Gene Expression Profiling after Radiation-Induced DNA Damage Is Strongly Predictive of BRCA1 Mutation Carrier Status

It is estimated that 5–10% of breast cancer patients develop the disease because of the presence of a mutation in a breast cancer predisposition gene (1). A significant proportion of this population (just under a half) has a mutation in one of the known breast cancer predisposition genes, BRCA1 or BRCA2. Besides the obviously disease-causing deleterious mutations, very often small alterations caused by a single base change (missense mutations) are found in these genes. Their functional effects are usually unknown and so they are termed variants of uncertain significance. Some of these variants of uncertain significance could also have a role in breast cancer predisposition, but it is not currently possible to establish their disease-causing effect. The available diagnostic tests for mutation analysis of BRCA1/2 are time and labor intensive, expensive, and none of them allow the identification of all types of mutation. Our aim in this preliminary study was to investigate if gene expression profiling could be used to distinguish between heterozygous BRCA1 mutation carriers and control samples from reduction mammoplasties with a very low chance of the presence of a BRCA1 mutation. If this were possible, then this would have the potential to develop a method to identify the presence of a BRCA1 defect.

The BRCA1 gene encodes a large nuclear protein (220 kDa) that has multiple possible functions, including DNA damage signaling, DNA repair, growth inhibition, and transcription regulation (2, 3). The involvement of BRCA1 in transcriptional regulation has been revealed in several studies showing direct interaction with other transcriptional activators and repressors such as STAT1, MYC, TP53, ZBRK1, and CDKN1B (4) It appears that BRCA1 is phosphorylated as a response to various DNA damaging agents by kinases such as CHEK2, ATM, and ATR, and this phosphorylation results in changes in its protein-protein interactions, which then can lead to regulatory changes in the expression of various target genes (5).

Microarray studies have been shown to be of great value in understanding the molecular biology of many diseases, and they have been successfully used to classify various tumors based on their clinical phenotype or genetic background (reviewed in Ref. 6). This approach has enabled classification of tumors and division into prognostic groups on the basis of their global patterns of gene expression. One of the first of these studies classified myeloid and lymphoblastic leukemias using a class discovery procedure (7) demonstrating the feasibility of this technique. Similarly, it has been shown that breast epithelial cell lines have an expression profile that is distinct from breast tumors and that breast tumor samples can be divided into different clinical outcome groups based on clustering algorithms using cDNA microarray data (8). Tumor classification using gene expression profiling has also been successfully used to distinguish breast cancer subclasses with important clinical implications (9) and to discriminate breast tumors on the basis of estrogen receptor status and lymph node status (10). BRCA1 and BRCA2 mutation status have also been shown to influence the gene expression profile of breast tumors (11). Various studies provide evidence that microarray analysis can be used to identify gene expression changes induced by drug treatments and radiation (12). Our aim in the present study was to assess the potential of gene expression profiling in discriminating between normal cells with BRCA1 mutations and controls after induced DNA damage.

Individuals who are heterozygous for BRCA1 germ-line mutations were identified from the BRCA1 and BRCA2 predictive testing program in the Royal Marsden Hospital NHS Trust/The Institute of Cancer Research, Cancer Genetics Carrier Clinic. Prophylactic mastectomy specimens were collected from nine individuals with a germ-line BRCA1 mutation, and short-term breast fibroblast cell lines were established. The mutation status of these samples is listed in Table 1. For our five control samples, we used similar cell lines established from reduction mammoplasty specimens. The healthy women undergoing reduction mammoplastic surgery had no family or personal history of any cancer. From the tissue specimens, short-term primary fibroblast cultures were established by standard methods. Confluent cells were irradiated with 15 Gy at a high dose rate (1 Gy/min) using a ⁶⁰Co source. Cells were maintained at 37 °C during irradiation. Total RNA was extracted from the cells 1 h after the treatment using an RNeasy kit (Qiagen, Valencia, CA). Reference RNA was pooled from three cancer cell lines (MDA-MB-231, A549, and DLD-1). RNA samples were reverse transcribed into cDNA fluorescently labeled with Cy3 and Cy5. Equal amounts (20 μg) of sample and reference cDNA were mixed and hybridized onto the microarrays. We used high density cDNA microarrays manufactured by the Section of Molecular Carcinogenesis, The Institute of Cancer Research, representing 5603 IMAGE cDNA clones. Details of the clone set and hybridization conditions are available online.⁹ Image acquisition and analysis were performed using GenePix 4000B scanner and GenePix Pro 4 (Axon Instruments, Inc., Union City, CA), respectively. Signal intensities for Cy3 and Cy5 channels were normalized based on the assumption that on average the genes have the same overall expression level in the experimental and the reference sample.

Of the 5603 cDNA clones represented on the arrays, 1967 clones were filtered out and used in the consequent analysis by the criteria that these data points showed signal intensity in at least one of the channels (red or green) 2-fold above background in a minimum of 12 of our 14 samples.

We analyzed our data first using supervised methods of class comparison and class prediction (13). For class comparison, we have used a recently described statistical method, significance analysis of microarrays (SAM; Ref. 14), to identify genes in which the levels of expression significantly correlate with the presence of a BRCA1 mutation. SAM uses t test statistics to compare the expression level of each gene in cells from BRCA1 mutation carriers and control individuals. Our aim was to find if there is a statistically significant difference between carriers and noncarriers in the expression level of a number of genes and to identify the genes that are most likely to show consistent differences. SAM, using repeated permutation, computes a score for each gene that measures the strength of its correlation with the presence of a mutation. A threshold value can be chosen, which will determine the false positive rate (false discovery rate) as estimated by repeated permutation and counting the number of genes selected as significant.

In an independent analysis to evaluate the predictive accuracy, we have used a support vector machine (SVM) classifier (15) with a linear kernel and feature ranking using the Fisher score. Predictive accuracy was determined across a range starting with all features (data from 1967 cDNA clones as mentioned above) followed by successive removal of the least discriminative feature (according to the Fisher score) through to the top two discriminative features. The test accuracy was evaluated using leave-one-out (LOO) cross-validation (15). To provide a baseline for comparison (for null prediction), we used a permutation test, i.e., the class labels were randomly shuffled 100 times, and the LOO test error was evaluated.

For hierarchical cluster analysis and principal component analysis, the Genesis software package (16) was used with the expression data of the significant genes identified by SAM and the discriminatory features identified by SVM as input.

We have analyzed the expression profiles of nine normal heterozygous BRCA1 mutation carrier breast fibroblast samples and compared these to the profiles of five reduction mammoplasty fibroblast samples with a very low probability of the presence of a BRCA1 mutation (the controls). All these samples were short-term primary cultures, and they were irradiated (15 Gy) to induce sublethal DNA damage. In a preliminary study, we also included nonirradiated samples; however, the analysis of this set showed no significant gene expression differences (data not shown), and we have therefore concluded to focus on the comparison of irradiated samples only. RNA samples were collected 1 h after the treatment for expression profiling. We also investigated the time course of expression profile changes on a smaller sample size at 15 min, 1 h, and 3 h after irradiation and found the most robust changes occurred at the 1-h time point and therefore chose this time point for the analysis of all samples. On the microarrays, 6503 cDNA clones were represented. Image analysis filtered the data as specified in “Methods and Materials,” and expression ratios of 1967 genes were included in the final data processing. We then carried out a class comparison analysis using SAM. This statistical method showed that there are significant differences between the two classes, BRCA1 mutation carriers and controls, and ranked the genes according to their contribution to the separation of the two classes.

Choosing a 20% false discovery rate, SAM called 113 genes significant; with a more stringent 10 or 5% false discovery rate, 47 and 22 genes, respectively, were selected as significant discriminatory features (supplementary data: Table A). Sometimes a smaller expression level change is more significant, as is shown in Table A, because the more consistent small changes mean a more persistent difference between the classes.

We then used these gene sets for unsupervised hierarchical clustering. The clustering dendogram based on 47 significant genes (10% false discovery rate) is shown in Fig. 1. Using the larger or smaller (SAM-113 or SAM-22) gene sets showed a very similar result (data not shown). This illustrates the clear separation of the two classes with the exception of one sample, which although from a mutation carrier, clusters with the controls at the extreme right hand site of the dendrogram. This individual has a substantial family history of breast cancer, so in this case, this is very likely to be a breast cancer causing mutation. However, there could be some other explanations for the clustering of this sample with the noncarriers such as an effect on gene expression levels by modifier genes in this particular family or the fact that this mutation leads to a truncation at the extreme COOH-terminal of the BRCA1 protein, which still maintains its transcription regulation function. This hypothesis, however, has to be tested with additional samples with this particular mutation.

Principal component analysis also separated the two classes with the input data of the gene sets as above. The plot shown in Fig. 2 is the result of this analysis using 47 significant genes (very similar results were obtained by the larger or smaller sets of 113 and 22 genes). This shows that samples from BRCA1 mutation carriers and controls are well separated with the exception of the same sample mentioned above (5328insC).

For supervised class prediction, we used a SVM classifier and LOO cross-validation to evaluate the test error. With the use of feature selection, the top 100 discriminatory genes were selected (supplementary data: Table B). We found that class prediction was most accurate if we used only 15 of the top ranked discriminatory features (ranked by the Fisher score). In this regime, the SVM learned with zero training error and achieved a minimum of two LOO test errors from 14 for a soft margin SVM and three errors from 14 for a hard margin SVM (15). To evaluate the significance of this result, we performed a permutation test in which the class labels were randomly shuffled 100 times, and the LOO test error was evaluated. This gave a test error of 7.00 ± 2.31 for the null hypothesis (no prediction ability) that is given along with the observed test error curve (supplementary data: Fig. 1.). Therefore, finding two or three test errors from 14 samples means that this is significant, rather than occurring by chance (P = 0.02–0.05). Because our dataset is small, a confirmation of this effect will require more samples for an independent test set. However, the consistent drop in the test error curve below that expected from the null hypothesis does indicate that this is a real effect. When we compared the discriminatory features from SVM (100 genes) and the significant genes from SAM (113 genes), we found that ∼80% of the genes (79 clones) were present in both lists, confirming their significant role in separating the two classes. In Table 2, the list of the 79 clones is presented with the SAM scores and Fischer scores (see “Methods and Materials”); the fold changes are also shown from the SAM analysis. These changes are mostly very subtle and often a smaller expression level change corresponds to a higher score if it is consistent because a more persistent small change can translate to a more significant value in this analysis.

Of these 79 cDNA clones, 5 are unknown expressed sequence tags, the remaining clones are known genes. Among them, several have a function in cell cycle regulation, DNA repair, and gene expression regulation. RAD51 and RAD23 were found to be significantly down-regulated in BRCA1 mutation carriers compared with controls after radiation-induced DNA damage. Because both genes have an important role in DNA damage repair, their down-regulation in BRCA1 mutation carriers could affect the DNA damage response in these cells. Interestingly, the CDKN1B/p27^Kip gene, which is a member of the cyclin-dependent kinase inhibitor family, was also significantly down-regulated in BRCA1 mutation carriers compared with controls. This gene has been previously shown to be transactivated by BRCA1 (17), and the CDKN1B protein level has been shown to be decreased during breast and ovarian tumor development (18, 19). We have found that ATR was similarly down-regulated in mutation carriers compared with the controls. This gene has been also linked with BRCA1 in the damage signaling pathway; there is evidence that ATR phosphorylates BRCA1 (20), but it is also possible that BRCA1 has an effect on the regulation of ATR as has been suggested in a recent study (21). It has been shown that BRCA1 facilitates the activity of ATR in its phosphorylation of various downstream elements. We propose that this scaffolding function of BRCA1 could be because of transcription activation.

In summary, we have shown that heterozygous BRCA1 mutation carrier fibroblasts have a distinctive gene expression phenotype after radiation-induced DNA damage. Our class prediction model was able to discriminate with 85% accuracy between normal breast fibroblasts from BRCA1 mutation carriers and fibroblasts from reduction mammoplasty specimens from women with no family history. Furthermore, the genes that exhibit significant changes in their expression level after radiation-induced DNA damage have been shown in other studies to be part of the functional pathways through which BRCA1 operates. This supports the use of this preliminary result as the basis for the development of a functional assay to assess the downstream effects of germ-line changes in the BRCA1 gene. Clearly, this initial study needs additional validation experiments, a larger sample set needs to be analyzed, and the prediction model will be tested on an independent sample set. However, our results provide evidence that expression profiling is a valuable approach in BRCA1 mutation testing, which could also be applied to BRCA2 testing. The use of fibroblasts is not the most practical as a routine screening test; however, the adaptation of these results to other cells types, e.g., lymphocytes, could lead to the development of a functional assay for BRCA1 germ-line mutation status. This would facilitate the diagnostic service in cancer genetics. Because the mutation test is expensive, samples could be prescreened by expression profiling to determine whom should be offered full mutation screening. Furthermore, it may be possible to distinguish those missense variants with functional significance from neutral polymorphisms, and it may be applicable to those families that harbor a regulatory mutation. This preliminary study has provided initial data that demonstrate that this approach is worth pursuing for clinical adaptation to functional genetic testing.

We thank all of the patients and their clinicians for participating in this study. We also thank Dr. Michael Steel for providing extra information on the family of one of our carrier samples. This work was generously supported by the legacy of the late Marion Silcock. We also thank the fund “Irene Granleese for Caroline and Olive” for their tremendous support. The microarrays were produced at the Cancer Research United Kingdom (formerly Cancer Research Campaign) Microarray laboratory at The Institute of Cancer Research, United Kingdom.