Characterization of twenty-five ovarian tumour cell lines that phenocopy primary tumours

Currently available human tumour cell line panels consist of a small number of lines in each lineage that generally fail to retain the phenotype of the original patient tumour. Here we develop a cell culture medium that enables us to routinely establish cell lines from diverse subtypes of human ovarian cancers with >95% efficiency. Importantly, the 25 new ovarian tumour cell lines described here retain the genomic landscape, histopathology and molecular features of the original tumours. Furthermore, the molecular profile and drug response of these cell lines correlate with distinct groups of primary tumours with different outcomes. Thus, tumour cell lines derived using this methodology represent a significantly improved platform to study human tumour pathophysiology and response to therapy.


Supplementary Figure 9 Copy number variations (CNV) of OCI cell lines are similar to ovarian tumors
In order to compare CNV patterns of OCI cells with the human ovarian tumors, we downloaded the TCGA CNV data from 100 randomly selected human ovarian tumors.
This comparison revealed that the overall CVN trends are very similar between OCI lines and TCGA ovarian tumors. In both data sets CNV trend is copy number gain in chromosomes 2, 19, 20 and copy number loss in chromosomes 4, 9, 13, 15, and 18. While copy number losses were predominant in short arm of chromosomes 3 and 8, gains are predominant in the long arm, and this pattern was replicated in OCI lines.
These results are consistent with the LOH comparison between OCI lines and their matched tumors and indicate that the genomic landscapes of primary tumors are preserved in OCI lines. The copy number analysis was performed using the Molecular Inversion Probe (MIP) 330k microarrays from Affymetrix on all 25 OCI lines. The results were merged into a copy number profile by using an algorithm that sums together the mean log2 intensity values in overlapping intervals. The red peaks indicate copy number gain and blue peaks indicate copy number loss. The complete dataset is available at GEO, accession number GSE40786.

Supplementary Figure 10
The Cluster 1 OCI cell lines are associated with distinct pathways The 558 transcripts up-regulated in Cluster-1 are associated with 37 core pathways in IPA (p < 0.05). For the analysis depicted in this figure we used Ingenuity Pathway Analysis (IPA) software to organize the 823 genes that are differentially expressed between Cluster 1 vs. in Cluster 2 (p < 0.05) (Figure 3). See Supplementary Table 10 for the top 25 pathways, and online Excel data file 3 for the full list of IPA enriched pathways.

Supplementary Figure 11
The Cluster 2 OCI cell lines are associated with distinct pathways The 265 transcripts up-regulated in Cluster-2 are associated with 37 core pathways in IPA (p < 0.05). See Supplementary Table 11 for the top 25 pathways and online Excel data file 3 for the full list of IPA enriched pathways.

Protein expression profiles of OCI cell lines are reproducible between replicates
The heatmap profiles derived from unsupervised clustering of data from RPPA analysis of OCI cell lines. Each column depicts a different antibody and each row depicts an individual replicate from each cell line (n = [3][4]. Tumor types are depicted in different colors; papillary serous (green), non-serous (orange). In order to ensure the reproducibility of the cell line phenotypes the protein extracts were prepared in triplicates from two different passages in two separate experiments. We observed that replicates of each cell line from different experiments and passages clustered together. See Supplementary Table 12 and online Excel file 6 for detailed histotype and replicate information and online Excel data file 5 for a complete list of antibodies.

Supplementary Figure 13
Protein expression profiles of OCI lines are stable during long term culture A) Ten OCI lines were cultured long term and triplicate aliquots were prepared at two different times with 9 to 48 passages separating them.
B) The unsupervised hierarchical clustering heatmap demonstrate that each cell line groups within the same cluster (1 vs. 2) regardless of passage number.
C) The cell lines, passages numbers and different replicates are listed in the order they clustered in the heatmap, demonstrating that the three replicates and two different passages of each cell line clustered together. The RPPA data set of 63 samples were probed with 218 antibodies.  50. Six replicates of the OCI (blue) and SOC (red) lines were plated in OCMI medium (3000 cells/well) in 96 well plates. Next, Taxol was added and metabolic activity was measured as 590/530 fluorescence via Alamar Blue after 5 days, which was used to estimate the LD50 for each line.

B-D)
Full dose-response curves of three separate experiments plotted as percent cell viability of vehicle control vs. Log Taxol concentration (nM); OCI lines (blue), SOC lines (red). The LD90 for all of the OCI lines was >800 nM; in contrast LD90 for the SOC lines ranged between <1 nM (A2780), 5.2 nM (TOV-112D), 24 nM (ES2), 158 nM (SKOV3) to 800 nM (OV90). Error bars represent standard error of the mean of six replicates.

Supplementary Figure 19
Vincristine response of cluster 1 OCI vs. Cluster 2 SOC/OCI cell line Dose-response curves of the three independent Vincristine sensitivity experiments demonstrate that the cluster 1 OCI cell lines (blue) are more sensitive to Vincristine compared to cluster 2 SOC cell lines (red). Six replicates of the OCI and SOC lines were plated in OCMI medium (3000 cells/well) in 96 well plates. Error bars represent standard error of the mean Both Taxol and Vincristine target the microtubules for their mechanism of action. However, while the vinca alkaloids bind stoichiometrically to tubulin subunits and induce polymer disassembly, the taxanes bind to polymerized tubulin and prevent disassembly. During cell division microtubules are assembled by polymerizing tubulin monomers, and are disassembled when they are no longer needed. Hence, both drugs block mitosis but through very different mechanisms.
Since taxanes and vinca alkaloids have very different mechanisms of actions, the control experiment with Vincristine suggests that the sensitivity of cluster 1 cell lines is related to a difference in tubulin function in cluster 1 vs. 2 cell lines and it is unlikely to be due to idiosyncratic metabolism, transport or inactivation of taxanes.
The gene expression microarray data for the Ince et al ovarian cancer cell lines were downloaded from the Gene Expression Omnibus (GSE40785). In total, 37 microarrays and 47,323 unique probes were retrieved. All data were log2 transformed with Microsoft Excel 2010, quantile normalized with R statistical software version 3.1 with the package 'ArrayTools'. The data were then filtered with Cluster v3.0 and all probes with standard deviations of less than 0.5 across all samples were removed, leaving 7,479 probes. These 7,479 probes from the 37 arrays were then median centered in Cluster v3.0 and hierarchical clustered, where two major groups of ovarian cancer cell lines were observed (Ince Cluster 1 and 2). The top 500 probes over-expressed in each group (FDR <5) were then used to generate gene signature scores for each tumor in the TCGA ovarian cancer gene expression dataset. Signature scores for other relevant gene expression signatures were also identified for each tumor in the TCGA dataset including the TCGA defined 'poor' and 'good' prognosis genes, the Tothill C1-C6 'up' genes, and a recently described 'vascular content' gene expression signature. All signature scores were identified by taking the mean value for all genes within a respective signature. Excel was then used to identify the Pearson correlation values for each of the signature scores. This analysis showed a very strong correlation with the genes most abundantly expressed in the Ince Cluster 1 ovarian cell lines and the Tothill C1 up genes (R value = 0.93). Interestingly, both these signatures also exhibited a strong correlation with genes that are abundantly expressed in vascular endothelial cells (>0.7) 8 , suggesting that both of these signatures define ovarian cancer cells with mesenchymal/endothelial attributes.  Table 4 The list of causes that prevent successful culture of tumor cells During the development and optimization of OCMI medium and culture methods, we failed to establish cell lines in some cases. In two cases, the tissue sample was necrotic and no viable cells were detected upon plating. In four cases the tissue sample was composed of normal and stromal tissues, and no tumor cells were detected. In two cases bacterial/fungal contamination was detected within the first weeks of culture. In these cases we had to terminate the cultures. In addition, during the preliminary phase, we used five endometrioid adenocarcinomas and one mucinous adenocarcinoma to optimize the medium formulation. These specimens did not survive due to suboptimal medium formulations and culture conditions such as 21% O2 that was initially used for endometrioid and mucinous tumors. Upon fully optimizing the culture methods, we failed to establish a long-term continuous culture only once with a viable sample that contained tumor tissue (OC-P10).     Table 6 The mitochondrial hyper-variable region 2 (HVR2) fingerprint for 16  Ins Loc .1 .2 .1 .   Supplementary Fig.  20 b). This result demonstrates that that the differences between cluster 1 vs. cluster 2 cell lines are not due to histotype differences. Furthermore, there were at least 5 to 36 passages between the RPPA analysis and Taxol/Cisplatin sensitivity experiments. Nevertheless, there was a close concordance between RPPA clusters and drug response, providing further evidence that the phenotype of OCI lines are stable over many passage in culture.  Table 15 List of proteins that are differentially expressed that are differentially expressed between cluster 1 vs. cluster 2 cell lines, corresponding to the RPPA heatmap in Figure 7B. The genes in italic in cluster 1 have been previously implicated in Taxol Supplementary Table 16 List of proteins with known Taxol resistance association that are differentially expressed between Cluster 1 vs. Cluster 2 OCI cell lines in RPPA analysis.

c.JUN pS73
Paclitaxel-resistant human ovariancCancer cells undergo c-Jun NH2-terminalkKinase-mediated apoptosis 19 . A physical interaction of Stat3 with c-Jun has been reported both in vitro and in vivo.

Crystalline
Crystalline interact with tubulin to regulate the equilibrium between tubulin and microtubules 21 .

Erg.1_2_3
Expression of EGR-1/p38MAPK plays a critical role in paclitaxel resistance of ovarian carcinoma cells 22

FAK _pY397
Docetaxel induces FAK cleavage in taxane-sensitive ovarian cancer cells but not in resistant cells 23 .

MAPK_ pT202
MEK inhibitor CI-1040 potentiates efficacy of Taxol in xenograft tumor modes 25 . RNAi screening identified Erk1 as enhancing paclitaxel activity 26 . Inactivation of ERK is necessary for the enhancement of paclitaxel cytotoxicity by U0126 25 .

N.Cadherin
N-Cadherin is over-expressed in Taxol resistant cells 27 .

PAX2
PAX2 expression correlated with resistance against apoptosis and proinvasive phenotype 29 .

PTEN
Silencing AKT in PTEN-mutated prostate cancer cells enhances the antitumor effects of taxol 31 .

SMAD3
Increased expression in Paclitaxel resistant cells 32 , SMAD3 binds to microtubules 33 . SMAD3 and SMAD4 cooperate with c-Jun/c-Fos to mediated transcription 34 .

SRC
Knockdown of Src enhanced paclitaxel-mediated growth inhibition in ovarian cancer cells 35,36 , SRC activates STAT3 37 . STAT3 siRNA inhibited Bcl-2 expression 38 . Bcl-2 Down-Regulation is associated with Paclitaxel resistance 39 . Constitutive activation of Stat3 by the Src causes growth of breast carcinoma cells 40 .

STAT3
STAT3 activation through Src leads to Taxol resistance 41 , STAT3 is activated by ERK1 and induces AKT. STAT3 binds the C-terminal tubulin 42 . Knockdown of Stat3 reduces AKT1 expression 43 STAT3 is induced by Src 44 .
The specific mRNA and protein expression differences cluster 1 vs. 2 are difficult to associate with specific genes associated with Cisplatin resistance, which can be due to the multitude of complex mechanisms associated with 45  can be established as cell lines using these standard cell culture media. Consistent with this, we failed to establish any permanent human breast or ovarian cancer cell lines using these media to culture cells from more than one hundred tumors. Therefore, we explored the use of a cell culture media that we had previously described in Ince et al,. 46 .
The chemically-defined cell culture medium we developed can support long-term growth of normal and transformed human breast cells without undefined components such as serum, feeder-layers, tissue extracts or pharmacological reagents 46 . Using a version of this medium optimized for normal cells (WIT-P), we were able to culture human breast epithelial cells (BPEC) for more than seventy population doublings during six months of continuous culture, a nearly 10 21 -fold expansion of cell number 46 . In contrast, in standard medium these normal breast epithelial cells ceased growing after several passages 46 . Importantly, it was recently shown that normal breast cells maintain high telomerase activity in culture in this medium 47 , which is not the case in standard media.
We initially tested a version of the medium (WIT-T) optimized for transformed human breast cells (BPLER) 46 to establish human ovarian tumor cell lines, but were unsuccessful with more than a dozen tumor samples.
Next, we examined whether modifications in the concentration of distinct components of the media would support the growth of primary ovarian tumor cells. First, we reasoned that low levels of serum may be required to mimic the physiologic environment of normal ovary, fallopian tube and ovarian cancers in the peritoneal cavity.
Since most normal epithelia are not in direct contact with blood or serum, we had used a chemically-defined serum free medium to culture normal breast epithelial cells.
However, the normal ovaries and fallopian tubes are normally surrounded by normal peritoneal fluid, which contains protein concentrations that are similar to serum levels, which is increased to above plasma levels in malignant ascites fluid associated with ovarian cancer [48][49][50] . Consistent with these observations, the addition of serum to the medium proved to be necessary, but not sufficient for growth of ovarian tumor cells; additional factors had to be optimized (Supplementary Fig. 1-2).
One of the difficulties associated with optimizing media is the non-obvious and synergistic combinatorial effects of individual components. The individual additives increased ovarian cancer culture success only incrementally in many cases. However, cell proliferation increased exponentially when all components were added at optimal concentrations (Supplementary Fig. 1-4).
As a result, after many years of optimization, we found that inclusion of insulin, hydrocortisone, EGF, and cholera toxin in addition to fetal bovine serum to media showed broad efficacy in supporting the growth of the serous ovarian cancers.
The non-serous ovarian tumor subtypes, such as endometrioid and mucinous cancers, express estrogen receptor (ER). In this case we found that addition of β-Estradiol (E2) is necessary to enhance the growth of these tumors subtypes (Supplementary Fig. 2).
We also had to optimize O 2 levels because while many papillary serous and clear cell subtypes proliferate best in ambient O 2 (18 to 21%), the endometrioid and mucinous subtypes proliferate best at low O 2 levels (5 to 10%) in general (see Supplemental A very time consuming aspect of the medium optimization process was the need to validate applicability of the medium and methods across the broad spectrum of specimens and cancer subtypes. We found that while some ingredients have little effect on culture efficiency for some samples, they are absolutely essential for others ( Supplementary Fig. 3). Importantly, the effects of removing these components become more apparent after multiple passages (Supplementary Fig. 4). Hence, effects of each ingredient has to be tested over many passages.
The cell attachment surface is also important; the standard tissue culture plates have a uniformly negatively charged surface which was not suitable for primary ovarian cells and produce variable results. In contrast we use a modified cell culture plastic with mixed positive and negative charges (Primaria, BD), which greatly helps in preserving cell morphology and heterogeneity of the original tumor. We found that while some cell lines proliferate best on standard plates others require Primaria plates (Supplementary Table 3).
All of these factors -the non-obvious nature of combinatorial outcomes, all-or-none synergistic effects, the need to test conditions in multiple passages and in multiple cell lines, and the very large number of conditions to test -precluded an incremental approach to medium development.
With the current formulation of OCMI, we were able to culture ovarian tumors with  Fig. 1-2) To our knowledge, none of the standard media support the culture of all of the existing SOC lines. Thus, it has been difficult to compare a large panel of SOC lines with each other. We also found that none of the OCI lines can be cultured in existing standard media ( Supplementary Fig. 5 A-D). In contrast, all of the SOC lines we tested can be cultured in OCMI medium (Supplementary Fig. 5 Fig. 6).
Consistent with our experience, others also recently reported that Y-27632 negatively impacts culture of hematopoietic progenitor cells, human adipose-derived stem cells and melanomas 2 . Thus, this is not an approach that can be used universally on all cells.
Sato et al. also reported a DMEM:F12 based medium for the culture of human small intestine, Barrett's esophagus and colon epithelium using drugs that inhibit of Alk and p38 52 . In the absence of these two kinase inhibitors, human intestinal cultures were growth arrested after 10 to 20 population doublings 52 . It is a significant concern that cell lines derived using the Liu or Sato et al., media will be 'addicted' to these drugs. In addition, the presence of feeder layers causes difficulties with the interpretation of cytotoxic assays since the observed results can be due to direct effects on cancer cells as well as indirect effects on the feeder layer. In contrast, we have no such drugs or feeder layers in our system, which is a significant improvement in new culture methods.

Tumor Tissue Collection and Clinical Information
All study procedures were approved by the Internal Review Boards of the Brigham and Women's Hospital and the University of Miami to collect discarded tissues with a written consent from all patients. In this initial study we concentrated on developing methods for successful culture of human ovarian tumors. For this purpose we used anonymized discarded human tissue and did not have access to clinical patient follow up information retrospectively. A prospective study with larger number of patients will be needed to examine the direct comparison of individual patients to treatment and in vitro response of their corresponding cell line, which is underway.

Establishment of Cell Lines
Tumors are complex tissues composed of many cell types including stromal cells such as fibroblasts, endothelial cells, leukocytes, macrophages as well as normal epithelial cells that are intermingled with tumor cells 53 . Among these, fibroblasts have historically been the easiest cells to grow in standard culture medium 54,55 . In general serum promotes fibroblast growth and inhibits epithelial cell proliferation. When tumor tissue is cultured in medium with high serum content, typically there is an exponential growth of fibroblasts such that in a few weeks the fibroblasts completely overtake the culture plate, and soon all other cell types including tumor cells are eliminated. For this reason we used low levels of serum (2 %) to culture ovarian tumor cells during the initial passages (1-5) to suppress fibroblast growth.
Another difference between epithelial cell and fibroblasts is adherence to tissue culture plastic; in general epithelial cells are more strongly adherent to the culture flasks and require higher concentrations of trypsin to release them. Thus, it is possible to treat the plates with diluted trypsin first (0.05%), which selectively removes stromal cells.
Afterwards, the epithelial cells that are still attached to the culture plate can be were treated with 0.25% trypsin for sub-culturing. OCMI was designed to support epithelial tumor cell proliferation and suppress fibroblast growth. However, in general it takes 4-6 passages with differential trypsinization to establish tumor cultures free of stromal and normal cell types. Afterwards the FBS levels were increased to 5% to increase tumor cell proliferation.
We observed a remarkable degree for consistency in the phenotype of OCI lines in long term culture. All OCI cell lines we cultured for at least 20-25 population doublings. In several cases, we carried out a formal population doubling analysis (Figure 1), which demonstrated that the OCI lines proliferate for at least 120 days (~60 population doublings). Importantly, there was a remarkable correlation among our results; even though the mRNA profiling (Figure 3), protein profiling (Figure 4) and drug sensitivity (Figure 7) experiments were carried out by different investigators at different times (passages 8-56) (Supplementary Fig. 12-13, Supplementary Table 12-14). These results indicate that OCI cell lines have a stable and reproducible phenotype.

Clonal Selection
Mindful of the possibility of clonal selection, we carefully monitored all OCI cultures for emergence of fast growing colonies, eliminated plates with too few starting cells and avoided partial trypsinization of plates during sub-culturing of OCI lines.

Measures of cell proliferation
In many previous reports, the cumulative number of cell passages has been used to indicate successful establishment of cell lines. However, it is important to note that the to be much greater with OCI lines (Figure 1 d). Sixty population doublings would be approximately equal to 10 20 -fold expansion in cell numbers (~ 100 quintillion cells) more than adequate for any research use.
The growth rate plateau that is seen during the culture of tumor cells in standard media is linked to the long lag time between the initial plating of tumor tissue and the emergence of a cell line. This is a significant variable in evaluating the efficiency and practicality of a culture system, and has significant implications for the quality of the cell lines. For example, Verschraegen et al., reported that on average it took more than five months (21 weeks) before tumor cells could be passaged for the first time, which is similar to our experience using RPMI medium (Figure 1 a) 56 .
In standard cell culture medium both normal and ovarian tumor cells are growth arrested within the initial several passages (Figure 1). Since the growth arrest due to telomere-shortening occurs typically after 50-70 passages, this type of early growth arrest is due to inadequate culturing conditions 57 .

Soft agar colony assay
In order to confirm that the OCI cell lines we established maintained their transformed phenotype in culture we carried out anchorage independent growth assays in soft agar.
Since normal cells are incapable of forming soft agar colonies, this is an excellent method to ensure that we have indeed established tumor cell lines. For these assays, well bottoms of a 12-well plate are sealed with 0.6% agar prepared in OCMI medium to prevent monolayer formation. Cells from established cultures (passage 6-8) are harvested. A single cell suspension in 0.4% agar in OCMI medium is added and allowed to set at room temperature, and placed in 37 C incubators with 5% CO2. The cells are fed with 0.4% agar in OCMI at 2 weeks, and colony formation is assessed 2-4 weeks after plating.

LOH Analysis
The genomic DNA of tumor tissues were extracted from paraffin sections or, when available, from fresh tissues. The fresh tumor tissues were homogenized directly in RLT+ cell lysis buffer (Qiagen). The DNA was extracted from the lysates using the Qiagen All-Prep mini kit. Briefly, DNA is cleaved with Sty1, and the fragments are PCR amplified. The purified products were further fragmented with DNaseI, biotinylated, hybridized to a chip, and fluorescently labeled with phycoerythrin-conjugated streptavidin with signal amplification. Inferred LOH analysis in Figure 2 was performed using dCHIP software and employed the hidden Markov model with a reference heterozygosity rate of 0.2.
LOH segment analyses in Supplementary Fig 2a and

Copy Number Analysis
The copy number analysis in Supplementary Fig. 9 was performed using the Molecular Inversion Probe (MIP) 330k microarrays from Affymetrix. The basic MIP assay has been previously described [58][59][60][61]   Interestingly, both these signatures also exhibited a strong correlation with genes that are abundantly expressed in vascular endothelial cells (>0.7), suggesting that both of these signatures define ovarian cancer cells with mesenchymal/endothelial attributes.
Next, the Tothill gene expression microarray data (GSE9899) and clinical data were downloaded and processed as above. Since the AOCS Tothill paper did not define which tumors they identified as belonging to groups C1-C6, we next aimed to categorize the 295 AOCS Tothill samples into one of six groups. Therefore, we used the k-Means function of Cluster v3.0, with similarity metric Euclidean distance, and ran 1000 permutations to identify these sample groupings. The Kaplan-Meier plot ( Figure 8) for overall survival was then generated with R using package Biobase and in Excel with Winstat.
Methods used for pathway enrichment analysis in Supplementary Fig. 10-11 For pathway analysis in Supplementary Figure  Protein expression analysis in Figure 3, 5, 7 and Supplementary Fig. 12-16 For the RPPA analysis in Figure 3 the cell lysates were immobilized on nitrocellulose coated slides, and each slide was incubated with an antibody specific for a protein of interest. The protein lysates were prepared in a lysis buffer containing SDS and protease inhibitors. Semi-confluent wells in 6-well plates were lysed in 125 uL lysis buffer on ice in triplicate (at least two different passages from the same cell line).

Cell Line Unique Identifier mtDNA and STR in Supplementary Tables 5-7
A common problem in cell culture is cross-contamination or misidentification of cells. In repeated studies since 1970s, it has been shown that 15-25 % of cell lines are contaminated with a second line, or is completely misidentified (Alston-Roberts, 2010).
In the 1970s and 1980s, Nelson-Rees showed that over 100 cancer cell lines were actually HeLa cells 66,67 . An effective cell culture quality and identity control is required in order to avoid inter-and intra-species contamination of cell lines and their further propagation and dissemination. However, vigilant monitoring against misidentification and cross-contamination is possible by developing a practical "unique identifier" for the cells by the establishing laboratory.
In Supplementary Table 5 and 6 we present mitochondrial DNA (mtDNA sequence) and short tandem repeat (STR) evidence that the cell lines examined in this manuscript are from unrelated individuals. Thus, the OCI cell lines can be verified by the recipient laboratories and can be monitored for purity and integrity. This will significantly reduce the incidence of cell line contamination and misidentification 68 . The control region of the human mtDNA is highly polymorphic due to a rapid rate of evolution. The mtDNA does not undergo recombination and is present in high copy number per cell. For this reason, its analysis is very useful for the identification of cell lines. . This match appears to be incidental because the SNP/LOH profile of OCI-C2p has a > 95% identity with its matching uncultured original tumor. All the SOC lines used in this study were also STR validated and matched the unique reference profile listed by ATCC for each line (data not shown).

mRNA profile of high grade serous vs. clear cell ovarian carcinomas
To determine if the OCI-CC and OCI-HGS cell lines reflect gene expression differences that are found within patients with CCOC and HGSOC, we identified a mRNA expression signature that significantly separated the CCOC and HGSOC tumors from Wu et al 72 . Signature values for each OCI cell line were then identified, box-andwhisker plot were generated, and ANOVA identified that the same genetic discriminator signature that was identified in patient tumors was also present in the OCI lines.
The gene expression microarray data from Wu et al., 72 were downloaded from the Gene Expression Omnibus (GSE6008). R software version 3.1 was then used to quantile normalize the data, collapse probes, and median center the data. Significance  (Ince dataset). Box-and-whisker plot were generated, and ANOVA identified that the same genetic discriminator signature that was identified in patient tumors was also present in the OCI lines

Drug sensitivity experiments
The relative sensitivities of OCI and SOC cell lines to Taxol was measured by seeding 3000 cells/well in six replicates in 96-well black-walled clear bottom Corning plates and allowing attachment in OCMI for 12h. Both OCI and SOC cell lines were cultured in the presence of Taxol dosages ranging from 1 to 800nM (or vehicle control) in OCMI medium for 120h. The fraction of metabolically active cells after drug treatment was measured by incubation with 2:10 (v/v) CellTiter-Blue reagent (Promega Cat# G8081) in media for 2h, and the reaction was stopped by addition of 3% SDS. Fluorescence was measured in SpectraMax M5 plate reader (Molecular Devices, CA) using SoftMax software (555EX/585EM). In case of high variation among calculated values for four independent assays, four additional independent assays were performed to allow defining and discarding outliers.

Lethal Dose Analysis:
Data was analyzed using GraphPad Prism 5 Software, and values were fit to a dose response-inhibition curve with variable slope (sigmoidal with four parameters). The cell viability as response r between bottom (B) and top (T) values (B < r < T) was assumed to depend on concentration (C) via a general Hill equation for inhibition as in equation (1) (1) ( )