PD-L1 as a response biomarker to chemotherapy plus immune checkpoint inhibitors in operable gastroesophageal adenocarcinoma: a meta-analysis of neoadjuvant clinical trials Authors: A Cammarota, SK Kamarajah, S Markar, EC Smyth Published: 2024 ABSTRACT AbstractBackgroundPerioperative chemoimmunotherapy [i.e. chemotherapy plus immune checkpoint inhibitor (ICI) (ChT–ICI)] is an investigational treatment for operable gastroesophageal adenocarcinoma (GEA) with unclear outcomes in an unselected population. We aimed to assess the cumulative treatment effect of ChT–ICI on pathological response rates, the surrogacy of pathological response rates on overall survival (OS), and the value of programmed death-ligand 1 (PD-L1) as a biomarker of response to ICIs in operable GEA.MethodsWe conducted a systematic review and meta-analysis of randomised clinical trials (RCTs) of ChT with or without ICIs with the primary outcome being pathological response rates [pathological complete response (pCR) and similar metrics]. A weighted linear regression model quantified the relationship between pathological response rates and OS using a determination coefficient (R2). A second meta-regression analysed the predictive effect of PD-L1 on pCR in trials of ICIs with ChT or chemoradiotherapy (CRT).ResultsA total of 18 records from 15 RCTs were included. Of 6624 patients, 5291 received ChT and 1333 ChT–ICI. ChT–ICI significantly improved pathological response rates [pooled odds ratio 3.18, 95% confidence interval (CI) 2.42-4.18, P < 0.0001] compared with ChT (pooled odds ratio 1.66, 95% CI 1.33-2.07, P < 0.0001). The correlation between pathological response and OS was low (R2 = 0.12) but improved in recent trials (R2 = 0.51) and those with ChT–biological agents, including ICIs (R2 = 0.79). In the second analysis (11 studies, 633 patients), PD-L1 ≥5 was a significant predictor of response both individually (estimate: 0.73, 95% CI 0.28-1.18, P = 0.001) and after accounting for the backbone treatment (estimate: 0.80, 95% CI 0.28-1.33, P = 0.003).ConclusionsPerioperative ChT–ICI improves pathological response rates in operable GEA and PD-L1 ≥5 is a significant biomarker of pCR, supporting stratification by PD-L1 and the design of biomarker-selected trials.Key words: immunotherapy, gastroesophageal adenocarcinoma, neoadjuvant treatment, pathological complete response, PD-L1, response biomarkers Introduction Over the past decade, neoadjuvant therapies such as CROSS and FLOT have significantly improved the management of operable gastroesophageal cancers, increasing 5-year overall survival (OS) rates to ∼50% overall and up to 80% in those with a pathological complete response (pCR).1,2 Despite this, recurrence rates remain high, prompting ongoing research into new treatment options. Neoadjuvant use of immunotherapy has improved survival in other tumour types, but it has unclear value in patients with operable gastroesophageal adenocarcinoma (GEA).3,4 Despite having established activity in the unresectable setting, immune checkpoint inhibitors (ICIs) yielded controversial results when used in the perioperative space in the recent KEYNOTE-585 trial where survival endpoints were not met following an earlier improvement of pathological outcomes.5, 6, 7, 8 In early-phase studies, pCR is a common endpoint to screen novel therapeutics allowing for a quick read-out of their biological activity. Recognised by the Food and Drug Administration (FDA) and European Medicines Agency (EMA) as a valid surrogate endpoint in breast cancer,9 pCR has also been adopted in trials for gastroesophageal cancers based on biological plausibility. Yet, the disconnection observed between pathological and clinical results across different trials casts doubt on the validity of pCR as a solid surrogate outcome in this space. Further, the variability in patient-level response to immunotherapy raises concerns around GEA biological heterogeneity and requests the identification of robust predictors to guide treatment decisions. Besides microsatellite instability (MSI), tissue expression of programmed death-ligand 1 (PD-L1) is a recognised biomarker of benefit from anti-PD-(L)1 agents across different solid tumours, including advanced GEA.10 However, PD-L1 has inherent heterogeneity issues and validated thresholds have been mostly tied to specific settings and treatments.11 It remains, thus, to be determined whether PD-L1 expression could serve as a biomarker of response to ICIs in operable GEA. Therefore, this study aimed to (i) carry out a systematic review and meta-analysis of the treatment effect of perioperative chemotherapy plus ICI (ChT–ICI) on pathological response outcomes; (ii) explore the surrogacy of pathological response on survival outcomes; and (iii) assess the value of PD-L1 as a predictor of treatment response in patients considered eligible for GEA resection receiving ICIs. Methods Systematic review and meta-analysis Search strategy and selection criteria We conducted a systematic review of randomised clinical trials (RCTs) testing neoadjuvant or perioperative ChT with or without ICIs in nonmetastatic GEA following the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guidelines.12 Institutional review board approval and informed consent were waived as our analyses did not contain individual patient data nor required the collection of new data. A comprehensive literature search was carried out for articles or conference abstracts published between 1 January 2006 and 15 January 2024, in MEDLINE, Embase, Scopus, Web of Science, and Cochrane Central Register of Clinical Trials. Conference abstracts were hand-selected from conferences’ websites. To be eligible RCTs had to test neoadjuvant or perioperative systemic treatments in both the experimental and comparator arms and report pathological response results for all arms. Studies testing systemic therapies against surgery alone, combination treatments with neoadjuvant radiotherapy, or enrolling a population with mixed histologies, such as GEA and squamous cell carcinoma, were excluded to reduce confounding factors. Study selection was restricted to English publications. The full information on the search criteria is available in Supplementary Table S1, available at https://doi.org/10.1016/j.esmogo.2024.100107. Two investigators (AC and SK) independently extracted data from each study, including the first author’s name, study design characteristics, trial treatment agents, patient demographics, duration of follow-up, and relevant outcomes. Quality assessment The selected studies were independently assessed for the risk of bias, adopting the Cochrane Risk of Bias tool RoB-2 recommended for randomised clinical studies (Supplementary Figure S1, available at https://doi.org/10.1016/j.esmogo.2024.100107). Disagreements were resolved through consensus or consultation with the other authors. Publication bias was inspected using funnel plots (Supplementary Figure S2, available at https://doi.org/10.1016/j.esmogo.2024.100107). Data extraction and statistical analysis We extracted and calculated the logit of odds ratios (ORs) and their respective 95% confidence intervals (CIs) from the included trials and carried out a meta-analysis to estimate the mean effect size across all included trials. The outcome of interest was pathological response defined as pCR or similar metrics, such as major pathological response, pathological subtotal response, or ypT0/ypT ≤1. These measures were considered together in the primary analysis. Heterogeneity across studies was assessed by I2 and between-subgroup homogeneity by Q. If I2 was >50% or the P value >0.10, the heterogeneity was considered substantial, and the random effect model was applied.13 Otherwise, the fixed-effect model was adopted. The model was estimated using the restricted maximum likelihood method. A subgroup analysis of trials of ChT–ICI versus trials of ChT with or without non-ICI agents was carried out to estimate the pooled effect of perioperative ICIs on pathological response. Sensitivity analyses restricted to (i) trials reporting pCR (excluding those reporting other metrics of pathological response); (ii) trials published since the standard-of-care FLOT regimen; (iii) studies evaluating ChT with or without ICIs only (excluding trials of non-ICI biological agents); and (iv) studies enrolling a rest-of-the-world (ROW; i.e. non-Asian only) population were carried out to further explore the impact of between-studies heterogeneity on the overall effect size. Association between pathological response and OS To evaluate the treatment effect on pathological response and median OS, we extracted the following data from the RCTs included in the meta-analysis that reported survival outcomes: (i) pathological response rates, expressed as the proportion of patients achieving pCR, major pathological response, pCR/pathological subtotal response, ypT0, or ypT ≤1 according to how pathological responses were recorded in each study; (ii) treatment effects, expressed as ORs for pathological response; and (iii) hazard ratios (HRs) and their respective 95% CIs for OS. The relationship between pathological response and OS was quantified using a determination coefficient (R2). In detail, a weighted linear regression model was fit with the natural logarithm of the OR [log(OR)] of pathological response as the independent variable and log(HR) of OS as the dependent variable. Data were weighted based on the sample size and variance of individual studies. R2 was computed for the overall model and described the proportion of variance in OS explained by the pathological response across studies. A nonparametric bootstrapping technique (1000 resamples) of the study data was used to calculate the distribution of the estimated regression coefficients and R2. Bias-corrected and accelerated CIs for R2 were derived from the bootstrap distribution as previously described.14 The strength of the association between pathological response and OS was interpreted according to prior frameworks proposed for the validation of surrogate endpoints by regulatory agencies. In detail, we considered R2 values ≥0.7 as representative of strong correlation, values between 0.69 and 0.5 as representative of moderate correlation, and values <0.5 as weak correlation according to Systematic Review and Recommendation for Reporting of Surrogate Endpoint Evaluation using Meta-analyses (ReSEEM) guidelines.15 Sensitivity analyses were carried out to assess the effect of different ethnicity (Asian/non-Asian), outcome measures (pCR/other pathological outcomes), treatment regimens (ChT with or without biological agents), and trial year (before or since FLOT) on the trial-level relationship between pathological response and OS. Meta-regression Search strategy and selection criteria To estimate the effect of PD-L1 expression on the pooled pCR rates, we conducted a second systematic review of clinical trials testing neoadjuvant or perioperative ICIs with ChT or chemoradiotherapy (CRT) in nonmetastatic GEA according to the PRISMA reporting guidelines.12 A comprehensive literature search was carried out for articles or conference abstracts published between 1 January 2006 and 15 January 2024, in MEDLINE, Embase, Scopus, Web of Science, and Cochrane Central Register of Clinical Trials. Conference abstracts were hand-selected from the conferences’ websites. Eligibility criteria were (i) phase II or III clinical trials; (ii) experimental arm testing ICIs with ChT or CRT as neoadjuvant or perioperative treatment; (iii) availability of pCR data; (iv) information on baseline PD-L1 expression on tissue; (v) histology restricted to adenocarcinoma; and (vi) setting restricted to nonmetastatic disease. Studies were excluded if enrolment was limited to a biomarker-selected population (i.e. MSI) and if published data lacked information on pCR for the overall population or by PD-L1 expression. Only studies published in English were included. When studies had a randomised design, only the outcomes of patients treated with ICIs were retrieved. The full information on the search criteria is available in Supplementary Table S2, available at https://doi.org/10.1016/j.esmogo.2024.100107. A summary of the study workflow for the meta-analysis and meta-regression is provided in Supplementary Figure S3, available at https://doi.org/10.1016/j.esmogo.2024.100107. Quality assessment Randomised studies were included in the meta-analysis and therefore assessed with the bias tool RoB-2 (Supplementary Figure S1, available at https://doi.org/10.1016/j.esmogo.2024.100107). The risk of bias in nonrandomised studies was evaluated by two investigators independently (AC and SK) using the Risk Of Bias In Non-randomised Studies of Interventions (ROBIN-I) tool (Supplementary Figure S4, available at https://doi.org/10.1016/j.esmogo.2024.100107). Disagreements were resolved through consensus or consultation with the other authors. Data extraction and statistical analysis We extracted and calculated the logit of ORs and 95% CIs from the included trials for the overall population and by PD-L1 expression. PD-L1 was quantified using a combined positive score (CPS) in all the included studies according to predefined cut-offs (1, 5, or 10). When pCR data by a certain cut-off were not directly presented in the selected publications, these were extracted or calculated from available proportions and/or rates. If pCR by PD-L1 expression at a specific cut-off was not reported, the available pCR outcomes were grouped into the closest relevant PD-L1 category (e.g. for studies not presenting results by PD-L1 at a cut-off of 5, pCR by PD-L1 <1 and ≥10 were included in the <5 and ≥5 subgroups, respectively). The trial-level association between pCR and PD-L1 expression at each cut-off (1,5, and 10) was assessed using Fisher’s exact test. A mixed-effects meta-regression model with restricted maximum likelihood was used to estimate the influence of PD-L1 expression (as a binary variable) on pCR—the main outcome—across all eligible trials either alone or with a second covariate, the backbone treatment (ChT or CRT). Four models were run to assess each moderator independently and together. The cut-off for PD-L1 expression was selected using the lowest threshold that showed a statistically significant association with pCR by Fisher’s exact test. Fixed-effect models were used to assess the impact of the two covariates, while random effects were specified for each study to account for between-studies heterogeneity and within-study comparisons. All statistical analyses were carried out with R statistical software version 4.3.0 (R Project for Statistical Computing, Vienna, Austria). All P values were two-sided and statistical significance was set at <0.05. Results Meta-analysis Study selection and study and participant characteristics From an initial search yielding 24 502 records, 18 351 records were excluded after screening titles and abstracts and 3337 duplicates were eliminated. This left 46 records for further evaluation, of which 28 were excluded due to insufficient relevant data. Ultimately, 18 records from 15 RCTs, with 3 of these trials having two experimental arms contributing to two comparisons, were included in the final qualitative and quantitative synthesis (Figure 1).1,8,16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29 All studies were randomised phase II, II/III, or III trials and enrolled patients with nonmetastatic disease and adenocarcinoma histology. The analysis comprised 6624 patients with gastric or gastroesophageal junction adenocarcinoma, of which 5291 received neoadjuvant or perioperative ChT and 1333 ChT–ICI. Detailed characteristics of the included RCTs are summarised in Supplementary Table S3, available at https://doi.org/10.1016/j.esmogo.2024.100107. Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) diagram for the studies included in the meta-analysis.aA total of 18 records from 15 randomised controlled trials (RCTs), of which 3 had two experimental arms contributing to two comparisons. Cumulative efficacy of neoadjuvant chemoimmunotherapy versus chemotherapy Neoadjuvant ChT–ICI showed a significant improvement in pathological response (pooled OR 3.18, 95% CI 2.42-4.18, P < 0.0001) with a larger effect compared with chemotherapy (pooled OR 1.66, 95% CI 1.33-2.07, P < 0.0001; Figure 2A). Results were consistent in sensitivity analyses restricted to studies reporting (i) strictly pCR, (ii) enrolling an ROW (i.e. non Asian-only) population, (iii) testing exclusively ChT with or without ICI (excluding studies of ChT with non-ICI biological agents), and (iv) assessing since-FLOT regimens (Figure 2B). Pathological response rate with perioperative/neoadjuvant chemotherapy versus chemoimmunotherapy. (A) Forest plot of a meta-analysis of phase II or III randomised clinical trials16, 17, 18, 19, 20, 21, 22, 23, 24,26, 27, 28, 32, 39, 40,32,39,40 testing perioperative/neoadjuvant chemotherapy with or without immune checkpoint inhibitor (ICI) versus an active comparator. Subgroup analysis by treatment agent in the experimental arm (chemotherapy versus chemotherapy plus ICI). Duplicated study names occur for trials that have more than two arms, thus contributing to more than one comparison. Squares indicate estimates; size of squares, study weights; whiskers, 95% confidence intervals (CIs); diamonds, mean estimates. (B) Sensitivity analyses. Pooled effect size [odds ratio (OR)] for pathological response in chemotherapy versus chemoimmunotherapy trials restricted to (i) studies reporting strictly pathological complete response (pCR) as the absence of viable tumour cells, (ii) assessing regimens tested since FLOT, (iii) testing exclusively chemotherapeutic combinations (excluding studies of chemotherapy plus non-ICI biological agents), and (iv) enrolling a rest of the world (ROW; i.e. non-Asian only) population. ChT, chemotherapy; CTR, control arm; EXP, experimental arm; FLOT, fluorouracil, leucovorin, oxaliplatin, docetaxel; OR; odds ratio for pathological response. Evaluation of pathological response as a surrogate endpoint of OS Relevant OS information was available for 10 of the 18 records included in the meta-analysis. The median follow-up time of the surviving patients among all the studies was 45.35 months (range 34.90-76.80 months). Across the entire dataset, pathological response demonstrated a low correlation (R2 = 0.12, 95% CI 0.00-0.57) with OS, with substantial variability across the studies (Figure 3A). Trial-level surrogacy of pathological complete response (pCR) on overall survival (OS). (A) Trial-level correlation between pCR and OS.1,16, 17, 18, 19, 20, 21,23,24,40 Treatment effects are expressed as log(OR) for pCR and log(HR) for OS. Each circle represents a study. The size of the circles represents the weight of the study, which is proportional to the number of patients enrolled. The dashed line represents the weighted regression line, which shows the effect on OS predicted by the observed effects on pCR. The grey shaded area represents the 95% confidence intervals (95% CIs) of the regression line. (B) The bubble plots display the trial-level correlation between pCR and OS in (from right to left and top to bottom) studies reporting pCR16, 17, 18, 19, 20, 21,23,24,40; rest of the world (i.e. non-Asian only) studies1,17,19, 20, 21,23,24,40; since-FLOT studies1,23,24,40; studies of chemotherapy plus a biological agent17,20,24,40. CI, confidence interval; FLOT, fluorouracil, leucovorin, oxaliplatin, docetaxel; HR, hazard ratio; log, natural logarithm; OR; odds ratio; R2, coefficient of determination. Acknowledging differences in the reported pathological outcomes, enrolled populations, and treatment regimens used across the included studies, we conducted four sensitivity analyses to evaluate the relationship between pathological response and OS in a more similar set of trials (Figure 3B). The influence of pathological response on OS was slightly more pronounced in pCR-only reporting studies (R2 = 0.29, 95% CI 0.01-0.70) and ROW studies (R2 = 0.32, 95% CI 0.01-0.82), indicating that pathological response could explain a larger proportion of the variance in OS within more homogeneous subgroups. A moderate-to-strong trial-level correlation of pathological response with OS was found in the studies conducted since FLOT, where R2 was 0.51 (95% CI 0.00-1.00), and in trials of ChT plus biological agents, including ICIs (R2 = 0.79, 95% CI 0.23-1.00). Further sensitivity analyses are reported in Supplementary Table S4, available at https://doi.org/10.1016/j.esmogo.2024.100107. Meta-regression Study selection and study and participant characteristics A total of 1086 records were retrieved through the initial database search. Of these, 128 duplicates were removed and 940 records were excluded after screening the title and abstract. The remaining 19 records were evaluated for eligibility, 8 of which were removed lacking relevant data, leaving 11 records for inclusion in the final qualitative and quantitative synthesis (Figure 4).25,29, 30, 31, 32, 33, 34, 35, 36, 37, 38 The selected studies were single-arm (n = 8) or randomised (n = 3) phase II or phase III trials and enrolled patients with nonmetastatic disease and adenocarcinoma histology. Studies enrolling biomarker-selected populations (e.g. MSI) were excluded. Overall, the analysis comprised 633 patients with gastric or gastroesophageal junction adenocarcinoma, receiving neoadjuvant or perioperative ChT–ICI (n = 492) or CRT–ICI (n = 141). Of those treated with ChT–ICI, 167 patients (26%) received an anti-PD-(L)1 agent combined with a regimen containing docetaxel, oxaliplatin, and fluoropyrimidines, including FLOT. Characteristics of the included studies are summarised in Supplementary Table S5, available at https://doi.org/10.1016/j.esmogo.2024.100107. Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) diagram for the studies included in the meta-regression. Association of PD-L1 expression with pathological complete response to neoadjuvant chemoimmunotherapy Among all the studies included, the pooled pCR rate was 24% (95% CI 0.21-0.27) and ranged from 22% (95% CI 0.19-0.25) in patients treated with ChT–ICI to 33% (95% CI 0.25-0.41) for those receiving CRT–ICI. The pooled effect was 29% (95% CI 0.24-0.35), 34% (95% CI 0.26-0.42), and 40% (95% CI 0.28-0.52) in tumours with baseline PD-L1 CPS ≥1, CPS ≥5, and CPS ≥10, respectively. PD-L1 expression was significantly associated with higher pCR rates at cut-offs of 5 and 10 (Figure 5A). Association of programmed death-ligand 1 (PD-L1) expression with pathological complete response (pCR) rate in trials of perioperative/neoadjuvant ICI + ChT or ICI + CRT. (A) Comparison of pCR rates in the entire population and by PD-L1 cut-offs using the chi-square test. Odds ratio (OR) reported for statistically significant (two-sided P < 0.05) comparisons. (B) Simple and multivariable linear regression models assessing PD-L1 combined positive score (CPS; cut-off 5) and treatment backbones (ChT versus CRT) as predictors of pCR (explanatory variables). The top table shows the estimates, 95% confidence intervals (CI), and P values for PD-L1 CPS ≥5 and CRT backbone, both individually and as an interaction term. The bottom plot shows the estimates and their 95% CIs from the interaction term in the multivariable model, with a vertical dashed line indicating no effect (estimate = 0). ChT, chemotherapy; CRT, chemoradiotherapy; ICI, immune checkpoint inhibitor. In a mixed-effect meta-regression model, PD-L1 ≥5 was a significant predictor of response both individually (estimate: 0.73, 95% CI 0.28-1.18, P = 0.001) and after accounting for the backbone treatment (estimate: 0.80, 95% CI 0.28-1.33, P = 0.003; Figure 5B and Supplementary Table S6, available at https://doi.org/10.1016/j.esmogo.2024.100107). Discussion Novel chemoimmunotherapy regimens are being actively investigated in operable GEA to overcome the traditionally low responses and high relapse rates achieved with current standard strategies. However, the cumulative effect of these strategies on patient-relevant outcomes is yet to be assessed. To the best of our knowledge, we presented the first comprehensive meta-analysis of neoadjuvant treatments in operable GEA using RCT data and including the most recent chemoimmunotherapy studies. Overall, we found that neoadjuvant chemoimmunotherapy has a consistently larger treatment effect on pCR compared with neoadjuvant chemotherapy. Improvement of pathological outcomes with chemoimmunotherapy was maintained irrespective of how the pathological outcome was reported (pCR versus other metrics), trial year (before versus since-FLOT studies), treatment regimen (chemotherapy only versus chemotherapy plus biological agent), or ethnicity (Asian versus ROW). Leveraging contemporary studies, our work also assessed the trial-level surrogacy of pathological response on OS in GEA for which evidence was lacking. Even though the correlation between pathological response and OS was overall low (R2 = 0.12, 95% CI 0.00-0.57), moderate-to-strong correlations were observed in more recent trials (R2 = 0.51, 95% CI 0.00-1.00) and with chemotherapy plus biological agents, including ICIs (R2 = 0.79, 95% CI 0.23-1.00). Intermediate endpoints such as pCR may not have sufficient strength to qualify as surrogate outcomes for OS but induction of higher pCR brings patient-level benefits. Retrospective analyses have identified pCR as a favourable prognostic factor in cohorts of patients with GEA treated with chemotherapy or, more recently, chemoimmunotherapy.39,40 Besides, pCR retains value in drug development by screening drugs to move forward into late-stage testing. Here, surrogacy may be context-dependent and the connection between early and late efficacy outcomes may be strengthened within treatment-sensitive subgroups. In our study, the improved relationship between pathological response and survival in trials conducted since FLOT may underlie a better assessment of pCR. In more recent studies, the evaluation included a thorough assessment not only of the resected primary tumour but also of locoregional lymph nodes that more often than not are responsible for disease relapse. However, it may also indicate the use of active treatment regimens within susceptible populations. Chemoimmunotherapy consistently induced higher rates of pCR with respect to chemotherapy alone in our analysis as in the KEYNOTE-585 trial where the survival endpoints numerically favoured the immunotherapy arm, despite not meeting the thresholds for statistical significance.8 In advanced GEA, PD-L1 CPS ≥5 and ≥10 are clinically validated cut-offs to identify patients obtaining higher benefits from anti-PD-(L)1 treatment. Similarly, in our work, patients considered eligible for resection receiving neoadjuvant ChT–ICI or CRT–ICI achieved significantly higher pCR rates when their tumours had baseline PD-L1 expression ≥5 and ≥10. Further, we provided evidence that PD-L1 CPS ≥5 qualifies as an independent biomarker of response to combination treatment with immunotherapy in the operable space. This result held true also after the backbone treatment type (chemotherapy or CRT) was included in our predictive model. We acknowledge that our study has a number of limitations. First, our analyses lack individual patient data, which could have allowed us to stratify outcomes by clinically relevant known prognostic factors. Second, chemotherapy regimens were heterogeneous and there was a relative paucity of RCTs evaluating ChT–ICI versus ChT (n = 5, allowing for six comparisons). With most of these ChT–ICI studies still due to report their outcomes, we were not able to carry out a direct comparison of the treatment effect between the two strategies with a meta-analytic approach. Yet, the chemotherapy schemes used belonged to largely overlapping pharmacological classes with differences mostly reflecting changes in treatment practices over time (i.e. the use of epirubicin-based regimen before FLOT) and across the world (e.g. the use of S-1 in Asia and 5-fluorouracil/capecitabine in other regions). Further, heterogeneity was low within subgroups in our study, and the treatment effect was consistent in favouring ChT–ICI in our sensitivity analyses which accounted for such differences. Third, the evaluation of the relationship between pCR and survival in chemoimmunotherapy trials was impeded by the immaturity of OS data in all but one of the RCTs included at the time this analysis was conducted. As surrogacy may be context- and intervention-dependent, re-evaluation, when mature data from ChT–ICI RCTs will be available, is warranted. Lastly, we had to leverage a second dataset of neoadjuvant trials of ICIs in GEA to assess pathological outcomes by PD-L1 expression. As a result of the nature of this second data collection (mostly early-phase single-arm studies), we could only compare pCR but not OS by PD-L1 levels; therefore our conclusions are limited by the extrapolation of the two main results of our work (improvement of pCR with ChT–ICI and incremental improvement of pCR with ChT–ICI at higher PD-L1 levels) from two distinct sets of studies. Although preliminary, we hope that our findings will encourage the reporting of the full dataset from KEYNOTE-585 and the other ongoing perioperative phase III studies with chemoimmunotherapy. Making outcomes stratified by PD-L1 cut-offs available to the scientific community, even if exploratory, will provide essential information to benchmark future biomarker-driven trials and guide rational clinical implementation. In conclusion, our study showed that adding ICIs to chemotherapy consistently increases pathological response rates in operable GEA with a benefit that is PD-L1 driven. Mature survival data and enrichment strategies within biomarker-selected trials may help establish the exact added value of ICIs in this context by focusing on a sensitive population sparing non-responders from potential toxicity and risks of delaying or missing surgery. Acknowledgments Funding None declared. Disclosure ECS reports grants and personal fees from BMS and AstraZeneca; personal fees from Amgen, Daiichi-Sankyo, Merck, Viracta, Astellas, Novartis, Pfizer, Zymeworks, and BeiGene; personal fees and nonfinancial support from Mirati; support by the NIHR Biomedical Research Centre at Oxford. All other authors have declared no conflicts of interest. Supplementary data Supplementary Figures and Tables References Associated Data Supplementary Materials Supplementary Figures and Tables