NanoString-based breast cancer risk prediction for women with sclerosing adenosis

We show that NanoString-based gene expression data can be used to model cancer risk for women with sclerosing adenosis, a common proliferative finding present in about ¼ of all benign biopsies. The expression analysis used highly fragmented RNA derived from archival FFPE benign tissue biopsies, and yet performed well with the NanoString assay. Models utilizing gene expression data performed better on average than either the BCRAT/Gail or BBD-BC models, and models including both gene expression and clinical predictor variables showed slightly improved performance compared to either gene expression or clinical predictor variables alone. Univariate modeling of randomly selected sets of the genes showed consistent association of five genes with case status, and of these five genes, two are therapeutically targetable cell surface receptors that have been implicated previously in cancer development or progression: DLK2, an effector of the NOTCH signaling pathway [8, 9], and KIT, a receptor for stem cell factor and other signaling molecules that is inhibited by imatinib [10, 11]. That, all models using DLK2 and KIT had positive coefficients for their expression values (indicating increased expression of the markers is associated with increased BC risk, Table 4) is consistent with their generally accepted roles in cancer development and progression. The other three genes that showed consistent expression across all models, SORBS2, RGS12, and EXOC6, have been investigated as predictive or prognostic cancer biomarkers [12‐15]. While it is unclear specifically how these latter three genes may contribute to breast cancer development or whether these molecules can be targeted therapeutically, consistently positive coefficients are suggestive of protumorigenic roles. Further investigations will be necessary to evaluate whether these risk markers are specific for women with SA or whether the biomarkers identified here are indicative of differential risk for all women with BBD.

Although the Breast Cancer Risk Assessment Tool (BCRAT, also referred to as the Gail model) [16] provides risk estimates at the population level, it is not as reliable when predicting risk for individual women [17, 18]. An individualized BC risk assessment model, designated the BBD-BC model, was recently developed for women with BBD and includes histologic features of the biopsy, including SA, as well as other demographic and clinical features. The BBD-BC model was found to provide improved performance for women with BBD as compared to the BCRAT model [18]. For women with SA in particular, we have found that risk stratification can be achieved by consideration of other histological and clinical features as well as expression of the proliferation marker Ki-67 [5, 19]. Our results here show that further improvements in individualized risk prediction can be obtained through examination of transcriptional biomarkers expressed in the benign breast biopsy tissue [19‐25].

Microarray-based gene expression platforms have been instrumental for advancing our understanding of breast cancer and treatment and for identifying prognostic and predictive gene signatures [26]. Although microarray-based methods work well with RNA derived from fresh or frozen samples, their mostly poorer performance with the highly fragmented RNA that is derived from FFPE biopsies has delayed their broad clinical implementation [27]. NanoString nCounter analysis methods quantify immobilized RNA using customized barcodes; because this method does not require library generation or polymerase action, it works well with FFPE-derived RNA. Compared to assays requiring fresh/frozen tissue, FFPE-based assays facilitate clinical implementation since no changes in sample collection and processing are needed. Our results presented here validate the feasibility of our overall goal to create an assay that incorporates NanoString-derived gene expression biomarkers with patient demographic information and pathological characteristics of the benign biopsy that can be applied to all women diagnosed with SA. Significant improvements in prediction ability will require application of these methods to larger patient cohorts and validation across multiple patient populations.

Strengths of our study include our focus on a SA, which is diagnosed in as many as 250,000–500,000 women per year in the United States; since SA is associated with a more than doubling of BC risk, the aggregate increased BC incidence following SA diagnosis is substantial. Moreover, unlike very high-risk lesions such as atypical hyperplasia, there are currently no clinical recommendations for women diagnosed with SA and no way to assess which of these women are at high risk and could thus benefit from interventions to reduce future BC incidence. Additionally, because SA is a cellular and homogenous lesion that can represent a substantial area of the biopsy tissue section, it represents an optimal target for methods that can identify risk signatures from RNA derived from entire tissue sections of FFPE samples, an approach that we feel will be necessary for broadest clinical translation for this patient population. Use of a NanoString-based risk signature assay offers the advantage of objective risk data and is independent of pathology interpretation. Moreover, our assay provides additional risk stratification when combined with standard clinical models, although as noted above, additional studies using larger patient cohorts will be necessary to optimally combine clinical information with transcriptional biomarkers for assessment of BC risk in patients with SA. The threshold used to determine case status from the predicted score can be optimized in future studies to reduce false positives and false negatives while balancing the consequences of each. Our study uses RNA derived from whole tissue sections rather than laser microdissected lesions; we believe this is most appropriate, since subsequent cancers derive from the complex tissue microenvironment, in which stromal factors are increasingly recognized as important in cancer progression [28]. Limitations include our focus on SA, which limits the application of our signature to this particular patient group, the relatively small number of events in each of the fivefold cross validations, the absence of a completely independent validation set and limited generalizability to women of European descent. Furthermore, our cases and controls were not matched, resulting in cases that were older and more likely to have atypia hyperplasia than controls; although this could induce potential biases, results were not substantially different when adjusted for clinical model predictions (which include age) or when restricted to subjects without atypia. Additionally, our approach used linear modeling to clarify application and interpretation; more sophisticated modeling methods that incorporate higher level feature interaction might further improve risk prediction, but these would require larger patient sample sizes.

In conclusion, we have found that the relative expression levels of a small set of genes, determined from RNA derived from FFPE-banked tissue biopsies and quantified using a clinically relevant transcriptional assay method, can be used to assess breast cancer risk for women with SA, which is found in more than 250,000 women per year in the US alone. Our results also identify specific genes that may influence breast cancer development, and thus represent potential targets for novel intervention strategies. Ultimate clinical translation of our approach will aid in decision-making for women with SA and their physicians, who would be better able to choose prevention strategies for women predicted to be at higher risk, and watchful waiting for those women predicted to be in lower risk categories [10, 11].