Immune checkpoint inhibitors in advanced gastroesophageal adenocarcinoma: a series of patient-level meta-analyses in different programmed death-ligand 1 subgroups

Abstract

Background

While the benefit of immune checkpoint inhibitors (ICI) is well established in programmed death-ligand 1 high (PD-L1_high) advanced gastroesophageal adenocarcinoma (GEAC), there remains significant controversy about their benefit in PD-L1_low GEAC. To elucidate the benefit of ICI in PD-L1_low and PD-L1_negative GEAC, we conducted an analysis leveraging individual patient data (IPD) extracted from Kaplan–Meier (KM) plots of pivotal trials.

Methods

KM curves from randomized clinical trials investigating the efficacy of ICI for advanced GEAC were extracted from published articles. IPD were extracted from the reported curves, and, in the case of unreported KM plots, KMSubtraction was used to retrieve survival data. A patient-level meta-analysis was conducted for PD-L1_low tumors.

Results

In the human epidermal growth factor receptor 2 (HER2)-negative setting, pooled PD-L1 combined positive score (CPS) 1-4 subgroup KM plots from KEYNOTE-859, CHECKMATE-649, and RATIONALE-305 showed a modest overall survival (OS) benefit with the addition of an anti-programmed cell death protein 1 (anti-PD-1) agent [hazard ratio (HR) 0.868, P = 0.018]. Similarly, a modest OS benefit was shown by our IPD meta-analysis of PD-L1 CPS 1-9 subgroups from KEYNOTE-859, KEYNOTE-062, and RATIONALE-305 (HR 0.840, P = 0.002.) Conversely, when CPS 5-9 subgroup KM plots from KEYNOTE-859 and RATIONALE-305 were pooled together, no significant OS benefit was found in the ICI-chemotherapy arm (HR 0.867, P = 0.181), although this subgroup was relatively small.

Conclusions

In PD-L1_low HER-2 negative GEAC, the benefit of first-line ICI is modest, yet significant. Further translational work is warranted to better select patients who could benefit from immunotherapy in this setting. Meanwhile, alternative therapeutic options such as zolbetuximab in Claudin18.2-positive disease must be taken into account.

Highlights

• •The benefit of first-line ICI in PD-L1_low HER2-negative GEAC is modest, yet significant.
• •The upfront choice of immunotherapy in patients with PD-L1_low should be discussed on a case-by-case basis.
• •Further translational work is warranted to better select patients who could benefit from immunotherapy in this setting.

Introduction

The standard of care for advanced human epidermal growth factor receptor 2 (HER2)-negative gastric/gastroesophageal adenocarcinoma (GEAC) involves an anti-programmed cell death protein 1 (anti-PD-1) agent in combination with doublet chemotherapy. However, there is notable variation in regulatory approvals across different countries. Based on the CHECKMATE-649 trial results, nivolumab plus chemotherapy was approved by the United States Food and Drug Administration (FDA) as the first-line treatment regardless of programmed death-ligand 1 (PD-L1) status,^1,2 and in patients with a PD-L1 combined positive score (CPS) ≥5 by the European Medicines Agency (EMA).³ The KEYNOTE-859 trial⁴ showed a statistically significant overall survival (OS) benefit with pembrolizumab in combination with platinum- and fluoropyrimidine-containing chemotherapy compared with chemotherapy alone in the same setting. The FDA has similarly approved this regimen regardless of PD-L1 status,⁵ whereas the EMA limited its approval for patients with a PD-L1 CPS ≥1.⁶ Based on the RATIONALE-305 trial results, the China National Medical Products Administration approved tislelizumab in combination with chemotherapy for patients with a PD-1 tumor area positivity ≥5%.⁷ More recently, provisional results of GEMSTONE-303 confirmed a benefit in OS from the addition of sugemalimab to chemotherapy in patients with a PD-L1 ≥5%, but not in the PD-L1 5%-9% subgroup.⁸ Notably, data from the original publications of these randomized controlled trials (RCTs) suggest that the magnitude of benefit progressively increases with higher PD-L1 cut-offs. However, data from KEYNOTE-859 on the CPS 1-4 and 5-9 subgroups, which were only subsequently published in the EMA documents,⁹ appear to contradict this observation.

With regards to HER2-positive GEAC, based on the first interim analysis of KEYNOTE-811, pembrolizumab in combination with trastuzumab and chemotherapy was initially approved by the FDA regardless of PD-L1 status. However, on 7 November 2023, based on the lack of benefit in PD-L1-negative patients shown by two subsequent interim analyses, the FDA restricted the approval to patients with CPS ≥1,¹⁰ thus aligning with the EMA approval.¹¹ However, the Kaplan–Meier (KM) plots for the subgroup of patients with PD-L1 CPS <1 have not been reported in a peer-reviewed journal yet and survival data at higher CPS cut-offs are not available.

Currently, data suggest there is little role for immune checkpoint inhibitors (ICIs) in patients with PD-L1-negative tumors, whereas the optimal treatment strategy when PD-L1 expression is low remains uncertain. In addition, KM plots for OS and progression-free survival (PFS) in different PD-L1 subgroups were not always reported.

To elucidate the survival benefits within the subgroup of patients with PD-L1_low GEAC, we conducted multiple analyses leveraging individual patient data (IPD) extracted from the reported and reconstructed KM plots of pivotal trials. In this study, we defined PD-L1 CPS 1-9 (and/or CPS 1-4 or CPS 5-9) as PD-L1_low to distinguish this subgroup from the subgroup of patients with PD-L1 CPS ≥10 (which we defined as PD-L1_high) for whom the efficacy of chemo-immunotherapy is ascertained. Our goal was to provide insights into the PD-L1 subgroups not previously reported.

Materials and methods

Search strategy, study selection, data extraction, and quality assessment

This study was conducted in line with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.¹² Search strategies were conducted on four online databases (PubMed, EMBASE, Web of Science, and Cochrane CENTRAL) from inception to 30 June 2024. The full search strategy is detailed in Supplementary Material 1, available at https://doi.org/10.1016/j.esmoop.2024.103962. Articles reporting phase III prospective RCTs investigating the efficacy of first-line immune checkpoint inhibition in HER2-negative and HER2-positive, advanced GEAC were included. Both original trial publications and conference abstracts [e.g. conference proceedings from the American Society of Clinical Oncology Annual Meeting (ASCO) and European Society for Medical Oncology Congress (ESMO)] were allowed. Permission to use was sought from the first author when data of interest were only available in conference presentations. This search was conducted without language restriction. Three authors independently filtered the title abstracts, followed by a full-text review. Discrepancies were resolved by consensus or in consultation with a senior author.

Data on included studies were extracted using a standardized form. The risk of bias was assessed using the revised Cochrane risk-of-bias tool for randomized trials (RoB2) which scores the risk of bias in five domains (randomization process, deviations from the intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result).¹³ Two authors independently carried out the assessment, and any discrepancies were resolved by a senior author.

Reconstruction of time-to-event outcomes

A graphical reconstructive algorithm^14,15 was used to estimate time-to-event outcomes from KM plots published from the KEYNOTE-859⁴ and KEYNOTE-811¹⁶ trials, as well as data from the RATIONALE-305^17,18 presented at the recent ESMO Gastrointestinal Cancer Congress 2024. Marginal hazard ratios (HRs) were computed using the Cox proportional hazards (PH) regression model.¹⁹ Outcomes of interest were OS, PFS, and duration of response (DOR). Reconstructed KM plots were compared with the original plots through visual inspection, marginal HRs, log-rank values, risk tables, and the median OS, PFS, and DOR.

KMSubtraction: matching of reconstructed patient data and derivation of unreported subgroups

With the reconstructed time-to-event outcomes from published overall and subgroup KM plots, bipartite matching of patient data was carried out using the Hungarian algorithm.²⁰ Matched patients are excluded, and the resultant data for the unreported subgroup are retrieved. For each implementation of the KMSubtraction algorithm,²¹ Markov chain Monte Carlo simulations were run with 1000 iterations to ascertain the error limits. In addition, the efficacy of the bipartite matching algorithm was evaluated using empirical cumulative distribution plots and Bland–Altman plots. The KM plots of matched cohorts were assessed to ensure near-complete overlap. Lastly, KM plots of unreported subgroups were then generated with the corresponding marginal HRs. All analyses were undertaken in R (R Foundation, Vienna, Austria), version 4.3.1, and the level of statistical significance is defined as a two-sided P < 0.05.

Meta-analysis of derived individual patient data of unreported subgroups

Following KMSubtraction, one-stage meta-analyses for OS were carried out for the following combinations of subgroups: (1) the PD-L1 CPS 1-9 subgroups of the KEYNOTE-859, KEYNOTE-062, and RATIONALE-305 trials; (2) the PD-L1 CPS 1-4 subgroups of the KEYNOTE-859, CHECKMATE-649, and RATIONALE-305 trials; and (3) the PD-L1 CPS 5-9 subgroup of the KEYNOTE-859 and RATIONALE-305 trials. The derivation of the unreported subgroup data for the KEYNOTE-062 and CHECKMATE-649 trials was conducted previously by Zhao et al.²² The Cox PH model was utilized for the primary analysis. To account for interstudy heterogeneity, the model was stratified by studies, thereby assuming a baseline hazard among patients unique to each study. The PH assumption was evaluated using the Grambsch–Therneau test. Where the PH assumption was violated, sensitivity analyses were undertaken utilizing restricted mean survival times.

Results

Study selection

Eight RCTs (KEYNOTE-062,²³ CHECKMATE-649,² KEYNOTE-590,²⁴ KEYNOTE-811,¹⁶ KEYNOTE-859,⁴ ORIENT-16,²⁵ RATIONALE-305,^17,18 and ATTRACTION-4²⁶) were included in this study (Figure 1). Three RCTs (KEYNOTE-590, ORIENT-16, and ATTRACTION-4) did not report KM plots for outcomes of interest in the PD-L1_low subgroups. The risk of bias in included studies was low (Supplementary Material 2, available at https://doi.org/10.1016/j.esmoop.2024.103962). A summary of trial characteristics and PD-L1 assays utilized and PD-L1 expression distribution across arms is reported in Table 1.

Figure 1.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchartdiagram. IPD, individual patient data; ITT, intention to treat; KM, Kaplan–Meier; PD-L1, programmed death-ligand 1; RCT, randomized controlled trial.

Table 1.

Baseline characteristics of included studies

Author/year	Trial	HER-2	Regions	Number of patients (CT-IT versus CT)	PD-L1 assay	CPS ≥10	CPS 1-9	CPS 5-9	CPS 1-4	Anti-PD-1 agent	CT backbone	Sited
Sun/2021	KEYNOTE-590a	–	Asia and ROW	99 versus 102	22C3	43 versus 54	NA	NA	NA	Pembrolizumab	Cisplatin + 5FU	Esophageal/GEJ (15:12:ESCC)
Shitara/2020	KEYNOTE0-62b	–	Asia and ROW	257 versus 250	22C3	99 versus 90	158 versus 160	NA	NA	Pembrolizumab	Cisplatin + 5FU	Gastric/GEJ (70:30)
Janjigian/2021	CHECKMATE-649c	–	Asia and ROW	789 versus 792	28.8	NA	NA	NA	168 versus 173	Nivolumab	Oxaliplatin + fp	Gastric/GEJ/EAC (70:18:12)
Kang/2022	ATTRACTION-4	–	Asia	362 versus 362	28.8	NA	NA	NA	NA	Nivolumab	Oxaliplatin + fp	Gastric/GEJ (65:8:UNK)
Janjigian/2023	KEYNOTE-811	+	Asia and ROW	350 versus 348	22C3	NA	NA	NA	NA	Pembrolizumab	Cisplatin or oxaliplatin + fp (+ Trastuzumab)	Gastric/GEJ (69:31)
Xu/2023	ORIENT-16	–	Asia	327 versus 323	22C3	146 versus 142	129 versus 129	51 versus 58	78 versus 71	Sintilimab	Oxaliplatin + fp	Gastric/GEJ (81:19)
Rha/2023	KEYNOTE-859	–	Asia and ROW	790 versus 789	22C3	279 versus 272	337 versus 345	110 versus 121	239 versus 229	Pembrolizumab	Cisplatin or oxaliplatin + fp	Gastric/GEJ (78:22)
Qiu/2024 + Moehler 2024	RATIONALE-305	–	Asia and ROW	501 versus 496	SP263	102 versus 151	269 versus 296	103 versus 131	166 versus 165	Tislelizumab	Cisplatin or oxaliplatin + fp	Gastric/GEJ (81:19)

5FU, 5-fluorouracil; CPS, combined positive score; CT, chemotherapy; EAC, esophageal adenocarcinoma; ESCC, esophageal squamous cell carcinoma; fp, fluoropyrimidine; GEJ, gastroesophageal junction; HER-2, human epidermal growth factor receptor 2; IT, immunotherapy; NA, not applicable; PD-1, programmed cell death protein 1; PD-L1, programmed death-ligand 1; ROW, rest of the world.

^aExcluding patients with ESCC.

^bExcluding the ‘pembrolizumab alone’ arm.

^cExcluding the ipilimumab + nivolumab arm.

^dPercentages refer to the experimental arm.

Evaluating the accuracy of reconstruction and matching

The reconstructive algorithm yielded IPD that provided similar HRs, log-rank values, and KM plots to those reported in KEYNOTE-811, KEYNOTE-859, and RATIONALE-305. A visual comparison between original and reconstructed KM plots is provided in Supplementary Material 3, available at https://doi.org/10.1016/j.esmoop.2024.103962. PD-L1 CPS 1-4 and 5-9 curves for KEYNOTE-859 were directly reconstructed from the EMA Keytruda Assessment Report (Procedure No. EMEA/H/C/003820/II/0135).⁹ The similarity between the original and reconstructed plots provided confidence to proceed with subsequent analyses.

The minimal cost bipartite matching algorithm produced optimal matching of the subgroup patients to the overall cohort. Each instance of KMSubtraction implementation is described in Supplementary Material 4, available at https://doi.org/10.1016/j.esmoop.2024.103962. There were minimal discrepancies between the matched pairs on empirical cumulative distributions and Bland–Altman plots, and the means of absolute differences in follow-up time approximated 0 (Supplementary Material 5, available at https://doi.org/10.1016/j.esmoop.2024.103962). A near-complete overlap was also observed on the KM plots of matched cohorts. Moreover, simulations using 1000 Monte Carlo iterations found negligible limits of error in each implementation of KMSubtraction (Supplementary Material 6, available at https://doi.org/10.1016/j.esmoop.2024.103962).

KMSubtraction-derived PD-L1_low or PD-L1_negative subgroups

KM plots of PD-L1_low or PD-L1_negative subgroups from KEYNOTE-811 (Supplementary Material 7A-B, available at https://doi.org/10.1016/j.esmoop.2024.103962), RATIONALE-305 (Figure 2), and KEYNOTE-859 (Figure 3) were retrieved with IPDfromKM and KMSubtraction and provided if not previously reported before.

Figure 2.

KMSubtraction-derived PD-L1_lowsubgroup Kaplan–Meier plots for RATIONALE-305. (A) RATIONALE-305 PD-L1 CPS 1-9 OS; (B) RATIONALE-305 PD-L1 CPS 1-4 OS; (C) RATIONALE-305 PD-L1 CPS 5-9 OS. CI, confidence interval; CPS, combined positive score; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1.

Figure 3.

KMSubtraction-derived PD-L1_lowor PD-L1_negativesubgroup Kaplan–Meier plots for KEYNOTE-859. (A) KEYNOTE-859 PD-L1 CPS <1 OS (PD-L1_negative). (B) KEYNOTE-859 PD-L1 CPS 1-9 OS (PD-L1_low). (C) KEYNOTE-859 PD-L1 CPS 1-4 OS (PD-L1_low). (D) KEYNOTE-859 PD-L1 CPS 5-9 OS (PD-L1_low). (E) KEYNOTE-859 PD-L1 CPS <1 PFS (PD-L1_negative). (F) KEYNOTE-859 PD-L1 CPS 1-9 PFS (PD-L1_low). (G) KEYNOTE-859 PD-L1 CPS 1-4 PFS (PD-L1_low). (H) KEYNOTE-859 PD-L1 CPS 5-9 PFS (PD-L1_low). CI, confidence interval; CPS, combined positive score; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1; PFS, progression-free survival. Note: For the purposes of this study, we defined PD-L1_negative as a PD-L1 expression <1, PD-L1_low as PD-L1 expression ranges excluding PD-L1 <1 tumors and PD-L1 ≥10, and PD-L1_high as a PD-L1 expression including ≥10.

In the KEYNOTE-811 PD-L1 CPS <1 subgroup, there were no significant differences in PFS between pembrolizumab–chemotherapy–trastuzumab and placebo–chemotherapy–trastuzumab [HR 1.011, 95% confidence interval (CI) 0.635-1.612, P = 0.962]. Regarding OS, there was a strong trend toward worse outcomes with the addition of pembrolizumab to trastuzumab and chemotherapy (HR 1.585, 95% CI 0.967-2.596, P = 0.065; Supplementary Material 7A-B, available at https://doi.org/10.1016/j.esmoop.2024.103962).

In the RATIONALE-305 PD-L1 CPS 1-9 subgroup, our retrieved KM curve did not show a statistically significant OS benefit for the chemo-immunotherapy arm compared with chemotherapy alone (HR 0.844, 95% CI 0.701-1.016, P = 0.073; Figure 2A). Similarly, no significant OS benefit was found in the PD-L1 CPS 1-4 subgroup (HR 0.843, 95% CI 0.669-1.062, P = 0.148; Figure 2B) and in the PD-L1 CPS 5-9 subgroup (HR 0.796, 95% CI 0.575-1.075, P = 0.131; Figure 2C).

In the KEYNOTE-859 PD-L1 CPS <1 subgroup, no significant differences in OS and PFS were obtained from the addition of pembrolizumab to chemotherapy (OS HR 0.933, 95% CI 0.738-1.181, P = 0.565; PFS HR 0.876, 95% CI 0.685-1.121, P = 0.292; Figure 3A and E, respectively). Similarly, no significant differences were found in the DOR between the two arms (HR 0.951, 95% CI 0.650-1.391, P = 0.795; Supplementary Material 8A, available at https://doi.org/10.1016/j.esmoop.2024.103962).

In the KEYNOTE-859 PD-L1 CPS 1-9 subgroup, our retrieved KM plot confirmed a sustained, yet modest, OS and PFS benefit in the pembrolizumab-chemotherapy arm compared with the placebo-chemotherapy arm (OS HR 0.827, 95% CI 0.702-0.974, P = 0.023; PFS HR 0.829, 95% CI 0.698-0.985, P = 0.032; Figure 3B and F, respectively). Similarly, the DOR was significantly improved in the pembrolizumab–chemotherapy arm (HR 0.751, 95% CI 0.571-0.987, P = 0.040; Supplementary Material 8B, available at https://doi.org/10.1016/j.esmoop.2024.103962). Patients with PD-L1 CPS 1-4 treated with pembrolizumab had a significant OS and PFS benefit (OS HR 0.785, 95% CI 0.643-0.959, P = 0.017; PFS HR 0.780, 95% CI 0.634-0.959, P = 0.018; Figure 3C and G, respectively). The PH assumption was not violated for both models. However, the 3-year OS was similar in the experimental versus control arm (11.4% versus 7.6%). Conversely, patients with PD-L1 CPS 5-9 did not have any significant survival benefit (OS HR 0.934, 95% CI 0.702-1.241, P = 0.636; PFS HR 0.908, 95% CI 0.676-1.221, P = 0.525) when treated with pembrolizumab (Figure 3D and H, respectively). However, the 3-year OS was 15.1% versus 3.2% in the experimental versus control arm, respectively.

Pooled analysis of PD-L1_low-expressing subgroups in HER2-negative GEAC

Considering the differential predictive value of PD-L1 expression in the PD-L1-1 CPS 1-4 and CPS 5-9 subgroups found in KEYNOTE-859, with the aim of increasing the statistical power, we pooled OS data with corresponding PD-L1 subgroups from CHECKMATE-649, KEYNOTE-062, and RATIONALE-305 on a one-stage model to further interrogate this finding. Subsequently, we carried out a two-stage IPD meta-analysis for each subgroup (Supplementary Material 9A-C, available at https://doi.org/10.1016/j.esmoop.2024.103962).

PD-L1 CPS 1-9 subgroup

PD-L1 1-9 subgroup KM plots from three trials, KEYNOTE-859, KEYNOTE-062 (previously reported in Zhao et al.²²), and RATIONALE-305, were pooled together. OS was found to be significantly superior with the addition of an anti-PD-1 agent (OS HR 0.840, 95% CI 0.753-0.937, P = 0.002; Figure 4A). However, the PH assumption was found to be violated (Grambsch–Therneau test, P = 0.046)

Figure 4.

Pooled analysis of PD-L1_lowsubgroups in human epidermal growth factor receptor 2 (HER2)-advanced GEAC. (A) OS outcomes for the PD-L1 CPS 1-9 subgroup comprising KEYNOTE-859, KEYNOTE-062, and RATIONALE-305. (B) OS outcomes for the PD-L1 CPS 1-4 subgroup comprising KEYNOTE-859, CHECKMATE-649, and RATIONALE-305. (C) OS outcomes for the PD-L1 CPS 5-9 subgroup comprising KEYNOTE-859 and RATIONALE-305. CI, confidence interval; CPS, combined positive score; GEAC, gastroesophageal adenocarcinoma; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1.

PD-L1 CPS 1-4 and 5-9 subgroups

PD-L1 CPS 1-4 subgroup KM plots from KEYNOTE-859, CHECKMATE-649 (previously reported in Zhao et al.²²), and RATIONALE-305 were pooled together and showed significant OS benefit with the addition of anti-PD-1 agents (OS HR 0.868, 95% CI 0.772-0.976, P = 0.018; Figure 4B). The PH assumption was not violated.

Conversely, when PD-L1 CPS 5-9 subgroup KM plots from KEYNOTE-859 and RATIONALE-305 were pooled together, no significant OS benefit was found with the addition of an anti-PD-1 agent (HR 0.867, 95% CI 0.702-1.069, P = 0.181; Figure 4C). However, the PH assumption was found to be violated (Grambsch–Therneau test, P = 0.028).

Conventional meta-analysis of RCTs pooled by chemotherapy regimen in HER-2-negative GEAC

To identify possible interactions between survival outcomes and chemotherapy regimens used in the different RCTs in the HER-2-negative setting, we carried out a conventional meta-analysis pooling trials by chemotherapy regimen (cisplatin-based regimens versus oxaliplatin-based regimens). OS and PFS improved consistently and similarly in the intention-to-treat population (P for heterogeneity in OS = 0.78; P for heterogeneity in PFS = 0.390) when comparing patients treated with cisplatin- (OS HR 0.82, 95% CI 0.72-0.95; PFS HR 0.78, 95% CI 0.67-0.90) and oxaliplatin-based chemotherapy (OS HR 0.81, 95% CI 0.76-0.85; PFS HR 0.74, 95% CI 0.69-0.80; Supplementary Material 10A, B, available at https://doi.org/10.1016/j.esmoop.2024.103962). Notably, the ipilimumab + nivolumab arm of CHECKMATE-649 proved to be an outlier with significantly worse PFS. Therefore this arm was excluded from the sensitivity analysis.

Discussion

Our analysis confirmed that PD-L1 CPS is a useful biomarker to predict the benefit from ICI in both HER2-positive and HER2-negative advanced GEAC, consistent with findings reported in a key meta-analysis by Yoon et al.²⁷

Our IPD meta-analysis in the CPS 1-9 subgroup (including KEYNOTE-859, KEYNOTE-062, and RATIONALE-305) showed a statistically significant benefit in the OS of an anti-PD-1 agent in combination with doublet chemotherapy versus chemotherapy alone.

Interestingly, the meta-analysis for the CPS 1-4 subgroup (including KEYNOTE-859, CHECKMATE-649, and RATIONALE-305) also showed a significant benefit for the chemo-immunotherapy arm. These results take on particular significance in light of the fact that CPS 5 is considered to be the most clinically reliable cut-off to date, according to the data from CHECKMATE-649 and the subsequent analysis by Zhao et al.²² However, as our meta-analysis shows, a picture may emerge in which the CHECKMATE-649 data on OS in patients with CPS 1-4 are an outlier. Further data are warranted to confirm or refute this observation. However, in both CPS 1-9 and CPS 1-4 subgroups, the OS benefit was overall modest and apparent only in the long-term outcomes and it would potentially correspond to grade 1 on the ESMO-Magnitude of Clinical Benefit Scale (MCBS) version 1.1.²⁸

With respect to the CPS 5-9 subgroup, the KEYNOTE-859 data did not show a significant benefit in OS and PFS for the chemo-immunotherapy arm, but a meaningful long-term difference was observed at the 3-year timepoint. Our meta-analysis (including KEYNOTE-859 and RATIONALE-305) showed a trend toward a modest OS benefit, although the results are still not significant. It seems plausible that the lack of significance may be related to the lower number of patients in this subgroup compared with the subgroup of patients with CPS 1-4. As such, the apparently counterintuitive findings of KEYNOTE-859 could be just a random observation because these subgroup analyses were not predefined in the original trial. Moreover, the well-known intertumoral and intratumoral heterogeneity of PD-L1 and the interpathologist and intrapathologist variability of PD-L1 scoring are likely even more pronounced when PD-L1 is expressed at low levels (i.e. CPS <10).²⁹ Therefore these results could be affected by a mismatch between PD-L1 scoring and actual PD-L1 expression.

While the survival benefit of chemo-immunotherapy in PD-L1_high is clear, we believe our current work strengthens the evidence that further translational research is needed to better select patients who could benefit from immunotherapy in PD-L1_low GEAC. We need additional biomarkers (microsatellite instability, tumor mutational burden, and immune signature) because the number needed to treat to have long-term benefit in this subgroup is still too high. Furthermore, the presence of alternative therapeutic options such as zolbetuximab in Claudin 18.2-positive disease must be considered. Claudin 18.2 expression has been shown to be highly prevalent in HER-2-negative, PD-L1-low GEAC samples, potentially providing a unique therapeutic target in this subset of patients.³⁰ In the GLOW study,³¹ 78.1% of patients had a CPS <5, while in the SPOTLIGHT study,³² this percentage reached 87%. In this regard, it will be interesting to see the results of the new cohorts of the ongoing phase II ILUSTRO trial³³ which will assess the efficacy of the combination chemotherapy–immunotherapy–zolbetuximab as first-line treatment in HER-2 negative GEAC. Meanwhile, we advocate for future immunotherapy trials for patients with GEAC to incorporate similar PD-L1 cut-offs and scoring systems and to provide full outcome datasets to peer-reviewed journals. In this context, the results of the ASPIRE study (Stomach cancer PD-L1 biomarker European Initiative) will be crucial for achieving harmonization of clinical use of immunohistochemistry assays and scoring methods for PD-L1.³⁴

Concerning the HER-2 positive setting, a possible trend toward a detrimental effect in OS has been shown by our retrieved KM plot in patients with CPS <1 treated with pembrolizumab plus trastuzumab–chemotherapy. The findings in this study suggest that the OS and PFS benefits associated with ICI-based regimens reported in the overall population might be attributed to positive outcomes specifically in the subgroups of patients with at least CPS ≥1 or an even higher PD-L1 expression. However, Merck recently announced that KEYNOTE-811 improved OS in the overall population with ‘the greatest benefit observed in patients with PD-L1 CPS >1’.³⁵ While we await the final follow-up analysis, our results clearly support the restriction of pembrolizumab approval by the FDA after >2 years of conditional approval in all comers.

Our study has objective limitations: Some of the individually reported subgroups may not be adequately powered to draw definitive conclusions specific to the original trials due to their small sizes. Despite methodological precautions taken to ensure close alignment between derived KM curves and HRs with reported HRs, we acknowledge these differences. These discrepancies may be attributed to subtle variations in censoring or patient-level covariates that are impossible to account for given the absence of participant-level data and the nature of univariate survival models. These analyses would certainly be more reliable if the authors of the original studies had directly published the data along with their respective KM plots.

In conclusion, we believe that in clinical practice the upfront choice of immunotherapy in patients with PD-L1_low should be discussed with individual patients considering the potential risk of toxicities, the financial costs, the expected benefit in the medium versus long-term, and presence of alternative therapeutic options such as zolbetuximab in Claudin18.2-positive disease. Considering the significant prevalence of immune-related toxicity associated with ICI-based regimens and the financial implications they pose on national health systems, it becomes imperative to prioritize the optimal selection of patients, aiming to minimize unnecessary treatments.³⁶

Disclosure

ES reports grants and personal fees from BMS and Astra Zeneca; personal fees from Amgen, Daiichi Sankyo, Merck, Viracta, Astellas, Novartis, Pfizer, Zymeworks, BeiGene, outside the submitted work; personal fees and nonfinancial support from Mirati; and has been the EORTC Gastric Cancer Taskforce Chair 2021-2024 UK & Ireland Oesophagogastric Cancer Group Trustee since 2022. KK has received consulting fees from AstraZeneca, Bayer, BeiGene/Novartis, Bristol-Myers Squibb, Merck Biopharma, Ono Pharmaceutical, and Roche; has received honoraria from Bristol-Myers Squibb and Ono Pharmaceutical; and has participated in a data safety monitoring board or advisory board for Bristol Myers Squibb, Merck Biopharma, and Ono Pharmaceutical. RS reports attending advisory board meetings for Bristol Myers Squibb, Merck, Eisai, Bayer, Taiho, Novartis, MSD, GSK, DKSH, Astellas, Pierre-Fabre, and Tavotek; receiving honoraria for talks from MSD, Eli Lilly, BMS, Roche, Taiho, Astra Zeneca, DKSH, Ipsen, Daiichi Sankyo, BeiGene, and Astellas; receiving travel support from Roche, Astra Zeneca, Taiho, Eisai, DKSH, and Ipsen; receiving research funding from Paxman Coolers, MSD, Natera, and CytoMed Therapeutics; and has patents pending with Auristone and Paxman. MM has received honoraria from Amgen, Roche/Genentech, Merck Serono, MSD Oncology, Bristol Myers Squibb, AstraZeneca/MedImmune, Servier, Pierre Fabre, Sanofi, Transcenta, Idience, Triptych Health Partners, Daiichi Sankyo Europe GmbH, Astellas Pharma, and Bayer Schering Pharma; has received consulting fees from Bayer, MSD, Merck Serono, Amgen, Taiho Pharmaceutical, Lilly, Servier, BeiGene, Bristol Myers Squibb, AstraZeneca, Nordic Group, and Daiichi Sankyo Europe GmbH; received institutional research funding from Amgen, Leap Therapeutics, Merck Serono, AstraZeneca, MSD, and Taiho Pharmaceutical; personal fees from Amgen, Merck Serono, Roche, Bayer, ASCO, German Cancer Society, MSD, ESMO, BeiGene, and Sanofi Pasteur. FP reports receiving institutional research grants from BMS, Incyte, Agenus, Amgen, Lilly, and AstraZeneca, and personal fees from BMS, MSD, Amgen, Merck-Serono, Pierre-Fabre, Servier, Bayer, Takeda, Astellas, Johnson & Johnson, Rottapharm, Ipsen, AstraZeneca, GSK, Daiichi-Sankyo, Seagen, Pfizer, Beigene, Jazz Pharmaceuticals, and Incyte. All other authors have declared no conflicts of interest.