Preclinical model of organotypic culture for pharmacodynamic profiling of human tumors

Abstract

Predicting drug response in cancer patients remains a major challenge in the clinic. We have perfected an ex vivo, reproducible, rapid and personalized culture method to investigate antitumoral pharmacological properties that preserves the original cancer microenvironment. Response to signal transduction inhibitors in cancer is determined not only by properties of the drug target but also by mutations in other signaling molecules and the tumor microenvironment. As a proof of concept, we, therefore, focused on the PI3K/Akt signaling pathway, because it plays a prominent role in cancer and its activity is affected by epithelial–stromal interactions. Our results show that this culture model preserves tissue 3D architecture, cell viability, pathway activity, and global gene-expression profiles up to 5 days ex vivo. In addition, we show pathway modulation in tumor cells resulting from pharmacologic intervention in ex vivo culture. This technology may have a significant impact on patient selection for clinical trials and in predicting response to small-molecule inhibitor therapy.

Predicting response to therapy in cancer patients based on cell line and xenograft studies has been traditionally very difficult. More recently, the limited success of current small-molecule inhibitors in many epithelial human cancers highlights the need to develop better techniques to more accurately predict response to therapy, preferably tailored to the individual cancer and its unique genetic and epigenetic alterations.

Understanding the effect of pathway perturbation on tumor response in the native tumor microenvironment is also of paramount importance. Tumor–stroma interactions have long been recognized as important facets in the pathogenesis and dissemination of malignancy. Significant evidence supporting the role of peritumoral tissues in tumor maintenance includes the presence of genetic mutations in the stroma of several types of cancers (1) and the role played by stromal cells in the acquisition of resistance to therapy (2).

Previous studies that use standard primary cell-culture systems and cell-line s.c. xenografts have advanced our understanding of tumor behavior (3); however, these methods have inherent limitations in evaluating the role of the tumor microenvironment in modulating carcinogenesis and tumor progression. As such, newer, more robust, and innovative in vitro models need to be developed. Ex vivo tissue slices seem to be the most promising of techniques (4, 5). The special advantage of this approach is the ability to both maintain organ and cellular architecture, while also preserving the integrity of the tumor–stroma interaction (6). This approach has been exploited with some success by other groups (6–14). Crucially, however, no study to date has systematically evaluated the efficacy of ex vivo organotypic culture in cancer using the slice technology on a heterogeneous pool of tumor samples and compared the effects of small-molecule inhibitors on specific signaling pathways.

We have perfected an ex vivo reproducible, rapid, and personalized culture method suitable to investigate epithelial–stromal interactions in tumor onset, progression, and resistance to therapy. Importantly, we show that our method allows for investigation of antitumoral pharmacological properties in a system that preserves the original cancer microenvironment. The technique that we have developed promises to shed light on the gap that currently exists between results in cell line/xenograft studies traditionally used to predict drug response and the actual efficacy in humans. It will provide guidelines for entry into clinical trials. Finally, it provides the oncologist data that are specific to the patient’s tumor. As a proof of principle, we have focused on the PI3K/Akt signaling pathway, because it plays a prominent role in cancer and its activity is characteristically affected by epithelial–stromal interactions (15, 16).

Results

Technical Improvements Aimed at Achieving Consistent and Reliable Results in the Ex Vivo Culture of Human Tumor Specimens.

This methodology takes advantage of the rapid sectioning of tumors immediately after harvesting from surgery. Briefly, the tumors were sliced on a vibratome at 400 μm thickness and cultured in 6-well plates for up to 120 h (Fig. 1). Details of the culture system are provided in Methods. To improve on the ex vivo tissue-slice culture technique previously used in other systems and adapt it to tumor tissues of various origin, extensive assay optimization was undertaken. This ranged from the use of advanced instrumentation to the adaptation of adequate culture conditions based on extensive experimental trials conducted before undertaking this study. Briefly, compared with previous instruments available, the instrumentation (Vibratome VT1200; Leica) allows for rapid cutting of each tumor tissue slice in 20–40 s with the significant advantage of controlling the thickness of the slice at a micrometer level (17); it also minimizes the vertical deflection of the blade, and therefore, it preserves the integrity of delicate specimens and ensures a higher number of viable cells on the section surface. In addition, with this technique, no agar or agarose embedding is necessary to successfully cut uneven specimens. The culture-plate inserts allow for prompt adherence of the slice to its surface without using an extracellular matrix coating, keeping the tissue at the interface between culture medium and a CO₂-rich environment. A small droplet of medium is usually added on top of the tissue slice to create a thin film of liquid that helps maintain the explant humid (Technical Note Number TN062; Millipore) (18). To avoid microbial contamination and achieve reliable slice cultures, we optimized media composition in terms of serum, growth factors, and antibiotic concentrations. The most favorable media composition is described in Methods. Generally, tumors with a solid consistency are better suited for this technique (e.g., lung, breast, and tumors with desmoplastic reaction in general). Friable tumors such as glioblastoma, however, are less suitable to analysis by organotypic culture. Baseline diversity among different organs with respect to biomarkers obviously exists and has to be taken into consideration in subsequent intratumoral and intertumoral comparisons.

Fig. 1.

Schematic overview of the experimental design.

Morphology, Proliferative Activity, and Viability Are Preserved in Organotypic Tissue Cultures.

Morphological integrity of tissues, defined as preservation of general architecture including epithelial structures and their spatial relationships to stroma, was confirmed in both normal and tumor samples up to 120 h of culture. Representative H&E images of colon and lung carcinomas are illustrated in Fig. 2A. The proliferative activity was assessed by Ki-67 immunostaining in a subset of tumor tissue cultures (n = 42). The percentage of proliferating cells was remarkably stable across all time points without a significant decrease up to 120 h (ANOVA test; P > 0.05) (Fig. 2B). Similarly, MTT [3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide] assay of tumor slices revealed no significant loss of viable cells up to 120 h of ex vivo culture (ANOVA test; P > 0.05) (Fig. 2C). Finally, the number of apoptotic cells, measured by TUNEL assay, did not change significantly over time (ANOVA test; P > 0.05) (Fig. 2D). Raw data for individual tumors are included in Tables S1 and S2.

Fig. 2.

Evaluation of tumor morphology, proliferation, viability, and apoptosis in organotypic tissue cultures. Tissue slices (400-μm thick) were cultured on organotypic supports in serum-supplemented media up to 120 h after surgical explants from patients. A selection of 42 cultured tumors was investigated for proliferative activity (Ki67 immunohistochemistry), tissue viability (MTT assay), and cell death (TUNEL assay) every 24 h. Two representative organotypic tissue cultures from colon and lung tumors are depicted in A. (Magnification: 250× and 100×, respectively.) Evaluation of Ki67 immunoreactivity in cultures (percent of positive cells relative to corresponding uncultured sample; T0) is shown in B, tissue viability is represented in C, and apoptosis is shown in D.

PI3K/Akt Pathway Analysis.

Gene expression of selected targets in the PI3K/Akt pathway was measured by qPCR in a subset of tissue cultures from 16 neoplastic samples (colon, lung, and prostate adenocarcinomas) and 7 matched normal counterparts. mRNA levels of PI3K, Akt1, ribosomal protein S6 kinase (p70S6K), and focal adhesion kinase (FAK-PTK2) were maintained at all time points compared with the initial values (Fig. 3A). Organ-specific expression levels of each gene are shown in Tables S3, S4, and S5.

Fig. 3.

PI3K–AKT pathway analysis in organotypic tissue cultures. The expression levels and the phosphorylation status of PI3K-pathway members were investigated in a subset of tissue cultures (16 neoplastic samples and 7 matched normal counterparts) up to 96 h. mRNA levels of PI3K, AKT1, and ribosomal protein S6 kinase were quantified by real-time RT-PCR (A). Averaged relative quantities of targets are indicated as fold changes relative to the corresponding T0 sample (RQ ± SEM). The active status of AKT and S6 ribosomal protein (S6RP) was investigated by phospho-immunohistochemistry (p-IHC) every 24 h. Immunoreactivity of tumoral and normal cultured slices (B and C, respectively) was compared with the matched T0 samples. (Magnification: 100×.) Results are indicated in histograms as averaged p-IHC score per time point (n fold T0 ± SEM).

We then sought to investigate the activation status of this signal transduction pathway by immunohistochemical analysis; analysis was performed on the formalin-fixed and paraffin-embedded slices using phospho-specific antibodies against the activated form of Akt and S6RP. Akt displayed constant levels of activation in tumor slices compared with the corresponding time 0 (T0) samples (ANOVA test; P > 0.05) (Fig. 3B). However, S6RP showed a cyclical pattern of activation in tumors; p-S6RP levels peaked at 48 h and 96 h of culture, whereas they were comparable with T0 sections at 72 h (Fig. 3B). Normal tissue slices expressed steady low levels of both phospho-proteins at all time points (Kruskal-Wallis test; P > 0.05) (Fig. 3C). An example of p-Akt and p-S6RP immunostaining in a representative matched normal/tumor lung slice is depicted in Fig. 3 B and C. Raw data for individual tumors are included in Table S1.

Gene-Expression Profiling of Tumor and Stromal Compartments Across Time.

As displayed in Fig. 4 A–C, more than one-half of all transcripts (i.e., ∼10,000 genes) show stable expression (false discovery rate [FDR]-adjusted P value ≤ 0.2) over time in both the tumor and stromal compartments. Despite moderate levels of mRNA available from the samples, we observed differential expression in many genes across tumor and stroma. The expression patterns of the top 143 differentially expressed genes that are significant at P value = 0.01 are plotted with a heatmap (Fig. 4D). Among the most significant 15 pathways identified by the method of Tian et al. (19) across tumor and stroma samples were genesets for MAP kinase, collagen synthesis, and extracellular matrix support. We also included various cancer-related genesets from the molecular signature database MSigDB in our analysis, such as non–small-cell lung cancer and phosphatase and tensin homolog deleted on chromosome 10 (PTEN) pathway genes. All of the depicted pathways exhibited differential mean profiles across the tumor and stroma samples and are shown in Fig. 4E and Fig. S1. The distinction is both biologically meaningful and statistically significant (Hotelling’s T² test with significance level of 0.01). Moreover, several distinctive markers of both epithelial [cytokeratin and epithelial membrane antigen (EMA)] and stromal (CD34, laminin, and vimentin) compartments were found to be stable over time by immunohistochemistry (Fig. 4F).

Fig. 4.

Gene-expression analysis of tumor and stromal samples over time. A and B show histograms of the FDR-adjusted P values indicating stability of gene expression over time for tumor and stromal samples, respectively. The smaller a gene’s P value, then the more conserved is its expression in time. Thus, more than one-half (∼10,000) of the total number of transcripts show conserved expression in our dataset. In C, bivariate distribution of tumor and stroma P values for every gene is plotted with a density map. The blue density depicts the concentration of genes at any given location. Clearly, most genes lie along the diagonal, indicating correlation in the stability of their tumor and stroma expressions. Notably, the cluster of dark points near the origin (0.0) represents stable expressions for more than one-half of the transcripts. In D, the expression of the top 143 differentially expressed genes that are significant at P value level 0.01 is plotted with a heatmap. The colors blue and red depict low and high expression, respectively. The plots in E show (blue line) the loess-smoothed expression profiles of two gene sets, non–small-cell lung cancer and extracellular matrix cellular constituent, distinctive of tumor and stromal compartments, respectively. The SE gray band and the cross-sample median line are shown for perspective. For each of these gene sets, the distinction between the mean profiles of tumor and stroma classes was found to be statistically significant. F shows microphotographs of a lung tumor at 0 h, 24 h, and 72 h of culture. H&E staining shows morphological integrity, CD34 highlights endothelial cells, laminin stains basal membranes, vimentin is an intermediate filament protein found in cells of mesenchymal origin, and cytokeratin and EMA label epithelial cells. (Magnification: 100×; Inset magnification: 250×.)

Targeted Therapy in ex Vivo Tissue Cultures.

When the reliability of our culture system was established, we sought to test the possibility of modulating the activation status of the pathway by using targeted inhibitors. After ex vivo treatment with the PI3K inhibitor LY294002, tumor slices showed a remarkable reduction of p-Akt and p-S6RP levels at all time points, as illustrated in Fig. 5 A–C (t test; P = 0.05 and P = 0.017, respectively). Importantly, this inhibitor induced a partial decrease in tissue proliferation and viability (Table S6). As expected, gene expression of the same targets was not affected by treatment with the LY294002 (Table S7). Given the degree of variability of the assay and the initial relatively high concentration of LY (50 μM), we then performed a dose-response experiment in three representative tumor tissue cultures (two lung and one colon adenocarcinomas). These results show that overall proliferation and apoptosis both partially decrease and increase respectively over time with increasing concentrations of LY294002 [for Ki-67 (Fig. 5D), t test; DMSO vs. LY 10 μM (P = 0.07), DMSO vs. LY 20 μM (P = 0.01), DMSO vs. LY 50 μM (P = 0.17); TUNEL information in Table S8]; this is accompanied by a corresponding reduction in p-Akt and p-S6RP protein levels at increasing concentrations of the drug (Fig. 5E). Moreover, we treated slices of dedifferentiated liposarcoma (DDLPS), a tumor known to amplify MDM2 in nearly 100% of cases, with the Mdm2 inhibitor Nutlin-3. Similar to the experiments with LY294002, we observed a reduction in proliferation as a result of treatment (assessed by BrdU) (Fig. S2). In addition, we found that p53 target genes p21 and Mdm2 were induced in DDLPS at 24 h after addition of drug as assessed by immunohistochemistry (IHC) (Fig. S2).

Fig. 5.

PI3K-targeted therapy with an LY294002 inhibitor. In A–C, tumor tissue cultures were incubated with 50 μm of LY29004 up to 96 h. Slices were harvested every 24 h, and drug effects on PI3K pathway (A, p-AKT; B, p-S6RP-IHC) were tested. Images of both phospho-protein immunostaining in a representative case of lung tumor are shown in C. (Magnification: 100×.) D and E show the results of a dose-response experiment on cell proliferation and PI3K pathway, respectively, using increasing doses of LY294002. (Magnification: 100×.)

Discussion

In this study, we describe ex vivo organotypic culture of human tumor specimens followed by fixation and reembedding of these tissue sections and analysis of various parameters in situ, such as expression of activated proteins as well as global gene-expression profiles. We show that this is a rapid, reproducible, and personalized method suitable to test therapeutic agents in viable tumor sections with preserved spatial relationships to stroma and their own microenvironment. The robust nature of our system is shown by the post-culture preservation of epithelial–stromal relationships, tumor tissue morphology, proliferation, and viability at all time points. We show that cancer-related pathways such as the PI3K/Akt signaling pathway remain functionally active and stable in ex vivo culture both at the transcriptional and protein level. Importantly, global gene-expression profiles specific to both tumor and stroma were found to be steady at the various time points, showing that an intact epithelial–stromal relationship is maintained over time.

Crucially, we show the utility of this technology in potentially predicting tumor sensitivity to drugs in a patient-specific manner (20). Untreated lung, prostate, and colon tumors showed constant levels of p-Akt and p-S6RP that decreased dramatically after ex vivo treatment with the PI3K inhibitor LY294002 in a dose-dependent manner. This inhibitor also predictably led to a reduction in proliferation rate and tumor-cell viability without affecting gene expression of the same targets. Importantly, we also highlight the extended use of this technology by assessing tumor response to another pharmacologic reagent (Mdm2 inhibitor Nutlin-3) that showed comparable efficacy. In addition, this methodology may be amenable to manipulation by small interfering RNA and/or monoclonal antibody treatment. One of the most important applications of this technology will be the assessment of the response as a function of the genetic background of the tumor, although here we focused on the appraisal of pathway activation and response to targeted drugs. Ultimately, answering the question of if ex vivo organotypic cultures will actually be predictive of the final therapeutic response in patients will require long-term follow-up and a direct comparison with clinical trial results using the same drug(s) ex vivo or in vivo.

In conclusion, we have showed that ex vivo organotypic short-term culture of tumor represents a means by which pathway activation and pharmacologic inhibition can be rapidly studied with respect to the native heterogeneity of a patient’s tumor. In an evolving era of targeted inhibitors, defining an in vivo signature of pathway activation remains challenging. The evaluation of functional and genetic alterations in all cellular compartments of heterogeneous solid tumors may help to identify patients that could benefit from targeted inhibitors.

Methods

Specimens.

Fresh tissue samples (n = 271) (Table S9) were procured immediately after surgical resection in San Paolo Hospital, Milan, Italy. Patients who received neoadjuvant chemotherapy and/or radiotherapy were excluded from the study. Informed consent was obtained from all patients. Tissue viability and histopathological diagnosis was confirmed by frozen section examination.

Fresh Tissue Sectioning.

A Vibratome VT1200 (Leica Microsystems) was used to cut thin (300–500 μm) slices from fresh tissue. Samples were soaked in ice-cold sterile balanced salt solution, orientated, mounted, and immobilized using cyanoacrylate glue. To preserve tissue integrity of hollow viscera before sectioning (i.e., gastrointestinal tract), tissue was mounted on polystyrene with the luminal surface facing the Vibratome blade. Slicing speed was optimized according to tissue density and type; in general, slower slicing speed was used on the softer tissues and vice versa (0.03–0.08 mm/s neoplastic tissue; 0.01–0.08 mm/s normal tissue). Vibration amplitude was set at 2.95–3.0 mm.

Organotypic Tissue Cultures.

Tissue slices were cultured on organotypic inserts for up to 120 h (two slices per insert; Millipore). Organotypic inserts are Teflon membranes with 0.4-μm pores that allow preservation of 3D tissue structure in culture. Tissue culture was performed at 37 °C in a 5% CO₂ humidified incubator using 1 ml of Ham F-12 media supplemented with 20% inactivated FBS (GIBCO), 100 U/mL penicillin (Invitrogen), 100 μg/mL streptomycin (GIBCO), 2.5 μg/mL amphoterycin B, and 100 μg/mL of kanamycin (Sigma Aldrich). Medium was changed every 2 days. Two tissue slices were harvested at baseline time (T0) and thereafter, at 24-h intervals; the first slice was snap-frozen for qPCR, and the second slice was formalin-fixed and paraffin-embedded for morphological (H&E) and IHC evaluation.

Tissue-Viability Analysis.

Tissue viability was assessed using an MTT 1-(4, 5-dimethyltiazol-2-yl)-3, 5-diphenylformazan assay (Sigma Aldrich). Tissue slices were incubated with 5 mg/mL of MTT at 37 °C for 4 h, harvested, and precipitated-salt extracted by incubation with 0.1 M HCl-isopropyl alcohol at room temperature for 25 min. A viability value was determined by dividing the optical density of the formazan at 570 nm by the dry weight of the explants. Baseline samples (T0) were used as calibrators (1×) to normalize intersample variation in absorbance readings, and tissue viability was expressed as a percentage of viability relative to T0 samples.

Morphological and IHC Evaluation.

H&E slides of formalin-fixed, paraffin-embedded (FFPE) material were used to assess the morphological integrity of tissue samples. Indirect immunoperoxidase analysis was performed on a representative population of tumor and matched normal cohorts for Ki67 (1:1,000; M7240; Dako), phospho-Akt (1:50; S473; Cell Signaling), and phospho-S6RP (1:100; Ser-240/244; Cell Signaling). To evaluate preservation of the epithelial–stromal relationship and assess integrity of each compartment over a time course, representative normal and tumoral tissue sections were stained for Laminin (1:50; 4C7; Dako), CD34 (1:100; QBEnd/10; Dako), cytokeratin (1:250; CKAE1-AE3; Dako), vimentin (1:5,000; V9; Dako), and EMA (1:2,000; E29; Dako). Briefly, 5-μm sections were cut, dewaxed, and incubated in absolute methanol solution with 0.3 mL of hydrogen peroxide for 30 min before antigen retrieval (3 5-min microwave cycles in sodium citrate buffer at pH 6.0). Sections were then treated with blocking serum for 10 min after which they were incubated for 2 h and up to 12 h with specific primary antibody, and protein detection was performed using peroxidase-conjugated specific secondary antibody (Dako REAL HRP secondary antibody System; Dako). Antigen–antibody complexes were detected by 3, 3′-diaminobenzidine (DAB) and counterstained with hematoxylin. Negative and positive controls were included. Each tissue section was independently evaluated by two pathologists (M.L. and G.F.). Apoptotic cells were identified by TUNEL method according to manufacturer’s protocol (ApopTag Peroxidase In Situ Apoptosis Detection Kit; Millipore).

Automated image analysis was performed using the Ariol system SL-50 (Applied Imaging); 1,000 nuclei were scored to quantify Ki-67, TUNEL, and BrdU staining. A two-score system for percentage of positive cells and intensity of staining was used to quantify p-Akt and p-S6RP immunostaining. The combined score was expressed in a scale of 0 (absent or weak staining and few positive cells) to 3 (strong staining and numerous positive cells). Data were normalized to baseline (T0) sample and expressed as the mean score ± SEM for each time point.

Real-Time qRT-PCR.

Total RNA was extracted from baseline (T0) and cultured samples using the RNeasy Mini Kit (Qiagen) according to the manufacturer’s protocol. Five hundred nanograms of purified RNA were retrotranscribed by Archive cDNA Reverse Transcription Kit, and 8 ng of cDNA were then amplified (TaqMan Technology) in a final PCR volume of 10 μL using a 7900HT instrument. All qPCR primers and probes were purchased as Assay-on-Demand (AB) (Table S10). The geometrical mean of the two most stable reference genes of three (hydroxymethylbilane synthase [HMBS] and ubiquitin C [UBC]) was used for data normalization and relative quantification (RQ) of targets expression. RQ of cultured slices is expressed as fold changes relative to baseline and indicated as RQ ± SEM. RQ values ≥ 2.5 or ≤ 0.4 were considered significant.

Small-Molecule Targeted Therapy.

Slices (300 μm) from a representative population of the tumor cohort (two lung, two prostate, and five colonic carcinomas) were cultured with 50 μM of PI3K inhibitor LY294002 (Sigma). DMSO replaced LY294002 in matched control samples. Two slices were harvested after 0 h, 24 h, 48 h, 72 h, and 96 h of ex vivo culture (T0, T24, T48, T72, and T96, respectively). Slices were assessed for morphological integrity, cell proliferation, apoptosis, gene expression, and phospho-activation of selected targets in the PI3K pathway, as described above. A dose–response curve was generated after culture of two lung adenocarcinomas and one colon adenocarinoma with 10 μM, 20 μM, and 50 μM of PI3K inhibitor LY294002 (Sigma) at T24 and T48, and morphological integrity, cell proliferation, apoptosis, and phospho-activation of selected targets in the PI3K pathway were evaluated as described above. In addition, 400-μm slices of two dedifferentiated liposarcomas were cut and cultured in the presence of 10 μM of the Mdm2 inhibitor Nutlin-3 (Cayman Chemical) for 24 h. Slices treated with DMSO were used as controls. Six hours before harvesting the tissue, 10 μM BrdU (Upstate) was added to the medium. Tissue slices were then formalin-fixed and paraffin-embedded for morphological and immunohistochemical evaluation (BrdU, 1:100; Becton Dickinson; p21, 1:100; BD Transduction; Mdm2, 1:50; Calbiochem), according to the methods described above.

cDNA-Mediated Annealing, Selection, Ligation, and Extension Assay.

Two cases of lung adenocarcinoma, cultured for up to 48 h and subsequently formalin-fixed and paraffin-embedded, were chosen to study global gene expression over time. Tumor and stroma foci (100% each respectively) were microscopically identified by a study pathologist (R.F.). An average of four 0.6-mm biopsy cores was taken from each area respectively per sample at T0, T24, and T48. Technical replicates were included in the analysis as well for a total of 12 samples consisting of 6 tumor and 6 stroma samples. Total RNA was extracted, cleaned up, and eluted using the RNeasy Mini Kit and RNeasy MinElute Cleanup Kit, respectively (Qiagen), according to the manufacturer’s protocol. The average total RNA yield per sample was 430 ng and ranged from 60 ng to 1,117 ng. cDNA-mediated annealing, selection, ligation, and extension (DASL) expression assay (Illumina) was performed as per manufacturer’s instructions (21) on a panel comprising 24,526 probes for a total of 18,631 genes (Human Ref-8 Expression; BeadChip).

Bioinformatic Analysis of DASL Gene-Expression Profiling.

Data were normalized using the cubic spline method available in GenePattern (22).

Permutation test for stability of gene expression.

To assess the stability of gene expression over time, we performed permutation testing separately for the tumor and stroma samples. For each gene g, the standard deviation σ_g of its expressions in six tumor samples, which spanned 0 h, 24 h, and 48 h, was computed as a measure of variation over time. We then generated a null population of 100,000 randomized tumor-expression profiles by permuting the expression values for each of the six samples 100,000 times, thus destroying any stability of expression over time in the null population. Then, we computed the P value for g’s conserved expression as the proportion of randomized profiles that had SD less than or equal to σ_g. Similarly, we also computed the P value for every gene's conserved expression over time in the stroma samples. Both the tumor and stromal P values were adjusted for FDR control by the Benjamini-Hochberg method.

Differential gene-expression analysis.

We applied two-sample t test to identify the differentially expressed genes between tumor and stroma samples. To discover the pathways that are statistically most significant across the tumor and stroma expression profiles, we used the method of Tian et al. (19) as implemented in the sigPathway package of BioConductor. In addition, we also searched the major molecular signature database MSigDB (23) for further cancer-related genesets and pathways. For each given pathway, we chose the top five differentially expressed genes and compared their loess-smoothed mean tumor and stroma profiles. Furthermore, to statistically evaluate the difference of the mean profiles of these top five genes between tumor and stroma, we used Hotelling's T² test using the R package hot (24).

Statistical Analysis.

GraphPad Prism 4.0 software was employed for all remaining statistical analyses. Each statistical test is specified in the text.