Chemical regulators of epithelial plasticity reveal a nuclear receptor pathway controlling myofibroblast differentiation

Plasticity in epithelial tissues relates to processes of embryonic development, tissue fibrosis and cancer progression. Pharmacological modulation of epithelial transitions during disease progression may thus be clinically useful. Using human keratinocytes and a robotic high-content imaging platform, we screened for chemical compounds that reverse transforming growth factor β (TGF-β)-induced epithelial-mesenchymal transition. In addition to TGF-β receptor kinase inhibitors, we identified small molecule epithelial plasticity modulators including a naturally occurring hydroxysterol agonist of the liver X receptors (LXRs), members of the nuclear receptor transcription factor family. Endogenous and synthetic LXR agonists tested in diverse cell models blocked α-smooth muscle actin expression, myofibroblast differentiation and function. Agonist-dependent LXR activity or LXR overexpression in the absence of ligand counteracted TGF-β-mediated myofibroblast terminal differentiation and collagen contraction. The protective effect of LXR agonists against TGF-β-induced pro-fibrotic activity raises the possibility that anti-lipidogenic therapy may be relevant in fibrotic disorders and advanced cancer.

the cells by adding 5 µl of a 1:100 dilution of the DMSO solution in medium using the 'Freedom Evo' pipetting robot (Tecan Group Ltd, Männedorf, Switzerland) equipped with a 96-head piece and with disposable tips (final concentration was depending on the library, but in most cases 10 µM or 16.6 µM). As positive controls GW6604 or LY-364947 were added in five concentrations (0.1, 0.3, 1.0, 3.0 and 10.0 µM) and staurosporine was added in two concentrations (0.001 and 0.1 µM). As negative control DMSO was added (0.1% v/v, n=29).
The cells were washed 72 h after seeding (48 h after compound addition) with PBS using the plate washer 'PW384' (Tecan Group Ltd, Männedorf, Switzerland) and fixed for 30 min by addition of para-formaldehyde (final concentration 4% (w/v) in PBS).
The proof-of-concept screen in 96-well format was done by seeding cells (3,000 in 100 µl/well) without TGF-β, replacing the medium 24 h later with medium containing TGF-β and adding 48 h after seeding seven concentrations of compounds in 50 µl of a 1:3 dilution of the DMSO solution (final concentration 41 nM -30 µM) using a manual multi-channel pipette.
As positive control GW6604 (3.3 µM) was added and as negative control DMSO was added (0.3% v/v, n=6). Ninety six hours after seeding (48 h after compound addition) the cells were washed once with PBS using a manual 8-needle adaptor and an automated multi-channel pipette. After fixation for 30 min with para-formaldehyde (final concentration 4% (w/v) in PBS), cells were washed twice with PBS. All further steps were performed manually in the same way as described further down for the robotic liquid handling, except that multi-channel pipettes and the 8-needle adaptor were used instead of the robotic automation.
For the robotic immunostaining procedure in 384-well plates the above mentioned automation (dispenser, pipetting robot and plate washer) were used. For immunostaining of fibronectin cells were permeabilised for 10 min with 1% Triton X-100 in PBS, blocked for 1 h with blocking solution (5% FBS/0.1 mM glycine in PBS), and incubated with the primary antibody (anti-fibronectin rabbit polyclonal, cat# F3648, Sigma-Aldrich Sweden AB, Stockholm, Sweden) using a dilution of 1:1,000 in blocking solution. After washing with PBS, the cells were incubated with a secondary antibody (anti-rabbit Alexa-568, cat# A10042, Molecular Probes, Life Technologies Corp., Foster City, CA, USA) using a dilution of 1:500 in blocking solution. Finally, the cells were washed again with PBS, and nuclei were stained with 1 µg/ml 4',6'-diamidino-2-phenylindole (DAPI) and sodium azide was added (0.02% w/v in PBS) to prevent contamination. Imaging plates were acquired automatically on the HCS microscope 'ArrayScan' (Cellomics/ThermoScientific, Thermo Fischer Scientific Inc., Waltham, MA, USA) using a 10× objective (NA = 0.3) taking 9 fields with 2 channels; for DAPI the filter was set to XF93-Hoechst and for fibronectin-568 the filter was set to XF93-TRITC.
For the counter-screen, cells were seeded without TGF-β, and 24 h later, they were treated with compounds for 2 h, before TGF-β was added to the cells. After 1 h of TGF-β treatment cells were fixed and stained as described above using the anti-p-Smad2 (rabbit polyclonal antibody, home-made, LICR-Uppsala, Sweden) 1,2 as a primary antibody. Plates were imaged as described above.
Image and data analysis of the high-content screen. Images of HaCaT cells taken on the 'ArrayScan' HCS platform were automatically analysed using the module 'Morphology Explorer' of the 'BioApplications' software package (Cellomics/ThermoScientific, Thermo Fischer Scientific Inc., Waltham, MA, USA). Two algorithms were optimised; the first was segmenting single nuclei, the second segmenting groups of cells as colonies. In total, nearly 500 measurement parameters were calculated. With the help of the open-source data analytics software KNIME 4 and the plug-ins HCS-Tools and Scripting Integration, that were necessary to handle screening data 5 , all parameters were evaluated during assay development for robustness, repeatability, and redundancy (manuscript in preparation). Finally, 18 parameters were selected to generate a multi-parametric profile for each compound (Fig. S3). Next the measurements of each plate were normalised by calculating the robust percentage-of-control (.poc) 6 : x rcpm .poc = x rcpm /median(x m [DMSO] p )*100, where x rcpm is the measurement x of parameter m of a well in row r, column c, and plate p, and x m [DMSO] p are the measurements x of parameter m of all DMSO negative control wells per plate p (here n = 29). Then all parameters were normalised for their distribution by calculating the z-score (.zscore): , where x rcpm is the plate normalised measurement x (.poc) of parameter m of a well in row r and column c and plate p, x m [DMSO] are the plate normalised measurements x (.poc) of parameter m of all DMSO negative control wells of the screen (here n = 5133), and sd is the standard deviation.
Last, a profile normalisation was done for clustering the data (.lennorm): x rcp .eucldist = √(x rcpm1 2 + x rcpm2 2 + … + x rcpmn 2 ), where x rcp .eucldist is the Euclidian distance x of the profile of a well in row r, column c, and plate p, and x rcpm is the plate normalised, zscore measurement x (.poc.zscore) of parameter m of a well in row r, column c, and plate p, with m 1 as the first parameter, and m n as the last parameter of the profile (here n = 18).
x rcpm .lennorm = x rcpm /x rcp .eucldist, where each parameter m is normalised to the Euclidian distance of the profile.
The profile normalised data were used to cluster all data with the k-means algorithm and evaluation of all control wells indicated that k = 5 performed best for this data set. The cluster 1, containing most of the EMT positive control wells, was filtered for weak phenotypes using a threshold of 15 on the Euclidian distance. Control wells and wells missing an annotation were removed from the list of hits and after aggregation of repeated plates 93 hits were defined.
The counter-screen was analyzed with the module 'CytNuc Translocation' of the 'BioApplications' and only one measurement parameter was selected for the p-Smad2 channel (CytNuc_MEAN_CircRingAvgIntenRatioCh2). The analysis of the verification and the validation screen was done as described above, except that the p-Smad2 parameter was considered separately. Taking these data together, a final list of 16 compounds, as shown in Fig. S2D, was obtained.
AG1523 cells were trypsinised, counted and seeded into a 1 mg/ml Type I collagen solution (PureCol, Advanced BioMatrix) in serum-free medium at a concentration of 1×10 5 cells/ml.
The collagen/cell suspension was vortexed, and 1 ml per well was added to the BSA-coated dishes and the solution was allowed to polymerise for 45 min at 37°C. Fresh medium containing the indicated treatment conditions was then added to the solidified collagen gels and plates were returned to the incubator. Collagen gel contraction was monitored over a period of 72 h and the surface area of contracted gels was measured using ImageJ software. In certain experiments, the protocol was modified by first transfecting MEFs with empty or LXRα expression vector for 24 h prior to trypsinising and counting the cells for the assay. In these experiments cells were maintained in full serum throughout and treated accordingly.
The  E-cadherin expression measured by cytoplasmic intensity was only slightly increased after treatment with GW6604 for 48 h revealing a worse signal-to-noise ratio than fibronectin. (C) As a poof-of-principle screen, a set of 78 compounds (mostly protein kinase inhibitors) were screened at 8 concentrations (4.6 nM -10 µM) using the high-content screening setup    selected not selected p-Smad2 influence confirmed tw ice, selected compound confirmed tw ice, conflict of interest confirmed tw ice, behaves like TGF R inhibitor confirmed tw ice, p-Smad2 unclear confirmed once, selected compound confirmed once, compound not selected confirmed once, behaves like TGF R inhibitor confirmed once, p-Smad2 unclear not confirmed, tw ice in EMT cluster not confirmed, different phenotype not confirmed, different/no phenotype not confirmed, no phenotype  The table lists all 93 primary hits with their library code names (1 st column), recovery in up to three independent repeats of the primary screen (A, B, C, where the number indicates the cluster after k-means clustering and 1 was the EMTspecific cluster, 2 nd to 4 th columns), phenotypic strength as Euclidian length of the parameter profile in each or the repeats (where > 15 was considered to be a hit, 5 th to 7 th columns), recovery in the two verification screens (where the number indicates the cluster after k-means clustering and 0 was the EMT-specific cluster, 8 th and 9 th column) and corresponding phenotypic strengths (where > 7.5 was considered to be a hit, 10 th and 11 th column) and description of the hit verification (12 th column). The colour coding of the latter (12 th ) column is explained below the table. The same 93 compounds with their code names are listed again (13 th column) and the colour code explained below the table indicates behavior of the compounds in the secondary screen where compounds influencing p-Smad2 nuclear localisation were discriminated. The mean scores (14 th column) and ranges (15 th and 16 th columns) of the cytoplasmic to nuclear ratios quantified for each compound are finally shown.  Fig. S5, 12 th column, and the small table below. The percent of compounds (out of 93 total) that corresponds to each colour is: 27% of hits were confirmed twice (dark orange), 11% of hits were confirmed once (orange), 5% of hits were not confirmed, but were grouped twice in the EMT cluster (light orange), 17% of hits were not confirmed due to absence of phenotype (green), and 40% of hits were not confirmed due to different phenotype (blue). ( Immunoblot for the indicated proteins and for Gapdh, which serves as protein loading control.