MATLAB and passed to ImageJ. The stack was z-projected and made into a 32-bit image, and the plugin OrientationJ was used to create a histogram of the fiber alignment, which was imported back to MATLAB and stored. To automate the data analysis process, the MATLAB code identifies the mode of the orientation angle (the angle that appears the most often in the fiber distribution), centers the histogram at the mode, assigns the mode to "0," calculates the percentage fraction of fibers within 20 degrees of the mode to provide a quantitative measure of fiber alignment, and exports the data to a table in Microsoft Excel. The average of 3-4 locations within each well was taken as the final value of alignment for that well. To quantify the number of nuclei per image, the Hoechst channel images were z-projected, and a Gaussian blur was performed with σ = 2. The images were thresholded and watershed segmentation was performed to separate overlapping nuclei. The number of nuclei in each image was counted and compared with the control, for a final metric of percent of control nuclei (Fig. 3).The custom MATLAB code is available on GitHub: https://github.com/jtsharick/HighThroughput_FDM_Screen/ blob/main/HT_Matrix_Analysis_Paper_FINAL.m

Kinase Inhibitor Screen Identifies Novel Regulators of Matrix Alignment
To test the discovery potential of the newly developed screening platform, we analyzed a commercially available library of 154 small-molecule kinase inhibitors (Cayman Chemical). This library is enriched in compounds targeting lipid, receptor, and non-receptor tyrosine and serine kinases, and targets more than 70 protein kinase families. Kinases are involved in a variety of signaling pathways known to affect matrix alignment, including cellular contractility (61,76), matrix adhesion (35), and fibroblast activation (77). In addition, kinases constitute nearly a quarter of the Lipinski-druggable genome (78), making them readily targetable by pharmacological inhibitors. We predicted that screening this library would identify compounds targeting known and novel regulators of matrix alignment.
Pten null murine fibroblasts were selected as the screen cells for this assay based on prior work, which demonstrated that these cells produce an aligned matrix in vitro and in vivo (61). Low PTEN expression in stromal fibroblasts has been associated with poor outcomes in multiple human tumors, including breast, prostate, pancreatic, and endometrial cancers (56,58,(79)(80)(81)(82)(83)(84). In vivo ablation of Pten (Fsp-Cre;Pten loxP/loxP ) in fibroblasts of mouse mammary tissue led to increased collagen deposition, collagen alignment, and tumorigenesis in the presence of the Neu oncogene (56,61). In vitro, CDM produced by Pten null (Pten −/− ) murine mammary fibroblasts results in a highly organized, aligned matrix compared with the largely disorganized matrix produced by Pten intact (wild-type) fibroblasts (61,73). As such, these cells are well suited for screening compounds that can "normalize" a highly aligned matrix.

Assay Performance
To ensure that the assay results were unaffected by solvent or plate positioning artifacts, Pten −/− murine fibroblasts were treated with the DMSO control or two different concentrations (1 and 10 μmol/L) of Y-27632, a ROCK inhibitor previously shown to significantly reduce matrix alignment in Pten −/− fibroblasts (61), in a single column across an entire 384-well plate. Matrix alignment was assessed by determining the fraction of fibers falling within 20 degrees of the mode matrix orientation. No significant differences were observed in the average values of any column with the same treatment (Fig. 4A). However, a dose-dependent decrease in matrix alignment with Y-27632 treatment was observed, indicating that the assay can reliably measure the effects of drug treatment on matrix alignment independent of plate position. Next, the effects of DMSO concentration, a common library solvent, on matrix alignment were examined. No significant differences in alignment were found across the 0%-2% range tested (Fig. 4B).

Assay Execution
Pten −/− fibroblasts were plated in 384-well plates on day 0 at 5 × 10 4 cells per well (to assure cells were confluent the next day), and the cell culture medium was replaced on days 1 and 3 to include ascorbic acid, fluorescently labeled fibronectin, and the kinase inhibitors or vehicle control. DMSO was used as a vehicle control at a final concentration of 0.1% and 10 μmol/L Y-27632 was used as a positive control to reduce matrix alignment. Y-27632 was used as a positive control for a reduction in matrix alignment, to mimic the disorganized matrix produced by normal cells as compared with the aligned matrices produced by CAFs (1,18,35). A total of 154 inhibitors in the Cayman kinase inhibitor library were screened at concentrations of 1 and 10 μmol/L. On day 5, the samples were fixed, and the nuclei were stained. Confocal images of the CDM and nuclei were FIGURE 3 Data analysis pipeline. Image analysis was automated using a MATLAB script. ImageJ/FIJI was run under the control of MATLAB using the Miji plugin. Images were imported to MATLAB, passed to ImageJ, and Z-projections of the nuclei channel were thresholded, made binary, segmented, and counted. Z-projections of the fibronectin channel were analyzed using OrientationJ, which outputs a histogram of fiber orientation. This output is passed back to MATLAB and the histogram is centered at the mode. The fraction of fibers within 20 degrees of the mode was determined as a metric of matrix alignment.
obtained using the automated imaging setup described above. The screen was repeated on three separate days.
The purpose of the HT CDM platform is to identify regulators of matrix organization; therefore, sufficient matrix must be present to measure matrix alignment. Toward this goal, wells with insufficient matrix, due to factors such as cell toxicity, deficient ECM production or fibrillogenesis, or imaging errors, were identified and excluded from further analysis (Fig. 5). Wells with insufficient matrix were defined as fields where the image contained more than 10% of pixels at the "background" level, where the background level in this study was defined as a region having an intensity level of 100 a.u. or less. The background level was determined by measuring the signal in portions of the imaged FOV that did not contain any matrix. Images that did not meet this threshold often lacked a matrix, large holes in the matrix, or imaging-related errors (e.g., out of focus, misalignment between the objective and matrix; Fig. 5A). Using these exclusion criteria, greater than 75% of the images contained sufficient matrix for further analysis of matrix alignment (Fig. 5B). Four FOV were imaged per well and at least three FOV were required to pass the above exclusion criteria for further matrix alignment analysis in that well. In addition to sufficient matrix, wells with high cell toxicity were excluded if the nuclei count was less than 40% of the control. On the basis of these exclusion criteria for insufficient matrix and cell toxicity, 229 of 924 total wells from all three screens were excluded from further analysis, with 152 wells excluded for insufficient matrix, 4 wells excluded for cell toxicity, and 77 that failed both criteria (Fig. 5D). Of the 288 excluded wells, 78% were from the conditions with the 10 μmol/L concentration of the kinase inhibitor.
Matrix alignment was quantified using the custom MATLAB script to determine the fraction of fibers falling within 20 degrees of the mode orientation angle of the matrix. A robust Z-score was calculated from the average matrix  alignment across the FOV for each condition in the screen. The robust Z-scores were then averaged across two or three screens to determine the hit ranking, excluding conditions that did not have sufficient FOV in two or three screens.
The results from the screen were ranked according to the average robust Zscore and included compounds that promoted (Z > 0, where 0 represents the DMSO control, n = 109), and which suppressed alignment (Z < 0, n = 212), although the distribution was significantly skewed with more inhibitors reducing alignment ( Fig. 6; Supplementary Fig. S1; Supplementary Table S1). This skew is likely library specific and may change with the composition of each library. The S-curves of the robust Z-scores from each of the three trials showed similar results ( Supplementary Fig. S2). Compounds were considered hits if the average robust Z-score was less than or equal to the known inhibitor Y-27632 (1 μm), resulting in 27 compounds and a hit rate of approximately 8% (Fig. 6A  in cancer cell morphology and migration in vitro. In addition, a 50% reduction in matrix alignment is similar in magnitude to the differences reported between normal cells and CAFs in the literature (1,18,35). The compounds did not need to meet this criterion at both concentrations (1 and 10 μm) to be considered a potential lead compound.

Kinase Inhibitor Screen Identifies Novel Regulators of Matrix Alignment
The hits identified from the screen fell into several broad categories, as depicted in Fig. 6C. Many of the top hits were related to PI3K and protein kinase B (PKB)/Akt signaling, including inhibitors targeting Akt (ARQ-092), GSK3β (kenpaullone), mTOR (PP242 and Torin 1), and PIM kinases (AZD 1208). MAPK signaling inhibitors were also among the top hits, including compounds targeting mitogen-linked kinases (URMC-099), ERK-1 (SC-1), Raf-1 (sorafenib), and p38 MAPK (SB 202190). Identification of these pathways is likely due to the use of Pten −/− fibroblasts, as PTEN is a known regulator of PI3K/Akt and MAPK signaling. In addition, the screen identified several inhibitors that target known modulators of matrix alignment related to cell-matrix adhesion and actomyosin contractility, including myosin light chain kinase (ML-9), focal adhesion kinase (PF-562271), and Rho kinase (Y-27632  White arrow indicates a hole in the matrix, which signifies poor fibrillogenesis. B, Histogram of percentage of pixels below background for every FOV imaged in this study (n = 4,336). C, To ensure sufficient fibrillogenesis and imaging quality for alignment analysis, the fraction of pixels in the fibronectin channel that fall below background noise levels (100 a.u) is quantified for each FOV. Those with >10% of pixels below background were flagged and excluded from subsequent analysis. Drug screen wells containing less than three valid FOV were excluded from subsequent analysis, and drugs with 0 or 1 valid wells across three replicate plates were also excluded. Images in C show FDM treated with 1μmol/L of PP-242 (hit #13 in kinase inhibitor screen). D, Of 924 total wells analyzed in this study, 152 were excluded for insufficient matrix, four were flagged for high toxicity, and another 77 were flagged for both conditions.

Screen Hits Correlate with Patient Survival
The HT in vitro matrix alignment kinase inhibitor screen suggested several novel regulators of matrix alignment. However, selecting the pathways to pursue further investigation and possible translational potential is not straightforward. One way to further narrow down these potential pathways is to use gene expression data from patient tumor samples to correlate pathways identified from the screen with patient outcomes.
Commonly utilized patient datasets, such as TCGA and METABRIC, focus on gene expression in tumor cells, and relatively few datasets exist that specifically examine the association between stromal gene expression and patient survival. However, several groups have generated datasets from patients with breast cancer, where the stroma surrounding the tumor was isolated using laser capture microdissection, and gene expression specifically in the stromal compartment was determined via microarrays (64,95). Using the dataset from Boersma and colleagues (64), we correlated the expression of the predicted gene targets from the hits of the inhibitor screen with breast cancer patient outcomes. The Boersma study investigated gene signatures of inflammatory breast cancer (15 patients) and invasive, noninflammatory breast cancer (35 patients) with poor disease outcomes (less than 5-year survival). The average overall survival for patients above and below the median gene expression was compared using a t test during a 5-year follow-up period. Several genes appeared highly associated with survival including AKT, PRKCA, MET, ERBB, NTRK, and MYLK (genes with significant changes in Table 2, complete analysis in Supplementary Table S2). ERBB has previously been identified as a critical mediator of fibrosis in several different systems (96)(97)(98)(99). Neurotrophins and their receptors (NTRK) are expressed in fibroblasts and have been found to play a role in dermal fibroblast activation to myofibroblasts and contribute to fibrosis in a number of tissues (100,101).
The prognostic significance of these six genes was then evaluated more rigorously using the Kaplan-Meier analysis (Fig. 7). For each gene, the patient population was stratified according to high or low expression levels along the median. The resulting curves were statistically compared using the log-rank test. High expression of MET and NTRK were associated with a significant decrease in patient survival. MYLK was not significant at α = 0.05; however, it showed a clear trend of decreased patient survival with high expression (P ∼ 0.07). A similar Kaplan-Meier analysis was performed for these genes using data from TCGA invasive breast carcinoma cohort ( Supplementary Fig. S3).
These analyses showed similar trends to the analyses of the breast cancer stroma samples but were not as strongly prognostic.

Discussion
Expression of several genes whose protein products are inhibited by hits from this screen appeared to be associated with poor patient outcome. c-MET is a receptor tyrosine kinase that is commonly upregulated in a wide variety of carcinomas and is activated in response to paracrine hepatocyte growth factor and η) whose individual contributions remain unclear (105). Interestingly, several PKC isozymes have been shown to play a role in human dermal fibroblast activation by TGFβ (106), suggesting that PKC may contribute to an activated CAF phenotype. PKCα was chosen for its association with patient survival in this study as it is most strongly inhibited by each of the inhibitors found as hits in this screen, although other isozymes may also play a role. A limiting factor of the analysis of the stromal specific dataset is that, despite the wealth of data showing that the stroma can greatly impact cancer progression, few datasets exist that specifically compare stromal gene expression with patient outcome. The stromal dataset examined here contained approximately 50 patients, which has limited statistical power to examine the association between gene expression and patient survival. Analysis of data from TCGA with a larger cohort showed similar trends to the analyses of the breast cancer stroma samples but were not as strongly prognostic, likely because TCGA samples are primarily comprised of tumor epithelia with limited amounts of associated stroma. Therefore, it is possible that hits in this screen influence patient survival but were not captured in this study. Future analyses using more and/or larger datasets may identify additional genes associated with patient outcomes related to the hits identified by this screen. Furthermore, overall survival is complex, and it is perhaps not surprising that more of the hits from a phenotypic screen of kinase inhibitors (in which many inhibitors have broad spectrum effects) did not directly translate to gene expression and correlation with patient survival. Further validation studies will be needed to identify the specific genes, proteins, and pathways that regulate matrix organization and eventually patient outcomes.
Several clinical trials are already in place for several of the compounds identified by this screen, supporting the potential of this approach to identify translatable targets (107)(108)(109). Two MET inhibitors, crizotinib and cabozantinib, are FDA approved for NSCLC and renal cell carcinoma, respectively, and a wide variety of other MET inhibitors are in clinical trials (110). The data presented here highlight the need to investigate the impact of these inhibitors on the tumor microenvironment and matrix organization. In addition, while specific kinases have been described in association with these inhibitors, small molecules can have off-target effects, and further studies are needed to conclusively prove the role of specific kinases and pathways in matrix alignment. Further studies of the effects of these inhibitors on normal fibroblast matrix production will also be critical, as normal fibroblasts can have an antitumorigenic effect, in part through production and maintenance of a tumor suppressive ECM (111).
Low stromal PTEN expression has previously been identified as a contributing factor to matrix alignment both in vitro and in vivo (61). Pten −/− fibroblasts were used as a model system to produce highly aligned matrices for kinase pharmacologic inhibitor library screening to identify modulators of matrix alignment. A variety of signaling pathways were suggested on the basis of hits from the screen, including several PI3K/AKT pathway inhibitors, which may reflect the Pten knockout model used to induce matrix alignment as PTEN normally acts as a negative regulator of the AKT pathway. A genetically engineered Pten null mouse model is already established and well characterized (56,57,61), enabling follow-up in vivo tumor studies to assess how hits from the screen impact matrix alignment in vivo, as well as disease progression (i.e., tumor growth, invasion, metastasis) and other components of the tumor microenvironment (i.e., suppressive effects of normal fibroblasts, angiogenesis, immune response).
It will also be important to determine whether the inhibitors and pathways identified here indicate broadly applicable mechanisms of matrix alignment or are specific to the Pten −/− model system. On the basis of the characterization studies presented here, the NIH/3T3 system stimulated with TGFβ1 is a promising alternative system that may be more broadly applicable. In addition, while two murine cell lines were used in this work, it will be important to validate these studies with human fibroblasts and primary human tumor CAFs, cell types which have previously been shown to be suitable for the production of CDMs and investigation of matrix alignment (19,112,113).
In addition to the pharmacologic inhibitor screening conducted in this study, the HTS matrix alignment platform can be easily adapted to other types of screening. Genetic screening (e.g., siRNA, short hairpin RNA, CRISPR-CAS9) is a powerful tool for identifying genes and pathways that are important for a given biological phenomenon. To adapt the HTS matrix alignment platform for genetic screening, the fibroblasts could be genetically modified in a 384-well plate prior to CDM production, and then the HTS alignment assay and analysis conducted. Future iterations of this platform could also be expanded to include high-content imaging, such as quantification of fibroblast activation markers like fibroblast activation protein (refs. 35, 114). In addition, while the HTS platform detailed here focused on ECM fiber organization and alignment, other ECM characteristics could be easily quantified within this framework, such as the overall amount of ECM (through fluorescence intensity or fiber density) as well as ECM fiber characteristics such as width and length. Increases in collagen fiber density, width, and length have been found to correlate with poor overall survival in patients with gastric cancer (115). Furthermore, collagen fiber length is an independent prognostic factor in head and neck and colorectal cancers and correlates with poor clinical outcomes in esophageal cancer (70).
The organization of ECM fibers in a tissue is a fundamental property critical for tissue and cell function. Changes in this organization, such as the fiber alignment observed in many cancers, can disrupt homeostasis and contribute to disease progression. Matrix alignment around tumors correlates with poor patient outcomes in many cancers, including breast cancer (8), PDAC (11), gastric cancer (115), and head and neck, colorectal, and esophageal cancers (70), indicating the importance of this phenomenon and the potentially broad impact of therapeutic agents capable of normalizing alignment. However, previous attempts to target ECM alignment in cancer have shown little clinical benefits. Thus, new methods are needed to identify additional modifiers of matrix alignment. Here, a novel screening platform was developed to enable high-content and HT analyses of ECM fiber alignment, which will facilitate unbiased highcontent screening campaigns and accelerate discovery in this area. Using this platform, several novel regulators of matrix organization were identified that were correlated with poor breast cancer patient outcomes. In particular, c-MET and NTRK1 were identified as regulators of matrix alignment and may serve as novel targets within the cancer stroma. Further efforts to screen different classes of proteins beyond kinases will be important to reveal novel signaling pathways that regulate matrix alignment.