Integrated Multi-Omics Approaches for Predicting Immune Checkpoint Inhibitor Response in NSCLC – Insights From Genomics, Proteomics, and Metabolomics

Introduction

Background

Lung cancer is the leading cause of cancer-related mortality worldwide, responsible for nearly one in five cancer deaths.¹ Non-small cell lung cancer (NSCLC) accounts for ~80% of cases, with adenocarcinoma, squamous cell carcinoma, and large cell carcinoma as the major subtypes.^2,3 Tobacco smoking remains the primary risk factor, while radon exposure is the second leading cause, particularly among non-smokers.³ At diagnosis, nearly half of cases present with metastatic disease, and five-year survival rates remain low, especially for advanced stages.^4,5 These figures emphasize the urgent need for improved therapies and reliable biomarkers to guide treatment selection.

Treatment for early-stage NSCLC (I–II) usually combines surgery, radiation, and platinum-based chemotherapy.⁶ Stage III patients without disease progression may receive immunotherapy, and stage IV patients benefit from molecularly targeted and immunotherapies, including cytokine therapy, adaptive cell transfer, and tumor vaccines.⁶

Immune checkpoint inhibitors (ICIs) are a type of immunotherapy that block inhibitory pathways used by cancer cells to escape the immune system, thereby helping the body’s immune cells recognize and attack tumors. ICIs have emerged as a promising option due to the limitations of earlier treatments.⁷ FDA-approved ICIs and their indications for treatment of NSCLC are summarized in Table 1.

Table 1 FDA-Approved Immune Checkpoint Inhibitors and Their Clinical Indications

Immune checkpoints regulate immune responses, preventing attacks on healthy cells, but their dysfunction contributes to NSCLC progression.⁷ Programmed Death-1 (PD-1), found on T cells, B cells, and natural killer cells, binds to Programmed Death-Ligand 1 (PD-L1) and PD-L2 on tumor and immune cells, sending inhibitory signals that suppress T cell activity and enable immune evasion.^8,9 PD-L1 is expressed in 53–62% of NSCLC cases¹⁰ and is often upregulated, contributing to immune evasion by allowing tumors to escape immune detection, particularly in advanced stages.^9,11 Another checkpoint, Cytotoxic T-Lymphocyte Antigen-4 (CTLA-4), competes with the co-stimulatory receptor CD28 for binding to B7 molecules on antigen-presenting cells. While CD28 promotes T cell activation, CTLA-4 inhibits it, dampening the immune response.^9,11,12 The mechanisms of immune checkpoint blockade are illustrated in Figure 1. Tumors exploit both PD-1/PD-L1 and CTLA-4 pathways to suppress immunity and promote growth.⁹ These biological insights have directly informed the therapeutic targeting of PD-1/PD-L1 and CTLA-4 in NSCLC.

Figure 1 Mechanism of immune checkpoint blockade. Professional antigen-presenting cells activate naïve T cells via co-stimulatory interactions between B7 (CD80/86)/CD28 and the MHC-II/TCR complex. CTLA-4 inhibitors enhance activation by preventing CTLA-4 from binding to B7 ligands. Activated effector T cells secrete cytokines such as IFNγ, amplifying the immune response against tumors. Tumor-expressed PD-L1 binds to PD-1 on T cells, suppressing activity even in the presence of tumor antigens. Regulatory T cells (Tregs) further inhibit T cell function, promoting an “exhausted” phenotype.PD-1/PD-L1 inhibitors restore anti-tumor immunity by blocking this suppressive interaction. Abbreviations: CTLA-4, cytotoxic T-lymphocyte-associated protein 4; MHC, major histocompatibility complex; PD-1, programmed cell death protein 1; PD-L1, programmed death ligand 1; TCR, T cell receptor. Note: Figure created with BioRender.

Building on this knowledge, the PD-1/PD-L1 and CTLA-4 pathways have become critical therapeutic targets, with ICIs significantly advancing the treatment of NSCLC by restoring T-cell-mediated anti-tumor immunity. ICIs, including anti-PD-1/PD-L1 and CTLA-4 inhibitors (eg, ipilimumab), are FDA-approved for neoadjuvant, adjuvant, and first-line use—either as monotherapy or in combination with platinum-based chemotherapy.^6,10,13–15 In advanced NSCLC, combining ICIs with chemotherapy has improved overall and progression-free survival.^16,17 Furthermore, early-stage trials support the integration of ICIs into neoadjuvant and adjuvant settings, demonstrating increased major and complete pathological response rates.^18–21 Yet only 20–30% of patients respond,^22–27 resistance is common, and immune-related toxicities occur.²⁸ This highlights the need for reliable biomarkers.

Among the biomarkers investigated, PD-L1 was the first to receive FDA approval.²⁹ High tumor proportion score (TPS ≥50%) is generally associated with better outcomes,^30–32 yet inconsistent findings across trials,^33–35 together with variability in immunohistochemistry assays and intratumoral heterogeneity,^36,37 limit its reliability. Microsatellite instability and mismatch repair deficiency (MSI/dMMR) were subsequently approved as predictive biomarkers for pembrolizumab in metastatic or unresectable tumors.³⁸ Although clinical trials demonstrated meaningful response rates,^39–41 their extremely low prevalence in NSCLC and methodological challenges constrain clinical applicability.⁴² More recently, tumor mutational burden (TMB) received tumor-agnostic approval in 2020.²⁹ Elevated TMB is linked to increased neoantigen load and better ICI response,^42–46 however, controversies persist regarding appropriate thresholds—≥10 mut/Mb in trials such as CheckMate 227 and 568^44,47,48 versus higher cut-offs of ≥16–20 mut/Mb or ≥200 nonsynonymous mutations by whole-exome sequencing.^49,50 Added to this are differences in sequencing platforms and bioinformatics pipelines that further undermine reproducibility.^51,52 Taken together, PD-L1, MSI/dMMR, and TMB represent the first generation of biomarkers that established proof of principle for biomarker-guided immunotherapy in NSCLC. Yet their inconsistent predictive performance, technical variability, and limited generalizability underscore the need for more robust and integrative strategies.

Several landmark clinical trials and large-scale consortia have driven biomarker development. The KEYNOTE trials validated PD-L1,^53,54 while CheckMate studies highlighted mutational load and combination regimens.^16,44 Large initiatives like Blueprint⁵⁵ and TRACERx⁵⁶ advanced assay harmonization and multi-omic profiling (Supplementary Table S1). Together, these efforts define the foundation of biomarker research in NSCLC, illustrating both the validation of PD-L1 and TMB and the collaborative infrastructure that enables cross-platform standardization.

Nevertheless, translating biomarkers into practice remains fraught with challenges. Clinically, immune-related adverse events, atypical responses, and comorbidities complicate care.⁵⁷ Technically, assay variability undermines reproducibility.^50,55,58,59 Economically, disparities in access and affordability remain pressing, with cost-effectiveness differing across therapy lines and health systems.⁶⁰ Debates also persist regarding the predictive reliability of PD-L1 and TMB. Some trials show strong associations between high PD-L1 expression and outcomes, while others demonstrate inconsistent value due to assay variability and heterogeneity.^55,59 For TMB, thresholds of ≥10 mut/Mb have been widely applied, yet other studies suggest higher cut-offs, further complicating clinical implementation.^44,47–50 MSI/dMMR, though highly predictive in colorectal cancer, remains exceedingly rare in NSCLC (<1%), raising questions about clinical utility.^61,62 Collectively, these discrepancies highlight the urgent need for more integrative biomarker strategies.

Emerging approaches aim to address these limitations by combining multiple modalities. Multi-omics analyses integrated with artificial intelligence (AI) are generating clinically actionable predictive models.^63,64 Blood-based proteomic platforms such as PROphet (Proteomic Prediction of immunotherapy benefit) have shown value alongside PD-L1 testing for treatment planning in metastatic NSCLC.⁶⁵ Advances in tumor microenvironment (TME) profiling, including spatial proteomics, add predictive value beyond bulk assays.^66,67 The gut microbiome has also emerged as a modulator of ICI efficacy, with dysbiosis and antibiotic exposure linked to reduced responses.⁶⁸

While tumor-intrinsic features such as PD-L1, TMB, and MSI/dMMR remain central to biomarker development, accumulating evidence shows the TME exerts decisive influence. High CD8⁺ T-cell infiltration, tertiary lymphoid structures, and favorable effector-to-suppressor ratios improve outcomes, whereas stromal exclusion, myeloid-derived suppressor cells, and regulatory T-cell dominance promote resistance.^69–71 Stromal remodeling and cytokine signaling dynamically modulate immune accessibility and checkpoint pathways. Multi-omics technologies provide powerful means to interrogate these interactions: proteomics captures cytokine and checkpoint profiles; metabolomics reveals nutrient competition and immunosuppressive metabolites; spatial proteomics and transcriptomics map immune–stromal crosstalk.^66,67 Integrating these TME-derived signals with tumor-intrinsic biomarkers is therefore essential to overcome the limitations of single-parameter approaches and establish robust composite predictors.

Several reviews have addressed NSCLC biomarkers, but with limitations. Bourbonne et al⁷² emphasized genomics and transcriptomics with little focus on proteomics and no discussion of metabolomics. Mei et al⁶⁴ integrated multi-omics with AI but underrepresented proteomics and excluded metabolomics. Yoon et al⁷³ presented a broad overview of metabolomics and advanced omics but lacked NSCLC-specific clinical focus. Other reviews emphasized single-cell or spatial omics⁷⁴ or circulating biomarkers⁷⁵ without systematic integration. In contrast, the present review provides a balanced synthesis of genomics, proteomics, and metabolomics, anchored in the TME context. By addressing both strengths and limitations of existing biomarkers, it highlights the need for integrated multi-omics approaches that combine tumor-intrinsic, TME, and host-related factors into clinically actionable models for NSCLC immunotherapy.

Objective

This review explores the predictive landscape of ICI response in NSCLC, with a focus on the emerging role of genomics, proteomics, and metabolomics in identifying novel biomarkers. By highlighting recent advancements and current challenges, the review aims to provide comprehensive insights into how integrative multi-omic approaches can enhance prediction accuracy, inform personalized immunotherapy, and ultimately improve patient outcomes.

Methods

A comprehensive literature search was performed to identify studies reporting on genomic, proteomic, or metabolomic biomarkers predictive of response to ICIs in NSCLC. The search was conducted on April 11, 2024 across three databases: PubMed, ClinicalTrials.gov, and Google Scholar, covering the period from 2016 to January 2025. The search strategy combined controlled vocabulary and free-text terms, using Boolean operators. For example, the PubMed search string was: (“immune checkpoint inhibitors” OR “ICIs”) AND (“NSCLC” OR “non-small cell lung cancer”) AND (“multi-omics” OR “genomics” OR “proteomics” OR “metabolomics”) AND (“response prediction” OR “predictive biomarkers” OR “tumor microenvironment”). The search strategy is summarized in Table 2, and the detailed PubMed Boolean string is provided in Supplementary Table S2.

Table 2 Summary of Search Strategy for Literature Retrieval

Eligibility criteria included: (1) human studies, (2) English language, (3) peer-reviewed articles, reviews, clinical trials, or meta-analyses, and (4) studies evaluating genomics, proteomics, or metabolomics in relation to ICI response in NSCLC. Exclusion criteria were: non-human studies, non-English articles, conference abstracts without full text, and case reports with fewer than 10 patients.

Study selection was conducted in two stages: (1) title and abstract screening, and (2) full-text review. The initial screening was performed by E. Elayeh, and independently reviewed by L. Dahabiyeh. Any disagreements were resolved by discussion until consensus was achieved. Duplicate records were removed prior to screening. To ensure methodological rigor, the review followed PRISMA 2020 guidelines for reporting. Risk of bias and study quality were assessed qualitatively according to study design, sample size, and biomarker validation stage.

To complement this qualitative assessment with a standardized approach, we further applied a structured evidence rubric across all three omics domains. Each genomic, proteomic, and metabolomic study was scored on a 0–14 scale derived from seven domains, with each domain contributing 0–2 points: (1) Effect robustness (0 = exploratory, 1 = association without replication, 2 = replicated or consistent across analyses); (2) Statistical validity (0 = descriptive only, 1 = p-value without correction, 2 = multivariate analysis or correction for multiple testing); (3) Clinical endpoint relevance (0 = surrogate only, 1 = intermediate endpoint such as ORR, 2 = hard endpoint such as PFS or OS with HR/CI); (4) Adjustment for confounders (0 = none, 1 = limited adjustment, 2 = comprehensive multivariate adjustment including PD-L1, TMB, stage, etc).; (5) Validation (0 = discovery only, 1 = internal/orthogonal validation, 2 = external cohort validation); (6) Cohort size (0 = <30, 1 = 30–49, 2 = ≥50 patients or multi-center); and (7) Data availability (0 = none, 1 = partial availability, 2 = fully deposited in repositories such as GEO, PRIDE, or MetaboLights). Scores were summed across domains to generate a total between 0 and 14.

For example, a proteomic study reporting differential plasma cytokines in 43 patients with multivariate Cox regression (adjusting for stage and performance status) and orthogonal validation by ELISA scored 13/14. In contrast, a small exploratory metabolomics study with fewer than 20 patients, univariate statistics, no adjustment for PD-L1 or TMB, and no validation scored 6/14.

Based on these standardized scores, all studies were retained in the main text to provide a comprehensive overview of the available evidence. To ensure clarity, we explicitly highlighted high-evidence studies (≥10/14) as robust predictors, while studies scoring <10/14 were clearly designated as exploratory or low-evidence. This approach preserved completeness while also guiding readers toward the most translationally relevant findings. In total, the screening process resulted in the inclusion of 51 eligible studies, comprising 33 genomic, 9 proteomic, and 9 metabolomic investigations.

Results: Omics Approaches Used to Predict Response to ICI Therapy in NSCLC

Omics studies including genomics, transcriptomics, proteomics and metabolomics, are advanced bioanalytical approaches that focus on the identification, quantification and characterization of the entire molecules “omes” within a specific system.⁷⁶ A genome is defined as the complete genetic material of an organism, while transcriptome is the complete set of RNA transcripts produced by the genome. A proteome refers to the complete set of proteins expressed by an organism, cell, or tissue and metabolome is the entire set of small-molecule chemicals (metabolites with molecular weight <1500 Da) found within a biological sample.⁷⁷

In the context of this review, these omics approaches were systematically evaluated for their ability to predict response to ICIs in NSCLC, and the results are presented below according to genomic, proteomic, and metabolomic domains.

Genomics

The use of biomarkers in the clinical care of cancer patients has grown in significance with the advent of genomic profiling technology and selective molecular targeted medicines. As predictive biomarkers for treatment decision-making, single-gene/protein or multi-gene “signature”-based assays have been developed to quantify particular molecular pathway dysregulations. It is also possible to include genome-based prognostic biomarkers for various cancer types into clinical prognostic staging systems or practice guidelines.⁷⁸

Several techniques have been used in identifying genomic alterations for prediction of response to ICI in NSCLC patients.

Next Generation Sequencing in genomics Analysis

Next-generation sequencing (NGS) is among the most frequently reported genomic techniques for predicting immune checkpoint inhibitor (ICI) response in NSCLC.^79–82 Characterized by high-throughput, multiplexed, and clonal sequencing, NGS offers significant advantages over traditional Sanger sequencing.⁷⁹ NGS technologies are broadly categorized into short-read (second-generation) and long-read (third-generation) sequencing.⁷⁹ Second-generation sequencing enables parallel sequencing of short (250–800 bp) clonally amplified DNA molecules, and involves a pipeline of library preparation, sequencing, and data analysis.⁸³ DNA sequencing approaches include Whole Genome Sequencing (WGS), Whole Exome Sequencing (WES), Epigenome Sequencing, and Targeted Sequencing (TS), depending on the template.⁸⁴ Long-read technologies, in contrast, can sequence native DNA fragments larger than 10 kb. Although early versions had high error rates, recent improvements have significantly enhanced their accuracy, making them useful for identifying genetic disorders.⁸⁵ The two primary long-read platforms are Oxford Nanopore Technology (ONT) and Pacific Biosciences (PacBio).

Potential Genomic Mutations for Prediction of Response to ICI in NSCLC

Cancer classification has shifted from tissue-based to genetics-based, driven by initiatives like The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium (ICGC), which have deepened understanding of genetic mechanisms, molecular subtypes, and tumor heterogeneity. Platforms such as cBioPortal (https://www.cbioportal.org/) support this shift by allowing exploration of mutation data across cancers. A key success of genomics-guided therapy is the use of vemurafenib in BRAF V600-mutant metastatic melanoma, present in around 50% of screened patients.⁸⁶ This paradigm demonstrates how genetic alterations can guide therapy, and a similar principle applies in NSCLC, where genomic context modulates ICI response.

To evaluate the robustness of available evidence, we systematically scored studies according to predefined criteria mentioned earlier. Most genomic studies (34/39) achieved high evidence scores (≥10), while four scored below this threshold.^87–90 High-scoring studies form the primary evidence base, and lower-scoring studies are presented as exploratory. To improve clarity, we summarized genomic predictors of ICI outcomes across two complementary main tables. Table 3 presents alterations associated with poor response and Table 4 highlights markers linked to favorable response. These tables provide a concise overview of candidate biomarkers, reported outcomes, validation status, and strength of evidence. Full quantitative details, including effect estimates (hazard ratios, odds ratios, confidence intervals, and p-values), are provided in Supplementary Table S3a (Excel file), which contain separate sheet for genomics studies. While the tables summarize the evidence base, Figure 2 provides a mechanistic overview that situates these alterations within broader genomic pathways.

Table 3 Genomic Predictors of Poor Response to ICIs in NSCLC

Table 4 Genomic Predictors of Favorable Response to ICIs in NSCLC

Figure 2 NGS workflow, pathways, and genetic mutations implicated in response to immune checkpoint inhibitor therapy in NSCLC. Abbreviations: ALK, anaplastic lymphoma kinase; EGFR, epidermal growth factor receptor; BRAF, v-raf murine sarcoma viral oncogene homolog B1; DLX2, distal-less homeobox 2; EPHA, Ephrin type-A receptor; HER2, human epidermal growth factor receptor 2; HSPG2, heparan sulfate proteoglycan; KRAS, Kirsten rat sarcoma viral oncogene homolog; MAPK, mitogen-activated protein kinase; MGA, MAX gene-associated; PI3K, phosphoinositide 3-kinase; TGFβ, transforming growth factor beta; TP53, tumor protein 53; UBE3A, ubiquitin protein ligase E3A. Note: Figure created with BioRender.

Predictors of Poor Response

KRAS Co-Mutations with STK11/LKB1

In KRAS (Kirsten rat sarcoma viral oncogene homolog)-mutant NSCLC, concurrent alterations in STK11 (serine/threonine kinase 11, also called LKB1) were consistently associated with primary resistance to PD-1/PD-L1 inhibitors.^108,109 These tumors displayed reduced PD-L1 expression, diminished CD8+ T-cell infiltration, and enrichment of immunosuppressive myeloid cells, consistent with an immune-cold phenotype that mediates checkpoint resistance.

KRAS Co-Mutations with KEAP1

Alessi 2021 reported that KEAP1 (Kelch-like ECH-associated protein 1) alterations, particularly in the context of KRAS-mutant NSCLC, were associated with inferior outcomes under checkpoint blockade. KEAP1 loss activates the NRF2 (nuclear factor erythroid 2–related factor 2) pathway, leading to altered oxidative metabolism and immune evasion, which together diminish immunotherapy benefit.⁹⁵

KRAS Subtype G12D and Other Non-G12C Variants

The G12D variant of KRAS has been reported as a marker of poor response to PD-(L)1 therapy. Stronger evidence indicates that G12D tumors show reduced tumor mutational burden (TMB), lower PD-L1 expression, and poor CD8+ T-cell infiltration, consistent with weaker immunogenicity.⁹¹ An exploratory cohort with limited evidence further suggested a similar association (score <10).⁹⁰ More broadly, non-G12C KRAS variants were less responsive to checkpoint blockade than KRAS G12C, highlighting the relevance of subtype-specific biology.⁹⁶

EGFR and ERBB2

Two large analyses reached similar conclusions regarding the limited benefit of immunotherapy in EGFR-mutant NSCLC. Lee et al (2017) demonstrated that EGFR-mutant patients derived no overall survival improvement with checkpoint inhibitors compared to docetaxel, while benefit was observed in the EGFR-wild-type subgroup.⁹² Lee et al (2018) confirmed these findings in a pooled cohort, further supporting EGFR mutation as a predictor of poor outcome under PD-1/PD-L1 blockade. Fang 2019 extended these results by showing that activating alterations in EGFR and ERBB2 (human epidermal growth factor receptor 2) were associated with shorter progression-free survival under immunotherapy.⁹³

TP53 Nonsense Mutations

Sun et al (2020) demonstrated that the nature of TP53 (tumor protein p53) alterations influences outcome. While some TP53 mutations aligned with inflamed phenotypes, nonsense or truncating variants were associated with impaired interferon signaling, lower PD-L1 expression, and reduced clinical benefit from checkpoint inhibition.⁹⁷

SWI/SNF Complex Mutations

Alessi et al (2021) also investigated the SWI/SNF (SWItch/Sucrose Non-Fermentable) chromatin remodeling complex, including ARID1A, ARID1B, ARID2, PBRM1, SMARCA4, and SMARCB1. While outcomes in unselected NSCLC were not significantly different between SWI/SNF-mutant and wild-type tumors, in the KRAS-mutant subset concurrent SWI/SNF alterations, particularly SMARCA4, were associated with significantly shorter progression-free and overall survival. These findings indicate that SWI/SNF disruption amplifies resistance to immune checkpoint inhibitors in specific molecular contexts.⁹⁵

Other Oncogenic Drivers: ALK, ROS1, RET, MET

Negrao et al (2021) reported that alterations in ALK (anaplastic lymphoma kinase), ROS1, RET, and MET were linked to poor outcomes under PD-1/PD-L1 blockade. These oncogene-driven tumors typically occur in never-smokers, are characterized by low TMB, and harbor immune-excluded microenvironments, explaining the lack of immunotherapy benefit.⁹⁶

Germline Immunogenetic Variants

Xin et al (2023) highlighted the role of host germline variants in shaping ICI outcomes. Mutant alleles in ATG7 (autophagy related 7), TLR4 (toll-like receptor 4), and CD274 (encoding PD-L1) were associated with early progression under PD-1/PD-L1 therapy. These variants likely impair autophagy-mediated immune activation, alter innate sensing, or disrupt PD-L1 regulation, collectively contributing to checkpoint resistance.⁹⁴

Predictors of Favorable Response

KRAS with TP53 (KP Subtype)

Several recurrent genomic alterations in NSCLC have been associated with improved outcomes under ICIs, reflecting their ability to shape an immune-active tumor microenvironment. Among the best studied is the co-occurrence of KRAS and TP53 mutations, often termed the KP subtype. KRAS mutations alone represent a heterogeneous group, but when accompanied by TP53 alterations they consistently align with favorable outcomes across multiple cohorts.^98,108,109 Mechanistically, the combined effect of KRAS-driven oncogenic signaling and partial loss of TP53 function appears to generate genomic instability, increased neoantigen load, and activation of interferon pathways, which together promote higher TMB, upregulation of PD-L1, and robust infiltration by cytotoxic CD8-positive T cells. These features collectively define an inflamed or “hot” tumor immune microenvironment that is permissive for checkpoint blockade.

TP53 Mutation (Context Beyond KRAS)

Beyond the KRAS context, TP53 mutations in general have shown favorable associations, although the subtype of TP53 alteration is critical. Missense mutations, which often produce a dysfunctional but still expressed protein, correlate with enhanced immune activation, interferon gamma signaling, and greater T-cell infiltration. These tumors frequently express PD-L1 at higher levels, consistent with an immune escape mechanism that is effectively targeted by PD-1/PD-L1 blockade.^12,110 An exploratory cohort with limited evidence further supported this association (score <10).⁸⁷ In contrast, nonsense or truncating TP53 mutations abolish protein function and have been associated with poor outcomes,⁹⁷ underscoring the need to differentiate TP53 subtypes when evaluating predictive value.

TP53 Mutation (Context Beyond KRAS)

KRAS mutations more broadly can be associated with benefit, particularly outside the unfavorable backgrounds of STK11/LKB1 or KEAP1. While meta-analyses have shown that EGFR-mutant tumors do not benefit, KRAS-mutant cases often align more closely with the genomic and immunologic features that support ICI efficacy.^93,111 Subtype analyses have highlighted that KRAS G12C can be a favorable predictor in specific clinical contexts. For example, in patients with high PD-L1 expression treated in the first-line setting, KRAS G12C has been associated with longer progression-free survival, suggesting that its unique mutational landscape confers greater immunogenicity.⁸⁰ The mechanism likely relates to the smoking-associated mutational background common in KRAS G12C tumors, which contributes to higher TMB and the presence of diverse neoantigens, thereby supporting immune recognition.

BRAF

BRAF mutations, particularly the V600E variant, have emerged as another favorable context. Although these cases remain relatively rare, stronger evidence suggests that BRAF V600E-mutant NSCLC responds more favorably to checkpoint blockade compared to other oncogenic drivers such as EGFR or ALK.⁹⁶ Additional exploratory studies with lower evidence levels also reported benefit (scores <10).^88,89 The biologic explanation may involve elevated TMB and activation of the MAPK pathway, which interacts with immune signaling pathways to promote tumor antigenicity.

The NOTCH Family

The NOTCH signaling family has also been implicated in favorable ICI response. Mutations in NOTCH1, NOTCH2, and NOTCH3 have been correlated with improved response rates and survival, while NOTCH4 mutations have been independently associated with longer survival in treated cohorts.^100,101 The NOTCH pathway regulates T-cell development and activation, and alterations within this pathway may enhance interferon responses and antigen presentation, explaining the observed clinical benefit.

EPHA

Favorable associations have also been identified with alterations in ephrin receptors, particularly EPHA5. Studies demonstrated that EPHA mutations were linked to higher response rates, longer progression-free survival, and improved overall survival.^106,112 Mechanistically, EPHA5-mutant tumors displayed increased TMB and enhanced immune infiltration, suggesting that loss of EPHA5 function creates a more immunogenic state.

The Hippo Signaling Pathway

The Hippo signaling pathway, which regulates organ size and tissue homeostasis through control of YAP and TAZ activity, has also been implicated. Mutations in Hippo pathway regulators such as FAT1, LATS1, LATS2, and NF2 were associated with favorable ICI outcomes.¹⁰⁴ Disruption of Hippo signaling may enhance tumor immunogenicity by increasing antigen presentation and interferon signaling, thereby facilitating immune-mediated tumor control.

RELN, HSPG2, PAK7, MGA

Other favorable mutations identified across multiple datasets include RELN, which encodes the extracellular matrix protein reelin; HSPG2, encoding the basement membrane proteoglycan perlecan; and PAK7, a serine/threonine kinase involved in cytoskeletal signaling. Each of these genes, when mutated, was associated with longer survival or higher response under PD-1/PD-L1 blockade.^81,102,105 The precise mechanisms vary but generally converge on immune activation, either through altered extracellular matrix signaling, improved T-cell access, or modulation of immune checkpoints. MGA (MAX gene associated) further supports this convergence by linking transcriptional deregulation to immune activation. MGA encodes a transcriptional repressor that interacts with the MAX protein, antagonizing the MYC oncogene. Loss of MGA function disrupts this balance, allowing unchecked MYC signaling, which promotes tumor progression. However, MGA-mutant NSCLCs have been reported to exhibit higher TMB, enriched interferon signaling, and increased infiltration by CD8+ T cells.¹⁰³ This inflamed phenotype provides a mechanistic basis for the improved outcomes observed in ICI-treated patients. In other words, while MGA loss deregulates growth control through MYC, it simultaneously generates a tumor environment more visible to the immune system, which makes ICIs more effective. UBE3A deletion has also been reported as a favorable marker, associated with longer progression-free and overall survival.¹⁰⁷ UBE3A encodes a ubiquitin ligase, and its loss may deregulate protein degradation pathways in ways that enhance immune recognition.

Multi-gene signatures integrating DNA damage response and immune signalling

Cucchiara et al⁹⁹ integrated MSK-IMPACT sequencing with network clustering to define composite predictive signatures in NSCLC patients treated with ICI. Two signatures emerged as favorable: Signature A, enriched for SWItch/Sucrose Non-Fermentable (SWI/SNF) chromatin-remodeling complex and ataxia telangiectasia mutated–target of rapamycin (ATM–TOR) DNA damage response pathways, and Signature B, enriched for NOTCH signaling, interleukin-7–Forkhead box O (IL7–FoxO), and phosphoinositide 3-kinase–Akt (PI3K–Akt) pathways, alongside epithelial–mesenchymal transition (EMT) and inflammatory modules. The presence of either signature was consistently associated with improved PFS and OS across discovery, internal testing, and external validation cohorts, highlighting both genomic instability–linked and immune signaling–linked mechanisms as potential drivers of immunotherapy sensitivity.

Taken together, favorable genomic predictors share several convergent mechanistic themes. The first theme is the inflamed tumor microenvironment. Alterations such as KRAS/TP53 co-mutations, MGA mutations, NOTCH family mutations, and RELN or EPHA5 alterations consistently align with higher levels of cytotoxic CD8-positive T-cell infiltration and activation of interferon signaling pathways. These changes create an immune-hot environment that supports recognition and elimination of tumor cells once the PD-1/PD-L1 axis is blocked. A second theme is enhanced antigenicity. TP53 missense mutations, BRAF V600E, Hippo pathway alterations involving FAT1, LATS1, LATS2, and NF2, as well as ephrin receptor mutations, are associated with higher TMB and a broader neoantigen repertoire. This increased antigenicity improves the likelihood of effective immune surveillance and augments response to checkpoint inhibition. A third theme is checkpoint pathway engagement. Many favorable genomic contexts, including KRAS/TP53, MGA, and NOTCH alterations, are correlated with elevated PD-L1 expression. This suggests that these tumors rely on the PD-1/PD-L1 axis for immune escape, making them particularly sensitive to therapeutic blockade of this pathway. By contrast, poor predictors demonstrate the opposite pattern. KRAS with STK11 or KEAP1 co-mutations, EGFR-driven tumors, and SWI/SNF or SMARCA4 alterations were all associated with immune-cold phenotypes characterized by reduced interferon signaling, lower PD-L1 expression, and exclusion of effector T cells from the tumor bed. Finally, the predictive impact of many alterations is context and treatment dependent. For instance, KRAS G12C was associated with longer progression-free survival in patients with high PD-L1 tumors treated with first-line immunotherapy, but no differential benefit in unselected settings. Similarly, TP53 mutations can predict improved outcomes when they are missense variants, but nonsense or truncating mutations confer resistance. These observations demonstrate that the predictive value of genomic alterations is not absolute, but rather interacts with co-mutation status, PD-L1 expression, TMB, and treatment regimen. Together, these mechanistic insights underscore that genomics stratifies NSCLC into immune-hot and immune-cold archetypes that diverge markedly in their sensitivity to immune checkpoint inhibition.

Heterogeneity and Limitations of Genomics

Genomic predictors of immunotherapy response in NSCLC demonstrate significant heterogeneity, which arises from multiple levels of variability across published studies. Differences in study design are a key driver: some investigations were retrospective analyses of real-world cohorts, while others were prospective clinical trials or exploratory biomarker studies nested within larger treatment protocols. The sequencing platforms employed also vary considerably, ranging from targeted next-generation sequencing panels of a few hundred genes to whole-exome sequencing, RNA sequencing, or broader multi-omics approaches. These methodological differences influence which mutations are captured, how TMB is quantified, and how immune-related gene expression is interpreted. Patient populations also differ substantially across studies, with variability in geography, ethnicity, smoking exposure, histologic subtypes, and prior lines of therapy.

Treatment regimens represent another source of heterogeneity. Some cohorts received anti–PD-1 or anti–PD-L1 monotherapy, whereas others were treated with chemo-immunotherapy combinations, or less commonly CTLA-4–based regimens. Clinical selection factors such as baseline PD-L1 expression and smoking history strongly influence both clinical outcomes and biomarker associations, yet these variables are not uniformly controlled or adjusted across studies. Cohort size further amplifies these discrepancies: small exploratory studies often report striking associations that do not reproduce in larger multicenter datasets, while very large cohorts may dilute signals by pooling heterogeneous patient subsets.

Technical variation also plays a role. Differences in bioinformatic pipelines, variant calling thresholds, and gene-level annotation can lead to discordant classifications of the same genomic alteration across datasets. Endpoint definitions are not always consistent, with some studies reporting objective response rate (ORR), others focusing on progression-free survival (PFS) or overall survival (OS), and some using composite or surrogate endpoints. This lack of uniformity complicates direct comparisons and meta-analyses.

Taken together, these sources of heterogeneity limit the generalizability of individual findings and underscore the importance of validation. Candidate genomic biomarkers require confirmation in large, uniformly treated cohorts with standardized assays, harmonized clinical annotation, and consistent endpoint definitions. Such efforts are essential to distinguish true predictive markers from context-dependent associations driven by study design, patient selection, or technical artifacts. Encouragingly, ongoing large-scale initiatives and prospective trials, such as MSK-IMPACT¹¹³ CheckMate 227,⁴⁴ and the POPLAR and OAK studies^10,114 as well as pan-cancer sequencing consortia, are beginning to address these gaps by providing standardized platforms and robust validation cohorts.

Proteomics

Proteomics focuses on the comprehensive examination of proteins, including their structures and activities.¹¹⁵ Analyzing an organism’s dynamic system of continually shifting protein sets, which vary not only across different developmental stages but also between tissues, cell types, and intracellular compartments, is a difficult challenge for proteomics. These dysregulated proteins might act as potential disease biomarkers.^115,116 Proteome analysis is now rapidly becoming more successful due to the significant advancements in mass spectrometry (MS) and the coupling of MS with separation methods as two-dimensional gel electrophoresis (2DGE) and Ultrahigh performance liquid chromatography (UHPLC).^117,118 Other techniques include protein microarrays, usually called protein chips, which are available in various formats. The four main types of protein microarrays used in cancer research are proteome, antibody, reverse-phase protein array (RPPA), and lectin microarrays. Table 5 presents proteomic predictors of ICI response in NSCLC, detailing the candidate proteins, reported clinical outcomes, validation status, and evidence scores. Full quantitative details, including effect estimates (hazard ratios, odds ratios, confidence intervals, and p-values), are provided in Supplementary Table S3b (Excel file), which contains separate sheet for proteomic studies.

Table 5 Proteomic Predictors of ICI Response in NSCLC

Techniques Utilized in Proteomics Analysis

The two primary techniques in proteomics analysis for predicting responses to ICIs are MS and protein microarrays. Below is a summary of each method, detailing their principles, subtypes, and respective advantages and disadvantages.

Mass Spectrometry

Mass spectrometry (MS) is a critical tool in proteomics, particularly when integrated with machine learning to identify predictive biomarkers for immunotherapy outcomes.¹²⁷ Two main quantification strategies are employed: untargeted (broad profiling) and targeted (focused quantification), using either bottom-up (peptide-based) or top-down (intact protein) approaches.¹²⁸ Common ionization methods include electrospray ionization (ESI) and matrix-assisted laser desorption/ionization (MALDI), enabling the analysis of large polar molecules, with ESI compatible with HPLC and MALDI reducing spectral complexity.¹²⁹ Key mass analyzers include Orbitrap, TOF, and quadrupole, often used in tandem (MS/MS) for detailed fragmentation and ion analysis.¹³⁰ In bottom-up proteomics, proteins are digested into peptides for MS analysis, offering high throughput but limited post-translational modifications (PTM) detail.^131,132 Top-down proteomics analyzes intact proteins, preserving PTMs and modifications, though it is best suited for small proteins and moderate throughput studies.¹³³ The typical workflow used in proteomics research is illustrated in Figure 3, which outlines key steps from sample preparation and protein digestion to mass spectrometry analysis and data interpretation.

Figure 3 Typical proteomics workflow using MS-based techniques. Abbreviations: SDS-PAGE, sodium dodecyl sulfate–polyacrylamide gel electrophoresis; RP-HPLC, reversed-phase high-performance liquid chromatography; LC/MS, liquid chromatography/mass spectrometry. Note: Figure created with BioRender.

Protein Microarrays

Protein microarrays (or chips) detect target proteins by immobilizing biomolecules on solid supports for biochemical analysis.¹³⁴ In cancer research and ICI response prediction, four main types are used: proteome, antibody, reverse-phase (RPPA), and lectin microarrays, offering high-throughput, system-level profiling.¹³⁵ Proteome arrays immobilize proteins from most ORFs and are made via expression/purification or in vitro transcription and translation (IVTT); while expression-based arrays are comprehensive but costly, IVTT offers a simpler, more economical option despite early-stage limitations.¹³⁵ Antibody microarrays use antigen-antibody specificity in either planar or bead-based formats. Planar arrays use slides or membranes with high sensitivity and throughput, while bead-based formats capture antibodies on microbeads.^136,137 These arrays enable proteome comparisons between cancer and healthy individuals, aiding biomarker discovery. Lectin microarrays target abnormal glycosylation, improving detection of tumor-associated glycoproteins and aiding early cancer diagnosis.¹³⁸ RPPA uses fixed lysates probed with specific antibodies to assess protein expression and PTMs across many samples with high sensitivity and multiplexing capabilities.^139–141

Proteomics Study for the Prediction of Response to ICI in NSCLC

The field of using protein (or peptide) biomarkers in clinical investigations is active and constantly expanding. In cancer research, protein biomarkers are used to guide treatment decisions by stratifying patients according to their likelihood of responding to specific medication¹⁴² and to aid in the early detection and prognosis of carcinoma.^143–150 One such example of proteomic use for prediction of response to ICI is the VeriStrat test, which is a blood-based proteomic test or host immune classifier (HIC) first developed in 2007 by Biodesix.¹⁵¹ It is used to predict and prognosticate outcomes for patients with advanced NSCLC. By evaluating the host’s response to the tumor using MALDI-TOF analysis, the test determines the degree of disease aggression and assigns patients to one of two groups: VeriStrat-Good (VS-G) or VeriStrat-Poor (VS-P). It has been demonstrated that the test is predictive for EGFR-targeting drugs in second-line treatment, following progression with or following platinum-based chemotherapy.¹⁵¹ According to a recent analysis of the INSIGHT clinical trial (NCT03289780), NSCLC patients can be accurately predicted to respond differently to ICI therapy based on their VS labels, VS Good and VS Poor, regardless of their PD-L1 expression.¹⁵² Moreover, individual studies and meta-analyses have demonstrated that patients classified as VS-G before treatment experience significantly better outcomes compared to those classified as VS-P across various treatment regimens and even without active therapy in advanced NSCLC.^153–156 Recently, the VS test has demonstrated promise as a prognostic marker for advanced NSCLC treated with immunotherapy.^152,157,158 It has also been found to be able to identify which patients would most benefit from the combination of ICI and chemotherapy.¹⁵²

Potential Protein Biomarkers for Prediction of Response to ICI in NSCLC

All nine proteomic studies included in this review achieved high evidence scores (≥10/14) according to our standardized rubric. This indicates robust statistical design, clinically relevant endpoints, and sufficient cohort sizes, allowing the following results to be interpreted with relatively high confidence.

Predictors of Favorable Response

Chemokines and Cytokines

Increased plasma levels of CXCL9, CXCL10, and interleukin-15 (IL-15) were associated with response to anti–PD-(L)1 therapy. These proteins recruit and sustain CD8+ T cells and natural killer (NK) cells, supporting a T-cell inflamed microenvironment. The same study also demonstrated that higher tumor necrosis factor superfamily ligand 14 (TNFSF14) and Fas ligand (FASLG) were associated with improved overall survival, and that high tumor PD-L1 expression (50–100%) predicted better progression-free and overall survival.¹¹⁹

Checkpoint-Related Proteins

Circulating soluble PD-1 (sPD-1), a decoy receptor form of PD-1, was enriched in responders and linked to improved overall survival in the monotherapy setting.¹²⁰

Proteomic Classifiers

A machine learning–based proteomic test termed PROphet (Proteomic Prediction of immunotherapy benefit), derived from 388 proteins, stratified patients into groups with differential benefit. The contributing proteins spanned immune-regulatory and stromal pathways, reflecting systemic influences on immunotherapy efficacy.⁶⁵ Mass spectrometry–based classifiers further identified “sensitive” patients with proteomic patterns enriched for antigen presentation and immune activation functions.¹²¹

Adhesion Molecules

High tumor-cell expression of CD44 predicted longer survival, consistent with its role in mediating immune–tumor interactions. In contrast, stromal expression of CD44 carried no predictive value.¹²³

Innate Immune Peptides: Neutrophil Defensins

Berghmans et al¹²⁶ applied MALDI-MSI on tumor biopsies from NSCLC patients treated with PD-1/PD-L1 blockade. They identified neutrophil defensins 1–3, antimicrobial peptides of the innate immune system, as being enriched in responders. Defensin signal distinguished tumor and immune compartments with area under the curve values around 0.74, and orthogonal immunohistochemistry confirmed defensin staining on the same biopsies. Although based on a small single-center cohort without external validation, these findings suggest that innate immune activation reflected by defensin expression may promote immunotherapy sensitivity.

Predictors of Poor Response

Chemokines and Cytokines

In a larger cohort, CXCL8 (interleukin-8) and CXCL10 were enriched in non-responders both at baseline and during treatment, correlating with poor survival. These findings directly contrast with the Swedish study, in which CXCL10 appeared favorable. This inconsistency likely reflects heterogeneity in cohort size, clinical characteristics, treatment regimens, and assay platforms, and highlights that CXCL10 may act as either a marker of effective T-cell trafficking or of chronic, non-productive inflammation depending on context.^119,122 The Swedish study further identified caspase-8 (CASP8), adenosine deaminase (ADA), mucin-16 (MUC16), CD244, interleukin-18 (IL-18), and angiopoietin-2 (ANGPT2) as enriched in non-responders, each associated with shorter survival outcomes.¹¹⁹

Intracellular and Alveolar Proteins

Non-responders also demonstrated enrichment of proteins such as yes-associated protein 1 (YAP1), decapping scavenger (DCPS), and stratifin (SFN), which reflect oncogenic stress and were linked to poor outcomes.¹²⁰

Inflammatory Mediators

Systemic inflammation was reflected by higher levels of serum amyloid A1 (SAA1), SAA2, and calcium-binding proteins S100A8 and S100A9, all of which amplify myeloid recruitment and immune escape.¹²⁵ Elevated interleukin-6 (IL-6), a driver of STAT3 signaling and immunosuppressive myeloid differentiation, further supported the link between inflammatory signaling and resistance.¹²²

Mixed or Context-Dependent Predictors

Autoantibodies

Autoantibodies against MYC-associated factor X (MAX) were enriched in responders, whereas those targeting TAP-binding protein (TAPBP) and DEAH-box helicase 29 (DHX29) showed inconsistent associations, underscoring the heterogeneity of humoral immune influences.¹²⁴

Soluble Checkpoints

Plasma levels of soluble immune checkpoints including PD-1, PD-L1, PD-L2, cytotoxic T-lymphocyte-associated protein 4 (CTLA-4), T-cell immunoglobulin and mucin domain-containing protein 3 (TIM-3), B- and T-lymphocyte attenuator (BTLA), herpesvirus entry mediator (HVEM), CD137 (4–1BB), and lymphocyte-activation gene 3 (LAG-3) were also studied. No consistent association with outcomes was observed.¹⁵⁹

Heterogeneity and Limitations in Proteomics

Proteomic predictors of immunotherapy response in NSCLC demonstrated substantial heterogeneity across studies. Sample sizes varied widely, ranging from fewer than 50 patients in some discovery cohorts^119,123 to over 140 patients in larger validation efforts.¹²² Clinical characteristics such as treatment line, PD-L1 expression, smoking status, and histology were not uniformly controlled, complicating cross-cohort comparisons. Different assay platforms were used, including proximity extension assay panels, multiplex ELISAs, and mass spectrometry–based discovery, each with distinct sensitivity and coverage. Endpoint definitions also differed, with some studies focusing on short-term response, others on PFS or OS, and statistical handling ranging from simple logistic regression to multivariate Cox models with internal cross-validation.

An illustrative example of heterogeneity is CXCL10, which was linked to favorable outcomes in the Swedish cohort but unfavorable in the Israeli cohort, highlighting how cohort size, assay design, and inflammatory context can yield opposing results. Similarly, circulating soluble checkpoint proteins showed inconsistent associations, likely reflecting differences in detection sensitivity and biological interpretation as either decoy receptors or tumor burden surrogates. Collectively, these discrepancies limit direct generalization of single markers. Larger multi-institutional studies, harmonized platforms, and standardized endpoints are required to validate proteomic predictors and establish their translational readiness.

Metabolomics

Metabolomics represents the functional readouts of a cellular state that can reveal changes in biological states.¹⁶⁰ In contrast to genes and proteins, which are influenced by epigenetic regulation and PTM, respectively, metabolites act as direct indicators of biological activity, making it simpler to associate them with phenotype.¹⁶¹ Because of their drastically changed metabolism, cancer cells create chemicals that are unique to and characteristic of non-physiological environments.¹⁶⁰ As metabolites can interact with and target various therapies, metabolomics is an important area for advancing the evaluation of immunotherapy and chemotherapy effectiveness.¹⁶² Techniques utilized in metabolomics analysis.

The primary analytical techniques used in metabolomics are MS and nuclear magnetic resonance spectroscopy (NMR). In most of the studies, MS is coupled to a separation technique such as chromatography (eg LC-MS or gas chromatography (GC-MS)) or capillary electrophoresis (eg CE-MS).¹⁶³

NMR

This technique has a number of advantages, including as high throughput, significant repeatability, simple metabolite identification, and non-destructive sample preparation.¹⁶⁴ The fact that NMR is not limited to the study of tissue or bio fluid extracts alone is a significant benefit.¹⁶⁵ Since NMR is not inherently destructive it can be used for metabolite imaging in living samples using magnetic resonance spectroscopy (MRS) and magnetic resonance imaging (MRI), unlike LC-MS and GC-MS based approaches. Nevertheless, NMR is 10–100 times less sensitive than LC-MS and GC-MS methods and provide less metabolome coverage (e.g LC-MS can provide details on 1000+ metabolites with concentrations >10 to 100 nM, while an NMR shows details of 50–200 metabolites with concentrations >1 µM).¹⁶⁶

LC-MS and GC-MS Analyses

Combining chromatographic separation with mass spectrometry improves metabolite detection by reducing sample complexity and ionization interference, enhancing sensitivity, signal consistency, and data reliability.¹⁶⁷ Advances in liquid chromatography, such as capillary monolithic chromatography and Ultra-Performance Liquid Chromatography, have enhanced analysis speed and resolution.¹⁶⁸ While gas chromatography offers higher resolution, liquid chromatography–mass spectrometry is preferred in metabolomics due to its ability to analyze a broader range of metabolites, including polar and thermally unstable molecules like amino acids and organic acids that may degrade during gas chromatography.^169,170 Liquid chromatography–mass spectrometry also simplifies sample preparation and enables higher sensitivity and specificity in complex biological matrices such as plasma, urine, and tissue extracts. The technique can be tailored to specific metabolite types by adjusting chromatographic settings, mobile phases, and ionization modes.^169,170

Several mass spectrometry analyzers are used in metabolomics: quadrupole mass spectrometers are suitable for targeted analysis; time-of-flight mass (TOF) spectrometers are ideal for untargeted studies; ion trap mass spectrometers allow detailed fragmentation; Orbitrap and Fourier transform ion cyclotron resonance mass spectrometers FT-ICR offer high mass accuracy and resolution for complex mixtures; and triple quadrupole mass spectrometers enable highly sensitive targeted quantification via multiple reaction monitoring.¹⁷¹

A typical metabolomics workflow begins with experimental design to select targeted or untargeted approaches. Metabolites are extracted using biphasic liquid methods, depending on the sample and study goals.^172,173 Profiling is done using nuclear magnetic resonance or liquid chromatography–mass spectrometry, followed by data processing with software like XCMS, MAVEN, or MZmine3.¹⁷⁴ Metabolite identification is based on standards, mass data, and tandem mass spectrometry, using databases such as the Human Metabolome Database and METLIN.¹⁷⁵ Statistical analyses identify biomarkers, followed by pathway analysis using resources like the Kyoto Encyclopedia of Genes and Genomes.^174,176 The overall steps involved in metabolomics research are summarized in Figure 4, which depicts the typical workflow from experimental design and metabolite extraction through profiling, data processing, compound identification, and pathway analysis.

Figure 4 Typical metabolomics workflow using MS and NMR-based techniques. Abbreviation: CSF, cerebrospinal fluid; NMR, nuclear magnetic resonance; LC/MS, liquid chromatography/mass spectrometry. Note: Figure created with BioRender.

Metabolomics Study for the Prediction of Response to ICI in NSCLC

Metabolomics studies have been widely employed in cancer research to investigate the underlying disturbed pathways^177–180 and identify metabolic biomarkers or patterns associated with cancer using different biological samples including serum,^181–185 plasma,¹⁸⁶ saliva,¹⁸⁷ urine,^188,189 and breath.^190,191

In contrast to proteomics, the metabolomics literature is more heterogeneous in evidence strength. Of the nine included studies, only four scored ≥10, while the remainder were exploratory and lower-evidence (<10). We therefore highlight high-scoring studies as primary findings and frame the others as preliminary signals requiring further validation. Table 6 presents metabolomic predictors of ICI response in NSCLC, detailing the candidate metabolites, reported clinical outcomes, validation status, and evidence scores. Full quantitative details, including effect estimates (hazard ratios, odds ratios, confidence intervals, and p-values), are provided in Supplementary Table S3c (Excel file), which contains separate sheet for metabolomic studies.

Table 6 Metabolomic Predictors of ICI Response in NSCLC

Potential Metabolite Biomarkers for Prediction of Response to ICI in NSCLC

Predictors of Favorable Response

Amino Acid Metabolism

Lower baseline alanine and pyruvate characterized responders, suggesting reduced glycolytic burden and diminished metabolic competition with lymphocytes (score ≥10).¹⁹² Other favorable metabolites included hippuric acid, butyrylcarnitine, cystine, and glutamic acid, reflecting host–microbial co-metabolism and antioxidant balance (scores ≥10).^193,194 Exploratory cohorts suggested that responders also showed higher baseline tryptophan and a higher histidine/phenylalanine ratio, consistent with preserved amino acid availability for effector T-cell proliferation (score <10).¹⁹⁶

Energy and Cofactor Metabolism

Nicotinic acid (vitamin B3) was enriched in responders, supporting NAD+ biosynthesis, redox balance, and T-cell metabolic fitness; this signal remains exploratory (score <10).¹⁹⁸

Microbiome-Derived Metabolites

High-evidence studies identified additional favorable metabolites noted above (hippuric acid, etc). (scores ≥10).^193,194 Exploratory cohorts reported short-chain fatty acids (SCFAs) such as propionic acid enriched in responders, consistent with enhanced dendritic cell activation and CD8+ effector function (scores <10).^159,198

Predictors of Poor Response

Tryptophan–kynurenine pathway Non-responders demonstrated lower baseline tryptophan and increased indoleamine 2,3-dioxygenase (IDO) activity with accumulation of downstream metabolites such as 3-hydroxyanthranilic acid (3-HAA), which can engage the aryl hydrocarbon receptor and promote T-cell exhaustion; these findings are exploratory (scores <10).^195,196

Energy and Cofactor Metabolism

Higher baseline and on-treatment alanine and pyruvate were linked to progression (score ≥10).¹⁹² Progressive disease was also associated with reduced threonine, urea, and oxalate, indicating disrupted nitrogen metabolism; this extends the pattern but is exploratory (score <10).¹⁹⁷

Lipid and Membrane Remodeling

Non-responders showed increased elaidic acid (a trans-monounsaturated fatty acid) and elevated phosphatidylcholines such as PC C38:0, suggesting intensified membrane synthesis, ER stress, and myeloid activation; these remain exploratory (scores <10).^196,197

Microbiome-Derived Metabolites

Exploratory cohorts reported enrichment of 2-pentanone, tridecane, and p-cresol in non-responders, metabolites linked to dysbiosis, oxidative stress, and impaired dendritic cell function (score <10).¹⁹⁸

Amino Acid Insufficiency Signatures

Reduced branched-chain amino acids (BCAA), which are needed for mTOR-dependent effector T-cell expansion, were observed in non-responders; these signals are exploratory (scores <10).^196,197

Mixed or Context-Dependent Predictors

Amino Acid and Tryptophan Metabolism

While reduced tryptophan and increased IDO activity were associated with poor outcomes in exploratory cohorts, some studies reported preserved tryptophan in responders without uniform changes in the kynurenine/tryptophan ratio, indicating that absolute availability and catabolic flux may both carry predictive value depending on matrix and timing (scores <10).^196,197

Energy Metabolism

Alanine and pyruvate trends were largely unfavorable, with context dependence between plasma versus serum and pre- versus on-treatment sampling; the dominant signal (higher alanine/pyruvate → non-response) is supported by higher-evidence work (score ≥10),¹⁹² with variability in exploratory cohorts (score <10).¹⁹⁷

Heterogeneity and Limitations in Metabolomics

Metabolomic studies also revealed marked heterogeneity in methodology and outcomes. Cohorts were generally small, ranging from fewer than 30 to just over 100 patients,^195,197 and biospecimens varied across plasma, serum, and fecal samples, which may capture distinct metabolic environments. Analytical platforms differed, including NMR, LC-MS, and GC-MS, with variability in compound detection and quantification. Reported endpoints spanned short-term response, PFS, and OS, and statistical strategies ranged from univariate comparisons to multivariate machine-learning models.

Contradictory findings emerged particularly in tryptophan metabolism. While lower tryptophan and increased IDO activity with accumulation of 3-hydroxyanthranilic acid were consistently poor predictors in several cohorts, other studies found that preserved tryptophan without consistent kynurenine elevation marked benefit. Similarly, alanine and pyruvate were unfavorable in most contexts but showed variability depending on biofluid source and sampling timepoint. Microbiome-derived metabolites also displayed cohort-specific patterns, with SCFAs generally favorable but other microbial products such as aldehydes, ketones, and phenols enriched in non-responders.

These inconsistencies reflect the influence of geographic and dietary backgrounds, microbiome diversity, and technical variability in metabolite detection. Limited external validation and small sample sizes further restrict generalizability. Standardization of metabolomics pipelines, integration with microbiome sequencing, and prospective validation in larger, uniformly treated NSCLC cohorts are needed to clarify the true predictive value of metabolic signatures.

Cross-Omics Integration of Genomic, Proteomic, and Metabolomic Predictors

In the following section, we integrate genomic, proteomic, and metabolomic evidence to highlight convergent pathways that shape ICI response in NSCLC. Cross-omics synthesis reveals recurring themes—such as interferon signaling, metabolic rewiring, and immune–stromal interactions—that transcend individual omics layers.

Interferon Signaling and T-Cell Inflammation

Multi-omics evidence converges on interferon signaling and T-cell inflammation as favorable predictors of ICI response. At the genomic level, KRAS/TP53 co-mutations have been associated with increased neoantigen load, interferon activation, and CD8⁺ infiltration^98,104 Proteomic studies identified higher levels of CXCL9, CXCL10, and IL-15 in responders, reflecting chemokine-driven T-cell recruitment and survival.¹¹⁹ Metabolomic evidence showed enrichment of short-chain fatty acids in responders, which promote dendritic cell activation and interferon-stimulated CD8⁺ effector function.^194,198

Adenosine Metabolism and Immune Suppression

Disruption of adenosine metabolism is consistently linked to poor outcomes across omics layers. Genomically, STK11/LKB1 co-mutations (the KL subtype) foster immune-cold phenotypes through adenosine accumulation and myeloid suppressor cell recruitment.^104,108 At the proteomic level, elevated adenosine deaminase (ADA) and CASP8 were observed in non-responders, consistent with disrupted purine metabolism and T-cell dysfunction.¹¹⁹ Metabolomic studies showed reduced branched-chain amino acids and accumulation of glycolytic intermediates such as alanine and pyruvate in non-responders, consistent with metabolic exhaustion that synergizes with adenosine-mediated suppression.^192,197

Tryptophan–Kynurenine Catabolism

Alterations in tryptophan metabolism emerged as another recurrent mechanism of ICI resistance. At the genomic level, alterations in IDO1 and IDO2, as well as aryl hydrocarbon receptor (AhR) activation signatures, have been reported in NSCLC and associated with immune evasion.^100,104 Proteomic studies showed elevated soluble immune checkpoints including PD-1, PD-L1, TIM-3, LAG-3, BTLA, and HVEM in non-responders, reflecting T-cell exhaustion linked to tryptophan depletion and kynurenine accumulation.^120,159 Metabolomic profiling further demonstrated lower baseline tryptophan and increased downstream products such as 3-hydroxyanthranilic acid (3-HAA) in non-responders.^195,196 Collectively, these cross-omics findings implicate the IDO–kynurenine axis as a central immunosuppressive pathway.

Angiogenesis and Stromal Remodeling

Angiogenesis and stromal remodeling are also recurrently associated with poor ICI outcomes. Genomic loss-of-function mutations in KEAP1 activate NRF2 signaling, impairing antigen presentation and promoting an immune-excluded microenvironment.^98,108 Proteomic studies reported elevated angiopoietin-2 (ANGPT2) and IL-6 in non-responders, linking angiogenesis and inflammatory signaling with poor prognosis.^119,122 Metabolomic profiles enriched for phosphatidylcholines and trans-fatty acids such as elaidic acid in non-responders further reflect endothelial activation and lipid remodeling that support stromal exclusion.^196,197

Although few individual markers have been validated consistently across multiple NSCLC cohorts, integration across omics layers highlights convergent biological pathways—interferon signaling, adenosine metabolism, tryptophan catabolism, and angiogenesis—that collectively distinguish immune-hot from immune-cold tumor states.

To consolidate these findings, Table 7 summarizes cross-omics evidence by integrating genomic, proteomic, and metabolomic predictors of ICI response in NSCLC, highlighting convergent pathways, reported outcomes, validation status, and evidence scores.

Table 7 Cross-Omics Integration of Genomic, Proteomic, and Metabolomic Predictors of Immune Checkpoint Inhibitor Response in NSCLC

Conclusion

In conclusion, ICIs have reshaped the therapeutic landscape of non-small cell lung cancer (NSCLC), yet accurately predicting benefit remains a major challenge. PD-L1 expression alone has limited precision, underscoring the need for broader predictive strategies. Evidence from 33 genomic, 9 proteomic, and 9 metabolomic studies demonstrates that multi-omics approaches can identify convergent biomarkers that stratify tumors into immune-hot versus immune-cold phenotypes.

Among genomic alterations, KRAS/TP53 co-mutations, NOTCH family variants, and BRAF V600E consistently associate with favorable outcomes, while EGFR, ALK/RET/ROS1 fusions, and KRAS with STK11 or KEAP1 predict resistance. In proteomics, chemokines such as CXCL9 and CXCL10 and soluble checkpoint proteins (sPD-1, sPD-L1, LAG-3) emerged as promising predictors, whereas acute-phase proteins (SAA1/2, S100A8/9) align with poor response. Metabolomic profiling implicated tryptophan and short-chain fatty acids as favorable, while 3-hydroxyanthranilic acid, pyruvate, and lipid metabolites marked resistance. These markers collectively point to pathways of antigenicity, interferon signaling, and metabolic immune regulation.

Although integrative multi-omics shows clear promise, only a subset of biomarkers have reached consistent validation across independent cohorts, and none yet meet the threshold for routine clinical use. Future work should prioritize prospective trials with pre-specified biomarker endpoints, harmonization of assay protocols and data pipelines, and incremental implementation in clinical laboratories. Ultimately, the integration of multi-omics with artificial intelligence may enable robust, individualized prediction of ICI outcomes and more precise treatment decisions for patients with NSCLC.

Future Directions

To advance biomarker development for NSCLC immunotherapy, several priorities emerge. First, prospective trials with pre-specified biomarker endpoints are essential to validate candidate genomic, proteomic, and metabolomic predictors in large, uniformly treated cohorts. Second, standardized assay protocols—including harmonized sequencing pipelines, proteomic panels, and metabolomics workflows—are needed to reduce variability and improve reproducibility across studies. Third, AI-driven integration pipelines that combine multi-omics data with clinical and radiomic features offer a promising approach to capture the complexity of tumor–immune interactions and generate clinically deployable prediction models. Beyond technical harmonization, cost–benefit assessments are required to ensure that proposed biomarkers can be implemented feasibly in routine practice. Finally, incremental implementation pathways should be pursued in clinical laboratories, beginning with targeted, low-cost assays and progressing to comprehensive multi-omics platforms as evidence and infrastructure mature.

A practical roadmap is also required to bridge discovery and clinical implementation. We propose a phased approach: initial analytical validation of candidate biomarkers using standardized protocols; subsequent development of targeted multi-omics panels in CLIA-certified laboratories to ensure regulatory compliance and reproducibility; and finally, prospective clinical trials incorporating these panels to establish predictive utility and cost-effectiveness. Early adoption may focus on single-analyte assays with the strongest evidence (eg, KRAS/TP53 co-mutations, CXCL9/10, tryptophan metabolites), which can be gradually expanded into multiplexed or integrative platforms. This stepwise strategy allows incremental integration into clinical workflows, while reducing barriers to implementation and ensuring that validated biomarkers can meaningfully guide patient management.