National Surgical Adjuvant Breast and Bowel Project (NSABP) Foundation Annual Progress Report: 2009 Formula Grant Reporting Period July 1, 2011 June 30, 2012 Formula Grant Overview The National Surgical Adjuvant Breast and Bowel Project (NSABP) Foundation, Inc. received $1,255,581 in formula funds for the grant award period January 1, 2010 through December 31, 2012. Accomplishments for the reporting period are described below. Research Project 1: Project Title and Purpose Discovery and Validation of MicroRNAs as Biomarkers in Breast and Colon Cancer - Messenger RNA (mrna) expression profiling has yielded useful prognostic tests for breast cancer (e.g., Oncotype DX and MammaPrint ), but has not been very useful in identifying definitive treatment targets. This is because mrna expression profiling does not necessarily provide insight about which changes may be susceptible to therapies. For colon cancer, no clinically useful gene-expression-based prognostic tests exist yet. MicroRNAs (mirnas) are a subset of RNAs which have regulatory functions and, so, could be viewed as master switches for many coordinately regulated genes. This project s purpose is to identify mirnas that are prognostic or predictive of benefit from systemic therapies for breast and colon cancers, using archived tumor tissues from completed NSABP clinical trials. Identified markers may point us to new treatment strategies. Anticipated Duration of Project 1/1/2010-12/31/2013 Project Overview The long-term goal of this research project is to develop biomarkers that will improve the treatment of colon and breast cancers. To achieve this goal, we propose to examine the expression of micrornas (mirnas), which are small, non-coding RNAs that control translation of many mrnas. mirna expression analysis may not only allow for the identification of better biomarkers but may provide a greater understanding of current expression data. One of the fundamental problems associated with the treatment of breast and colon cancers is clinical heterogeneity. Gene-expression mrna signatures that distinguish several subtypes of breast cancer have been identified, and, while no specific signatures currently exist for colon NSABP Foundation - 2009 Formula Grant Page 1
cancer, the molecular and clinical heterogeneity of colon tumors is known. The subtyping of breast cancers has vastly improved breast cancer treatment because different subtypes respond differently to different breast cancer therapies. Current molecular signatures that distinguish these subtypes of breast cancer involve the profiling of a large number of genes and have primarily been identified by whole genome mrna expression analysis. Whole genome expression analyses are too costly and too complicated, and have too much background interference to be used as routine clinical tests. mirnas may be able to give prognostic and predictive information because they act as major switches that regulate the expression of many mrnas. It may be possible to better characterize a tumor with a small number of mirnas than is possible with mrnas. Colon cancer lacks biomarkers for identifying tumors that have a high risk for recurring, and predicting response to chemotherapy is not possible. Given these considerations, we have adopted the following specific aims: 1) Discover and validate prognostic mirna markers for breast cancer patients with resectable tumors and histologically positive axillary lymph node(s); 2) Discover and validate predictive and prognostic mirna markers for colon cancer patients with stage II and III resectable colon cancer; and 3) Initiate the development of an integrated colon cancer molecular database using mirna data as well as mrna and mutation data from the NSABP clinical trial C-07 samples which will provide unique information for hypothesis generation. Principal Investigator Katherine L. Pogue-Geile, PhD Assistant Director of Molecular Profiling NSABP Foundation, Inc. Four Allegheny Center, 5 th Floor Pittsburgh, PA 15212 Other Participating Researchers Soonmyung Paik, MD; Patrick Gavin, BS; Chung Kim, MD employed by NSABP Foundation Expected Research Outcomes and Benefits In this project, we will examine the expression of mirnas in breast and colon cancers. The project has the following expected outcomes: 1) To identify the gene expression profile of mirnas in a large number of colon and breast cancer samples; 2) To determine the prognostic significance of mirnas in colon and breast cancers; 3) To determine the predictive significance of mirnas in colon and breast tumor responses to specific anti-cancer therapies; and 4) To acquire new information for the development of a valuable integrated molecular database of colon tumors. The C-07 colon tumors that will be interrogated for mirna NSABP Foundation - 2009 Formula Grant Page 2
expression have already been profiled for mrna gene expression by Illumina DASL, arrays and profiled for critical cancer mutations in KRAS, NRAS, PIK3CA, BRAF and MET. This information along with standard histological and clinical information makes this a valuable resource for the discovery of the mechanisms that determine the success or failure of specific cancer treatments and to generate hypotheses regarding control of gene expression and its role in cancer development. Potential benefits include the following: 1) Improved prognostication for patients with diagnoses of breast or colon cancer; 2) Improved prediction of response of patients with colon cancer to oxaliplatin; and 3) Generation of information for the development of integrated molecular databases. C-07 mirna data will add to other molecular information on the same clinical samples, including data on whole genome mrna expression and somatic cancer mutations. This database can be used for hypothesis-generation as it will provide a large database for correlating these molecular characteristics. Summary of Research Completed The Specific Aims addressed in this report period were: 1. Completed mirna expression profiles of 914 cases included in the validation cohort. 2. Identified prognostic and predictive mirnas (mirs) in the discovery cohort of C-08 and began building prognostic models. 3. Identified bevacizumab-predictive mirs and began building predictive models. To achieve the overall goal of this proposal, which was to improve the prognostication and the treatment of colon and breast cancer, we have chosen to focus our efforts on colon cancer both for pragmatic and scientific reasons. Expression profiling of all well-annotated mirs in a large cohort of cases and using the best technology required more budget expenditure than we had originally planned which then necessitated that we choose only one cancer. To maximize the possibility of developing a clinically meaningful signature, we decided to focus on colon cancer because much less molecular data have been collected for colon cancer compared to that for breast cancer. Breast cancer already has prognostic signatures that are used clinically to identify patients who will receive benefit from chemotherapy. We also decided to focus on clinical trial NSABP C-08 because this trial included a control arm of patients who were treated with current standard chemotherapy (mfolfox), and an investigational arm that tested whether the addition of the targeted drug bevacizumab to standard chemotherapy improved the disease-free survival of patients with Stages II or III colon cancer. Even though bevacizumab did not show benefit when added to mfolfox when the entire cohort of C-08 patients was evaluated, it is still possible that a subset of patients did receive benefit from bevacizumab treatment and that these patients could be identified through a molecular test. One of the goals of this project is to develop a mirrna expression signature that is able to identify such a subset. Furthermore, the control arm of C-08 may allow us to develop a prognostic signature that would identify patients who would need more than standard chemotherapy. I. Mir Profiling of the Validation Cohort We have completed the mir expression profiles from 914 cases included in the C-08 validation cohort. This validation cohort consists of cases that were not included in the C-08 discovery NSABP Foundation - 2009 Formula Grant Page 3
cohort. Mir expression profiles were obtained using the same technology (the TaqMan Array Human MicroRNA Cards) that was used for the expression profiling of the discovery cohort in order to eliminate cross-platform differences between the discovery and validation data sets. Currently, we are performing a quality control (QC) assessment of the validation cohort and may repeat profiling of a few samples that failed our QC. We are still evaluating methods for normalization to compensate for the batch effects, which distinguish the discovery and validation cohorts. Thus, the wet lab portion of the profiling of the validation cohort is nearly completed, but additional bioinformatics work must be done before the validation cohort is complete. II. Statistical Methods of the Mir Expression Data from the C-08 Discovery Cohort We have begun to analyze the mir expression data from the C-08 discovery cohort, which consists of approximately 1000 randomly selected C-08 cases profiled with 754 mirs. C-08 samples were chosen randomly within strata defined by treatment and N-Stage (N0, N1, N2). The proportion of patients in each strata of the sample is the same as the proportion of patients in each strata of the eligible population. A. Quality Control and Normalization Mirs were interrogated with TaqMan Array Human MicroRNA Cards (from Applied Biosystems), which includes 2 cards (A and B). Each card interrogates 381 different mirs. Different QC metrics were required for the two cards because the manufacturer intentionally biased the A card to include more prevalent, better annotated mirs. Quality control of A cards was performed by removing cards with fewer than 150 mirs detected, or more precisely, any A card in which there were fewer than 150 mirs with a Ct value below 37 was removed from the data set. Quality control of B cards was performed by removing cases with an average Ct for U6 snrna of more than 20. The two metrics are highly correlated. Cards failing these criteria were repeated for cases with available RNA. Adequate results were obtained for 966 of 1000 discovery cases for the A card and 953 of 998, for the B card. We selected mirs for normalization by analyzing our data with NormFinder and genorm algorithm. The geometric mean of the 5 normalizers was used to normalize the data. The mirs used for the A card were hsa_mir_140_3p_4395345, hsa_mir_152_4395170, hsa_mir_103_4373158, hsa_mir_99a_4373008, hsa_mir_93_4373302, and for the B card were U6, hsa_mir_26b 002444, hsa_mir_18a 002423, hsa_mir_214 002293, and hsa_mir_93 002139. Missing values are indicated in the raw data as not defined if their Ct value was 37 cycles or above. B. Identification of Prognostic and Predictive Mirs The prognostic/predictive effect of each mir marker was determined in Cox regression models, which were adjusted for age, sex, nodal status, and t stage. Normalized Ct values were categorized with quartiles to test for possible non-linear effects. Mirs were screened based on the minimum p-values among the three calculated p-values for each subgroup based on the quartile of Ct values for each mir. We also applied a 10-fold jack-knifing technique to determine whether the association p-values were robust. The associations of the p-values for each mir calculated from 10 subsets of the data were averaged. The clinical endpoint for these analyses was time to recurrence (TTR). Forty predictive mirs with minimum p-values below 0.05 were NSABP Foundation - 2009 Formula Grant Page 4
identified. Table 1 shows the p-values for the top 10 mirs, but the mir names have been changed to anonomous names a-k for proprietary reasons. C. Building and Evaluating Prognostic and Predictive Models We are currently exploring different methods for the building of prognostic and predictive models and are evaluating these models in an effort to pick the best model for validation. Below is a summary of our current efforts to build prognostic and predictive models; however, we have not revealed the identity of specific mirs because these models may be patented. This process is ongoing and we are still building and evaluating these models. Predictive Model 1. The p-values for the top 10 mirs, which were identified as predictive by a categorical analysis, are shown in Table 1, but the mir names have been changed for proprietary reasons. These mirs also were examined for their distinct predictive effect using visual inspection of Kaplan-Meier plots and STEPP analysis. The top 3 mirs (mir-a, mir-b and mir-c) based on the p-value also had a distinct predictive effect. Continuous (linear) interaction p-values for mirs were also calculated using TTR as the clinical endpoint and using a 10-fold jack-knife process. The top four mirs based on the predictive continuous analysis were mir_d (p=0.0162), mir_c (p=0.0269), mir_e (p=0.0434) and mir_f (p=0.0451). The Kaplan-Meier plots were generated for each of these mirs and only the top 2 mirs (mir_d and mir_c) rendered a relatively distinct predictive effect. The top mirs based on the p-values determined by continuous analysis (mir-c and mir-d) and categorical analysis (mir-a, -b, -c) were included in a predictive model. These 4 mirs (mir-c was identified by both continuous and categorical analyses) were used to build a model applying Wald's method. Other classification methods, including average linkage and centroid methods were also applied. All of these methods generated 5 clusters with a differential response to bevacizumab. On the basis of this result, the patients were re-categorized into the following three groups: 1+2, 3, 4+5 (Fig. 1). Only group 1 patients appear to receive benefit from bevacizumab; the adjusted hazard ratio for group 1 was 0.61 (p=0.038). Patients in group 2 and group 3 do not appear to receive benefit, and those in group 2 may have received harm. The interaction p-value for this model was significant (p= 0.018). Even if the number of clusters was reduced, none of the mirs were removed for discrimination for the 3 groups when the stepwise discriminant analysis was performed. When linear discriminant analysis was applied with cross validation, the linear discriminant function with the 4 mirs classified patients into 3 groups with a 10% misclassification rate. Predicitive Model 2 was generated using mirs that have been described to play a role in the immune system. Patients were clustered into 5 groups. Results of this model are shown in Fig. 2. Global p-value for these 5 clusters was 0.011, and 0.0061 if cluster 1 and 2, and cluster 3 and 4 were combined. This model based on this subgrouping was more significant than Predictive Model 1. D. Prognostic Models A prognostic model was built based on the Super Principal Component method and using only the patients who were in the C-08 control group. Mirs included in this analysis had a missing NSABP Foundation - 2009 Formula Grant Page 5
rate of less than 5%. First, ten-fold cross-validation was used to determine the optimum number of genes and Principal Components (PCs) to be included in the models. The median p-values for prognostic effect of the prognostic index (PI) were used in the COX model. The PI is the estimated exponentiated survival time estimated by the model, built in the training set, which is divided into 10 different training sets. The clinical endpoint is TTR. The optimum number of genes and PCs were determined to be 4 and 3, respectively, but only one PC was included in the model because only the first PC contributed to the model (Fig. 3). The top 10 mirs selected in each of the each of the 10 cycles during 10-fold cross validation are shown in Table 2. Mirs included in all 10 cycles have a cv (cross validation) rate of 1. To select the best model, we built 4 models using all samples and the following clinical covariates: nodal status, gender, and age. The models were 1) a model with only the covariates; 2) a model with 1 PC and the top 4 mirs based on their p-values; 3) a model with the covariates and the top mir based on p-value (hsa_mir-l) using forward selection; and 4) a model with covariates and PC1 and the top 4 mirs based on the cv rate (Table 2). These models were evaluated by several statistical measures including a score for deviance. Based on the statistical measures, the best model was model 4. When this model was applied to all samples, the prognostic p-value was 0.0004, and, 0.018 in the control samples when adjusted for treatment effect. III. Summary In conclusion, we have nearly completed the wet lab portion of the mir expression profiling of the C-08 validation cohort in this year. Using the separate, non-overlapping discovery cohort, we have shown that it is possible to build prognostic and bevacizumab-predictive models using mir expression data. We are currently evaluating these models in the discovery cohort. NSABP Foundation - 2009 Formula Grant Page 6
Table 1. Predictive mirs mir minpred_catp hsa_mir_a 0.0021 hsa_mir_b 0.0024 hsa_mir_c 0.0031 hsa_mir_d 0.0035 hsa_mir_e 0.0051 hsa_mir_f 0.0056 hsa_mir_g 0.0056 hsa_mir_h 0.0057 hsa_mir_i 0.0077 hsa_mir_j 0.0083 hsa_mir_k 0.0111 Table 2 Prognostic mirs for Model Buiding mir Pr > ChiSq cvrate hsa_mir-l 0.0009 1 hsa_mir_m 0.004 1 hsa_mir_n 0.0118 1 hsa_mir_o 0.0128 0.8 hsa_mir_p 0.0141 0.8 hsa_mir_q 0.0144 0.8 hsa_mir_r 0.0171 0.8 hsa_mir_s 0.0197 1 hsa_mir_t 0.0223 0.9 hsa_mir_u 0.0241 0.9 NSABP Foundation - 2009 Formula Grant Page 7
Predictive Model 1. Kaplan-Meier Plots of Response (TTR) to Bevacizumab by Cluster Groups NSABP Foundation - 2009 Formula Grant Page 8
Predictive Model 2. Kaplan-Meier Plots of Response (TTR) to Bevacizumab by Cluster Groups NSABP Foundation - 2009 Formula Grant Page 9
Model Building for Prognostic Model. Determination of Optimum Number of Principal Components and Genes NSABP Foundation - 2009 Formula Grant Page 10