Exhausted T cell signature predicts immunotherapy response in ER-positive breast cancer

Study objectives

The primary endpoints of this study were to define the optimal approach for epigenetic priming in breast cancer patients on the basis of overall response and to define safety and tolerability. The secondary endpoints were to assess duration of response and the impact of vorinostat on PD-L1 expression in ER-positive tumors. Finally, the exploratory endpoints include post-treatment assessment of TILs population, inflammatory T-cell signature and PBMCs histone acetylation as a pharmacodynamic response marker for HDACi treatment.

Baseline characteristics and patient disposition

Thirty-four heavily pretreated ER-positive breast cancer patients [median age 56 years (32–81)] were enrolled and received at least one dose of therapy. The median number of prior lines of systemic treatment in this metastatic setting was 5 (range 2–13), with 85% of patients having received at least 3 lines of therapy, including hormonal therapy. Only 2 patients had tumors with more than 1% of PD-L1 expression (Supplementary Fig. 1). Due to limited ORR and insufficient clinical benefit in an unselected patient population, the accrual was halted prior to completion of target enrollment, in accordance with the study’s protocol. Baseline characteristics for the patients are shown in Table 1.

Efficacy and toxicity analyses

Thirty-four patients were treated on this trial prior to stopping the study. Six of the 34 patients never received pembrolizumab due to progression prior to scheduled infusion (Fig. 1). One patient was taken off study at week 3, without evidence of progression, due to immune-related grade 3 hepatitis (Fig. 1). A complete response (CR) was seen in one of the 27 (3.7%) evaluable patients who received all three agents with a clinical benefits rate [CBR: CR, partial response (PR) and stable disease (SD) > 6 months] of 5/27 patients (18.5%) and an objective response rate of 1/27 (3.7%). None of the patients with clinical benefit had discernible PD-L1 staining in their tumor. The patient with the CR had her treatment stopped after 10 months due to a disabling stroke. Upon a follow up visit at 21.7 months, she had not progressed and was lost to follow up afterwards. The median duration of treatment for all patients was 3.4 months (range 1–22+ months). These findings suggested that epigenetic modulation or priming did not benefit this group of heavily pretreated, ER-positive patients. The two patients on study with detectable PD-L1 expression experienced no benefit from therapy.

Overall the combination treatment was well tolerated and the addition of the HDACi did not worsen pembrolizumab or tamoxifen adverse events. Grade 3/4 toxicities included elevated transaminitis (9%, including immune hepatitis requiring treatment discontinuation), fatigue (6%), hyponatremia (6%), thrombocytopenia (6%) and anorexia (3%), as well as a patient with a disabling stroke of unclear relatedness, requiring treatment discontinuation. Grade 2 immune-related toxicities included pneumonitis, hypothyroidism, colitis, fatigue and thrombocytopenia (Table 2).

Correlative analysis and immune response modulation

The role of TILs and partially exhausted PD-1⁺/CTLA4⁺ CD8⁺ T-cells²⁸ and their epigenetic modulation were part of preplanned correlative endpoints in blood and tumor samples for this study, as well as comprehensive T-cell immune-profiling. The proportion of responders (CR, PR and SD > 6 months) and duration of response was markedly improved in patients with an increased number of exhausted cytotoxic CD8⁺ T lymphocytes (CTLs defined as CTLA4⁺/PD-1⁺ CD8⁺ T-cells) in either blood or tumor, with longer median progression-free survival (PFS) (8.6 versus 2.8 months, HR 0.32, p = 0.0098) (Fig. 2a, b and Supplementary Fig. 2). Similarly, longer median OS was observed, but did not reach statistical significance (29.50 versus 15.23 months, HR = 0.51 and p = 0.12). The percentage of exhausted CTLs was significantly higher in patients with response versus non-responders (41.03 vs 3.93%, p < 0.0001) (Fig. 2c, d). Clinical benefit (CR, PR and SD > 6 months) and PFS strongly correlated with a higher percentage of partially exhausted CTLs in tumor samples (Fig. 2e). This correlation remained significant (with exhausted CTLs in responders and non-responders, 12.97% vs 5.47% respectively, p = 0.032) in peripheral blood samples (Supplementary Figs. 3 and 2f). Pearson correlation analysis of the relationship between peripheral blood and tumors was performed in both non-responders and responders. For the non-responders, a significant p-value of 0.0178 was found (Supplementary Fig. 4). None of the evaluable non-responders had a CTLA4⁺/PD-1⁺ exhausted CTL signature in tumor samples, rendering the positive and negative predictive value of the test at 100%. The negative predictive value for blood was 91%. On the other hand, correlative analysis could not be performed in the responders, due to lack of tumor tissue that was available for only 3 of the responders. The percentage of exhausted CTLs was not affected by the treatment in either responders or non-responders (Fig. 3a).

Several reports support the correlation of total TILs to immunotherapy response. The overall number of CD3⁺ cells in tumor samples was very low in all patients, with median number of CD3⁺ cells relative to tumor cells at baseline (median 0.17% range 0.005–1.6, N = 22), which was not affected by treatment (median 0.16%, range 0.03–0.56, p = not significant) (Fig. 3b). Similarly, CD45⁺ cell infiltration was low (median 0.22%, range 0.006–1.89%) (Fig. 3c).

The overall percentage of CD4⁺ and CD8⁺ T-cells was not affected by either pembrolizumab or vorinostat treatment and did not predict response in tumor (Fig. 3d).

Our data did not suggest activation or proliferation of CD8⁺ T-cells, measured by HLA-DR and Ki67 expression, respectively^28–30, either at baseline or following therapy in tumor or blood (Fig. 3e, f and Supplementary Fig. 5). PD-L1 tumor cell expression was absent in all but 2 patients (both non-responders) at baseline, and unlike observed in TNBC preclinical models²², no changes in PD-L1 expression were observed in any of the patients (Fig. 2a, Supplementary Fig. 1).

Epigenetic modulation of CD4⁺ regulatory T-cells (Tregs) and histone acetylation

Multiple preclinical studies suggested that epigenetic modulation could impact the efficacy of immune checkpoint inhibitors. HDAC inhibition, and in particular vorinostat, has shown to significantly reduce Treg numbers and activity^22,27. Previous reports have further linked the increased presence of Tregs in breast cancer with an aggressive phenotype and reduced OS^31–33. Our preclinical in vivo and in vitro studies showed that HDACi, including vorinostat, potentiated immune checkpoint inhibitors in in vitro models, reduced tumor burden and prolonged the survival of tumor bearing 4T1 mice by decreasing tumor infiltrating Tregs²². In this study, responders had more (but not reaching significance) Tregs at baseline (Foxp3⁺ CD4⁺: 15.8% vs 11.5%, p = not significant, Fig. 4a).

However, treatment significantly reduced activated Tregs in the tumor microenvironment: Foxp3⁺CTLA4⁺ CD4⁺ Tregs were reduced in all patients, from 11.8% to 2.9%, p = 0.0067 (Fig. 4b). This reduction was much more prominent in the patients with clinical benefit, with a decrease from 10.4% to 2.5% (p = 0.034) in non-responders compared to a drop from 20.3% to 4.2% (p = 0.031) in responders. Our preclinical data in mice suggested that this was due to vorinostat treatment²². Neither HLA-DR nor Ki67 expression in Tregs changed with vorinostat treatment, suggesting the treatment directly affected Treg numbers, without modulating their activity, that was comparable in responders and non-responders (Fig. 4c, d).

To further associate a causal role of vorinostat-induced depletion of Tregs and its relevance to immune checkpoint inhibitor response in this study, we correlated sustained histone H3 and lysine acetylation in peripheral blood mononuclear cells (PBMCs) with clinical benefit and time to progression^{24,25,34–36}. Using acetylation as a pharmacodynamic measure of HDAC activity, our data further supported that both increased histone H3 and lysine acetylation significantly correlated with time to progression (Fig. 4e, f).

Preclinical in vivo data further suggested a select expansion of CD8⁺ T-cells in the tumor microenvironment²², however, this was not seen in patients. CD8⁺ T-cells were very few in all tumors at baseline, which did not increase with vorinostat treatment at the approved dose. Similarly, treatment did not affect tumor residing or peripheral CD3⁺, CD4⁺, CD8⁺ T-cells or their ratio in tumor or blood (Figs. 3d, 5a, b). Our data suggested a strong correlation between response and an increase in exhausted CD8⁺ T-cells in both tumors and blood. However, peripheral blood Tregs were low (2.95% versus 1.54% responders versus non-responders, respectively, p=not significant) and neither their number nor activity were affected by vorinostat on day 12 or after 3 cycles of therapy (Fig. 5c–f).

Patient selection

Pre and postmenopausal women or men with metastatic ER-positive breast cancer, after progression on at least one hormonal therapy in the metastatic setting and any number of prior chemo- or hormonal therapies, including tamoxifen, were eligible for this trial. Measurable disease, defined by RECIST 1.1⁴⁸, and pre and post treatment biopsies were required if tumors were accessible and safe for biopsy, together with an ECOG Performance Status ≤ 2 and adequate organ function (absolute neutrophil count ≥ 1.5 ×10⁹/L; hemoglobin ≥ 9 g/dL; platelets ≥ 100 ×10⁹/L; AST and ALT ≤ 2.5 × Upper Limit Normal (ULN) or ≤ 5.0 × ULN if liver tumor is present; serum total bilirubin ≤ 1.5 × ULN; Serum creatinine ≤ 1.5 × ULN, or 24-h clearance ≥ 60 ml/min). All participants provided written informed consent. The protocol, the proposed informed consent form and all forms of participant information related to the study were reviewed and approved by the UCSF CHR (UCSF Institutional Review Board), the Helen Diller Family Comprehensive Cancer Center Site Committee and by the Protocol Review Committee (PRC).

The trial was initially designed to enroll 87 patients. Using an Optimal Simon’s two-stage design⁴⁹, the null hypothesis (H_0) that the true response rate is 10% was to be tested against a one-sided alternative. The trial protocol stipulated that in the first stage, 10 patients would be accrued, if there were 0 or 1 objective responses at 24 weeks in these 10 patients, accrual to this arm would be stopped. Otherwise, 19 additional patients would be accrued for a total of n = 29, with an estimated ORR ≥ 30%. This study was designed to have an alpha of 0.0471, and a power of 81%.

The first and last patient were enrolled in May 2015 and January 2017 respectively. As the trial enrolled 34 patients it became clear that unselected patients would not benefit. A decision was made by the study team and accepted by the institutional review board and data safety committee that it would not be in the best interest of the patients to continue this study for lack of efficacy.

Study treatment, drug administration, safety and response assessment

Patients received 400 mg of vorinostat once daily orally for 5 days/week and 20 mg tamoxifen daily in 21-day cycles. Pembrolizumab at 200 mg was given intravenously on day 1, week 1 on cycle 1 (Arm A) or cycle 2 (Arm B) to investigate potential benefit of epigenetic priming to vorinostat until disease progression or unacceptable toxicity. To allow for epigenetic priming, patients initially received vorinostat and tamoxifen for 3 or 9 weeks prior to pembrolizumab dosing. This changed to 3 weeks or no priming due to very rapid progression in the first 5 patients who came off the study before receiving pembrolizumab.

Toxicity was assessed continuously according to the NCI Common Terminology Criteria for Adverse Events Version 4.0 (CTCAE v4.0).

Correlative studies: histone acetylation

Histone H3 and total lysine acetylation were evaluated in PBMCs obtained at baseline, on day 10(+2) of cycle 1 and at the end of cycle 3 (day 17–19) post-treatment, as a pharmacological marker of HDACi activity. Briefly, cells were spun at 2500 rpm for 10 min, separated from plasma and the pellets were stored at −80 °C until analysis. Once all the samples for one patient were collected, lymphocytes and red blood cells were separated with lysis buffer (BD Pharm Lyse™), following antibody staining to perform flow cytometry analysis. The following antibodies were used: Ghost/Dye v450 viability dye (#13-0863 Tonbo Bioscence); anti-human acetyl histone H3 (Cell signaling cod. #9649); anti-human Acetyl-Lysine (Cell signaling cod. #9441); followed by secondary antibody goat anti-rabbit IgG H + L-AlexaFluor488 (Life Technologies, cod. A11008). Flow cytometry staining protocol was based on a variation of the method of Ronzoni et al.⁵⁰. Briefly, lymphocytes were incubated in Ghost/Dye v450 for 30 min at 4 °C in the dark, then fixed with Paraformaldehyde (PFA) for 15 min at 4 °C in the dark and permeabilized for 5 min at RT in a solution of 0.4% triton X-100 in 0.1% BSA in PBS 1×. Cells were then incubated in primary antibodies, followed by secondary antibody, both for 1 h at RT in the dark. Cells were re-suspended in cold PBS-EDTA 3 mM and kept at 4 °C until analysis. Samples were acquired on BD FACSVerse Flow cytometer with BD FACSuite software. Gates were determined using isotype control antibody staining; live/dead, positive and negative gates were based on II only non-specific staining.

Correlative studies: immunophenotyping

Multiparameter flow cytometry was performed on pretreatment (baseline) and post-treatment (cycle 3 day 17–19) fresh biopsies of metastatic sites, when accessible and safe, and on PBMCs (collected on pretreatment, day 10–12 of cycle 1 and day 17–19 of cycle 3).

Freshly isolated samples were minced and digested overnight with buffer consisting of collagenase type IV (Worthington Biochemical Corp., #4188;), DNAse (Sigma-Aldrich, #DN25), 1% penicillin-streptavidin (CCF, #INVZR124), 1% 1 nM Na-pyruvate (Millipore, #TMS-005-C), 1% HEPES (Gibco, #15630-080), 1x non-essential amino acids (Gibco, #11140-050) and 11nM-mercapto-ethanol (Sigma-Aldrich, #M6250) in RPMI medium (Gibco, # 21870092) and collected as single-cell suspensions. PBMCs were suspended in 10% FBS, 1% HEPES, and 1% penicillin-streptavidin in RPMI medium. Approximately 3 × 10⁶ cells were characterized by a multiparameter flow cytometry analysis performed on a BD LSRFortessa with BD FACSDiva software and analyzed by FlowJo software (BD Biosciences). The flow cytometry immunophenotyping panel was designed based on previous findings²⁸ and includes the antibodies used, listed in Supplementary Table 1. To standardize voltages over time, Sphero Ultra Rainbow Beads (Spherotech) were used to calibrate and normalize to baseline intensity. Gates were determined using fluorescent minus one (FMO), isotype control antibodies staining and an internal negative control cell population (PD-1 and CTLA-4 expression on CD3- CD4- CD8- cells). Gating strategy is shown in Supplementary Fig. 6.

PD-L1 staining by immunohistochemistry

H&E staining and PD-L1 (clone SP263; #790-4905, Roche) protein expression on freshly isolated tumor tissue from pre-treatment and post-cycle 3 biopsies were performed immunohistochemically by the UCSF Tissue Core. A study-dedicated pathologist blindly performed the scoring of PD-L1, evaluating the % of positive cells present in each tissue (Supplementary Fig. 1).

Efficacy assessment and statistical methods

The primary endpoints were the evaluation of the safety and tolerability of the combination treatment, the evaluation of the objective response (ORR), defined as complete response (CR) and partial response (PR), and clinical benefits rate [CBR = CR + PR + stable disease (SD) > 6months] by RECIST 1.1 and irRECIST criteria. Secondary endpoints included progression-free survival (PFS), immune response, and overall survival (OS). Correlative endpoints include epigenetic modulation and immune modulation as well as their interaction, together with the evaluation of changes in histone acetylation.

The study employed a 2-arm Simon two-stage design with early stopping rules. All statistical evaluations were performed with Prism software version 8.0 (GraphPad Software, Inc). Data are expressed as mean with standard deviation (±SD) indicated. Data were assessed for normal Gaussian distribution with Shapiro–Wilk test and then analyzed by Mann-Whitney test, Kruskal–Wallis test, Wilcoxon test, ANOVA test or T test, as appropriate. The level of statistical difference was defined as *p ≤ 0.05; **p ≤ 0.01; ***p ≤ 0.001; ****p ≤ 0.0001; NS = not significant.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.