Abstract
Gene function in cancer is often cell type-specific. The epithelial
cell-specific transcription factor ELF3 is a documented tumor
suppressor in many epithelial tumors yet displays oncogenic properties
in others. Here, we show that ELF3 is an oncogene in the adenocarcinoma
subtype of lung cancer (LUAD), providing genetic, functional, and
clinical evidence of subtype specificity. We discover a region of focal
amplification at chromosome 1q32.1 encompassing the ELF3 locus in LUAD
which is absent in the squamous subtype. Gene dosage and promoter
hypomethylation affect the locus in up to 80% of LUAD analyzed. ELF3
expression was required for tumor growth and a pan-cancer expression
network analysis supports its subtype and tissue specificity. We
further show that ELF3 displays strong prognostic value in LUAD but not
LUSC. We conclude that, contrary to many other tumors of epithelial
origin, ELF3 is an oncogene and putative therapeutic target in LUAD.
Subject terms: Cancer, Cancer genetics, Cancer genomics, Cancer models,
Lung cancer
__________________________________________________________________
Tissue context can dictate why a gene can have seemingly opposing
functions in different settings. ELF3 is tumor suppressive in many
cancers of epithelial origin but in lung cancer, the authors describe
an oncogenic role in the adenocarcinoma histology of non-small cell
lung cancer.
Introduction
There is increasing evidence of transcription factors exhibiting both
tumor suppressive and oncogenic functions across cancers despite a
common epithelial origin. For example, in lung cancer, the
transcription factor Nuclear Factor I B (NFIB) has both tumor
suppressive^[64]1 and oncogenic behaviors depending on the histological
subtype^[65]2,[66]3. Such dual functions are often attributed to
DNA-level events specific to tumors arising from distinct cell types
and lineages. Deleterious mutations in the epithelial-specific E74 Like
ETS Transcription Factor 3 (ELF3) have been described as tumourigenic
in many epithelial tumors;^[67]4–[68]12 in contrast, gene amplification
and oncogenic activity has also been reported in others^[69]13,[70]14.
In lung cancer, the evidence suggests an oncogenic function as
overexpression has been reported in non-small cell lung cancer (NSCLC)
and a mouse lung tumourigenesis model^[71]15–[72]18. Furthermore, the
ELF3 locus is located within a region of recurrent DNA-level gain in
non-small cell lung cancer on chromosome 1q32.1, suggesting a genetic
mechanism of selection^[73]19–[74]21. Contrary to many other
recurrently gained or amplified regions in lung cancer, the gene target
or targets of chromosome 1q gain remain elusive.
In this study, we have comprehensively analyzed 1835 human clinical
samples of NSCLC and identify a disparate ELF3 expression pattern in
the adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC)
histological subtypes. This overexpression is driven not only by a
lineage-specific focal amplification event, but also via alternate
DNA-level mechanisms, detected collectively in ~ 80% of LUAD and which
hold prognostic value uniquely in LUAD. Functional experiments in cell
and animal models confirm the oncogenic behavior of ELF3 in LUAD. We
further discover ELF3 signaling networks to be highly tissue-specific,
accounting for its conflicting functions across epithelial cancers. We
demonstrate the duality of the ELF3 transcription factor and solidify
its role as an oncogene in LUAD.
Results
ELF3 is frequently overexpressed in lung adenocarcinoma
To investigate the oncogenic potential of ELF3 in NSCLC, we searched
for recurrent DNA-level alterations that could indicate selection. ELF3
mutations were uncommon in NSCLC, with a 1.4% mutation incidence in The
Cancer Genome Atlas (TCGA) cohort (n = 408) (Supplementary Fig. [75]1a)
and 1.0% in COSMIC v84 (n = 1302). We then searched for evidence of
focal amplification at the ELF3 locus and uncovered a narrow peak at
1q32.1 that encompassed ELF3, providing the first evidence of oncogenic
selection. Interestingly, this focal amplification event was specific
to LUAD histology, corroborating previous reports of lineage-specific
frequent 1q gain in LUAD^[76]19–[77]22 (Fig. [78]1a). This subtype
specificity held true at the expression level, as ELF3 expression was
significantly higher in LUAD compared with LUSC in three independent
cohorts (Fig. [79]1b, Supplementary Table [80]1). Immunohistochemistry
data also showed elevated ELF3 protein expression in LUAD compared with
LUSC (n = 236) (Fig. [81]1f, g, Supplementary Fig. [82]2). Similarly,
ELF3 expression was significantly positively correlated with LUAD
lineage markers and negatively correlated with LUSC markers
(Supplementary Table [83]2). When compared with adjacent non-malignant
lung tissue, ELF3 was significantly overexpressed in LUAD in the BC
Cancer Agency (BCCA) (n = 83 pairs, Wilcoxon sign-rank p = 1.64E-21)
and TCGA data sets (n = 571, Mann–Whitney U test p = 1.54E-07)
(Fig. [84]1c), but not differentially expressed in LUSC (Supplementary
Fig. [85]1b). Furthermore, TCGA LUAD samples with higher tumor cell
purity exhibited higher ELF3 expression, consistent with overexpression
in tumor cells (Fig. [86]1d). In a pairwise analysis, greater than
twofold overexpression was detected in 73 and 40% of BCCA and TCGA
data, respectively (Fig. [87]1e).
Fig. 1.
[88]Fig. 1
[89]Open in a new tab
ELF3 is located within a region of focal amplification in lung
adenocarcinoma. a Comparison of significantly focally amplified regions
on chromosome 1q in LUAD and LUSC. b Comparison of ELF3 expression
between LUAD (red) and LUSC (green) in TCGA (n = 1017), Samsung Medical
Centre (n = 138, [90]GSE8894), and Duke University data sets (n = 111,
[91]GSE3141) by Mann–Whitney U test. c Box and whiskers plots of log[2]
ELF3 expression in 83 paired cases of non-malignant lung (blue) and
lung adenocarcinoma (red) in the BCCA cohort (Wilcoxon sign-rank test),
and unpaired cases of 58 non-malignant lung and 513 lung adenocarcinoma
samples in the TCGA cohort (Mann–Whitney U test). d Mean log[2] ELF3
expression in non-malignant lung (blue) and in LUAD (red) grouped by
purity estimate in the TCGA cohort^[92]55. e Distribution of ELF3
expression fold change in LUAD compared with paired non-malignant lung
in the BCCA and TCGA cohort. f Representative ELF3 immunohistochemistry
images from a tissue microarray of LUAD (left panel) and LUSC (right
panel) (scale bar = 100 µm). g Box and whiskers plots of ELF3
immunohistochemistry (IHC) score as determined by pathologist review
(n = 236, Mann–Whitney U test). Score was calculated by multiplying the
percent of positive cells by the stain intensity (+1, +2, +3). In all
box and whisker plots, center line represents the median, box bounds
indicate the 25th and 75th percentiles, and whiskers extend from
minimum to maximum. LUAD = lung adenocarcinoma, LUSC = lung squamous
cell carcinoma, TMA = tissue microarray.
Genetic and epigenetic determinants of ELF3 overexpression
Focal ELF3 amplification was detected in 14% of LUAD, and therefore
could not explain the observed frequency of overexpression. Further
investigations into broad copy number alterations and local promoter
methylation changes that could explain additional cases of
overexpression revealed disruption of the ELF3 locus in ~ 80% of LUAD
(BCCA = 79%, TCGA-60 = 83%) (Fig. [93]2a). ELF3 expression correlated
strongly with methylation of CpG probe cg12970084, as well as with gene
dosage. Expression was significantly increased in tumors with genetic
or epigenetic disruption compared to those without (Mann–Whitney U test
p = 5.17E-07) (Fig. [94]2b–d, Supplementary Figs. [95]3, [96]4). A
previous study identified SMAD4 as a direct transcriptional repressor
of ELF3, and ERBB2 signaling as an activator of ELF3 expression^[97]17.
Given that ERBB2 is a known oncogenic driver of LUAD, we assessed the
influence of inactivating SMAD4 and activating ERBB2 mutations on ELF3
expression in tumors lacking locus disruption and find that ELF3
expression was elevated in these samples (Mann–Whitney U test p = 0.06)
(Supplementary Fig. [98]5). Therefore, ELF3 overexpression frequently
observed in LUAD is explained predominantly by direct ELF3 locus
alterations and by mutations in upstream regulators.
Fig. 2.
[99]Fig. 2
[100]Open in a new tab
Genetic and epigenetic alterations at the ELF3 locus. a Case-by-case
representation of ELF3 locus deregulation in the BCCA data set (n = 83
pairs) by genetic and epigenetic events. The frequencies of copy number
gain or amplification, and promoter hypomethylation are indicated, as
well as the cumulative frequency of ELF3 DNA-level alterations in the
BCCA data set. These frequencies are also summarized to the right of
the plot for the TCGA (n = 420) and TCGA-60 (n = 252) data sets (TCGA
cases with ≥ 60% tumor cellularity, see Methods). b Scatter plot of
ELF3 promoter methylation (x axis) and log[2] ELF3 expression (y axis)
across 452 LUAD (gray) and 21 non-malignant lung (blue) samples
(Spearman’s correlation). c Log[2] ELF3 expression as a function of DNA
copy number alteration in LUAD (n = 420). ELF3 expression is increased
with activating events (Mann–Whitney U test). d Comparison of log[2]
ELF3 expression between LUAD with gain and/or promoter hypomethylation
(blue) compared with those without (black) (n = 420, Mann–Whitney U
test). In all box and whisker plots, center line represents the median,
box bounds indicate the 25th and 75th percentiles, and whiskers extend
from minimum to maximum. LUAD = lung adenocarcinoma,
AMP = amplification.
ELF3 disruption occurs across molecular subtypes
As molecular cancer subtype information is valuable in guiding clinical
management, we examined the effect of driver mutation status on ELF3
expression and searched for patterns of co-occurrence or mutual
exclusivity at the DNA-level. High ELF3 expression occurred
irrespective of KRAS or EGFR status in three independent cohorts
(Fig. [101]3a). Similarly, we observed no association of ELF3
expression with KRAS, EGFR, or ALK in a panel of 25 LUAD cell lines
(Supplementary Fig. [102]6). Mutations in known LUAD drivers^[103]23
(KRAS, EGFR, BRAF, RIT1, ERBB2, MAP2K1, NRAS, HRAS, ERBB2
amplification, MET amplification) and tumor suppressors (TP53, KEAP1,
STK11/LKB1, NF1) occurred in both tumors with ELF3 locus alterations
and those without (Fig. [104]3b). There was a statistical enrichment of
KRAS (Fisher’s exact p = 0.0011), STK11 (p = 0.0023), and KEAP1
(p = 0.0031) mutations in the ELF3-altered group, whereas MET and ERBB2
mutations and amplifications were enriched in cases without ELF3
disruption (p = 0.0009 and p = 0.0019, respectively). This is in
agreement with previous associations of KRAS mutation and chromosome 1q
gain, and correlations between KRAS, STK11, and KEAP1
mutations^[105]24,[106]25. Overall, we conclude that ELF3 holds
biological relevance across LUAD molecular subtypes and proceed to
pursue the oncogenic function of ELF3 function in representative cell
models.
Fig. 3.
[107]Fig. 3
[108]Open in a new tab
High ELF3 expression is not dependent on the molecular subtype of lung
adenocarcinoma. a Comparison of log[2] ELF3 expression and ELF3
immunohistochemistry (IHC) score between cases with or without driver
mutations in EGFR and KRAS (BCCA n = 83, TCGA n = 484, Tissue
Microarray n = 161, Mann–Whitney U test). In all box and whisker plots,
center line represents the median, box bounds indicate the 25th and
75th percentiles, and whiskers extend from minimum to maximum. b
Case-by-case representation of clinical features (upper panel) and
genomic alterations (lower panel) across 420 LUAD from the TCGA data
set dichotomized into those with and without DNA-level alterations at
the ELF3 locus. Mutations and amplifications in prominent oncogenes and
mutations in tumor suppressor genes are displayed.
ELF3 regulates cancer phenotypes of lung cell lines
ELF3 expression was stably inhibited by lentiviral-mediated delivery of
five shRNA vectors in the LUAD cell line, HCC827, which exhibits high
ELF3 expression. The optimal two shRNAs were identified, and cell
viability was assessed by MTT assay. ELF3 knockdown (shELF3) cell lines
demonstrated significantly reduced viability as compared with their
isogenic empty vector controls (Supplementary Fig. [109]7a, b). With
this preliminary observation that ELF3 knockdown reduces oncogenic
phenotypes, we expanded our experiments to include four additional cell
lines harboring diverse molecular drivers (A549, H1395, H1993, and
H1819 (Supplementary Fig. [110]7c, d)). For subsequent molecular
assays, the shRNA with the highest degree of knockdown was selected and
compared with empty vector control (HCC827 shRNA-1, A549 shRNA-1, H1395
shRNA-5, H1993 shRNA-5, H1819 shRNA-5). Four of five lines with ELF3
knockdown (shELF3) exhibited reduced proliferation as measured by BrdU
incorporation assay, and all evaluable lines demonstrated a
significantly reduced ability to form colonies in soft agar
(Fig. [111]4a–c). To account for potential off-target effects, we
validated the effect of ELF3 knockdown on cell viability using an siRNA
pool (five siRNAs) on A549 cells and further tested specificity by
including a cell line with low ELF3 expression (H2030). This
observation was specific to LUAD cell lines with high ELF3 expression,
as viability of A549 cells was significantly decreased while H2030
cells were not affected by ELF3 siRNA-mediated inhibition
(Supplementary Fig. [112]7f, g).
Fig. 4.
[113]Fig. 4
[114]Open in a new tab
Manipulation of ELF3 expression regulates oncogenic phenotypes.
Histogram summarizing the effect of shRNA-mediated ELF3 inhibition on
a, b soft agar colony formation and c cell proliferation in lung
adenocarcinoma cell lines, and d the effect of forced ELF3
overexpression (OE) on cell proliferation in HBEC-KT cell lines with
and without p53 knockdown (HBEC-KT53) and/or induction of KRAS
(HBEC-KTR, HBEC-KTR53). e Tumor growth of isogenic HCC827 shELF3 and
control cells in NOD-SCID mice (n = 12). f Quantification of shELF3 and
control vector DNA copy number in endpoint tumors compared with input
material. g Putative model of clonal drift throughout xenograft tumor
growth. At Day 0 we assume a mixed population of cells with low (blue),
medium (pink), and high (orange) vector copy numbers (CN) for both
populations. Over time, the relative proportion of low shELF3 vector CN
clones dominates owing to the growth advantage provided by ELF3
expression. h Tumor growth curve of clonal populations of isogenic A549
shELF3 and control cells in NOD-SCID mice (n = 24). i ELF3 expression
of input clones as measured by immunoblot. j Control and absent shELF3
tumors at endpoint. k Cell viability as measured by annexin/PI staining
of control (black bar) and clonal shELF3 (blue bar) LUAD cell lines
cultured in complete (10% FBS, solid fill) or serum starved (0% FBS,
hatched fill) media (H1993 n = 1 replicates owing to viability issues).
l qPCR of ELF3 mRNA expression in control and shELF3 clonal cell lines
with forced ELF3-OE or empty vector control (lacZ). Histograms
summarizing the effect of ELF3 overexpression on colony formation, in
terms of number m and size n of colonies formed (paired two-tailed
Student’s t test). o Cell proliferation shows rescued growth rate with
ELF3-OE in A549 cells (Wilcoxon test). p Tumor growth of shELF3.cl-lacZ
and shELF3.cl-OE A549 cells in NRG mice (n = 8). In all box and whisker
plots, center line represents the median, box bounds indicate the 25th
and 75th percentiles, and whiskers extend from minimum to maximum, all
histograms display the mean + SEM (n = 3 biological replicates unless
otherwise stated). CN = copy number, +p < 0.10, *p < 0.05, **p < 0.01,
***p < 0.001, ****p < 0.0001, paired two-tailed Student’s t test.
To examine the impact of ELF3 overexpression in a non-malignant
setting, ELF3 was stably overexpressed (ELF3-OE) in immortalized but
untransformed human bronchial epithelial cells (HBEC-KT), which do not
normally express ELF3, and compared to cells transformed with an empty
vector control (Supplementary Fig. [115]7e). Consistent with shELF3,
ELF3-OE cells displayed significantly increased proliferation
(Fig. [116]4d), but overexpression alone was insufficient to transform
cells in soft agar colony formation assays. As cell transformation
requires multiple oncogenic manipulations, we assessed the ability of
ELF3 to transform HBEC-KTs in the background of additional molecular
alterations. We stably overexpressed ELF3 in HBEC-KTs with p53
knockdown (HBEC-KT53), with oncogenic RAS expression (HBEC-KTR), and
with both alterations (HBEC-KTR53), again comparing to empty vector
control^[117]26. We first assessed proliferation as above and found
ELF3 overexpression to significantly decrease the proliferative
capacity of HBEC-KT53 cells, whereas no significant difference was
observed in HBEC-KTR or HBEC-KTR53 cells (Fig. [118]4d). We then
assessed cell transformation by colony formation assay. No sizable
colonies were detected in any condition following three biological
replicates, leading us to conclude ELF3 is unable to transform HBEC-KT
cells with these specific molecular alterations in vitro. It is evident
that ELF3 is capable of regulating oncogenic phenotypes; however, these
results indicate the functional importance of ELF3 may arise during a
later stage of tumor development or on a background of further genetic
alterations.
Clonal elimination of ELF3 abolishes tumor growth
The effect of ELF3 inhibition on tumor growth was examined in vivo
using a xenograft model of HCC827 shELF3 (shRNA-1) cells and isogenic
controls (n = 12 NOD-SCID mice). Although growth of polyclonal shELF3
tumors was initially slowed, it eventually reached control levels
(Fig. [119]4e). Analysis of endpoint tumors showed that knockdown was
not maintained (Supplementary Fig. [120]8a), and clones within the
shELF3 tumors that had fewer copies of the shELF3 vector (and higher
ELF3 expression) out-competed clones with increased copies of shELF3
vector (Fig. [121]4f–g). This suggests strong selective pressure
maintains ELF3 expression throughout xenograft tumor growth.
Based on the tendency of polyclonal shELF3 cell populations to restore
ELF3 expression, we established clonal shELF3 and controls for all cell
lines and selected the best knockdown clones for molecular assays.
Interestingly, the shELF3 clones were morphologically distinct and
demonstrated reduced viability from their control counterparts (Fig.
[122]4k, Supplementary Fig. [123]9). The effect of ELF3 inhibition on
tumor growth was once again assessed using clonal populations of cells
(A549 shRNA-1, and isogenic control) injected into the flanks of
NOD-SCID mice (n = 24). Clonal A549 shELF3 cells did not express
detectable levels of ELF3 at time of injection (Fig. [124]4i,
Supplementary Fig. 8b). Control xenografts formed a large tumor mass
over time, whereas shELF3 clones remarkably showed no evidence of
growth over the course of the experiment (Fig. [125]4h–j). At endpoint,
small shELF3 masses were identified in six mice (25%). These small
masses had restored human ELF3 expression to control levels, indicating
the strong selective pressure and requirement for ELF3 in this
xenograft model. This role in tumor initiation is functionally
reminiscent of a previously established role in regulating stemness via
NOTCH3^[126]27. However, we did not detect NOTCH3 upregulation,
implicating alternative signaling networks (Supplementary
Fig. [127]8b–e). Nevertheless, we establish the requirement of ELF3
expression for tumor initiation and growth.
The potential for off-target effects was investigated by rescuing ELF3
expression in our clonal A549 shELF3 cell lines; in addition, ELF3 was
overexpressed in the corresponding clonal control. This overexpression
resulted in a 20% rescue in the shELF3 cells, which restored the
morphology to the control state, and a fourfold increase relative to
control cells (Fig. [128]4l, Supplementary Fig. [129]10). The effect of
ELF3 overexpression in control A549 cells expressing ELF3 was modest
and did not significantly affect colony formation but increased cell
proliferation (Wilcoxon p < 0.0001). In contrast, ELF3 rescue in A549
shELF3 clones increased the size and number of colonies formed, and the
proliferation rate of cells (Fig. [130]4m–o). The influence of ELF3
overexpression on A549 control and clonal shELF3 xenograft growth was
assessed by subcutaneous flank injection (n = 8 mice). Although ELF3
overexpression in A549 control cells did not significantly increase
xenograft tumor growth, the modest ELF3 rescue in clonal A549 shELF3
cells restored the ability to establish tumors in vivo (Fig. [131]4p,
Supplementary Fig. [132]10). Interestingly, ELF3 expression in tumors
at endpoint had increased eightfold compared with the injected cell
lines, providing further evidence of its importance to in vivo growth
(Supplementary Fig. [133]10). These results confirm the requirement of
ELF3 for A549 xenograft growth and further support its role as an
oncogene in LUAD.
ELF3 is associated with broad transcriptional reprogramming
Our results point to a clear oncogenic role in LUAD, countering
classifications of ELF3 as a tumor suppressor gene in other epithelial
tissues^[134]12. With mixed reports of gene behavior, we sought to
define the tissue specificity of ELF3 interaction networks. We
leveraged gene expression profiles of 13 tissues to construct ELF3
protein–protein interaction (PPI) networks, and assessed their
disruption in a pan-cancer analysis^[135]28. We discovered remarkable
tissue specificity of ELF3 PPI networks, with the largest disrupted
network observed in lung (Fig. [136]5a, Supplementary Fig. [137]11). A
NSCLC-specific analysis indicated the majority (25/33) of ELF3 PPI
disruptions occurred in LUAD as compared with LUSC and lung large cell
carcinoma (Fig. [138]5b) prompting a detailed investigation of altered
PPIs and pathways in this subtype^[139]29.
Fig. 5.
[140]Fig. 5
[141]Open in a new tab
Pan-cancer analysis of ELF3 protein–protein interaction networks
reveals tissue specificity. a Tissues are ordered vertically according
to decreasing number of altered ELF3 PPIs. Disrupted PPIs were
identified by comparing cancer profiles with non-malignant profiles in
the same tissue types, with PPIs gained in cancer shown in red, and
those lost in cancer shown in blue. The largest disrupted network was
in lung cancer, with 111 PPIs gained and 85 PPIs lost in cancer.
Considering all altered PPIs, 57% were specific to one tissue and only
15 PPIs were gained or lost in more than three tissues. b The largest
disrupted network was in lung adenocarcinoma (LUAD), with 14 lost and 9
gained PPIs in cancer. The only altered PPI common to LUAD and squamous
cell carcinoma (LUSC) was CCL11, whereas MYC, GLI2, and NKX2–1 were
common to LUAD and large cell carcinoma (LULC). The protein names of
the altered PPIs are indicated at the bottom of the figure. c
Comparison of altered ELF3 PPIs between TCGA LUAD data and isogenic
A549 cells (KRAS mutant samples only). Twenty-nine out of 156 PPI
partners were significantly deregulated in LUAD with high ELF3
expression and are color-coded by their respective GO Molecular
Function. Proteins at the left-side show significantly upregulated
(up-triangles) mRNAs when ELF3 is highly expressed in TCGA, and
right-side proteins represent significantly downregulated
(down-triangles) mRNAs when ELF3 is highly expressed in TCGA. Edge
color represents positive co-expression (red edges) or negative
co-expression (blue edges), and edge thickness is proportional to the
number of data sets supporting the edge. Node outline color represents
up- or down-regulation in A549 control cells compared with clonal
shELF3 cells (n = 3 biological replicates). Therefore, bright green
outline around up-triangles, and gray outline around down-triangles
indicate consistency between TCGA and isogenic cell line data. Node
size is proportional to the total number of altered PPIs, and node
highlight size is proportional to the FDR corrected p value of
expression fold change in isogenic cell lines, whereas blue bars
indicate the fold change value. Circles in the center indicate ELF3
partners whose differential expression was not statistically
significant. PPI = protein–protein interaction.
ELF3 PPI networks disrupted by high ELF3 expression were analyzed in
LUAD as compared with non-malignant lung tissue, and LUAD dichotomized
by KRAS mutation status based on the DNA-level association of ELF3
disruption in KRAS mutant LUAD. A total of 69 deregulated PPIs were
supported by at least one analysis (Supplementary
Figs. [142]12–[143]14, Supplementary Data [144]2). PPIs that were
supported across all analyses included ERBB3, ETS1, and TIMP3, whereas
those supported by at least two analyses included ARHGEF6, CCL11,
CLDN4, ELK3, ERBB2, FLI1, GLI2, KDM5B, NFKB1, PMF1, ROBO1, SLIT2,
SPDEF, SPI1, SPINT1, TAF9, TRADD, ZEB1, androgen receptor (AR), and
INPP5D. Importantly, ELF3 PPI network data from KRAS mutant LUAD was
compared against whole-genome expression data generated from isogenic
A549 control and shELF3 clonal cell lines. Of the proteins with
deregulated PPIs identified in TCGA, 26 were available on the A549 gene
expression microarray. Twenty-one out of these 26 demonstrated
significant levels of differential expression as a result of isogenic
ELF3 expression manipulation and were considered validated
(Fig. [145]5c, Supplementary Data [146]2).
Pathway analysis identified deregulation of not only those consistent
with previously established ELF3-related signaling events—for example,
IL-1β, NFκB, p38, and JNK signaling in inflammation^[147]30–[148]33,
ETS transcription factors in MAPK signaling^[149]34,[150]35, and NOTCH
and WNT in cancer stem cells and colorectal
cancer^[151]14,[152]27,[153]36—but other pathways that agreed with
phenotypes established in our isogenic systems and pointed to
previously undescribed functions. These included cell cycle, apoptosis,
adhesion, and motility functions (organization, junction, adherens,
cadherin), as well as AR signaling (Supplementary Fig. [154]12 and
Supplementary Data [155]1). Interestingly, AR-ELF3 was a LUAD-specific
interaction, and high expression of AR has been reported in NSCLC and
associated with growth potential in murine models^[156]37
(Supplementary Fig. [157]13).
Moreover, our network analyses predict other transcription factors as
predominantly altered ELF3-binding partners in LUAD (45–56%), including
the LUAD lineage-specific oncogene NKX2–1 (PPI in LUAD)^[158]38 and
epithelial-to-mesenchymal transition-promoting ZEB1 (anti-correlated
with ELF3 expression) (Fig. [159]5, Supplementary Figs. [160]13,
[161]14 and Supplementary Data [162]2). Similarly, we find positive
associations between ELF3 and E-Cadherin, an epithelial marker, in our
isogenic cell models and TCGA protein expression data (Supplementary
Fig. [163]15). ELF3-altered partners have varied subcellular
localization, raising the question of ELF3 function in alternative
cellular compartments. We investigated ELF3 localization in human
tumors (immunohistochemistry) and cell lines (immunofluorescence) and
observe both nuclear and cytoplasmic localization (Supplementary
Figs. [164]2, [165]16). It is possible ELF3 has as-of-yet
uncharacterized functions, such as its ability to transform breast
cells through an unknown cytoplasmic mechanism^[166]39.
High ELF3 expression is associated with poor survival
Finally, we investigated the prognostic utility of ELF3 in NSCLC
(n = 1715)^[167]40. High ELF3 expression was associated with poor
overall survival (OS) in NSCLC patients (log-rank p = 4.65E-04), an
association which improved in Stage I patients (log-rank p = 2.30E-05)
(Fig. [168]6). However, our study supports an increased biological
relevance in the LUAD subtype of NSCLC. Agreeing with this, high ELF3
expression was highly significantly correlated with OS in LUAD
(log-rank p = 5.09E-07; Stage I p = 1.00E-06), whereas no significant
distinction was observed in LUSC (Fig. [169]6). These data further
demonstrate the clinical relevance of ELF3 expression in LUAD.
Fig. 6.
[170]Fig. 6
[171]Open in a new tab
Prognostic relevance of ELF3 expression in non-small cell lung cancer.
Kaplan–Meier survival curves comparing overall survival of lung cancer
patients with high or low ELF3 expression (top and bottom tertiles,
log-rank p values). All Stages: NSCLC n = 1926, p = 4.65E-04, Hazard
Ratio (HR) (95% confidence interval) = 1.32 (1.13–1.55); LUAD n = 720,
p = 5.09E-07, HR = 2.08 (1.57–2.76); LUSC n = 524, p = 0.229, HR = 1.20
(0.89–1.60). Stage I: NSCLC n = 577, p = 2.30E-05, HR = 2.11
(1.51–2.95); LUAD n = 370, p = 1.00E-06, HR = 3.66 (2.27–5.88); LUSC
n = 172, p = 0.357, HR = 1.27 (0.76–2.14). NSCLC = non-small cell lung
cancer; LUAD = lung adenocarcinoma; LUSC = lung squamous cell
carcinoma.
Discussion
Here, we find that ELF3 is overexpressed in LUAD compared with
non-malignant lung and LUSC, and promotes oncogenic phenotypes
including the requirement for tumor growth in vivo. These data refine
a recent study implicating ELF3 as an oncogene in NSCLC^[172]18,
uncovering DNA-level mechanisms driving overexpression and a strong
survival association specifically in the LUAD histology. Interestingly,
the low mutation rate of ELF3 in lung cancer has allowed it to elude
sequencing-based screens of recurrently altered oncogenes; the high
frequency of alternative DNA-level disruptions in upwards of 80% of
LUAD underscores the importance of interrogating multiple ‘omics'
levels. Furthermore, this ETS transcription factor appears to follow
divergent paths in cancer with contrasting genetic mechanisms of
disruption. Although behavior and alteration mechanisms similar to
LUAD, including gene amplification, have been observed in colorectal
and breast cancer^[173]13,[174]14,[175]39, contrasting recurrent
deleterious mutations in ELF3 have been identified in biliary tract
cancer, mucinous ovarian carcinoma, and cancers of the cervix, stomach,
and bladder, indicating a tumor suppressive role in these cancer
types^[176]4–[177]8.
This observed diversity across epithelial malignancies is further
exacerbated by highly tissue-specific co-expression networks revealed
in our pan-cancer analysis. Interestingly, the largest ELF3-related
network was identified in lung cancer, more specifically LUAD. This
could point to not only the significant functional role in malignant
LUAD but also a larger biological role in normal lung biology. Although
ELF3 is currently poorly characterized in this regard, what is known
ties ELF3 to both fetal lung development and airway tissue repair,
functions that are often co-opted by cancer cells. ELF3 is highly
expressed in human fetal lung tissue^[178]41, and in mice Elf3 knockout
induces embryonic lethality^[179]42. In the adult airway, Elf3
regulates the kinetics of tissue repair following Clara cell-specific
injury, which is a putative cell-of-origin for LUAD^[180]43. Mirroring
these studies, we find that ELF3 cancer-specific networks include
NXK2–1, a regulator of fetal lung development and oncogenic potential
in LUAD^[181]38. We also find ELF3 regulates proliferation and
apoptosis both experimentally and through pathway analysis. This broad
reprogramming of ELF3-centric expression networks in LUAD underpins its
significance to this disease subtype and is representative of
large-scale reprogramming of cell states that result from the
deregulation of transcription factors^[182]44–[183]46. Indeed, the
development of therapeutic strategies to inhibit oncogenic
transcription factors such as ETS family members is an area of active
research^[184]47. The specific relevance of ELF3 to LUAD is highlighted
by its subtype-specific association with poor patient outcome; however,
we caution that ELF3 alterations are enriched in KRAS, STK11, and KEAP1
mutant tumors, all of which are themselves associated with poor
prognosis. Further studies are warranted to uncouple the effects of
these genetic alterations and to define the mechanistic role of ELF3 in
LUAD which may clarify its association with patient survival.
Chromosome 1q is a region of recurrent gain in NSCLC that is most
prominently associated with LUAD and disease
aggressiveness^[185]19–[186]22. Although putative oncogenic targets of
this event have been proposed, no conclusive oncogene in this region
has been identified. These putative gene targets have been found to
regulate complementary tumor hallmarks including autophagy and immune
phenotypes^[187]48–[188]50. Our discovery of a region of focal
amplification at the ELF3 locus, coupled with our findings that ELF3
knockdown is highly selected against in a polyclonal tumor growth model
and completely abrogates tumor growth in a clonal setting highlights
the importance of this oncogene. Importantly, we further discover
additional mechanisms of ELF3 overexpression including DNA copy number
gain, promoter hypomethlyation, and mutations in upstream regulators in
up to 80% of LUAD and across molecular subtypes. Based on the strong
phenotype observed in our in vivo models, the potential for broad
applicability to the most common subtype of LUAD, and the association
of improved survival for patients with low ELF3 expression, the
feasibility of therapeutic inhibition should be explored.
Methods
Patient tissue accrual
Tumor and non-malignant lung tissues from the British Columbia Cancer
Agency (BCCA) cohort were collected from treatment-naive patients at
the time of surgical resection and frozen in liquid nitrogen. Tissues
were obtained from the Tumor Tissue Repository of the British Columbia
Cancer Agency or Vancouver General Hospital under informed written
patient consent, relevant ethical regulations, and with approval from
the University of British Columbia—BC Cancer Agency Research Ethics
Board. Hematoxylin and eosin staining was performed for each malignant
and non-malignant sample and reviewed by a pathologist. Subsequently,
tumor specimens were microdissected to contain at least 70% tumor cell
content and non-malignant samples were verified to be histologically
normal. DNA and RNA from alternating sections were extracted using
standard protocols. Tissue microarrays from the Dalhousie University
cohort were made from surgically resected lung cancer specimens from
treatment-naive patients under informed written patient consent and
with approval from the Nova Scotia Health Authority Research Ethics
Board.
Immunohistochemistry
Formalin-fixed paraffin embedded (FFPE) tissues were deparaffinised,
and antigen retrieval was performed in decloaking chamber plus with
Diva decloaker (Biocare Medical, Pacheco, CA, USA). Endogenous
peroxidise blocking with peroxidazed-1 (Biocare Medical) and
non-specific blocking with background sniper 1 (Biocare Medical) was
performed in the Intellipath FLX. Next, primary antibody (1:75
dilution, anti-ELF3 HPA003479, Sigma-Aldrich, Oakville, ON, Canada) was
added, followed by Rb-HRP polymer and DAB chromogen (Biocare Medical).
Slides were counterstained with CAT hematoxylin (1:5 dilution, Biocare
Medical). Slides were imaged on a Pannoramic Digital Slide Scanner (3D
Histotech) and image analysis conducted using Pannoramic Viewer.
Tissue profiling
Gene expression profiles of 83 paired LUAD and NM tissues were
generated on the Illumina HT-12 Whole Genome 6, v3 BeadChip array
according to the manufacturer’s instructions (Illumina, San Diego, CA,
USA). Bead-level data were pre-processed using the R package mbcb (ver.
2.11.0) to perform background correction and probe summarization. Data
were then quantile normalized and log[2] transformed. Pairwise analysis
of log[2]expression values comparing LUAD and NM was performed using
the Wilcoxon sign-rank test. Fold change was calculated by subtracting
the normal log[2](expression) value from the paired malignant
log[2](expression) value, and transforming the log[2](fold change)
value: Fold change = 2^(log[2](fold change)). Paired LUSC and NM
tissues were obtained as described for LUAD. Extracted RNA was
subjected to RNA sequencing on the Illumina HiSeq 2000 sequencing
platform following library construction and bar-coding using a plate
based protocol developed at the British Columbia Genome Science
Centre^[189]51. For DNA copy number analysis, DNA was hybridized to
Affymetrix Genome-Wide Human SNP Array 6.0 arrays according to the
manufacturer’s instructions (Affymetrix, Santa Clara, CA, USA). Raw CEL
probe intensity files were processed and normalized using Partek
Genomics Suite 6.5. Probe sequence, fragment length, GC content, and
background adjustments were applied to correct for biases in signal
intensities. Tumor copy number profiles were processed using the
corresponding non-malignant copy number profile as a baseline.
Thresholds for DNA copy number alterations were applied as follows:
copy number loss < 1.7, copy number gain > 2.3. For methylation
analysis, DNA was bisulfite converted and hybridized to the Illumina
Infinium Human Methylation 27 array. Raw methylation data were
corrected for color bias and normalized using SSN normalization with
the Bioconductor package lumi in R statistical computing software.
Probe hypermethylation was defined as a change in β-Value (∆βV) ≥ 0.15
(at least 15% more methylated in tumor), whereas probe hypomethylation
was defined as a ∆βV ≤ −0.15 (at least 15% less methylated in tumor). A
two-group comparison between LUAD and NM was performed for each probe
of interest using a Mann–Whitney U test. For all statistical analyses,
a p value of < 0.05 was considered significant.
The Cancer Genomic Atlas data
For DNA copy number analysis, level 3 whole- genome SNP 6.0 copy number
segmentation files (GRCg37/hg19) were downloaded from the TCGA Data
Portal [190]https://tcga-data.nci.nih.gov/tcga/tcgaDownload.jsp.
Specifically, 551 LUAD (tissue codes 01 and 02) *.nocnv_hg19.seg.txt
files, from which a fixed set of germline-variable probes have been
removed, were assembled into a master segmentation file for GISTIC 2.0
analysis^[191]52. GISTIC 2.0 analysis was run on the Broad server using
the following Hg19 URL marker file:
[192]ftp://ftp.broadinstitute.org/pub/GISTIC2.0/hg19_support/genome.inf
o.6.0_hg19.na31_minus_frequent_nan_probes_sorted_2.1.txt. Amplification
and deletion thresholds were increased to 0.3; default settings were
used for all other parameters. For gene expression analysis, level 3
LUAD gene expression (IlluminaHiSeq) data from 513 LUAD and 58 NM lung
tissue specimens was downloaded from CancerBrowser
([193]https://genome-cancer.ucsc.edu/proj/site/hgHeatmap/). Fifty-seven
LUAD cases had paired NM expression. Fold change was calculated with a
cutoff of ±2, and expression was compared using a paired sign-rank
test. For methylation analysis, level 3 Illumina Infinium
HumanMethylation450K data from 460 LUAD and 32 NM specimens was
downloaded from CancerBrowser. For gene mutation analysis, level 2
Mutation Annotation Format (MAF) files for 543 LUAD were downloaded
from the TCGA Data Portal. For the genes KRAS, BRAF, EGFR, ERBB2,
MAP2K1, MET, HRAS, and NRAS, only annotated driver mutations were
considered^[194]23. For all other genes analyzed, silent mutations were
removed. For protein expression analysis, TCGA normalized protein
expression data for E-cadherin was downloaded from
cBioPortal^[195]53,[196]54. Lung tumor specimens from The Cancer Genome
Atlas are composed of a varying degrees of tumor cell contents^[197]55.
Owing to potential dilution of expression or methylation signals by
stromal cells, a second TCGA data set was generated that comprised of
samples with at least 60% tumor cell content, termed the TCGA-60 data
set for tumor purity analysis.
Cell line experiments
Cell lines were obtained from American Type Culture Collection (ATCC,
Manassas, VA, USA) and maintained according to ATCC guidelines. All
cell lines were routinely monitored for mycoplasma contamination. LUAD
cell lines were cultured in RPMI 1640 (Gibco–ThermoFisher, Waltham, MA,
USA) supplemented with 10% FBS and maintained in a humidified 5% CO[2]
incubator. Human Bronchial Epithelial Cells are untransformed cells
that have been immortalized by expression of cyclin-dependent kinase
(Cdk) 4 and humantelomerase reverse transcriptase (hTERT) (HBEC-KT),
and have been further altered to harbor p53 knockdown, oncogenic RAS
expression (HBEC-KTR), and with both alterations
(HBEC-KTR53)^[198]26,[199]56. All HBEC-KT cells were cultured in
Opti-MEM media (Gibco–ThermoFisher, Waltham, MA, USA) supplemented with
0.0002 ng/μl EGF and 30 μg/ml BPE. Stable knockdown of ELF3 was
achieved by lentiviral transfection of vectors encoding shRNA inserts
directed against ELF3 mRNA as well as a puromycin resistance selectable
marker (Sigma-Aldrich). Five shRNA clones were tested and the clone
with the best knockdown of ELF3 transcript and protein levels in each
cell line was selected for phenotypic assays. Virus was prepared for
ELF3-shRNAs and a control with no shRNA insert. LUAD cell lines were
transfected, and after 24 h media was replaced with
puromycin-containing media. Following complete puromycin-induced death
of control cells after 3–5 days, transfected cells were cultured in
puromycin-media for an additional 7 days. Knockdown efficiencies were
quantified by qRT-PCR (TaqMan—Applied Biosystems, Carlsbad, CA) using
18 S as an endogenous control: Hs00963881_m1 (ELF3) and Hs99999901_s1
(18 S), and verified at the protein level by western blot (described
below). Stable overexpression of ELF3 in HBEC-KT and A549 cell lines
was achieved by lentiviral delivery of a vector containing the full ORF
with a blasticidin resistance selectable marker or an empty vector
control (Invitrogen, Carlsbad, CA, USA). For immunoblotting, whole-cell
extracts were prepared in radioimmunoprecipitation assay (RIPA) lysis
buffer (20 mm Tris-HCl, pH 7.5; 150 mm NaCl; 0.5% DOC, 1% NP-40 and
0.1% SDS). For cytoplasmic and nuclear fractionation, cells were first
pelleted and resuspended in Buffer A (10 mm HEPES, pH 7.9; 1.5 mm
MgCl2; 10 mm KCl). After 10 min of incubation on ice, lysates were
pelleted and supernatant was collected as the cytoplasmic extract. The
pellet was washed two times with Buffer A and resuspended in Buffer C
(20 mm HEPES, pH 7.9; 25% glycerol; 420 mm NaCl; 1.5 mm MgCl[2] and
0.2 mm EDTA). Tubes were incubated for 20 min on ice and then
centrifuged to clear the nuclear extract. All lysis buffers were
supplemented with protease and phosphatase inhibitors (2 mm Na3VO4;
1 mm NaF; 2 mm β-glycerolphosphate; 0.2 mm PMSF; 0.5 mm DTT and
Complete protease inhibitor cocktail (Roche Diagnostics, Laval, QC,
Canada)). Protein concentrations were determined by the BCA Protein
Assay (Pierce, Rockford, IL, USA) according to the manufacturer's
recommendations. Immunoblot analysis was performed on cell-equivalent
lysates subjected to sodium dodecyl sulfate-polyacrylamide gel
electrophoresis and electrophoretic transfer to polyvinylidene
difluoride (Bio-Rad Laboratories, Mississauga, ON, Canada). Membranes
were probed with anti-ELF3 (1:1000 dilution, Abcam anti-ESE1 antibody
[EPESER1] (ab133621), Toronto, ON, Canada) and anti-Histone H3 (1:2000
dilution, #9715, Cell Signaling Technology, Danvers, MA, USA). For
immunofluorescence analysis of 2-D monolayer cultures, cells were fixed
in paraformaldehyde for 10 min Cells were then permeabilized using
0.05% Triton in PBS for 8 min After washing with 0.25% Tween in PBS,
PBS containing 1% BSA was added to cells for 30 min to block
non-specific interactions. Subsequently, anti-ELF3 (1:200,
Sigma-Aldrich (HPA003479)) was added for 16 h at 4 degrees Celsius in
1% BSA in PBS. After washing with PBS, secondary antibodies were added
for 30 min in 1% BSA in PBS at room temperature. Staining of
filamentous actin was performed by adding rhodamine-conjugated
phalloidin (1:400, Invitrogen, Burlington, ON, Canada). Coverslips were
mounted in Fluoroshield Mounting Medium with DAPI to stain DNA (Abcam,
Toronto, ON, Canada). Cell images were acquired with a Zeiss Colibri
fluorescence microscope, AxioCam MRm camera and AxioVision Rel. 4.8
software (Carl Zeiss Canada Ltd., Toronto, ON, Canada). Cell viability
of shRNA transfected HCC827 cells and isogenic controls was quantified
calorimetrically using MTT reagent daily for 5 days. Cells were
incubated for 4 h with MTT at 37 degrees Celcius before the reaction
was ended by the addition of 20% SDS. Plates were left at room
temperature overnight before scanning. Cell viability of A549 and H2030
cells was quantified calorimetrically by Alamar Blue 48 h post
treatment with non-targeting control or pooled ELF3-targeting siRNAs.
Cell proliferation was quantified using the BD PharmingenTM Apoptosis,
DNA Damage and Cell Proliferation Kit (BD Bioscience, Mississauga, ON,
Canada). Cells were incubated with BrdU for 8–24 h, and processed
according to the manufacturer's instructions. In brief, cells were
fixed, permeabilized, and treated with DNase I before staining with
PerCP-CyTM5.5 Mouse Anti-BrdU antibody and DAPI (1 μg/ml). Cells were
analyzed using the BD FACS CantoTM II cell analyzer (BD Bioscience).
Experiments were repeated three times. Cell proliferation was also
measured using the IncuCyte Live Cell Analysis system (Essen
BioScience, Ann Arbor, MI, USA). For each cell line, a mask was created
at the first measurement to determine average cell size, and applied to
images taken every 2 h for 7 days to count the number of cells at each
time point. Measurements were taken in technical sextet, and
transformed standard deviation is show for a representative biological
triplicate. Measurements of cell number at each time point were
normalized to starting confluence of, as well as to plateau of growth
achieved after 5–6 days (representing 100% confluence). For colony
formation, single cell suspensions were prepared in growth media
supplemented with 20% fetal bovine serum, and 0.3% low-melting point
agarose (Invitrogen, Carlsbad, CA, USA). One milliliter of cell
suspension was plated onto an equal volume of supplemented media with a
0.5% low-melting point agarose concentration. Each cell line was seeded
in triplicate in 12-well plates and cultured for 14–21 days at 37˚C.
Experiments were repeated three times. For HBEC-KT colony formation
assays, 1000 viable cells were suspended and plated in 0.37% agar in
Opti-MEM medium supplemented with 20% fetal bovine serum and 50 μg/mL
bovine pituitary extract with 5 ng/mL EGF in triplicate 12-well plates,
and were layered over a 0.50% agar base in the same medium as the one
used for suspending the cells26. Cell apoptosis was quantified using
the BD PharmingenTM Annexin V Apoptosis Detection Kit I according to
the manufacturer’s instructions (BD Bioscience). Cells were grown in
complete or serum free media for 72 h prior to cell processing,
staining, and analysis by flow cytometry on the BD FACS CantoTM II cell
analyzer (BD Bioscience). Live cells that were adherent at time of
collection were used as a gating control.
Xenograft tumor growth
All animal protocols complied with relevant ethical regulations and
were approved by the Animal Care Committee of the University of British
Columbia (Vancouver, British Columbia, Canada). ELF3 knockdown cell
lines and controls were subcutaneously injected (A549: 2.5 × 10^6 cells
per site; HCC827: 5 × 10^6 cells per site) into the left and right
flanks of 6–8 week old NOD-SCID mice. Tumor volume was measured several
times weekly until a total tumor burden of 1500 mm^3 was achieved or
tumors became ulcerated, at which point mice were euthanized. ELF3
overexpression cell lines and controls were subcutaneously injected
(A549: 1.0 × 10^6 cells per site) into the left and right flanks of 6–8
week old NRG mice. Tumor volume was measured several times weekly until
a total tumor burden of 1500 mm^3 was achieved or tumors became
ulcerated, at which point mice were euthanized. For shELF3.cl-lacZ and
shELF3.cl-OE xenografts (injected at 1.5 × 10^6 cells per site), the
slow growth rate resulted in a study endpoint whereby tumors reached a
volume from which RNA could be extracted. DNA from FFPE xenografts was
extracted using the Biostic FFPE Tissue DNA Isolation Kit (MO BIO
Laboratories, Inc., Carlsbad, CA, USA) and cleaned up using the
MiniElute Reaction Cleanup Kit (Qiagen, Hilden, Germany). DNA from cell
line input was extracted using the DNeasy Blood & Tissue Kit (Qiagen,
Hilden, Germany). DNA copies of lentiviral vector were quantified using
Taqman Copy Number Assays, with primers designed to amplify the
puromycin resistance cassette and calibrated against TFRC:
Hs02677106_cn (diploid in HCC827); diploid reference assay for all
samples was SFTPB: Hs01649948_cn. Copy number was assessed using
CopyCaller Sofware v2.0 (Applied Biosystems, Carlsbad, CA, USA). After
tissue homogenization, RNA was extracted according to standard Trizol
protocols. Relative human cell-specific expression of ELF3
(Hs00963882_g1 – no cross reactivity with mouse) and NOTCH3
(Hs01128537_m1 – no cross reactivity with mouse) was quantified by
qRT-PCR (TaqMan—Applied Biosystems) using mouse Actb (Mm04394036_g1)
and human Actb (Hs99999903_m1) as endogenous controls. Whole-cell
extracts were prepared in RIPA lysis buffer (20 mm Tris-HCl, pH 7.5;
150 mm NaCl; 0.5% DOC, 1% NP-40 and 0.1% SDS) and supplemented with
protease and phosphatase inhibitors (2 mm Na[3]VO[4]; 1 mm NaF; 2 mm
β-glycerolphosphate; 0.2 mm PMSF; 0.5 mm DTT and Complete protease
inhibitor cocktail (Roche Diagnostics, Laval, QC, Canada)). All TaqMan
assays were read in a 7500 Fast Real-Time PCR System (Applied
Biosystems).
PPI network and pathway analysis
For physical PPIs, we obtained 189 interacting partners of ELF3 and an
additional 2526 PPIs among them (experimentally detected or
computationally predicted interactions) from IID (ver. 4_2017;
[200]http://ophid.utoronto.ca/iid)^[201]28. Next, we annotated these
PPIs with differential gene expression and differential gene
co-expression data to construct tissue-specific and
NSCLC-subtype-specific ELF3 interaction networks for further analyses.
To investigate differential gene co-expression networks, gene
expression data (HG-U133plus2 and HG-U133a chips) was downloaded from
Gene Expression Omnibus (GEO, [202]https://www.ncbi.nlm.nih.gov/geo/)
and analyzed in batches, described below. Data were normalized using
MAS5 method using Bioconductor package (Affy package version 1.48.0)
available in R 3.2.3. For the pan-cancer analysis, data sets comprising
samples from more than one tissue were separated into tissue-specific
subsets. These data covered 1858 tumor (T) and 1026 non-malignant (N)
treatment-naive patient samples across 13 tissue sites, including lung
(n = 188 N, n = 510 T); breast (n = 86 N, n = 93 T); cervix (n = 30 N,
n = 52 T); colon (n = 117 N, n = 345 T; endometrium (n = 21 N,
n = 58 T; esophagus (n = 132 N, n = 159 T); kidney (n = 102 N,
n = 181 T; liver (n = 20 N, n = 20 T); mouth (n = 67 N, n = 70 T);
ovary (n = 49 N, n = 71 T); pancreas (n = 61 N, n = 73 T); prostate
(n = 96 N, n = 143 T); thyroid (n = 57 N, n = 84 T). Lung
subtype-specific analysis consisted of adenocarcinoma (n = 166 N,
n = 398 T); squamous cell carcinoma (n = 35 N, n = 71 T); and large
cell carcinoma (n = 33 N, n = 41 T).
Pan-cancer: [203]GSE19383, [204]GSE26910, [205]GSE3744, [206]GSE5764,
[207]GSE20437, [208]GSE5462, [209]GSE6883, [210]GSE9574, [211]GSE9750,
[212]GSE20916, [213]GSE8671, [214]GSE41258, [215]GSE5364,
[216]GSE11024, [217]GSE14762, [218]GSE21816, [219]GSE7023,
[220]GSE8271, [221]GSE6280, [222]GSE6344, [223]GSE781, [224]GSE29721,
[225]GSE14520, [226]GSE31908, [227]GSE14407, [228]GSE15578,
[229]GSE18520, [230]GSE36668, [231]GSE38666, [232]GSE15471,
[233]GSE16515, [234]GSE22780, [235]GSE17951, [236]GSE32448,
[237]GSE32982, [238]GSE3325, [239]GSE6956, [240]GSE29265, [241]GSE3467,
[242]GSE3678, [243]GSE6004, [244]GSE27155, [245]GSE17025,
[246]GSE20347, [247]GSE23400, [248]GSE29001, [249]GSE30784,
[250]GSE31056, [251]GSE33426, [252]GSE38129, [253]GSE53757,
[254]GSE64985, [255]GSE7305, [256]GSE7307, [257]GSE7803.
Non-small cell lung cancer: [258]GSE31210, [259]GSE10245,
[260]GSE19188, [261]GSE28571, [262]GSE31908, [263]GSE7670,
[264]GSE10072, [265]GSE5364.
Next, for each pair of N and T samples, we calculated a Pearson
Correlation Coefficient to generate a gene co-expression matrix across
samples in N (i.e., ρ[N]) and across samples in T (i.e., ρ[T]). In the
pan-cancer analysis, pairs of N and T samples were grouped by tissue.
In the lung cancer subtype analysis, pairs of N and T samples were
grouped by histological subtype. In the lung adenocarcinoma analysis,
we restructured samples into 11 pairs of sample sets, each containing
at least 20 T and 20 N samples. For each analysis we calculated
differential gene co-expression matrix for each N and T pair:
[MATH:
Diff
(N,T)<
/mrow>=ρN−ρT :MATH]
1
We used the top 1% of values in the Diff[(N,T)] matrix and overlaid it
on the physical PPI network around ELF3 to define altered PPIs in
cancer for each tissue. Next, we separated annotated PPIs into those
lost in cancer (
[MATH:
ρN>ρT :MATH]
) and those gained in cancer (
[MATH:
ρT>ρN :MATH]
).
Differential gene expression network: we used level 3 mRNA expression
data of 498 LUAD patient samples from TCGA. We removed genes with more
than 10% missing values. For 15,448 remaining genes, we replaced
missing values with average of available values across all samples of
each gene. In all, 125 samples in the top quartile of ELF3 mRNA
expression distribution were used as ELF3^high and 125 samples in the
bottom quartile of this distribution were used as ELF3^low samples.
ELF3 expression of all samples in ELF3^high set was at least two times
of its expression in all samples in ELF3^low set. Next, we stratified
each group of samples into two groups: (1) KRAS^mut samples (41
ELF3^high and 31 ELF3^low samples) and KRAS^wt samples (84 ELF3^high
and 94 ELF3^low samples).
For each of the KRAS^mut and KRAS^wt groups, we calculated differential
mRNA expression between ELF3^high and ELF3^low samples (two-tailed t
test). After correcting raw p values for multiple hypothesis testing
using Benjamini–Hochberg (false discovery rate) method we used genes
with corrected p value < 0.05 and fold change of at least 1.2 as
differentially expressed genes to annotate proteins in the first level
network of ELF3.
For isogenic cell lines, we used mRNA expression of three replicates of
shELF3 A549 cell line and three control samples (explained above).
Owing to low analysis power which is a result of small number of
samples, we did not use standard gene expression analysis on cell
lines, rather, we used difference of geometric mean and raw p value
across genes between shELF3 and control lines to further validate our
TCGA KRASm^mut network analysis.
In order to summarize consistency and rank proteins based on their
alterations across different conditions we defined three support score
components as the following: a protein has full support in a data set
if its alteration passes the defined threshold; a protein has partial
support in a data set if it satisfies two conditions: (a) it has full
support through another data set, and (b) its alteration level in this
data set is slightly less than the defined threshold (for example,
TIMP3 is partially supported in TCGA-LUAD-KRAS^wt set, since it has
full support from GEO and TCGA-LUAD-KRAS^mut (further supported through
cell lines), and its differential expression q value is slightly above
0.05 (between 0.05 and 0.06) in TCGA-LUAD-KRAS^wt), and a gene has no
support if it is present neither fully nor partially supported.
Full support from each data set for a protein increases its support
score by 2 points, partial support increases its support score by 1
point and no support has value 0. The final support score for each
protein is calculated by adding up these three score components.
For pathway analysis, we used pathDIP (version 2.5;
[266]http://ophid.utoronto.ca/pathDIP)^[267]29, extended pathways based
on experimental and computational PPIs at cutoff association score of
0.95, to perform pathway enrichment analysis across the above three
pairs of differential networks (GEO N vs T, TCGA KRAS^mut ELF3^low vs
ELF3^high, and TCGA KRAS^wt ELF3^low vs ELF3^high). Next, we used term
enrichment analysis tool in pathDIP to further summarize enriched
pathways (pathways with corrected q value < 0.05). The results are
presented in Supplementary Fig. [268]12, and additional relevant
details are available in Supplementary Data [269]1.
Survival analysis
Meta data for Kaplan–Meier survival analysis were obtained from
[270]http://kmplot.com/analysis/index.php?p=service&cancer=lung^[271]40
. Samples were ranked by ELF3 expression and survival curves of the top
and bottom expression tertiles were compared by log-rank analysis.
Reporting summary
Further information on research design is available in the [272]Nature
Research Reporting Summary linked to this article.
Supplementary information
[273]Supplementary Information^ (2.6MB, pdf)
[274]41467_2019_13295_MOESM2_ESM.pdf^ (82.5KB, pdf)
Description of Additional Supplementary Files
[275]Supplementary Data 1^ (32KB, pdf)
[276]Supplementary Data 2^ (22.2KB, pdf)
[277]Reporting Summary^ (89.4KB, pdf)
Acknowledgements