Preclinical optimization of a broad-spectrum anti-bladder cancer tri-drug regimen via the Feedback System Control (FSC) platform

Therapeutic outcomes of combination chemotherapy have not significantly advanced during the past decades. This has been attributed to the formidable challenges of optimizing drug combinations. Testing a matrix of all possible combinations of doses and agents in a single cell line is unfeasible due to the virtually infinite number of possibilities. We utilized the Feedback System Control (FSC) platform, a phenotype oriented approach to test 100 options among 15,625 possible combinations in four rounds of assaying to identify an optimal tri-drug combination in eight distinct chemoresistant bladder cancer cell lines. This combination killed between 82.86% and 99.52% of BCa cells, but only 47.47% of the immortalized benign bladder epithelial cells. Preclinical in vivo verification revealed its markedly enhanced anti-tumor efficacy as compared to its bi- or mono-drug components in cell line-derived tumor xenografts. The collective response of these pathways to component drugs was both cell type- and drug type specific. However, the entire spectrum of pathways triggered by the tri-drug regimen was similar in all four cancer cell lines, explaining its broad spectrum killing of BCa lines, which did not occur with its component drugs. Our findings here suggest that the FSC platform holdspromise for optimization of anti-cancer combination chemotherapy.


Results
Bladder cancer cell lines vary considerably in their chemoresistance properties. We used the MTT (Test by thiazolyl blue tetrazolium blue) based assay to determine the dose required for 50% of cells to be killed (IC50) by following six anti-bladder cancer drugs (https://www.nccn.org/store/login/login. aspx? ReturnURL = http://www.nccn.org/professionals/physician_gls/pdf/bladder.pdf):Pirarubicin (Pi), Paclitaxel (Pa), Epirubicin Hydrochloride (EH), Cisplatin (Ci), Gemcitabine (Ge), and Mitomycin (Mi) in eight human bladder cancer cell lines (7 TCC lines: T24, Biu87, EJ, J82, UM-UC-3, 5637, and H-bc and one squamous carcinoma cell line (SCaBER) and an immortalized benign cell line (SV-HUC-1) ( Fig. 2A-F). The relative IC50 (fold) of each drug of all nine cell lines was normalized to the lowest IC50 in the most sensitive cell line (Fig. 2G), to describe the single drug resistance of cell lines. The overall chemoresistance of each cell line, a "chemoresistance index" (the numerator of the total relative IC50 over the number of drugs) (Fig. 2G) was calculated to rank the overall resistance state of each cell line to these six chemotherapeutics. The results from the IC50 profiling of these nine cell lines (Fig. 1) illustrated the considerable heterogeneity in drug resistance among BCa cell lines, despite an exception: an almost uniform sensitivity of all cell lines to Cisplatin. 5637 exhibited the highest chemo-sensitivity to four of six drugs and overall chemo-sensitivity for all six drugs ("chemoresistance" index: 1.62). H-bc was the most chemoresistant to three of six drugs, with an exceedingly high relative IC50 to GE (3976.1), which makes it the most chemoresistant (thechemoresistance index: 678.68). It was the second highest in the cancer cell line (thechemoresistance index: 19.2) when GE was taken out of consideration (Fig. 2G), to SCaBER ("chemoresistance index": 33.88). SV-HUC-1, an immortalized benign epithelial cell line exhibited the highest IC50 to each of these five drugs and therefore the highest chemoresistance index: 147.28, which is almost 3.5 fold higher than SCaBER's ( Fig. 2G). Therefore, these five drugs have a high level of selective cancer cell killing capability.  The relative cell survival (%) of the drug treated over the no drug treated cells (Y-axis) was plotted against the dose (ng/mL, lg) of the following six drugs (X-axis): Pi, Pa, EH, Ge, Ci, and Mi, respectively, and the curves were plotted for SCaBER, T24, Biu87, J82, EJ, 5637, UM-UC-3, H-bc, and SV-HUC-1 cell lines, respectively (panel A-F). (G) The relative IC 50 (fold) of each drug with the lowest IC 50 (ng/mL, lg) as reference was calculated.
consists of 4 iterative operations: 1) starting with a randomly selected drug-dose combination, 2) experimental assessment of combination-mediated cell killing efficacy, 3) optimization of the combinations via the DE algorithm, and 4) completion of the feedback control loop by testing the newly selected drug-dose combination until an optimum is reached (Fig. 3).
Twenty six-drug combinations at the indicated doses were initially generated through a random number generator (using MATLAB MathWorks©, Natick, MA, U.S.) (The upper half of Table 1). The cell survival (% over the no drug control) of treated cells was determined and then used for a calculation of the average cumulative cell survival (%, ACS: the sum of cell survival (%) per combination divided by the number of cell lines tested). ACS < 30% was used to separate the effective from the not effective combinations. Nine of the twenty combinations tested in the 1 st round of iteration fell into this category (the lower half of Table 1).
The ACS of each combination of the 1 st round testing was fed into a DE algorithm scheme to generate a new list of combinations for the next round of testing. Four (4) of the 20six-drug combinations tested in the 2 nd round of iteration were effective (Table S2). To reduce the number of drugs in combinations from six to four or five, a zero dose option was introduced in the 3 rd and the 4 th round of testing for four or five drug combinations. The number of the effective combinations was raised from the 2 nd round (25%, 4 out of 20) (Table S2) to the 3 rd round (45%, 9 of 20) (Table S3), and the 4 th round (82.5%, 33 out of 40) of testing (Table 2), demonstrating improved optimization with the "evolution" of drug combinations (Fig. S2). More specifically, combination #40 (EH 250 ng/mL, highest dose; Ci, and 1500 ng/mL highest dose; Mi, and 200 ng/mL at the second highest dose, and Pi at the lowest dose, 0.32 ng/mL) produced the lowest ACS (13.60%) ( Table 2). Pi, in this combination was at the lowest concentration and did not have any apparent impact on maintaining the efficient cell killing abilities of this regimen, as suggested by a comparative analysis with other Pi inclusive combinations (Table 1,2, Table S2-3). Therefore, we removed Pi and selected the combination 40 minus Pi (EH/Ci/Mi) as the candidate combination for further refinement.
Statistical modeling analysis confirmed EH/Ci/M as the optimal tri-drug combination. For additional support, we performed statistical modeling 28 analysis of all the features of the 100 combinations (Tables 1-2 and S1-2) from testing of the following four cell lines: 5637 (sensitive), H-bc and UM-UC-3 (resistant), and Biu87 (moderately resistant) (Fig. 3).
For a given cancer cell line treated by six (k) different drugs (Pi, Pa, EH, Ci, Ge, and Mi), the system output, y, can be expressed by a 2 nd order polynomial regression model of drug doses x i . The data confirm that higher than 2 nd order terms do not make significant contributions to the efficacy, as was previously demonstrated in a lung cancer model 29 .
where β 0 , β i , β ii and β ij are the intercept, linear, quadratic, and bilinear (or interaction) terms 30 . The "observed" values represent the experimental data (ACS) and the "fitted" results represent the model predictions for the same drug combinations. A 2 nd order polynomial regression model fits the experimental results of Biu87, 5637, UM-UC-3, and H-bc which were plotted for comparison (Fig. 3). We further examined the regression coefficients in the statistical modeling for each of these four cell lines. For instance, the polynomial regression model for Biu87 is: where Biu87ACS represents the ACS for Biu87 cells under different drug treatments and Pi, Pa, EH, Ci, Ge, and Mi in the equation (2) represent the absolute dosages of each drug (ng/mL). The model has a R 2 value of 0.7015 and p-value equal to 2.439e-15, which indicates a good fit between the model and experimental observations. The chemotherapeutic agent Ge has a negative linear regression coefficient but the largest positive quadratic regression coefficient. This implies that a dose increase of Ge is less likely to enhance drug-triggered cell death when it is used in combination with other drugs. Ge also  EH  250  250  50  50  250  50  50  50  250  0  2  2  2  2  50  2  50  50  0  0   Ci  1500  12  60  300  1500  60  1500  60  1500 1500  300  1500 1500  0  0  60  60  300 1500 300 has the largest positive two-drug interaction coefficient with Mi, implying that Ge and Mi would have antagonistic effects in reducing cell survival. Therefore, we eliminated Ge from further consideration. The regression model for the H-bc cell line is:  Table 3) were determined in the following TCC cell lines: Biu87, T24, 5637, UM-UC-3, and H-bc. Again, the EH/Ci/Mi regimen (1500 ng/mL Ci, 200 ng/mL Mi and 250 ng/mL EH,) showed the lowest ACS value (7.97%) among all of the four tri-drug regimens (combinations [11][12][13][14] which was finally examined in all eight BCa cell lines (Table S3). The EH/Ci/Mi regimen (ACS: 6.32%, Table S4) not only delivered more cell killing than other tri-drug regimens: Pi/Ci/Mi and Pa/Ci/Mi (Table S4), but was also more effective than the component drugs at four fold doses: Ci (6000 ng/mL, ACS: 16.60% ± 13.53%), Mi (800 ng/mL, ACS: 40.49% ± 21.63%) or EH (1000 ng/mL, ACS: 9.11% ± 7.29%), respectively (or Pi, 160 ng/ml, 21.32 ± 12.55%, or Pa, 640 ng/ml, 20.89 ± 12.73% Fig. S4A). This observation suggests a synergistic action of the three drugs when they are used in combination. Furthermore, the EH/Ci/Mi regimen possesses a broad spectrum of anti-bladder cancer activity, by killing between 82.86% and 99.52% of BCa cells in comparison to only 47.47% of SV-HUC-1 cells (Table S4). However, Ci is the most toxic to the immortalized normal cell line (SV-HUC-1), resulting in 73.79% cell death in comparison with EH (38.11%), Mi (8.83%), Pi (8.21%), Pa (19.83%), and EH/Ci/Mi (52.53%) cell death. Taking both the anti-BCa cell line (ACS%) and toxicity on the immortalized epithelial cells (Cell survival %) into account, we created the therapeutic window (TW) ( Table S4 and Fig. S3), referring the difference (%) in the treatment trigger cell survival between SV-HUC-1 and BCa cell lines and the therapeutic window index (TW index), referring to the difference (%) between the therapeutic window and the ACS % to rank the EH/ CI/Mi regimen and three component drug at a threefold higher dose (Table S4 and (Table S5-S9). The relative pathway activity between more resistant cell lines: Biu87 (Table  S6), UM-UC-3 (Table S7), H-bc (Table S8), and SV-HUC-1 (Table S9) over the most chemo-sensitive cell line, 5637 (Table S5) was determined and analyzed (Fig. 4D). Eight pathways exhibited higher activities in Biu87 and SV-HUC-1 than in 5637 and 7 pathways showed higher activity in UM-UC-3 and H-bc cells than in 5637, by more than 0.5 fold (Fig. S2D). Therefore, a higher level of the basal activity of most pathways tested positively correlated with the chemoresistant state of BCa cells.
We then determined the correlation coefficient between the overall activity of these nine pathways and the chemoresistant state of these five cell lines. The pathway activity was digitized as follows: (1) an active state: the signaling pathway with an activity higher by 0.5 fold than in 5637; (− 1) an inactive state: the signaling pathway with an activity lower by 0.5 fold than 5637; and (0) an inert state: the signaling pathway with an activity being neither of the above. The overall state of activity of these pathways in each cell line were summarized (Fig. 4E) and the correlation coefficient of each pair of cell lines was calculated (Fig. 4F) (see Supplementary material 4 the Pearson correlation analysis).
The correlation coefficients of the overall nine pathway activity between 5637 and four chemoresistant cell lines were negative (− 0.397, − 0.5, − 0.5, and − 0.397), and indicating the distinct state of 5637 from the other four cell lines, these figures in each pair were close to 0.99 (Fig. 4F). It was anticipated that the drug-triggered pathway activities also reflect the extent of chemoresistance of the cells. Cells were Scientific RepoRts | 5:11464 | DOi: 10.1038/srep11464 subjected to a 6 hrs treatment with either the EH/Ci/Mi regimen (EH,250 ng/mL, Ci, 1500 ng/mL and Mi, 200 ng/mL) or each component drugs at a fourfold dose: EH (1000 ng/mL), Ci (6000 ng/mL) and Mi (800 ng/mL) as well as two other drugs. The EH/Ci/Mi regimen's ACS for BCa cells was significantly lower than that of each fourfold dosed component drugs (and the other two drugs), while it was higher than a fold dose of Ci for SV-HUC-1 cells than a fourfold dosed Ci (Table S4). The 6 hrs post-drug-triggered pathway activities were measured (Table S5-S9), normalized with the basal level activities (Table 4), digitalized (Table 5), and summarized (Table 6) for the correlation analysis of pathway activities by each single drug treatment and the EH/Ci/Mi regimen (Table 7). Taking 0.917 as the cutoff, the EH/Ci/Mi regimen's effect on the pathway activity was the same as that of all the three component drugs in 5637 cells. On the contrary, only one component drug (Mi, in UM-UC-3, EH, in H-bc) or two drugs (Mi and EH, in Biu87) triggered pathway activity that was the same as the EH/Ci/Mi regimen's. This observation suggests that the ability of the EH/Ci/Mi regimen to evoke the pathway response by one or more fourfold dosed single drugs may form the mechanistic basis for its a broad spectrum of cancer cell killing. Although the analysis of this level grouped 5637 and SV-HUC-1 (the most resistant) cell lines together, the detailed difference is obvious. The numbers of drug-activated pathways were from 5-8 in 5637 in comparison to 3-4 in SV-HUC-1 cells, while the numbers of inert pathways to drug treatment in 5637 were 1-3 versus 5-6 in SV-HUC-1 cells (Table 5 and Table 6).
The EH/Ci/Mi regimen inhibited the in vivo growth of bladder cancer cell line derived tumor xenografts in nude mice more effectively than its mono-drug or bi-drug counterparts. We then compared the in vivo anticancer activity of the EH/Ci/Mi regimen with its mono-and bi-drug counterparts on the 5637, UM-UC-3, or Biu87 cells derived tumor xenografts in nude mice. Since there is no established conversion rule for the drug dose in cell culture to in nude mice, we used 75 μ g per mouse for the mono-drug treatments and 25 μ g per drug per mouse for both bi-drug and tri-drug  0  0  160  0  0  0  32  0  0  0  0  160  0  0  0  0  160   EH  0  250  0  0  0  50  0  0  0  0  250  0  0  0  0  250  0   Ci  1500  1500  1500  1500  1500  1500  1500  1500  1500  1500  1500  1500  1500  1500  300  300  300   Ge  0  0  0  0  5  0  0  0  1  0  0  0  0  5  0  0  0   Mi  1000  1000  1000  1000  1000  1000  1000  1000  1000  200  200  200  200  200  1000  1000  1000 Cell survival(%)          Table 5. Digitalized analysis of relative drug trigged pathway activities over the no-drug level of each cell line. The pathways with their activity higher by 0.5 fold or more are marked as 1 in yellow, and these with activity lower by 0.5 fold are marked as − 1 in green.  combinations, based on the maximum tolerated drug dose recommended for these drugs [31][32][33] . To minimize the inter-mouse bias's effect on tumor growth, three different BCa cell line-derived tumor xenografts were individually subcutaneously established on the back of the twelve mice (Fig.5A). A group of three mice were intraperitoneally injected with PBS (Phosphate balanced solution, the no-drug control), Ci, Ci/Mi, and the EH/Ci/Mi regimen on day 7 (after injection of the cancer cell lines), and at a three-day interval four more times. The tumor volumes were measured once every three days and were normalized by the volume on day 7. As shown in Fig. 5B (Fig. 5B). The relative tumor volume of 5637 cells in the EH/ Ci/Mi group (2.54) was lower than that in Ci (6.2) and Ci/Mi groups (4.95) (Fig.5B). The conclusion of the better anti-cancer effect of the EH/Ci/Mi regimen than Ci and Ci/Mi regimens was further supported by the different tumor masses on day 28 when the animal study ended. Taking the tri-drug group as 1, relative tumor weights in the PBS group were 5.85, 1.14, and 1.84 for UM-UC-3, 5637, and Biu87 derived tumor, and in Ci and Ci/Mi groups were 3.37, 1.79, and 1.48 and 4.87, 1.88, and 1.67, for UM-UC-3, 5637, and Biu87 derived tumor, respectively (Fig. 5D). Elevated levels of both Ki67 and CD34 proteins indicate active cell proliferation 34,35 and increased blood vasculature formation 36 in the tumor mass. The percentage of Ki67-positive cells and the numbers of CD34 positive vascular structures in tumor xenografts were reduced by Ci, Ci/Mi, and EH/Ci/Mi treatments relative to PBS controls ( Fig. 5F-I).
The highest reduction was with the tri-drug combination and the lowest with the single drugs, which is consistent with the effects on tumor growth of this tri-drug combination.

Discussion
Comprehensive cancer genomic studies in the last few years have repeatedly demonstrated that genetic heterogeneity in cancer cells differentiates not only in tumors from different individuals, but also different lesions, different part in a single lesion, or different cancer cells in a cancer patient 37,38 . This 39 has been regarded as the major cause of failure in cancer treatment. Therefore, optimizing a multi-drug regimen capable of killing a broad spectrum of cancer cells with different single drug-resistant profiles should be done on a panel of established cancer cell lines to capture disease diversity in clinics. BCa consists of two major pathological subclasses, approximately 90% of cases in transitional cell carcinoma and 6% of cases in squamous cell carcinoma. In this study, seven TCC and one squamous carcinoma cell lines that dramatically vary in chemoresistance to five of six drug exposures (Fig. 2G) were used, with the intent to better reflect the heterogeneous spectrum of BCa. We used the FSC platform, DE algorithm guided experimental testing (Fig.1) to assess the cell killing capability of 100 out of 15,625 possible combinations of a set of six drugs at five doses in four rounds of testing. A tri-drug regimen (EH/Ci/Mi, 1500 ng/ mL Ci, 200 ng/mL Mi, and 250 ng/mL EH) was able to kill 82.86% to 99.52% of cancer cells of 7 TCC and, 93.75% of 1 squamous BCa cell line in comparison with 52.53% of the immortalized normal cells (Table S4). The ideal state of the EH/Ci/M regimen was confirmed from a statistic modeling analysis of the ACS readouts of 100 combinations tested (Fig. S1). This tri-drug regimen killed more BCa cells at a 3 fold lower concentration than that of each component drugs (Table S4 and Fig. S3), indicating a synergistic interaction among these three drugs when used together. We also showed that the EH/Ci/Mi regimen was more potent than mono-drug and bi-drug combinations in inhibiting the growth of tumor xenografts derived from 5637, UM-UC-3, and Biu87 cell lines in nude mice (Fig. 5). As shown in Table  S4, the tri-drug combination killed bladder cancer cells more efficiently than single components, and provided a kill range from 0.48% (H-bc) to 17.14% (EJ). Interestingly, the tri-drug combination dramatically killed a high rate of resistant cell H-bc, and thus may be able to cure some therapeutic resistant breast cancers. Since the EH/CI/Mi regimen was the most favorable with TW (41.15%) and TW index: (34.83%, and Table S4 and Fig. S3), it is expected to be able treat a broad spectrum of bladder cancer patients than any single drug. The current anti-bladder cancer combination chemotherapies are chiefly cisplatin-based: CMV (cisplatin, methotrexate, and vinblastine), M-VAC (methotrexate, vinblastine, adriamycin, and cisplatin) 40    and GC (gemcitabine plus cisplatin) [41][42][43] . Recent reports suggest a combined use of EH and Mi for BCa treatment [44][45][46] . EH's mechanism of action is different from Ci and Mi 46-48 that covalently bind DNA and repress both replication and transcription of DNA, leading to cell death [49][50][51][52] . As an intercalating agent, EH affects the secondary structure of DNA and therefore inhibits transcription and induces apoptosis 53,54 . Thus, use of these drugs together at a lower dose can simultaneously target different mechanisms for cancer cell killing along with a reduced level of cytotoxicity of the normal cells as described in this study.
To understand why the EH/Ci/Mi regimen is capable of killing a broad spectrum of BCa cells but each component drug at a three fold higher dose failed, we analyzed the basal and drug-triggered activities of nine cancer associated signaling pathways in four bladder cancer cell lines (5637, UM-UC-3, Biu87, and H-bc) and an immortalized non-transformed bladder epithelial cell line (SV-HUC-1) ( Table. S4): DNA damage response, Hypoxia, ER Stress, Heat Shock, Wnt, Notch, Cell cycling, Myc/Max, and MAPK/ERK pathways (Table 4-7 and Table S5-S9). Consistent with the chemoresistance of these cell lines, the basal level activities of these signaling pathways were collectively higher in chemoresistant cell lines (UM-UC-3, Biu87, H-bc, and SV-HUC-1) than the most sensitive cell line, 5637 ( Fig. 4 and Table S5-S9). We also found that the subsets of pathways activated by the tri-drug combination were similar to at least the set of pathways triggered by the dominant drug in all cell lines tested (Table 4-7 and Table. S5-S9). This observation offers a mechanistic explanation for why the EH/Ci/Mi regimen can effectively kill all the cancer cell lines tested, but the component drugs at a fourfold doses failed.
By identifying the EH/Ci/Mi regimen as a particular example for discovering effective BCa specific chemotherapeutic regimens, we demonstrated both robustness and general applicability of the FSC platform to develop effective drug (chemotherapeutics, biologicals, and target therapeutics) combination therapy for cancer. . Drug induced cell death was determined by a thiazolyl blue tetrazolium blue (MTT, 490 nm reading) based cell proliferation assay in the cultured cells following treatment by various concentrations of drugs for 72 hours, as previously described 60 . The relative chemoresistance to each drug of all cell lines was calculated with the lowest IC 50 as a reference. The overall chemoresistance of each cell line was described as the "chemoresistance index", the numerator of the total relative IC 50 to all the drugs over the number of drugs to rank the overall resistance state of each cell line to these six chemotherapeutics collectively.

Optimization of the tri-drug regimen via the FSC platform in eight BCa cell lines. "Differential
Evolution" algorithm guided experimental optimization (DE) consists for three iterative operations: generating drug combinations for experimental testing, acquiring the efficacy of the generated drug combinations from biological assays, and recalculating further improved drug combinations using a feedback search algorithm. We first tested twenty drug combinations at randomized doses. These drug combinations were generated from a random integer matrix generator, coded in MATLABTM language. These twenty drug combinations were then experimentally tested for cancer cell cytotoxicity. The experimental data were then subjected to calculation with a DE algorithm for new drug combinations based on three major mathematical calculations: "Mutation", "Crossover" and "Selection". The equations for each step can be found in the main text. The equations for each step are coded in MATLABTM language.
Statistical Modeling on the relationships between cell survival and drug dose. Statistical analysis of experimentally measured cell survival rates was done for each individual cell line. The readouts (ACS) from 100 tested options for each cell line were pooled together and a polynomial regression model was built. The polynomial regression modeling is commonly used to mathematically describe the relationship between Y's (in our case, the cell death) and X's (in our case, the drug doses) using the R language environment. The polynomial regression model was computed by function "g = lm{stats}" and simplified by function "g = step(stats)" in R programming. The pathway analysis. The pathway analysis was performed with the Qiagen's Cignal Finder Pathway Reporter Array systems according to the manufacturer's instructions. The pathway-focused dual-luciferase reporter construct is a mixture of an inducible transcription factor responsive firefly luciferase reporter and constitutively expressing Renilla luciferase construct. Transcriptional response element (TRE) sequence response elements to each specific transcriptional factor triggered by the different signaling pathways are positioned before the TATA box of the firefly luciferase reporter gene. The extent of activity of the reporter reflects the level of activity of the particular signaling pathway. The CMV controlled Renilla luciferase expression vector is co-transfected into the cell as an internal control. A basic reporter construct that lacks the transcriptional factor binding sequence and could not be induced by any transcription factors is used as a negative control. Cells transfected with the CMV directed firefly luciferase reporter gene serve as a positive control. Briefly, cells seeded in 6-well plates were transfected with Cignal finder pathway arrays according to the manufacturer's instructions (Qiagen) 61 . Then cells were seeded onto 96-well plates. Ci (6000 ng/mL), Mi (800 ng/mL), EH (1000 ng/mL), or the Ci/Mi/EH combination (Ci, 1500 ng/mL, Mi, 800 ng/mL and EH, 1000 ng/mL) were then added to the cells 24 h post-transfection. After 6 h of incubation with the different drugs at 37 ºC, cells were lyzed and a dual luciferase assay (for firefly and Renilla luciferase) was performed using a Promega GloMax 20/20 illuminometer. The firefly luciferase activities were normalized over the Renilla luciferase activities after background activity determined by luciferase activities of negative controls was subtracted. A triple test of cell survival under these drug doses was also performed by MTT assay following 18 h incubation at 37 ºC.
Pearson correlation analysis of the signaling pathway data in chemoresistant distinct cell lines. The levels of pathway activities were digitized as follows: activated (1): if pathway activity was higher than 1.5-fold over the activity in 5637 cells (Table 5) or than the no drug control; repressed (− 1), if the activity was 0.5-fold lower than in 5637 cells (Table 6) or than the no drug control, and null (0) if neither of the above. Correlation coefficients between two cell lines or between drug treated cells and untreated cells were calculated ( Table 7). The definition and calculation equations for correlation coefficients can be found in the main text. To calculate the correlation coefficients between the two cell lines we used R language statistics. Function "cor{stats}" was used to generate the correlation matrix between multiple cell lines.
The in vivo study. Cells (1 × 10 6-7th /100 μ L) embedded in BD Matrigel ™ Matrix 62 were subcutaneously injected at three sites in the back of each mouse (3-4 mice per treatment). On days 7-12 following cell injection, the animals were treated with the different drugs by an intraperitoneal injection, which was repeated every three days. Tumor volumes were monitored/calculated using the equation V = W 2 L × 0.5 (where W and L represent the largest and second largest tumor diameters (mm)) and then plotted. The tumor mass were weighed, and fixed/sliced for immunohistochemistry after the animals were sacrificed. Immunohistochemical analysis. Immunostaining was performed on 5 μ m slices of formalin fixed paraffin-embedded tumor xenografts using antibodies from Gene Tech Company (Shanghai, China) against both Ki67 and CD34. Pictures were taken by a LEICA DM 4000B microscope and the relative level of each protein was calculated using LEICA software. The relative staining per antibody (in percentage) in the mock treated over chemotherapeutic treated tumors was calculated and plotted. Statistical Analysis. Two sets of tumor xenograft/nude mice studies were carried out. The in vivo study involved four groups with three mice per group. Both mean ± S.D of tumor volume and weight were calculated. Student's t-test was used for statistical comparisons. P values less than 0.05 were considered statistically significant.