The ability of gene signatures to predict complex biological phenomena appears to be limited, and some biological endpoints have been shown to be inherently difficult to predict regardless of the study design and bioinformatics methods employed 2 , 13 , The predictive signatures generated thus far have either not been validated in subsequent studies or offered limited predictive value in addition to that provided by standard clinico-pathological parameters 1 , 4 , 15 - This limited success in the development of predictive signatures can be attributed to biological phenomena and technical issues, including pharmacokinetics variability that may not be entirely captured by expression profiling of primary tumors reviewed in 18 , weakly informative features i.
These convergent phenotypes pose a challenge for the development of predictive signatures, as tumors with different resistance mechanisms may display either completely different or only partially overlapping gene expression patterns 4 , 17 , 25 , and conventional methods of genome-wide microarray analysis may only be able to identify genes significantly altered in the majority of therapy-resistant or sensitive tumors in a given dataset In studies aiming to derive gene expression predictors of response, resistant samples have been treated as a single, homogeneous group without the knowledge of the underlying mechanisms of resistance Hence, we sought to determine the impact of the existence of multiple mechanisms of resistance to a hypothetical therapeutic agent or combinatorial therapy on the performance of predictive gene signatures.
Furthermore, we assessed in actual datasets of breast cancer patients who underwent neoadjuvant chemotherapy whether sub-stratification of the chemotherapy-resistant cases improved the performance of the predictive signatures generated. We selected nine breast cancer gene expression datasets generated on the Affymetrix Ua2 platform comprising 1, cases from Haibe-Kains et al.
The resulting merged dataset showed no signs of bias resulting from batch effects data not shown. Next, we generated bioinformatically perturbed datasets using this merged dataset by spiking in arbitrarily-selected resistance-associated gene expression changes i. Using this approach, we have defined bioinformatically the genes associated with resistance to the hypothetical drug or combinatorial therapy, and the cases classified as resistant or sensitive.
For each combination of s, v and n , we repeated the perturbation steps to generate bioinformatically-perturbed datasets. Using the same methods, we also simulated datasets for other scenarios. First, we generated iterations where there were 2, 3, 4 or 5 resistance mechanisms n , for which the proportions of resistant cases driven by a pre-determined number of resistance mechanisms i.
Second, we generated iterations where there were 2, 3, 4 or 5 resistance mechanisms n , for which the proportions of resistant cases were identical in the training and test sets, however, the proportion of cases driven by a given mechanism of resistance was randomly and independently allocated for the training and test sets. Perturbed datasets were generated using microarray-based gene expression profiles of 1, breast cancer cases analyzed with the Affymetrix Ua2 platform.
For illustration purposes, we assumed up to three resistance mechanisms i. Predictive signature models were derived by ranking the features probes by t-tests using the CMA package. The top features were then used as the predictive gene signature for diagonal linear discriminant analysis DLDA or supervised principal components superPC classification. Validation of the predictive gene signature was performed by stratified 3-fold Monte-Carlo cross-validation, repeated 50 iterations.
Comparing the predicted and actual classes, we calculated the area under curve of receiver operating characteristic curves, sensitivity, specificity, accuracy, positive predictive value and negative predictive for each predictive gene signature.
For each combination of variables, we repeated the spiking-in and classification up to times. The top features were then used as the predictive signature for DLDA. Feature selection was performed by ranking the features using Wald score. The top features were selected as the gene predictive signature, the optimal number of principal components up to 3 was selected by cross-validation of the training set and a predictive signature was defined by superPC.
For both DLDA and superPC, validation of the predictive signatures was performed by 50 iterations of 3-fold Monte-Carlo cross-validation MCCV , stratified to preserve the proportions of the different groups of sensitive and resistant cases. Semi-stratified 3-fold MCCV was performed when only the sensitive to resistant ratio had to be preserved but not the proportion of cases driven by a given mechanism of resistance between training and test sets.
For each analysis performed, as performance indicators, we measured the area under curve AUC of the receiver operating characteristic ROC curves, sensitivity, specificity, accuracy, positive predictive value PPV and negative predictive value NPV by taking the median of the MCCV repeats, and selected distributions for illustrative purposes. We performed two types of statistical analyses.
First, we performed a trend test to calculate the statistical significance of the linear slope fitted to the logits of the AUCs y — dependent values , the logits of the AUCs being inverse variance weighed and the independent values set to the integers 1,2,3,4 or 5 or 1,2,3 or 4 depending on the number of data points.
This test is used to calculate the statistical significance of increasing the number of resistance mechanisms. Standard errors for differences were calculated by dividing the difference between the confidence limits and the mean by 1. On this basis, this rule was employed to define statistically significant differences between different classifiers generated. To assess the impact of multiple resistance subgroups on predictive signature performance, we employed two actual i.
Normalized gene expression data from these studies were obtained from Hatzis et al. To avoid the impact of proliferation-related genes on the ability to define chemotherapy response predictors, only ER-negative breast cancers were included in the analysis, as these consistently display high levels of proliferation-related genes 1.
Predictive signatures were derived using pCR as a surrogate for sensitivity to the chemotherapy regimen. Briefly, COPA transformation was performed on normalized expression values.
Using the COPA-transformed scores, we defined over-expressed resistant outliers as features greater than the 75 th percentile plus 1. Furthermore, only candidate features that displayed at least a 2-fold difference between the mean expression of the resistant outliers and the mean expression of the sensitive cases were included.
Using the same approach, candidate features using sensitive outliers were also identified by comparing the sensitive cases to the resistant cases. The features up- and down-regulated in the resistant cases and those up- and down-regulated in the sensitive cases were combined and ranked by the difference in expression between the outliers and the control group i. Features were selected using t-tests comparing all sensitive vs resistant cases within each sub-cohort of the training set, and selected features were merged from the individual sub-cohorts by ranking according to the t-statistics.
In the mixed approaches, to overcome the potential overlap of predictive genes identified by t-test and mCOPA or by t-test and the methods employed for signature generation using the clinical parameters, the genes that compose the final signature were selected by iteratively adding one feature at a time such that the proportion of genes not shared by the two sources of features was maintained. Validation of the predictive signatures was performed by leave-one-out cross-validation LOOCV of the training set.
The R scripts and codes employed for the analyses described are available as a Supplementary file. In a scenario where distinct and equally prevalent mechanisms of resistance would result in optimal i. In an ideal setting i. Employing a more realistic clinical estimate i. Increasing the proportion of resistant cases from the ideal to the clinically-realistic settings i.
For each combination of s, n and v , we repeated the spiking and classification times. Impact of distinct mechanisms of resistance on the area under the curve AUC of receiver operating characteristic ROC curves derived with the predictive gene signatures generated.
Classification was performed using diagonal linear discriminant analysis DLDA. Gene expression changes associated with sensitivity or resistance to a given therapeutic intervention have been shown often to be weaker than 2. The impact of multiple mechanisms of resistance on the performance of the predictive signatures was less pronounced but still statistically significant when the signature was strong i.
Consistent with the notion that 2-fold expression changes are optimal 13 , we observed that reducing the signature strength from 2-fold to 1.
We also investigated scenarios where the different mechanisms of resistance had an uneven and randomly determined prevalence, but identical distributions in the training and test sets. As observed when each resistance mechanism was equally distributed in the resistant population, increasing the number of unevenly distributed resistance mechanisms reduced the AUCs.
Chemotherapy forms an important part of a successful treatment regimen; however, half of the patients may fail to benefit from this, as a result of drug resistance 3. Thus, chemoresistance constitutes a major clinical obstacle for the successful treatment of breast cancer. Studies have shown that miRNAs are often aberrantly expressed in human cancers and are associated with tumorigenesis, metastasis, invasiveness and drug resistance 5 — 8.
For example, miR is downregulated in gastric cancer and inhibits cell migration and invasion through targeting CSF1 5. Most recently, Zhang et al identified five miRNAs miRb-5p, miRp, miRp, miRb-5p and miRp as candidate blood biomarkers in breast cancer patients 9.
However, studies concerning chemoresistance-associated miRNAs in breast cancer based on human tissues are scarce. The genes targeted by DE-miRNAs were predicted, and their potential functions were analyzed by functional and pathway enrichment analysis. Furthermore, a protein-protein interaction PPI network of the predicted target genes was constructed. Potential transcription factors that may regulate the target genes were screened. Through these comprehensive bioinformatic analyses, the present study aimed to explore the molecular mechanisms underlying breast cancer chemoresistance and identify important miRNA therapeutic targets.
The chemotherapy drugs used for treating the breast cancer patients are provided in Fig. A Data are presented as a heat map. FC, fold change. B The chemotherapy drugs used for treating the breast cancer patients. Data were analyzed by subtracting the background and then the signals were normalized using a locally-weighted regression LOWESS filter Mean values of each group were used in the cluster analysis by HemI software Heatmap Illustrator, version 1. The miRWalk2. In the present study, miRWalk2.
Only the common target genes predicted by at least nine types of software were selected, which were defined as potential target genes. The degree of connectivity in networks was analyzed using Cytoscape software version 3. GO functional and KEGG pathway enrichment analyses were performed on the aforementioned potential target genes. The enriched GO functions for the target genes are presented in Tables I and II , including the positive regulation of gene expression, positive regulation of transcription DNA-templated and positive regulation of cellular biosynthetic process in the biological process BP category; cell junction and nuclear transcriptional repressor complex in the cellular component CC category; and transcription factor activity and enzyme binding in the molecular function MF category.
For downregulated miRNAs Fig. The top 10 enriched pathways are presented. For better visualization, the top 10 hub nodes with higher degrees were screened Table III. As shown in Fig.
The top 10 most significant transcription factors are presented. Validated miRNA-gene interactions in breast carcinoma. B The validated miRNA-gene network was constructed.
Chemoresistance is a major limitation for breast cancer therapy. In the present study, bioinformatics analyses were performed to investigate microRNA miRNA -mediated mechanism of breast cancer chemoresistance and to identify molecular targets. Then, by the target gene software of miRWalk2. The enrichment and function analyses showed that these target genes may participate in many important cancer-related biological processes, molecular functions and signaling pathways.
Among the dysregulated miRNAs, miRa-5p upregulation and miR downregulation were found to have the greatest expression fold change between chemoresistant and chemosensitive tissues.
This is the first report to link these two miRNAs with breast cancer chemotherapy resistance. The most enriched GO terms were significantly associated with regulation processes at the BP level, and transcription activity at the MF level, respectively. These data suggest that ONC has a dose-dependent growth inhibitory effect on metastatic cells and only a moderate growth inhibitory effect on nonmetastatic cells.
Cells were plated in a well plates, cell viability was assayed by MTT and detected using spectrophotometer. To gain a better understanding of the cytotoxic effect of ONC on these cell lines, we used flow cytometry to evaluate cell proliferation stages in response to ONC treatment.
As indicated in Fig. Relative to the vehicle-treated control, LST cells treated with ONC showed a significant, tenfold increase in apoptotic cells and 3.
In contrast, nonmetastatic SW cells were more resistant to the drug treatment, with apoptotic and dead cells increasing by only 2. The studied colorectal cancer cell lines exhibited a differential response to ONC treatment, suggesting a unique mechanism of action that may be related to the metastatic transformation of LST cells.
Gaining an understanding of these mechanisms will provide new insights into the effectiveness of ONC treatment. This critical differentially expressed transcript filter is shown in volcano plots in Fig. In total, we detected 1, and 1, upregulated and downregulated gene transcripts, respectively, in ONCtreated metastatic LST cells relative to the expression in vehicle-treated cells Fig.
In comparison, reduced numbers of differentially regulated transcripts were observed in nonmetastatic SW cells post-ONC treatment, i.
Each dot represents one gene that had detectable expression in either cell line in response to ONC treatment. The software Affymetrix Expression Console Version 1. Initially, these pathways were classified into major network mechanisms including oncogenesis, cell cycle, cellular metabolic pathways, DNA repair, micro-RNAs, and stress; the latter was affiliated only with ONCtreated nonmetastatic SW cells Supplemental Table 2.
In comparison, the overall number of regulated signaling pathways and associated genes in ONCtreated metastatic LST cells was higher than that observed in treated nonmetastatic SW cells. Detailed analysis of the total gene expression profile associated with each signaling pathway revealed remarkable diversity between the metastatic and nonmetastatic cancer cell lines. In drug-treated LST cells, we observed a notable global downregulation of genes associated with oncogenesis, cell cycle, and DNA repair networks.
Whereas, cell homeostasis networks, such as cellular metabolic pathways and micro-RNAs, showed a comparable number of upregulated or downregulated genes Supplemental Table 1. Surprisingly, ONCtreated SW cells showed fewer regulated genes that were almost equivalently upregulated or downregulated, at least in part, for the studied networks. Notably, a large number of stress response network genes were upregulated only in the ONCtreated nonmetastatic SW cells Supplemental Table 2.
In response to ONC treatment, observed differences between the two cell lines implied the existence of differentially regulated mechanisms. Accordingly, we performed a comparative meta-analysis of all the differentially expressed genes and their influence on signaling pathways. We used a computational method that considered the interplay between the gene products in the pathway in response to the drug treatment and scored a predicted functional perturbation for each protein Supplemental Figure S1 ; the data were then further adjusted by Bonferroni corrections.
This approach predicts functional results for the microarray data. For instance; the apoptosis map generated from extrinsic and intrinsic gene expression changes in response to ONC treatment does not explain the moderate apoptotic phenotype in nonmetastatic SW cells compared with the phenotype in metastatic LST cells Supplemental Figure S2 A and S3 A, respectively.
On the other hand, the predicted functional perturbation changes in the apoptotic pathway clearly indicate that the observed phenotype in SW cells is due to a moderate induction of apoptotic genes e. This approach predicts the hidden functional effects of altered upstream regulatory genes Supplemental Figure S1 B and S3 B. This map also shows that, upon ONC treatment, only the intrinsic apoptotic pathway is affected in nonmetastatic SW cells whereas both the extrinsic and intrinsic apoptotic pathway effectors are increased in metastatic LST cells; hence, LST cells show a higher apoptotic fraction see Fig.
Interestingly, when the computational methods were applied to both cell types, nine pathways were significantly perturbed in SW cells posttreatment with ONC, whereas 44 signaling pathways were perturbed in the metastatic cell line LST Fig.
Figure 3 B shows the pathways that are commonly perturbed in both cell lines, while Table 1 details the 35 pathways that are perturbed only in LST. Meta-analysis alignment of differentially expressed common pathways in response to ONC treatment. A Venn diagram illustrating the number of signaling pathways differentially perturbed in response to ONC treatment in nonmetastatic and metastatic cell lines.
The overlapping area indicates the number of signaling pathways commonly altered in both cell types. B Nine common pathways modified in both cell types in response to ONC treatment and the number of altered genes per pathway.
Detailed analysis of the p -values revealed that ONC treatment profoundly influenced genes associated with cell cycle signaling and ER- processing proteins, especially in metastatic LST cells. Similarly, the other seven signaling pathways were also differentially regulated in either cell type in response to the drug treatment; of particular interest, genes associated with metabolic pathways, autophagy, and necroptosis may explain the observed phenotype shown in Fig.
As previously mentioned, 35 signaling pathways were significantly perturbed in the genes differentially regulated in metastatic LST cells but not nonmetastatic SW cells Table 1. Of particular interest, changes to gene expression in the cellular senescence and colorectal cancer signaling pathways were most pronounced with notably low p -values.
In addition, p53 signaling, DNA replication, and other metabolic signaling pathways were significantly decreased but to a lesser extent than the earlier described pathways. We also performed comparative analysis of all data to identify genes associated with the differentially regulated pathways and subsequently ranked these genes in accordance with their p -values Fig.
Data analysis revealed that 2, and 3, genes were differentially regulated in the nonmetastatic SW and metastatic LST cell lines, respectively. Of these, 2, were found to be commonly impaired in both cell types, albeit with varying p -values Fig.
The top 15 genes that were significantly upregulated in either cell line in response to ONC are listed in Fig. Since the cell lines are carcinogenic in nature, such a transcript should have been detected in our microarray data. ASNS asparagine synthetase is involved in asparagine synthesis and facilitates progression through G1 phase of the cell cycle Fig.
Meta-analysis Alignment of the differentially expressed common genes in response to ONC treatment. A A Venn diagram illustrating the total number of genes perturbed in response to ONC treatment in the nonmetastatic and metastatic cell lines. The overlapping areas indicate the numbers of genes that are commonly altered in both cell types.
B Top ranked transcripts differentially regulated in LST and SW cells in response to ONC treatment are shown, with significantly low p -values in green and fold expression in red color. Our initial results indicated that ONC induces apoptosis differently in both cell lines and since CHOP is a critical regulatory factor for this pathway therefore, we focused our study on the regulatory mechanisms associated with this gene.
Mapping the differentially expressed gene transcripts detected in microarray data onto the map of the ER protein processing pathway 30 , 31 revealed similar, but nonidentical, regulatory mechanisms upstream of CHOP for each treated cell line Fig. In addition, transcripts of downstream antiapoptotic BCL2 were significantly downregulated in metastatic LST cells, relative to their expression in nonmetastatic SW cells, post-ONC treatment.
B Pathway analysis of genes differentially expressed in nonmetastatic colorectal cancer SW cells treated with ONC The microarray gene expression data were validated by western blot analysis for selected proteins associated with the upregulation of CHOP.
The upstream CHOP regulatory proteins associated with different ER signaling pathways were also studied and showed similar patterns of expression to those observed in the microarray study. Thus, microarray analysis data and protein expression levels of the studied signaling network markers were found to be alignment after ONC treatments. The software ImageJ V1. Western blot analysis of nonmetastatic SW cells treated with or without ONC, at varying concentrations.
Defects in RNA alternative splicing are a hallmark of cancerous cells. Many RNA splicing regulators have been studied as tumor suppressors or are associated with drug resistance 32 , 33 , To confirm these data experimentally, we performed quantitative RT-PCR and fractionated the products on a bioanalyzer. This is an additional indication of the multifaceted mechanisms of action of ONC as an anticancer drug. Alternative spliced variants of CHOP. CHOP was amplified using specific primers flanking the exons 1 and 4.
PCR products were fractionated on Bioanalyzer microgel. Alternative spliced isoforms were mapped according to their fragment size. Using nonmetastatic SW and metastatic LST colorectal cancer cell lines, we identified signaling pathways that were differently perturbed when cells were treated with ONC Furthermore, the complexity of the CHOP regulatory mechanism in the metastatic cell line and the subsequent downregulation of BCL2 may explain the observed proliferation arrest and high apoptosis rates in ONCtreated LST cells relative to the response in nonmetastatic SW cells.
Our observations are in accordance with previous studies that highlighted the role played by ONC in mediating the ER stress response in breast cancer cells 12 , 39 and high-grade central nervous system glioblastoma Lev et al. Active eIF2a attenuates protein synthesis and reduces protein-processing workload on the stressed ER 41 , 42 , Taken together, these findings suggest that the cellular response to ONC treatment is cell type-dependent, but that the overall mechanisms are associated with ER stress and unfolded protein response signaling.
Gene expression profiles in colorectal cancer cells revealed that ONC downregulates genes associated with energy metabolism. Specifically, ONC reduced the gene expression of citrate carrier SLC25A1 and fumarate hydratase FH that regulate the mitochondrial metabolite carrier and substrate metabolism, respectively. Thus, ONC is involved in reducing metabolic pathways that may cause energy stress in cancer cells.
Similarly, Ishida et al. ONC can also reduce mitochondrial respiration in breast cancer cells, which may lead to energy stress reducing ATP and result in apoptosis In our analysis, we identified several spliced variants of CHOP that were differentially expressed in ONCtreated metastatic and nonmetastatic colorectal cell lines.
Notably, exon 2 was found to be the target for the CHOP splicing mechanism. In summary, the efficacy and outcome of cancer treatment is dependent on the stage of the disease. Differences between nonmetastatic and metastatic cancer cells are associated with cellular plasticity and metabolic reprogramming 52 , which lead to differential responses to chemotherapy as observed here and in other studies. These microarray-based results are confirmed with quantitative, real-time PCR and immunofluorescence.
The functional effect of drug efflux in MPNST-derived cells is confirmed using in vitro growth inhibition assays. Alternative therapeutics supported by the molecular-guided therapy predictions are reported and tested in MPNST-derived cells.
0コメント