Clinical Efficacy and Whole-Exome Sequencing of Liquid Biopsies in a Phase IB/II Study of Bazedoxifene and Palbociclib in Advanced Hormone Receptor–Positive Breast Cancer

CDK4/6 inhibitors (CDK4/6i) with endocrine therapy (ET) are the standard of care in hormone receptor–positive/HER2-negative (HR⁺/HER2⁻) advanced breast cancer (1). The cyclin D1–dependent kinases CDK4 and CDK6 phosphorylate the pRB tumor suppressor, enabling progression through the cell cycle. Cyclin D1 is a key transcriptional target of the estrogen receptor (ER), and the convergence of the ER transcriptional network and cyclin D1-CDK4/6-Rb1 axis contributes to the synergistic activity of ET and CDK4/6i, such as palbociclib (2–4). This synergy is a key component of the clinical efficacy of palbociclib, which has limited clinical benefit as a single agent (5).

The importance of the synergy between ET and palbociclib suggests that improved targeting of ER may enhance the efficacy of regimens combining ET and palbociclib. This notion is supported by recent studies demonstrating that several mechanisms of resistance to ET are also mechanisms of resistance to the combination of ET and palbociclib (6, 7). Importantly, in the randomized phase III PADA-1 trial, patients with HR⁺/HER2⁻ metastatic breast cancer with an ESR1-activating ligand-binding domain (LBD) mutation, which confers resistance to ET, had double the odds of disease progression while on treatment with an aromatase inhibitor (AI) plus palbociclib compared with patients with wild-type (WT) ESR1 (8). Furthermore, in the PADA-1 trial, patients with rising plasma circulating tumor DNA (ctDNA) ESR1 mutations had improved progression-free survival (PFS) when treatment was switched to fulvestrant and palbociclib versus continued treatment with an AI and palbociclib (9). In the PALOMA-3 trial, the Y537S ESR1 LBD mutation was enriched after the acquisition of resistance to fulvestrant alone or to the combination of palbociclib plus fulvestrant in metastatic disease (10). Collectively, these results suggest that better targeting of ER can improve outcomes with the combination of a CDK4/6i and ET, and the Y537S ESR1 mutation has a role in the acquisition of resistance to fulvestrant and palbociclib.

Bazedoxifene is a selective ER modulator (SERM)-selective ER degrader (SERD) hybrid (SSH; i.e., has ER degrader properties in addition to the SERM activity; refs. 11–13). Bazedoxifene has favorable effects on bone and lipid metabolism without adversely affecting the uterus (14, 15). It is the only SERM that is approved in combination with conjugated estrogens for the prevention of postmenopausal symptoms and osteoporosis based on large randomized clinical trials in healthy postmenopausal women (16). Preclinical studies showed that bazedoxifene has antitumor activity in HR⁺ breast cancer models, including models of resistance to tamoxifen, resistance to estrogen deprivation, and models harboring a Y537S ER LBD mutation (11, 12, 17, 18). Preclinical studies showed that bazedoxifene had greater inhibitory effects on cell growth and ER-mediated transcription compared with tamoxifen or fulvestrant in the presence of the Y537S ESR1 mutations (13, 19). Computational predictions and, more recently, the crystal structure of mutant LBD bound to bazedoxifene showed that the Y537S-mutant LBD bound to bazedoxifene leads to the formation of a new S537-E380 hydrogen bond that stabilizes helix 12 (H12) in the activating function-2 (AF-2) cleft of the LBD, and this enables the inhibition of coactivator binding that results in transcriptional inhibition. In contrast, several other SERMS and SERDs, including tamoxifen and fulvestrant, were shown to have limited transcriptional inhibitory activity in the presence of the Y537S mutation. This is likely because H12 is pushed out from the AF-2 cleft of the Y537S-mutant LBD after interacting with tamoxifen or fulvestrant, and this allows coactivator binding and consequently a decreased response to these drugs (13, 19).

Taken together, based on the preclinical data of bazedoxifene in endocrine-resistant HR⁺ breast cancer models, the safety profile of bazedoxifene in randomized clinical trials, and the rationale of combining palbociclib with next-generation ETs, we conducted the first clinical trial with the combination of palbociclib and bazedoxifene. This was a phase Ib/II study designed to investigate the efficacy and safety of palbociclib with bazedoxifene for the treatment of patients with HR⁺/HER2⁻ advanced breast cancer who had progressed on prior ET. In addition, we performed serial ctDNA whole-exome sequencing (WES) to study genetic mechanisms of resistance and the tumor clonal architecture and evolution during treatment with palbociclib and bazedoxifene.

We enrolled female or male patients ages 18 years or older with histologically confirmed invasive breast cancer that was metastatic or unresectable locally advanced disease. Primary or invasive disease was confirmed to be HR⁺ (ER >10% and/or progesterone receptor>10%) and HER2⁻ according the American Society of Clinical Oncology/College of Pathology guidelines (20) [defined as 0–1+ by IHC or FISH-negative (HER2 copy number < 6 and HER2/CEP17 ratio < 2.0)]. Measurable disease was required per RECIST 1.1 (21). Bone-only disease was allowed if there was evidence of lytic lesions (22). Progressive disease on prior ET was required; this was defined as relapse while on adjuvant ET or within 1 year of completion of adjuvant ET, or progression through at least one line of ET in metastatic or locally advanced breast cancer. Postmenopausal and premenopausal women who have been on a GnRH agonist were eligible. Prior treatment with up to two lines of chemotherapy in advanced disease was allowed. The protocol was amended to allow up to one line of prior chemotherapy when the 18th patient was enrolled. There was no limit to the number of prior endocrine therapies.

The study was a phase IB/II, single-arm, single-institution, and open-label study. The objectives were to evaluate the safety and antitumor activity of the combination of palbociclib and bazedoxifene. The primary endpoint was clinical benefit rate (CBR), defined as the proportion of patients with complete response (CR), partial response (PR), or stable disease (SD) at week 24 [per RECIST version 1.1 (21)]. A CBR of 20% or less would not be of clinical interest based on a conservative estimate of the percentage of patients who progressed on adjuvant ET or progressed on ET or palbociclib as a single agent in metastatic disease (5). Secondary endpoints included objective response rate, PFS, overall survival (OS), and safety. Objectives were determined in the intention-to-treat (ITT) population and the evaluable cohort who reached the first restaging timepoint.

All patients received bazedoxifene 40 mg orally once daily on days 1–28 and palbociclib 125 mg orally once daily on days 1–21 of a 28-day cycle. The bazedoxifene dose of 40 mg was based on published safety data for this dose (23). The first 6 patients enrolled completed a safety run-in, and accrual was paused for safety assessment after these patients completed at least one treatment cycle. Response and progression were evaluated using RECIST 1.1 guidelines (21). For patients with bone-only disease, evaluation was done per MD Anderson bone tumor response criteria (22). After a baseline scan, tumor response was assessed every 8 weeks, with a PR or CR requiring confirmation within 4–6 weeks. After cycle 6, tumor measurements were performed every 12 weeks. Safety was evaluated according to NCI-CTCAE version 4.0. Exploratory translational objectives were to investigate the efficacy of the combination of bazedoxifene and palbociclib in relation to genomic findings in ctDNA.

The study followed a Simon “optimal” two-stage design. On the basis of the published activity of palbociclib as a single agent, a CBR of 20% or less was considered not to be of clinical interest and was the null hypothesis, and a CBR of ≥ 30% was considered worthy of further investigation. The sample size of 37 patients was chosen to have 90% power to declare the combination effective at this rate while controlling for no more than 10% one-sided type I error under the null hypothesis. The first stage of the study included 17 evaluable patients, and accrual was continued to the second stage after the confirmation that at least 4 patients met the primary endpoint.

CBR was reported as a proportion and 95% confidence interval (CI) accounting for the two-stage design. Associations between CBR and clinical and/or correlative factors were assessed using the Wilcoxon test for continuous variables and Fisher exact test for categorical variables.

Median PFS in months was estimated using the Kaplan–Meier estimation method and compared between clinical and/or correlative factors via log-rank test. HRs were obtained using Cox proportional hazards models. All tests were two sided with a type I error of 0.05. All analyses were carried out via SAS, V9.4 (SAS Institute) and R, V4.0.2.

Whole venous blood (6–10 mL) was collected from patients into ethylenediaminetetraacetic acid vacutainer tubes (Becton and Dickinson). Blood was processed within 2 hours of collection. Whole blood was centrifuged for 10 minutes at 1,200 × g and the plasma supernatant was cleared by centrifugation for 10 minutes at 3,000 × g. Cell-free DNA (cfDNA was extracted from 2 to 7 mL of plasma and eluted into 40–80 μL of resuspension buffer using the Qiagen Circulating DNA kit on the QIAsymphony liquid handling system. Germline DNA (gDNA) was extracted using the Qiasymphony DSP DNA midi kit on the QIAsymphony liquid handling system. Quantification of extracted cfDNA and gDNA was performed using the PicoGreen (Life Technologies) assay on a Hamilton STAR-line liquid handling system. Library preparation was performed using a commercially available kit (KAPA HyperPrep Kit with Library Amplification product KK8504) and IDT's duplex UMI adapters. Unique 8-base dual index sequences embedded within the p5 and p7 primers (purchased from IDT) were added during PCR. Enzymatic clean-up was performed using Beckman Coultier AMPure XP beads. Samples were sequenced using Illumina sequencing technology.

Sequenced reads in both ultra-low passage whole-genome sequencing (ULP-WGS) and WES libraries were mapped to the human reference genome (hg19) with BWA-MEM (version 0.17.7; RRID: SCR_010910). PCR duplicates were flagged by MarkDuplicates in Picard tool (version 2.26.10; RRID: SCR_006525). ichorCNA (version 0.1; ref. 24) was run for estimating tumor purity estimation. We applied the hg19 panel of normals provided in the ichorCNA github repo (inst/extdata/HD_ULP_PoN_1Mb_median _normAutosome_mapScoreFiltered_median.rds).

Somatic mutations on the exome data were called by The Cancer Genome Analysis (TCGA) WES Characterization Pipeline (the CGA pipeline). The CGA pipeline first runs deTiN (25) and ContEst (ref. 26; RRID: SCR_000595) for estimating potential tumor-in-normal contamination and cross-patient contamination, and calls single-nucleotide variants (SNV) and indels with MuTect (ref. 27; RRID: SCR_000559) and Strelka (ref. 28; RRID: SCR_005109), respectively. Out of the 68 exome samples, two samples were processed as tumor-only due to lack of the matched normals. For the tumor-only runs, CalculateContamination in GATK4 (version 4.0.5.1; RRID: SCR_001876) was used for estimating the cross-patient contamination. For the indel calling step in the tumor-only runs, an unmatched normal sample was used for Strelka, which requires a matched normal input. Potential artifacts were filtered with two panels of normals (created from 298 normal samples and 8,334 TCGA normal samples, respectively), the BLAT (RRID: SCR_011919) read-realignment filter, and the read orientation bias filters. The mutations in the tumor-only runs were further filtered if the called mutations overlapped with the germline mutations with the population allele frequencies ≥5% in the gnomAD database (ref. 29; RRID: SCR_014964). Mutations were denoted as pathogenic or likely pathogenic based on OncoKB (ref. 30; RRID: SCR_014782).

Mutational signatures were extracted using two algorithms: (i) hierarchical Dirichlet process (HDP; https://github.com/nicolaroberts/hdp) based on the Bayesian hierarchical Dirichlet process, and (ii) SignatureAnalyzer based on Bayesian non-negative matrix factorization (31). The Gibbs sampler was run with 20,000 burn-in iterations (parameter “burnin”). With a spacing of 200 iterations (parameter “space”), 100 iterations were collected (parameter “n”). After each Gibbs sampling iteration, three iterations of concentration parameter sampling were performed (parameter “cpiter”). To deconvolute composite signatures and to equate obtained HDP signatures to reference signatures, we used an expectation-maximization algorithm to deconstruct these signatures into reference constituents (31). Because of the low burden of indels, we excluded a mismatch repair deficiency signature (SBS6) and homologous repair deficiency signature (SBS3) from the candidate list using SigFit (32).

Copy-number alterations in the WES samples were detected by GATK-CNV (version 4.1.2.0; RRID: SCR_001876) with -number-of-changepoints-penalty-factor = 2.0. As the postfiltering steps, a panel of normals created from 298 normal samples was used, and potential germline copy-number alterations were discarded. The allelic copy-number alteration outputs from ModelSegments in GATK-CNV (version 4.1.2.0) were used for estimating tumor purity with ABSOLUTE (ref. 33; RRID: SCR_005198). The estimated tumor purity between the exomes (ABSOLUTE) and the ULP-WGS data (ichorCNA) was highly correlated (Pearson correlation coefficient = 0.98). Tumor mutation burden (TMB) was calculated as the number of coding mutations per Mb of covered exons.

The genomic sites of detected mutations across serial WES in a patient were force-called by investigating the number of reference and alternate bases at the sites. Somatic copy-number alterations (SCNA) in the serial WES were also grouped across the timepoints. PhylogicNDT was run with 500 iterations of Markov chain Monte Carlo Gibbs sampler on the force-called mutations and SCNAs to assign mutations to clonal clusters for estimating the phylogeny (34).

The study was conducted in accordance with the International Conference on Harmonization Good Clinical Practice Standards and the Declaration of Helsinki. Institutional review board approval was obtained at Dana-Farber/Harvard Cancer Center (DF/HCC). The study was registered in ClinicalTrials.gov (NCT02448771). The DF/HCC Data and Safety Monitoring Committee, which is composed of clinical specialists with experience in oncology and who had no direct relationship with the study, reviewed and monitored toxicity and accrual data from the study. Participants provided written informed consent prior to the performance of any protocol specific procedures or assessments. The study was an investigator-initiated trial funded by Pfizer. Palbociclib and bazedoxifene were supplied by the manufacturer (Pfizer). The funder had no role in data collection, data analysis, or data interpretation.

TCGA pan-cancer data are available through a data portal (https://gdc.cancer.gov/node/905/; https://gdc.cancer.gov/about-data/ publications/pancanatlas). The WES data from this study was deposited in dbGap accession number phs002802.v1.p1.

Between July 2015 and July 2017, we enrolled 36 patients with advanced HR⁺/HER2⁻ breast cancer (34 females and 2 males; Supplementary Fig. S1). The study met its primary endpoint after enrolling 36 patients. Patient demographics and baseline characteristics are in Table 1 and Supplementary Table S1 (Supplementary Table S1 includes the study representative information). All patients had ER⁺, HER2⁻ disease. The median age was 59.5 years (range, 37–79 years). Most of the patients had liver metastases (64%, 23/36) and 33% of patients (12/36) had lung metastases. Only 6% of patients (2/36) had bone-only disease. All patients had disease progression while on ET or disease recurrence within 1 year of completion of adjuvant ET. Most patients were heavily pretreated, with 50% of patients (18/36) having received two or more prior lines of ET and 52% of patients (19/36) having received up to two lines of chemotherapy in advanced disease. Nearly all patients (28/36) had disease progression while on treatment with an AI or disease recurrence within 1 year of completion of adjuvant treatment with an AI. There were 7 (19%) patients who received prior treatment with fulvestrant and 4 (11%) patients who received prior treatment with everolimus.

At the data cutoff in August 2020, with a median follow-up of 29.6 months, all patients had discontinued palbociclib and bazedoxifene treatment. A median of 4 cycles were given per patient (range, 1–34). The first 6 patients who were included in the safety run-in did not experience any dose-limiting toxicities (DLT) during their first treatment cycle. Thus, all patients were given 125 mg palbociclib for the remainder of the study. Neutropenia (61%) and fatigue (22%) were the most common adverse events associated with study treatment. A total of 16 patients (44%) experienced grade 3 neutropenia, and 1 patient (3%) experienced grade 4 neutropenia (Table 2; Supplementary Table S2). There were no deaths attributed to the study treatment. Most patients discontinued treatment due to disease progression (83%). Only 1 patient (3%) discontinued treatment due to protocol-specified unacceptable toxicity (neutropenia). Palbociclib dose reductions were required in 14 (38.9%) patients due to toxicity. Doses were held in 23 (63.9%) patients for a median of 10 days (range, 1–45 days). There were no bazedoxifene dose reductions, and bazedoxifene doses were held in 13 (36.1%) patients for a median of 12 days (range, 5–22 days; Supplementary Tables S3 and S4).

Of the 36 patients, 1 patient was unevaluable (this patient developed grade 4 neutropenia during the first cycle) and was included in the ITT analysis (Table 3). In the first stage of the study, 17 evaluable patients were enrolled; 5 of these patients achieved a clinical benefit (SD ≥ 24 weeks). The study continued to the second stage and accrued 19 additional patients. Accounting for the two-stage design, the best overall response in the ITT analysis was SD in 20 patients (20/36, 55.6%), progressive disease in 14 patients (14/36, 38.9%) and a confirmed PR in 1 patient. Among the patients with SD, 3 patients had unconfirmed PR (Fig. 1A). The CBR at ≥ 24 weeks was 33.3% (12/36; 95% CI, 9.8–56.4) accounting for two-stage design. This exceeded the predetermined boundary of at least 11 patients experiencing clinical benefit and therefore met the primary endpoint. These included 1 patient with a confirmed PR and 11 patients with SD ≥ 24 weeks. When including only evaluable patients, the CBR at ≥ 24 weeks was 34.3% (12/35; 95% CI, 20.7–58.9) accounting for two-stage design. Of the patients who achieved a clinical benefit at ≥ 24 weeks, 4 patients (4/36, 11.1%) had an exceptional clinical benefit with disease stability for ≥ 12 months (Fig. 1A and B; Table 3). The median PFS was 3.6 months (95% CI, 2.0–7.2; Fig. 1C), and the median OS was 26.5 months (95% CI, 20.7–not reached; Fig. 1D). There was no association between prior treatment regimens in the metastatic setting and the clinical benefit, including; prior tamoxifen treatment [prior tamoxifen, 30.8% (8/26) vs. no prior tamoxifen, 40% (4/10); P = 0.7, Fisher exact test for all comparisons in this section], prior fulvestrant [prior fulvestrant, 28.6% (2/7) vs. no prior fulvestrant, 34.5% (10/29); P = 0.8], or prior chemotherapy treatment [prior chemotherapy, 36.8% (7/19) vs. no prior chemotherapy, 29.4% (5/17); P = 0.7]. There was also no association between the number of prior treatment regimens in the metastatic setting and clinical benefit [0–1 lines of treatment, 33.3% (3/9) vs. ≥ 2 lines of treatment, 33.3% (9/27); P = 1.0]. However, liver metastases versus no liver metastases [17.4% (4/23) vs. 61.5% (8/13), respectively], or three to four organs with metastatic involvement versus two organs versus one organ with metastatic involvement [14.2% (3/21) vs. 60% (6/10) vs. 60% (3/5), respectively] were associated with decreased clinical benefit (P = 0.01).

Plasma samples were collected at baseline (n = 36), day 1 of cycle 2 (C2D1; n = 33) and at the time of disease progression at the end of treatment (EOT; n = 33) to identify genetic alterations associated with primary or acquired resistance to bazedoxifene and palbociclib (Supplementary Fig. S2A and S2B). All samples (n = 102) were subjected to ULP-WGS (mean coverage 0.19–0.57X) to determine cell-free DNA (cfDNA) tumor fraction (TF; see Patients and Methods) for the selection of samples with detectable ctDNA for subsequent WES, and to determine the relationship between the TF and clinical outcomes (Supplementary Fig. S3A). The median tumor fraction in all samples was 0.04 [range (0–0.85)]. Baseline cfDNA TF was higher in patients with liver metastases [median = 0.07 (0–0.85)] versus patients without liver metastases [median = 0.04 (0–0.3)] (Wilcoxon P = 0.01; Supplementary Fig. S3B) and patients who had three to four organs involved with metastases [median = 0.06 (0–0.85)] versus one organ involved with metastases [median = 0.03 (0–0.04)] (Wilcoxon P = 0.03). Moreover, baseline cfDNA TF correlated with PFS; patients with a high baseline cfDNA TF (≥0.1) had a lower PFS (n = 12, median PFS = 1.94 months) compared with patients with a low cfDNA TF (<0.1; n = 24, median PFS = 5.56 months; HR = 2.29; 95% CI, 1.08–4.89; P = 0.027, log-rank test; Supplementary Fig. S3C). There was no significant association between the changes in cfDNA TF after 1 month of treatment and PFS [ratio between C2D1/baseline TF > 1 (n = 12), median PFS = 2.04 months; vs. C2D1/baseline TF ≤1 (n = 19), median PFS = 3.67 months, HR = 0.9; 95% CI, 0.4–1.9; P = 0.71] (Supplementary Fig. S3D). Comparison of the TF at C2D1 and the EOT showed an increase in TF in most cases (58%, 18/31). The TF was unchanged in 7 patients (23%) and in 6 patients (19%) there was a decrease at the EOT. In most of the cases in which the TF decreased at the EOT, the absolute change was marginal and the TF either remained low (in the range of 0.03) or very high (above 0.45) except for one patient (patient 25) who had a decrease in the TF from 0.6 at C2D1 to 0.045 at the EOT. This patient had an increase in the size of the liver metastases but a mixed response of bone metastases with evidence of decreased fluorodeoxyglucose uptake in several bone metastases based on a PET-CT scan. Thus, although the early cfDNA TF dynamics were not indicative of PFS, the cfDNA TF was overall reflective of metastatic burden and in most cases of disease progression.

Samples with detectable ctDNA (cfDNA TF > 0.03) were subjected to WES (mean target coverage = 90–461X, median = 193X). WES was performed in 68 samples [baseline 28/36 (78%), C2D1 19/34 (56%) and EOT 23/32 (78%)] (Mutation Annotation Format in Supplementary table S5). These included 20 matched pairs of baseline and EOT samples. Sample-to-sample correlation of the somatic SNVs (SSNV) in the 68 samples showed that the matched samples from an individual patient had the strongest correlation in nearly all cases except for two samples that had a low number of mutations (Supplementary Fig. S4A). As expected, most of the SSNVs in exonic regions were unique to a single patient (4,205 of 4,227 nonsilent SSNVs, 99.5%; 3,190 of 3,211 silent SSNVs, 99.3%). The most frequent SSNVs were oncogenic hotspot mutations in ESR1 and PIK3CA (Supplementary Fig. S4B).

We next focused on known cancer-related genes (list of genes and the detected amino acid changes in Supplementary Table S6). At baseline, the most frequent mutations in the cancer-related genes included pathogenic or likely-pathogenic mutations in ESR1 (Fig. 2B), pathogenic PIK3CA mutations (Fig. 2D), EPAS1 (none of these are known to be pathogenic) and TP53 mutations (Supplementary Table S6). Other baseline pathogenic or likely pathogenic mutations identified in at least 2 patients included AKT1 and KMT2C mutations (Figs. 2A–C; Supplementary Table S6). Pathogenic ESR1 mutations were found in 32% (9/28) of baseline samples, with the missense mutations Y537 and D538 accounting for the majority of the ESR1 mutations (Fig. 2B). ESR1 mutations were detected in patients who had or did not have a clinical benefit at ≥ 24 weeks (Fig. 2A), and there was no association between the ESR1 mutations and PFS [WT median PFS = 3.6 months (n = 19), mutant median PFS = 2.0 months (n = 9); HR = 1.0; 95% CI, 0.4–2.4; P = 0.94, log-rank test for all comparisons in this section] (Fig. 2D). In contrast, baseline hotspot PIK3CA mutations were not detected in patients who had a clinical benefit at ≥ 24 weeks (Fig. 2A), and patients with a baseline PIK3CA mutation had a lower PFS compared with patients with WT PIK3CA at baseline [WT median PFS = 3.9 months (n = 22), mutant median PFS = 1.8 months (n = 6); HR = 0.2; 95% CI, 0.06–0.6, P = 0.0019] (Fig. 2E). This difference remained significant in a multivariate analysis correcting for significant clinical variables including the presence of liver metastases versus no liver metastases and 1 versus three to four organs with metastases (mutant vs. WT HR = 3.82; 95% CI, 1.18–12.33; P = 0.025; Supplementary Table S7). In addition, patients who had a baseline truncating mutation, including mostly pathogenic or likely pathogenic mutations in tumor suppressor genes (TP53, MEN1, RB1, CDKN1B, NF1, TP53, BP1, TP63, SMAD2, SMAD4, ARID1A, and KMT2C), had a worse PFS in a univariate analysis [no truncating mutation PFS = 3.8 months (n = 21), truncating mutation median PFS = 1.7 months (n = 7), HR = 0.3; 95% CI, 0.1–0.7; P = 0.006] and in a multivariate analysis (mutant vs. WT HR = 3.01; 95% CI, 1.09–8.3; P = 0.034; Fig. 2F; Supplementary Table S7).

Consistent with the baseline ctDNA, the most common SSNVs in cancer-related genes at the EOT were PIK3CA (32%, 7/22) and ESR1 (27%, 6/22), followed by PTEN, TP53, and GATA3 (Fig. 2G). Comparison of the matched baseline and EOT ctDNA pairs showed a high correlation between the frequency of the mutations found in at least two different patients (r = 0.7, P = 0.01, Pearson correlation; Fig. 2H). However, when comparing EOT ctDNA versus baseline ctDNA in individual patients, several patients gained actionable driver mutations at the EOT that are suggestive of putative mechanisms of acquired resistance and offer potential therapeutic targets. As examples, one patient (patient 30) who was on treatment for 7 months gained two BRAF missense mutations (G469A and D287H) and 2 hotspot PIK3CA mutations (E365K and N1044K). A second patient (patient 10) was on treatment for 3 months and gained pathogenic mutations in ERBB2 (S310Y) and PIK3CA (H1047L), and a missense mutation in FGFR2 (L309V; Supplementary Fig. S5).

SCNA analysis was performed in 30 plasma ctDNA samples from 16 patients with a TF ≥ 0.15 (baseline n = 12, C2D1 n = 9, EOT n = 9; Fig. 3A). Two patients (patient 1 and patient 14) were found to have genome doubling, which is associated with poor outcomes (35). In all samples, there were SCNAs in at least two cancer-related driver genes. In addition, in all samples we identified SCNAs in at least one gene related to a potential mechanism of resistance to ET and/or CDK4/6i. These included SCNAs in CCND1, CCNE2, MYC, ESR1, CDKN2B, CDKN2A, FOXA1, PIK3CA, AURKA, AURKB, CDKN1A, FGFR1, CDK6, CDK4, IGF1R, or PTEN (7, 36–38). Amplifications or gains in CCND1 and other genes that are within the same genomic region on 11q13.3 (FGF19, FGF3, and FGF4) were among the most common SCNAs. In addition, ELF3 amplifications or gains were identified in most of the patients (69%, 11/16). In line with the high frequency of ELF3 SCNAs that we identified, amplifications of ELF3 are more frequent in ER⁺/HER2⁻ metastatic breast cancers in the Metastatic Breast Cancer Project (39, 40) versus the ER⁺/HER2⁻ primary breast cancers in TCGA cohort (ref. 41; 31%, 22/70 in metastatic tumors vs. 6%, 29/471 in primary tumors, P < 0.0001, Fisher exact test). Overall, amplifications and losses at baseline or C2D1 compared with EOT within the matched pairs were comparable. Only a small number SCNA were acquired at the EOT. As an example, one patient (Patient 4) acquired losses of CDKN2A and CDKN2B (genes that encode the cell-cycle inhibitors P16 and P15, respectively, and are both in the same region in 9p).

We next assessed additional genomic features, including TMB (42, 43) and mutational signatures (44, 45) in the 30 ctDNA WES with sufficient TF. These features are associated with breast cancer outcomes and could potentially predict sensitivity or resistance to bazedoxifene plus palbociclib. The median TMBs at baseline (n = 12) and EOT (n = 9) were comparable and low [baseline median of 1.3 mutations/Mb, range (0.74–19.5); EOT median of 0.87 mutations/Mb, range (0.73–12.0); Wilcoxon P = 0.84] (Fig. 3B) and there was no significant difference in the TMB in patients who had a clinical benefit compared with patients who did not across all timepoints (Wilcoxon P = 0.7).

As for the mutational signatures in the cohort (Fig. 3C), the most common mutational signatures (SBS1, SBS5, and SBS40) can be attributed to clock-like and aging processes. The patients with a high TMB (Patient 2, 10, and 30) had enriched APOBEC signatures (SBS2 and SBS13). The same set of APOBEC signatures was detected in patients 21 and 31. Overall, most of the patients with APOBEC signatures (4 of 5 patients) did not have a clinical benefit from bazedoxifene and palbociclib. Albeit a small number of patients, these results are consistent with previous studies showing that the APOBEC signature is associated with endocrine resistance (44) and a high TMB (46). We also detected the mutational signature (SBS17b) that is associated with exposure to chemotherapeutic agents, 5-fluorouracil and capecitabine in a patient who received treatment with capecitabine followed by eribulin prior to enrollment on the study.

To better understand how tumor heterogeneity and the subclonal evolution contribute to treatment resistance, we performed an evolutionary analysis (see Patients and Methods). Seven patients had sufficient ctDNA (cfDNA TF ≥ 0.15) from three timepoints or at baseline and EOT (Fig. 4A–D; Supplementary Fig. S6A–S6C). In at least 3 of these patients the clonal evolution helped to shed light on potential mechanisms of treatment resistance. Patients 17 (Fig. 4A) and 2 (Fig. 4B) had clonal PIK3CA and ESR1 pathogenic mutations. Patient 17 had a subclone with losses of the tumor suppressor genes, RB1 and PBRM1, and the abundance of this subclone increased at the time of disease progression [cancer cell fraction (CCF) of 0.10 at baseline increased to 0.52 at EOT]. Patient 2 had two major subclone branches (light blue and red in Fig. 4B). Interestingly, within less than 2 months at the time of disease progression, the subclone with an ESR1 D538G and an NF1 truncal mutation increased from a CCF of 0.37 at baseline to 1. Loss of neurofibromin was shown to be a mechanism of treatment resistance in HR⁺ breast cancer (47). On the other hand, the second subclone with a SMAD2-truncating mutation and VHL missense mutation decreased from a CCF of 0.45 at baseline to 0.02 at EOT and is likely not a key driver of the disease progression.

Patient 1 had rapid disease progression within less than a month and was found to have a clonal pathogenic RB1 mutation (E54*) at baseline and at EOT (Fig. 4C). Although this RB1 mutation is monoallelic, it has been shown that the loss of even one copy of Rb1 can lead to genomic instability (48). Patient 1 and patient 5 had clonal ZFHX3 (zinc finger homeobox 3) mutations (Fig. 4C and D). In addition, three other patients in our study had nonsilent ZFHX3 mutations (clonal status could not be evaluated in these other patients), making ZFHX3 a common mutation in our dataset [16% (5/31)]. In contrast, ZFHX3 mutations, mainly deletions and truncating mutations, were found in only 4% of the ER⁺ breast cancers in TCGA cohort (39–41).

We provide the first study report of the clinical efficacy and safety of bazedoxifene in combination with palbociclib for the treatment of endocrine-resistant HR⁺/HER2⁻ advanced breast cancer. We observed SD as the best response in 56% of the ITT patient population and 57% in the evaluable patient population. The CBR ≥ 24 weeks was 33% and 34% in the ITT and the evaluable patients, respectively, in a heavily pretreated patient population. The CBR at 24 weeks exceeded the predefined threshold, meeting the primary endpoint of the study.

In this study, >50% of the patients had prior chemotherapy in the metastatic setting, and >40% of the patients had at least two prior lines of ET. Prior chemotherapy is associated with decreased benefit from subsequent treatments (49, 50). In addition, most of the patients had liver metastases, which is also associated with worse outcomes in HR⁺ advanced breast cancer (51). Importantly, the results of this study compare favorably to a phase II study of palbociclib as monotherapy in patients with HR⁺ advanced breast cancer, in which the CBR ≥ 24 weeks was 21% (7/33; ref. 5), providing evidence for the clinical efficacy of bazedoxifene in this setting.

The need for new ETs when combined with a CDK4/6i that better target the ER mutations in advanced breast cancer is supported by the recent results from the PADA-1 clinical trial. In PADA-1, switching from an AI plus palbociclib to fulvestrant plus palbociclib in patients with HR⁺ advanced breast cancer and rising levels of ctDNA ESR1 mutations, resulted in improved PFS (9). In retrospective analyses, the presence of an ESR1 mutation did not lead to inferior outcomes with the combination of fulvestrant and a CDK4/6i (52). However, preclinical studies (53) and clinical studies (54, 55) showed decreased activity of single-agent fulvestrant in the presence of the ESR1 mutations, particularly the Y537S ER mutation. On the other hand, preclinical studies showed that bazedoxifene has superior activity in inhibiting the Y537S ER mutation compared with fulvestrant and biophysical and structural studies provided mechanistic insights explaining these differences (13, 18). In line with these preclinical studies, our clinical trial showed that the presence of an ESR1 mutation was not associated with decreased benefit from the combination of bazedoxifene and palbociclib. Taken together, these published clinical and preclinical studies, along with the results from our clinical trial support a future randomized clinical trial comparing the combination bazedoxifene plus a CDK4/6i versus fulvestrant plus a CDK4/6i in patients with advanced breast cancer who have rising levels of ctDNA ESR1 mutations.

The EMERALD trial was a randomized phase III study investigating the new a SSH, elacetsrant, versus standard-of-care ET (AI or fulvestrant; ref. 56). In this trial, patients with HR⁺ advanced breast cancer who progressed on prior treatment with ET and a CDK4/6i and were randomized to receive elacestrant had improved PFS compared with patients that received standard-of-care ET. The magnitude of this improvement was higher in patients that had an ESR1 mutation. These results provided evidence for improved clinical outcomes with a single-agent next-generation ET, particularly in patients that have disease that remains ER dependent and support further development of this class of drugs. However, the median PFS was limited in the EMERALD trial and future studies will need to identify the patients who continue to benefit from ET after treatment with CDK4/6i and investigate novel treatment combination that incorporate the next-generation ETs. Increasing the dose of ET is another potential strategy to improve efficacy from ET, especially in the presence of the Y537S ER mutation, which is characterized by decreased affinity to multiple SERMs and SERDS (19, 57). Because bazedoxifene was well tolerated in large clinical trials with healthy women and in our study in advanced breast cancer, there were no significant gastrointestinal or other DLTs, a trial to investigate the pharmacokinetics, safety, and efficacy of higher doses of bazedoxifene in patients with HR⁺ advanced breast cancer and an ESR1 mutation is warranted.

We performed a prospective analysis of serial ctDNA WES to search for genetic determinants of sensitivity and resistance to the combination of bazedoxifene and palbociclib in advanced breast cancer after multiple lines of treatment. We showed that pathogenic PIK3CA mutations were associated with decreased PFS. Preclinical studies support the role of the PI3-kinase pathway in endocrine resistance (58, 59), and the α-specific PI3-kinase inhibitor alpelisib in combination with fulvestrant is approved for the treatment of PIK3CA-mutant HR⁺ metastatic breast cancer (60). A preclinical study showed that inhibition of PI3 kinase can prevent adaptive resistance to palbociclib, suggesting that the activation of the PI3-kinase pathway is involved in resistance to palbociclib (61). In the phase III PALOMA-III trial in which patients were randomized to fulvestrant plus palbociclib or fulvestrant plus placebo, there was an enrichment of plasma ctDNA PIK3CA mutations at the EOT. However, baseline plasma ctDNA PIK3CA mutations were not predictive of decreased PFS (10, 62). Our results suggest that in HR⁺ advanced breast cancer, the pathogenicity of the PIK3CA mutations increases in more advanced metastatic disease after multiple lines of treatment. This is in keeping with previous studies that showed that in metastatic disease, the presence of a PIK3CA mutation is associated with treatment resistance and tumor progression (63), whereas in early-stage, treatment-naïve HR⁺ breast cancer, PIK3CA mutations are associated with improved outcomes (64, 65).

The ctDNA WES analyses also identified genetic alterations that are potential novel drivers of disease progression and treatment resistance. As an example, we identified a high rate of ELF3 SCNAs. ELF3 is a member of the ETS transcription factors and is located within 1q32, a region commonly gained in breast cancer. In preclinical studies, ELF3 was shown to be a transcriptional repressor of ER, suggesting a possible role in endocrine resistance (66). We also identified a relatively high rate of ZFHX3 mutations. ZFHX3 encodes the transcription factor AT-motif binding factor-1 (ATBF1), which has key roles in protein–protein interactions and is a tumor suppressor in prostate cancer (67). In HR⁺ breast cancer, ATBF1 inhibits ER-mediated transcription (68), implying that aberrations in this gene may have a role in resistance to ET.

We acknowledge several limitations in our study, including the small number of patients that limits our ability to make strong conclusions about the efficacy of the combination of bazedoxfene and palbociclib, and the determinants of sensitivity or resistance. In addition, not all the plasma samples had sufficient ctDNA TF for TMB, SCNV, mutational signature, and phylogenetic analyses. Nonetheless, herein we demonstrate the utility and breadth of data that can be gleaned from WES of ctDNA. In summary, our results showed: (i) the safety of bazedoxifene in combination with palbociclib; (ii) a signal of the efficacy of bazedoxifene in HR⁺ breast cancer; (iii) PIK3CA mutations are a candidate biomarker of resistance to this combination that can be applicable to other next-generation ETs; and (iv) the advantages of WES of ctDNA that enabled us to capture the tumor heterogeneity and subclonal evolution during disease progression. WES also allowed us to identify potential novel genomic mechanisms of tumor progression that warrant future investigation.

I. Leshchiner reports personal fees and nonfinancial support from ennov1, LLC, and NoRD Bio, Inc and personal fees from PACT Pharma, Inc. outside the submitted work. S.M. Tolaney reports grants and personal fees from Pfizer during the conduct of the study as well as grants and personal fees from AstraZeneca, Merck, Genentech/Roche, Novartis, SeaGen, Eli Lilly, Gilead, Bristol-Myers Squibb, Eisai, and Sanofi; grants from Exelixis, Nektar, Cyclacel, and NanoString; and personal fees from Daiichi Sankyo, Athenex, CytomX, Certara, Mersana Therapeutics, Ellipses Pharma, 4D Pharma, Infinity Therapeutics, OncoSec Medical Inc, Chugai Pharmaceuticals, BeyondSpring Pharmaceuticals, OncXerna, Zymeworks, Zentalis, Blueprint Medicines, Reveal Genomics, ARC Therapeutics, Myovant, Umoja Buopharma, and Menarini/Stemline outside the submitted work. D.P. McDonnell reports personal fees from Pfizer and grants from NIH during the conduct of the study as well as grants, personal fees, and other support from Zentalis Pharmaceuticals and RAPPTA Therapeutics; other support from G1 Therapeutics and XRAD Therapeutics; and grants and personal fees from BMS outside the submitted work. In addition, D.P. McDonnell has a patent for a method of treating cancer using selective estrogen receptor modulators (US10,071,066, US10,420,734 issued, licensed, and with royalties paid from Radius Health) and a patent for lasofoxifene for treatment of breast cancer (US10,258,604, US10,905659, issued, licensed, and with royalties paid from Sermonix). I.E. Krop reports personal fees from AstraZeneca, Daiichi Sankyo, Genentech/Roche, Novartis, Merck, and Macrogenics and grants from Genentech/Roche and Pfizer outside the submitted work. E.P. Winer reports grants from Maor Foundation during the conduct of the study as well as other support from Carrick Therapeutics, Genentech/Roche, Jounce Therapeutics, and Leap Therapeutics outside the submitted work. G. Getz reports personal fees from Scorpion Therapeutics and grants from IBM during the conduct of the study as well as grants from Pharmacyclics outside the submitted work; in addition, G Getz is an inventor on patent applications related to MSMuTect, MSMutSig, MSIDetect, Polysolver, and SignatureAnalyzer-GPU and is a founder of, is a consultant for, and holds privately equity in Scorpion Therapeutics. R. Jeselsohn reports grants from Pfizer during the conduct of the study as well as grants from Lilly and personal fees from GE Healthcare outside the submitted work. No disclosures were reported by the other authors.

J. Tsuji: Formal analysis, investigation, methodology, writing–original draft, writing–review and editing. T. Li: Formal analysis, visualization, methodology, writing–original draft, writing–review and editing. A. Grinshpun: Data curation, investigation, visualization, writing–review and editing. T. Coorens: Formal analysis, visualization, writing–original draft, writing–review and editing. D. Russo: Formal analysis, visualization, writing–original draft, writing–review and editing. L. Anderson: Data curation, project administration, writing–review and editing. R. Rees: Data curation, writing–review and editing. A. Nardone: Data curation, writing–review and editing. C. Patterson: Data curation, project administration, writing–review and editing. N.J. Lennon: Supervision, writing–review and editing. C. Cibulskis: Supervision, writing–review and editing. I. Leshchiner: Formal analysis, writing–review and editing. N. Tayob: Formal analysis, supervision, validation, writing–review and editing. S.M. Tolaney: Supervision, writing–review and editing. N. Tung: Supervision, investigation, writing–review and editing. D.P. McDonnell: Conceptualization, investigation, writing–review and editing. I.E. Krop: Conceptualization, supervision, investigation, writing–review and editing. E.P. Winer: Conceptualization, supervision, investigation, writing–review and editing. C. Stewart: Formal analysis, supervision, writing–original draft, writing–review and editing. G. Getz: Conceptualization, supervision, investigation, visualization, methodology, writing–original draft, writing–review and editing. R. Jeselsohn: Conceptualization, supervision, funding acquisition, investigation, visualization, writing–original draft, writing–review and editing.

The study was conducted with funding from Pfizer. R. Jeselsohn is funded by the Claudia Adams Barr Program and an R01 CA237414-01.

We thank Myles Brown for helpful discussion. We thank Sam Pollock and Fanny Dao for assistance in sequencing data management. G. Getz was partially funded by the Paul C. Zamecnik Chair in Oncology at the Massachusetts General Hospital Cancer Center.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.