Introduction

Immune checkpoint inhibitors (ICIs) in combination with chemotherapy are now standard of care in the definitive treatment of resectable non-small cell lung cancer (NSCLC). We previously reported findings from our phase II study evaluating the combination of atezolizumab, carboplatin, and nab-paclitaxel in resectable NSCLC. Hypothesizing that the addition of atezolizumab would confer improvement in the major pathological response (MPR) rate from 22% historically to 44%, the primary endpoint was met with our observations of an MPR rate of 57% and pathological complete response (pCR) rate of 30% in the intention-to-treat population.1

This study, together with findings from the NADIM study,2 provided early evidence of the clinical benefit of neoadjuvant chemoimmunotherapy—findings that have since been confirmed in recent randomized phase III trials, including CheckMate 816,3 Keynote 6714, and the AEGEAN study.5 Results from these phase III studies, some of which incorporated ICIs into both the neoadjuvant and adjuvant settings, have led to improvements in event-free and most recently overall survival (OS).4 5 Nonetheless, questions remain regarding which patients are most likely to benefit from ICIs and the relative contributions of neoadjuvant versus perioperative ICI. Furthermore, identification of predictive biomarkers to guide strategies to refine our management remains an unmet clinical need.

To date, subgroup and exploratory analyses from the above studies have nominated smoking status, degree of pathological response,6 7 pretreatment tumor PD-L1 score, and post-treatment clearance of circulating tumor DNA8 as possible predictive biomarkers of those most likely to benefit from early introduction of ICIs, but robust, patient-level predictors and a detailed mechanistic understanding remain elusive.

Here, we provide an update on our phase II investigator-initiated study investigating the combination of neoadjuvant atezolizumab with platinum-doublet chemotherapy in NSCLC patients with stage IB-IIIA disease. We report clinical outcomes after 3 years of follow-up, correlated with major findings from whole exome sequencing (WES), RNA sequencing, and multiplexed immunofluorescent measurement of immune cell markers.

Materials and methodsPatients

Eligibility criteria for this study have been previously described.1 Briefly, eligible patients were prior or current smokers 18 years or older with American Joint Committee on Cancer (AJCC seventh edition)—defined stage IB-IIIA NSCLC that was deemed surgically resectable by a thoracic surgeon. All patients had baseline core tumor biopsy assessable for PD-L1 expression (22C3, Dako), Eastern Cooperative Oncology Group performance status 0 or 1, and RECIST (V.1.1) measurable disease.

Trial design and treatment

This study was an open-label, phase II trial investigating the combination oimmunotherapy with atezolizumab. Eligible patients received four cycles of intravenously infused atezolizumab 1200 mg on day 1, nab-paclitaxel 100 mg/m2 on days 1, 8, and 15, and carboplatin area under the curve 5 on day 1 of each 21-day cycle prior to undergoing surgical resection. Patients experiencing progression of disease (PD) on interval chest CT scan after the first two treatment cycles, as well as those patients unable to tolerate all four cycles of treatment, were able to proceed directly to surgery. Postoperative course entailed clinic visits at 2–4 weeks and 3 months after surgery; no further systemic therapy was planned. Imaging surveillance was performed at the 3-month visit and thereafter every 4–6 months as per standard of care.

Pathological response to treatment was assessed by local pathologists, who measured the percentage of viable residual tumor in resected primary tumors from each patient at the time of surgery. MPR was defined as the presence of 10% or less viable residual tumor; pCR was defined as the presence of 0% residual tumor. Translational analyses are reported only for cases where patients underwent complete resection; the four patients (indicated as “missing” in table 1) who did not were excluded from this analysis.

Table 1

Patient characteristics (N=30)

Outcomes

The primary endpoint was the proportion of patients achieving MPR, whereas secondary endpoints included objective response rate (ORR, RECIST V.1.1), disease-free survival (DFS), and OS. Safety was monitored throughout the study, with adverse events assessed at each visit according to the Common Terminology Criteria for Adverse Events (V.4.0). No patients experienced serious or persistent treatment-related or immune-related adverse events prohibiting or delaying surgery by more than 37 days beyond the preplanned date of surgery, so the early stopping rule of 3 or more such instances was not invoked, and the study accrued to completion of enrolment.

Statistical analysis

A Simon’s optimal two-stage design was used to assess MPR as the primary endpoint, hypothesizing that the addition of atezolizumab would increase the MPR rate from 22% (historical control) to 44% for 83% Power at a type I error of 0.05. Achievement of MPR by more than 4 of the first 18 patients enrolled permitted enrolment of an additional 12 patients for a total of 30 evaluable patients.

Data were summarized descriptively using frequency (percentage) for categorical variables, and using mean (SD), median (IQR) for continuous variables. Overall and DFS probabilities were estimated using Kaplan-Meier method and compared across various groups of interest, using Cox proportional hazards regression models. Comparisons of interest included MPR, pCR, radiologic response type, PD-L1 score with a cut-off of 1, and PD-L1 score with a cut-off of 50. All analyses were completed in R V.4.1.2, with a type I error set at 0.05.

Whole exome sequencing

Genomic DNA (gDNA) was isolated from formalin-fixed, paraffin-embedded (FFPE) tissue slides, macrodissected for tumor-specific areas, using the QIAmp DNA FFPE Kit (Cat# 56404). The FFPE was deparaffinized with xylene and lysed under denaturing conditions with proteinase

gDNA was also isolated from peripheral blood mononuclear cells (PBMCs) and blood using the QIAamp DNA Mini Kit (cat# 51104). The blood and PBMC samples were treated with Proteinase K, RNAse and lysed. The purified gDNA samples were assessed for quality and quantified using fluorescence (Qubit) and stored at −20°C. gDNA (100 ng) was fragmented to ~150 bp using a Covaris instrument. Libraries were prepared using the Swift Accel-next generation sequencing (NGS) 2S Hyb DNA library kit (cat# 23096, Swift) along with the Swift 2S SureSelect XT Compatibility Module (#27496, Swift).

After DNA was dephosphorylated and ligated, the DNA fragments were amplified by PCR. Total of 500 ng of enriched library was used in the hybridization and captured with the SureSelect All Exon V.6 (Agilent Technologies, 5190-8865) bait. Following hybridization, the captured libraries were purified and amplified by PCR. Library fragment size and quality were assessed using the Tapestation and quantified by qPCR. Normalized libraries are pooled and sequenced on a Novaseq 6000 using 2×100 bp paired-end sequencing.

Somatic variant calling was performed following GATK Best Practices. Fastq file quality checks were done by FastQC (V.0.11.9). Reads were aligned to the human reference genome (Genome Reference Consortium Human Build 38 (GRCh38)) using bwamem V.0.7.15-r1140 (parameters: ‘-K 100000000 -p -v 3 -t 16’). Duplicate reads in the resulting BAM file were marked using PicardTools MarkDuplicates (V.2.18) and bases were recalibrated using the GATK ApplyBSQR function (V.4.1.4.1). Variations were called using the default options in Strelka (V.1.0.15), LoFreq (V.2.1.3.1), and Mutect2 from GATK 4.1.4.1 (parameters:′—downsampling-stride 20—max-reads—per-alignment-start 6—max-suspicious-reads-per-alignment-start 6′) in comparison of tumor sequence to a paired normal tissue sequence. Tumor/normal pair confirmation was provided by NGSCheckmate (V.1.0.0). The consequences of each mutation were determined using Ensembl Variant Effect Predictor (Release 99). Merging and intersecting VCF files from each variant caller was performed with bcftools (V.1.10) and variants from at least two of the three callers were retained. Merged VCF files were transformed to MAF files using vcf2maf (V.1.6.21).

Tumor purity was bioinformatically calculated with Titan software9 and samples with less than 15% were removed prior to analysis. Copy number variants at the gene, bin, and segment level were called using CNVkit V.0.9.9. The default parameters for all steps except the genemetrics call were used, where the parameters of ′genemetrics −t 0 m 1′ were used.10 11 Prior to analyses, gene-level CNA results were filtered to prevent low coverage artifacts (CNVkit log2 <−5, chrY alterations in females). CNVkit log2 cut-offs were used to call categorical amplification (log2>=1) and deletion (log2 <=−1) calls.

Multiplexed immunofluorescence

Roche Tissue Diagnostics (RTD, Ventana developed two multiplexed immunofluorescence (IF)) panels—panel 1: PD- L1(SP263)/CD3/CD8/CD68/FoxP3, panel 2: MHCI/B2M/MHCII/CD11c/panCK (see panel information in online supplemental table 1).

Paraffin embedded (FFPE) tissue slides for each specimen were stained and scanned for whole slide image analysis using the HALO (Indica Labs) platform scoring algorithms. HALO Image Analysis included artifact (folds, necrosis, red blood cell or eosinophils) removal, classification of tumor from stroma, cell detection (DAPI nuclei staining and positive staining of biomarkers), batch run outputs and generating readout data (cell counts, rations, density, average distances, number of cells within a distance, area ratios and intensity). A pathologist reviewed the matched H&E-stained tissue and annotated tumor and peritumor content (500 µm). Any specimen images that failed QC (eg, due to systemic issues) were reanalyzed using a refined scoring algorithm as necessary).

RNA sequencing

Total RNA, including miRNA, was isolated using Qiagen miRNeasy FFPE kit (Cat # 217504) from FFPE tissue slides macrodissected for tumor-specific areas. The FFPE was deparaffinized with xylene and lysed under denaturing conditions with proteinase K, heated to reverse formalin cross-linking and then treated with DNase to eliminate all genomic DNA. The purified RNA samples were assessed for quality and quantified using Qubit or the Agilent Bioanalyzer, and stored at −80°C. Sequencing Library was generated using the Illumina TruSeq RNA Access method (hybridization-based protocol) which involved the preparation of the total RNA library and coding RNA library enrichment.

Libraries were sequenced using the Illumina sequencing-by-synthesis platform (Illumina TruSeq RNA Access), with a sequencing protocol of 5 0 bp paired-end sequencing and total read depth of 40M reads per sample. Samples with the recommended input of 100 ng RNA were moved forward with library preparation and samples with %DV200<30 are not recommended for analysis by this method.

RNA-sequencing data were analyzed using HTSeqGenie12 in BioConductor13 as follows: first, reads with low nucleotide qualities (70% of bases with quality <23) or matches to rRNA and adapter sequences were removed. The remaining reads were aligned to the human reference genome (human: GRCh38.p10) using GSNAP (PMID:20147302, 27008021) version ‘2013–10- 10-v2’, allowing maximum of two mismatches per 75 base sequence (parameters: ‘-M 2 n 10—B 2—i 1 N 1 w 200000—E 1—pairmax-rna=200000—clip-overlap'). Transcript annotation was based on the Gencode genes database (human: GENCODE 27). To quantify gene expression levels, the number of reads mapping unambiguously to the exons of each gene was calculated. CPM was calculated by deriving the proportion of reads for each gene within each sample, and then scaling the resulting proportion by a factor of one million.

Batch effects due to sequencing date and library preparation kit were removed using the R package limma43. After discarding genes not present at ≥0.5 c.p.m. in ≥10% of samples (due to low abundance), 17,729 genes remained for analysis. xCell (V.1.3) was used to estimate immune cell abundance. To estimate STK11 and KEAP1 “deficiency” scores, a gene list was compiled based on previous studies.14–16 Normalized Z scores for each gene on the list were summed to calculate the scores.

ResultsPatients and treatment

Across 3 sites, a total of 30 evaluable patients were enrolled among 39 assessed for eligibility between May 26, 2016 and March 1, 2019. This included 1 replaced non-evaluable patient who, after beginning study treatment, was found to have primary colon cancer. Adenocarcinoma (ADC) was the dominant histology, accounting for 17 of 30 patients’ tumors; 12 (40%) had squamous cell carcinoma (SCC), and 1 (3%) had a large cell neuroendocrine tumor on final pathology (initial biopsy had revealed ADC).

Baseline clinical pathological characteristics have not changed since the original publication. Clinical outcomes of interest are summarized in table 1. Pretreatment PD-L1 tumor proportion score (TPS) levels measured by IHC were <1% in 11 patients (39.3% of 28 evaluable), 1%–50% in 9 patients (32.1%), and >50% in 8 patients (28.6%), with 2 patients having unevaluable PD-L1. Standard of care NGS was performed on a subset of ADCs and one large cell neuroendocrine carcinoma. In total, 18 patients had a pathological response assessment and tumors had adequate tumor content from either a pretreatment or post-treatment sample to comment on WES results. Biomarker manifest is available as online supplemental figure S1 and online supplemental table 2).

Censoring for the current analysis occurred on October 15, 2022. The median follow-up time at the date of censoring was 39.52 months (30.99–49.42 IQR). In the intention-to-treat population, 17/30 patients (57%; 95% CI 37% to 75%) experienced MPR and pCR was observed in 10/30 patients (33%; 95% CI 17% to 53%). Of note, four patients’ tumors were unable to be assessed for pathological response. Similar rates of pathological response were observed across the dominant lung cancer histologies. Of 15 evaluable ADCs, 8 demonstrated MPR and 5 demonstrated pCR; of 10 evaluable SCCs, MPR was observed in eight and pCR was observed in 5. Two patients experienced progressive disease (PD) per RECIST V.1.1. As of the date of censoring, 17 of 30 patients had either died or had disease recurrence. The median DFS was 34.5 months (95% CI 18.4 to not reached). Landmark DFS rates at 12 months, 24 months, 36 months were 80%, 60%, and 49.2%, respectively (figure 1A). The median OS was 55.8 months (95% CI 43.6 to not reached). Landmark OS rates at 12 months, 24 months, and 36 months were 96.7%, 80%, and 76.7%, respectively (figure 1B).

Figure 1

Disease-free and overall survival. (A) Disease-free survival probability. (B) Overall survival probability.

Clinical outcomes relative to pathological response are summarized in figure 2. MPR did not significantly associate with either DFS or OS by Cox proportional hazards regression. The 17 patients who achieved MPR experienced a median DFS of 50.5 months (95% CI 20.7 to not reached) and median OS of 58.0 months (95% CI 58 to not reached). The nine patients who did not achieve MPR experienced median DFS of 55 months (95% CI 14.1 to not reached) and OS of 55.8 months (95% CI 43.6 to not reached).

Figure 2

Swimmer’s plot indicating patient recurrence and survival status. Patients are grouped by pathological response status. Follow-up status is indicated by icons within each row. Patient and tumor characteristics are depicted by squares in each row.

pCR similarly did not significantly associate with either DFS (HR 0.45, 95% CI 0.12 to 1.65) or OS (HR 0.22, 95% CI 0.03 to 1.83), though there was a trend toward improved outcomes among patients who experienced pCR. Only 3 of the 10 patients who achieved pCR had experienced disease recurrence and one of these patients died; median DFS and OS have, therefore, not yet been reached.

Radiographically, 9 patients had stable disease, and 19 achieved a partial response (no patients had a complete response), consistent with an overall response rate of 63.3% and disease control rate of 93.3% in the intention-to-treat population (n=30). Two patients experienced PD, with one patient proceeding to surgery without subsequent recurrence and the other developing brain metastases and ultimately dying. Although radiographic response did not correlate significantly with survival endpoints, there was a numerical difference (mOS: SD 47 months (43.6 to not reached), PR 58 months (55.8 to not reached)); median DFS among patients who achieved PR was 50.5 months (20.71 to not reached) and 18 months (9.93 to not reached) for SD. Median OS among patients who achieved PR was 58 months (55.8 to not reached) and 47 months (43.6 to not reached) for those with SD For DFS, HR=0.58, 95% CI (0.21 to 1.64). Correlation between degree of pathological regression and radiographic response appeared to be limited due to small sample size (−0.4442654 (−0.70521927, –0.07730212) (online supplemental figure S2).

PD-L1 level did not significantly associate with either DFS or OS at TPS cut-off of 1% or 50% (online supplemental figure S3). Tumor mutational burden did not significantly associate with either pathological regression (p=0.3) or radiographic response (p=0.6 (online supplemental figure S4). All patients with either MPR or pCR and WES data available (10/17) had missense variants in TP53 detected on WES; 4 patients without MPR/pCR had TP53 alterations. Although we previously reported two of four patients with EGFR mutations experienced a pCR, one of these patients turned out to have two genetically distinct tumors—the tumor harboring the EGFR mutation did not meet a pCR or even MPR, but the other tumor, which harbored a mutation in STK11, did.

Patients with recurrent disease to the brain

Six patients on study were found to have recurrent disease to the brain, of whom five died during the follow-up period. Median DFS for patients with recurrent disease to the brain was 12.4 months (95% CI 9.93 to 32.5); median OS was 45.3 months (95% CI 22.7 to 55.8). Three of the six patients who had brain-recurrent disease had achieved MPR (two of these were pCR); none of the pathological or radiographic measures of response strongly associated with DFS or OS among patients who developed brain metastases (figure 3). Pretreatment tumor samples were available for multiplexed IF from all patients with brain-recurrent disease. PD- L1+cell density was higher in patients with CNS recurrence (online supplemental figure S5A).

Figure 3

Patient-level associations of molecular features with clinical outcomes for genes of interest. Variants are pooled from pre-tx and post-tx samples (N=21 patients). Per cent of patients with each alteration indicated to the right of the oncoprint, along with a count of alteration types (stacked bar graph).

Consistent with previous findings, there was a trend toward increased incidence of brain metastases in patients whose tumors had deleterious mutations in STK 11 (figure 4A) and KEAP1 (figure 4B); however, this did not reach statistical significance.17 Because STK11 and KEAP1 are both located on chromosome 19p and are commonly codeleted, we sought to determine whether copy number of these genes pretreatment is also associated with brain-recurrent disease as in previous studies.17 STK11 copy number was significantly lower pretreatment in patients with brain metastases (figure 4C); this trend was not observed with KEAP1 copy number (figure 4D).

Figure 4

Associations between copy number status of STK11 and KEAP1 with clinical outcome. (A, B) Brain metastasis status stratified by STK11 and KEAP1 mutation status. (C, D) Brain metastasis status by STK11 and KEAP1 copy number. (E, F) Histoplot comparing copy number of STK11 (E) and KEAP1 (F) in patients with pCR, MPR, and those without. MPR, major pathological response. * = p value < 0.05

Translational findings

Robust biomarkers for predicting benefit from PD-(L)1 inhibitors in NSCLC remain limited,18 so we next wanted to identify genetic alterations in pretreatment samples that could predict pathological response. STK11 and KEAP1 mutations have previously been found to be associated with poor response to immunotherapy.19 In our cohort, patients whose tumors had mutations in either STK11 (N=5) or KEAP1 (N=7) did not experience statistically worse pathological response (online supplemental figure S5B and S5C). STK11 activity has previously been found to be associated with poor response,14 which we also observe in our cohort (online supplemental figure S5D). KEAP1 activity has previously been found to be associated with poor response regardless of treatment option,20 but correlations were not observed in this dataset (online supplemental figure S5E).

We wondered whether copy number changes in these genes correlated with clinical outcomes beyond recurrent disease to the brain. Interestingly, patients who experienced MPR had statistically significantly higher relative copy number of STK11 compared with patients without MPR (p=0.01, figure 4E); for KEAP1 association between copy number and pathological response trended similarly but was not statistically significant (figure 4F). Copy number of STK11 or KEAP1 did not associate with DFS or OS.

We next analyzed immune infiltration from transcriptomic data, using the xCELL algorithm21 to estimate enrichment of different immune cell types. 20 pretreatment and 16 post-treatment samples had RNA sequencing data for these analyses. Pretreatment, high TH2 fractions correlated with better pathological response (figure 5A). Post-treatment, high macrophage (M2), eosinophil and basophil fractions correlated with worse pathological response; plasma cell fraction correlated with better response (figure 5A). Interestingly, some of these correlations differed depending on tumor type. In ADC, tumors with higher macrophage fractions pretreatment trended toward better response (figure 5B). In SCC, the opposite trend is observed—macrophage fraction significantly correlates with worse response (p=0.0278, figure 5C).

Figure 5

Immune cell analysis by QIF and transcriptomic deconvolution. (A) Heat plot depicts magnitude and direction of associations between deconvolved immune cell populations and pathological regression. (B, C) Pathological response versus estimated macrophage fractions. Representative images of adenocarcinoma (D) and squamous cell carcinoma (E). (F, G) Pathological response versus CD8+T cell density. QIF, quantitative immunofluorescent. * = p value < 0.05, ** = p value < 0.01

To better characterize these immune differences, two multiplexed IF panels were used to quantify immune cell subsets in the 13 pretreatment and 15 post-treatment tumor samples available (5 were matched pretreatment/post-treatment samples). The first panel examined markers to delineate T cells (CD3, CD8), PD-L1, Tregs (FoxP3), and macrophages (CD68); the second elucidated markers of antigen presentation machinery (MHC-I, MHC-II, beta 2 microglobulin) and myeloid cell differentiation (CD11c) relative to tumor (pan-cytokeratin) (online supplemental table 1, figure 5D,E).

From the quantitative IF (QIF) data, we also uncovered key differences between ADC and SCC in biomarker response. ADCs with higher CD8+T cell infiltration post-treatment had overall better pathological response (figure 5F; p=0.0083), consistent with other tumor types and NSCLC studies.22 However, SCCs did not show this trend (figure 5G). In total, these findings suggest that immune infiltration as a biomarker of ICI response might be tumor-type specific for NSCLC.

Discussion

After 39.5 months of median follow-up, survival data from this first reported study of neoadjuvant chemoimmunotherapy in resectable NSCLC confirms 3-year DFS and OS rates on par with the practice-changing studies that have established immunotherapy in the perioperative management of this disease.

As with all immunotherapy studies in NSCLC management, select patients may experience prolonged benefits; however, identifying these patients remains a key challenge. This study and others23 ,24nominated degrees of pathological response (MPR, pCR) as primary endpoints in the hopes that these would serve as reasonable surrogates of improved survival outcomes. Although the primary endpoint for this study was met (MPR=57%), the associations between pathological endpoints and DFS and OS, respectively, were not statistically significant. Pathological response also did not associate significantly with disease recurrent to the brain. These findings are not unexpected given our small sample size, and in KEYNOTE-6714, there was a suggestion of curve separation based on pCR status in an exploratory analysis.

Importantly, all these studies point toward the need to further refine biomarkers in order to intensify or deescalate treatment appropriately. Both AEGEAN5 and CheckMate-81625 have reported on ctDNA clearance, with the immunotherapy arms having increased rates of clearance, and patients meeting pCR also with increased rates of clearance. One can perhaps envision a paradigm where a composite of treatment response, incorporating clinical, radiographic, pathological, and peripheral blood features, may offer patient-level management insights.

Although our study is small and consequently not all molecular subtypes are represented with adequate power (ie, lower frequency of non-squamous KRAS-mutant cases than historically reported), paired samples were available to analyzed across multiple modalities, which is a strength of neoadjuvant studies. It has been observed in the literature that STK11-mutant and KEAP1-mutant lung cancers tend to contain fewer immune cells capable of effectuating antitumor activity and are less likely to respond to immunotherapy.19 26 For us, clinical outcomes did not associate significantly with STK11 or KEAP1 mutation status, but patients whose tumors featured reduced copy number of these genes were less likely to achieve MPR. Tumors with low copy number of STK11 and KEAP1 had reduced density of cytotoxic T cells, which were less frequently co-localized with M1-like macrophages and Treg. This finding highlights that it may not be only deleterious mutations, but possibly also copy number events involving these genes, that confer a pro-tumorigenic immune state. That reduced copy number of these genes is more common than deleterious mutations in lung cancer may expand our ability to identify patients at increased risk of diminished benefit from immune checkpoint blockade from careful analysis of pretreatment specimens. Are these patients more likely to benefit from the enhanced immune activation, effector memory, and cytotoxic T cell function as observed with the incorporation of CTLA-4 inhibition in the NEOSTAR neoadjuvant study?27 Further research to characterize the clinical and biological significance of reduced copy number of STK11 and KEAP1 in lung cancer is warranted, particularly in larger datasets adequately representative of KRAS mutations in the broader population.

Our translational analysis of immune cells provides further support for differential immune composition and clinical significance by tumor histology in lung cancer. It was noted in translational analysis of the LCMC3 cohort (treated with atezolizumab monotherapy) that transcripts of PD-L1 and immunoglobulin-like transcript 2 (ILT2) significantly associated with pathological response in non-squamous tumors.23 In our study, we observed across both QIF and transcriptomic analysis that CD8+T cell infiltration associated significantly with pathological response in ADCs but not SCC. Increased M1-like or CD68+macrophage density trended with pathological regression in ADCs. In SCC, there was not a significant association between M1-like macrophage density and outcome by QIF, but in fact the opposite trend was seen by transcriptomic analysis. These findings are concordant with our prior study of nearly 800 NSCLCs across all stages of disease in which levels of CD68 (as measured by QIF) were shown to have differential prognostic association with respect to histology, with high-CD68 SCCs trending toward worse outcomes.22 This distinction within NSCLC has important implications for identifying patients most likely to respond to ICI and should be closely studied in translational samples collected from larger perioperative studies.

Taken together, our survival rates at 3 years are compatible with those seen from other perioperative studies of ICIs in resectable NSCLC; inconsistent survival predictivity of pathological response in our study most likely reflects smaller sample size and inclusion of both lung cancer histologies. As the field encounters the critical patient selection questions necessary to further refine treatment paradigms, our integrated analysis nominates new questions for deeper inquiry: what is the significance of discrepant immune cell composition and benefit with respect to NSCLC histology? Can the phenotype of relative immune exclusion and diminished benefit from chemoimmunotherapy identified by mutations in STK11 and KEAP1 genes be identified more broadly on the level of copy number changes? Further research to characterize the clinical and biological significance of reduced copy number of STK11 and KEAP1 in lung cancer is warranted, particularly in larger datasets adequately representative of KRAS mutations in the broader population. We look forward to continuing the pursuit of these and other essential gaps in the clinical and translational literature over the years to come.

Data availability statement

Data are available on reasonable request. We have included the most relevant data in the supplemental materials and the previously published manuscript for this trial.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

This study involves human participants and was approved by Columbia IRB AAAQ3153. Participants gave informed consent to participate in the study before taking part.

Acknowledgments

We would like to thank all the patients and families involved with this trial.