Determinants of Response to Sequential Pembrolizumab with Trastuzumab plus Platinum/5-FU in HER2-Positive Gastric Cancer: A Phase II Chemoimmunotherapy Trial

Abstract

Purpose:

Adding pembrolizumab to first-line fluoropyrimidine (5-FU)/platinum chemotherapy plus trastuzumab improves outcomes in advanced HER2⁺ gastroesophageal adenocarcinomas, but the benefit is largely confined to dual HER2⁺ and PD-L1⁺ patients. To assess the contributions of components, we conducted a phase II trial evaluating 5-FU/platinum/trastuzumab and added pembrolizumab in cycle 2 in patients with metastatic HER2⁺ disease.

Patients and Methods:

Treatment-naïve patients with advanced HER2⁺ gastroesophageal cancer underwent a baseline biopsy and received a single dose of 5-FU/platinum with trastuzumab followed by repeat biopsy. Pembrolizumab was added, and a third biopsy was performed after six cycles. The primary endpoint was the objective response rate. Secondary endpoints included progression-free and overall survival. Exploratory biomarker analysis and dynamic changes in HER2 and PD-L1 were prespecified.

Results:

Sixteen patients were enrolled. The objective response rate was 69%, and the median progression-free survival was 11.9 months. Serial whole-exome, single-cell RNA, T-cell receptor sequencing, and spatial transcriptomics from pretreatment and on-treatment samples revealed early trastuzumab-induced NK cell infiltration in HER2⁺ tumor beds and an increase in Fc receptor gamma III expression in macrophages, suggesting that trastuzumab directs Fc receptor–mediated antibody-dependent cytotoxicity. This favorable remodeling was enhanced by the addition of pembrolizumab, primarily in PD-L1⁺ samples. We observed TGF-β signaling in HER2-negative tumor regions, which was associated with nonresponder status.

Conclusions:

These data highlight the biology of intratumoral heterogeneity and the impact of tumor and immune cell features on clinical outcomes and may partly explain the lesser magnitude of pembrolizumab benefit in HER2⁺ and PD-L1–negative subgroups.

Translational Relevance.

Intratumoral heterogeneity is common across epithelial cancers and represents a barrier to targeted immunotherapies. In a phase II trial of treatment-naïve advanced HER2⁺ gastric cancers, we demonstrate distinct spatially organized biology in HER2+/− and PD-L1+/− tumor regions which may underly clinical response and resistance. These data underscore the importance of understanding biological differences among patient subgroups within our standard of care therapies.

Introduction

HER2 (ERBB2) amplification or overexpression occurs in 15% to 20% of advanced gastric or gastroesophageal junction (GEJ) adenocarcinomas (1, 2). Addition of the anti-HER2 mAb trastuzumab to first-line platinum/fluoropyrimidine (5-FU) chemotherapy improved survival in HER2-positive (HER2⁺) gastric/GEJ cancer (3). The phase III KEYNOTE-811 study recently showed that adding pembrolizumab to first-line trastuzumab and 5-FU/platinum chemotherapy in HER2⁺ gastric/GEJ cancer improved objective response rate (ORR), progression-free survival (PFS), and overall survival (OS) in the intent-to-treat populations (4–6). Although the median PFS and OS were improved in the intent-to-treat population (10.0 vs. 8.1 months; HR, 0.73; P = 0.0002), no improvement in PFS or OS was observed in HER2⁺ and PD-L1–negative patients [PD-L1^–; combined positive score (CPS) <1; PFS 9.5 vs. 9.5 months; HR, 1.03], suggesting differences between double-positive (HER2⁺/PD-L1⁺) and single-positive (HER2⁺/PD-L1^–) HER2⁺ populations (5). The biological hypothesis for cooperation between trastuzumab and anti–PD-1 therapies comes from preclinical observations that trastuzumab therapy increases HER2 cross-presentation on dendritic cells and upregulates T-cell costimulatory molecules via NK cell engagement and IFN-γ secretion (7–10). However, this has not yet been shown in patient samples and there are no studies in HER2⁺ gastroesophageal cancers exploring whether differential transcriptional programs are operating in HER2⁺ and HER2⁻ regions (or PD-L1⁺ vs. PD-L1^–) in the clinical entity defined as HER2⁺ gastric/GEJ adenocarcinoma.

The composition and spatial organization of the tumor microenvironment (TME) are increasingly being appreciated as a determinant of response to biological therapies, including anti–PD-1 agents (11–13). We previously used serial whole-exome sequencing (WES), whole-transcriptome sequencing (WTS), and single-cell RNA sequencing (scRNA-seq) coupled with multiplex IHC to assess the impact of cytotoxic chemotherapy on the TME in gastric cancer (14–16). Our prior phase II trial data were focused on first-line HER2⁻ patients and we observed coordinated and differential TME remodeling between responder and nonresponder patients. In the present study, we used a suite of high-resolution approaches to dissect the TME composition under the therapeutic pressure of sequential HER2 and PD-1 blockade in the HER2⁺ cohort of a multicohort phase II investigator-sponsored trial (NCT04249739). Our central hypothesis was that intrinsic and adaptive features of tumor and immune cells in the TME would determine which patients derive the greatest benefit from combination therapy with anti–PD-1 with trastuzumab. Our aim was to help define the biological mechanisms underlying the clinical observation that adding pembrolizumab to 5-FU/platinum/trastuzumab preferentially benefits patients with dual HER2⁺/PD-L1⁺ tumors (5).

Patients and Methods

Study design and participants

We conducted an open-label nonrandomized investigator-initiated phase II trial in first-line, previously untreated, advanced gastric cancer in Korea. The trial included two parallel cohorts of HER2⁺ and HER2^– patients (ClinicalTrials.gov identifier: NCT04249739). Patients in the HER2⁺ cohort were required to meet the following criteria: (i) at least 19 years of age, (ii) histologically confirmed diagnosis of unresectable, metastatic HER2⁺ gastric cancer [defined per American Society of Clinical Oncology/College of American Pathologists (CAP) guidelines as IHC 3+ or 2+ with a HER2:CEP17 FISH ratio ≥2.0], (iii) adequate organ function per protocol, and (iv) Eastern Cooperative Oncology Group performance status of 0 or 1. The trial protocol was approved by the Institutional Review Board of Samsung Medical Center (no. 2019-11-089) and was conducted in accordance with the Declaration of Helsinki and the Guidelines for Good Clinical Practice. All patients provided written informed consent before enrollment. The detailed protocol is provided in Supplementary Material.

Treatments

All patients received standard trastuzumab at an initial loading dose of 8 mg/kg intravenously, followed by 6 mg/kg every 3 weeks concurrently with 5-FU/platinum chemotherapy. Cisplatin was administered at a dosage of 80 mg/m² intravenously on day 1 for up to eight cycles, and capecitabine 1,000 mg/m² was administered orally twice daily on days 1 to 14 of each 21-day cycle. Pembrolizumab 200 mg intravenously was started at the second cycle [i.e., 3 weeks after the first cycle of cisplatin/capecitabine/trastuzumab (XPH)] and subsequently given every 3 weeks for up to 35 cycles. Capecitabine and trastuzumab were maintained for up to 35 cycles or until disease progression, unacceptable toxicity, or at patient or physician discretion. All patients underwent a pretreatment endoscopic biopsy of the primary tumor (baseline, BL), a repeat biopsy of the same region of the primary tumor after a single dose of chemotherapy plus trastuzumab (follow-up 1, FU1), and a subsequent primary tumor biopsy at the same region six cycles later, reflecting the combination of 5-FU/platinum/trastuzumab + pembrolizumab (follow-up 2, FU2). For patients who remained on therapy, a final primary tumor biopsy was performed at 1 year on treatment (follow-up 3, FU3), independent of their response status (Figs. 1 and 2A).

Figure 1.

Patient disposition in a two-cohort phase II sequential chemoimmunotherapy trials in first-line advanced HER2⁺ and HER2⁻ gastric cancer. HER2⁺ outcomes are reported here; HER2⁻ is included for completeness and used correlative comparisons.

Figure 2.

Sample collection schema and clinical outcomes in a phase II sequential chemoimmunotherapy trial. A, Sample collection schedule in a phase II sequential chemoimmunotherapy trial. B, Waterfall plot with best overall change from baseline among enrolled patients. C, Swimmer plot demonstrating response durability. D, Spider plot demonstrating response kinetics among enrolled patients. E, Patient-level response data by PD-L1 CPS. F, Relationship between the degree of PD-L1 expression and change in tumor volume among PD-L1⁺ patients (Spearman correlation, P < 0.05).

Endpoints

The primary endpoint was the ORR according to RECIST v1.1. Key secondary endpoints included the disease control rate (DCR), response duration, PFS, and OS. Prespecified exploratory biomarker analysis and assessment of dynamic changes in HER2 and PD-L1 expression at serial time points are outlined in the protocol. All biomarker results should be considered exploratory.

Response evaluation

At enrollment, patients were required to have measurable disease according to RECIST v1.1. All patients underwent radiographic evaluation and tumor measurements at BL and every 6 weeks. The treatment response was classified as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). The ORR was defined as the percentage of patients who achieved a CR or PR among all patients. The DCR was defined as the percentage of patients with a CR, PR, or SD. Responder patients were defined as CR/PR and nonresponders as SD/PD according to RECIST v1.1 definitions.

Statistical analysis

A single-stage study design was used to determine the sample size based on the hypothesis that an observed ORR <40% suggests a lack of activity and an improvement in the ORR to ≥60% is of clinical interest. A sample size of 65 patients would provide 90% power to detect the difference at a one-sided significance level of 5%. Considering a 20% drop-out, we planned to recruit up to 78 patients in the multicohort phase II trial (ClinicalTrials.gov ID: NCT04249739). Given that the rate of HER2⁺ gastric cancer is 20%, the HER2⁺ cohort would have to include 15 patients. The HER2⁻ gastric cancer cohort has been described in a recent publication (14) and here we describe the HER2⁺ cohort. Survival was analyzed using Kaplan–Meier plots, Cox proportional hazards models, and the log-rank test to determine P-values. All statistical analyses were conducted using R (v4.1.2, R Foundation for Statistical Computing, www.R-project.org). Two independent groups were compared using the Mann–Whitney U test. All variables between pre- and on-treatment were compared using the Wilcoxon signed-rank test. All P values were two-sided, and results were determined to be significant at P < 0.05. The relationship between variables was quantified using the Pearson correlation, and the coefficient of determination (R) was calculated to assess the strength and direction of correlation.

Tumor sample collection

All primary tissues were collected from patients who had biopsies available, as outlined in Fig. 2A. In an effort to control for intratumoral heterogeneity, tumor tissues were obtained from the primary tumor, which was endoscopically mapped to ensure that the same region was sampled at each time point. Tumor DNA and RNA were extracted using a QIAamp Mini Kit (Qiagen) in accordance with the manufacturer’s instructions for exome and transcriptome sequencing if tumor purity was determined to be >40% following pathologic evaluations. An ND1000 spectrophotometer (NanoDrop Technologies, Thermo Fisher Scientific) was used to measure DNA/RNA concentrations and 260/280 nm and 260/230 nm ratios, and DNA/RNA was quantified using a Qubit fluorometer (Life Technologies, RRID: SCR_020311).

PD-L1 IHC

Tissue samples were cut into 4-μm sections, placed on Superfrost Plus Microscope Slides (Thermo Fisher Scientific), and allowed to dry at 60°C for 1 hour. The Dako PD-L1 IHC 22C3 pharmDx Kit (Agilent Technologies) was used for IHC staining in a Dako Autostainer Link 48 system (Agilent, cat. #SK006, RRID: AB_2889976) using the EnVision FLEX Visualization System. The sections were counterstained with hematoxylin in accordance with the manufacturer’s instructions. PD-L1 protein expression was calculated as the number of PD-L1–stained cells (tumor cells and immune cells) divided by the total number of viable tumor cells multiplied by 100. Samples were deemed PD-L1⁺ if CPS ≥1, consistent with the manufacturer’s guidelines and phase III clinical trials (5, 17).

WES and WTS

Genomic DNA (gDNA) was extracted from the tumor tissues and matched blood samples using a QIAamp DNA Blood kit (Qiagen). Standard exome capture libraries were generated using 1 μg of gDNA and the Agilent SureSelect Target Enrichment procedure for Illumina paired-end sequencing library construction. The SureSelect Human All Exon V6 probe set was used for each sample. The DNA was qualified and quantified using agarose gel electrophoresis and PicoGreen, respectively. As per the manufacturer’s instructions, 1 μg of gDNA was diluted in elution buffer and sheared into 150 to 200 bp fragments using an LE220 ultrasonication device (Covaris). The fragments were ligated with Agilent adapters and amplified with PCR. The libraries were quantified using TapeStation DNA ScreenTape D1000 (Agilent).

For exome capture, according to the Agilent SureSelect Target Enrichment protocol, 250 ng of the DNA library was mixed with hybridization buffer, blocking mixes, RNase block, and 5 μg of SureSelect All Exon Capture Library. The capture baits were hybridized at 65°C for 24 hours in a thermal cycler with the lid heated at 105°C. The extracted DNA was cleaned and amplified. The final purified product was certified using TapeStation DNA ScreenTape D1000 and quantified using KAPA Library Quantification kits for Illumina sequencing platforms according to the qPCR Quantification Protocol Guide. Flow-cell clusters were generated by multiplexing samples and using TruSeq Rapid SBS and Cluster kits (Illumina). The Illumina HiSeq2500 platform (Illumina) was used for paired-end (2 × 100 bp) sequencing.

ScRNA-seq and single-cell T-cell receptor sequencing

To prepare single-cell suspensions, tumor tissues were dissociated using a gentleMACS Dissociator and tumor-infiltrating lymphocyte kit (both from Miltenyi Biotec) according to the manufacturer’s protocol. The cells were cryopreserved in liquid nitrogen until use. All samples showed a viability of approximately 90% after thawing. The 5′ gene expression profiling was conducted using Chromium Single Cell V(D)J Solution from 10× Genomics according to the manufacturer’s instructions. For each sample, up to 8,000 cells were loaded onto a cartridge (10× Genomics). Cell-barcoded 5′ gene expression libraries were constructed using Chromium Next GEM Single Cell 5′ Reagent kits and sequenced on a NovaSeq 6000 platform (Illumina, RRID: SCR_016387) at a depth of approximately 50,000 reads per cell. CellRanger (10× Genomics) was used to map the libraries to the GRCh38 human reference genome. T cells were enriched prior to scRNA-seq.

WES data analysis

WES data were aligned to the GRCh37 human reference genome using BWA-MEM (18). The Genome Analysis Toolkit (v4.1.1.02) was used for preprocessing to create BAM files for analysis. The preprocessing comprised base recalibration, indel realignment, and duplicate marking. To increase the sensitivity for detecting both lower and higher allele frequencies of somatic variants in a given tumor and associated normal BAM data at the genomic locus, we used union variant calls from two tools, MuTect2 (19) and Strelka2 (20) with default parameters. Both variant callers were run using data for known polymorphic sites from HapMap (phase III; ref. 21), 1000G (phase I; ref. 22), and Single Nucleotide Polymorphism Database (v138; ref. 23). Using the GRCh37 database, variants filtered based on a depth of ≥5 and ≥2 alternative alleles were annotated using Ensembl Variant Effect Predictor (v87; RRID: SCR_005194; ref. 24).

WTS data analysis

We used STAR (v2.6.1; ref. 25) to match the annotated RNA sequence reads to the human reference genome (GRCh38) after annotating them using ENSEMBL (RRID: SCR_002344; v98). Using the parameters suggested in the Genotype-Tissue Expression (GTEx) project, we quantified gene expression as transcripts per million using RSEM (v1.3.1; ref. 26). Expression signatures for immune cell and CD8 T-cell classification were computed using gene set variation analysis (ref. 27).

scRNA-seq analysis

The scRNA-seq data from all samples were merged in R v.4.1 using the Seurat package v.4.0 (28). Doublets were excluded using Scrublet (RRID: SCR_018098; ref. 29). In addition, cells exhibiting a high mitochondrial read fraction (>30%) and low-quality libraries (<400 genes) were excluded. After scaling and standardizing the data from each sample, principal component analysis was performed. Batch correction was conducted using Harmony (RRID: SCR_022206; ref. 30). Based on previous research, we used shared closest neighbor modularity optimization–based clustering to categorize the cell types (14). To visualize Uniform Manifold Approximation and Projection plots and T-cell and macrophage lineages, we employed Monocle3 (RRID: SCR_018685; ref. 31) v.1.2.8. The UCell R package was used to estimate module scores of tumor-reactive T cells within the tumor and in adjacent and distant healthy tissues.

T-cell receptor repertoire analysis

We profiled the T-cell receptor (TCR) repertoire from bulk RNA sequences using TRUST4 (RRID: SCR_026162; ref. 32) with default parameters. To compare CDR3 clonotypes among the tumor samples, we quantified clonal abundance using the TRUST4 report. Samples with less than five diverse TCRβ clonotypes were excluded to increase accuracy and avoid ambiguous results. In addition, we measured clonal abundance from the filtered contig annotation data and profiled the TCR repertoire using single-cell TCR sequences. Unclear contigs mismatching with T cells were excluded for accuracy. To calculate TCRβ clonality, we used the following clonality score equation developed from the Shannon clonality:

$$1\mathrm{-}\frac{\mathrm{-}{\displaystyle \mathrm{\sum }}_{i\mathrm{=}1}^n{X}_i{\log }_2\left[X_i\right]}{{\log }_{2}n}$$

in which X_i is the frequency of the TCR sequence and n is the number of unique TCR sequences in the repertoire.

Digital spatial profiling and data processing

Thirteen HER2⁺ gastric cancer tumors were processed into formalin-fixed, paraffin-embedded blocks that were cut into 5-µm-thick sections, which were mounted on glass slides according to the manufacturer’s instructions. After baking at 60°C for 2 hours, the tissue sections were deparaffinized using Neo-Clear (Sigma-Aldrich) and rehydrated in descending grades of ethanol. Antigen retrieval was achieved using Tris-EDTA at high pressure and high temperature. Next, the sections were incubated with a whole-transcriptome analysis panel (NanoString Technologies, RRID: SCR_023912) in a humid chamber at 37°C overnight. The sections were stained for morphologic markers pan cytokeratin (panCK; NanoString Technologies), CD45 (NanoString Technologies), and HER2 (BioLegend, cat. #324412, RRID: AB_2262300). Nuclei were stained with SYTO13 green fluorescent nucleic acid stain (NanoString Technologies). The stained sections were scanned using a GeoMx Digital spatial profiling system (33). Regions of interest (ROI) were selected based on morphologic marker signals. We generated areas of illumination (AOI) within each ROI to profile the architectures of HER2⁺ and HER2^– tumor cells and immune cells. For each patient, we selected 1 to 10 ROIs. The sections were sequenced and subjected to barcode-conjugated RNA ISH (33). The nuclear, epithelial, immune, and HER2⁺ compartments were delineated based on the above four markers. Immunofluorescence imaging, ROI selection, segmentation into marker-specific AOIs, and spatially indexed barcode cleavage and collection were performed using a GeoMx DSP Spatial Profiler (NanoString Technologies, RRID: SCR_021660). The exposure times were 100 ms for SYTO13 and 300 ms for panCK, CD45, and CD340. Spatially indexed oligomers were analyzed in 2 to 10 ROIs and 6 to 22 AOIs per specimen. Illumina adapter sequences and dual sample indices were added using PCR amplification during library preparation in accordance with the manufacturer’s recommendations. All Whole Transcriptome Analysis (WTA) AOIs were sequenced with at least 150 to 200 reads per square micron of AOI using a NextSeq 550 System (RRID: SCR_016381). After sequencing, FASTQ data for digital spatial profiling (DSP) were combined to create count matrices. Unique Molecular Identifier (UMI) and molecular target tag sequences were used to compute deduplicated sequencing counts. The limit of quantitation was estimated as the geometric mean plus two standard deviations of the negative control probes. The datasets were normalized using upper quartile (Q3) normalization, and targets that fell below the quantitation limit were discarded.

Data availability

The DNA and RNA sequencing data generated in this study are publicly available in the European Nucleotide Archive under accession PRJEB60680. All other data used for figures and tables are provided within the article and Supplementary Tables and Figures. The NanoString DSP data are available upon written request from the corresponding authors.

Results

Patients

Between June 2021 and February 2023, 16 patients with HER2⁺ gastric cancer were enrolled and treated. The enrollment schema and sample collection schedule are shown in Figs. 1 and 2A, respectively. Patient and tumor characteristics are summarized in Table 1. The mean age was 60 years (range, 36–76 years), and 87.5% of patients were male. At BL pretreatment analysis, all patients had IHC3⁺ HER2 expression and HER2 amplification. No patient had mismatch repair deficiency; one patient (S93) was microsatellite instability indeterminate. Prior to therapy, 13 patients (81.25%) had PD-L1⁺ (CPS ≥1) tumors and three patients had PD-L1⁻ (CPS <1) tumors. The median CPS score was 5 (range, 2–100; Table 1).

Table 1.

Clinicopathologic features and patient characteristics from enrolled HER2⁺ patients (N = 16).

Characteristic	Value
Age, years [mean ± SD (range)]	60.2 ± 11.92 (36–76)
Sex
Male	14 (87.5%)
Female	2 (12.5%)
ECOG PS
0	2 (13%)
1	14 (87%)
Primary tumor location
Cardia	3 (19%)
Body	7 (44%)
Angle	1 (6%)
Antrum	5 (31%)
Pathology
Moderately differentiated	10 (63%)
Poorly differentiated	5 (31%)
Signet ring cell carcinoma	1 (6%)
Prior gastrectomy
Curative intent	2 (13%)
HER2 status
IHC 3+	16 (100%)
EBV positivity
Positive	1 (6%)
Negative	15 (94%)
PD-L1 positivity (22C3)
Positive (CPS ≥1)	13 (81%)
Median CPS (range)	5 (2–100)
CPS ≥10	5 (31.3%)
Negative (CPS = 0)	3 (19%)
MSI status
MSS	15 (94%)
Indeterminate	1 (6%)
Sites of metastatic lesion
Liver	9 (56%)
Lymph node	11 (69%)
Peritoneum	7 (44%)

Abbreviations: EBV, Epstein-Barr virus; ECOG, Eastern Cooperative Oncology Group; MSI, microsatellite instability; MSS, microsatellite stable; PS, performance status.

Clinical outcomes

All patients had measurable lesions, and the ORR (primary endpoint) was 68.8% [11 of 16 patients; 95% confidence interval (CI), 54.3–95.9], with a DCR of 87.5% (14 of 16 patients; 95% CI, 69.7–99.8; Fig. 2B; Table 2). A numerical tumor decrease was observed in all patients (Fig. 2B), and nine (56.3%) patients achieved a >50% target reduction (Fig. 2B–D). One patient (S45) showed PD at the first response evaluation, with newly developed bone metastasis and progression of pleural seeding despite marked shrinkage of the target lesion. Patient S74 showed a CR of the measurable lesion but still had nonmeasurable lesions. At the data cutoff on August 1, 2024, the median PFS was 11.9 months (95% CI, 5.12–18.28) and the median OS was not reached. Among responders, the median time to response was 1.2 months (range, 1.0–2.6) and the median duration of response was 5.5 months (2.5–15.7). The 11 patients who achieved PR (responders) had longer PFS (log-rank test, 12.1 vs. 4.0 months, P < 0.01) and OS [nonresponders (NR) vs. 9.2 months, P < 0.05] than nonresponders (SD and PD; Supplementary Fig. S1A and S1B).

Table 2.

Clinical activity of pembrolizumab plus trastuzumab and chemotherapy in patients with HER2-positive gastric cancer.

Best response	XPHP (N = 16)
	Numbers	%
CR	—	—
PR	11	68.75
SD	3	18.75
PD	2	12.5
ORRa (95% CI)	13	68.75 (54.3–95.9)
DCRb (95% CI)	14	87.50 (69.7–99.8)

Abbreviations: CR, complete response; DCR, disease control rate; ORR, objective response rate; PD, progressive disease; PR, partial response; SD, stable disease; XPHP, capecitabine, cisplatin, trastuzumab, pembrolizumab.

^aORR is defined as the proportion of patients with CR or PR as best overall response.

^bDCR is defined as the proportion of patients with CR, PR, or SD as best overall response.

Among all patients enrolled (N = 16), five patients had died, and three patients continued to receive the study treatment at data cutoff (Fig. 2C and D). Treatment was ongoing in three patients, and a patient (S92) who achieved PR stopped treatment after cycle 7 because of planned conversion gastrectomy. Seven patients discontinued treatments because of disease progression, and 11 patients were alive at the data cutoff. Eleven (84.6%) of thirteen patients who had PD-L1⁺ tumors were responders, whereas all patients with PD-L1^– (N = 3, 100%) were nonresponders (Fig. 2E). Among PD-L1⁺ tumors, the degree of PD-L1 expression correlated with the radiographic change in tumor volume (Fig. 2F; Spearman correlation, P < 0.05). In line with published data, patients with PD-L1⁺ tumors had longer PFS (log-rank test, 12.1 vs. 4.0 months, P < 0.05) and OS (NR vs. 9.2 months, P > 0.05) than patients with PD-L1⁻ tumors (N = 3). Among responders, the median OS of patients with PD-L1⁺ tumors was not significantly higher than PD-L1⁻ tumors (Supplementary Fig. S1C and S1D). There were no new safety signals, and adverse events of any grade, most commonly, nausea (81%), anorexia (44%), hand–foot syndrome (44%), and a decreased neutrophil count (44%), occurred in all patients. Events of grade 3 or higher, most commonly, a decreased neutrophil count (38%), occurred in seven (44%) patients (Supplementary Table S1). No treatment-related fatal adverse event was reported.

Biomarker and molecular evolution during chemoimmunotherapy

The pretreatment genomic landscape of HER2⁺ gastric cancer in our cohort was in line with the literature (1, 34–37). Serial biopsies were collected from the primary tumor in the same endoscopically mapped tumor region for nine patients. Among them, five patients remained HER2⁺ (IHC 3+; 63%), whereas HER2 expression was no longer observed in three patients (38%) after trastuzumab in line with the literature (Supplementary Table S2; refs. 38, 39). At BL, 13 of 16 (82%) patients were PD-L1⁺ (CPS ≥1). At FU2 (after pembrolizumab addition), loss of PD-L1 expression was observed in two of nine patients (22%) and PD-L1 expression was numerically decreased in four of nine patients (44%) compared with the levels at BL. Across the entire cohort, the mean PD-L1 CPS score decreased, but not significantly, after XPH + pembrolizumab treatment (from 18.6 at BL to 7.6 at follow-up, P = 0.175). No associations between HER2 and PD-L1 expression dynamics and outcomes were observed.

Trastuzumab induces Fc receptor–mediated cytotoxicity and activates an innate immune response

The primary mechanism of action of trastuzumab in gastric/GEJ cancer is thought to involve Fc receptor (FcR)–mediated antibody-dependent cellular cytotoxicity (ADCC) via NK cell engagement, and FcR⁺ immune cells are required for the tumor response to trastuzumab (10, 40,41). Leveraging the previously reported HER2⁻ patients, we compared matched time points at BL and FU1 (after one dose of platinum/5-FU in HER2⁻ and one dose of platinum/5-FU/trastuzumab in HER2⁺) to gauge differential local immune responses in HER2⁺ and HER2⁻ gastric cancer (Supplementary Fig. S2A). We analyzed paired scRNA-seq data from 145 samples in total (BL tumor, N = 35), BL normal (adjacent normal, N = 12 and distant normal, N = 12), FU1 tumor (N = 34), FU1 normal (adjacent normal, N = 12 and distant normal, N = 12), and FU2 tumor (N = 28). We performed semi-unsupervised learning according to our previous scRNA-seq study (14) and defined six major cell types (tumor, stromal, T lymphocytes, NK, B lymphocytes, and myeloid cells) among 373,229 cells (Supplementary Fig. S2B and S2C). From these data, we observed a higher NK:tumor cell ratio in HER2⁺ patients (N = 6) after one cycle of platinum/5-FU/trastuzumab versus HER2⁻ patients (N = 29) after one cycle 5-FU/platinum (Fig. 3A; Wilcoxon signed-rank test, P = 0.063). The NK:tumor cell ratio did not significantly differ between HER2⁺ and HER2^– at BL, but we observed NK cell recruitment during treatment, which is a known feature of ADCC, presumably related to trastuzumab. To broadly assess features associated with immune activation, we performed WTS and conducted gene set variation analysis of samples at BL, FU1, FU2, and FU3 for immunogenic cell death, pro-inflammatory macrophages, and cyclic GMP–AMP synthase (cGAS) and STING (also known as TMEM173) signatures. Activation of these pathways was higher in HER2⁺ tumors than in HER2⁻ tumors treated with the same classes of chemotherapy (5-FU/platinum; Fig. 3B–D). Notably, at FU2 (after the addition of pembrolizumab), cGAS–STING pathway expression was significantly increased in HER2⁺ tumors only (Fig. 3D; Wilcoxon signed-rank test, BL vs. FU1, P < 0.05; FU1 vs. FU2, P < 0.01; and BL vs. FU2, P = 0.001). We considered the HER2⁻ cohort with matched sample collection as a strength to allow validation, as there are no other public scRNA-seq datasets available for HER2⁺ patients.

Figure 3.

Trastuzumab plus 5-FU/platinum chemotherapy induces features of ADCC after a single dose in HER2⁺ GC. A, NK cell increase after a single dose of 5-FU/platinum/trastuzumab in HER2⁺ GC by scRNA-seq (Wilcoxon signed-rank test). B, Induction of immunogenic cell death module score during treatment in HER2⁻ and HER2⁺ GC (Wilcoxon signed-rank test). C, Pro-inflammatory macrophage module score evolution during sequential chemoimmunotherapy (Wilcoxon signed-rank test). D, cGAS–STING changes in HER2⁺ and HER2⁻ GC during therapy (Wilcoxon signed-rank test). E, UMAP embedding of scRNA-seq–defined TAM clustering and changes between BL and FU1. M1-like and M2-like clusters labeled. F, Changes in scRNA-seq–derived M1/(M1 + M2) ratio after a single dose of 5-FU/platinum (HER2^–) or 5-FU/platinum + trastuzumab (HER2⁺) therapy. G, Increase in Fc gamma receptor IIIb and IIIa expression on macrophages after trastuzumab treatment (Wilcoxon signed-rank test). H, FCGR3A and FCGR3B expression level across major cell lineages highlights FCGR3B expression primarily in macrophage subsets. DC, dendritic cell; GC, gastric cancer; MAIT, mucosal-associated invariant T cell; UMAP, Uniform Manifold Approximation and Projection.

Next, we refined the tumor-associated macrophage (TAM) clusters into five main subgroups using a broad M1-like and M2-like classification as used in our previous study (Fig. 3E; Supplementary Fig. S2D). HER2⁺ patients showed an increased M1/(M1 + M2) ratio after one XPH cycle (Fig. 3F; Wilcoxon signed-rank test, P = 0.038) and a higher ratio at BL than HER2⁻ patients (Wilcoxon signed-rank test, P = 0.356). There were no baseline differences in M1- or M2-like TAMs or other cell types between HER2⁺ and HER2⁻ cohorts (Supplementary Fig. S3A–S3C). Notably, cells from HER2⁺ samples at FU1 contributed primarily to cluster 1 (Fig. 3E). This cluster was marked by high expression of FCGR3B (also known as FcRIIIb and CD16b), which is the IgG receptor and an important modulator of HER2-directed antibody therapy (Supplementary Fig. S2D; ref. 42). This observation was supported when exploring FcR expression across major cell lineages in the HER2⁺ patients (Fig. 3G and H). As HER2 or broader MAPK/RTK pathway inhibition may influence the TME composition, we assessed the expression of MAPK/RTK genes and observed no difference between HER2⁺ and HER2⁻ patients (Supplementary Fig S3D), but we noted a decrease in expression after trastuzumab in both responders and nonresponders, which was only sustained in responding patients (Supplementary Fig S3E). Although limited by patient numbers, our observations suggest trastuzumab-driven ADCC may contribute to treatment efficacy in HER2⁺ patients (Fig. 3G; Wilcoxon signed-rank test, P < 0.001).

Dual HER2⁺ and PD-L1⁺ tumor cells colocalize with tumor-reactive T cells

To assess the interplay among and spatial orientation of HER2, PD-L1, and immune cell subsets during therapy, we conducted DSP (NanoString GeoMx human whole-transcriptome atlas) of primary tumors from 13 HER2⁺ patients (total 54 ROIs). We selected ROIs based on four-color immunofluorescence staining of nuclei, panCK, HER2, and CD45 (Fig. 4A; Supplementary Fig. S4A). For each patient, we selected 1 to 10 ROIs (23 ROIs in total in nonresponders and 31 in responders). PD-L1 expression based on clinical IHC (gold standard benchmark) was used to classify samples as PD-L1⁺ or PD-L1^–. Despite all samples being HER2 IHC 3+ according to standard IHC, HER2⁺ and HER2⁻ tumor cells were detected in all samples based on HER2 RNA expression consistent with transcriptional heterogeneity. CD274 (PD-L1) mRNA levels were higher in HER2⁺ tumor cells than in HER2⁻ tumor cells in PD-L1⁺ patients (Fig. 4B; Wilcoxon signed-rank test, P = 0.003). We next measured CD274 expression over time in HER2⁺ tumor and immune cells. CD274 expression increased after 5-FU/platinum/trastuzumab and after subsequent pembrolizumab addition, but only in IHC-defined PD-L1⁻ patients (N = 3; Fig. 4C and D; Wilcoxon signed-rank test, HER2⁺ tumor BL vs. FU2, P = 0.025 and Wilcoxon signed-rank test, immune cells BL vs. FU2, P = 0.08). Consistent with the correlation between CD274 expression and PD-L1 IHC, the CD274 mRNA level at BL in PD-L1⁺ patients was higher than that at BL in PD-L1⁻ patients (Fig. 4C and D; Mann–Whitney U test, HER2⁺ tumor cells, HER2^– vs. HER2⁺, P = 0.04 and Mann–Whitney U test, immune cells HER2⁻ vs. HER2⁺, P = 0.04). Although the sample size was small (PD-L1⁻ patients, N = 3, and PD-L1⁺ patients, N = 10), this result implies that chemoimmunotherapy with trastuzumab may induce PD-L1 upregulation, possibly via cGAS–STING–mediated IFN-γ signaling.

Figure 4.

DSP highlights interplay between HER2, PD-L1, and immune cell subsets during therapy. A, Representative whole slide from patient S69 demonstrating ROI selection for DSP analysis. B, Differences in DSP-derived tumoral CD274 expression among HER2⁺ and HER2⁻ cells in patient determined to be PD-L1⁺ or PD-L1⁻ by IHC testing (gold standard). C, Changes in CD274 expression in HER2⁺ tumor cells over time in patient determined to be PD-L1⁺ or PD-L1⁻ by IHC testing (gold standard). D, Changes in CD274 expression in CD45⁺ immune cells over time in patient determined to be PD-L1⁺ or PD-L1⁻ by IHC testing (gold standard). E, CD8 TCR gene score differences in pretreatment BL and during therapy in HER2⁺ GC. F, Relationship between PD-L1 expression level by IHC (22C3, CPS testing) and CD8 TCR signaling gene score (GSVA) with clinical response in HER2⁺ GC (Spearman correlation). GC, gastric cancer.

To gain a deeper understanding of the association between the PD-L1 CPS and tumor-specific immune response, we examined curated gene scores for CD8 TCR signaling (MSigDB; ref. 43). CD8 TCR signaling was higher at BL and FU1 in PD-L1⁺ patients than in PD-L1⁻ (Fig. 4E; Mann–Whitney U test, PD-L1⁻ at BL vs. PD-L1⁺ at BL, P = 0.026; PD-L1⁻ at FU1 vs. PD-L1⁺ at FU1, P = 0.006). The IHC-defined (gold standard) PD-L1 CPS score was significantly correlated with the CD8 TCR signaling gene score (Fig. 4F; Spearman correlation, P < 0.05). Notably, among PD-L1⁺ patients, those who had a CPS ≥15 tended to show the highest CD8 TCR signaling expression and particularly good outcomes. Together, these findings suggest that in advanced HER2⁺ gastric cancer, PD-L1 IHC likely reflects an intrinsic immunogenicity as evidenced by enhanced CD8 TCR signaling, which is a more favorable background to introduce T cell–centric anti–PD-1 approaches like pembrolizumab.

Given the roles of TAMs in modulating tumor immune responses and FcR-mediated phagocytosis, we hypothesized that M2-like macrophage predominance may underlie the limited response in some patients, even in PD-L1⁺ patients. First, we examined gene scores for FCGR-dependent phagocytosis from our single-cell data. FcR-dependent phagocytosis scores were higher at BL and FU1 in PD-L1⁺ patients than in PD-L1⁻ patients (Supplementary Fig. S4B; Mann–Whitney U test, PD-L1^– at BL vs. PD-L1⁺ at BL, P = 0.038; PD-L1^– at FU1 vs. PD-L1⁺ at FU1, P = 0.0006). Among PD-L1⁺ patients, the FCGR score increased during treatment [Supplementary Fig. S4B, median g-score at BL, 2.28 (range, 0.73–3.00); median g-score at FU1, 2.44 (0.77–3.06); median g-score at FU2, 2.49 (1.63–3.11); and median g-score at FU3, 2.64 (1.20–3.16)]. Although an overall relationship existed between PD-L1 expression and FCGR TAM scores, two of 13 patients (15%; patients S90 and S52) had a CPS >5 but were primary resistant to triplet combination. Patient S90 had a CPS score of five but had high expression of the immunosuppressive marker CD163 (classic M2-like marker) in multiple ROIs according to the DSP data (Supplementary Fig. S4C). Collectively, these findings suggest that even within the favorable PD-L1⁺ group, the composition of immune cells around the tumor is heterogeneous, and the interplay between CD8 T cells and M2-like macrophages and the tumor may influence the clinical response to treatment.

Tumor-reactive T cells are expanded during treatment in HER2⁺/PD-L1⁺ tumors

To investigate the associations between CD8 TCR signaling and CD8 T-cell phenotypes and HER2 and PD-L1 protein expression, we selected immune regions in the DSP data. Effector CD8 signatures were enhanced in immune cells (Fig. 5A, Wilcoxon signed-rank test, PD-L1⁻ patients BL vs. FU1, P = 0.023 and PD-L1⁺ patients BL vs. FU2, P = 0.041). Exhausted CD8 signatures were increased only in PD-L1⁺ patients after treatment with 5-FU/platinum/trastuzumab (Wilcoxon signed-rank test, BL vs. FU1, P = 0.043 and BL vs. FU2, P = 0.028). Similarly, a positive correlation between exhausted CD8 and CD8 TCR signaling was observed only in PD-L1⁺ patients (Fig. 5B; Spearman correlation, P = 6.8e−05). Analysis of the bulk TCR data using the TRUST4 algorithm revealed that exhausted CD8 T cells were correlated with TCR repertoire clonality (Fig. 5C; Spearman correlation, P = 0.037). To elucidate the relationship between CD8 TCR and PD-L1⁺, we profiled T cells by matching the scRNA-seq and paired single-cell TCR sequencing datasets to tumor time points and sample sources (Supplementary Fig. S5A–S5C). Most expanded T-cell clones reside in the CD8 clusters (Supplementary Fig. S5C; ref. 44).

Figure 5.

Tumor-reactive T-cell subsets are expanded after trastuzumab and pembrolizumab in HER2⁺ GC. A, DSP-derived changes in major T-cell subsets during therapy stratified by PD-L1 status in HER2⁺ GC. B, DSP-derived CD8 TCR signaling is correlated with exhausted T-cell phenotype in PD-L1⁺ patients. C, Bulk TCR sequencing relationship between clonality and exhaustion score in HER2⁺ GC. D, CD8 T-cell subset proportions obtained from scRNA-seq and shown by timepoint in therapy. E, Comparison of tumor-reactive T-cell module scoring between PD-L1⁺ and PD-L1⁻ patients (Mann–Whitney U test). F, Difference in CXCL13 expression, derived from DSP, between clinical responder and nonresponder patients (Mann–Whitney U test). GC, gastric cancer.

We next explored T-cell evolution during sequential therapy in HER2⁺ gastric cancer. We isolated granular CD8 T-cell subsets, which revealed that the PD-L1⁺ group had expanded populations of exhausted CD8 T cells during treatment [Fig. 5D; median proportion of PD-L1^– group (N = 2) at BL, 12.42% range, (7.94%–16.90%); at FU1, 9.98% (7.18%–12.78%); and at FU2, 4.35% vs. median proportion of PD-L1⁺ group (N = 4) at BL, 24.08% (17.85%–37.93%); at FU1, 17% (13.73%–23.77%); and at FU2, 23.25% (16.35%–30.15%)]. We leveraged several validated gene signatures used to predict tumor-reactive T-cell populations to assess differences in the abundance of the tumor-reactive T-cell population over time (31). We first confirmed that tumor-reactive CD8 T cells were enriched in tumor versus adjacent and distant normal tissues based on a scoring using the UCell algorithm (Supplementary Fig. S5D; Wilcoxon signed-rank test, tumor vs. adjacent normal, P < 0.001; tumor vs. distant normal, P < 0.001; and adjacent normal vs. distant normal, P < 0.001). Corroborating a greater pre-existing immune response in PD-L1⁺/HER2⁺ gastric cancer, the tumor-reactive module score at BL was higher in the PD-L1⁺ group (Fig. 5E; Mann–Whitney U test, P = 0.002). As CXCL13 expression is the strongest feature predicting tumor-reactive T cells, we explored CXCL13 expression differences between clinical responder and nonresponder patients. CXCL13 expression was higher in responders than in nonresponders and largely confined to T-cell subsets vs other immune cells (Fig. 5F; Supplementary Fig. S4E–S4H; Mann–Whitney U test, P < 0.05; refs. 45, 46). We additionally compared patient-level DSP-derived expression of other validated tumor-reactive markers, including ENTPD1 (CD39) and PDCD1 (PD1; Supplementary Fig. S5F). Finally, we compared tumor reactivity and TCR clonality in the T-cell compartment among individual patients. Changes in the tumor-reactive T-cell signature paralleled scRNA-seq–derived TCRβ clonality during treatment (Supplementary Fig. S5I and S5J). Although responders (patients S39 and S40) had increased tumor-reactive scores and TCRβ clonality after adding pembrolizumab (FU2), one nonresponder (S47) had a decreased tumor-reactive score and TCRβ clonality at FU2. As a pre-existing TCR repertoire is associated with immunotherapy response in melanoma, we explored patient-level TCR repertoire evolution in our cohort (47). We only had scRNA-seq data at FU2 (after pembrolizumab addition) from patients who had complete serial time points (BL, FU1, and FU2). The proportion of the pre-existing TCR repertoire increased in all patients after pembrolizumab addition (Supplementary Fig. S4K). Although one nonresponder (S47) shared <5% of the pre-existing TCR clones between BL and FU2, responders (S39 and S40) shared a larger portion of clones during treatment, suggesting expansion of a tumor-specific population (Supplementary Fig. S4L).

Elevated ERBB2 expression correlates with response, and TGF-β signaling is increased in HER2^– regions in nonresponders

We analyzed the somatic genomic landscapes of eight patients [five responders (S39, S40, S51, S61, and S65) and three nonresponders (S45, S47, and S52)]. Overall, genomic changes were limited over time, particularly among nonresponders (Fig. 6A). Similarly, the tumor mutation burden did not significantly decrease after one cycle of XPH (Fig. 6B; Wilcoxon signed-rank test, BL vs. FU1, P = 0.072). We examined HER2 copy number variation (CNV) and found a significant reduction in log₂(CNV) after adding pembrolizumab (Fig. 6C; Wilcoxon signed-rank test, BL vs. FU2, P = 0.018), which was consistent with tumor responses. To assess the concordance between HER2 CNV and HER2 protein expression based on multiplex immunofluorescence, we segmented the panCK⁺/HER2⁻ compartment from the panCK⁺/HER2⁺ compartment (Supplementary Fig. S6A). We examined serial time points in representative nonresponders (S45 and S47) and found limited changes in HER2 CNV and HER2 expression during treatment (Supplementary Fig. S6B and S6C). Conversely, representative responders (S39 and S40) showed clearance of detectable HER2 CNV and regression of HER2⁺ tumors in multiplex immunofluorescence at paired time points (Supplementary Fig. S6D and S6E). To orthogonally validate these findings, we analyzed scRNA-seq data from pre- and on-treatment samples. The results suggested that higher ERBB2 expression at BL was associated with favorable outcomes (Fig. 6D; Mann–Whitney U test, P < 2.2e−16; Supplementary Fig. S7A). During treatment (FU1 and FU2), ERBB2 expression decreased in responders (Fig. 6D; Wilcoxon signed-rank test, BL vs. FU1, P < 2.2e−16 and BL vs. FU2, P < 2.2e−16). In line with the scRNA-seq–derived findings, BL ERBB2 expression in HER2⁺ tumor cells was higher in responders than in nonresponders in our spatial transcriptomic DSP data (Fig. 6E; Mann–Whitney U test, P = 0.005). Following therapy, ERBB2 expression was decreased in both responders and nonresponders, which was largely consistent with antitumor activity, even when a clinical response per RECIST v1.1 was not achieved (Wilcoxon signed-rank test, P < 0.05). Similarly, bulk RNA-seq data indicated that ERBB2 expression above the median transcripts per million of 175.43 was prognostic in the overall population (Supplementary Fig. S7B; log-rank test, NR vs. 3.1 months, P = 0.004).

Figure 6.

Molecular features associated with response to 5-FU/platinum with trastuzumab and pembrolizumab in HER2⁺ GC. A, Tile plot derived from WES data highlighting genomic alterations across time. B, Changes in WES-derived tumor mutation burden (TMB) during therapy. C, Differences in HER2 CNV calls over time among enrolled patients with available WES data. D, ERBB2 (HER2) expression level from scRNA-seq differences between responder and nonresponder patients. E, Expression differences in ERBB2 from DSP during therapy and stratified by clinical response in HER2⁺ GC. F, Differences in the TGF-β signaling module score in HER2⁻ tumor cells determined by DSP between responder and nonresponder patients. G, Machine learning model–derived key features differentiating responder and nonresponder patients with HER2⁺ GC. GC, gastric cancer.

Consistent with the intratumoral HER2 transcriptional heterogeneity, we observed HER2⁺ and HER^– regions in DSP. To understand the transcriptional programs operating in HER2⁻ (according to expression) tumor clones within the HER2⁻ compartment defined by the spatial data, we performed gene set enrichment analysis (Supplementary Tables S3 and S4). In the HER2⁻ compartment, TGF-β signaling was enriched in nonresponder patients at BL and persisted after the addition of pembrolizumab (FU2; Fig. 6F and G; Mann–Whitney U test, P < 0.05). Within the HER2⁺ compartment, we also observed persistent TGF-β signaling in nonresponders at FU2 (Supplementary Fig. S7C, Mann–Whitney U test, P < 0.05). Finally, using a machine learning model, we assessed the strongest features distinguishing responder from nonresponder patients in our data. TGF-β signaling in HER2⁻ tumor regions, ERBB2 expression level in HER2⁺ tumor cells, PD-L1 CPS, and the tumor-reactive T-cell marker (CXCL13) expression in immune cells were most strongly associated with the clinical response (Fig. 6G).

Discussion

Patients with HER2⁺ gastroesophageal cancers have achieved high response rates and improved survival with trastuzumab + 5-FU/platinum-based therapies (3). In dual HER2⁺ and PD-L1⁺ patients, the addition of pembrolizumab improved PFS and OS in the phase III KEYNOTE-811 trial, but survival benefit was limited to dual HER2⁺/PD-L1⁺ patients and the US FDA labeling restricts this combination to dual HER2⁺ and PD-L1⁺ patients (5, 6). As we move toward building on the standards established in the KEYNOTE-811 trial, understanding the mediators of response and resistance is key, especially considering the differential benefit between dual-positive (HER2⁺/PD-L1⁺) and single-positive (HER2⁺/PD-L1-) patients. To help understand the biology underlying the differential clinical activity, we conducted a parallel cohort phase II trial in first-line advanced gastric cancer with sequential pembrolizumab and comprehensive correlative studies using serial biopsies. The ORR of 68.8% and median PFS of 11.9 months are consistent with previous trial data, and our patient population is representative of HER2⁺ gastric cancer (Supplemental Table S5; refs. 4, 5, 35, 48). To our knowledge, this is the first in-depth multiparametric molecular characterization of HER2⁺ gastric cancer using tumor samples collected serially during treatment with 5-FU/platinum/trastuzumab plus pembrolizumab and the first reported scRNA-seq and spatial transcriptomic data in HER2⁺ gastric cancer.

Although HER2 overexpression is associated with increases in downstream PI3K and MAPK activity, trastuzumab is not a potent inhibitor of HER2 signaling (49). In fact, the primary mechanisms of trastuzumab activity are likely ADCC and complement-dependent cytotoxicity (41, 50, 51). Therefore, we hypothesized that measures of ADCC induction would be increased after treatment with trastuzumab. Similarly, patients with an existing favorable TME composition may be primed to experience the greatest benefit when anti–PD-1 is added to a monoclonal antibody such as trastuzumab (52). We demonstrated that in patients with HER2⁺ gastric cancer, treatment with a single dose of trastuzumab with 5-FU/platinum chemotherapy (prior to pembrolizumab) induces features of ADCC to a greater extent than in HER2⁻ patients. Similarly, immunogenic cell death and cGAS–STING signatures were significantly induced in HER2⁺ patients treated with a single cycle of trastuzumab with 5-FU/platinum chemotherapy and could plausibly lead to secondary IFN-γ–mediated PD-L1 upregulation. Notably, after one dose of trastuzumab + 5-FU/platinum, FCGR3A and FCGR3B expression in macrophages increased only in HER2⁺ patients, suggesting that trastuzumab induces features of ADCC. Owing to the lack of functional data, we cannot rule out our alternative explanations, including whether other factors such as chemotherapy and/or secondary effects of HER2 signaling inhibition with trastuzumab (we observed minimal changes in MAPK) are responsible for the observed NK cell recruitment, and functional studies are beyond the scope of this clinical trial and correlatives.

Consistent with previous data, we observed a high rate of clinically defined PD-L1 positivity (13 of 16 patients, 81%), and we newly observed a linear relationship between the degree of PD-L1 expression and tumor volume decrease during therapy (5, 35, 36). Although the correlation between PD-L1 expression according to IHC and CD274 (PD-L1) mRNA expression is imperfect, we observed an increase in CD274 expression in immune and tumor cells after adding pembrolizumab to trastuzumab and 5-FU/platinum in patients who were PD-L1⁻ at baseline. This suggests an induction of PD-L1 expression following trastuzumab-containing therapy, which is supported by the orthogonal observation of changes in PD-L1 CPS scores after treatment in some small series and preclinical data observing PD-L1 upregulation after trastuzumab (8, 10, 53). As T cells are the primary ultimate effectors of the antitumor response to immune checkpoint inhibition, we interrogated the relationship between CD8 TCR signaling signatures and PD-L1 expression and observed higher baseline TCR signaling and greater induction of TCR signaling in PD-L1⁺/HER2⁺ patients. Using validated signatures to predict tumor-reactive T cells (44–46, 54–57), we confirmed an increase in tumor-reactive T-cell signatures in PD-L1⁺/HER2⁺ versus PD-L1^–/HER2⁺ gastric cancer based on both scRNA-seq and DSP data. Our sample size limited serial scoring for tumor-reactive T-cell modules, but in the samples available, an increase was observed after the addition of pembrolizumab in HER2⁺/PD-L1⁺ patients but not in HER2⁺/PD-L1⁻ patients. Although biologically plausible, these observations should be considered highly preliminary based on the small sample size with TCR data (n = 2 PD-L1⁻ patients and n = 4 PD-L1⁺ patients) and limited number of clinically defined PD-L1⁻ patients. Although we used multiple published methods, we fully acknowledge that predicting tumor-reactive T-cell populations remains imperfect and substantial validation is needed to confirm our findings.

We leveraged the preserved architectural information from spatial transcriptomic data to segment compartments with and without HER2 expression, and heterogeneous signals were observed in all samples, consistent with well-established intratumoral heterogeneity by traditional IHC (58, 59). The relationship between spatially resolved HER2 expression in our data (higher expression in responders) is consistent with the WES data showing greater benefit in HER2-amplified patients and likely reflects more homogeneous tumors and/or greater oncogenic reliance on HER2. After segmenting the HER2 expressing and non-expressing compartments, we observed a greater degree of TGF-β pathway expression in nonresponders than in responders at BL and after the addition of pembrolizumab. Although the patient sample size is small, the use of multiple spatially distinct regions per patient partly addresses heterogeneity and may increase the generalizability of our observation. However, this observation should be interpreted with caution given the small sample size, although mutations known to activate TGF-β have been seen at resistance to the same therapy in the phase II PANTHERA trial (35).

Existing data suggest that despite HER2⁺ biomarker enrichment, approximately 30% of patients are intrinsically resistant to trastuzumab-based combinations. All three PD-L1⁻ patients and two PD-L1⁺ patients in our trial were nonresponders. Consistent with our previous scRNA-seq data from HER2⁻ patients, patient S90 (PD-L1⁺, CPS = 5) had high expression of the M2-like macrophage marker CD163 in multiple ROIs in DSP, which may partly explain the lack of treatment response. M2-like TAMs interact closely with cancer cells and the TME by producing TGF-β and can drive epithelial to mesenchymal transition and limit T-cell activation (60–62). Although confirmation in larger series is required, this pilot observation suggests that even within the most favorable dual PD-L1⁺/HER2⁺ tumors, the composition of immune cells surrounding the tumor exhibits heterogeneity and informs the treatment response. Although nongenomic features were the focus of our analysis, our finding is consistent with data from pretreatment samples in similarly treated patients that revealed that co-existing ERBB2 amplification and a high neoantigen load were positively associated with survival (35, 63, 64). We added some granularity using scRNA-seq and DSP, which showed that responders have higher baseline ERBB2 expression in tumor cells than nonresponders. In nonresponders, HER2 expression remained elevated whereas in responders, ERBB2 levels decreased during treatment with pembrolizumab + 5-FU/platinum/trastuzumab. Loss of HER2 expression (or elimination of HER2⁺ clones) after first-line treatment with trastuzumab is reported in approximately 30% of patients (39, 65), which is consistent with our findings. Similarly, the rates of PD-L1 discordance over time are consistent with findings in small series of gastric/GEJ cancer samples separated in time (66, 67). The development of formal heterogeneity metrics may be an important future direction for patient stratification.

Our trial and correlative analyses are limited by the small sample size, particularly in the nonresponder group. The correlative analyses should be considered exploratory, and we recognize we are underpowered for several comparisons. We partly address this by integrating a large HER2⁻ dataset from the HER2⁻ cohort of the same trial, which enabled additional comparisons and validation. Similarly, we attempted, via endoscopic mapping, to biopsy the same region of the primary tumor over time but recognize that a single biopsy may not be representative of global tumor biology, especially with respect to our TCR observations. However, in clinical practice, it is rarely feasible to undertake multiple biopsies in a serial manner and feel that our observations are biologically plausible. We also recognize the requirement for functional validation of our observations; however, appropriate immunocompetent models are limited.

In conclusion, our clinical trial results confirmed the activity of 5-FU/platinum/trastuzumab plus pembrolizumab in patients with HER2⁺ gastric cancer and support a model in which PD-L1⁺ tumors exhibit heightened ADCC and enhanced M1 macrophage recruitment during therapy. Conversely, the ADCC response to trastuzumab is less in tumors lacking PD-L1 expression, providing a plausible component to explaining the absence of PFS and OS benefits in PD-L1^– patients in the KEYNOTE-811 trial (5).

Supplementary Material

Supplementary Figure S1

Progression free and overall survival in a frontline phase II sequential chemoimmunotherapy trial in HER2+ GC.

Supplementary Figure S2

Graphical sample collection and analysis framework in HER2+ GC.

Supplementary Figure S3

TME composition and evolutions in HER2+ and HER2- gastric cancer.

Supplementary Figure S4

Relationship between macrophages phagocytosis and PD-L1 expression in HER2+ GC.

Supplementary Figure S5

T-cell subsets and tumor-reactive T-cells in advanced HER2+ GC

Supplementary Figure S6

Spatial analysis of HER2+ and HER2- tumor compartments within HER2+ gastric cancer.

Supplementary Figure S7

HER2 and TGF-beta signaling associations with outcomes in HER2+ gastric cancer.

Supplementary Table S1

Adverse Event Data

Supplementary Table S2

HER2 and PD-L1 changes over time

Supplementary Table S3

Differentially expressed genes in HER2+ vs HER2- clusters from digital spatial profiling.

Supplementary Table S4

HER2 (ERBB2) immune (CD45+) and tumor (EPCAM+) quantification from digital spatial profiling.

Supplementary Table S5

Representativeness of study participants in a phase II frontline gastric cancer trial.

Trial Protocol S1

Clinical Trial Protocol

Authors’ Disclosures

A. Mehta reports personal fees from Third Rock Ventures and The Column Group outside the submitted work. S.J. Klempner reports personal fees from Gilead, AstraZeneca, Astellas, Amgen, Elevation Oncology, Taiho Oncology, Boehringer Ingelheim, I-Mab, Daiichi Sankyo, Merck, Eisai, Bristol Myers Squibb, Novartis, BeiGene, and Mersana outside the submitted work, as well as being an uncompensated member of the NCCN gastric and esophageal guidelines committee and of the medical advisory board for Debbie’s Dream Foundation and for Hope for Stomach Cancer. No disclosures were reported by the other authors.

Authors’ Contributions

S.H. Lim: Conceptualization, formal analysis, investigation, writing–original draft, writing–review and editing. M. An: Formal analysis, validation, methodology, writing–original draft, writing–review and editing. H. Lee: Conceptualization, formal analysis, investigation, writing–original draft, writing–review and editing. Y.J. Heo: Data curation, formal analysis, investigation, methodology, writing–original draft, writing–review and editing. B.-H. Min: Resources, investigation, writing–review and editing. A. Mehta: Formal analysis, methodology, writing–original draft, writing–review and editing. S. Wright: Data curation, formal analysis, visualization, writing–review and editing. K. Kim: Formal analysis, validation, investigation, methodology, writing–review and editing. S.T. Kim: Conceptualization, resources, formal analysis, investigation, writing–original draft, writing–review and editing. S.J. Klempner: Conceptualization, formal analysis, supervision, writing–original draft, writing–review and editing. J. Lee: Conceptualization, resources, formal analysis, supervision, investigation, writing–original draft, writing–review and editing.