Acquired mutations and transcriptional remodeling in long-term estrogen-deprived locoregional breast cancer recurrences

Institutional Review Board approval from both participating institutions (University of Pittsburgh IRB# PRO15050502, The Charité IRB Office) was obtained prior to initiating the study. Inclusion criteria for this study were (1) patients harbored patient-matched formalin-fixed paraffin-embedded (FFPE) tissue from primary breast cancers and local recurrences, (2) biospecimens contained macrodissectable regions with sufficient tumor cellularity, and (3) disease was treated continuously with a form of estrogen-depleting therapy prior to the recurrence. Biospecimens were reviewed by a trained molecular pathologist to confirm pathology, to quantify tumor cellularity, and to highlight regions of relatively high tumor cellularity for macrodissection. If a slide region harbored sufficient, microscopically verifiable adjacent normal cells, this region was also dissected and processed for downstream analyses. Between four to ten (depending on tumor size) 10-μm FFPE sections immediately adjacent to the H&E-analyzed section were pooled and underwent dual DNA/RNA extraction using Qiagen’s AllPrep kit. Nucleic acids were quantified fluorometrically with a Qubit 2.0 Fluorometer and quality assessed with an Agilent 4200 TapeStation Instrument prior to sequencing.

RNA-seq library preparation was performed for all 12 cases using approximately 100 ng of RNA and Illumina’s TruSight RNA Pan-Cancer (1385 targets) protocol. DNA-seq library preparation was performed for 10 (6 with associated normal tissue, 2 cases were excluded based on limiting DNA content) cases using no less than 30 ng of DNA and Illumina’s TruSeq Exome protocol with TruSight RNA Pan-Cancer probes for hybridization-based capture. Indexed, pooled libraries were then sequenced on Medium Output flow cells using an Illumina NextSeq 500 system (paired-end reads, 2 × 75 bp). A target of 5–10 million reads per sample was used to plan indexing and sequencing runs for RNA sequencing, and a target of 10–15 million reads was used for DNA sequencing. Additional RNA-seq data derived from luminal metastatic breast cancers were obtained from the MET500 cohort (n = 47) [28] and our own University of Pittsburgh’s Women’s Cancer Research Institute (WCRC) cohort (n = 89) [29‐31]. All RNA-sequencing FASTQ files were quantified with k-mer-based lightweight-alignment (Salmon, quasi-mapping mode, 31-kmer index using GRCh38 Ensembl v82 transcript annotations, seqBias and gcBias corrections) [32]. tumorMatch was used to validate sequencing pairs were patient-matched as done previously [30].

RNA-seq read counts and mapping percentages were calculated (Data Supplement S1) and transcript abundance estimates were collapsed to gene level with tximport [33]. Log2-transformed TMM-normalized CPM (log2normCPM) values were implemented for subsequent analyses [34, 35]. DNA-seq reads were aligned with bwa –mem (v.0.7.13) to an hg19 reference, sorted with samtools (v1.3), duplicates marked and removed with picardtools (v1.140), and local realignment performed with GATK (v3.4-46) [36‐38]. Average coverage depth for the processed bam file was calculated using GATK’s DiagnoseTargets and the Illumina Pan-Can bed file (Data Supplement S2). Metrics for average coverage values across all target intervals were plotted with ggplot2. Publicly available RNA sequencing or microarray data of ER-positive breast cancer cell lines in the absence or presence of E2 were collected from GEO, including MCF7_1 (GSE89888), MCF7_2 (GSE51403), MCF7_3 (GSE78286), MCF7_4 (GSE94493), T47D_1 (GSE89888), T47D_2 (GSE3834), T47D_3 (GSE3834), T47D_4 (GSE108304), ZR75-1 (GSE61368), BT474 (GSE3834), MM134 (GSE50695), and SUM44 (GSE50695). Cells were deprived in E2-free media, stimulated with vehicle (Veh) or 1 nM/10 nM E2 for 16 or 24 h, then processed from gene expression profiling. KLK7 and PROM1 Log2 (TPM+1) expression value were compared between Veh and E2-treated group of each cell line.

To determine enriched variants in recurrences versus patient-matched primary tumors, VarScan2 was implemented [39]. More specifically, primary and recurrent samtools mpileup files derived from processed bam files were input into VarScan2 using somatic mode, with somatic p values representing the significance of a particular variant being acquired or enriched in the recurrence [SS = 1 or SS = 2]. Tumor purity estimates, as assessed by a molecular pathologist, were included in VarScan2 to correct contaminating normal cell influence on allele frequencies. The minimum coverage for a variant to be considered was 40X, with a minimum allele frequency (AF) of 0.05 in either the primary or recurrence and a minimum of 5 reads supporting the variant. Germline variants were determined for cases containing a matched normal (ERLR_01, ERLR_02, ERLR_07, ERLR_08, ERLR_12, and ERLR_15) using VarScan2’s germline mode with the same parameters. VCF output files were then imported into R using the VariantAnnotation package [40]. If a normal sample was available for the case, all germline variants (AF > 0.30) were excluded from subsequent analyses. Additionally, to limit technical artifacts especially considering specimens were formalin-fixed paraffin-embedded [41], a “blacklist” of variants was created including those called in at least 3 of the normal samples. Germline and blacklist-removed variants were then annotated with Annovar [42]. Lastly, to call recurrence-enriched, potentially pathogenic variants, the following inclusion criteria were enacted: (1) VarScan2 somatic p value < 0.05, (2) > 2-fold gain in allele frequency in the recurrence versus the primary, (3) minimum AF of 0.10 in the recurrence, (4) non-silent, and (5) an ExAC AF < 0.01 considering some samples were without a paired normal [43]. These non-silent, enriched, potentially pathogenic variants were then plotted using the OncoPrint function in ComplexHeatmaps [44]. A Pearson R correlation was calculated between the frequency of enriched variants and disease-free survival. PIK3CA mutations were visualized with IGV (2.3.60) [45] and variant allele frequencies were derived from VarScan2.

RNA-seq reads covering mutation sites called from DNA-seq of the corresponding sample were extracted from bam file and counted. Variants with at least 2 supporting reads containing the altered allele and with AF greater than 0.05 in either primary or recurrence were considered.

To estimate copy number ratios, CNVkit was implemented on processed bam files using default settings and the -drop-low-coverage option [46]. A pool of bam files from adjacent normal tissue, sequenced in the same manner, was used as a reference. Probe- and segment-level copy number estimates were finalized with CNVkit’s call function, which utilizes circular binary segmentation [47]. To adjust for tumor purity and normal contamination, the –m clonal option was used with tumor purities from pathologic evaluations. Copy number ratios were then plotted with the heatmap function, and copy number values were assessed and plotted with ggplot2. Gene-level copy number estimates represent the mean copy number call across all probe targets.

Hierarchical clustering was performed using the heatmap.3 function (https://​raw.​githubuserconten​t.​com/​obigriffith/​biostar-tutorials/​master/​Heatmaps/​heatmap.​3.​R) in R on log2normCPM values of the top 10% most variable genes (defined by IQR) with 1 minus Pearson correlations as distance measurements and the “average” agglomeration method. Differential expression between primary and recurrent tumors was analyzed with limma. Raw counts were input into the voom function and quantile normalized prior to fitting the linear model and performing the empirical Bayes method for differential expression [48, 49]. The linear model was fitted with a design that accounts for the paired nature of the cohort (model = ~Patient+Tissue [primary or recurrence]). Outlier expression gains and losses were determined for each patient by discretely categorizing genes into one of 5 categories. If log2FC values (i.e., recurrence log2normCPM – primary log2normCPM) for a given gene were less than Q1 – (1.5 × IQR) or Q1 – (3 × IQR), using case-specific log2FC values for all genes as the distribution, that gene was deemed an “Outlier Loss” or “Extreme Loss” respectively. If log2FC values calculated were greater than Q3 + (1.5 × IQR) or Q3 + (3 × IQR), it was deemed an “Outlier Gain” or “Extreme Gain” respectively. All other genes with intermediate fold changes were classified as “Stable.” To determine subtype expression of KLK7, PROM1, and NDRG1, normalized microarray expression data along with PAM50 calls was obtained from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) through Synapse (https://​www.​synapse.​org/​, Synapse ID: syn1688369), following IRB approval for data access from the University of Pittsburgh [50]. Overlap with genes in long-term estrogen-deprived, ER-positive breast cancer lines (HCC1428, MCF7, T47D, ZR75.1) was performed by running a separate differential expression analysis (LTED vs. parental lines) on microarray data with limma [49, 51]. Dysregulated gene overlap was designated if the nominal p value and FDR-adjusted p value were both < 0.05 in the local recurrence and LTED differential expression analysis, respectively. Binary dichotomization of METABRIC samples using NDRG1 expression (> 50th percentile, < 50th percentile) and log-rank testing were used to assess significant differences in disease-specific survival (DSS) and then Kaplan-Meier curves were plotted with survminer [52, 53].

Dual hybrid-capture DNA/RNA sequencing was performed for 1385 cancer genes on 12 paired primary tumors and local recurrences from patients with ER-positive breast cancers that underwent continuous endocrine therapy (Table 1). RNA-seq data (Supplementary Table 1) underwent unsupervised hierarchical clustering of normalized RNA expression values which showed most patient-matched pairs clustered transcriptionally with their matched primary—regardless of the length of disease-free survival (Fig. 1a). Unlike a previous transcriptome-wide analysis of primary breast cancers and matched bone metastases [30], there was no significant correlation in pair transcriptional similarity and time to recurrence—although a trend towards negative correlation was observed (Pearson R = − 0.37, p value = 0.236). Only a single recurrence showed marked transcriptional deviation from its matched primary (ERLR_03_R1); whereby it lost ER positivity and gained HER2 positivity clinically. Copy number alterations (CNAs) between primary and recurrences were analyzed in the targeted capture regions for 10 cases (Supplementary Figure 1 and Supplementary Figure 2). Similar to expression, CNAs were largely consistent among the recurrences when compared to their matched primary (Fig. 1b). Two exceptions were recurrences from cases ERLR_01 and ERLR_03, which showed distinct copy number profiles from the matched primary tumors with poor correlation between primary and recurrence CNA values versus all other cases (Supplementary Figure 3). Notably, unlike case ERLR_03, ERLR_01 interestingly retained a similar expression profile despite a distinct CNA profile. An analysis of shared variants validated both DNA and RNA extracts originated from the same patient (Supplementary Figure 4), excluding the possibility of sample mixup. ERBB2 copy number values correlated well with RNA expression (Pearson R 0.92), as did other amplified genes including CDK12 and CCND1 (Fig. 1c, Data Supplement S3).

A total of 406 distinct, presumed-somatic nonsynonymous mutations were detected in either a primary or recurrence at an AF > 5% among the 10 DNA-sequenced cases (Data Supplement S4). To assess if there are shared DNA mutations acquired in recurrences, an analysis of enriched single nucleotide variants (SNVs) was performed which showed 56 statistically enriched SNVs in local recurrences versus matched primary tumors (Fig. 2a, Data Supplement S5). SNVs in two genes were found to be enriched in more than one case (n = 2 [20%]), AKAP9 (R3320W, S319*) and KMT2C (T1969I, Y366N, R894Q). The recurrent mutations did not exhibit features suggesting functional selection, such as being within a conserved functional domain or within a COSMIC [54] hotspot region, making it difficult to assess if these are pathogenic. Other case-specific, n-of-one enriched mutations included nonsense mutations in ARID1A (Q1424*, case ERLR_20, primary AF 0.5%, recurrence AF 16.5%) and BRD1 (Q467*, case ERLR_01, primary AF 0.93, recurrence AF 57.88%), an acquired TP53 mutation (S241C, case ERLR_03, primary AF 0.0%, recurrence AF 53.4%), and an enriched NCOR2 mutation (A4942C, case ERLR_08, primary AF 4.4%, recurrence AF 19.4%). In case ERLR_01, an enrichment of a suite of three somatic mutations in PIK3CA was observed (E542K, Q546K, E726K) in the recurrence (Fig. 2b). Notably, the number of enriched, non-silent SNVs ranged from 0 to 13 and was positively correlated with clinical time to recurrence (Fig. 2c). No acquired ESR1 mutations were observed. These mutations were examined in the corresponding RNA-seq data to determine if they are expressed. Out of 633 total mutations—considering some of the 406 distinct mutations were present in both matched tumors—315 were detected in RNA with at least 2 supporting reads of the altered allele and an AF ≥ 5%. Allele frequencies called from DNA-seq and RNA-seq data correlated well (Supplementary Figure 5, Pearson R = 0.609, p value = 2.57e−33). Noteworthy, out of the 56 enriched SNVs in recurrence at the DNA level, 31 distinct mutations can be detected with confidence at the RNA level (Data Supplement S6)—including AKAP9, KMT2C, ARID1A, BRD1, and TP53 as discussed above.

A differential expression analysis, comparing all primary tumors versus all local recurrences, yielded no genes passing an FDR-corrected p value of less than 0.05—which is perhaps expected given the heterogeneity of clinical specimens and the limited number of cases (Data Supplement S7). Nonetheless, 71 genes with an average, voom normalized expression value of 2 or greater, a nominal p value of less than 0.05, and a log2 fold-change greater than ± 0.5 were identified (Table 2). Some of these genes, including the upregulation of EPOR, NDRG1, IDH2, CEBPA, and PTPA and downregulation of ESR1, IGF1R, NFKB1, and RUNX2, are also differentially expressed in long-term estrogen-deprived ER-positive cell lines (Supplementary Figure 6) [55].

To further explore major expression changes that may be driving recurrence and therapy resistance, an outlier expression analysis was performed using gene-level fold-change values of each patient-matched case (Data Supplement S8). Unlike non-silent SNVs, recurrent transcriptional gains and losses were common (Fig. 3a). These included gains and losses in shared pathway members, notably NTRKs and SFRPs, respectively; targetable upregulation of growth factor pathway mediators such as FGFR4 and EGF; and outlier gains in the CDK regulator CCNE1. Three of 12 cases also shared outlier expression gains in TERT, with case ERLR_14 harboring a particularly extreme enrichment from near undetectable levels in the primary tumor (Fig. 3b). Case ERLR_03’s recurrence, which was most dissimilar to its patient-matched pair transcriptionally, showed extreme loss and gain of ESR1 and ERBB2, respectively. CNA analysis confirmed recurrence-specific ERBB2 amplification and is consistent with previous studies of endocrine therapy-treated breast cancers selecting for HER2 signaling in more advanced tumors. The most recurrent outlier loss involved ESR1.

Five cases showed outlier expression losses of ESR1 (Fig. 4a). Despite estrogen receptor being the driver of ER-positive breast cancer and a major regulator of transcription, counterintuitively, 4 of 5 of the recurrences which lost ESR1 expression generally retained the expression profile of their patient-matched primary (Fig. 1a). Importantly, many of these cases also harbored very similar CNA profiles (Fig. 1b), implying the recurrences were derived from a continuous clonal lineage as opposed to being completely distinct breast cancers. Thus, to explore the transcriptional consequences of acquired ESR1 loss in ER-positive disease and identify potential bypass mechanisms driving ER− independence, a differential expression analysis was performed on the subset of pairs with outlier ESR1 expression losses. This analysis revealed several recurrently dysregulated genes in ESR1-depleted recurrences (Fig. 4b, Data Supplement S9). Two standout genes, KLK7 and PROM1, showed the highest degree of fold change with a log2 fold-change increase of 5.4 and 3.9 respectively—with some tumors exhibiting changes from near undetectable levels to high expression (Fig. 4c). These two genes are more commonly expressed in basal cancers, with PROM1 being a cancer lineage stem cell marker and perhaps negatively regulated by E2 stimulation (Supplementary Figure 7) [56]. Other genes with significant log2 fold changes > 1 included drug targets such as FGFR4, KIT, IGF1R, and BCL-2 (Table 2). NDRG1, a particularly compelling candidate since it also showed upregulation in LTED breast cancer models, was further interrogated using METABRIC data. Like PROM1 and KLK7, NDRG1 is most highly expressed in basal breast cancers; yet, when expressed in ER-positive primary tumors, NDRG1 confers significantly worse disease-specific survival outcomes (Supplementary Figure 8). Notably, none of the top differentially regulated genes in ESR1-depleted recurrences showed statistically significant changes in expression in non-ESR1-depleted recurrences (Supplementary Figure 9). To determine the prevalence of ESR1-depleted in other endocrine-resistant cohorts, we analyzed RNA-seq expression data from the MET500 and our own WCRC cohorts. Although we found most metastatic tumors with PAM50 luminal classifications had high expression of ESR1 as expected, there was a subset of metastatic luminal tumors which harbored very low levels of ESR1 expression (Fig. 4d)—suggesting ESR1 depletion in endocrine-resistant luminal tumors is not restricted to local recurrences and may also be a feature of more advanced disease.