Exploratory plasma ctDNA genomic biomarkers identified by whole-exome sequencing and a novel bioinformatics pipeline in advanced driver-negative NSCLC Authors: Luis Posado-Domínguez, Álvaro López-Gutiérrez, Edel del Barco Morillo, Juan Carlos Redondo-González, Marco Hernández-Pérez, Noelia Egido-Iglesias, Ángel Canal-Alonso, Laura Corvo-Félix, Aline Rodríguez-Françoso, Lorena Bellido-Hernández, Emilio Fonseca-Sánchez, Juan Jesús Cruz-Hernández, Juan Manuel Corchado, Juan Luis García Hernández, Alejandro Olivares-Hernández Published: 2026 ABSTRACT Immune checkpoint inhibitors (ICIs) have improved outcomes in advanced non-small-cell lung cancer (NSCLC), but predictive biomarkers remain suboptimal. Blood-based tumour mutational burden (bTMB) captures part of the signal, yet its clinical performance is inconsistent. Whole-exome sequencing (WES) of plasma-derived ctDNA, analysed through an automated, AI-enabled platform (AIRGenomics), may reveal broader genomic patterns associated with response or resistance to immunotherapy. We conducted a prospective observational study of 37 patients with advanced NSCLC without EGFR-activating mutations or ALK/ROS1 rearrangements treated in first line with immunotherapy or chemo-immunotherapy. Baseline plasma ctDNA was analysed by WES and processed with the AIRGenomics platform (Nextflow-based pipeline for QC, alignment, somatic/germline calling, CNV and annotation) including an AI-based pathogenicity model. bTMB was calculated as somatic mutations/Mb. Unsupervised clustering was performed according to PD-L1 status and progression-free survival (PFS). Survival was assessed with Kaplan–Meier and Cox models. Median bTMB was 12.31 mut/Mb. Higher bTMB was associated with tumours with PD-L1 ≥ 50% and adenocarcinomas but not with overall survival (OS) or PFS and showed limited discrimination for response (AUC 0.328). Cluster analysis by PD-L1/PFS identified recurrently altered genes (including KMT2C, CEP89 and TPSB2). Univariable survival analysis revealed 11 genes associated in mutated status with worse OS and PFS; among them, CYP4F2 (OS wild-type -WT- median not reached vs. mutated 9 months; p = 0.011), ARSD (OS WT median not reached vs. mutated 13 months; p = 0.017) and TPSB2 (OS WT 24 months vs. mutated 1,5 months; p = 0.007) were selected for multivariable modelling. In the Cox model, CYP4F2 (HR = 2,846; IC 95%: 1,102–7,352; p = 0,031) and TPSB2 (HR = 3,089; IC95%: 1,053–9,060; p = 0,040) were independently biomarkers associated with shorter OS (χ² =13,128; p = 0.004), and CYP4F2 (HR = 3,167; IC95%: 1,384–7,244; p = 0,006) remained an independent predictor of shorter PFS (χ² =11.116; p = 0.011). This proof-of-concept study demonstrates that WES of ctDNA processed through the AIRGenomics platform is viable in real-world cases of advanced NSCLC treated with immunotherapy, detecting new potential candidate genes and pathways like predictive biomarkers, such as CYP4F2, ARSD, and TPSB2.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-55868-0. Introduction Lung cancer is the malignancy with the highest incidence and mortality worldwide, with approximately 2.4 million new diagnoses and 1.8 million deaths per year. Non-small-cell lung cancer (NSCLC) accounts for around 80% of all lung cancers1. Over the past decade, the natural history of NSCLC has been transformed by the introduction of immunotherapy with immune checkpoint inhibitors (ICI) and targeted agents, which have, for example, enabled long-term survival rates of 20–30% even in the metastatic setting2,3. However, it is important to recognise that a substantial proportion of patients do not derive benefit from ICI, either initially or during treatment, and ultimately have a prognosis comparable to that observed in the pre-ICI era4–6. Immunotherapy, in the form of ICI, has been incorporated across all treatment settings in NSCLC, from the adjuvant and neoadjuvant contexts to advanced disease. The observed response rates and the proportion of long-term survivors have led to ICI becoming the standard of care for these tumours7. Nevertheless, despite favourable outcomes in a sizeable subgroup of patients, there remains a population that does not respond to ICI and experiences progression or even hyperprogression, with very limited short-term survival8. The reported frequency of lack of response is highly variable, with hyperprogression rates ranging, for example, from 4% to 20%9,10. Multiple biomarkers are currently under investigation in an attempt to predict ICI response in NSCLC. These include genomic biomarkers (such as STK11 or KEAP1), microbiome alterations, and epigenetic changes in genes regulating the immune system or the tumour microenvironment11. However, results have often been contradictory and difficult to extrapolate to routine clinical practice12. Among the emerging predictive biomarkers with potential clinical application, genomic markers are arguably the most promising, owing to technological advances in molecular and bioinformatic analysis13. For example, biomarkers that capture tumour genomic heterogeneity are highly attractive for predicting the clinical benefit of ICI in patients with NSCLC. Increased chromosomal instability (CIN) and intratumoural heterogeneity (ITH) have been associated with tumour immune escape and greater immune evasion14. However, the predictive potential of these biomarkers has not yet been validated. In addition, several studies have already evaluated global circulating tumour DNA (ctDNA) levels as a marker of response to ICI, based on changes in ctDNA burden over time15. Despite this strong theoretical rationale, very few studies have assessed the exome in ctDNA, largely because of its complexity, which lies more in the bioinformatic than in the molecular analysis. Using advanced bioinformatic tools, including artificial intelligence (AI)-based approaches, it is now possible not only to extract and interrogate whole-exome sequencing (WES) data, but also to integrate clinical and molecular information in order to identify and characterise ctDNA-based genomic biomarkers predictive of response to ICI16,17. The aim of the present study arises from this rationale: to apply bioinformatic tools to the clinico-molecular analysis of ctDNA in patients with advanced NSCLC treated with ICI. Methods Design, population and sample This is an observational, multidisciplinary and prospective study conducted at the Medical Oncology Department of the University Hospital of Salamanca (CAUSA, Salamanca, Spain), BISITE group of the University of Salamanca and the Cancer Research Center of Salamanca (CIC). The patient cohort was established through these hospitals’ discharge records collected between 1 Decembre 2022 and 31 June 2025. Primary and secondary objectives The primary objective of the study is to identify and characterise predictive genomic biomarkers, together with their functional pathways, associated with response to first-line ICI in advanced NSCLC without actionable driver mutations, through the application of bioinformatic tools to ctDNA whole-exome sequencing (WES) obtained from liquid biopsy. The secondary objectives are1: to define a ctDNA-based genetic signature capable of predicting response to ICI in advanced NSCLC, and2 to demonstrate the clinical utility of liquid biopsy by applying this approach in a real-world cohort of patients with advanced NSCLC treated with first-line ICI. Recruitment, follow-up and variables of the study The inclusion criteria for the study were as follows: Age ≥ 18 years and ability to understand and sign informed consent for the use of clinical and molecular data.Diagnosis of advanced non-small-cell lung cancer with no prior ICI treatment.Absence of EGFR-activating mutations or ALK/ROS1 rearrangements. Absence of EGFR-activating mutations or ALK/ROS1 rearrangements. Other genomic alterations, including MET exon 14 skipping mutations, KRAS G12C mutations, or ERBB2 exon 20 insertions, were not considered exclusion criteria, as no standard targeted therapies were systematically used in the first-line setting for these alterations in this cohort.First-line treatment with ICI or chemo-ICI in the advanced setting, in accordance with clinical guidelines.Peripheral blood sampling for ctDNA liquid biopsy prior to initiation of first-line treatment in the advanced setting.Availability of an electronic medical record containing all information required for statistical analysis. Age ≥ 18 years and ability to understand and sign informed consent for the use of clinical and molecular data. Age ≥ 18 years and ability to understand and sign informed consent for the use of clinical and molecular data. Diagnosis of advanced non-small-cell lung cancer with no prior ICI treatment. Diagnosis of advanced non-small-cell lung cancer with no prior ICI treatment. Absence of EGFR-activating mutations or ALK/ROS1 rearrangements. Absence of EGFR-activating mutations or ALK/ROS1 rearrangements. Other genomic alterations, including MET exon 14 skipping mutations, KRAS G12C mutations, or ERBB2 exon 20 insertions, were not considered exclusion criteria, as no standard targeted therapies were systematically used in the first-line setting for these alterations in this cohort. Absence of EGFR-activating mutations or ALK/ROS1 rearrangements. Absence of EGFR-activating mutations or ALK/ROS1 rearrangements. Other genomic alterations, including MET exon 14 skipping mutations, KRAS G12C mutations, or ERBB2 exon 20 insertions, were not considered exclusion criteria, as no standard targeted therapies were systematically used in the first-line setting for these alterations in this cohort. First-line treatment with ICI or chemo-ICI in the advanced setting, in accordance with clinical guidelines. First-line treatment with ICI or chemo-ICI in the advanced setting, in accordance with clinical guidelines. Peripheral blood sampling for ctDNA liquid biopsy prior to initiation of first-line treatment in the advanced setting. Peripheral blood sampling for ctDNA liquid biopsy prior to initiation of first-line treatment in the advanced setting. Availability of an electronic medical record containing all information required for statistical analysis. Availability of an electronic medical record containing all information required for statistical analysis. The exclusion criteria for the study were as follows: Lack of patient understanding of the objectives and implications of the study.Previous treatment with ICI in any setting.History of a prior cancer requiring systemic therapy.Presence of EGFR-activating mutations or ALK/ROS1 rearrangements. Lack of patient understanding of the objectives and implications of the study. Lack of patient understanding of the objectives and implications of the study. Previous treatment with ICI in any setting. Previous treatment with ICI in any setting. History of a prior cancer requiring systemic therapy. History of a prior cancer requiring systemic therapy. Presence of EGFR-activating mutations or ALK/ROS1 rearrangements. Presence of EGFR-activating mutations or ALK/ROS1 rearrangements. Peripheral blood sampling for ctDNA analysis performed after initiation of ICI. Clinical and demographic variables were collected from the electronic medical records of patients included by each patient’s clinicians. The clinical variables collected were age, sex, risk factors (smoking with a pack-years index > 10 and alcoholism), tumour histology, tumour stage, NGS molecular analysis results, treatment received, presence of brain metastases, treatment dose, toxicity, toxicity grade (CTCAE v5.0), overall survival (OS) in months, progression-free survival (PFS) in months, response (RECIST1.1) and ORR. OS was defined as the time in months from the start of treatment to the death or loss of follow-up of the patient. PFS was defined as the time in months from the start of treatment to progression or loss to follow-up. Toxicity was measured using the Common Terminology Criteria for Adverse Events v.5.0 classification. Response was defined using the Response Evaluation Criteria in Solid Tumours v1.1 criteria. ORR was defined as the sum of complete and partial responses. Extraction of ctDNA by liquid biopsy A total of 37 patients with advanced NSCLC without actionable driver mutations who were due to receive first-line treatment with ICI or chemo-ICI were recruited. Prior to the initiation of systemic treatment in the advanced setting, ctDNA was obtained by liquid biopsy (peripheral blood). In most patients, blood samples were collected on the same day as treatment initiation, within a window ranging from day − 10 to day 0. Two 10 mL EDTA blood tubes were drawn. Plasma was obtained by centrifugation of whole blood for 10 min at 1,900 × g, followed by a second centrifugation for 10 min at 18,000 × g in the presence of EDTA and proteinase K. Plasma samples were then processed using the QIAamp Circulating Nucleic Acid Kit (55114, Qiagen, Valencia, CA). On average, 20–80 ng of ctDNA were isolated from 4 to 5 mL of plasma. The extracted samples were stored at − 80 °C until further use. Genomic and bioinformatic analysis of the ctDNA by WES WES was used to identify variants in plasma-derived ctDNA isolated from the cohort with advanced NSCLC. Exome library preparation for sequencing workflow utilized the SureSelect Human All Exon V6 library capture kit (Agilent Technologies, CA, USA) in a NovaSeq6000 (Illumina, San Diego, CA, USA), using paired-end reads (150 × 2). The average sequencing output was approximately 18 Gb per sample. Library preparations quantified and quantity of individual sequencing libraries was was performed at Macrogen Inc. (Seoul, South Korea). Sequencing data was processed using the AirGenomics platform18. The end-to-end pipeline covers quality control, read alignment, variant calling, annotation, and clinical prioritization. Raw FASTQ files were subjected to an initial quality assessment to evaluate base quality, coverage depth, and potential contamination. High- quality reads were aligned to the human reference genome (hg19) using optimized algorithms for somatic variant detection in low-frequency contexts. Variant calling was performed for single nucleotide variants (SNVs), small insertions/deletions (indels) and structural variants (SVs), followed by annotation through curated databases (e.g. ClinVar, COSMIC, gnomAD, dbSNP). Germline variants were filtered based on genotype quality (GQ ≥ 20), allele frequency (AF ≥ 0.45 for heterozygous, AF ≥ 0.9 for homozygous), and standard quality metrics (QD ≥ 2.0; QUAL ≥ 30.0; MQ ≥ 40.0; SOR ≤ 3.0; FS ≤ 60.0), with additional CNN-based refinement. Somatic variants required a PASS filter status, base quality Q ≥ 10 and support from at least half of the callers used, excluding non-callable regions and low-quality calls. Structural variants were detected using complementary algorithms with adaptive filtering and annotated to retain high-confidence calls. Copy number variants (CNVs) were called using specific thresholds: log₂ ratio ≤ − 1.1 (copy number 0), ≤ − 0.25 (copy number 1), ≥ 0.2 (copy number 2), and ≥ 0.7 (copy number 3). At the final stage of the workflow, somatic variants were evaluated through a custom pathogenicity prediction tool18. This model employs a Feature Tokenizer Transformer architecture and a semi-supervised learning approach to derive a quantitative pathogenicity score, which was used as a complementary filter to refine variant prioritization and support the interpretation of clinically actionable mutations. Prioritization was performed in accordance with ACMG/AMP guidelines, considering allele frequency, predicted pathogenicity, and clinical relevance. Clinic-molecular statistical analysis First, in order to describe the molecular characteristics of the cohort and facilitate biomarker discovery, four differentiation clusters were defined. These were: cluster 1, poor response to ICI and PD-L1 ≥ 50%; cluster 2, poor response to ICI and PD-L1 < 50%; cluster 3, good response to ICI and PD-L1 < 50%; and cluster 4, good response to ICI and PD-L1 ≥ 50%. Subsequently, for the univariable and multivariable survival analyses according to the genomic characteristics of ctDNA, clinical and molecular data from the bioinformatic study were integrated. Two response groups to first-line ICI or chemo-ICI in advanced NSCLC were defined on the basis of progression-free survival (PFS). A cut-off of 9 months PFS was used to subdivide patients, in line with published first-line ICI studies in advanced NSCLC19. This subdivision into two groups was chosen in accordance with the available sample size. Statistical analysis was performed by cross-referencing molecular data with clinical data from the patients included. The software used was SPSS v.28 (IBM®). Survival and correlation analysis was performed using Kaplan-Meier (Log-rank test and Breslow test) and Chi-squared test (95% confidence intervals “CIs”). Breslow test was utilised in these curves with cross-linking. Hazard Ratio (HR) was stablished by Cox Regression. Missing data were handled with variable-by-variable exclusion, and the total number of cases analysed was reported in each table. No imputation methods were used due to the limited extent of missing data. Statistical significance was p < 0.05. Role of the funding source and ethic The study was conducted in accordance with the Declaration of Helsinki and approved by the local Ethics Committee at the University Hospital of Salamanca (Salamanca, Spain) (protocol code PI 2022 10 1155 and date of approval 29/11/2022). Written informed consent for sample collection and subsequent research use was obtained from all participants and/or their legal guardians. None other external funding supported the study. Results General characteristics of the sample A total of 37 patients were included (75.7% male and 24.3% female). The majority were smokers (33/37, 89.2%). The most frequent histology was adenocarcinoma (27/37, 73%), followed by squamous cell carcinoma (10/37, 27%). Overall, 19 patients (51.4%) had tumours with PD-L1 expression < 50%, 17 patients (45.9%) had PD-L1 expression ≥ 50%, and PD-L1 status was unknown in 1 patient. The most frequent molecular alteration identified was in KRAS (27%), with the G12C subtype accounting for 10.8%. The next most frequent alterations were in TP53 (13.5%) and MET (8.1%). Clinically, central nervous system involvement was present in 8 patients (21.6%). Seventeen patients (40.5%) received ICI monotherapy, whereas 20 patients (59.5%) were treated with chemo-ICI combinations. In all cases, treatment consisted of pembrolizumab, either alone or in combination with platinum-based doublet chemotherapy. A total of 8 patients (21.6%) experienced immune-related toxicity, which was grade > 3 in 3 patients (8.1%). Median overall survival (OS) for the whole cohort was 20 months (95% CI 17.0–23.0), and median progression-free survival (PFS) was 9 months (95% CI 7.8–10.2). In patients with high PD-L1 expression (≥ 50%), median OS was 22 months (95% CI 18.0–25.0) and median PFS was 9.5 months (95% CI 8.0–10.8). In patients with low PD-L1 expression (< 50%), median OS was 15.5 months (95% CI 12.0–18.5) and median PFS was 8.5 months (95% CI 7.0–9.5). The overall response rate (ORR) for the entire cohort was 59.5% (56.8% partial response and 2.7% complete response). Among patients with high PD-L1 expression, ORR was 64.7%, compared with 57.9% in those with low PD-L1 expression (Table 1). Table 1General characteristics of the sample.PatientAgeSexSmokerHistologyDriver mutationsCNS mtsPD-L1CT-IO/IO (1 L)Toxicity grade ≥ 3RECIST1.1OSPFSCLUSTER159MYesADC-NO90IONOPR991275MYesADC-NO0CT-IONOPR992361MYesADC-NO10CT-IONOPR29293472MYesADC-NO70IONOPR50234574FYesADC-NO80CT-IONOPR2281676MYesUD-YES75IONOCR28284759MYesADC-NO0CT-IONOPR2892886MYesADCMET Ex14YES80IONOPD221957MNoADCERBB2 Ex20insYES10CT-IONOPR28731073FYesADC-NO0CT-IOYESSD281531178MNoSQ-NO2CT-IOYESPR24821273MYesADCKRAS G12CYES50I-ONOPD2211360FYesSQ-NO50IONOSD10811467MYesADC-NO10CT-IONOPR272731579MYesSQ-NOUknownCT-IONOSD12821663FYesADC-NO80IONOPD25411772MYesSQ-NO15CT-IONOPR22821869MYesADC-NO50IONOPR565641956MYesSQ-NO15CT-IONOPD3322073MYesUD-NO5CT-IONOPD6322173MYesSQ-NO0CT-IONOSD12822262FYesADC-NO20CT-IONOSD242432377FYesADCKRAS G12CYES5CT-IONOPD5422476MYesSQ-NO70IONOPR22612566FYesADCMET Ex14YES60IONOPR221042679MYesSQKRAS G12CNO60IONOPR191942765MYesADC-NO75IONOPR222242884MYesSQ-NO75IONOPR222242973MYesADC-NO0CT-IONOPR222033081FNoADCERBB2 Ex20insNO60IONOPD13213165MYesADC-YES20CT-IONOSD13823282MYesADC-NO20CT-IONOPD6323373MYesADC-NO50CT-IONOPD2213457MYesADC-YES80IOYESPR202043574FNoSQMET Ex14NO20CT-IONOPR151433650MYesADCKRAS G12CNO0CT-IONOPR161633775MYesADC-NO5CT-IONOPR14143*Abbreviations: CNS mts, Central Nervous System Metastases; CT-IO, chemo-immunotherapy; IO, immunotherapy; 1 L, first-line treatment; M, male; F, female; ADC adenocarcinoma, UD, undifferentiated; SQ, squamous; CR, complete response; PR, partial response; SD, stable disease; PD, progression disease. NA, not available. Table 1General characteristics of the sample.PatientAgeSexSmokerHistologyDriver mutationsCNS mtsPD-L1CT-IO/IO (1 L)Toxicity grade ≥ 3RECIST1.1OSPFSCLUSTER159MYesADC-NO90IONOPR991275MYesADC-NO0CT-IONOPR992361MYesADC-NO10CT-IONOPR29293472MYesADC-NO70IONOPR50234574FYesADC-NO80CT-IONOPR2281676MYesUD-YES75IONOCR28284759MYesADC-NO0CT-IONOPR2892886MYesADCMET Ex14YES80IONOPD221957MNoADCERBB2 Ex20insYES10CT-IONOPR28731073FYesADC-NO0CT-IOYESSD281531178MNoSQ-NO2CT-IOYESPR24821273MYesADCKRAS G12CYES50I-ONOPD2211360FYesSQ-NO50IONOSD10811467MYesADC-NO10CT-IONOPR272731579MYesSQ-NOUknownCT-IONOSD12821663FYesADC-NO80IONOPD25411772MYesSQ-NO15CT-IONOPR22821869MYesADC-NO50IONOPR565641956MYesSQ-NO15CT-IONOPD3322073MYesUD-NO5CT-IONOPD6322173MYesSQ-NO0CT-IONOSD12822262FYesADC-NO20CT-IONOSD242432377FYesADCKRAS G12CYES5CT-IONOPD5422476MYesSQ-NO70IONOPR22612566FYesADCMET Ex14YES60IONOPR221042679MYesSQKRAS G12CNO60IONOPR191942765MYesADC-NO75IONOPR222242884MYesSQ-NO75IONOPR222242973MYesADC-NO0CT-IONOPR222033081FNoADCERBB2 Ex20insNO60IONOPD13213165MYesADC-YES20CT-IONOSD13823282MYesADC-NO20CT-IONOPD6323373MYesADC-NO50CT-IONOPD2213457MYesADC-YES80IOYESPR202043574FNoSQMET Ex14NO20CT-IONOPR151433650MYesADCKRAS G12CNO0CT-IONOPR161633775MYesADC-NO5CT-IONOPR14143*Abbreviations: CNS mts, Central Nervous System Metastases; CT-IO, chemo-immunotherapy; IO, immunotherapy; 1 L, first-line treatment; M, male; F, female; ADC adenocarcinoma, UD, undifferentiated; SQ, squamous; CR, complete response; PR, partial response; SD, stable disease; PD, progression disease. NA, not available. General characteristics of the sample. *Abbreviations: CNS mts, Central Nervous System Metastases; CT-IO, chemo-immunotherapy; IO, immunotherapy; 1 L, first-line treatment; M, male; F, female; ADC adenocarcinoma, UD, undifferentiated; SQ, squamous; CR, complete response; PR, partial response; SD, stable disease; PD, progression disease. NA, not available. Molecular and genomic characteristics of the sample When the cohort was subdivided into four differentiation clusters based on ICI response and PD-L1 expression, a total of seven genes with variants present at a frequency greater than 50% were identified as statistically significantly associated with the different clusters (Fisher’s exact test). These genes were KMT2C, CEP89, ANKRD36C, SORD, MST1, ZDHHC11 and TPSB2. The absence of alterations in KMT2C and CEP89 was associated with a favourable response to ICI in terms of PFS in patients with high PD-L1 expression. Conversely, alterations in TPSB2 were significantly associated with cluster 1, characterised by lack of response to ICI in tumours with high PD-L1 expression (Table 2 and SM1). However, after adjustment for clinico-genomic covariates, none of these variants retained statistical significance following false discovery rate (FDR) correction with a p-value threshold of < 0.05. Table 2Distribution of ARSD, CYP4F2 and TPSB2 by clusters. Table 2Distribution of ARSD, CYP4F2 and TPSB2 by clusters. Distribution of ARSD, CYP4F2 and TPSB2 by clusters. Analysis of copy number variations (CNVs) according to ICI response showed that, in cluster 1, 67% of tumours harboured gains at 5p13.2 and deletions at 15q25.3. In cluster 2, the most frequent CNVs (72.7%) were gains at 21p11.1 and 8q21.2. In cluster 3, 78% of tumours exhibited gains at 19p13.12 and 20q11. Finally, in cluster 4, 63% of tumours showed gains at 20q11.21. No statistically significant differences in CNVs were observed between clusters (p > 0.05). We also examined actionable genes in NSCLC and genes associated with predisposition or resistance to ICI in the ctDNA samples. No EGFR mutations or ALK/ROS1 fusions were detected in ctDNA. A single STK11 mutation was identified in a patient with PD-L1 ≥ 50%, who exhibited an unfavourable outcome on pembrolizumab monotherapy (OS 9 months). Analysis of liquid biopsy-calculated TMB The median liquid biopsy–calculated tumour mutational burden (bTMB) in the cohort was 12.31 mutations/Mb (range 5.97–71.29). Patients alive at last follow-up had a significantly higher bTMB compared with those who had died (18.9 vs. 11.3 mutations/Mb; p = 0.041). The median bTMB in cluster 1 (poor response to ICI and PD-L1 ≥ 50%) was approximately double that observed in the other clusters (15.66 vs. 8.29 in cluster 2, 9.57 in cluster 3 and 7.48 in cluster 4). No statistically significant differences were observed between clusters (p > 0.05) (Fig. 1). Fig. 1bTMB by clusters (PD-L1/PFS). bTMB by clusters (PD-L1/PFS). With regard to histological subtype, squamous tumours had a median bTMB of 7.11 (6.37–31.21), compared with 12.32 (5.97–71.29) in adenocarcinomas. In tumours with PD-L1 ≥ 50%, the median bTMB was 12.44 (6.68–71.29) versus 8.30 (5.97–45.78) in tumours with PD-L1 < 50%. In our cohort, bTMB showed low discriminatory power for predicting response to ICI in NSCLC (AUC = 0.328; 95% CI: 0.150–0.506; p = 0.075), suggesting that bTMB alone is not a robust predictive biomarker. Univariable survival analysis of genomic biomarkers A total of 11 genes were found to be associated, in univariable analysis, with resistance to first-line ICI in advanced NSCLC. These genes were CCZ1B, ARSD, NP1PB5, PABPC3, TRIM49, TBC1D28, KRT40, APOB, CYP4F2, RGPD4 and TPSB2. All of these genes were significantly associated with worse OS and PFS. Of the 11 genes associated with survival, the only one previously identified at cluster level as differentiating response between groups was TPSB2. With respect to OS, the most pronounced differences were observed for ARSD p.(Phe223Ile), which was mutated in 15 of 37 patients (40.54%); CYP4F2 p.(Leu519Met), mutated in 11 of 37 patients (29.7%); and TPSB2 p.(Val76Leu), mutated in 6 of 37 patients (16.2%). Median OS in ARSD wild-type -WT- tumours was not reached, compared with 13 months in ARSD-mutated cases (p = 0.017). For CYP4F2, median OS was not reached in WT tumours and 9 months in mutated tumours (p = 0.011). For TPSB2, median OS was 24 months in WT versus 1.5 months in mutated tumours (p = 0.007) (Fig. 2). Fig. 2Kaplan-Meier curves of overall survival by the main genes and CNVs alterations (ARSD, CYP4F2, TPSB2 and 21p11.1). Kaplan-Meier curves of overall survival by the main genes and CNVs alterations (ARSD, CYP4F2, TPSB2 and 21p11.1). Similarly, PFS analysis for ARSD, CYP4F2 and TPSB2 also demonstrated unfavourable outcomes in the presence of the mutations. Median PFS was not reached in ARSD WT tumours versus 8 months in ARSD-mutated cases (p = 0.016). For CYP4F2, median PFS was not reached in WT versus 3 months in mutated tumours (p = 0.004). For TPSB2, median PFS was 9 months in WT versus 1,5 months in mutated tumours (p = 0.003, Breslow test) (Fig. 3). Fig. 3Kaplan-Meier curves of progression-free survival by the main genes and CNVs alterations (ARSD, CYP4F2, TPSB2 and 21p11.1). Kaplan-Meier curves of progression-free survival by the main genes and CNVs alterations (ARSD, CYP4F2, TPSB2 and 21p11.1). Among the CNVs analysed, only gains at 21p11.1, which were more frequent in cluster 2 (poor response to ICI in tumours with PD-L1 < 50%), were associated with survival. Tumours harbouring CNVs at 21p11.1 had a median OS of 13 months, compared with not reached in tumours without these alterations (HR = 4.32; p = 0.010). No statistically significant differences in PFS were observed (8 vs. 10 months, respectively; p = 0.609). Multivariable analysis of genomic biomarkers In the multivariable Cox model, which included the three variants showing the best performance in the univariable survival analysis (TPSB2, CYP4F2 and ARSD), overall model fit was statistically significant for both OS (χ² = 13.128; p = 0.004) and PFS (χ² = 11.116; p = 0.011). For OS, the presence of CYP4F2 alterations was associated with an increased risk of death (HR = 2.846; 95% CI: 1.102–7.352; p = 0.031), as were alterations in TPSB2 (HR = 3.089; 95% CI: 1.053–9.060; p = 0.040). However, alterations in ARSD did not reach statistical significance (HR = 2.609; 95% CI: 0.997–6.827; p = 0.051). For PFS, only CYP4F2 mutations emerged as an independent predictive biomarker of survival (HR = 3.167; 95% CI: 1.384–7.244; p = 0.006), whereas TPSB2 (HR = 2.073; 95% CI: 0.693–6.201; p = 0.192) and ARSD (HR = 1.707; 95% CI: 0.752–3.877; p = 0.201) did not reach statistical significance (Table 3). Table 3Multivariable Cox regression model including CYP4F2, TPSB2 and ARSD, which showed the most robust results in univariable survival analysis.GeneHR (OS)IC95% (OS)p (OS)HR (PFS)IC95% (PFS)p (PFS) CYP4F2 2,8461,102–7,3520,0313,1671,384–7,2440,006 TPSB2 3,0891,053–9,0600,0402,0730,693–6,2010,192 ARSD 2,6090,997–6,8270,0511,7070,752–3,8770,201 Table 3Multivariable Cox regression model including CYP4F2, TPSB2 and ARSD, which showed the most robust results in univariable survival analysis.GeneHR (OS)IC95% (OS)p (OS)HR (PFS)IC95% (PFS)p (PFS) CYP4F2 2,8461,102–7,3520,0313,1671,384–7,2440,006 TPSB2 3,0891,053–9,0600,0402,0730,693–6,2010,192 ARSD 2,6090,997–6,8270,0511,7070,752–3,8770,201 Multivariable Cox regression model including CYP4F2, TPSB2 and ARSD, which showed the most robust results in univariable survival analysis. Discussion In this cohort of 37 patients with advanced NSCLC without actionable driver mutations treated with first-line chemo-ICI or ICI, we explored the value of baseline ctDNA WES as a tool to identify genomic biomarkers predictive of response or resistance to ICI. The results should be interpreted as exploratory, given that the predictive value of individual variants is limited; rather, the study demonstrates the feasibility of this approach and generates biologically plausible signals that warrant further investigation as potential modulators of ICI response. First, we observed that a higher bTMB was not significantly associated with improved OS or PFS, although tumours with PD-L1 ≥ 50% and adenocarcinomas tended to show higher bTMB values. The discriminatory ability of bTMB to distinguish responders from non-responders was low (AUC 0.328), suggesting that bTMB alone is not a robust biomarker in this setting. The published literature on this topic is contradictory: some studies indicate that bTMB is an independent biomarker of benefit from ICI in advanced NSCLC, such as the phase 2 B-F1RST trial by Kim ES et al.20, whereas others, including the randomised phase 3 BFAST cohort C trial did not demonstrate that bTMB could predict response to atezolizumab in advanced NSCLC21. Our results are consistent with previous reports demonstrating that tumour mutational burden alone does not uniformly predict response to immunotherapy and should be interpreted within the broader biological and immune context of each tumour, as previously described by McGrail et al.22. Taken together, these findings support the notion that bTMB should be interpreted within multivariable models rather than as a standalone marker. Second, bioinformatic analysis of ctDNA by WES enabled the identification of genomic patterns associated with response and resistance beyond bTMB. The cluster analysis based on PD-L1 and PFS identified a group of recurrently altered genes (KMT2C, CEP89, ANKRD36C, SORD, MST1, ZDHHC11, TPSB2) with variant frequencies above 50% and differential associations across clusters. Notably, several of these alterations were observed across multiple clusters, which limits their specificity and supports their interpretation as exploratory, hypothesis-generating signals rather than definitive predictive biomarkers. The absence of alterations in KMT2C and CEP89 was associated with better PFS in PD-L1-high tumours, whereas TPSB2 mutations were concentrated in the poor-response cluster with PD-L1 ≥ 50%. This suggests that specific genomic profiles may modulate treatment outcome even in “PD-L1-high” tumours, which would be expected to be more sensitive to ICI. Beyond the cluster analysis, univariable survival analysis identified 11 genes associated with worse OS and PFS, including ARSD, CYP4F2 and TPSB2, with marked reductions in median OS and PFS among variant carriers. It should be noted that CYP4F2 is known to be a relatively polymorphic gene, raising the possibility that some of the detected variants may reflect population-level variability rather than tumour-specific alterations23. In the multivariable Cox model incorporating these three candidate variants, CYP4F2 and TPSB2 were independently associated with poorer OS, whereas only CYP4F2 retained an independent association with worse PFS. Although the cluster analysis was corrected for FDR, the survival analyses were not formally adjusted for multiple testing; these associations should therefore be considered exploratory and hypothesis-generating, rather than validated biomarkers. From a biological perspective, several of these signals are plausible and potentially relevant for future clinical application. The three genes that emerge most prominently in our analysis—CYP4F2, ARSD and TPSB2—can be framed within a shared biological axis centred on innate–stromal modulation of the tumour microenvironment. CYP4F2 catalyses the formation of 20-HETE, and the CYP4F2–20-HETE–GPR75 axis has been shown to promote an immunosuppressive stromal phenotype in NSCLC, characterised by increased PD-L1 expression, IL-6/TGF-β production and angiogenesis, together with reduced sensitivity to anti-PD-1 therapy21. ARSD has been described in glioma as a modulator of the tumour microenvironment that activates the JAK2/STAT3 pathway and favours infiltration by M2/TAM macrophages, thereby contributing to local immunosuppression24. TPSB2 encodes a mast cell–specific tryptase; mast cell tryptases have been implicated in angiogenesis, extracellular matrix remodelling and regulation of immune cell infiltration in several tumour types, and infiltration by tryptase-positive mast cells has been associated with worse prognosis in multiple settings25. Although the available evidence arises from different models and tumour types, collectively it suggests that alterations in CYP4F2, ARSD and TPSB2 may mark a TME subtype characterised by STAT3 activation, 20-HETE production and mast cell activity, with a prominent angiogenic and immunosuppressive component—features that are consistent with the reduced efficacy of ICI observed in our variant carriers. The distribution of genomic alterations observed in this cohort differs from that reported in large tissue-based series, particularly regarding TP53 mutations, which are typically more frequent than KRAS alterations in lung adenocarcinoma. This discrepancy may be explained by the small sample size and the use of plasma-derived ctDNA, which may not fully capture the tumour genomic landscape due to sensitivity limitations and biological variability in ctDNA shedding. Although STK11 and KEAP1 alterations were not prevalent in our cohort, their established biological and clinical relevance warrants consideration. These alterations have been consistently associated with primary resistance to immune checkpoint inhibitors and with an “immune-desert” tumour phenotype, characterised by reduced immune infiltration and limited anti-tumour immune activity26,27. Thus, this study proposes biologically plausible genes and functional pathways (lipid metabolism via CYP4F2/20-HETE, the sulfatase ARSD and the tryptase TPSB2) identified in a real-world setting using bioinformatic analysis of ctDNA WES in a clinically relevant context. The analytical platform employed allowed integration of quality control, variant calling, CNV assessment and annotation, as well as the application of an artificial intelligence–based pathogenicity prediction model. While the data do not support the immediate use of these genes as individual biomarkers for treatment selection in first-line advanced NSCLC, they do suggest that incorporating genomic information beyond TMB could enhance prediction of ICI benefit. This study has several important limitations. First, the sample size of 37 patients, although clinically homogeneous, limits statistical power. Second, the analysis was performed on plasma without matched tumour tissue, which raises the possibility that some variants classified as somatic may in fact be germline or arise from clonal haematopoiesis, despite filtering by population frequency, allelic fraction and quality metrics. The FDR correction applied at the exome-wide level reduces the risk of false positives but also diminishes our ability to detect moderate effect sizes; the fact that no variant remained significant after FDR correction reinforces the exploratory nature of this work. In addition, ctDNA-based approaches have inherent sensitivity limitations, which depend on tumour shedding and ctDNA fraction; in low-shedding tumours, relevant genomic alterations may not be detected, potentially leading to an underestimation of tumour heterogeneity. Finally, there are no subsequent follow-up samples analysed after response or progression on immunotherapy, so it is not possible to determine tumour dynamics. Overall, this genomic analysis serves as a proof of concept for a new bioinformatics tool for ctDNA analysis with artificial intelligence applications. Conclusions Our study shows that whole-exome sequencing of ctDNA using the AIRGenomics platform, with an automated pipeline and AI-assisted variant interpretation, is feasible in clinical practice and can provide information beyond PD-L1 and bTMB in advanced NSCLC treated with first-line ICI. Although we identified several genomic pathways involving CYP4F2, ARSD, and TPSB2 as potential biomarkers of poorer outcomes, linked to immunometabolic processes and the tumour microenvironment, no single variant remained robustly significant after adjustment for multiple testing and consideration of the plasma-only study design. Therefore, these findings should be considered hypothesis-generating and require independent validation in prospective studies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1