Genomic characterization of de novo metastatic breast cancer

Abstract

Background:

In contrast to recurrence after initial diagnosis of stage I-III breast cancer (recurrent metastatic breast cancer, rMBC), de novo metastatic breast cancer (dnMBC) represents a unique setting to elucidate metastatic drivers in the absence of treatment selection. We present the genomic landscape of dnMBC and association with overall survival (OS).

Experimental Design:

Targeted DNA sequencing (OncoPanel) was prospectively performed on either primary or metastatic tumors from 926 patients (212 dnMBC; 714 rMBC). Single-nucleotide variants, copy number variations and tumor mutational burden (TMB) in treatment-naïve dnMBC primary tumors were compared to primary tumors in patients who ultimately developed rMBC, and correlated with OS across all dnMBC.

Results:

When comparing primary tumors by subtype, MYB amplification was enriched in triple-negative dnMBC vs. rMBC (21.1% vs. 0%, p=0.0005, q=0.111). Mutations in KMTD2, SETD2 and PIK3CA were more prevalent, and TP53 and BRCA1 less prevalent, in primary HR+/HER2− tumors of dnMBC vs. rMBC, though not significant after multiple comparison adjustment. Alterations associated with shorter OS in dnMBC included TP53 (wild-type: 79.7 months; altered: 44.2 months; p=0.008, q=0.107), MYC (79.7 vs. 23.3 months; p=0.0003, q=0.011) and cell-cycle (122.7 vs. 54.9 months; p=0.034, q=0.245) pathway genes. High TMB correlated with better OS in triple-negative dnMBC (p=0.041).

Conclusions:

Genomic differences between treatment-naïve dnMBC and primary tumors of patients who developed rMBC may provide insight into mechanisms underlying metastatic potential and differential therapeutic sensitivity in dnMBC. Alterations associated with poor OS in dnMBC highlight the need for novel approaches to overcome potential intrinsic resistance to current treatments.

Introduction

Despite improvements in treatment options, metastatic breast cancer remains largely incurable and is responsible for over 600,000 deaths annually around the globe.¹ The Cancer Genome Atlas (TCGA) and other groups have identified recurrent genetic alterations in breast cancer, though only TP53, PIK3CA, and GATA3 are present at greater than 10% frequency across all subtypes combined and tested samples included mostly primary tumors in patients with early breast cancer, many of whom will likely be cured of their disease.^2,3

Genomic characterization of metastatic tumors has demonstrated differences compared to matched primary tumors as well as to primary tumors in TCGA and other similar datasets.^4–8 However, outside of breast cancer subtype (e.g. hormone receptor and HER2 status), genomic features associated with the early development of metastatic disease have not been well elucidated to date.

Although only 3–6% of all patients diagnosed with breast cancer in high-resource countries present with de novo metastatic breast cancer (dnMBC), overall dnMBC accounts for 20–50% of all metastatic breast cancer cases worldwide, depending upon tumor subtype.^9–12 dnMBC represents a unique setting to elucidate metastatic drivers in the absence of treatment selection. Prolonged overall survival (OS) from the time of metastatic diagnosis has been reported in dnMBC compared to patients with recurrent metastatic breast cancer (rMBC).^10,12,13 It remains unclear whether differences in tumor biology between dnMBC and rMBC drive the earlier onset of metastatic disease in dnMBC and whether there are genomic features in dnMBC that confer a survival advantage compared to rMBC. A next-generation sequencing (NGS) panel encompassing several hundred genes was introduced at the Dana-Farber/Brigham and Women’s Cancer Center in 2013. Since its introduction, the consistent genotyping of metastatic patients with dnMBC and rMBC has allowed for comparison between the two and for adequate follow-up to assess the impact of different genomic alterations on patient outcomes. Here, we sought to evaluate whether treatment-naïve primary tumors in patients with dnMBC harbor distinct genomic features compared to primary tumors in patients who later develop rMBC, and to determine the prognostic value of alterations identified in dnMBC.

Methods

Patient Selection and Clinical Data Abstraction

The Dana-Farber/Harvard Cancer Center (DF/HCC) institutional review board (IRB) approved the protocols (DF/HCC IRB #11–104 and #17–000, PROFILE) which allow tumor genomic profiling with a NGS panel (OncoPanel) and linkage with clinical outcomes. In addition, a dedicated protocol (DF/HCC IRB #17–482) allowed for collection of additional clinical data linked to OncoPanel data in patients with breast cancer. We included all patients with metastatic or locally advanced unresectable breast cancer who provided written informed consent to DF/HCC IRB #11–104 and who underwent successful sequencing with OncoPanel of either primary or metastatic tumor between June 26, 2013 and August 31, 2016 to allow for sufficient follow-up time for survival (Supplementary Fig. S1A). Patients with breast cancer who did not have evidence of metastatic disease were excluded. Only one sample per patient was included in the analysis; thus, data were not available from paired primary and metastatic tumors.

In patients initially diagnosed with stage I-III breast cancer who subsequently developed rMBC, we classified primary tumors as those collected as part of initial treatment (diagnostic tissue biopsy or surgical resection) and involving breast and/or ipsilateral axillary lymph nodes. In dnMBC patients, primary (breast and/or ipsilateral axillary lymph nodes) and metastatic tumors were classified as treatment-naïve if the sample tested was collected no more than 30 days after starting first-line systemic therapy. Tumors were classified according to receptor status at time of initial diagnosis. For additional details, please see Supplementary Methods.

Genomic Analysis

OncoPanel testing

Starting in June 2013, tumor specimens were tested using OncoPanel to detect somatic mutations, copy number variations and structural variants. Genomic testing on formalin-fixed, paraffin-embedded (FFPE) tissue was performed centrally within the Center for Advanced Molecular Diagnostics at Brigham and Women’s Hospital, a CLIA-certified laboratory environment, according to published methods.¹⁴ Per protocol, paired germline samples were not analyzed. All FFPE tumor tissue samples underwent histopathologic review prior to DNA extraction to determine adequacy and tumor cellularity, and to mark areas for tumor enrichment. Only cases with at least 20% tumor cellularity were selected. Regions of adequate cancer of at least 3 mm in size were either manually dissected from unstained sections or cored directly from the paraffin block. DNA extraction was performed using standard methods.¹⁵ Test results were reviewed by laboratory staff and interpreted and reported by board-certified pathologists. 200 ng of DNA were used for library preparation (with a low input threshold of 50 ng). DNA was analyzed using a solution-phase Agilent SureSelect hybrid capture kit and an Illumina HiSeq 2500 sequencer, with 2×100 paired-end reads to a mean target coverage of 187X unique, high-quality, mapped reads per sample (range 50–844X; 50X minimum) required to pass. From August 2013 to July 2014, OncoPanel Version 1.0 included the full coding regions of 275 genes plus selected intronic regions of 30 genes for rearrangement detection. Starting in August 2014, OncoPanel Version 2.0 included the full coding regions of 300 genes plus selected intronic regions across 35 genes (Supplementary Tables S1–2). Data were analyzed by an internally-developed bioinformatics pipeline, as detailed below.

Filtering of OncoPanel data

OncoPanel is a targeted next-generation sequencing panel performed on tumor-only DNA without paired germline DNA control. The automated pipeline calls somatic nucleotide variants with Mutect (v1.1.4) (RRID: SCR_000559) and small insertion and deletions (indels) with the Genome Analysis Toolkit (GATK-version 1.6–5-g557da77) (RRID: SCR_001876). Since tumor tissues were tested without a paired normal, additional informatics steps were taken to identify common single nucleotide polymorphisms (SNPs), as previously described.¹⁶ Variants were filtered to exclude those that occur at a populational frequency of greater than 0.1% in the Exome Sequencing Project (ESP) database (RRID: SCR_012761) and those in an in-house panel of control samples (non-neoplastic FFPE liver and blood samples) were also filtered. Any filtered variants that were reported in the COSMIC database (RRID: SCR_002260) more than twice were rescued and presented for manual review. During manual review, variants with evidence of being recurrently identified in tumors or other in vitro data that showed biological significance are assigned “tiers” based on a 5-tier system. Any synonymous mutations or mutations found in 3´ and 5´ untranslated regions were excluded. Although the majority of germline variants are removed, a small number of rare germline variants remain in the final dataset.

Single-nucleotide variants (SNV) and gene copy number variations (CNV):

For SNV and gene CNV calling, only genes common to both OncoPanel Versions 1.0 and 2.0 were considered. For comparison of the frequency of SNV between two groups, genes that were found to be mutated in at least 1% of the overall cohort (n=177 genes) were included. First, an unfiltered SNV analysis was performed, in which all mutations in a given gene were included except synonymous, intronic and UTR variants. A second exploratory analysis was performed according to the putative functional implication of the mutation based on OncoKB annotation. Lists of tumor suppressor and oncogenes along with specific annotated variants were downloaded from OncoKB on April 15, 2019 (v1.19) (RRID: SCR_014782). Of 277 genes that were covered in both OncoPanel Versions 1.0 and 2.0, OncoKB annotation was available for 221 genes. We included all single-nucleotide and in-frame insertion and deletion variants annotated as oncogenic or likely oncogenic in OncoKB, along with known loss-of-function mutations (splice site, frameshift, nonsense or nonstop mutations) in genes listed in OncoKB as tumor suppressor genes. For gene CNV, high amplifications and deep deletions in a given gene were defined as previously described.¹⁴ Low level gene copy number gain or loss were excluded.

Oncogenic cell pathway analysis:

Exploratory analyses by oncogenic cell pathways were performed for SNV and CNV in each gene if their functional consequence matched previously reported gain-of-function or loss-of-function annotation (Supplementary Table S3).¹⁷ Genomic alterations in genes were distributed into the following cell pathways: Receptor tyrosine kinase (RTK)-Ras, PI3K, TP53, Cell-cycle, Myc, Notch, Wnt, TGF-β, NRF-2 and HIPPO.

Arm-level copy number changes

Arm-level copy number variant calls were generated using a previous version of an in-house algorithm specific for panel copy number segment files (ASCETS; latest version available online at https://github.com/beroukhim-lab/ascets) followed by manual review. Briefly, a threshold above which genomic regions could be determined to be amplified or deleted was determined by modeling the noise in the copy number segments in the cohort. Chromosome arms were classified as amplified or deleted if more than 70% of the bases within the covered regions of each arm were altered at a level above the identified threshold.

Statistical Analysis

Odds ratio (OR) was computed to determine the association between groups of interest and the mutant or wild-type status of the gene. Patients were classified as altered or wild-type based on the presence of any mutation (coding non-synonymous mutations and mutations predicted to potentially disrupt splicing) or CNV in a given gene or cell pathway depending on the comparison of interest. Enrichment in a given group was determined through Fisher´s exact test for categorical data. For continuous data, a two-tailed unpaired t-test or Mann-Whitney U-test was used to determine if there was a significant difference between two groups. A separate exploratory analysis was also performed according to the putative functional implication of the mutation based on OncoKB annotation. All single-nucleotide and in-frame insertion and deletion variants annotated as oncogenic or likely oncogenic in OncoKB and known loss-of-function mutations in genes listed in OncoKB as tumor suppressor genes were included.

Overall survival (OS) was defined as time from diagnosis of metastatic breast cancer to death or censored at the last known vital status date. Kaplan-Meier product limit estimates were reported for groups defined by stage, tumor characteristics, and genomic alterations. Under the Cox proportional hazards model, the log-rank Score test was used to determine whether there was a significant difference in OS between two groups, and medians of the groups and hazard ratio (HR) with 95% confidence interval (CI) were computed. For OS analyses that included the overall dnMBC group (n=212), a multivariate analysis was performed including receptor subtype as a covariate. For multivariate models, adjusted HR with 95% CI and Wald p-values for the relevant model coefficients were reported.

For each set of comparisons between groups of interest and genomic alterations, false discovery rates (FDR) were computed using Benjamini-Hochberg q-values.

Tumor mutation burden calculation

Tumor mutation burden (TMB) was computed among patients with OncoPanel results by counting the number of exonic mutations that were reported, summing this number for each patient, and normalizing each patient sum by the size of the exonic bait-set of the panel used (V1.0: 757,787 bases; V2.0: 831,033 bases). The size of the exonic bait-set increased over the first and second versions of the OncoPanel assay design (+9.67% for V2.0 when compared to V1.0). FDR adjustment for TMB analyses was not performed given the limited number of comparisons.

Signature analysis

For signature analysis, we expanded the dataset of mostly non-synonymous, exonic variants to include any SNV detected with at least 30 read coverage depth by the OncoPanel analysis pipeline. To clean this dataset of likely germline variants, we first removed variants present in the gnomAD database (accessed Jan 31, 2019; RRID: SCR_014964) at a population allele frequency greater than 0.1%¹⁸ or in the ClinVar database (accessed Dec 19, 2018; RRID: SCR_006169) with a Benign or Likely Benign annotation.¹⁹ We then removed variants that showed a germ-like Gaussian distribution (based on at least 4 variants), or artifact-like distribution (below 10% in all of its samples). We also removed variants present at a VAF between 45 and 55%, or greater than 95%; although these may contain some somatic variants with concurrent loss of heterozygosity, this VAF range also represents the majority of germline events. The mutational spectrum of these removed variants was very similar to the previously published human germline mutational spectrum.²⁰ After these extensive germline filtering steps, there were on average 14.8 variants present per sample.

We restricted our signature analysis to samples with over 15 variants (Supplementary Fig. S2), resulting in 681 samples. Signature analysis was performed using the multiple linear regression method deconstructSigs²¹ using the set of 30 well-established COSMIC signatures²² to detect presence of the apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like (APOBEC)-associated signatures (COSMIC Signatures 2 and 13).

Results

Study population

Clinicopathological features of the study cohort are summarized in Table 1 and Supplementary Table S4. A total of 926 samples were sequenced, with one sample tested per patient (n=212, dnMBC; n=714, rMBC). Median age at initial diagnosis in the dnMBC group was 50.6 years and 47.6 years in the rMBC group. In the patients who developed rMBC, the median time from initial diagnosis to metastatic diagnosis was 46.7 months (range 3.7–377.7 months), and 72 (10.1%) of the metastatic recurrences in the rMBC group occurred within the first 12 months of diagnosis. Ductal histology at diagnosis was more prevalent in rMBC than dnMBC (74.6% vs. 68.9%, respectively), in contrast to lobular carcinoma present in 16.5% of dnMBC and 9.9% of rMBC. Histologic grade at initial diagnosis did not significantly differ between dnMBC and rMBC, with approximately 50% of poorly differentiated tumors in both groups. In dnMBC, 168 samples were treatment-naïve (primary, n=145; metastasis, n=23) and 44 were obtained more than 30 days (max: 3,516 days) after starting first-line systemic therapy (primary, n=23; metastasis, n=21). In rMBC, patients had a primary (n=418), metastatic (n=273), local recurrence (n=11) or unspecified (n=12) site of disease tested. As expected,¹⁰ subtype distribution differed significantly between rMBC and dnMBC, including fewer triple-negative breast cancer (TNBC, 11.8% vs. 21.6%) and greater HR+/HER2− (64.2% vs. 53.1%) and HER2+ (24.1% vs. 15.1%) representation in dnMBC compared to rMBC.

Table 1.

Clinicopathological characteristics of study population according to stage at initial diagnosis.

Characteristic	Stage at initial diagnosis				All (n=926)		p-value
	Stage 0-III (n=714)		Stage IV (de novo MBC) (n=212)
	N	%	N	%	N	%
Age at initial diagnosis, median (range)	47.6a (17.7–82.7)		50.6 (18.4–83.8)		48.1 (17.7–83.8)		9.55e–05
Race
Caucasian	627	87.8	188	88.7	815	88.0	0.066
African American	26	3.6	12	5.7	38	4.1
Asian or Pacific Islander	21	2.9	8	3.8	29	3.1
Other	22	3.1	0	0.0	22	2.4
Unknown	18	2.5	4	1.9	22	2.4
Sex
Female	710	99.4	209	98.6	919	99.2	0.418
Male	4	0.6	3	1.4	7	0.8
Ethnicity
Spanish/Hispanic	28	3.9	7	3.3	35	3.8	0.264
Non-Spanish; non-Hispanic	630	88.2	195	92.0	825	89.1
Unknown	56	7.8	10	4.7	66	7.1
Histologic Subtype
DCIS	14	2.0	0	0	14	1.5	0.006
Invasive Ductal (IDC)	533	74.6	146	68.9	679	73.3
Invasive Lobular (ILC)	71	9.9	35	16.5	106	11.4
Mixed (IDC and ILC)	54	7.6	26	12.3	80	8.6
Tubular	2	0.3	0	0	2	0.2
Mucinous	4	0.6	1	0.5	5	0.5
Other	9	1.3	1	0.5	10	1.1
Unknown	27	3.8	3	1.4	30	3.2
Histologic Grade
Well differentiated	41	5.7	11	5.2	52	5.6	0.175
Moderately differentiated	255	35.7	88	41.5	343	37.0
Poorly differentiated	365	51.1	105	49.5	470	50.8
Unknown	53	7.4	8	3.8	61	6.6
Subtype according to HR/HER2 status at initial diagnosis (primary)
HER2+/HR+	72	10.1	35	16.5	107	11.6	2.65e–08
HER2+/HR−	36	5.0	16	7.5	52	5.6
HER2−/HR+	379	53.1	136	64.2	515	55.6
TNBC	154	21.6	25	11.8	179	19.3
Unknown	73	10.2	0	0	73	7.9
Subtype according to HR/HER2 status at metastatic diagnosisb
HER2+/HR+	39	5.5	35	16.5	74	8.0	<2.2e–16
HER2+/HR−	41	5.7	16	7.5	57	6.2
HER2−/HR+	340	47.6	136	64.2	476	51.4
TNBC	142	19.9	25	11.8	167	18.0
Unknown	152	21.3	0	0	152	16.4
Stage at initial diagnosis
DCISc	14	2.0	0	0	14	1.5	-
Stage I	128	17.9	0	0	128	13.8
Stage II	329	46.1	0	0	329	35.5
Stage III	237	33.2	0	0	237	25.6
Stage IV	0	0	212	100.0	212	22.9
Unspecifiedd	6	0.8	0	0	6	0.6
Disease-free interval (until diagnosis of metastatic recurrence)
De novo MBC (stage IV at diagnosis)	0	0	212	100.0	212	22.9	-
≤2 years	191	26.8	0	0	191	20.6
>2 years	523	73.2	0	0	523	56.5
Type of sample tested
Primary	418	58.5	168	79.2	586	63.3	4.13e–07
Metastatic Recurrence	273	38.2	44	20.8	317	34.2
Local Recurrence	11	1.5	0	0	11	1.2
Unspecified	12	1.7	0	0	12	1.3
OncoPanel Version
1.0	150	21.0	40	18.9	190	20.5	0.561
2.0	564	79.0	172	81.1	736	79.5
Vital status
Alive	476	66.7	159	75.0	635	68.6	0.027
Dead	238	33.3	53	25.0	291	31.4

^a.Median age at initial diagnosis of breast cancer was 47.6 years in patients who later developed rMBC. In this group, the median age at diagnosis of metastatic recurrence was 53.6 years.

^b.In three patients who presented with initial diagnosis of HR+/HER2− dnMBC, OncoPanel was performed on a metastatic site of disease (collected > 180 days after initial diagnosis) that was TNBC (n=2) or HR+/HER2+ (n=1). In two patients with HR+/HER2+ dnMBC at initial diagnosis, OncoPanel was performed on a metastatic sample (collected >180 days after initial diagnosis) that was found to be HR−/HER2+.

^c.Patients with history of DCIS prior to the diagnosis of metastatic breast cancer (n=14) had a metastatic site of disease tested with OncoPanel.

^d.Cases are confirmed as early-stage invasive breast cancers, however, there was insufficient information available to classify by anatomic stage.

Overall Landscape of Mutations and CNV in Patients with Metastatic Breast Cancer

Before analyzing differences between dnMBC and rMBC, we performed a descriptive, non-comparative overview of the landscape of mutations and CNV in the entire metastatic breast cancer cohort. Across all 926 samples, a total of 6,119 mutations were detected. Median variant allelic fraction (VAF) was 36.6% (range: 2.4–100; IQR: 19.4–50.2). The mutation and CNV landscape according to stage at initial diagnosis (0-III, rMBC patients; IV, dnMBC patients), breast cancer subtype, and type of sample tested (primary tumor versus recurrent/metastatic site of disease) is shown in Figures 1–2 and Supplementary Tables S5–7.

Figure 1. Frequency of genomic alterations observed in samples tested with OncoPanel according to subtype at diagnosis.

1A) Gene mutations per breast cancer subtype at initial diagnosis. The top 15 most frequently mutated genes in the overall cohort and additional genes of interest are displayed. On the left, alterations observed in patients with DCIS/stage I-III breast cancer at diagnosis who later developed recurrent metastatic breast cancer (rMBC) (n=641; 73 samples excluded due to unspecified subtype at initial diagnosis). On the right, alterations observed in patients with stage IV breast cancer (de novo metastatic breast cancer, dnMBC) at diagnosis (n=212). Oncogenic mutations are depicted opaque in each subtype color-coded bar; variants of unknown significance (VUS)/passenger mutations are partially transparent.

1B) Gene copy number variations (CNV) per breast cancer subtype at initial diagnosis, including the top 15 most frequently altered genes and additional genes of interest. High amplifications (A) and deep deletions (D) are displayed for all genes (ERBB2 single copy number gain not included).

Figure 2. Mutation and copy number variation (CNV) landscape in patients with metastatic breast cancer tested with OncoPanel: de novo (dnMBC) and recurrent (rMBC).

2A) Co-mutation/gene CNV plot for samples tested with OncoPanel with clinical annotation for subtype and type of sample tested (n=914; 12 samples from overall cohort excluded due to unspecified type of sample tested). Distribution of breast cancer subtype at initial diagnosis (based on HR/HER2 status), stage at diagnosis (DCIS/stage I-III [on the left] vs. stage IV [on the right]) and type of sample tested (primary vs. recurrent/metastatic tumor) are depicted at the top of the figure, in addition to tumor mutational burden (TMB, adjusted by log-scale). Top 15 genes with the highest frequency of alterations and other genes of interest are displayed on the left (genes covered by both OncoPanel V1.0 and V2.0).

2B) Heatmap of gene CNV across samples tested with OncoPanel with clinical annotation for type of sample tested (n=914).

Descriptive analysis of all 212 dnMBC samples (regardless of type or timing of sample tested, and unfiltered by putative functionality), showed that TP53 was frequently mutated in TNBC (23/25, 92.0%), HR−/HER2+ (12/16, 75.0%), and HR+/HER2+ (23/35, 65.7%), and less common in HR+/HER2− tumors (27/136, 19.9%). In HR+/HER2− tumors, the most frequently mutated genes were PIK3CA (57/136, 41.9%) and CDH1 (33/136, 24.3%). Of the 33 CDH1-mutated samples, 31 were of lobular or mixed ductal/lobular histology. Across treatment-naïve HR+/HER2− samples (n=105), three (2.9%) activating ESR1 mutations (E380Q) were identified, in addition to another mutation (missense of unknown significance, L354Q) affecting the ligand-binding domain and one missense mutation in a non-functional domain, G278V; none were identified in untreated HR+/HER2+ dnMBC tumors (Supplementary Fig. S3).

Gene CNV in dnMBC and rMBC are included in Figures 1B and 2A. The heatmap of gene CNV according to stage at diagnosis is shown in Figure 2B. CCND1 amplifications in HR+/HER2− patients were observed in 17.6% (24/136) of dnMBC, and 16.6% (63/379) of rMBC. We detected 1q/16q loss most frequently in HR+/HER2− dnMBC (16/54, 29.6%), similar to HR+/HER2− tumors in rMBC (43/182, 23.6%). FGFR1 amplification was seen in 14.7% (20/136) of HR+/HER2− tumors and 4.0% (1/25) of TNBC in dnMBC, and in 10.8% (41/379) and 3.2% (5/154) in rMBC, respectively.

Genomic differences between primary and metastatic samples in patients with rMBC have been previously reported, particularly in HR+ breast cancer in which acquired alterations in the metastatic setting may reflect mechanisms of resistance to endocrine therapy.^6–8 Thus, we restricted enrichment analyses between dnMBC and rMBC to primary tumors, and excluded metastatic samples in rMBC patients. Also, due to differences in subtype distribution between rMBC and dnMBC which could confound comparisons between these groups as a whole, all comparative analyses between cohorts were subsequently performed according to receptor subtype.

NGS Detects Enrichment of Genetic Alterations in Treatment-naïve dnMBC

First, we aimed to investigate differences in tumor biology that may drive the earlier onset of metastatic disease in dnMBC compared to rMBC. To identify alterations enriched in dnMBC, we compared primary tumors from patients initially presenting with stage I-III breast cancer who later developed recurrence to untreated primary tumors in patients with dnMBC, and then within each breast cancer subtype (Supplementary Fig. S1B). As detailed above, given that the genomic landscape of primary tumors may differ from metastatic sites of disease, only primary tumors (breast or ipsilateral lymph node, as defined in the Methods) were included in this enrichment analysis. Also, considering that systemic therapy may lead to clonal selection, only treatment-naïve primary tumors from dnMBC patients (n=145, all breast cancer subtypes) were included.

In TNBC, MYB amplification was significantly enriched in de novo (n=19) vs. stage I-III (n=101) primary tumors (21.1% [4/19] vs. 0% [0/101]; p=0.0005, q=0.111). All MYB amplifications were identified as focal (Figure 3). In addition, we observed CIITA mutations enriched in de novo vs. stage I-III primary TNBC (26.3% [5/19] vs. 1.0% [1/101]; p=0.0003, q=0.233). However, CIITA mutations in triple-negative dnMBC were missense variants of unknown significance that did not affect functional domains.

Figure 3. MYB amplifications detected with OncoPanel in patients with de novo metastatic breast cancer (dnMBC).

Displayed here is the absolute copy number at each exonic probe on the 6q arm for each of the 5 dnMBC samples (4 TNBC; 1 HR+/HER2−) with MYB amplification. The dotted line represents the centromere.

Recognizing that our power after correcting for multiple testing is limited by sample size and the prevalence of alterations, we also present findings with unadjusted p-value <0.05, for hypothesis generation (Table 2). In HR+/HER2− tumors, KMT2D mutations were present in 14.6% (13/89) of dnMBC vs. 6.0% (14/235) stage I-III primaries (p=0.022). Mutations in SETD2 were also found at a higher frequency in dnMBC HR+/HER2− primaries (9.0% [8/89] vs. 2.1% [5/235], p=0.009). In contrast, proportionally fewer TP53 (21.3% [19/89] vs. 32.3% [76/235], p=0.056) and BRCA1 (0% [0/89] vs. 7.7% [18/235], p=0.005) mutations were observed in HR+/HER2− dnMBC. In HR−/HER2+ tumors, IKZF3 mutations were frequently present in dnMBC versus rMBC (23.1% [3/13] vs. 0% [0/22], p=0.044). When restricting the analysis to likely oncogenic mutations (OncoKB annotated hotspots or loss-of-function alterations in known tumor suppressor genes), only differences in TP53 (11.2% [10/89] vs. 25.1% [59/235], p=0.006) and PIK3CA (41.6% [37/89] vs. 29.8% [70/235], p=0.048) in the HR+/HER2− subtype reached p-value significance, suggesting the presence of a predominant luminal-A-like phenotype in HR+/HER2− dnMBC compared to luminal-B-like tumors in patients who develop rMBC.

Table 2.

Prevalence of genomic alterations in primary tumors of patients initially diagnosed with stage I-III breast cancer versus treatment-naïve primary tumors of patients diagnosed with de novo metastatic breast cancer (dnMBC). Alterations include gene mutations (without OncoKB filtering), gene copy number variations (high amplification or deep deletion) and chromosome arm level changes. Displayed here are genes/chromosome arms with unadjusted p values < 0.05 and other genes of interest; highlighted in blue are genes with adjusted q values < 0.25. Numbers of samples available for each comparison are displayed. Comparison of median TMB (mutations/megabase) is also included (Wilcoxon, significant p < 0.05).

SUBTYPE AND ALTERATIONS	Primary tumors in Stage I-III		Primary tumors in dnMBC		Unadjusted p value	FDR q value	Odds Ratio
HR+/HER2−
MUTATION	(n=235)		(n=89)
	N	%	N	%
BRCA1	18	7.7	0	0	0.005	1	Inf
SETD2	5	2.1	8	9.0	0.009	1	0.2
KMT2D	14	6.0	13	14.6	0.022	1	0.4
CDKN1B	4	1.7	6	6.7	0.029	1	0.2
HNF1A	3	1.3	5	5.6	0.039	1	0.2
BRD4	3	1.3	5	5.6	0.039	1	0.2
TP53	76	32.3	19	21.3	0.056	1	1.8
TMB (median)	7.220		7.220		0.167
CHROMOSOME ARM	(n=106)		(n=41)
	N	%	N	%
5p gain	12	11.3	11	26.8	0.040	0.980	0.4
20q gain	27	25.5	4	9.8	0.043	0.980	3.1
HR+/HER2+
MUTATION	(n=36)		(n=24)
	N	%	N	%
NOTCH1	0	0	3	12.5	0.059	1	0
TMB (median)	6.618		7.220		0.581
CHROMOSOME ARM	(n=16)		(n=13)
	N	%	N	%
15q gain	0	0	4	30.8	0.030	0.980	0
16q loss	7	43.8	1	7.7	0.044	1	8.7
HR-/HER2+
MUTATION	(n=22)		(n=13)
	N	%	N	%
IKZF3	0	0	3	23.1	0.044	1	0
ARID1B	1	4.5	4	30.8	0.052	1	0.1
TMB (median)	7.822		6.598		0.129
ALL HER2+
MUTATION	(n=58)		(n=37)
	N	%	N	%
CDC73	0	0	3	8.1	0.056	1	0
EPHA5	0	0	3	8.1	0.056	1	0
IKZF3	0	0	3	8.1	0.056	1	0
MTOR	0	0	3	8.1	0.056	1	0
TMB (median)	7.220		7.220		0.824
CHROMOSOME ARM	(n=26)		(n=21)
	N	%	N	%
16q loss	12	46.2	3	14.3	0.029	1	5.0
15q gain	0	0	4	19.0	0.034	0.980	0
7p gain	0	0	4	19.0	0.034	0.980	0
TNBC
MUTATION	(n=101)		(n=19)
	N	%	N	%
CIITA	1	1.0	5	26.3	3.3E-04	0.233	0.03
TMB (median)	7.220		7.220		0.846
HIGH GENE AMPLIFICATION	(n=101)		(n=19)
	N	%	N	%
MYB	0	0	4	21.1	4.7E-04	0.111	0
CHROMOSOME ARM	(n=45)		(n=9)
	N	%	N	%
5p loss	8	17.8	5	55.6	0.028	1	0.2

Similar results were observed in an exploratory analysis when comparing mutations and CNV in all treatment-naïve dnMBC samples (adding 23 untreated metastases) with primary tumors of rMBC patients (Supplementary Table S8). Genomic alterations found in untreated dnMBC metastases were sorted by cell pathway (Supplementary Fig. S4).¹⁷ Activating alterations in the MAP-kinase pathway associated with resistance to endocrine therapy, such as FGFR2 and NRAS amplification or truncating NF1 mutations,^5,23,24 were observed in treatment-naïve HR+ metastases (n=20), albeit these findings should be interpreted with caution given the sample size and that most of these alterations were detected in only one sample.

Novel Associations Between Genomic Alterations and Survival in dnMBC

Next, we studied the association between genomic features and patient outcomes in dnMBC, and aimed to determine whether there are specific alterations in dnMBC that may confer a survival advantage compared to patients with rMBC. At the time of the data cut-off analysis, 635 patients were alive and 291 had died in the overall cohort. Given that all patients in this cohort were diagnosed with metastatic disease, for survival analyses in the dnMBC and rMBC cohorts, OS was defined as the time from metastatic diagnosis to death or censored at the last known vital status date. The median distribution of time from metastatic diagnosis to date of last follow-up in patients who were alive (censored) at last follow-up, was 22.0 months (range: 0.3–252.6 months) in the rMBC cohort (476 patients alive at last follow-up) and 23.4 months (0.3–168.5) in the dnMBC cohort (159 patients alive at last follow-up).

Comparison of time from metastatic diagnosis to last vital status in rMBC and dnMBC confirmed significantly longer survival in dnMBC patients (Kaplan-Meier estimates of median OS: rMBC 59.9 months vs. dnMBC 78.6 months, log-rank p=0.0056; Supplementary Fig. S5A). Within subtypes, this OS improvement in dnMBC versus rMBC was seen in HR+/HER2−, with a similar trend in the HER2+ subtype but not in TNBC (Supplementary Fig. S5B). Mutations in the rMBC cohort that correlated with poor OS after FDR correction (q-value <0.25) were those found in TP53 and RB1, in contrast to patients with NOTCH1, ATM or GATA3 mutations who had improved OS (Supplementary Fig. S6). No significant OS difference was observed between rMBC with or without MYC amplification (46.5 vs. 57.0 months; p=0.533, q=0.815).

Recognizing that survival analyses for rMBC could be biased based on timing of metastatic biopsies and the impact of intervening treatments on genomic findings and outcomes,²⁵ we then restricted survival analyses to the dnMBC population. As expected, OS was longest for patients with HER2+ dnMBC (median: not reached, NR) and shortest for those with TNBC (median: 20.0 months).

Across all dnMBC, we confirmed the previously reported²⁶ correlation in MBC between TP53 mutation and worse OS (median OS: wild-type, 93.1 months; altered, 37.8 months; p=0.001; q=0.123) (Supplementary Fig. S7A, Supplementary Table S9). Several potential novel associations were identified on univariate analysis, including mutations in MLH1 (mismatch repair pathway) and FANCC (Fanconi anemia pathway involved in DNA damage repair) that predicted poor OS. Three of the five MLH1-mutant samples were hypermutated (TMB range: 9.683–12.104 mut/Mb) but none harbored indels in homopolymer regions that would suggest microsatellite instability. Mutations in KMT2D (gene encoding a histone methyltransferase involved in epigenetic modulation and, as described above, enriched in de novo HR+/HER2− primaries) predicted improved OS in dnMBC with a consistent trend in each subtype.

We observed significant associations between amplification of MYC, RAD21 and MYB, as well as deletion of CDKN2A/CDKN2B, and worse OS (q<0.1) (Supplementary Fig. S7A). Several gene CNV that predicted poor OS in dnMBC were located in 8q21–24 amplicon, including MYC (8q24.21), RAD21 (8q24.21), EXT1 (8q24.11), NBN (8q21.3) and PTK2 (8q24.3). Co-occurrence of these events suggests that the observed prognostic effect may not necessarily be driven by amplification of all genes (Supplementary Fig. S8). Copy number gain of IKZF3 (17q12-q21), associated with improved OS, always co-occurred (n=40) with ERBB2 (17q12).

3q loss was significantly associated with improved OS across the dnMBC cohort (median OS: altered, NR; wild-type, 59.6 months; p=0.001, q=0.032) (Supplementary Fig. S7B). Copy number changes per chromosome arm that were significantly associated with OS at a univariate level can be found in Supplementary Table S9, all of which were unique to dnMBC, with the exception of 17q loss which also predicted poor OS when detected in rMBC.

Considering that the distribution of receptor subtype could impact the association of alterations with OS in the overall dnMBC cohort, a multivariate analysis adjusting for receptor status was performed (Supplementary Table S10). Of the molecular alterations that were prognostic at the univariate level after correcting for multiple comparisons, TP53 mutation, amplification of RAD21, MYC, NBN, EXT1 and PTK2, and deletion of CDKN2A and CDKN2B retained significance after adjusting for breast cancer subtype.

Gene mutations and CNV were grouped by cell pathways as previously described¹⁷. We observed significantly worse OS among dnMBC with alterations in the following pathways: MYC (median OS: wild-type, 79.7 months; altered, 23.3 months; p=0.0003; q=0.011), Cell-cycle (122.7 vs. 54.9 months; p=0.034, q=0.245) and TP53 (79.7 vs. 44.2 months; p=0.008, q=0.107) (Figure 4B; Supplementary Fig. S9A). After correcting for subtype in a Cox proportional hazards model, the MYC and Cell-cycle pathways retained statistical significance (Wald p-value for MYC alteration model coefficient=2.7E-04; Wald p-value for Cell-cycle alteration model coefficient=0.024) but the TP53 pathway which, in addition to TP53, includes alterations in MDM2, MDM4, ATM and CHEK2, did not remain significant (Wald p-value for TP53 alteration model coefficient=0.126). When analyzing the association between OS and cell pathways within each breast cancer subtype, MYC alterations were significantly prognostic in HR+ subtypes, regardless of HER2, but not in TNBC. (Supplementary Fig. S9B).

Figure 4. Genomic alterations associated with overall survival (OS) in patients with de novo metastatic breast cancer (dnMBC) tested with OncoPanel.

4A) High tumor mutational burden (TMB) was associated with improved OS in triple-negative dnMBC. Left: Kaplan-Meier curves for OS according to TMB quartiles across the overall dnMBC cohort. Median OS (from lowest to highest TMB quartile): Q1, 79.7 months; Q2, 93.1 months; Q3, 78.6 months; Q4, 59.6 months (p=0.977). Right: Kaplan-Meier curves according to TMB quartiles in patients with triple-negative dnMBC. Median OS (from lowest to highest TMB quartile): Q1, 13.3 months; Q2, 21.8 months; Q3, 16.4 months; Q4, 42.6 months (p=0.041). Hazard ratio (HR) Q4 vs. Q1: 0.06 (95% Confidence Interval [CI]: 0.005–0.60); HR Q3 vs. Q1: 0.33 (95% CI: 0.06–1.60); HR Q2 vs. Q1: 0.48 (95% CI: 0.12–1.98).

4B) Kaplan-Meier curves for OS in patients with dnMBC according to the presence or absence of alterations in oncogenic cell pathways. Pathways that reached statistical significance in the overall dnMBC cohort on univariate analysis: MYC pathway (median OS wild-type [n=199]: 79.7 months; median altered [n=13]: 23.3 months; log-rank p=2.9E-04; q=0.011; HR 4.56, 95% CI 2.02–10.32); Cell-cycle pathway (median wild-type [n=161]: 122.7 months; median altered [n=51]: 54.9 months; p=0.034; q=0.245; HR 2.00, 95% CI 1.09–3.65); TP53 pathway (median wild-type [n=147]: 79.7 months; median altered [n=65]: 44.2 months; p=0.008; q=0.107; HR 1.78, 95% CI 0.85–3.74). Of these, after adjusting for subtype, MYC (Wald p-value for MYC pathway alteration model coefficient = 2.7E-04) and Cell-cycle (Wald p-value for Cell-cycle pathway alteration model coefficient = 0.024) pathways maintained significance.

Associations between individual genomic alterations and OS were also evaluated specifically within each receptor subtype group and are shown in Supplementary Table S11, although interpretation of results is limited by the number of altered versus wild-type samples, particularly in HER2+ and triple-negative dnMBC. We also investigated OS associations when restricting samples to dnMBC treatment-naïve primary tumors (n=145); similar trends in OS were seen for individual alterations and those grouped by oncogenic pathways (Supplementary Fig. S10), albeit several did not maintain significance after FDR correction, likely due to smaller sample sizes.

Tumor Mutational Burden (TMB) and Mutational Signatures

Median TMB was 7.2 mut/Mb in both dnMBC (range: 0.0–38.5) and rMBC (range: 1.2–108.3). In dnMBC, median TMB in each subtype was: HR+/HER2− (n=136) 7.2 mut/Mb, HR−/HER2+ (n=16) 6.0 mut/Mb, HR+/HER2+ (n=35) 8.4 mut/Mb, and TNBC (n=25) 8.4 mut/Mb. No significant differences in TMB were observed between primary stage I-III tumors and dnMBC treatment-naïve primary tumors across subtypes (Table 2). Comparison of treatment-naïve dnMBC samples (primary or metastasis) and metastatic samples in rMBC also revealed no significant differences in median TMB per subtype (data not shown).

Across all 926 patients, and within the dnMBC cohort, the top TMB quartile corresponded to a cutoff of ≥9.627 mut/Mb. In rMBC, there was no significant association between metastatic survival and TMB quartiles across the cohort (p=0.38) or within subtypes, including TNBC (Supplementary Fig. S11). Similar to rMBC, we did not observe consistent relationships between OS and TMB quartiles (p=0.977) across the entire dnMBC cohort. However, when analyzed by subtype, TNBC patients with tumors in the highest TMB quartile had improved OS compared to the other quartiles (p=0.041) (Figure 4A, Supplementary Fig. S12). A similar, non-significant trend was observed in HER2+ dnMBC, whereas the highest TMB quartile in HR+/HER2− patients had numerically inferior OS.

Finally, based on prior data suggesting an increase in APOBEC-mediated mutagenesis between primary and metastatic tumors in patients who develop rMBC,^8,27 we explored mutational signatures in dnMBC. No significant difference in APOBEC-induced mutations was noted between stage I-III primary tumors in rMBC patients and treatment-naïve dnMBC primaries, whether including all patients (12.9% [54/418] vs. 17.2% [25/145], respectively; p=0.165) or restricting to samples with higher mutational burden (>15 variants; 17.1% [50/292] vs. 21.6% [24/111]; p=0.248) (Supplementary Fig. S13A). The percentage of APOBEC-induced mutations was also not significantly different between treatment-naïve primary and metastatic dnMBC samples (Supplementary Fig. S13B).

Discussion

We report results from the molecular characterization of patients with de novo metastatic breast cancer compared to rMBC. Prior studies portraying the genomic landscape of metastatic breast cancer have primarily focused on rMBC and its comparison to untreated primary breast tumors,^5–8 with limited dnMBC cases across many publicly available datasets. Here, approximately 80% of dnMBC patients had a treatment-naïve sample tested allowing us to study genomic alterations in the absence of treatment selection and representing one of the largest DNA sequencing cohorts to date of treatment-naïve dnMBC to our knowledge.

MYB amplification was significantly enriched in triple-negative dnMBC primary tumors. In addition to the four treatment-naïve triple-negative dnMBC samples found to harbor MYB amplification, five other MYB amplifications were identified in the overall cohort, all in samples collected at the time of metastatic disease (treatment-naïve HR+/HER2− dnMBC, n=1; metastasis in HR+/HER2− rMBC, n=2; metastasis in TNBC rMBC, n=1) except for one primary HR+/HER2− sample in a patient who later developed rMBC. In preclinical models, c-Myb has been described to promote breast cancer invasion through activation of the beta-catenin pathway, and loss of c-Myb suppresses metastatic dissemination.²⁸ Our findings suggest that MYB amplifications may contribute to de novo metastatic potential, particularly in TNBC.

In primary HR+/HER2− tumors, we found a numerically greater prevalence of alterations in genes involved in epigenetic modulation (e.g. KMT2D, SETD2, BRD4) and fewer mutations in DNA damage repair genes (TP53, BRCA1) in dnMBC compared to rMBC. Functional annotation of hotspots demonstrated more PIK3CA-mutant and fewer TP53-mutant samples in dnMBC, suggesting a predominant luminal-A phenotype among HR+/HER2− dnMBC, in contrast to luminal-B-like tumors in patients who develop rMBC. Almost half of dnMBC treatment-naïve HR+/HER2− samples harbored a PIK3CA mutation, identifying potential candidates for the PI3K inhibitor alpelisib.²⁹

Mutational load influences survival across several cancer types in patients treated with immune checkpoint inhibitors.³⁰ In our cohort of over 700 rMBC patients, the top TMB quartile (cutoff: 9.6 mut/Mb) did not correlate with survival, regardless of subtype. In contrast, patients with triple-negative dnMBC whose tumor was in the highest TMB quartile had significantly longer OS compared to patients with lower TMB, whereas the highest TMB quartile in HR+/HER2− had numerically inferior OS. None of the dnMBC TNBC patients received first-line immune checkpoint inhibition. Interestingly, the subtype-specific association between high TMB and OS mimics the difference observed in early breast cancer, where increased tumor-infiltrating lymphocytes predict improved disease-free survival in TNBC and HER2+ subtypes, but not in luminal HER2- breast cancer.³¹ Integration of genomic data with immune infiltration, determined by transcriptomic analysis³² or histologic assessment, may help refine the selection of patients with high TMB who are more or less likely to experience better outcomes.

Improved OS in dnMBC compared to rMBC has been previously documented.^10,12,13 This was confirmed in our dataset across the overall dnMBC cohort and in HR+/HER2− dnMBC, with a similar trend in HER2+ subtypes. At DFCI, all patients with metastatic breast cancer are approached for participation in research OncoPanel testing to identify potentially actionable alterations for clinical trial eligibility and targeted therapies. The median age of patients seen at DFCI is younger than population-based estimates and might have contributed to longer OS in the cohort, although this would not have been specific to dnMBC. To date, most studies evaluating the correlation between genomic features and survival have been conducted in patients with early-stage breast cancer,^33,34 with limited data available on the prognostic and predictive value of these alterations when detected in the metastatic setting. In a previously published cohort of approximately 250 patients with MBC (less than 20% dnMBC) who underwent NGS, mostly on archival primary tumors, OS from time of MBC diagnosis was significantly shorter in the presence of TP53 mutations in the HR+ subtype but this was not observed in TNBC or HER2+ patients, and no other significant associations were noted.²⁶ A retrospective case-control study evaluating TCGA data from primary tumors in 17 treatment-naïve dnMBC patients and 49 rMBC patients found that dnMBC was more likely to be HR-positive and have decreased lymphocytic infiltrate, with downregulation of chemotaxis, TNFα and IL-17 signaling.³⁵ GATA3 and ABL2 alterations were associated with poor survival in dnMBC but not rMBC patients. However, small sample sizes and lack of adjustment by breast cancer subtype limit the interpretation of these findings, as immune infiltration, prevalence of gene alterations and clinical outcomes differ between breast cancer subtypes. Poor OS in rMBC compared to dnMBC could also be attributed to acquired mutations or CNV in the metastasis, that potentially reflect mechanisms of resistance to therapy that are not present in the primary tumor. To address this question, analysis of paired primary and metastatic tumors in a large rMBC cohort would be needed.

Here, we confirm that TP53 is prognostic in both rMBC and dnMBC HR+/HER2− patients, and a similar non-significant trend was observed in HER2+ but not TNBC. In addition, we report several novel associations, including alterations in MYC and cell-cycle pathways predicting poor OS in dnMBC. MYC overexpression in stage I-III breast cancer has been associated with decreased relapse-free and breast cancer-specific survival.^34,36 We found that MYC amplification correlated with poor OS in dnMBC but not in rMBC, indicating that the prognostic impact of MYC alterations may depend on the moment of detection, classifying high-risk patients at an earlier rather than later timepoint of the disease. Regardless of subtype, the prevalence of MYC amplifications did not differ between primary tumors in dnMBC versus rMBC and, thus, it is unclear if MYC drives earlier onset of metastatic disease or whether the difference in outcomes is mostly attributed to therapeutic drug resistance. Notably, RAD21 (also located in 8q24) was co-amplified in 42% (5/12) of MYC-amplified dnMBC, and RAD21 overexpression by immunohistochemistry has been described to confer chemoresistance in high-grade early-stage breast cancer.³⁷ Finally, we observed a significant association between 3q loss and improved OS in dnMBC, with a consistent trend across subtypes. Gene expression analyses have suggested that 3q genes may contribute to regulation of immune response pathways.^38,39 This region includes genes such as PIK3CA, PRKCI, CTNNB1 and BCL6 (Supplementary Table S12). In ovarian cancer models, PRKCI upregulates TNFα production, increasing myeloid-derived suppressor cells and depleting cytotoxic T-cells.⁴⁰ Active β-catenin decreases CD8+ T-cells in melanoma through suppression of CCL4 transcription and impaired dendritic cell recruitment.⁴¹ Loss of 3q may thus lead to a less immunosuppressive microenvironment that correlates with improved outcomes although this hypothesis requires further validation.

Strengths of our study include the large sample size and detailed clinical annotation, with high representation of dnMBC. In contrast to TCGA, all patients in our cohort, regardless of the type of sample tested, had metastatic disease at the time of OncoPanel testing. More recent studies have also emphasized patients with metastatic breast cancer.^6–8,42 However, some metastatic breast cancer genomic datasets do not specify stage at diagnosis. Others include dnMBC but, based on publicly available data, the timing of the primary or metastatic sample tested with NGS cannot be discerned and, therefore, results may reflect acquired alterations in a pretreated rather than treatment-naïve setting. The limited representation of treatment-naïve dnMBC and linked clinical annotation in publicly available datasets underscores the need for improved automated methods to facilitate electronic medical record abstraction and for collaborative data sharing efforts to help address relevant questions in smaller patient subsets.

Our study had several limitations. We focused only on mutations and CNV, and did not include analysis of tumor microenvironment, host or immune factors. Clinicopathologic features such as receptor subtype and histology differed between dnMBC and rMBC groups. To control for differences in tumor biology and prior treatment exposure, enrichment analyses were performed between primary tumors within each receptor subtype, which limited sample sizes and power to detect significant differences. Survival analyses performed in the overall dnMBC cohort were adjusted for receptor status to account for the breast cancer subtype distribution in this group.

Clinical implementation of OncoPanel involved testing of one sample per patient, and thus, we did not have access to paired primary and metastatic samples for this study. In the absence of paired tumor tissue samples, blood biopsies have the potential to capture the genomic heterogeneity at a given time point.⁴³ Future efforts could focus on cfDNA analysis in patients in whom blood was collected at diagnosis of dnMBC, and at later time points, to evaluate changes in cfDNA after exposure to subsequent lines of therapy in the dnMBC population.

Targeted NGS panel sequencing is a high-throughput, cost-effective tool that has been widely implemented in clinical practice to identify potential candidates for matched targeted therapies or as molecular prescreening for participation in clinical trials.⁴⁴ While targeted gene panels provide greater depth of coverage in regions of interest and have the advantage of generally shorter turnaround time of results, there are limitations compared to more comprehensive platforms such as exome or genome sequencing, which are mostly reserved for research purposes. Whole exome sequencing (WES) remains the gold standard for calculating TMB. Estimation of TMB with targeted gene panels is influenced by the size of the panel and more likely to overestimate TMB compared to WES, particularly in tumors with lower mutational burden.^45,46 Mutational burden derived from targeted gene panels with sufficient genomic coverage has been shown to significantly correlate with exome-derived TMB estimates, but the deviation between both methods significantly increases with panels with less than 0.5 Mb sequenced.⁴⁵ Both versions of OncoPanel are above this 0.5 Mb threshold (V1.0: 0.7 Mb; V2.0: 0.82 Mb), and significant correlation was observed in TCGA between exome- and panel-derived TMB in breast cancer samples when using whole exome mutations versus exome mutations in only OncoPanel-targeted regions to estimate TMB (Supplementary Fig. S14A). Panel-estimated TMB in TCGA breast cancer samples was also highly correlated when restricting exome variants to genomic regions covered by OncoPanel V1.0 compared with V2.0-targeted regions (Supplementary Fig. S14B). Similarly, in TCGA breast cancer dataset, APOBEC signature contribution derived from exome variants in only OncoPanel-targeted regions accurately predicted the presence or absence of APOBEC signature (compared to whole exome-derived APOBEC signature contribution) when restricting the analysis to samples with more than 15 variants (Supplementary Fig. S2). As the costs of sequencing decrease, replacement of targeted panel sequencing with WES has the potential to provide in-depth genomic data at a larger scale, and may improve the accuracy of TMB measurements and other mutational signatures.

An important limitation of OncoPanel testing is the lack of matched normal DNA. To account for this, additional informatics steps were taken to identify and filter common SNPs, followed by manual review by a molecular pathologist to consider tumor purity, ploidy, allele fraction, and tumor type contextual information. Despite filtering, results may contain rare germline variants not found in dbSNP/ExAC, which could explain the higher TMB compared to other studies. With tumor-only sequencing, it can also be challenging to distinguish somatic mutations from those associated with age-related clonal hematopoiesis.⁴⁷ Alterations found to be enriched in treatment-naïve dnMBC (e.g. greater prevalence of SETD2, KMT2D or BRD4 mutations) could be due to differences in the prevalence of clonal hematopoiesis (CH) mutations if, for example, dnMBC primary tumor samples were obtained from a cohort of older age than the time of primary tumor collection in the rMBC cohort. However, most dnMBC samples were collected at initial diagnosis, at a median age of 50.6 years (compared to the median age at diagnosis of early-stage breast cancer, 47.6 years, in patients who later developed rMBC). Furthermore, prior studies suggest that relatively few patients (5%) have CH-related mutations misattributed as tumor-derived in panel sequencing, making this phenomenon unlikely to have significantly impacted our findings.⁴⁷ Finally, given the heterogeneity of treatments received, we did not have the power to explore correlations between molecular alterations and specific systemic therapies. Considering the complexity of tumor subsets, frequency of alterations and therapies received, combining multiple datasets with relevant clinical annotation will be required to definitively answer these questions.

In summary, genomic profiling of treatment-naïve dnMBC breast cancer reveals differences compared to primary tumors of patients who developed rMBC. A higher prevalence of PIK3CA mutations in HR+/HER2− dnMBC versus rMBC supports early NGS testing in clinical practice to consider treatment with targeted PI3K inhibition, and suggests a functional role for PIK3CA in mediating metastatic spread. Ongoing clinical trials evaluating the efficacy of chemotherapy with or without immune checkpoint inhibition may clarify the prognostic and predictive role of high mutational burden across breast cancer subtypes. Alterations associated with poor OS in dnMBC identify a patient population in whom novel therapeutic approaches are warranted, including strategies to overcome potential intrinsic resistance to CDK4/6 inhibitors in patients with cell-cycle pathway altered genes. Future studies will also focus on specific pathways and treatments to further elucidate the relationships between data generated from multi-omics profiling and clinical outcomes.

Statement of Translational Relevance.

To date it is unclear whether differences in tumor biology drive the earlier onset of metastatic disease in de novo metastatic breast cancer (dnMBC) and whether there are genomic features in dnMBC that confer a survival advantage compared to recurrent metastatic breast cancer (rMBC). Integrating targeted DNA sequencing (OncoPanel) and detailed clinical annotation, distinct genomic features are found in treatment-naïve primary tumors of dnMBC patients compared to primary tumors in patients who ultimately developed rMBC. PIK3CA mutations are prevalent (41.9%) in HR+/HER2− dnMBC, suggesting a high proportion of potential candidates for PI3K inhibition among this subgroup. High mutational load correlates with improved survival in triple-negative dnMBC, but not rMBC. Alterations in TP53, MYC and cell-cycle pathway genes are associated with poor prognosis in dnMBC, identifying a patient population in whom novel therapeutic approaches are warranted, including strategies to overcome potential intrinsic resistance to currently approved treatments, such as CDK4/6 inhibitors.