Abstract B2-51: SpEeCH: Quantifying Spatial Expression of Clonal Heterogeneity in breast cancer

Abstract

Abstract In routine breast cancer clinical practice, multiple serial sections are stained for guiding therapy. These include the most widely used hematoxylin & eosin (H&E) as well as established immunohistochemical biomarkers such as estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2) and increasingly so, Ki67, collectively known as the IHC4. Analysis of these sections involves visual assessment which is labor intensive and inefficient. More importantly, it ignores the wealth of spatial information in histological samples that can be used to better understand clonal heterogeneity, which may potentially be responsible for mixed response to treatment and drug resistance. Material and Methods: We have developed an automated image registration algorithm that aligns H&E and IHC4 serial section images from the same tumor by employing a combination of rigid and non-rigid registration. This automated image registration algorithm subsequently enables us to develop SpEeCH (spatial expression of clonal heterogeneity), a novel method to identify spatial clusters with differential expression. At the core of SpEeCH is a clustering method that exploits spatial cellular characteristics of tumor regions to identify spatial clonal clusters. Gaussian mixture Model (GMM) is used in combination with the Bayesian information criterion and resampling to identify the optimal number of stable clusters. As GMM ignores local spatial dependency which is an important tumor characteristic, we are also comparing GMM with conditional random fields that are undirected graphical models to account for local dependencies instead of full joint distribution. To discover spatial clusters, we applied SpEeCH to a set of 13 high-grade surgically resected breast tumors with serial sections of H&E and IHC4. Tumor regions were defined as 5 μm x 5 μm squares, resulting in >10,000 regions for clustering in each tumor. Clinical implications of the spatial clusters were verified in an independent dataset (METABRIC) that consists of gene expression whole-tumor profiles and disease-specific survival data of 2,000 breast cancer tumors. Results: By employing SpEeCH, we identified five stable spatial clusters based on histological section of H&E and IHC4 in the discovery set, with significantly different disease-specific survival validated in METABRIC gene expression data (p=0, log-rank test). This analysis revealed a group of poor prognosis ER-/HER2+/Ki67+ patients (p=0, hazard ratio = 4.4, 95% confidence interval = 3.0-6.3, log-rank test compared to a good prognosis group ER+/HER2-/Ki67-). Marked differences in prognosis among these groups highlight the significance of identified spatial clusters. For example, one of the tumors in the discovery set contains both ER-/HER2+/Ki67+ and ER+/HER2-/Ki67- clones. Therefore, such clonal heterogeneity, which is ignored during diagnosis, may help explain treatment resistance. To this end, our observation confirmed the clinical relevance of spatial clusters by leveraging a large-scale dataset under the assumption that such data represents signal from major clones. We propose that the number of clonal clusters and their spatial relationship within a tumor can be used as measures of intra-tumor heterogeneity. Their efficacy is currently being evaluated on independent clinical datasets. Conclusion: Large numbers of histology samples are available from clinical diagnostics and research laboratories, yet there is a lack of tools that facilitate objective analysis of tumor heterogeneity in these samples. Compared to high-cost multimarker immunofluorescence approaches, our automated image and statistical analysis pipeline enables efficient and effective analysis of available histology samples, which is critical for understanding cancers as highly heterogeneous diseases. Citation Format: Andreas Heindl, Adnan Mujahid Khan, Yinyin Yuan. SpEeCH: Quantifying Spatial Expression of Clonal Heterogeneity in breast cancer. [abstract]. In: Proceedings of the AACR Special Conference on Computational and Systems Biology of Cancer; Feb 8-11 2015; San Francisco, CA. Philadelphia (PA): AACR; Cancer Res 2015;75(22 Suppl 2):Abstract nr B2-51.