Tumor-Type–Specific Tumor Mutational Burden Cutoffs for Improved Immune Checkpoint Inhibitor Outcome Prediction: Large-Scale Analysis with Real-World Targeted Next-Generation Sequencing Data

1. Study cohorts

A total of 9,459 patients with histologically confirmed cancer were included who underwent tumor-only targeted NGS at Asan Medical Center (AMC) between February 2021 and July 2023. Genomic profiling was conducted using the OncoPanel v4.5 (OP_v4.5), comprising 86,168 probes targeting a total of 369 genes, including complete exons from 244 genes, partial intronic regions of 14 genes frequently involved in solid tumor-associated rearrangements, and 181 regions designed for genetic fingerprinting and hotspot single-nucleotide polymorphism detection. All somatic variants underwent manual curation, were formatted as Mutation Annotation Format files, and were subsequently used for downstream analyses. The cohort spanned 27 tumor types, classified according to the OncoTree taxonomy (https://oncotree.mskcc.org). Among these, 1,017 patients received ICI treatment—primarily pembrolizumab (n=774) and atezolizumab (n=221)—with 13 patients receiving both sequentially. Based on the timing of sequencing relative to treatment, ICI-treated patients were classified into the NGS before ICI (n=860, 84.6%) or NGS after ICI (n=157, 15.4%) groups (Fig. 1A), following the clinical classification strategy previously described by Aggarwal et al. [19]. The study protocol was approved by the Institutional Review Board of AMC (2024-0248), and all data were fully de-identified prior to analysis.

Fig. 1.

Study overview and derivation of tumor-type–specific tumor mutational burden–high (TMB-H) cutoffs. (A) Overview of the study cohort and classification of the immune checkpoint inhibitor (ICI)–treated patients based on the sequencing-to-treatment intervals. (B) Distribution of the ICI-treated cases across 27 cancer types in the Asan Medical Center (AMC) cohort. Lung cancer accounted for the highest proportion (29.9%), followed by gastric/esophageal, biliary tract, and pancreatic cancers. (C) Pre-processing and stratification workflow for TMB-H classification. Synonymous mutations, likely false positives, and germline variants were removed using an internal database. Three TMB-H definitions were compared: a fixed threshold (≥ 10 mut/Mb), tumor-type–specific interquartile range (IQR)–based cutoffs, and survival-optimized thresholds. (D) Range of IQR multipliers (0.1-1.5) tested to define upper-bound TMB-H thresholds after excluding hypermutated cases (microsatellite instability–high [MSI-H] or POLE/POLD1; n=233). The 0.1×IQR multiplier was selected based on optimal concordance with survival. (E) Survival-based optimization of the TMB cutoffs in lung cancer was performed. Moreover, Kaplan-Meier (KM) analysis was performed across cutoffs ranging from 10 to 40 mut/Mb, and the cutoff resulting in the lowest survival p-value was identified as optimal. CNS, central nervous system; FDA, U.S. Food and Drug Administration; SNV, single-nucleotide variant.

2. TMB calculation and pre-processing

TMB was defined as the number of somatic non-synonymous variants per megabase (mut/Mb). The exonic region covered by OP_v4.5 was estimated to span 0.71 Mb. Prior to TMB calculation, synonymous mutations and artifacts were removed.

Germline filtering for tumor-only data was performed using a three-step pipeline.

First, a population-frequency filter removed variants with allele frequency ≥ 1% in any of gnomAD, ExAC, or the Korean Variant Archive (KOVA). Second, we excluded variants present in an internal panel-of-normals generated on the same OP_v4.5 assay. Third, we constructed a cohort-derived “germline-like” variant list from the full AMC cohort (9,459 tumor profiles) by flagging variants that simultaneously satisfied all four criteria: overall incidence ≥ 0.1%, mean variant allele frequency (VAF) 35%-60%, minimum VAF ≥ 20% across carriers, and no “oncogenic/likely oncogenic” annotation in OncoKB. Variants passing these criteria were removed from sample-level calls before TMB estimation (S1A Fig.). Additionally, variants with mean VAF ≥ 75% (with minimum VAF ≥ 20% and not annotated as oncogenic/likely oncogenic in OncoKB) were also included as germline variants. After filtering, a total of 85,469 single nucleotide variants and Indels remained for TMB estimation.

3. Derivation and validation of IQR-based TMB-H cutoffs

To define tumor type-specific TMB-H cutoffs, we first excluded cases with MSI-H status (n=205) or POLE/POLD1 mutations (n=28) within each cancer type, as their exceptionally high TMB values could disproportionately skew the distribution. For each tumor type, the TMB distribution (excluding MSI-H and POLE/POLD1 cases) was characterized by calculating the IQR. The upper cutoff was defined as Q3+k×IQR, where Q3 is the 75th percentile and k is a multiplier (tested from 0.1 to 1.5). The final k= 0.1 was selected based on concordance with survival outcomes. This IQR-based approach follows the framework proposed by Fernandez et al. [20], which highlighted the utility of distribution-adjusted cutoffs to account for tumor-specific variability in TMB. The prognostic value of these cutoffs was evaluated using Kaplan-Meier (KM) and Cox regression analyses for PFS and OS across the full cohort and clinical subgroups (pre-ICI and NGS before ICI). Finally, we compared panel-based TMB estimate with TCGA WES data (10,443 cases across 23 cancer types) as a reference, and additionally validated them against the Memorial Sloan Kettering Cancer Center (MSKCC) panel cohort (25,040 cases across five cancer types), which is based on matched tumor-normal sequencing. Both datasets were obtained from cBioPortal (https://www.cbioportal.org/).

4. Classification of TMB-H and hypermutated samples

Samples were classified as TMB-H using three methods: a universal threshold (≥ 10 mut/Mb), an IQR-based tumor-type–specific cutoff, and outcome-optimized thresholds based on the OS/PFS in ICI-treated cases. TMB-H cases were further classified into three subgroups: hypermutation, MSI-H, and ‘Others’ groups. Hypermutation was defined as either (1) TMB ≥ 50 mut/Mb with POLE/POLD1 exonuclease domain mutations or (2) TMB 25-50 mut/Mb with POLE/POLD1 mutations previously reported in hypermutated phenotype cases. MSI-H status was primarily determined using indel index-based method, which has been validated in previous tumor-only targeted NGS studies as a reliable surrogate for MSI detection [21]. MSI-H status was defined as ≥ 20 variants with an indel index ≥ 20%, or ≥ 10 variants with an indel index ≥ 40%. Samples fulfilling neither the hypermutation nor MSI-H criteria were classified into the ‘Others’ group.

5. Survival outcome definition

Due to the absence of standardized radiographic progression data, proxy-PFS definitions were used based on treatment duration and follow-up patterns. Proxy progression was defined based on ICI treatment duration and follow-up criteria. The patients were classified as having progressive disease (PD) if they survived ≥ 30 days after the final ICI dose with no further treatment. In cases without further treatment or death, censoring was applied at the last known follow-up time. Both PFS and OS were calculated from the date of first ICI administration to the date of proxy PD, death, or last follow-up, whichever came first. This proxy-PFS definition is consistent with that used by Aggarwal et al. [19]. To ensure comparability, all survival times were restricted to a maximum follow-up of 24 months.

6. Statistical analysis

All statistical analyses were performed using R ver. 4.5.0 (R Foundation for Statistical Computing). KM curves were generated for the PFS and OS, and log-rank tests were used to assess differences between groups. Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs), with adjustment for relevant baseline clinical covariates such as age, sex, and tumor histology where appropriate. Categorical variables were compared using chi-square or Fisher’s exact tests, and continuous variables using Student’s t test or Wilcoxon rank-sum test, as appropriate.

7. PD-L1 expression analysis

PD-L1 expression was assessed by IHC using the 22C3 pharmDx antibody and scored using the tumor proportion score (TPS). Among the ICI-treated patients with lung cancer having available PD-L1 data (n=488), the tumors were initially categorized into four groups: negative (0%), weak positive (1%-49%), strong positive (50%-89%), and very high (≥ 90%). For outcome analysis, the cases were dichotomized into very high (TPS ≥ 90%) versus non–very high (TPS < 90%). KM analyses were conducted to compare the OS and PFS between these two groups and evaluate the prognostic relevance of very high PD-L1 expression in ICI-treated lung cancer.

8. Subclonality assessment

Subclonality in TMB-H samples was evaluated using the clone ratio-based approach proposed by He et al. [22]. The clone ratio was defined as the VAF of each mutation divided by the maximum VAF observed in the sample. Based on the proportion of variants with a clone ratio ≤ 0.1, cases were classified as clonal (0%), low-subclonal (1 to < 20%), and high-subclonal (≥ 20%) categories. The threshold of 20% for high subclonality was set according to the upper quartile of the clone ratio distribution observed in the full AMC cohort, thereby reflecting relative subclonal diversity within the real-world population. Survival outcomes (OS and PFS) were analyzed within the TMB-H group, focusing on the prognostic implications of high subclonality. To further assess the predictive performance of the TMB status, we excluded high-subclonality samples from the ‘Others’ group and compared the OS and PFS based on the TMB status.

1. Concordance between IQR-based and survival-optimized TMB-H cutoffs

Among the 1,017 patients who received ICI treatment, bladder/urinary tract cancer accounted for the highest proportion (n=112, 37.8%), followed by lung cancer (n=569, 29.9%) and skin cancer (n=49, 4.8%) (Fig. 1B). Detailed clinical characteristics regarding the ICI-treated cases, including age at diagnosis, sex, and cancer types, are summarized in S2 Table.

To determine the appropriate TMB-H classification thresh olds, we compared three approaches: the conventional fixed cutoff of 10 mut/Mb, the tumor-type–specific IQR-based cutoffs, and the OS/PFS-optimized cutoffs. The IQR-based method, applied after excluding MSI-H and POLE/POLD1-mutated samples, demonstrated superior consistency and generalizability across the tumor types (Fig. 1C and D).

Before defining tumor-type–specific cutoffs, we first assessed the effect of germline variant filtering on tumor-only TMB estimates. Across the full cohort, the median TMB decreased from 12.7 mut/Mb to 9.87 mut/Mb after filtering (p < 0.05), and this reduction was consistently observed across multiple cancer types (S1B and S1C Fig.). This approach reduced systematic inflation of tumor-only TMB values, providing a more robust baseline for subsequent cutoff derivation.

Survival-based optimization was performed to assess the validity of the IQR-based TMB-H cutoff. Using KM and log-rank tests, we identified OS- and PFS-optimized cutoffs in both MSI-H/POLE/POLD1-excluded and MSI-H/POLE/POLD1-included cohorts. Although lung cancer was presented as a representative example (Fig. 1E), the OS- and PFS-based cutoffs were moderately concordant across most cancer types. In particular, the IQR-based cutoff closely matched the PFS-optimized threshold in bladder/urinary tract, colorectal, and uterine cancers (S3 Fig.). Based on this consistency, we selected the IQR-based threshold (0.1×IQR above the 75th percentile) as the final TMB-H definition for subsequent analyses.

2. TMB distribution across tumor types

TMB distribution was analyzed across 27 cancer types in the full cohort, regardless of the ICI treatment status. Substantial variability was observed in both median TMB values and IQR-based TMB-H thresholds across the tumor types. The highest mutational burdens were observed in the bladder/urinary tract (median, 12.7 mut/Mb; cutoff, 19.2), lung (11.3; 16.2), and bowel (12.7; 16.1) (Fig. 2A).

Fig. 2.

Pan-cancer landscape of tumor mutational burden (TMB) and validation using The Cancer Genome Atlas (TCGA) cohort. (A) Distribution of TMB across 27 cancer types within the Asan Medical Center (AMC) cohort. Black dots indicate the median TMB for each tumor type. Solid and dashed vertical lines represent TMB-high cutoffs derived using interquartile range multipliers of 0.1 and 1.5, respectively. The number of total samples, immune checkpoint inhibitor (ICI)–treated cases (atezolizumab or pembrolizumab), overall survival events, and progression-free survival (PFS) events are annotated for each cancer type. (B) Violin plots comparing panel-based TMB from the AMC cohort (targeted next-generation sequencing) with whole-exome sequencing (WES)–derived TMB from the TCGA cohort across 22 matched cancer types. Diamonds represent the median values, and the total case numbers per cancer type are displayed below each plot. (C) Correlation of the median TMB values between the AMC and TCGA cohorts. A moderate positive correlation was observed (Spearman’s ρ=0.66, p < 0.001), supporting the concordance between panel- and WES-based TMB estimates. CNS, central nervous system.

Application of the conventional fixed cutoff of 10 mut/Mb resulted in apparent overclassification of the TMB-H status in certain tumor types—particularly bladder and bowel cancers—highlighting potential overestimation (S4 Fig.). These observations supported the use of tumor-type–specific cutoffs for improved classification accuracy.

Comparisons between panel-based TMB estimates from the AMC cohort and WES-based TMB from the TCGA cohort across 23 matched tumor types revealed substantially higher TMB values in the AMC dataset (Fig. 2B). These differences may reflect both the tumor-only sequencing design, which cannot fully remove rare germline variants, and the enrichment of cancer-associated genes in the targeted panel. Despite germline filtering, residual inflation remained (median TMB reduced from 12.7 to 9.87 mut/Mb) (S1 Fig.). Accordingly, the discrepancy with TCGA values should be considered substantial. Nevertheless, a correlation remained significant (ρ=0.63, p < 0.001), supporting the relative robustness of panel-based TMB estimates.

Notably, TCGA skin cancer exhibited a higher median WES-TMB, consistent with better capture of ultraviolet-induced mutational signatures, whereas AMC uterine cancers showed elevated panel TMB due to a small number of hyper-mutated outliers. Despite these discrepancies, a moderate correlation between AMC and TCGA median TMB values was observed (Spearman’s ρ=0.66, p < 0.001) (Fig. 2C), supporting the robustness of panel-based TMB estimation for real-world clinical application.

We also compared panel-based TMB distribution in AMC (n=4,064) with the MSK-IMPACT cohort (n=25,040), which is based on matched tumor-normal sequencing. These analyses demonstrated strong concordance of median TMB values across these cancer types (Spearman’s rho=0.97, p=0.005) (S5 Fig.), supporting the robustness of our panel-derived TMB estimates across independent large-scale datasets.

3. Predictive value of TMB-H for ICI treatment outcomes

The prognostic performance of TMB-H classification was evaluated using two thresholds: the conventional fixed cutoff of 10 mut/Mb and the individualized IQR-based cutoffs. In the total ICI-treated cohort (n=1,017), IQR-defined TMB-H was significantly associated with prolonged PFS (HR, 0.85; 95% CI, 0.73 to 0.98; p=0.022). Similar associations were observed in the NGS before ICI subgroup (n=860; HR, 0.86; 95% CI, 0.73 to 1.00; p=0.049) and the pre-ICI subgroup (n=523; HR, 0.80; 95% CI, 0.66 to 0.97; p=0.026). In contrast, the fixed 10 mut/Mb cutoff showed no significant association with PFS in the overall cohort (HR, 0.92; 95% CI, 0.81 to 1.04; p=0.181) or either the NGS before ICI subgroup (HR, 0.93; 95% CI, 0.81 to 1.08; p=0.344) or pre-ICI subgroup (HR, 0.93; 95% CI, 0.78 to 1.11; p=0.455) (S6 Fig.).

Cancer-type-specific analyses were conducted for nine tumor types with ≥ 20 PFS events (Fig. 3A-C). In the pancancer setting, IQR-defined TMB-H status was significantly associated with longer PFS (HR, 0.80; 95% CI, 0.68 to 0.93; p=0.003), demonstrating consistent findings across both the NGS before ICI subgroup (HR, 0.81; 95% CI, 0.69 to 0.96; p=0.012) and pre-ICI subgroup (HR, 0.74; 95% CI, 0.60 to 0.91; p=0.008) (Fig. 3B). Moreover, significant PFS benefits were observed in bladder/urinary tract (p=0.014), colorectal (p=0.013), and uterine cancers (p=0.006), while lung cancer showed a borderline trend toward significance (p=0.076). No significant associations were observed in skin cancer (p=0.667), esophageal/gastric cancer (p=0.177), or biliary tract cancer (p=0.590) (Fig. 3C). These findings were further supported by univariate Cox regression models (Fig. 3D). Additional analyses were performed for 14 rare or less-represented cancer types with fewer than 20 PFS events, showing an overall trend toward improved PFS for IQR-defined TMB-H in aggregated and subgroup analyses, despite limited sample sizes (S7 Fig.). In contrast, for tumor types with intrinsically low TMB such as brain and thyroid cancers, the IQR-based cutoffs did not yield meaningful survival associations in ICI-treated cohorts.

Fig. 3.

Prognostic value of interquartile range (IQR)–based tumor mutational burden-high (TMB-H) across cancer types. (A) Kaplan-Meier (KM) plots for progression-free survival (PFS) comparing the IQR-based TMB-H and TMB-low (TMB-L) groups in the total immune checkpoint inhibitor (ICI)–treated cohort (n=1,017), next-generation sequencing (NGS) before ICI subgroup (n=771), and pre-ICI subgroup (n=464). Only cancer types with ≥ 20 PFS events were included. (B) Forest plot of Cox regression analyses comparing the PFS between the TMB-H and TMB-L groups, adjusted for age at diagnosis and sex, in the total, NGS before ICI, and pre-ICI cohorts across nine selected cancer types. (C) KM curves for the 24-month PFS in nine cancer types within the total cohort, comparing IQR-based TMB-H and TMB-L. (D) Forest plot of unadjusted Cox regression analyses for PFS across the same nine cancer types. CI, confidence interval.

The survival advantage of IQR-based TMB-H extended to OS, with consistent trends observed in both the total and lung cancer cohorts, and in both the NGS before ICI and preICI subgroups (S8 Fig.).

In addition, stratification by the top 20% TMB per tumor type, as previously reported by Samstein et al. [23], revealed an association with survival (log-rank p=0.091 for OS, p=0.065 for PFS) (S9 Fig.). Notably, IQR-based cutoff demonstrated clearer separation (log-rank p=0.038 for OS, p=0.035 for PFS).

4. Integrative analysis of PD-L1 expression and TMB in lung cancer

The predictive interaction between TMB and PD-L1 expression was examined in patients with lung cancer treated with ICI (n=488), using PD-L1 IHC data obtained with the 22C3 pharmDx assay (Fig. 4). Tumors were classified into four groups based on the TPS: negative (0%), weakly positive (1%-49%), strongly positive (50%-89%), and very high (≥ 90%) (S10A Fig.). For survival comparisons, the patients were stratified into a very high PD-L1 group (TPS ≥ 90%) versus a non–very high group (TPS < 90%).

Fig. 4.

Integrative analysis of tumor mutational burden (TMB) and programmed cell death-ligand 1 (PD-L1) expression in immune checkpoint inhibitor–treated lung cancer. (A) Left: Distribution of panel-based TMB across four PD-L1 tumor proportion score (TPS) groups: negative (< 1%), weak positive (1%-49%), strong positive (50%-89%), and very high (≥ 90%). Right: Binary comparison of TMB between the very high (≥ 90%) and non–very high (< 90%) PD-L1 groups; one hypermutated case was excluded. (B) Proportion of cases with very high PD-L1 expression in the TMB-high (TMB-H) versus TMB-low (TMB-L) groups. (C) Kaplan-Meier (KM) plots of the progression-free survival (PFS) stratified by TMB-H versus TMB-L and very high versus non–very high PD-L1 expression in the total lung cancer cohort. Cox regression hazard ratios (HRs) and log-rank p-values are annotated within each plot. (D) KM plots showing the PFS across four combined biomarker groups: TMB-H/very high PD-L1, TMB-H/non–very high PD-L1, TMB-L/very high PD-L1, and TMB-L/non–very high PD-L1. (E) Same analyses as in (C), performed in the pre-ICI lung cancer subgroup (n=254). (F) KM plots of the PFS across combined biomarker groups in the pre-ICI subgroup, demonstrating the strongest benefit in TMB-H/very high PD-L1 cases.

The median TMB was highest in the very high PD-L1 group (15.5 mut/Mb); however, the overall differences among the four TPS-defined groups did not reach statistical significance (ANOVA, p=0.059). A binary comparison revealed significantly elevated TMB in the very high group compared to the non–very high group (p=0.027, t test) (Fig. 4A). A hyper-mutated outlier in the PD-L1–negative group was excluded from the primary analysis; results including this sample are provided in S10B Fig. The proportion of very high PD-L1 expressers was greater in the TMB-H group than in the TMB-low (TMB-L) group (28.5% vs. 18.4%) (Fig. 4B), suggesting partial co-enrichment of high TMB and PD-L1 elevation.

In the full lung cancer cohort, both TMB-H (HR, 0.83; 95% CI, 0.67 to 1.03; p=0.085) and very high PD-L1 expression (HR, 0.73; 95% CI, 0.58 to 0.92; p=0.009) were individually associated with improved PFS (Fig. 4C). Combined biomarker analysis showed that cases with both TMB-H and very high PD-L1 expression had the most favorable PFS (HR, 0.64; 95% CI, 0.44 to 0.93; p=0.021), followed by TMB-H/non–very high, TMB-L/very high, and TMB-L/non–very high subgroups (Fig. 4D). In the pre-ICI subgroup (n=254), individual associations remained significant for both TMB-H (HR, 0.73; p=0.009) and very high PD-L1 expression (HR, 0.68; p=0.015) (Fig. 4E). The combined TMB-H/very high PD-L1 subgroup demonstrated the most prolonged PFS (HR, 0.48; 95% CI, 0.30 to 0.78; p=0.003) (Fig. 4F).

5. Prognostic impact of high subclonality in TMB-H tumors

To evaluate the influence of subclonal diversity on ICI response, TMB-H tumors were stratified into hypermutated (MSI-H or POLE/POLD1-mutated; n=41) and non-hypermutated ‘Others’ (n=242) groups (Fig. 5A). Although differences were not statistically significant (OS: p=0.310; PFS: p=0.389), the hypermutated group demonstrated a trend toward longer OS and PFS compared to the ‘Others’ group (Fig. 5B). Subclonality was further examined using clone ratio-based metrics. High-subclonality cases (≥ 20% of variants with a clone ratio ≤ 0.1) were slightly more frequent in the hypermutated group than in the ‘Others’ group (9.8% vs. 7.9%), although the small sample size limited definitive comparisons (Fig. 5C and D). Survival analysis stratified by subclonality revealed no significant prognostic impact within the hypermutated group, whereas in the ‘Others’ group, high subclonality was associated with a shorter OS, particularly in the pre-ICI subgroup (log-rank p=0.032) (Fig. 5E).

Fig. 5.

Prognostic impact of subclonality in tumor mutational burden–high (TMB-H) tumors and refinement of biomarker analysis. (A) Distribution of microsatellite instability–high (MSI-H), POLE/POLD1-mutated, and Others among the TMB-H tumors, including their proportions in the immune checkpoint inhibitor (ICI)–treated subgroup. (B) Kaplan-Meier (KM) curves for the overall survival (OS) and progression-free survival (PFS) comparing MSI-H/POLE versus others within the TMB-H group. Log-rank test p-values are shown in each plot. (C) Schematic representation of subclonality assessment based on clone ratio, defined as each variant allele frequency (VAF) divided by the maximum VAF in each sample (AF/max(AF)). (D) Comparison of subclonality groups (clonal, low-subclonal, high-subclonal) between the MSI-H/POLE and Others within TMB-H tumors. (E) KM analyses of OS and PFS classified by the subclonality (clonal+low-subclonal [LS] vs. high-subclonal [HS]) in the MSI-H/POLE and others groups, shown separately for the total and pre-ICI subgroups. (F) Heatmap (left) showing the proportion of high-subclonality cases across cancer types in the TMB-H group. The box plots (right) illustrate the distribution of subclonal fractions among the non-clonal samples (excluding fully clonal samples with subclonal ratio=0). CNS, central nervous system. (G) KM plot for the PFS outcomes in ICI-treated lung cancer after excluding high-subclonality samples from the TMB-H/others group (n=9). Improved prognostic discrimination was observed across combined TMB-H and programmed cell death-ligand 1 (PD-L1) expression categories.

Subclonality distributions across tumor types revealed marked heterogeneity. Among the TMB-H tumors, bladder/urinary tract (22.2%), colorectal (21.1%), and skin (18.2%) cancers exhibited higher proportions of high-subclonality cases (Fig. 5F). Boxplot analyses further highlighted wide intertumoral variability in clone ratios, especially in lung, bladder, and biliary tract cancers. To assess the combined predictive value of TMB-H and very high PD-L1 expression, high-subclonality samples within the ‘Others’ group (n=9) were excluded. Compared to the original analysis (Fig. 4C-F), exclusion of these cases resulted in stronger prognostic separation (log-rank p=0.021) (Fig. 5G). Consistent findings were observed in the pre-ICI subgroup (S11 Fig.), where the combination of TMB-H and very high PD-L1 expression was associated with the most favorable outcomes (p=0.013). Moreover, the exclusion of high-subclonality cases enhanced prognostic separation, especially in the combined biomarker group.