Title:
PREDICTIVE RADIOSENSITIVITY NETWORK MODEL
Kind Code:
A1


Abstract:
This invention is a model that simulates the complexity of biological signaling in a cell in response to radiation therapy. Using gene expression profiles and radiation survival assays in an algorithm, a systems model was generated of the radiosensitivity network. The network consists of ten highly interconnected genetic hubs with significant signal redundancy. The model was validated with in vitro tests perturbing network components, correctly predicting radiation sensitivity 2/3 times. The model's clinical relevance was shown by linking clinical radiosensitivity targets to the model network. Clinical applications were confirmed by testing model predictions against clinical response to preoperative radiochemotherapy in patients with rectal or esophageal cancer.



Inventors:
Torres-roca, Javier F. (St. Petersburg, FL, US)
Eschrich, Steven (Lakeland, FL, US)
Application Number:
12/053796
Publication Date:
09/25/2008
Filing Date:
03/24/2008
Assignee:
University of South Florida (Tampa, FL, US)
H. Lee Moffitt Cancer Center and Research Institute, Inc. (Tampa, FL, US)
Primary Class:
Other Classes:
703/2, 702/179
International Classes:
G06F17/18; G06F17/10; G06F19/00
View Patent Images:



Primary Examiner:
SIMS, JASON M
Attorney, Agent or Firm:
FISH & RICHARDSON P.C. (BO) (MINNEAPOLIS, MN, US)
Claims:
What is claimed is:

1. A method of generating a radiation network model for predicting cellular radiation sensitivity comprising the steps of: developing a multivariate linear regression model of radiosensitivity and gene expression comprising: establishing the radiation sensitivity of at least one cell line; establishing the genomic expression of the at least one cell line; selecting expressed genes based on statistical relevance to radiation induction; and identifying at least one gene of interest, expressed by the at least one cell line, predictive of a radiation response using the regression model; identifying radiosensitivity network components using the multivariate linear regression model, comprising: identifying genes reactive to radiation induction using the multivariate linear regression model; and identifying interconnected genes; and incorporating biological interactions of common radiation response elements with the radiosensitivity network components.

2. The method of claim 1, wherein the radiation sensitivity of the cell line is established by the survival fraction of the cell line after exposure to about 2 Gy of radiation.

3. The method of claim 1, wherein genomic expression of the at least one cell line is established from a microarray.

4. The method of claim 3, wherein the microarray data is translated using a blast program to match consensus sequences of gene probesets.

5. The method of claim 1, wherein the radiosensitivity network components are identified by: plotting the interconnection of gene reactive to radiation induction; and selecting genes with at least 5 connections to other genes.

6. The method of claim 1, wherein the common radiation response elements are major hubs of the radiation network model comprising: genes with at least 5 connections to other genes; and no more than 50% of the edges hidden within the network.

7. The method of claim 4, wherein the radiosensitivity network components are selected from the group consisting of: c-jun, HDAC-1, RelA, PKC, SUMO-1, c-Abl, STAT-1, AR, CDK1, and IRF1.

8. The method of claim 1, wherein the selected expressed genes are identified by measuring the correlation between the expression of the gene of interest and the radiation sensitivity of the cell line expressing the gene.

9. The method of claim 1, wherein the biological interactions of common radiation response elements are incorporated by: creating a linear model for each gene in the dataset; analyzing the variability of radiation response in multiple cell lines.

10. The method of claim 1, wherein the common radiation response elements are selected from the group consisting of: tissue origin, ras mutation status, p53 status, tissue origin interaction with gene expression, ras mutation status interaction with gene expression, and p53 status interaction with gene expression.

11. The method of claim 7, wherein the genomic data is normalized by ranking each gene using the radiosensitivity network components.

12. A method of predicting a clinical response to anticancer therapy of a patient in need thereof comprising: obtaining a sample of target cells from the patient; establishing the genomic expression of the sample; and applying the genomic expression of the sample to a multivariate linear regression model of treatment sensitivity whereby a high expression value correlates with a treatment sensitive phenotype thereby predicting the clinical response to treatment.

13. The method of claim 12, wherein the anticancer therapy is selected from the group consisting of radiation therapy, radiochemotherapy, and chemotherapy.

14. The method of claim 12, wherein the sample of target cells comprises cancer cells.

15. The method of claim 12, wherein the multivariate linear regression model is created comprising the steps of: developing a multivariate linear regression model of radiosensitivity and gene expression comprising: establishing the radiation sensitivity of at least one cell line; establishing the genomic expression of the at least one cell line; selecting expressed genes based on statistical relevance to radiation induction; and identifying at least one gene of interest, expressed by the at least one cell line, predictive of a radiation response using a regression model; identifying radiosensitivity network components using the multivariate linear regression model, comprising: identifying genes reactive to radiation induction using the multivariate linear regression model; and identifying interconnected genes; and incorporating biological interactions of common radiation response elements with the radiosensitivity network components.

16. The method of claim 15, wherein genomic expression of the cell line is established from a microarray.

17. The method of claim 15, wherein the radiosensitivity network components are identified by: plotting the interconnection of gene reactive to radiation induction; and selecting genes with at least 5 connections to other genes.

18. The method of claim 17, wherein the radiosensitivity network components are selected from the group consisting of: c-jun, HDAC-1, RelA, PKC, SUMO-1, c-Abl, STAT-1, AR, CDK1, and IRF1.

19. The method of claim 15, wherein the selected expressed genes are identified by measuring the correlation between the expression of the gene of interest and the radiation sensitivity of the cell line expressing the gene.

20. The method of claim 15, wherein the common radiation response elements are selected from the group consisting of: tissue origin, ras mutation status, p53 status, tissue origin interaction with gene expression, ras mutation status interaction with gene expression, and p53 status interaction with gene expression.

Description:

CROSS REFERENCE TO RELATED APPLICATIONS

This application claims priority to currently pending U.S. Provisional Patent Application 60/896,350, entitled, “Radiation Response System Model”, filed Mar. 22, 2007, and pending U.S. Provisional Patent Application 60/896,550, entitled, “Radiation Response System Model”, filed Mar. 23, 2007 the contents of which are herein incorporated by reference.

STATEMENT OF GOVERNMENT INTEREST

This invention was made with Government support under Grant No. K08 CA 108926 awarded by the National Cancer Institute. The Government has certain rights in the invention.

FIELD OF INVENTION

This invention relates to cancer treatment. Specifically this invention is a predictive model of cancer radiosensitization.

BACKGROUND AND SUMMARY OF THE INVENTION

Understanding the biological networks that regulate oncogenic events and influence the inherent radiosensitivity of tumors is central to the development of personalized treatment strategies in radiation oncology, including targeted and improved therapeutic interventions. During the last two decades, many key components and signaling pathways in the oncogenic network have been elucidated by studying radiophenotypic changes after network components are perturbed. However, the dynamics of network component interactions have remained mostly undefined, largely due to lack of accurate testing methods.

The generation of high-throughput datasets in the “omic” era has been central to the development of a systems-view of complex biological systems. In systems biology, the goal is to understand the dynamics of the system and how components interact during operation (H. Kitano, Computational Systems Biology, Nature 420:6912, 2002, 206-210) (H. Kitano, Systems Biology: A Brief Overview, Science 295:5560, 2002, 1662-1664) (L. Hood, J. Heath, Systems Biology and New Technologies Enable Predictive and Preventative Medicine, Science, 306:5696, 2004, 640-643) (L. Hood, R. Perlmutter, The Impact of Systems Approaches on Biological Problems in Drug Discovery, Nat. Biotech., 22:10, 2004, 1215-1217). To study complex biological interactions within a network model, novel methods are needed. A central experimental approach in molecular biology has focused on studying biological systems after components are perturbed by activation/inactivation. A problem of this approach is that it is unable to capture and study the continuous nature of many phenotypic features in diseased and normal states. An alternative approach is a systems-view of biological networks where the focus is on understanding the dynamics and structure of the system of interest. A common feature of systems biology is the development of dry computational models which exploit comprehensive datasets of high-throughput measurements. A common denominator in these models is that biological hypothesis can be generated for testing in “wet” experiments, thus allowing the validation of the models and the dynamics studied. Computational models have been key in the development of central concepts in neurobiology.

SUMMARY OF THE INVENTION

Provided is a mathematical model that facilitates the study of radiation response by providing a systems view of the radiosensitivity network. The model predictions were tested against several biological and clinical endpoints. The systems-based approach improves the ability to define network dynamics and structure, allows the visualization of network topology, and provides a framework to understand its operation, thus leading to a better understanding of the variables that drive radiation sensitivity. Furthermore, as the model accounts for network interactions, the model to captures the variability of radiation response across biological/clinical conditions. This allows the possibility of developing an accurate predictive model of clinical radiosensitivity, a major clinical goal in radiation oncology.

A multivariate linear regression model of gene expression and radiosensitivity (SF2) was developed in a 48 cell line dataset within the context of an accurate radiation sensitivity classifier. A literature review of 35 cell lines in the dataset identified the radiation sensitivity of the cells using SF2 data. Radiation sensitivity for the remaining cell lines was established by establishing the genomic expression of each cell line. After a baseline gene expression dataset was obtained, the cells were irradiated with 2Gy and difference in gene expression determined using microarrays, which allowed for selection of expressed genes based on the gene's statistical correlation between the expression of the gene of interest and the radiation sensitivity of the cell line expressing the gene

The genetic information for the cell line dataset was analyzed using the developed regression model, thereby identifying at least one gene of interest, which is predictive of a radiation response. The regression model identified 500 genes reactive to radiation induction. The 500 genes were analyzed using GeneGo to map interconnections between the genes, and identified a network of interacting genes. This data was further restricted by selecting genes with at least 5 connections to other genes and no more than 50% of the edges hidden within the network.

A series of dynamic cellular states were defined by incorporating biological interactions that has been shown to perturb radiosensitivity. The biological interactions of these common radiation response elements were defined by gene models using the best linear fit model and analyzing the variability of radiation response in multiple cell lines to identify the significant response elements.

The model was then applied to each gene in an expanded genomic/SF2 database of 48 cell lines. This design of an in silico model includes a diverse group of cancer cell lines and favors the identification of genes that are important across cell lines and are more likely central components of the radiation signaling network. The mathematical model also allows the development of biological predictions that can be confirmed by in vitro experiments. This allows feedback into whether the interpretations of the mathematics represented in the model are of biological value. Although the presence of multiple cell lines accounted for component variability, we sought to integrate actual dynamic states of the network. The hypothetical dynamic states were defined by incorporating into the linear regression model biological variables that have been reported to perturb the radiation response network: TO, ras status (mut/wt) and p53 status along with gene expression.

The resulting predictive algorithm identified five components of functional/biological relevance to the network that proved best at building the most accurate predictor, genes rbap48, topi, rgs19, r5pia and an unknown gene. rbap48 and rgs-19 were biologically-validated as network components. siRNA knockdown of rbap48 resulted in radioresistance in HCT-116 cell lines, while overexpression of rgs-19 led to radiosensitization of MDA-MB23 1 cell lines, both observations were consistent with model predictions. In contrast, overexpression of r5pia resulted in no radiophenotypic change of MALME-3M melanoma cells. Finally, top-1 is a clinically validated radiosensitizing target in current clinical use. Thus, we conclude that the linear regression model is reasonably accurate at identifying radiosensitivity network components.

The invention allows for predicting a clinical response to radiation therapy of a patient. Samples of the target cells were collected from the patients. The genomic expression of the collected sample was determined by microarray analysis and the data applied to the network model. High expression values correlated with a radiosensitive phenotype and predicted the clinical response to treatment with radiation therapy. In a specific embodiment, the target cells comprise cancerous cells.

A mathematical model has been developed to represent the topology of the radiation response network. The model identifies novel components of the radiation network as well as integrates dynamics and variability into biological predictions. Both of these abilities have been biologically validated. The model is also envisioned useful in biomarker discovery, allowing biomarkers of response or of radiophenotype to be identified using the model. The model is also useful for clinical trial designs. Network architecture proposed by the model has resulted in identified nodes, which allow for drug designs to specifically target those nodes. This is also useful to guide clinicians in proposing novel combinations of known drugs in clinical trials. Additionally the model may provide an approach to dissect the complexity of network operation. For example, a model detailing the contribution of each hub in the network to final system output can be derived from our database. Finally, a similar model is useful in identifying chemotherapy response using cellular/genomic databases.

BRIEF DESCRIPTION OF THE DRAWINGS

For a fuller understanding of the invention, reference should be made to the following detailed description, taken in connection with the accompanying drawings, in which:

FIG. 1 is an illustration showing a flowchart of the multivariate linear regression model classifier algorithm.

FIG. 2 is a graph showing the radiation response illustration showing the linear regression model output using RbAp48. The output predicts knockout of RbAp48 will result in radioresistance.

FIG. 3 shows the experimental outcome of knocking out RbAp48 in HCT116 cells. (A) Cell survival rates are graphed based on radiation exposure, and (B) a western blot of RbAp48 protein after siRNA transfection.

FIG. 4 is a graph showing the radiation response illustration showing the linear regression model output using RGS-19. The output predicts knockout of RGS-19 will result in radiosensitivity.

FIG. 5 shows the experimental outcome of knocking out RGS-19 in MDA231 cells. (A) Cell survival rates are graphed based on radiation exposure, and (B) a western blot of RGS-19 protein after siRNA transfection.

FIG. 6 is a graph showing the radiation response illustration showing the linear regression model output using Ribose 5 Phosphate Isomerase A (R5PIA). The output predicts overexpression of R5PIA will result in radiosensitivity.

FIG. 7 shows the experimental outcome of knocking out R5PIA in SKMEL28 cells. (A) Cell survival rates are graphed based on total radiation exposure, and (B) a western blot of RGS-19 protein after siRNA transfection.

FIG. 8 is a table of the significant pathways defined by GeneGO MetaCore analysis for the top 500 genes identified by linear fit.

FIG. 9 is a table of selected pathways from the GeneGO MetaCore analysis of significant terms using ANOVA.

FIG. 10 is a table of significant pathways found in Dynamic State 2.

FIG. 11 is a table of significant pathways found in Dynamic State 3.

FIG. 12 is a table of significant pathways found in Dynamic State 4.

FIG. 13 is an illustration showing the network model of the radiation response. The topology of major hubs is shown.

FIG. 14 is a table showing the gene distribution of the data probeset against the dynamic states.

FIG. 15 is a table showing the radiation network hub genes. Genes in gray were used as central hubs for the classifier. The probesets used on each platform are listed for each hub.

FIG. 16 is a table showing the network model predictions for three cancer types.

FIG. 17 is a table of experimental data for three cancer types. The data validates the model predictions, seen in FIG. 16.

FIG. 18 is a graph showing leave-one-out cross-validation results for hub-based classifier on dataset cell lines.

FIG. 19 is a table showing leave-one-out cross-validation predictions on the dataset cell lines using a rank-based linear classifier on the proposed radiation network hubs.

FIG. 20 shows topotecan radiation sensitivity predictions and results for rectal cancer patients. (A) A table of rectal cancer samples shows the radiation sensitivity using survival fractions and clinical response, and (B) a graph of predicted outcomes of rectal cancer radiotherapy, as defined by the network model.

FIG. 21 shows radiation sensitivity predictions and results for esophageal cancer patients. (A) A table of rectal cancer samples shows the radiation sensitivity using survival fractions and clinical response, and (B) a graph of predicted outcomes of esophageal cancer radiotherapy, as defined by the network model.

FIG. 22 is a graph showing a summary of predicted responses for both rectal and esophageal cancer radiosensitivity. As seen, the model noted significant radiation sensitivity response between responders and non-responders.

FIG. 23 is a graph showing the summary of experimental data from rectal and esophageal cancer radiosensitivity.

FIG. 24 is a table showing the manner in which radiosensitization targets are linked to the radiation sensitivity network.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT

A multivariate linear regression model of gene expression and radiosensitivity (SF2) was developed in a 35 cell line database within the context of an accurate radiation sensitivity classifier. The clinical value of a radiosensitivity predictive model is significant, therefore an understanding the intricacies of its operation were critical. The predictive algorithm identified five components of functional/biological relevance to the network that proved best at building the most accurate predictor, genes rbap48, top1, rgs19, r5pia and an unknown gene. FIG. 1 shows a schematic representation of the classifier algorithm. As shown in FIGS. 2-5, rbap48 and rgs-19 were biologically-validated as network components. Consistent with model predictions, depicted in FIG. 2, siRNA knockdown of rbap48 in HCT-116 cells, seen in FIG. 3(b), resulted in radioresistance as seen in FIG. 3(a). Next, G-protein signaling regulator rgs-19 was overexpressed in MDA-MB231 cell lines, seen in FIG. 5(b). Overexpression led to enhanced radiosensitization of the cell lines compared to EV-transfected controls, as seen in FIG. 5(a). This was consistent with model predictions, seen in FIG. 4. In contrast, overexpression of r5pia in MALME-3M melanoma cell lines, seen in FIG. 7(b), resulted in no radiophenotypic change, as seen in FIG. 7(a). This result was contrary to the predicted response, seen in FIG. 6. Finally, top-1 is a clinically validated radiosensitizing target in current clinical use, and confirmed that the linear regression model is reasonably accurate at identifying radiosensitivity network components.

The model was more fully developed to map the radiation sensitivity network by incorporating biological interactions with the genomic/SF2 database. A linear model was created for each gene in the cell line dataset. Common elements of radiation response were analyzed for variability introduced by multiple cell lines in the classifier to explicitly model the dynamic states. The dynamic states were models incorporating biological variables that have been reported to perturb the radiosensitivity network: tissue of origin (TO), ras status (mut/wt) and p53 status along with gene expression. The resulting model:


SF2=k0+k1(yx)+k2(TO)+k3(ras status)+k4(p53 status)+k5(yx)(TO)+k6(yx)(ras status)+k7(TO)(ras status)+k8(yx)(p53 status)+k9(TO)(p53)+k10(ras status)(p53 status)+k11(yx)(TO)(ras status)+k12(yx)(ras status)(p53 status)+k13(TO)(ras status)(p53status)+k14(yx)(TO)(ras status)(p53 status)

represents 14 different potential dynamic states based on the four chosen variables and interactions between those variables. Original cell line data was created on HU6800 GeneChips while the newer patient data was created on HG-U133Plus Chips. The probesets were translated using a blast program to identify the best U133Plus probeset match to the consensus sequence from which the 6800 probeset was designed using Affymetrix software. The 500 genes identified with the smallest sum of squared residuals for the developed linear model were further analyzed using Analysis of Variance (ANOVA) to determine the significant terms (e.g. gene, ras status) and correlation to SF2 across cell lines. The model produced four significant dynamic states in the radiosensitivity network, reduced from the 14 hypothetical states. TO and ras status and their interaction with gene expression proved to be key variables in defining the four states. Interestingly, the prostate cancer TO term grouped separately. In contrast p53 was not found to be a significant factor in the analysis. Cell lines grouped in the three states mainly distinguished by the presence of a mutated ras. The ras wt population was divided in two groups (NSCLC and Ovarian vs. Other TO). The ras term was dominant, therefore cell lines with mutated ras grouped closer than cell lines from the same TO, as exemplified by breast cancer cell lines (HS578T, MDAMB231) grouped together with other cell lines that shared this biological feature.

To explore the functional difference in the dynamic states, a pathway analysis was performed. GeneGO MetaCore identified a series of significant pathways shared by the 500 genes, depicted in FIG. 8. As seen in FIG. 9, key biological differences exist across dynamic states in the network. For example, dynamic state 2 represented pathways in metabolism, hypoxia and Akt, seen in FIG. 10. Dynamic state 3 represented 29 pathways, 11 of which were cell cycle related, seen in FIG. 11. Finally, dynamic state 4 was the most functionally diverse, representing pathways in DNA repair, cell cycle regulation, adhesion, apoptosis, immune response and protein kinase cascades, seen in FIG. 12. While many of these pathways have been implicated in the regulation of radiation response, the model evidences the importance of each pathway's dependence on the biological context that defines the dynamics of the network.

To visualize the network proposed by the mathematical model, the primary interconnections of the original 500 genes selected using literature-based annotations were plotted using GeneGO, seen in FIG. 13. The gene probesets were loaded into GeneGO MetaCore and analyzed for overexpression in various cellular pathways defined by the dynamic states, seen in FIG. 14. Hubs were defined within the gene network as a node consisting of at least 5 connections to other genes, seen in FIG. 15. All hubs with more than 5 connections and less than 50% of edges hidden within the network were chosen as the major hubs for classification purposes. This network model, shown in FIG. 13, proposes ten central hubs: c-jun, HDAC-1, RelA (p65 subunit of NFKB), PKC, SUMO-1, c-Abl, STAT-1, AR, CDK1 and IRF1, seen in FIG. 15. Remarkably, each of these hubs is reportedly involved in radiation signaling and 6/10 (HDAC1, NF-KB, c-ABL, STAT1, AR, SUMO-1) have been proposed as targets for radiosensitizer development. Additionally, the model proposes significant cross-talk among the central hubs, consistent with a robust system with significant signal redundancy. It should be noted that these hubs would not be identified using the correlation of gene expression to SF2 as the median R of these hubs is 0.02.

Because the hub classifier was applied to datasets generated from differing GeneChip platform and technology, genes were normalized using a rank-based approach. Gene expression was ranked for each gene per sample using the identified ten hubs.

The model was analyzed by testing the effect of c-jun knockdown on radiosensitivity, thereby determining whether biologically-relevant network dynamics and interactions were being captured. Selection of c-jun was due to the fact that c-jun is a central hub and an AP-1 regulated pathway was the only commonality between the three main dynamic states. Importantly, c-jun has been shown to play a role as an early response gene in the initial stages of radiation response. The model predicted c-jun knockdown would cause differing results, based on the biological context as defined by TO. The predictions and experimental outcomes, using a linear fit for c-jun gene expression to SF2 stratified by TO, are shown in FIG. 16. As seen in FIG. 17, c-jun siRNA was transfected into 8 different cell lines, representing the three tissues types selected: NSCLC, colon and breast. Downregulation of c-jun resulted in induction of radiation resistance in NSCLC cell lines, consistent with the linear regression curve derived from the model. Cell survival curves in both A549 and H460 cell lines confirmed these observations. Furthermore, the radiophenotype of colon cancer cell lines (when considered as a group) was unaffected by c-jun downregulation, also supporting the model. However, it should be noted that in HCT-116 cells, c-jun downregulation led to radioresistance (p=0.52). Additionally, radiation response in breast cancer cell lines was unchanged by c-jun siRNA transfection, while the linear model predicted radiosensitization. The model was experimentally validated in two of three tested instances, for lung cancer (radioresistance, p=0.005) and colon cancer (no change).

The experiments supported the model's ability to capture the influence of biological context on network outcome. However, because radiosensitivity prediction is linked to biological contexts, predicitive features changed depending on expression context. A hub-based gene expression classifier was built to estimate the predictive accuracy of the network model. A linear regression model was developed along with support-vector machines for comparison, however, the linear regression model found the most accurate at 30/48 (62.5%), seen in FIG. 18. Further, the rank-based dataset normalization yielded a more accurate classifier than using actual gene expression values, shown in FIG. 19.

To determine the clinical relevance of the model, it was used to predict clinical response in 14 patients with locally-advanced rectal cancer treated with preoperative concurrent radiochemotherapy. Pre-treatment samples from the patients were arrayed on the HG-U133Plus platform. The tumors were staged at initial biopsy with ultrasound and later stages using pathological information from surgical resection. Downstaging in the T stage from the TNM staging system translated R (response) in the dataset, while no change or progressive disease was recorded as NR (no response). Data was processed using gcRMA using the Bioconductor implementation. Gene expression values for the 10 hubs were converted to ranks and SF2 values were generated from the model created using cell line data, depicted in FIG. 20 (mean predicted SF2 R vs. NR 0.31 vs 0.45, p=0.03). Responders were further tested for significantly lower predicted SF2 using a one sided Wilcoxon rank-sum test (P=0.02964). The 10 gene model was further tested in a cohort of 12 patients with esophageal cancer also treated with preoperative radiochemotherapy. A pre-treatment biopsy was collected from the patients and tissue arrayed on the HG-U133Plus platform. The entire dataset was processed (22 patient samples), though only 12 esophageal cancer samples with chemoradiation response were available. Chips were normalized using RMA in the GENE implementation (Eschrich, 2007). Gene expression values for the ten hubs were converted to rank values and the SF2 values were generated from the model created using cell line data. Similar to the rectal cancer cohort, responders were predicted to be more radiosensitive than non-responders as determined by predicted SF2, seen in FIG. 21 (0.34 vs. 0.48, p=0.05). For both patient cohorts, rectal and esophageal cancer, the model predictions significantly separated pathological responders (R) from non-responders (NR), seen in FIGS. 22 and 23. A test of significantly lower predicted SF2 values in the CR group was performed using a one-sided Wilcoxon rank-sum test (P=0.05303). These results are encouraging since no esophageal cancer cell lines were included in the original database, suggesting that the model is capturing central common aspects of the radiosensitivity network that are of clinical relevance.

The model was further analyzed against ten known radiosensitizer drug targets, both in clinical development or routine clinical use. All drug targets are linked by primary interconnection to at least one central hub of the model, seen in FIG. 24, supporting the clinical relevance of the radiosensitity network model. Moreover, the model revealed that the targets interference with only a minority of the hubs, suggesting the current clinical approach to radiosensitization is inefficient at disrupting the radiosensitivity network.

A fundamental objective of the field of systems biology is to develop an understanding of the dynamics and structure of complex biological systems. The presented model integrates both of these elements and represents an important advance in the understanding of the radiation response regulatory network.

The mathematical model proposes a highly interconnected network topology with ten central hubs and significant signal redundancy. The redundancy explains why targeting a single hub could lead to different or inconsistent system outputs (i.e c-jun knockout), as phenotypic responses may be driven by competing signal networks. The complex combination of signals is consistent with the continuous nature of radiation response, providing a framework to explain individual response variability. The hubs identified by the model have been shown important in the regulation of radioresponse. All targets connected via at least one of hub, supporting the biological validity of the model. In contrast, 20 alternative networks were developed using chance for feature selection. The mathematical model outperformed all alternative chance networks in all instances, when target connectivity and hub's relevance in radioresponse were used as benchmarks for comparison.

An advantage of the mathematical model is that it considers the inherent individual variability that exists in the response to therapeutic agents. Furthermore, biological variables that may define specific resistance/sensitivity phenotypes can be included, allowing the model to capture several signaling states in the network. This last concept has been proposed to explain the lack of commonality between validated disease-specific molecular signatures in clinical oncology. The model can identify novel network components and integrate complex interactions and dynamics into biological predictions. Finally, it provides a network architecture that allows hypothesis development, extending from basic radiation molecular biology to hypothesis with a direct impact in clinical radiation oncology.

Material and Methods

Cell lines—Cell lines were obtained directly from the National Cancer Institute (NCI). Cells were cultured in RPMI 1640 media supplemented with glutamine (2 mM), antibiotics (penicillin/streptomycin, 10 U/ml) and Heat inactivated Fetal Bovine Serum (10%) at 37° C. with an atmosphere of 5% CO2.

Radiation Survival Assays (SF2)—The SF2 of cell lines used in the classifier were obtained from the literature in 23 of the 48 cell lines in our analysis. For cell lines obtained from the literature, papers (published before 2004) were used that reported on clonogenic assays that had been performed without the use of any substrate (i.e. agar) and that required cells to be in log phase at the time of irradiation. The cell lines also needed at least two reported values in the literature by different laboratories. Mean SF2 values were determined for each cell line and used for the generation of the model. The remaining 25 cell lines (MCF-7, MDA-MB-435, KM-12, HOP62, H23, BT549, MDA-MB-231, HCT116, HT29, H460, OVCAR5 and PC3) SF2 values were determined in the lab. Clonogenic survival assays after 2 Gy of radiation were performed as previously described (J. Staunton, D. Slonim, Chemosensitivity Prediction by Transcriptional Profiling, Proc. Nat. Acad. Sci., 98:19, 10787-10792). Plating efficiency for each cell line was determined, prior to SF2 determination. Cells were plated so that 50-100 colonies would form per plate and incubated overnight at 37° C. overnight to allow for adherence. Cells were then radiated with 2 Gy using a Cesium Irradiator (J. L Sheperd, Model I 68A, San Fernando, Calif.). Exposure time was adjusted for decay every three months. After irradiation cells were incubated for 10-14 days at 37° C. before being stained with crystal violet. Only colonies with at least 50 cells were counted. SF2 was determined by the following formula:


SF2=number of colonies/total number of cells plated×plating efficiency.

Microarrays—Gene expression profiles were from Affymetrix HU6800 chips (7,129 genes) or from a previously published study (J. Torres-Roca, S. Eschrich, et al., Prediction of Radiation Sensitivity Using a Gene Expression Classifier, Cancer Res., 65:16, 7169-7176). The gene expression data had been previously preprocessed using the Affymetrix MAS 4.0 algorithm in average difference units. Negative expression values were set to zero and the chips were normalized to the same mean intensity.

siRNA transfection. 3×105 Hs-RbAp48-hi cells in 2 mL antibiotic-free complete medium were plated in each well of a six-well plate and after 24 h of incubation were transfected following the basic dharmaFECT transfection protocol (Dharmacon, Inc., Lafayette, Colo.) with either a pool of 4 negative control siRNAs (siRNA pool) or RbAp48 siRNA designed by Dharmacon's SMARTpool technology both at 100 nM final concentration. 48 hours after transfection, cells were lysed for Western blotting, to confirm the knockdown of RbAp48, or plated in coverslips for immunofluorescence.

The disclosure of all publications cited above are expressly incorporated herein by reference, each in its entirety, to the same extent as if each were incorporated by reference individually.

It will be seen that the advantages set forth above, and those made apparent from the foregoing description, are efficiently attained and since certain changes may be made in the above construction without departing from the scope of the invention, it is intended that all matters contained in the foregoing description or shown in the accompanying drawings shall be interpreted as illustrative and not in a limiting sense.

It is also to be understood that the following claims are intended to cover all of the generic and specific features of the invention herein described, and all statements of the scope of the invention which, as a matter of language, might be said to fall therebetween. Now that the invention has been described,