Published in Volume
124, Issue 2
(February 3, 2014)J Clin Invest.
Copyright © 2014, American Society for Clinical
Taxonomy of breast cancer based on normal cell phenotype predicts
1Department of Pathology, Brigham and Women’s Hospital and
Harvard Medical School, Boston, Massachusetts, USA.
Pathology, Interdisciplinary Stem Cell Institute, Braman Family Breast Cancer Institute,
and Sylvester Comprehensive Cancer Center, Miller School of Medicine, University of
Miami, Miami, Florida, USA.
3Department of Epidemiology, Harvard School
of Public Health, Boston, Massachusetts, USA.
4Channing Division of
Network Medicine, Department of Medicine, Brigham and Women’s Hospital and
Harvard Medical School, Boston, Massachusetts, USA.
Comprehensive Cancer Center, University of North Carolina at Chapel Hill, Chapel Hill,
North Carolina, USA.
6Department of Biostatistics and Computational
Biology, Dana-Farber Cancer Institute, Harvard School of Public Health, Boston,
7Department of Pathology, Beth Israel Deaconess
Medical Center and Harvard Medical School, Boston, Massachusetts, USA.
Address correspondence to: Tan A. Ince, Department of Pathology, Braman Family
Breast Cancer Institute, Interdisciplinary Stem Cell Institute, and Sylvester
Comprehensive Cancer Center, University of Miami Miller School of Medicine, BRB, Room
907, Miami, Florida 33136, USA. Phone: 305.243.1782; Fax: 305.243.9376; E-mail:
First published January 27, 2014
Received for publication May 8,
2013, and accepted in revised form October 17,
Accurate classification is essential for understanding the pathophysiology of a
disease and can inform therapeutic choices. For hematopoietic malignancies, a
classification scheme based on the phenotypic similarity between tumor cells and
normal cells has been successfully used to define tumor subtypes; however, use of
normal cell types as a reference by which to classify solid tumors has not been
widely emulated, in part due to more limited understanding of epithelial cell
differentiation compared with hematopoiesis. To provide a better definition of the
subtypes of epithelial cells comprising the breast epithelium, we performed a
systematic analysis of a large set of breast epithelial markers in more than 15,000
normal breast cells, which identified 11 differentiation states for normal luminal
cells. We then applied information from this analysis to classify human breast tumors
based on normal cell types into 4 major subtypes, HR0–HR3, which were
differentiated by vitamin D, androgen, and estrogen hormone receptor (HR) expression.
Examination of 3,157 human breast tumors revealed that these HR subtypes were
distinct from the current classification scheme, which is based on estrogen receptor,
progesterone receptor, and human epidermal growth factor receptor 2. Patient outcomes
were best when tumors expressed all 3 hormone receptors (subtype HR3) and worst when
they expressed none of the receptors (subtype HR0). Together, these data provide an
ontological classification scheme associated with patient survival differences and
provides actionable insights for treating breast tumors.
Common classification terminology is necessary for medical progress. Over the past 2
centuries, normal tissue morphology and function has been successfully used as a
reference point to define various diseases. Most notably, such an approach has been used
to classify hematopoietic tumors, such as lymphomas and leukemias (1). The discovery of the morphologic and molecular resemblance of
various subtypes of leukemias and lymphomas to particular normal hematopoietic cell
types was critical in this process.
Based on this insight, hematopoietic malignancies have been classified as B cell and T
cell neoplasms (e.g., small lymphocytic, large B cell, lymphoblastic, follicular, and
mantle cell) that resemble specific normal cell types. Similarly, myeloproliferative
diseases are classified as neutrophilic, granulocytic, lymphoblastic, prolymphocytic,
myeloid, promyelocytic, monocytic, erythrocytic, basophilic, and megakaryoblastic
neoplasms. Some of the most notable and earliest strides against cancers have been made
in the treatment of hematopoietic malignancies (2). While many factors have contributed to this success, the accurate
classification of hematopoietic malignancies played an important role. The
identification of cell-type specific cluster of differentiation (CD) markers on the
surface of these cells permitted efficient immunophenotyping (3). These CD markers were later used to identify lymphomas and
leukemias with a phenotype nearly identical to a specific normal cell type, allowing the
development of the current classification system of these diseases (4). Despite major successes in rationally classifying
and treating hematological malignancies, the use of normal cell types to classify solid
tumors has not been widely emulated. A major reason for this has been our lack of
understanding of the diversity of cell types in most solid tissues.
Characterization of normal cell subtypes in solid tissues has been challenging. Until
recently, only 2 cell types have been morphologically described in the human breast: the
inner luminal cells and the outer myoepithelial cells (5). This limited understanding of the cell types comprising the breast ducts
has precluded the development of a normal cell type–based classification system.
While there has been more recent interest in normal breast cell subtypes, this research
has been difficult to correlate with existing human breast tumor phenotypes (6). Numerous markers have been used to describe
normal human mammary stem/progenitor cells, including
CD44hiCD24lo, aldehyde dehydrogenase–high
(ALDHhi), CD10+, Ep-CAM+MUC1–,
and Ep-CAMhiCD49f+. Whether these stem/progenitor cell markers
identify the same cell populations remains unknown. Furthermore, Tlsty and colleagues
discovered that human breast cells can exhibit extensive lineage plasticity (7), which may explain why marker profiles have been
difficult to associate with distinct tumor subtypes.
Clinically, human breast cancers are grouped into 3 categories based on the presence of
estrogen receptor (ER+), progesterone receptor (PR+), and human
epidermal growth factor receptor 2 (HER2+), or by their absence in
triple-negative breast cancers (TNBCs; i.e.,
In the research setting, mRNA profiles have been used to define prognostic subtypes of
breast cancer: luminal A, luminal B, basal-like, claudin-low, and HER2-like (8). DNA methylation patterns have also been used to
identify 5 distinct DNA methylation groups (9),
and 10 different breast cancer clusters have been identified in a genome-driven
integrated classification system, each associated with distinct clinical outcomes (10, 11).
Several additional mRNA expression–based molecular prognostic panels, such as
Oncotype Dx, PAM50, and MammaPrint, have also emerged with potential clinical utility
The main evidence supporting the importance of each of these molecular subtypes has been
identification of patient groups with different outcomes. Hence, it is important to
recognize that these molecular subtypes are prognostic categories, different from
disease taxonomy. Therefore, while these molecular prognostic tools have been useful in
the research setting, they have not produced a commonly agreed-upon new system of
classification that is uniformly used in the clinic. This is partly because each
molecular platform appears to produce a different prognostic classification. A breast
cancer task force recently concluded that at the moment, molecular tools do not provide
sufficiently robust information beyond histological type, grade, and ER/PR/HER2 status
(13). Thus, these molecular tests are not
routinely performed at most institutions (14).
It is increasingly becoming clear that a more fundamental breast cancer classification
system, one that does not conflate prognostic categories with diagnostic categories, is
needed. Ideally, such a system should be robust and not change depending on which
technological platform is used to classify breast cancer. Inspired by the classification
of hematopoietic malignancies, we hypothesized that differentiation states of normal
cell populations in normal human breast may provide such a reference classification
system for human breast tumors.
The normal human breast is composed of milk-producing lobules and interlobular ducts
that transport the milk to the nipple (Supplemental Figure 1A; supplemental material
available online with this article; doi: 10.1172/JCI70941DS1).
This anatomical distinction is important for understanding breast cancer, because in
addition to ER/PR/HER2 status, human breast tumors are classified by pathologists on
morphological grounds, either as ductal carcinomas or as
lobular carcinomas, for reasons unrelated to their cell of origin.
This arcane terminology has resulted in a common misconception that ductal and lobular
breast cancers initiate in the normal ducts and lobules, respectively. However, despite
their names, almost all of the early progression steps for both tumor types almost
exclusively involve the breast lobules. Thus, in the present study, we specifically
examined the normal cells in the lobules using immunohistochemical (IHC) staining, which
preserves tissue architecture and allows for discrimination of ducts, lobules, and
different layers of the epithelium (see below). For a list of the 37 primary antibodies
used in these studies, see Supplemental Table 1.
Analysis of CD markers and intermediate filaments in normal human breast. An ideal cell type–specific immunostain marker should have a bimodal
expression pattern (i.e., one subpopulation is clearly negative, and the other
strongly positive). While CD markers have been useful in isolating breast cell types
using FACS, we found that they had a gradient-type expression pattern in situ that
limited their utility to define cell subtypes using semiquantitative methods such as
IHC (Supplemental Figure 1).
In an attempt to identify molecules with bimodal expression patterns in normal human
breast, we examined the expression of intermediate filaments. These molecules are
differentially expressed in distinct cell types, and their expression is both tissue-
and cell type–specific. Furthermore, it has been well recognized that cell
type–specific expression of intermediate filaments is preserved in tumors and
can be used to determine the tissue origin of tumors (15). We found that keratin 5 (K5), K7, K8, K14, K17, K18, and K19 were
useful in identifying subpopulations of human breast cells, because they were
expressed in a bimodal pattern (Supplemental Figure 2).
Next, we subjected normal breast tissues from 36 breast reduction mammoplasty
procedures to IHC with K5, K7, K8, K14, K17, K18, K19, CD10, SMA, and p63. Normal
breast lobules and ducts are lined by a bilayer epithelium, consisting of an inner
layer of milk-producing luminal cells and an outer layer of supportive myoepithelial
cells. As previously shown (16), we found that
K7, K18, and claudin-4 (Cld-4) were expressed in all luminal cells, but not in
myoepithelial cells (Figure 1A and Supplemental
Figure 2A). In contrast, CD10, SMA, and p63 were expressed in all myoepithelial
cells, but not in luminal cells (Supplemental Figure 2). Thus, these markers
constitute a pan-luminal versus pan-myoepithelial panel. Interestingly, in some
lobules, luminal cells were K19– (Figure 1B); thus, K19 was not a pan-luminal marker.
Expression of intermediate filaments and ER in normal human breast. Single and double IHC with immunoperoxidase
(A–E, G, I, and
K) and merged IHC images (F and J) of
normal human FFPE sections are shown. (A) K7/18 (brown).
(B) K18 (red) and K19 (brown). (C) K5/14 (brown).
(D) CD10 (red) and K14 (brown). (E) K5/14 (brown) and
SMA (red). (F) K18 (green) and K14 (red). Merged
K14+K18+ appears yellow. (G) K5/14 (red) and
ER (brown). We designated this population of cells K5/14/17+ because
the tissue sections were not stained simultaneously with these markers.
(H) Differentiation states of normal luminal epithelial cells,
based on expression of ER and keratins. (I) Ki67 (brown) and K5/14
(blue). (J) ER (green) and Ki67 (red). (K) K18 (red) and
Ki67 (brown). (L) Differentiation states of normal luminal epithelial
cells, based on ER, keratins, and Ki67. Representative images were selected from
multiple patient samples (n = 36). Original magnification,
×20 (A); ×40 (B); ×200
(F); ×400 (C, G, and
I–K); ×600 (D and
E). See http://sylvester.org/ince
for additional high-resolution images.
In human skin, K5/14/17 are exclusively expressed in the basal layers; in mouse
mammary tissue, they are expressed in the myoepithelial layer (Supplemental Figure
2C). Hence, these keratins are usually referred to as basal
keratins. However, in normal human breast tissue, K5/14/17 were expressed in
both luminal and basal layers, depending on location. In the interlobular ducts,
K5/14/17 were expressed in the myoepithelial (basal) layer, as expected (Supplemental
Figure 1B and Supplemental Figure 2, D–F). However, in the lobules, the site
where precursor lesions develop, K5/14/17 were expressed in the luminal layer (Figure
1, C–E, Supplemental Figure 1J,
Supplemental Figure 2, G–I, and ref. 6). We confirmed the luminal nature of these cells with double IHC, which
demonstrated that the K5+, K14+, or K17+ cells were
Ki67/ER– (Figure 1,
G–J) and CD10/SMA/K17– (Supplemental Figure 2,
J–L) and were located above the CD10/SMA/K17+ myoepithelial cell
layer (Supplemental Figure 2, J–L). We did not find luminal K5+
cells in the mouse breast (Supplemental Figure 2C).
We identified luminal K5+, K14+, or K17+ cells in
all 36 patients examined; thus, this was a robust and highly reproducible luminal
cell subpopulation. Interestingly, while some lobules had a small percent of luminal
K5+, K14+, or K17+ cells, adjacent lobules were
entirely composed of K5+, K14+, or K17+ luminal
cells (Supplemental Figure 2, M–P).
When 2 different cell lineages are defined by mutually exclusive expression of
markers, coexpression of these markers in the same cell has been used as evidence of
“stemness.” Previously, coexpression of K5/14/17 with K7/8/18 has
been interpreted as evidence for bipotential cells. Here, however, some lobules were
entirely composed of K14+K18+ or K5+K18+
double-positive cells in nearly every tissue section examined (Figure 1F and Supplemental Figure 2, Q–U). On
average, 36% of luminal cells were K14+K18+ (n
= 746), and 16% were K5+K18+ (n = 1,339).
Importantly, K5/14+ cells also expressed MUC1, a marker of luminal
differentiation (Supplemental Figure 2V). It would be extremely unusual to find an
epithelial tissue entirely composed of progenitor/stem cells. Thus, our results
indicated that the luminal layer cells coexpressing K5/14/17 with K18/19 are more
consistent with a differentiated luminal cell variety (6, 17).
Analysis of hormone receptors in normal human breast. Having identified 2 subtypes of luminal layer cells based on K5/14/17 expression, we
next characterized the expression of hormone receptors (HRs) in these cells, because
they are involved in differentiation and some have a bimodal expression pattern.
In an initial survey, 3 receptors — ER, androgen receptor (AR), and vitamin D
receptor (VDR) — stood out with distinct bimodal expression patterns. Many of
the other HRs (i.e., TRHα, TRHβ, PTH1R, OXTR, SSTR1, SSTR2, SSTR3,
SSTR5, RARα, RARβ, RXRα, and RXRβ) did not appear to
have a bimodal expression pattern. Because PR expression tracks with ER expression,
we did not include PR in this study.
Next, we carried out double IHC on normal breast sections and counted cells in 5
different sections for coexpression of various markers (Supplemental Table 2). Double
IHC demonstrated that all ER+ cells were luminal and did not overlap with
K5/14/17+ luminal cells (<0.3% overlap, n = 3,313)
or with Ki67+ proliferating cells (0.1% overlap, n =
1,206) (Figure 1, G and J, and Supplemental
Table 2). Nearly all proliferating Ki67+ cells were K18+
luminal cells that were negative for the myoepithelial markers CD10 (0.5% overlap,
n = 1,084) and K5/14/17 (0%–1.9% overlap,
n = 1,078) (Figure 1K and
Supplemental Table 2). These results allowed us to define 4 mutually exclusive
subsets of luminal cells in normal human breast that were all positive for the
pan-luminal markers K7 and K18 (Figure 1L): (a)
ER+ cells, (b) K5/14/17+ cells, (c)
ER–K5/14/17– cells, and (d) Ki67+
Double IHC demonstrated that all AR+ cells were luminal, and they were
also mutually exclusive with K5/14+ cells (0.0% overlap,
n = 789) and Ki67+ cells (0.0% overlap,
n = 698) (Figure 2, A and B,
and Supplemental Table 2). AR+ cells partially overlapped with
ER+ cells (44% overlap, n = 429) (Figure 2C and Supplemental Table 2). These results allowed
us to describe 3 subsets of HR+ cells: ER+, AR+, and
ER+AR+ (Figure 2D).
Double IHC demonstrated that VDR+ cells were exclusively in the luminal
layer as well, with no overlap with CD10+ myoepithelial cells or
proliferating Ki67+ cells (0.0% overlap, n = 179), but
they did partially overlap with K5/14+ cells (15%–23% overlap,
n = 266), AR+ cells (16%–35% overlap,
n = 835), and ER+ cells (22%–74% overlap,
n = 749) (Figure 2,
E–I, and Supplemental Table 2).
Expression of intermediate filaments, ER, AR, and VDR in normal human
breast. Double IHC (A and J) and merged images (B,
C, E–I, and
K–M) of normal human breast FFPE sections, as
well as differentiation states of luminal (D and N) and
myoepithelial (O) cell types, are shown. (A) K5/14 (red)
and AR (brown). (B) AR (green) and Ki67 (red). (C) ER
(green) and AR (red). Merged ER+AR+ appears yellow.
(D) Differentiation states of normal luminal epithelial cells
based on presence of ER, keratins, Ki67, and AR. (E) CD10 (green) and
VDR (red). (F) VDR (red) and Ki67 (green). (G) K5
(green) and VDR (red). (H) AR (green) and VDR (red). Merged
AR+VDR+ appears yellow. (I) ER (green) and
VDR (red). Merged ER+VDR+ appears yellow. (J)
CD10 (red) and Ki67 (brown). (K) ER (green), AR (red), and VDR
(blue). Merged ER+AR+ appears yellow; merged
ER+VDR+ appears purple. (L) ER (green), AR
(green), and VDR (red) shown individually. In the merged image,
ER+AR+VDR+ (i.e., HR3) appears white.
(M) HR3 (green), Ki67 (red), and DAPI (blue; nuclear marker).
(N and O) Differentiation states of normal luminal
(N) and myoepithelial (O) breast cells based on the
full marker panel. Representative images were selected from multiple patient
samples (n = 36). Original magnification, ×200
and M); ×400 (L). See http://sylvester.org/ince
for additional high-resolution images.
Triple IHC also demonstrated the presence of triple-HR+ cells (i.e.,
ER+AR+VDR+; Figure 2, K and L). These results allowed us to describe 7 subsets of
HR+ cells in the luminal layer of human breast lobules: ER+,
AR+, VDR+, ER+AR+,
ER+VDR+, AR+VDR+, and
ER+AR+VDR+ (Figure 2N). Interestingly, only VDR+ cells substantially overlapped
with K5/14/17+ luminal cells, and the proliferating
K18+Ki67+ luminal cells were
Cumulatively, in the luminal layer of normal human breast, we were able to define 11
differentiation states (Table 1), including 3
HR– states (collectively designated group HR0, states
which were either K5/14/18– (L2; 52%–83%) or
K5/14/18+ (L3; 17%–48%, n = 2,085), and 8
HR+ states, grouped as single-HR+ (HR1, states
L4–L7; ER+, AR+, or VDR+),
double-HR+ (HR2, states L8–L10; ER+AR+,
ER+VDR+, or AR+VDR+), or
triple-HR+ (HR3, state L11;
Cellular differentiation states in normal human breast lobules
In the myoepithelial layer, all cells expressed CD10, SMA, and p63, with 2 subtypes,
K5/14/17– and K5/14/17+ (designated My1 and My2,
respectively; Table 1). Proliferating cells
were very uncommon in the myoepithelial layer; CD10 and Ki67 overlapped in only 0.5%
of the cells (Figure 2J, Table 1, and Supplemental Table 2).
Simultaneous examination of 12 markers in normal human breast with a novel
multiplex immunofluorescence method. In the above experiments, we were able to stain the same formalin-fixed,
paraffin-embedded (FFPE) section with up to 3 different antibodies simultaneously. A
greater number of antibodies is difficult to multiplex by conventional methods, for
multiple reasons (see Supplemental Methods).
To confirm simultaneous coexpression patterns predicted by double and triple IHC for
all 12 different markers (ER, AR, VDR, K5, K7, K8/18, Cld-4, CD10, SMA, Ki67,
NaKATPase, and DAPI), we wanted to examine their expression in the same cells.
Recently, a new technology has been developed by GE Healthcare that allows for
immunofluorescence (IF) of the same tissue section with more than 10 different
antibodies serially (known as multiplex IF; ref. 18), which was used to confirm all of our results (Figure 3, Supplemental Figure 3, and Supplemental
Multiplex IF of 12 markers in normal human breast. (A–I) 1 FFPE section of normal breast epithelium
was stained serially with each antibody for the markers (A)
pan-keratin (Pan-K, green), (B) K18 (red), (C) K5 (red),
(D) DAPI (blue), (E) ER (green), (F) AR
(green), (G) VDR (red), (H) Ki67 (red), and
(I) SMA (green). (J–O) The
individual IF staining images were merged to reveal the coexpression pattern of
all markers in each cell. (J) K5 (red) and SMA (green).
(K) K5 (red) and K18 (green). (L) ER (red), AR
(green), and K5 (blue). (M) VDR (red) and ER (green).
(N) VDR (red) and AR (green). (O) AR (red), ER (green),
and VDR (blue). (P) Differentiation states of normal luminal breast
cells based on the full marker panel. Representative images were acquired using
multiplex IF technology (GE Healthcare). Original magnification, ×200
(A–O). See http://sylvester.org/ince for
additional high-resolution images, including K7, Cld-4, NaKATPase, and CD10
We used image analysis software for quantitative analysis of our multiplex IF for ER,
AR, VDR, K5, and Ki67 in individual cells. Each cell was numbered by the image
analysis software, and the fluorescent signal specifically from the luminal
epithelium was measured for each marker. We plotted the results for each marker as a
percentage of total fluorescence for each cell. This analysis allowed us to correlate
the expression of these markers in >300–500 individual cells in lobules
from 8 different patients (Figure 4 and
Supplemental Figure 3C). Based on the double IHC analysis (Figures 1 and 2), we
had deduced that there were inverse correlations between Ki67 and K5, between Ki67
and ER/AR/VDR, and between K5 and ER/AR (Supplemental Table 2). Multiplex IF allowed
us to demonstrate all of these complex trends for the first time in individual cells
(Figure 4 and Supplemental Figure 3).
Multiplex analysis of 12 markers in normal human breast. Histograms of relative ER, AR, VDR, K5, and Ki67 expression in each luminal cell
in normal human breast lobules 1–4. Cell number is plotted against percent
contribution of each marker to total fluorescence of each cell. See Supplemental
Figure 3 for additional lobules.
These observations highlighted 4 predominant, mutually exclusive differentiation
patterns in the luminal layer: a HR+ state, a proliferative state
(Ki67+), and 2 HR– states, one K5+ and
the other K5– (Figure 4 and
Supplemental Table 2). Consistent with this, we observed that Ki67+ and
K5+ cells were rare in lobules that were enriched in HR+
cells (Supplemental Figure 3D). When K5+ cells expanded, HR+
and Ki67+ cells decreased, and in highly proliferative areas, there were
very few HR+ or K5+ cells (Supplemental Figure 3D). Hence, it
appears that a given cell can exist in only one of these differentiation states at
Analysis of HR+ and HR– cell types in
ER+ breast tumors. The remarkable heterogeneity observed at the single-cell level in normal breast
epithelium was reminiscent of the distinct features of normal hematopoietic cell
populations. Because hematological malignancies maintain normal cell type–
and differentiation-specific patterns, we next asked whether breast tumors do as
We evaluated the staining pattern of 12 protein markers in 20 full FFPE sections
using IHC, then confirmed the results using tissue microarrays (TMAs) that contained
216 tumors (51 ER+, 46 HER2+, and 119 TNBC) (Figure 5). Staining was scored by combining percent
expression and staining intensity on a 0–25 expression scale (Supplemental
Identification of normal cellular phenotypes in human breast tumors. Heat maps of Cld-4, K7, K18, VDR, AR, K5, K14, CD10, SMA, p63, PR, ER, and HER2
protein levels in 216 human breast cancer tumors, separated into (A)
ER+ (n = 51), (B) HER2+
(n = 46), and (C) TNBC (n =
119). Luminal markers (Cld-4, K7, K18, VDR, and AR) and basal markers (CD10, SMA,
and p63) are indicated. TNBCs are separated into luminal 1 (LM1;
K5/14–), luminal 2 (LM2; K5/14+), and mixed (M;
expressing both luminal and myoepithelial markers) subtypes. TMA sections were
subjected to IHC and scored using light microscopy on a scale of 0 (blue, low
expression) to 25 (yellow, high expression), with white denoting intermediate
expression. Corresponding normal cell counterparts are illustrated next to each
In both tissue sets, we observed that all ER+ human breast cancers
strongly expressed multiple pan-luminal markers (Cld-4, K7, and K18), whereas none
were positive for pan-myoepithelial markers (CD10, SMA, and p63) (Figure 5A). All ER+ breast cancers were
K5/14–. Interestingly, the majority of ER+ tumors
were VDR+ (93%), and two-thirds were AR+ (59%) (Figure 5A). This pattern was identical to that of normal
breast ER+ cells, which could coexpress AR or VDR, but were very rarely
K5/14/17+ or CD10/SMA+. These results indicate that all
ER+ tumors have a luminal phenotype identical to HR+ normal
luminal cell type L4, L8, L9, or L11 (Figure 5A
and Supplemental Table 3). Intriguingly, we also observed that, similar to normal
tissues, most proliferating tumor cells (Ki67+) were
ER–AR–VDR–, with focal
VDR+ proliferating tumor cells (Supplemental Figure 5).
Analysis of HR+ and HR– cell types in
HER2+ breast tumors. In HER2+ tumors, we observed strong expression of multiple pan-luminal
markers (Cld-4, K7, and K18) and none of the pan-myoepithelial markers (CD10, SMA,
and p63) (Figure 5B). Nearly all
HER2+ tumors (44 of 46) had a luminal phenotype identical to that of
HR+ normal breast cells (i.e., L4–L11). A minority of
HER2+ tumors (2 of 46) were similar to HR– cells
(Figure 5B and Supplemental Table 5).
Analysis of HR+ and HR– cell types in TNBCs. TNBCs are defined as
ER–PR–HER2–. We examined
119 TNBCs for the expression of HRs and keratin markers, which revealed 3 major
subgroups (Figure 5C and ref. 19). Nearly 66% of TNBCs (78 of 119) had a pure
luminal phenotype, positive for pan-luminal markers and negative for
pan-myoepithelial markers; of these, 37 were identical to K5/14–
HR0 luminal cells (designated luminal 1), and 41 were identical to K5/14+
HR0 luminal cells (luminal 2) (Figure 5C and
Supplemental Table 5). All remaining TNBCs (33%, n = 41) strongly
expressed luminal markers (Cld-4, K7, AR, VDR, and K18), but 38 of these tumors also
expressed myoepithelial markers (CD10, SMA, and p63) (Figure 5C), consistent with a mixed phenotype.
In summary, 95% of human breast tumors were phenotypically identical to one of the
normal luminal breast cell subtypes (Supplemental Table 3), similar to lymphomas and
leukemias. For the remaining 5% (HR0 tumors with a mixed phenotype), it is possible
that their normal counterparts are rare progenitor cells with a mixed
luminal/myoepithelial phenotype (20), or that
these tumors exhibit an altered phenotype due to mutations that result in
inappropriate expression of these markers.
Expression of normal basal versus luminal-specific mRNAs in TNBC. The cell of origin of TNBC has been of great interest recently (21). As mentioned above, K5/14/17 are expressed in the basal
layers of human skin and rodent mammary glands (Supplemental Figure 2C). Thus, these
keratins have been commonly referred to as basal keratins in the literature (17), and TNBCs that express them have been called
basal-like carcinoma (BLC) (22). Consequently,
some have suggested that these tumors are similar to myoepithelial (basal) cells of
the normal breast. However, as we demonstrated here, K5/14/17 were predominantly
expressed in the luminal layer of normal human breast lobules, and K5/6+
BLCs expressed markers identical to those of L3 and L7 luminal cells (Table 1 and Supplemental Table 3).
Because the TNBC/BLC category was based on mRNA expression in microarray analysis
(23–25), we also carried out an analysis of mRNA in normal human breast cells,
by combining results from 3 different studies that profiled highly purified luminal
versus myoepithelial cells (26–28). We found that 131 mRNAs were identified as
luminal-specific and 90 as myoepithelial-specific in at least 2 of the 3 datasets
(Supplemental Table 4), providing a strong consensus signature distinguishing normal
luminal versus myoepithelial cells. Next, we examined the expression of these genes
in basal-like and non-basal-like human breast tumors (29–32).
Interestingly, no significant correlation was observed between basal-like tumors in
these cohorts and the expression signature of normal basal/myoepithelial cells
(P = 0.22, Fisher exact test; Supplemental Figure 6A). Thus, the
differentiation state of BLC is most similar to K5/14/17/18+ normal
luminal cells of the breast (L3; Supplemental Table 5), and the name
basal-like is probably not an accurate description of neither
their differentiation state nor their cell of origin (6, 17, 21, 33, 34). In some cohorts, patients with basal-like
tumors have a worse outcome than those with TNBC tumors (22); here, we did not observe a significant difference between
K5/6+ versus K5/6– TNBC patients (Supplemental
Distribution of HR0–HR3 breast tumor phenotypes in the NHS
cohort. Based on the above results, we hypothesized that human breast tumors can be
classified according to normal breast differentiation states and tested this
hypothesis using a breast cancer cohort from the Nurses’ Health Study (NHS),
with >25 years of follow-up from a large number of patients (n =
1,731) (35–37). We conducted IHC of NHS TMAs with ER, PR, HER2, VDR, AR,
K8/18/Cld-4, K5/6, and CD10/SMA/p63 antibodies and scored them semi-quantitatively
into 4 categories based on normal tissue differentiation: HR3
(ER+AR+VDR+), HR2 (ER+AR+,
AR+VDR+, or ER+VDR+), HR1
(ER+, VDR+, or AR+), and HR0
Importantly, the 4 HR categories are different from the current ER+,
HER2+, and TNBC classification. For example, based on standard
classification, 75% of NHS study patient tumors were ER+
(n = 1,356), 10% were HER2+ (n =
177), and 15% were TNBC (n = 253) (Figure 6A and Supplemental Table 5). These were reclassified as 58.1% HR3
(n = 1,006), 24.8% HR2 (n = 429), 10.7% HR1
(n = 185), and 6.4% HR0 (n = 111) (Figure 6B), because each standard breast cancer subtype
was composed of multiple HR groups: of ER+ tumors, 75.1% were HR3, 23.4%
were HR2, and 1.5% were HR1; of HER2+ tumors, 29.4% were HR3, 43.5% were
HR2, 22.0% were HR1, and 5.1% were HR0; of TNBC tumors, 36.8% were HR0, 44.6% were
HR1 (i.e., AR+ or VDR+), and 18.6% were HR2 (i.e.,
AR+VDR+) (Supplemental Figure 7A and Supplemental Table 5).
Thus, our HR-based classification approach does not merely rename existing groups,
but organizes tumors in a new way.
Normal cell subtype-based classification identifies breast cancers with
different outcomes. (A) Distribution of ER+, HER2+ and TNBC cases
from the full panel of NHS cases analyzed in this study. (B)
Reclassification of ER+, HER2+, and TNBC human breast tumors
from the full panel of NHS cases analyzed in this study as HR3
(ER+AR+, AR+VDR+, or
ER+VDR+), HR1 (ER+, VDR+, or
AR+), and HR0
tumors were divided into the 4 HR0–HR3 categories based on normal tissue
differentiation (see Supplemental Table 3). (C) Kaplan-Meier analysis
for overall survival of all individuals with invasive breast cancer from the NHS,
scored by IHC. (D) Kaplan-Meier analysis of relapse-free survival for
all invasive breast cancers from an 855-patient breast tumor dataset (38). Tumors were ranked according to gene
expression values for ER, AR, and VDR, scored as high or low based on a 50% cutoff
point, and assembled based on HR status (HR0, n = 141; HR1,
n = 287; HR2, n = 284; HR3,
n = 143).
Analysis of breast cancer outcomes based on normal cell lineage
phenotypes. We next investigated whether the HR0–HR3 categories correlated with breast
cancer survival and found a strong association between the total number of positive
receptors and outcome. Kaplan-Meier analyses of the NHS cohort showed that patients
with HR3 tumors had the best survival, those with HR1 tumors had the worst survival,
and those with HR2 tumors had intermediate survival (P < 0.0001;
Figure 6C). In multivariate analysis, these
differences remained significant: compared with HR3 tumors, the relative hazard ratio
(RHR) for HR2 tumors was 2.9 (95% CI, 1.60–5.21); for HR1 tumors, the RHR was
5.3 (95% CI, 2.77–9.97), and for HR0, the RHR was 6.9 (95% CI,
3.37–14.39) (Supplemental Table 6).
Interestingly, the HR0 group had a biphasic outcome curve similar to that of HR1
tumors, with the worst outcome during the first 5 years, followed by a flat curve
thereafter (Figure 6C), consistent with an
excellent outcome. Thus, we reevaluated the association stratified by time, before
and after a 5-year cutoff. During the first 5 years, HR3 tumors had the best outcome;
compared with HR3, HR2 tumors had a worse outcome (RHR, 1.69; 95% CI,
1.14–2.50), and HR1 tumors (RHR, 2.44; 95% CI, 1.55–3.84) and HR0
tumors (RHR, 2.7; 95% CI, 1.56–4.70) had the worst outcome
(P < 0.0001; Supplemental Figure 7, B and C, and Supplemental
Table 6). After 5 years, there was no significant difference among HR3, HR2, and HR1
(P > 0.5), but HR0 had a better outcome (RHR, 0.34;
P = 0.02; Supplemental Figure 7D). Analyzing the HER2 groups
separately did not change these results (Supplemental Figure 7E). In a multivariate
analysis, these differences remained significant even after accounting for other
factors, such as age, stage, grade, HER2 status, treatment, and radiation
(Supplemental Tables 5 and 6).
We evaluated our survival results at the mRNA expression level by examining a
meta-dataset of gene expression TMAs from 855 human breast tumors (38). Kaplan-Meier analyses for relapse-free
survival showed that women with HR3 tumors had the best outcome, HR1 and HR0 tumors
were the most aggressive, and HR2 tumors were intermediate between these groups.
Unlike the IHC-based HR categories, which had significant overall survival
differences (P < 0.0001; Figure 6C), there was a more modest overall relapse-free survival difference
among the mRNA-based HR groups (P = 0.13; Figure 6D). However, lung metastasis relapse-free survival
differences among mRNA-based HR groups were significantly different
(P = 0.0014; Supplemental Figure 7F). Taken together, these data
support a correlation of ER/AR/VDR with tumor differentiation state: more
differentiation correlates with less aggressive behavior. Importantly, these results
suggest that measurement of ER, AR, and VDR protein levels may be more relevant than
Analysis of HR+ and HR– cell types in breast cancer
cell lines. We next examined whether the HR0–HR3 phenotypes are preserved in breast
cancer cell lines. Publicly available mRNA expression data from more than 50 breast
cancer cell lines were analyzed (39), which
revealed that most tumor cell lines fell into one of the normal cell-of-origin
We found that the HR+ pure luminal breast cancer cell lines
(ER/AR/VDR+; n = 16) rarely expressed K5/14/CD10/SMA,
as expected (Supplemental Figure 8A). This was also true for the HER2+
tumor cells lines (n = 13), which rarely expressed K5/14/CD10/SMA,
but were occasionally AR/VDR+, as expected. 7 cell lines had a TNBC
luminal 2 phenotype (BT-20, SUM149, HCC38, HCC-1187, BPLER, HCC-1143, HCC-1500), and
6 cell lines had a TNBC mixed phenotype (MDA-MB-468, HCC1937, HCC-70, HMLER,
HCC-3153, HCC2157) (Supplemental Figure 8A). We also confirmed these phenotypes at
the protein level in breast cancer cell lines, in order to select a subset of lines
that closely conformed to in vivo HR phenotypes (Supplemental Figure 8, B and C).
This set of breast cancer cell lines was then examined for in vitro drug response
Interestingly, 9 cell lines that are frequently used as models of human TNBC
(MDA-MB-231, SUM-159PT, MDA-MB-157, MDA-MB-436, HBL100, BT549, SUM1315M02,
MDA-MB-435, and HS578T) had expression profiles that were not present either in
normal breast cells or in human breast cancers (i.e., negative for most HRs and
epithelial markers; Supplemental Figure 8A). Since this profile is almost never seen
in vivo, either these cells have lost their original phenotype, or they were derived
from very rare tumor types, cautioning against their frequent use.
Response of breast cancer cell lines to HR inhibition. The HR0–HR3 classification of breast cancers not only correlated with
clinically significant outcome groups, but may also provide insights about how the
treatment of these patients might be personalized. For example, we envisioned that
HR3 tumors might be treated using triple-hormone therapy (ER antagonists plus AR and
VDR agonists), and tested some of these concepts in breast cancer cell lines.
There are currently very few effective treatments against TNBCs, because they are
ER– and HER2–. However, because 63% of TNBCs
were AR+, VDR+, or AR+VDR+, hormone
treatment might be possible in a majority of TNBCs, in combination with chemotherapy.
Cell lines BT-20, MDA-MB-468, and SUM159 only expressed VDR, corresponding to the
HR1/TNBC phenotype. We found that combining the VDR agonist calcitriol with taxol
additively inhibited proliferation of these HR1 breast cancer cells more effectively
than either drug alone (Supplemental Figure 9A).
A similar combination treatment strategy can be also employed in ER+ tumor
cells; for example, the HR2 ZR75B cell line was ER+VDR+, and
combining calcitriol with low doses of the ER antagonist ICI182,780 (0.5 nM)
additively inhibited proliferation of these cells (Supplemental Figure 9B). In
another example, the combination of the AR agonist R1881 (50 nM) with calcitriol (50
nM) additively inhibited proliferation of the HR3 breast cancer cell line T47D
(Supplemental Figure 9C).
In HER2+ breast cancer cells, we observed that the combination of the AR
antagonist flutamide (45 μM) and the HER2 inhibitor lapatinib (0.5
μM) additively inhibited proliferation of the HR2/HER2+ cell line
MDA-MB-453 (Supplemental Figure 9D). Similarly, the combination of ICI182,780 (0.5
nM) and lapatinib (10 nM) additively inhibited proliferation of the
HR3/HER2+ cell line BT474 (Supplemental Figure 9E). In control
experiments, no inhibition was observed with the ER antagonist ICI182,780 in HR2
AR+VDR+ (i.e., ER–) MDA-MB-453 cells or
with the VDR agonist calcitriol in HR2 AR+ER+ (i.e.,
VDR–) BT549 cells (Supplemental Figure 9, F and G).
Because nearly 95% of HER2 tumors expressed at least 1 HR, and 29% expressed all 3
HRs, these results indicate that hormone treatment might also be possible in a
majority of HER2 tumors in combination with anti-HER2 therapy.
In the present study, we analyzed more than 15,000 normal breast cells and described 11
previously undefined cell subtypes in the luminal layer of human breast, L1–L11.
These 11 normal breast cell types conformed to 4 novel hormonal differentiation groups,
HR0–HR3. Analysis of 3,157 human breast tumors revealed that each tumor was
similar to 1 of the 11 normal cell types and HR groups. Almost none of the breast
cancers exhibited a pure basal-like phenotype, as defined by the expression of true
myoepithelial markers and the absence of any luminal markers. Nearly all human breast
tumors replicated one of the normal hormonal differentiation groups (HR0–HR3),
and we found that these subgroups exhibited significant survival differences. Our
ontological classification scheme provides actionable hormonal treatment strategies for
all subtypes of human breast cancer.
Taxonomy dilemma: lumpers versus splitters. Historically, the challenges of taxonomy have led to 2 opposing taxonomic approaches:
lumping and splitting. Lumpers prefer a few categories marked by
large similarities that have clear practical utility; splitters tend
to describe as many entities as possible with small differences, regardless of
practical utility. Charles Darwin has been credited with using the terms first, when
he wrote in a letter, “Those who make many species are the
‘splitters,’ and those who make few are the
‘lumpers’” (40). In
medicine, this divide is exacerbated when a clear mechanistic understanding of a
disease entity is incomplete.
High-throughput technologies — such as mRNA, miRNA, SNP, and epigenetic
profiling as well as next-generation sequencing — have unveiled a complex
heterogeneity of breast tumors in individual patients (9, 10, 22, 41, 42). This has led some to advocate designation of
categories based on mutations and genetic alterations (which, at its logical extreme,
would place each patient into his own unique category), even as the difficulties of
such an approach are increasingly becoming evident (43). As we currently stand, without a clear intellectual underpinning of
the origins of breast tumors, a consensus has yet to emerge regarding how many types
of breast cancers there are, or should be, and how they should be appropriately
lumped or split.
While -omics approaches have tremendous molecular resolution, at the
anatomic level, they are hindered by several limitations (21, 22). To begin, -omics
approaches have low morphologic resolution. In most cases, RNA/DNA is isolated from
mm3–cm3 tissue fragments containing tumor cells with
heterogeneous molecular characteristics, admixed with various normal epithelial,
stromal, vascular, and inflammatory cells. An additional concern is the
self-referential nature of many such tumor datasets, as they are predominantly
focused on comparing tumors among each other, not against normal cell counterparts.
Finally, although cataloging mutations remains an excellent method by which to
identify genomic differences between tumors, it is ill-suited for finding
similarities between tumors and normal tissues (10, 22, 41, 42). These
shortcomings result in a loss of spatial, architectural, and tumor lineage
An alternative method of classifying tumors is an ontological approach, which focuses
on defining tumor subtypes based on their similarities with specific normal cell
origin subtype, akin to evolutionary biology, in which subspecies are identified
based on the degree of similarities to common ancestors. Immunostains are currently
the most powerful means by which to define similarities between tumors and specific
normal cell types, and they resolve many of the issues raised regarding the -omics
approaches. As in situ methods preserve tissue architecture, they provide high
morphologic resolution at the subcellular level. Additionally, the normal cells
provide an unchanging gold-standard internal control and provide a means to describe
similarities between normal and tumor cells that is simply not possible using
mutational analysis. Lastly, one of the most important differences between the
unsupervised clustering (-omics) and the normal cell type–based (ontological)
tumor classification methods is that the latter is hypothesis based, which can be
tested and refuted or confirmed.
All of the markers used in this study have been examined previously by other
investigators individually. However, to our knowledge, this is the first time they
have been examined all together within the framework of a comprehensive taxonomic
assessment and, moreover, evaluated within the same tissue sections. The latter was
made possible by use of the novel matrix IF technology that was recently
It is important to note that this is just the beginning of efforts to identify
markers that define the cells comprising the luminal layer of the breast epithelium.
It took several decades to accurately describe all the relevant cell types in the
hematopoietic system, with the aid of an ever-increasing number of cell
type–specific markers. Likewise, in the case of human breast, additional
markers will undoubtedly refine and expand the classification system we propose
Our results indicate that the composition of normal breast epithelium is much more
complex than previously appreciated. While one might ponder the need for 11 cellular
subtypes within the normal breast luminal epithelium, it is well worth considering
the evolutionary importance of mammary gland development. This seminal event required
emergence of a diverse range of cell types that together produce milk, an extremely
complex substance essential for the survival of all mammals.
Potential diagnostic, prognostic, and treatment implications. Admittedly, the HR0–HR3 categories describe only one aspect of a disease as
complex as breast cancer, and we do not propose using this classification alone or in
place of existing approaches. Rather, we imagine that the HR0–HR3 categories
can be used to refine the ER+, HER2+, and TNBC classification
presently in use. In the future, this approach can be improved by more detailed
descriptions of normal cell types in human breast and by incorporating this
information into clinical evaluation along with integrated molecular genetic
Nevertheless, even at this early stage, we believe this cell-based classification
approach has produced some actionable insights. One of the most intriguing
implications of our work is the possibility of expanding the patient population being
treated with hormone therapy, by targeting AR and VDR in conjunction with ER (44).
Genotype, cell of origin, and tumor phenotype. The outcome differences of tumors that arise from the same organ can be associated
with genetic differences, presenting an attractive paradigm by which to guide the
design of personalized cancer therapeutics (45).
An emerging and complementary hypothesis is that phenotypic differences among
distinct subtypes of tumors arising in a single tissue may also be imposed by
cell-autonomous factors unique to the cell of origin (16, 17, 46). For example, we previously demonstrated that while some
normal cell types gave rise to highly tumorigenic and metastatic adenocarcinomas,
other breast cell types that were isolated from the same patient and transformed with
identical oncogenes gave rise to cells that were morphologically distinct, weakly
tumorigenic, and nonmetastatic (16, 47–49). Others have also suggested that the same oncogenes can have vastly
different phenotypic consequences depending on the cell of origin (50, 51).
In light of these observations, we contemplate whether the better outcome of
HR+ breast tumors may be due to their cell of origin and
differentiation lineage. As mutations or amplifications of ER, AR, or VDR genes are
very rare in breast cancer, it is likely that HR+ tumors have high HR
expression because they either arose in an already HR+ normal cell, or
they arose in a HR– precursor that was preordained to
differentiate into a HR+ phenotype.
Ontological taxonomy of tumors. In summary, the use of in situ stains and normal cells as a reference point for
classifying tumors solves several of the issues that were raised for -omics
approaches: (a) tissue architecture is preserved, (b) morphologic resolution is high,
(c) normal cells provide an unchanging gold-standard reference, and (d) phenotypic
similarity between normal and tumor cells is maintained. In addition, using normal
cell subtypes as a reference point in tumor classification addresses the question of
lumping versus splitting. Each normal tissue is designed for a specific function, and
each cell subtype is designed to perform different components of this function. Since
these functions are finite, the maximum number of biologically important normal cell
types is limited, unchanging, and able to be precisely defined. Thus, this method
constrains the arbitrary splitting of tumors into endless subclasses. This
ontological approach provides a durable infrastructure and context within which
molecular data may be appropriately placed in the right cellular context and
At first, the molecular heterogeneity of breast cancer appears difficult to reconcile
with the robust phenotypic subtypes we observed. However, it is the signaling
pathways, not the individual genes, that are responsible for the phenotypes of tumors
(43). Thus, a deeper pathway-based
understanding will be necessary to correlate the cell type–based subtypes and
molecular heterogeneity of breast tumors.
We perceive the cell type–based ontological and high-throughput molecular
-omics approaches as complimentary methods. Current in situ examination methods, such
as IF and IHC, have several shortcomings: they are low-throughput and
semiquantitative and do not allow examination of thousands of proteins
simultaneously. As mentioned above, the high-throughput molecular approaches have low
morphologic resolution. An ideal approach in the future should combine the
subcellular resolution of immunostains with the power of high-content, multigene
molecular approaches for studying tumors. We believe that such molecular imaging
technologies represent the next frontier in cancer research, and the multiplex IF
technique used herein is just one example of how such technologies may advance our
understanding of cancer.
Cell culture. For cell propagation and studying effects of ER and VDR modulators on cell
proliferation, see Supplemental Methods.
IHC, IF, and NHS analysis. The NHS is a prospective cohort study initiated in 1976 (52). For antibodies and conditions of FFPE section staining, see
Supplemental Table 1. For image acquisition and processing; sample scoring and data
display in heat maps; and study design, population, and analysis, see Supplemental
Multiplex IF. For antibodies and conditions of FFPE section staining, see Supplemental Table 1. For
sequential quenching of fluorescence signal, data acquisition, and multiplex IF image
analysis, see Supplemental Methods.
Gene expression analysis and cell lines. For analysis of mRNA expression profiling data of cancer cell lines (39), see Supplemental Methods.
Cell proliferation assays. For analysis of cell line proliferation responses to calcitriol, taxol, ICI182,780,
and R1881, see Supplemental Methods.
Statistics. Event outcomes were compared using Kaplan-Meier analysis, and P
values were determined with the log-rank test. Tumor expression of myoepithelial and
luminal genes (26–28) were explored in frozen robust multiarray
analysis–normalized (53) gene
expression data (GEO accession nos. GSE3744, GSE4922, GSE6532, and GSE7390) using
hierarchical cluster analysis (Pearson r, average linkage) using the
Bioconductor package MADE4 (54). Global test,
available from the Bioconductor package globaltest (55), was used to assess the association between gene expression and
luminal or myoepithelial classification. Tumors were classified as basal-like or
non-basal-like as previously described (56),
except where basal-like subtype classification was provided by the authors (29). Global test (55) was used to determine association between individual genes and
basal-like/non-basal-like division. A P value less than 0.05 was
considered significant. For statistical methods for NHS protein expression and
outcome data analysis, UNC mRNA expression and outcome data analysis, and additional
details, see Supplemental Methods.
Study approval. FFPE blocks from surgical resection specimens of normal breast tissue and of breast
tumors were obtained from the archives of Brigham and Women’s Hospital (BWH)
in accordance with the regulations for excess tissue use stipulated by the BWH IRB.
For TMAs used in this study, see Supplemental Methods.
High-resolution images. High-resolution files corresponding to all image panels in Figures 1–3,
Supplemental Figures 1–3 and 5, and Supplemental Table 2 are available at
View Supplemental data
View Supplemental table
The authors acknowledge funding support from Breast Cancer Research Foundation and Play
for P.I.N.K. (to T.A. Ince), Kristin Jordahl and friends (to T.A. Ince); NCI grant
R01-CA146445-01, NIH Roadmap Epigenomics Project (to T.A. Ince); Department of Defense
CDMRP BCRP grant W8 1XWH-08-1-0282,BC-07456 (to T.A. Ince); DHHS and GlaxoSmithKline
grants WE234 and EPI40307 (to R.M. Tamimi); NIH grant K08NS064168 (to S. Santagata); V
Foundation for Cancer Research (to S. Santagata); DFCI Women’s Cancers Program
(to A.C. Culhane and M. Schwede); and Claudia Adams Barr Program in Innovative Basic
Cancer Research (to A.C. Culhane). We thank Q. Li, S. Dinn, A. Santamaria-Pang, and F.
Ginty (GE Global Research Center) as well as T.D. Tlsty, J.S. Brugge, J.M. Slingerland,
R.D. Cardiff, C. Gomez, M.D. Pegram, S.A. Borowsky, M. Nadji, and M. Jorda for helpful
Conflict of interest: Sandro Santagata is a cofounder of and scientific
advisor to Bayesian Diagnostics. Tan A. Ince was a scientific advisor to 3DM Inc.
(2007–2012) and a consultant to Stemgent (2008–2010).
Citation for this article:J Clin Invest. 2014;124(2):859–870. doi:10.1172/JCI70941.
See the related Commentary beginning on page 478.
Swerdlow SH, et al.WHO Classification of Tumours of Haematopoietic and Lymphoid Tissue.4th ed. Lyon, France: World Health Organization; 2008.
Hamilton A, Gallipoli P, Nicholson E, Holyoake TL. Targeted therapy in haematological malignancies. J Pathol. 2010;220(4):404–418.
Wood GS, Warnke RA. The immunologic phenotyping of bone marrow biopsies and aspirates:
frozen section techniques. Blood. 1982;59(5):913–922.
Drexler HG. Classification of acute myeloid leukemias--a comparison of FAB and
immunophenotyping. Leukemia. 1987;1(10):697–705.
Jones C, et al. Expression profiling of purified normal human luminal and
myoepithelial breast cells: identification of novel prognostic markers for breast
cancer. Cancer Res. 2004;64(9):3037–3045.
Gusterson BA, Ross DT, Heath VJ, Stein T. Basal cytokeratins and their relationship to the cellular origin and
functional classification of breast cancer. Breast Cancer Res. 2005;7(4):143–148.
Roy S, et al. Rare somatic cells from human breast tissue exhibit extensive lineage
plasticity. Proc Natl Acad Sci U S A. 2013;110(12):4598–4603.
Prat A, Perou CM. Deconstructing the molecular portraits of breast
cancer. Mol Oncol. 2010;5(1):5–23.
Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast
tumours. Nature. 2012;490(7418):61–70.
Curtis C, et al. The genomic and transcriptomic architecture of 2,000 breast tumours
reveals novel subgroups. Nature. 2012;486(7403):346–352.
Dawson SJ, Rueda OM, Aparicio S, Caldas C. A new genome-driven integrated classification of breast cancer and its
implications. EMBO J. 2013;32(5):617–628.
Prat A, Ellis MJ, Perou CM. Practical implications of gene-expression-based assays for breast
oncologists. Nat Rev Clin Oncol. 2012;9(1):48–57.
Guiu S, et al. Molecular subclasses of breast cancer: how do we define them? The
IMPAKT 2012 Working Group Statement. Ann Oncol. 2012;23(12):2997–3006.
Schnitt SJ. Classification and prognosis of invasive breast cancer: from
morphology to molecular taxonomy. Mod Pathol. 2010;23(suppl 2):S60–S64.
Chu PG, Weiss LM. Keratin expression in human tissues and neoplasms. Histopathology. 2002;40(5):403–439.
Ince TA, et al. Transformation of different human breast epithelial cell types leads
to distinct tumor phenotypes. Cancer Cell. 2007;12(2):160–170.
Molyneux G, et al. BRCA1 basal-like breast cancers originate from luminal epithelial
progenitors and not from basal stem cells. Cell Stem Cell. 2010;7(3):403–417.
Ginty F, et al. The relative distribution of membranous and cytoplasmic met is a
prognostic indicator in stage I and II colon cancer. Clin Cancer Res. 2008;14(12):3814–3822.
Collins LC, et al. Basal cytokeratin and epidermal growth factor receptor expression are
not predictive of BRCA1 mutation status in women with triple-negative breast
cancers. Am J Surg Pathol. 2009;33(7):1093–1097.
Boecker W, Buerger H. Evidence of progenitor cells of glandular and myoepithelial cell
lineages in the human adult female breast epithelium: a new progenitor (adult
stem) cell concept. Cell Prolif. 2003;36(suppl 1):73–84.
Molyneux G, Smalley MJ. The cell of origin of BRCA1 mutation-associated breast cancer: a
cautionary tale of gene expression profiling. J Mammary Gland Biol Neoplasia. 2011;16(1):51–55.
Lavasani MA, Moinfar F. Molecular classification of breast carcinomas with particular emphasis
on “basal-like” carcinoma: a critical review. J Biophotonics. 2012;5(4):345–366.
Livasy CA, et al. Phenotypic evaluation of the basal-like subtype of invasive breast
carcinoma. Mod Pathol. 2006;19(2):264–271.
Sorlie T, et al. Gene expression patterns of breast carcinomas distinguish tumor
subclasses with clinical implications. Proc Natl Acad Sci U S A. 2001;98(19):10869–10874.
Sorlie T, et al. Repeated observation of breast tumor subtypes in independent gene
expression data sets. Proc Natl Acad Sci U S A. 2003;100(14):8418–8423.
Grigoriadis A, et al. Establishment of the epithelial-specific transcriptome of normal and
malignant human breast cells based on MPSS and array expression
data. Breast Cancer Res. 2006;8(5):R56.
Lakhani SR, et al. Prediction of BRCA1 status in patients with breast cancer using
estrogen receptor and basal phenotype. Clin Cancer Res. 2005;11(14):5175–5180.
Raouf A, et al. Transcriptome analysis of the normal human mammary cell commitment and
differentiation process. Cell Stem Cell. 2008;3(1):109–118.
Richardson AL, et al. X chromosomal abnormalities in basal-like human breast
cancer. Cancer Cell. 2006;9(2):121–132.
Loi S, et al. Predicting prognosis using molecular profiling in estrogen
receptor-positive breast cancer treated with tamoxifen. BMC Genomics. 2008;9:239.
Ivshina AV, et al. Genetic reclassification of histologic grade delineates new clinical
subtypes of breast cancer. Cancer Res. 2006;66(21):10292–10301.
Desmedt C, et al. Strong time dependence of the 76-gene prognostic signature for
node-negative breast cancer patients in the TRANSBIG multicenter independent
validation series. Clin Cancer Res. 2007;13(11):3207–3214.
Gusterson B. Do ‘basal-like’ breast cancers really
exist? Nat Rev Cancer. 2009;9(2):128–134.
Lim E, et al. Aberrant luminal progenitors as the candidate target population for
basal tumor development in BRCA1 mutation carriers. Nat Med. 2009;15(8):907–913.
Collins LC, Cole KS, Marotti JD, Hu R, Schnitt SJ, Tamimi RM. Androgen receptor expression in breast cancer in relation to molecular
phenotype: results from the Nurses’ Health Study. Mod Pathol. 2011;24(7):924–931.
Hu R, et al. Androgen receptor expression and breast cancer survival in
postmenopausal women. Clin Cancer Res. 2011;17(7):1867–1874.
Santagata S, et al. High levels of nuclear heat-shock factor 1 (HSF1) are associated with
poor prognosis in breast cancer. Proc Natl Acad Sci U S A. 2011;108(45):18378–18383.
Harrell JC, et al. Genomic analysis identifies unique signatures predictive of brain,
lung, and liver relapse. Breast Cancer Res Treat. 2012;132(2):523–535.
Neve RM, et al. A collection of breast cancer cell lines for the study of functionally
distinct cancer subtypes. Cancer Cell. 2006;10(6):515–527.
Darwin C, Darwin F. The life and letters of Charles Darwin: Including an autobiographical
chapter.Vol. 1. New York, NY: D. Appleton and Co.; 1911.
Banerji S, et al. Sequence analysis of mutations and translocations across breast cancer
subtypes. Nature. 2012;486(7403):405–409.
Stephens PJ, et al. Complex landscapes of somatic rearrangement in human breast cancer
genomes. Nature. 2009;462(7276):1005–1010.
Yaffe MB. The scientific drunk and the lamppost: massive sequencing efforts in
cancer discovery and treatment. Sci Signal. 2013;6(269):pe13.
Ni M, et al. Targeting androgen receptor in estrogen receptor-negative breast
cancer. Cancer Cell. 2011;20(1):119–131.
MacConaill LE, et al. Profiling critical cancer gene mutations in clinical tumor
samples. PLoS One. 2009;4(11):e7887.
Yalcin-Ozuysal O, Brisken C. From normal cell types to malignant phenotypes. Breast Cancer Res. 2009;11(6):306.
Merritt MA, t al. Gene expression signature of normal cell-of-origin predicts ovarian
tumor outcomes. PLoS One. 2013;8(11):e80314.
Godar S, et al. Growth-inhibitory and tumor- suppressive functions of p53 depend on
its repression of CD44 expression. Cell. 2008;134(1):62–73.
McAllister SS, et al. Systemic endocrine instigation of indolent tumor growth requires
osteopontin. Cell. 2008;133(6):994–1005.
Vogelstein B, Kinzler KW. Cancer genes and the pathways they control. Nat Med. 2004;10(8):789–799.
Gupta GP, Massague J. Cancer metastasis: building a framework. Cell. 2006;127(4):679–695.
Tamimi RM, et al. Comparison of molecular phenotypes of ductal carcinoma in situ and
invasive breast cancer. Breast Cancer Res. 2008;10(4):R67.
McCall MN, Bolstad BM, Irizarry RA. Frozen robust multi-array analysis (fRMA). Biostatistics. 2010;11(2):242–253.
Culhane AC, Thioulouse J, Perriere G, Higgins DG. MADE4: an R package for multivariate analysis of gene expression
data. Bioinformatics. 2005;21(11):2789–2790.
Goeman JJ, van de Geer SA, de Kort F, van Houwelingen HC. A global test for groups of genes: testing association with a clinical
outcome. Bioinformatics. 2004;20(1):93–99.
Culhane AC, Quackenbush J. Confounding effects in “A six-gene signature predicting breast
cancer lung metastasis”. Cancer Res. 2009;69(18):7480–7485.