PD-L1 phenotype classification based on expression in tumor and immune cells as a potential biomarker for optimizing anti-PD-1/CTLA-4 immunotherapies in NSCLC.

Background
Several immune checkpoint molecules have recently been identified as promising targets for cancer immunotherapy. Programmed cell death-ligand 1 (PD-L1), one such molecule, is expressed not only on tumor cells (TCs) but also on immune cells (ICs), including antigen-presenting cells (APCs).1 PD-L1 expression is primarily induced by inflammatory cytokines in the tumor immune microenvironment (TIME) and suppresses T-cell activity through interactions with programmed cell death protein 1 (PD-1).2 3 Pembrolizumab, an anti-PD-1 antibody, has been approved for use as first-line immunotherapy for advanced non-small cell lung cancer (NSCLC) with high PD-L1 expression (Tumor Proportion Score (TPS) ≥50%) based on previous clinical trials.4 Furthermore, pembrolizumab-based chemoimmunotherapies (Pembro) have demonstrated superior therapeutic efficacy over chemotherapy alone for advanced NSCLC, regardless of PD-L1 expression on TCs.5 6 As a result, these regimens have become the standard of care for patients with advanced NSCLC.
Cytotoxic T-lymphocyte-associated antigen 4 (CTLA-4), another immune checkpoint molecule, is expressed on activated T cells. CTLA-4 competes with CD28 for binding to B7-1/2 on APCs, thereby inhibiting T-cell activation.7 8 Ipilimumab, an anti-CTLA-4 antibody, was first approved for advanced melanoma based on outcomes from a phase 3 trial.9 In advanced NSCLC, nivolumab plus ipilimumab with or without two cycles of platinum-doublet chemotherapy (Nivo+Ipi) demonstrated superior efficacy compared with platinum-doublet chemotherapy alone, regardless of PD-L1 expression on TCs.1013 Notably, nivolumab plus ipilimumab yielded improved clinical outcomes compared with nivolumab monotherapy in PD-L1-negative patients, although the study was not powered for this subgroup analysis.10 12 Recent reports have described that KEAP1 and/or STK11 mutations are associated with resistance to PD-1/PD-L1 blockade through a suppressed TIME,14 15 but with additional benefit from anti-CTLA-4 treatment on PD-1/PD-L1 blockade.16 17 Nevertheless, practical biomarkers to identify patients who would benefit from these combination immunotherapies rather than PD-1/PD-L1 blockade alone remain undefined.
This multicenter retrospective cohort study aimed to compare the efficacy of first-line Pembro and Nivo+Ipi, according to PD-L1 expression status on TCs and ICs. In addition, multiomics analyses, including PD-L1 expression status, whole-exome sequencing (WES), and RNA sequencing (RNA-seq), were conducted on tumor samples from treatment-naïve patients with NSCLC to elucidate the relationships among PD-L1 phenotypes, genetic alterations, and gene expression profiles.

Methods

Patients
This study retrospectively reviewed patients with advanced NSCLCs treated at a participating institute between January 2019 and December 2023 with Pembro or Nivo+Ipi, as initial systemic treatment. Inclusion required evaluable PD-L1 expression status on TCs (TPS by the 22C3 assay) and ICs (IC score by the SP142 assay) (figure 1). Patients were excluded if they lacked (1) driver oncogene status or (2) evaluable lesions. The following data were collected for eligible patients: age at treatment initiation, sex, smoking status, Eastern Cooperative Oncology Group performance status (ECOG-PS), histological diagnosis, driver mutations, PD-L1 expression status evaluated as PD-L1 TPS and IC score, location of metastasis, tumor response, progression-free survival (PFS), immune-related adverse event (irAE) incidence, and irAE locations. Treatment efficacy was evaluated according to the Response Evaluation Criteria in Solid Tumors V.1.1.18 Driver-oncogenes were detected using the Oncomine Dx Target Test, Amoy Dx, Lung Cancer Compact Panel Test, or companion diagnostics for each driver mutation, such as real-time PCR and immunohistochemistry (IHC), in clinical settings.
In addition, WES and RNA-seq analyses were conducted in treatment-naïve NSCLCs with sufficient frozen tissue samples, available TPS/IC scores, and informed consent, constituting the sequencing cohort (figure 1). Furthermore, we conducted multiplex IHC (mIHC) staining in available samples from the sequencing cohort.

Immunohistochemical staining and scoring systems
IHC staining for PD-L1 was automatically performed using two different platforms: the DAKO Autostainer Link 48 platform for the 22C3 assay to evaluate TPS and the VENTANA Bench Mark ULTRA IHC/ISH system for the SP142 assay to assess TC and IC scores. These were evaluated using formalin-fixed, paraffin-embedded specimens containing more than 100 viable TCs. Membranous staining of TCs was considered positive. TPS was calculated as the percentage of TCs with positive PD-L1 staining, with no expression defined as TPS <1%, low expression as TPS 1–<50%, and high expression as TPS ≥50%.4 For analysis, TPS <50% was defined as “low TPS” and TPS ≥50% as “high TPS.” The IC score was calculated as the area occupied by positive ICs per whole tumor area, including tumor stroma. IC scores were defined as follows: 0 (0–<1%), 1 (1–<5%), 2 (5–<10%), and 3 (≥10%).19 IC score 0 was categorized as “low IC score,” while IC scores ≥1 were categorized as “high IC score.”

DNA extraction and WES analysis
WES was conducted using 200 ng of genomic DNA isolated from snap-frozen cancerous and non-cancerous tissues obtained from patients with NSCLCs. Exome capture was performed using the Twist Comprehensive Exome Panel or Twist Exome V.2.0 plus Comprehensive Exome Spike-in (Twist Bioscience HQ, California, USA), according to the manufacturer’s instructions. Sequencing was performed on the NovaSeq6000 platform (Illumina, San Diego, California, USA) using 2×150 bp paired-end reads, resulting in approximately 100-fold genome coverage. Basic alignment and sequence quality control were conducted according to the Genomic Analysis Toolkit (GATK)4 best practices pipeline.20 The resulting reads from WES were aligned to hg38, using Parabricks V.3.1.3 (NVIDIA), which delivers the high-speed analysis recommended by the GATK, with graphics processing unit (GPU) acceleration for WES samples.21 Somatic single-nucleotide variants (SNVs) were called using Mutect2 (GATK V.4.1.2.0),22 and small insertions/deletions (IDs) were called using Mutect2 and Strelka2 (GitHub, San Francisco, California, USA).23

Classification of oncogenic/pathogenic mutations and analyses
Somatic mutations were selected using the following criteria: (1) a variant allele frequency of somatic mutations of >1% in tumor tissues, (2) removal of single-nucleotide polymorphisms if they had a threshold allele frequency of ≥0.01 in either the NHLBI GO Exome Sequencing Project (ESP6500) (http://evs.gs.washington.edu/EVS/) or in the Integrative Japanese Genome Variation Database (iJGVD, 20181105) (https://ijgvd.megabank.tohoku.ac.jp/), and (3) mutations were registered as “pathogenic/likely pathogenic variants” in ClinVar or as “oncogenic/likely oncogenic variants” in the OncoKB dataset (http://oncokb.org). All selected variants were manually checked using Integrative Genomics Viewer V.2.94.24 25 Enriched gene detections were conducted using the Oncodrive algorithm for single-cohort analysis and the mafcompare function, both of which are included in the maftools R package.26 Assignment of each mutation to COSMIC signatures and estimation of these proportions were conducted using the palimpsest R package.27

HLA typing and neoantigen prediction
HLA typing was performed with sequence reads from WES using the HLA-HD (V.1.4.0) algorithm.28 Classical alleles of class I (HLA-A, HLA-B, and HLA-C) and class II HLA genes (HLA-DRB1—5, HLA-DQA1/DQB1, HLA-DPA1/DPB1) were estimated. Neoantigens were predicted using the pVAC-Seq (V.4.0.10) pipeline.29 The NetMHCpan and NetMHCIIpan algorithms were used to estimate binding affinity.30 As recommended, variants were annotated for wild-type and mutant peptide sequences using the variant effect predictor (V.86) from Ensembl.31 Epitopes with a binding affinity inhibitory concentration (IC50) of ≤500 nM were considered potential neoantigens that bind to HLA alleles, and those with a low expression level (measured as transcripts per million (TPM) <1) were excluded. Immunogenic neoantigens were defined as those meeting all the following thresholds: (1) IC50 MT <500, (2) IC50 MT/WT <1, and (3) Tier=pass or subclonal.

RNA-seq and data analysis
Total RNA was extracted from frozen tumor samples using the TRIzol reagent (Invitrogen, Waltham, Massachusetts, USA) or the AllPrep DNA/RNA Mini Kit (Qiagen, Venlo, Netherlands). It was purified using the RNeasy Kit (Qiagen), and aliquots (10 ng) of the purified material were used to prepare RNA-seq libraries with the SMART-Seq Stranded Kit (Takara Bio, Japan). The resultant libraries were subjected to paired-end sequencing of 150 bp reads using the NovaSeq6000 or X Plus systems (Illumina). Raw reads were aligned to the reference human genome (UCSC-Build38). Read mapping was performed using STAR V.2.4.2 a with the human genome (GRCh38) and transcriptome data (GENCODE V.41) as reference datasets.32 33 RSEM was used to estimate the expression levels of individual transcripts by assembling the transcripts using BAM files created via STAR.34

Differential gene expression analysis and gene set enrichment analysis
Significantly enriched gene expression between the groups was detected by differential gene expression analysis using the Deseq2 R package (V.1.10.1).35 In differential gene expression analysis, significant results were identified as those with a false discovery ratio <0.25. Gene set enrichment analyses (GSEA) of HALLMARK gene sets and the Kyoto Encyclopedia of Genes and Genomes (KEGG) database were performed using the clusterProfiler R package.36 In GSEA, significance was defined as an adjusted p value <0.05.

Signature, deconvolution, and TCR repertoire analyses
The T-cell-inflamed gene expression profile score was calculated using TPM as previously described.37 IC deconvolution analysis for estimating the IC fraction in these tumors was performed using CIBERSORTx (https://cibersortx.stanford.edu/), running in absolute mode using the LM22 signature over 1,000 permutations.38 The T cell receptor (TCR) repertoire was calculated from RNA-seq data using MiXCR.39

Multiplex immunohistochemistry staining
For mIHC staining, 4 μm-thick sections were prepared. Multiplexed fluorescent IHC was performed using the Opal 7-Color Manual IHC Kit (NEL811001KT and NEL861001KT, Akoya Biosciences, Marlborough, Massachusetts, USA) after deparaffinization and heat-induced epitope retrieval, according to the manufacturer’s instructions. Briefly, the tissue sections were stained with the following anti-human monoclonal antibodies: anti-CD8 mAb (M7103, Agilent, Santa Clara, California, USA), anti-FOXP3 mAb (ab96048, Abcam), anti-cytokeratin mAb (ab80826, Abcam), anti-PD-1 mAb (ab137132, Abcam), anti-CD11c mAb (ab52632, Abcam), and anti-CD86 mAb (91882, Cell Signaling Technology, Danvers, Massachusetts, USA). After each primary staining, we applied the Opal Polymer Horseradish Peroxidase and Opal Fluorophore Kits using the tyramide signal amplification method, which amplifies IHC detection by covalently depositing fluorescent molecules in close proximity to targeted antigens. After the secondary staining with Opal fluorophores, the slides were mounted with ProLong Diamond Antifade Mountant containing 4',6-diamidino-2-phenylindole (DAPI, P36966, Thermo Fisher Scientific, Waltham, Massachusetts, USA).

Slide scanning and image analysis for mIHC
Images of entire stained slides were captured using the PhenoImager Fusion (Akoya Biosciences, Marlborough, Massachusetts, USA). The median of five regions containing TCs from each slide was captured using inForm V.2.8 (Akoya Biosciences). The captured images were analyzed using HALO V.3.6 multiplex image analysis software (Indica Labs, Corrales, New Mexico, USA). All slides were manually annotated to identify parenchymal regions containing tumor areas, and regions with insufficient tumor area were excluded from the following analyses (cut-off: tumor area <10% or <0.1 mm2). The intensity thresholds for the stained markers were adjusted visually, and the average densities of the stained cells (cells/mm2) were calculated automatically. The density of tumor-infiltrating cells was obtained from the mIHC data. Nearest-neighbor analysis was performed for all slides with sufficient tumor area, and the average distance (μm) between the stained cells was calculated in each case. Two researchers (JM and AF) evaluated the stained slides independently.

Statistical analysis
Categorical variables were analyzed using Fisher’s exact test. Continuous variables were analyzed using the Mann-Whitney U or Kruskal-Wallis test. Correlations between two variables were assessed using Spearman’s rank correlation coefficient. Survival curves were estimated using the Kaplan-Meier method and compared using the log-rank test. Univariate Cox proportional hazards regression analysis was performed for PFS. When the proportional hazards assumption was not met for any survival curve (Schoenfeld individual test: p<0.05), restricted mean survival time (RMST) ratios were additionally estimated using the survRM2 R package.40 Statistical significance was set at p<0.05 or adjusted p<0.05. Multiple testing correction was performed using the Benjamini-Hochberg method for GSEA and Oncodrive analyses, while the Bonferroni method was applied for other analyses. All statistical analyses were conducted using R software (R Foundation for Statistical Computing, Vienna, Austria).

Results

Patient characteristics
A total of 198 patients were included: 137 in the Pembro cohort and 61 in the Nivo+Ipi cohort (figure 1). The patient characteristics of these cohorts are summarized in table 1. No significant differences were observed between the cohorts in histology (squamous cell carcinoma (SQC) vs non-SQC), IC score, liver-metastasis frequency, or ECOG-PS. However, the Nivo+Ipi cohort had significantly lower TPS and a higher frequency of brain metastasis (BM) than the Pembro cohort (median TPS: 0% vs 10%, p<0.001; brain-metastasis rate: 26% vs 10%, p=0.005, table 1). In the whole of both cohorts, TPS was higher in tumors with higher IC score (median TPS: 0% in IC score 0, 15% in IC score 1, 20% in IC score 2, 80% in IC score 3, p<0.001 (Kruskal-Wallis test), (online supplemental figure S1a). However, 29% of patients with low TPS (<50%) had a high IC score (≥1), and 26% of those with high TPS (≥50%) had a low IC score (0, online supplemental figure S1b).

Differential activity between pembrolizumab-based chemoimmunotherapies and nivolumab plus ipilimumab-based immunotherapies according to PD-L1 expression on tumor and immune cells
In the efficacy analyses, the median PFS (mPFS) was 7.3 and 4.0 months in the Pembro and Nivo+Ipi cohorts, respectively. In the Pembro cohort, patients with high TPS had significantly longer PFS than those with low TPS (mPFS: 8.1 vs 7.1 months, p=0.02, HR 0.59 (95% CI 0.38 to 0.92), figure 2a); however, no significant difference in PFS was observed according to IC score (low IC score vs high IC score, mPFS: 6.4 vs 7.8 months, p=0.11, HR 0.72 (95% CI 0.49 to 1.07), figure 2b). In the Nivo+Ipi cohort, there was no significant difference in PFS according to TPS (low TPS vs high TPS, mPFS: 4.0 vs 4.0 months, p=0.26, HR 1.95 (95% CI 0.60 to 6.4), figure 2c), although the small number of high TPS patients (n=3) precludes firm conclusions. In contrast, patients with high IC scores had significantly longer PFS than those with low IC scores (mPFS: 7.7 vs 2.8 months, p=0.04, HR 0.53 (95% CI 0.28 to 0.96), figure 2d). Moreover, multivariable Cox regression analysis for PFS in the Nivo+Ipi cohort showed no statistically significant covariates, including BM, histology, and ECOG-PS (BM status: HR 0.96 (95% CI 0.45 to 2.0), p=0.91; ECOG-PS: HR 2.9 (95% CI 0.59 to 14.4), p=0.19, (online supplemental table S1).
Next, clinical outcomes were compared between the two groups stratified by IC score among patients with low TPS. Among patients with low TPS/high IC scores, the mPFS was 6.6 months in the Pembro cohort and 12.4 months in the Nivo+Ipi cohort (HR 0.56 (95% CI 0.28 to 1.13), p=0.10). Among patients with low TPS/low IC scores, the mPFS was 7.1 months in the Pembro cohort and 2.7 months in the Nivo+Ipi cohort (HR 1.24 (95% CI 0.79 to 1.99), p=0.35). Since the proportional hazards assumption was not met between the survival curves of the two cohorts (Schoenfeld individual test: p<0.05), RMST estimation was applied for superior detection power to assess delayed therapeutic effects. Among patients with low TPS/high IC scores, Nivo+Ipi achieved significantly greater RMST than Pembro in the later phase (RMSTNivo+Ipi/RMSTPembro (24 months)=1.5 (95% CI 1.00 to 2.32), p=0.049, figure 3a), while no significant difference was observed between treatments in those with low TPS/low IC scores at any phase (Nivo+Ipi vs Pembro; RMSTNivo+Ipi/RMSTPembro (24 months)=0.82 (95% CI 0.54 to 1.26), p=0.37, figure 3b). These results suggest that patients with low TPS/high IC scores may be better suited for Nivo+Ipi than for Pembro, given that Nivo+Ipi provides a more durable response in this subgroup.

Genomic profiles related to each PD-L1 expression status
To investigate the relationship between genomic profiles and TPS/IC score, treatment-naïve NSCLC samples with available TPS and IC scores that had undergone WES and RNA-seq were reviewed. A total of 152 samples were included: 139 with both WES and RNA-seq data and 13 with RNA-seq data only (‘Sequencing cohort’ in figure 1). The tumor characteristics of this cohort are summarized in online supplemental table S2. The onco-plot of these 139 tumors including all mutations and OncoKB-defined cancer gene are shown in figure 4, online supplemental figure S2a. Many frequently identified mutations were consistent with previous studies except for the EGFR mutation, which may reflect differences in ethnicity. No SNVs/INDELs were significantly associated with TPS or IC scores, and TPS showed no significant correlation with tumor mutation burden (TMB) (R=−0.14, p=0.09; online supplemental figure S2b). However, in tumors with low TPS, TMB was higher in tumors with higher IC score (median TMB (mTMB, mut/Mb): 1.7 in IC score 0, 3.3 in IC score 1, 18.1 in IC score 2, and 25.0 in IC score 3, p<0.001 (Kruskal-Wallis test); left in figure 4b), in contrast to those with high TPS (mTMB: 1.6 in IC score 0, 2.4 in IC score 1, 1.8 in IC score 2, 1.8 in IC score 3, p=0.81; right in figure 4b).
According to these findings, we further assessed TMB based on PD-L1 expression patterns classified by TPS and IC score (ie, PD-L1 phenotypes): Phenotype I (P-I)=low TPS/low IC score, Phenotype II (P-II)=low TPS/high IC score, and Phenotype III (P-III)=high TPS/any IC score (figure 4c). Representative IHC images of each phenotype are shown in figure 4d. Based on this classification, P-II tumors exhibited a significantly higher TMB than the other phenotypes (mTMB: 1.7 /mb in P-I tumors, 18.2 /mb in P-II tumors, and 1.9 /mb in P-III tumors, p<0.001 (Kruskal-Wallis test); figure 4e). To validate actual tumor immunogenicity, we estimated immunogenic neoantigen counts using HLA typing from WES and expression data from RNA-seq (neoantigen prediction algorithm shown in online supplemental figure S2c). P-II tumors exhibited the greatest diversity of immunogenic neoantigens for both HLA class I and II (figure 4f). To explore the underlying causes of increased TMB in P-II tumors, we estimated proportions of each COSMIC signature assignment based on the PD-L1 phenotypes. High TMB in P-II tumors consisted of clock-like signatures and not homologous recombination repair deficiency (HRD) or mismatch repair deficiency (MMR-D) signatures (online supplemental figure S3a). In fact, HRD/MMR-D-related mutations were not enriched in P-II tumors compared with the other tumors (online supplemental figure S3b,c). Furthermore, there was no enrichment of KEAP1 and STK11 mutations in P-II tumors, which are associated with additional benefit of anti-CTLA-4 antibody on PD-1/PD-L1 blockade (online supplemental figure S3c).
These findings suggest a distinct tumor subgroup among tumors with low TPS, characterized by high TMB and tumor immunogenicity without genomic instability and these tumors could be identified by PD-L1 expression on ICs.

Tumor immunological microenvironment related to each PD-L1 expression status
We investigated how the TIME in P-II tumors differs from that in P-I and P-III tumors using expressional data from 152 samples. GSEA revealed that among tumors with low TPS, P-II tumors exhibited significant activation of various pathways associated with antitumor immune responses, including antigen presentation, TCR signaling, and T cell differentiation, compared with P-I tumors (HALLMARK: figure 5a–i, KEGG pathways: online supplemental figure S4–i). In contrast, when comparing P-II and P-III tumors, no significant differences were found in the enrichment of pathways involved in PD-1/PD-L1 interactions and immune response (figures 5a–iii, and 4–iii).
In the deconvolution analysis, P-II and P-III tumors had significantly higher fractions of activated CD4+ T cells and M1 macrophages compared with P-I tumors. P-II tumors had equivalent CD8+ T cell fractions to P-I tumors (figure 5b). Additionally, P-II tumors had the highest regulatory T cell (Treg) fraction compared with both P-I and P-III tumors. In the TCR repertoire analysis, P-II tumors had significantly greater TCR diversity than P-I tumors, while P-III tumors showed no significant increase (figure 5c). In the signature analysis, P-II tumors exhibited an activated TIME compared with P-I tumors, similar to P-III tumors (figure 5d). Interferon-gamma expression showed the same tendency as TIME (figure 5e). Nevertheless, the expression of CD274, which encodes PD-L1, was significantly lower in P-II tumors than in P-III tumors (figure 5f). In summary, P-II tumors exhibited an activated TIME compared with P-I tumors among those with low TPS, similar to P-III tumors. On the other hand, P-II tumors were characterized by significant Treg enrichment with suppressed CD274 expression compared with P-III tumors.
Finally, to validate these distinct TIMEs observed in P-II tumors, we performed mIHC staining in available 15 samples in the sequencing cohort (5 samples in each PD-L1 phenotype, online supplemental table S3). The representative mIHC pictures are shown in online supplemental figure S5a). In the density analysis, P-II tumors had significantly higher CD11c+CD86+ APC density (median APC density (/mm2): P-I/P-II/P-III; 7.4/281.1/11.3, p=0.01 (Kruskal-Wallis test), online supplemental figure S5b), and a trend toward higher Treg density than P-I and P-III tumors (median Treg density (/mm2): 10.7/62.3/31.2, p=0.09). In addition, PD1+FOXP3+ Treg density was significantly higher in P-II tumors than P-I and P-III tumors (median PD-1+FOXP3+ Treg density (/mm2): 2.5/31.2/2.4; p=0.04). In the nearest-neighbor analysis, distance between Treg and CD8+ T cells was numerically shorter in P-II than P-I and P-III, despite no significant difference (median Treg–CD8+ T cells distance (μm): 171.3/83.5/179.0; p=0.28, (online supplemental figure S5c). These results were consistent with the results observed in efficacy and sequence analysis.

Discussion
This study demonstrated that PD-L1 expression on ICs (IC score by the SP142 assay) served as a predictive biomarker for the efficacy of Nivo+Ipi treatment, whereas expression on TCs (TPS by the 22C3 assay) served as a predictive biomarker for Pembro treatment efficacy. Tumors with low TPS/high IC scores showed a more durable response to Nivo+Ipi than to Pembro. Additionally, multiomics analyses revealed that tumors with low TPS/high IC scores had an activated TIME comparable to that of those with high TPS, but were characterized by significantly more diverse immunogenic neoantigens and a higher Treg fraction. This suggests that additional anti-CTLA-4 therapy to PD-1 inhibition could be required to induce a durable immune response for these tumors.
PD-L1 TPS has been established as a predictive biomarker for anti-PD-1/PD-L1 immunotherapies, such as pembrolizumab, nivolumab, and atezolizumab, in patients with advanced NSCLC.41 In contrast, anti-PD-1/CTLA-4 combination immunotherapies, including nivolumab plus ipilimumab, have shown durable responses regardless of TPS.1013 However, adding ipilimumab to pembrolizumab did not improve efficacy compared with pembrolizumab monotherapy in patients with treatment-naïve metastatic NSCLC with high TPS.42 Therefore, determining the patient subgroups that may benefit from the addition of anti-CTLA-4 antibodies to PD-1/PD-L1 blockade based on TPS status alone remains a clinical challenge.
On the other hand, the status of PD-L1 expression on ICs (IC score) in addition to PD-L1 expression on TCs (TC score) has been validated as part of the predictive biomarker profile for atezolizumab-based immunotherapies in patients with advanced NSCLC.19 43 A recent study showed that tumors with high PD-L1 expression on ICs and low expression on TCs possess an immune-rich microenvironment, similar to those with high PD-L1 expression on both TCs and ICs.44 45 This suggests that even among tumors with low PD-L1 expression on TCs, the tumor immune landscape may vary depending on PD-L1 expression on ICs. In fact, tumors with low TPS/high IC scores exhibited a more activated TIME compared with those with low TPS/low IC scores, resembling the TIME of high TPS tumors.
Multiomics analysis indicated that tumors with low TPS/high IC scores were uniquely characterized by a high TMB, Treg enrichment, and suppressed CD274 expression compared with high TPS tumors. In general, tumor-associated macrophages (TAM), which constitute a major component of PD-L1-expressing ICs,1 intrinsically express PD-L1 during antigen presentation.46 47 This suggests that the IC score in low TPS tumors could reflect increased TAM activity involved in presenting a diverse neoantigen repertoire, which is supported by distinct APC enrichment observed in mIHC. Moreover, diverse neoantigens promote Treg differentiation from CD4+ T cells via MHC class II-mediated presentation.48 This supports the hypothesis that tumors with high TMB may rely on Treg-mediated immune suppression, not only on the PD-1/PD-L1 axis, against activated anticancer immunity. In this context, considering several reports that ipilimumab exerts anticancer effects via Treg depletion,49 50 ipilimumab addition to the PD-1/PD-L1 blockade may be a reasonable option for these tumors. The present study demonstrated that patients with tumors exhibiting low TPS/high IC scores exhibited the highest TMB among all PD-L1 phenotypes and derived significantly greater benefit from Nivo+Ipi than from Pembro. This finding is consistent with a subgroup analysis from the CheckMate 227 and 848 trials, which demonstrated that nivolumab plus ipilimumab provided superior clinical benefit compared with nivolumab alone in patients with TMB-high advanced NSCLC or other solid tumors.51 52
This study has several limitations. First, the sample size of the Nivo+Ipi cohort was relatively small, even though it was derived from a multicenter cohort study. Especially, a small number of patients with high TPS (n=3) in the Nivo+Ipi cohort precluded validation of the findings from the KEYNOTE-598 trial. Second, PD-L1 expression in ICs showed relatively poor concordance among each assay in the previous studies.53 54 In this study, the SP142 assay was adopted because this is the only approved assay evaluating PD-L1 expression in ICs. Thus, although there are further needs for validating IC score for the PD-L1 expression in ICs by the other assays, our findings are available in the clinical setting. Third, this study lacked experimental validation of immunosuppressive mechanisms mediated by Treg in P-II tumors. Although our multiomics analysis using sequencing and mIHC staining showed P-II tumors exhibit distinct TIME, how Treg suppressed anticancer immunity remains unclear. To address these points, further studies such as spatial transcriptome analysis which could visualize actual interactions between Treg and effector T cells are required. Despite these limitations, our findings demonstrate potential applicability of PD-L1 phenotype classification in identifying suitable candidates for Nivo+Ipi treatment over Pembro. Larger prospective studies are essential to confirm the clinical utility of PD-L1 phenotype classification in improving treatment outcomes for patients with NSCLC.

Conclusion
Nivo+Ipi induced a superior and more durable response than Pembro in patients with NSCLC who had low TPS/high IC scores. These tumors were characterized by high TMB and enriched Treg fractions compared with other tumors. The PD-L1 phenotype classification based on TPS and IC score could assist in identifying optimal candidates for addition of anti-CTLA-4 antibody to PD-1/PD-L1 blockade among patients with advanced NSCLC.

Supplementary material
10.1136/jitc-2025-012880online supplemental file 110.1136/jitc-2025-012880online supplemental file 210.1136/jitc-2025-012880online supplemental file 3