Abstract

   The grade of a cancer is a measure of the cancer's malignancy level,
   and the stage of a cancer refers to the size and the extent that the
   cancer has spread. Here we present a computational method for
   prediction of gene signatures and blood/urine protein markers for
   breast cancer grades and stages based on RNA-seq data, which are
   retrieved from the TCGA breast cancer dataset and cover 111 pairs of
   disease and matching adjacent noncancerous tissues with
   pathologists-assigned stages and grades. By applying a differential
   expression and an SVM-based classification approach, we found that 324
   and 227 genes in cancer have their expression levels consistently
   up-regulated vs. their matching controls in a grade- and
   stage-dependent manner, respectively. By using these genes, we
   predicted a 9-gene panel as a gene signature for distinguishing poorly
   differentiated from moderately and well differentiated breast cancers,
   and a 19-gene panel as a gene signature for discriminating between the
   moderately and well differentiated breast cancers. Similarly, a 30-gene
   panel and a 21-gene panel are predicted as gene signatures for
   distinguishing advanced stage (stages III-IV) from early stage (stages
   I-II) cancer samples and for distinguishing stage II from stage I
   samples, respectively. We expect these gene panels can be used as
   gene-expression signatures for cancer grade and stage classification.
   In addition, of the 324 grade-dependent genes, 188 and 66 encode
   proteins that are predicted to be blood-secretory and urine-excretory,
   respectively; and of the 227 stage-dependent genes, 123 and 51 encode
   proteins predicted to be blood-secretory and urine-excretory,
   respectively. We anticipate that some combinations of these blood and
   urine proteins could serve as markers for monitoring breast cancer at
   specific grades and stages through blood and urine tests.

Introduction

   Breast cancer is a major threat to women's health, accounting for 22.9%
   of cancer cases in women [[31]1]. According to the World Cancer Report
   [[32]1], 458,503 cases of breast cancer–associated deaths worldwide
   were reported in 2008, which represents 13.7% of cancer-related deaths
   in women. It has been generally understood that breast cancer, probably
   other cancer types as well, of different stages and different grades
   require different treatment plans. For example, breast-conserving
   surgery plus radiation therapy is effective for most patients with
   early stage breast cancers [[33]2] while systemic therapy are generally
   needed for advanced stage patients, such as hormone or chemo therapy,
   in addition to cancer-removal surgery and radiation. In addition,
   cancer grades are strongly associated with prognosis [[34]3].
   Specifically, more differentiated cancer grades tend to have more
   favorable prognosis. Clearly, correct classification of the grade and
   stage of a cancer has significant implications in determination of the
   treatment plan for a patient.

   Cancer stages are used to reflect the size of a cancer tumor and its
   extent of invasion. It has been traditionally determined by cancer
   pathologists based on tumor size, nodal spread and metastasis [[35]4].
   In the recent past, molecular level information has been incorporated
   into the decision process of cancer staging, using markers such as
   alpha-fetoprotein and lactate dehydrogenase for determination of germ
   cell tumors [[36]5]. A widely used system for cancer staging is that
   the cancer tissues are classed into four stages, namely I, II, III and
   IV, with a higher stage representing a more advanced cancer.

   Cancer grading is a measure of the malignancy and aggressiveness
   independent of stage. Unlike staging, cancer grading has been
   predominantly done through visual inspection of the cell morphology and
   tissue structure [[37]3], generally lacking in using molecular level
   information. Compared to stage determination, it is a less developed
   area in cancer classification. Currently there is no universal grading
   system for all cancer types, instead research communities of a few
   cancer types each have developed their own grading systems such as the
   one for breast cancer developed by Bloom and Richardson [[38]6], the
   Gleason system for prostate cancer [[39]7] and the Fuhrman method for
   kidney cancer [[40]8]. While there are some differences in the detailed
   classification criteria, these grading systems generally classify
   cancer tissues to four grades: well differentiated (WD), moderately
   differentiated (MD), poorly differentiated (PD) and undifferentiated
   (UD).

   A number of computational studies have been published on cancer staging
   and grading prediction based on transcriptomic data. For example, Cui
   et al have reported a 198-gene and a 10-gene panel for grading and
   staging prediction of gastric cancers, respectively [[41]9]. For breast
   cancer, a grade index based on the expressions of 97 genes in cancer
   tissues was previously developed to classify patients with grade 2
   tumors into two subgroups with high versus low risks of recurrence
   [[42]10]. However, markers so developed have had only limited
   applications since tissue-based gene-expression data are generally not
   available for most patients [[43]11, [44]12]. Hence, it is essential to
   extend tissue-based gene markers to markers that can be measured using
   blood or urine samples of patients [[45]13, [46]14], the challenge of
   which is to predict reliably which of the overly expressed proteins in
   cancer tissues can be secreted into blood and further into urine.

   In this study, we conducted a computational analysis tissue-based
   gene-expression data to identify possible gene signatures and
   blood/urine proteins markers for breast cancer grading and staging
   prediction. The following represents the unique contributions by this
   study, to the best of our knowledge: (1) RNA-seq-based gene-expression
   signatures for breast cancer grading and staging prediction; and (2)
   predicted potential marker proteins for cancer staging and grading that
   can be measured by using blood and urine samples. Clearly, this work
   represents only a pilot study for prediction of blood and urine marker
   proteins for breast cancer grading and staging. We expect that
   follow-up studies will demonstrate the feasibility of the predicted
   signature genes and protein markers.

Results

A. Identification of gene signatures for breast cancer

(1) Identification of gene groups whose expressions distinguish breast cancer
from other cancers

   Gene-expression data of 111 paired of breast cancer and adjacent
   control tissue samples were retrieved from the TCGA database [[47]15],
   where each gene-expression dataset covers 20,501 human genes measured
   using RNA-seq. 5,562 differentially expressed genes between cancer and
   matching control tissues were identified using the following procedure:
   the expression levels of a gene in cancer show at least 2-fold change
   from the matching control tissues with the q-value < 0.05 to control
   the False Discovery Rate (FDR) (see [48]Material and Methods). Among
   the 5,562 genes, 2,078 were up-regulated and 853 of them were found to
   be up-regulated in less than three out of 12 other cancer types that
   were examined in our study as references, hence making them as good