Skip to main content
NCBI home page
Nature Communications logoLink to Nature Communications
. 2026 Aug 31;17:10348. doi: 10.1038/s41467-026-77375-6

Clinical attribute-based risk matrix for prognostication in immunotherapy-treated non-small cell lung cancer

Lu Wang 1,2,#, Nan Xu 1,#, Hairui Wang 3,#, Bosinan Chen 4,#, Qiuju Sheng 5,#, Jie Zhang 6,#, Zhaoshuo Liu 7, Taixue An 8, Guanchao Ye 9, Jianing Sun 10, Zhijian Zhang 11, Qiuyang Hou 12, Bin Liu 13, Song Chen 14,✉, Li Yu 15,✉, Jiangdian Song 1,✉
PMCID: PMC13623942  PMID: 42811057

Abstract

Biomarkers used to guide immunotherapy in non-small cell lung cancer (NSCLC) are invasive, variably available, and insufficient to capture clinical and regional heterogeneity, leaving many patients inadequately stratified for prognosis. We develop a clinical attribute–based framework that leverages routinely collected variables to estimate survival risk, even when baseline data are incomplete. We analyze 17 051 patients with NSCLC treated with immunotherapy, assembled from published cohorts and tertiary centers across multiple regions, and evaluate seventeen common clinical attributes using a two-stage survival modeling strategy with external validation in Western and East Asian populations. Clinical attributes exhibit context-dependent prognostic effects across variable-defined strata and survival endpoints, with performance status and disease stage emerging as the most consistent determinants of survival risk prediction, alongside region-specific modifiers. Here we show that this “start-anywhere” clinical risk matrix provides robust, population-aware prognostic stratification and complements biomarkers for assessing immunotherapy outcomes in patients with NSCLC.

Subject terms: Prognostic markers, Immunotherapy, Cancer immunotherapy, Non-small-cell lung cancer

Current biomarkers used to guide immunotherapy in non-small cell lung cancer (NSCLC) do not adequately stratify patients for prognosis. Here, the authors show that CLARITY-ICI Matrix, a clinically deployable, missing-data-tolerant framework, enables population-aware stratification of survival risk in immunotherapy-treated NSCLC using routinely available clinical attributes and complements biomarkers.

Introduction

Immune checkpoint inhibitors (ICIs) have transformed the therapeutic landscape for non-small cell lung cancer (NSCLC), with patients having high programmed death-ligand 1 (PD-L1) levels (≥50%) or tumor mutation burden (TMB, >10 mut/Mb) obtaining high survival benefit1. However, these molecular markers leave a clinically heterogeneous proportion of the population unclassified2–4. Next-generation genomic, transcriptomic, and spatial microenvironmental biomarkers remain investigational and are constrained by tissue requirements, turnaround time, and cost, limiting their application outside specialized medical centers5,6. In daily clinical practice, clinicians often rely on routinely available clinical information; however, most multivariable scores lose utility when one or more baseline variables are missing, which is common in the real world.

Increasing evidence links common clinical attributes to ICI activity7,8. Demography (age9, sex10, race/region11, smoking status12), functional status (Eastern Cooperative Oncology Group [ECOG] performance status13), tumor burden (clinical stage14, longest tumor diameter15), treatment history (prior treatment history9, treatment line16, and treatment regimen16), and laboratory or pathological indices (histological subtype17, fraction of genome altered [FGA]18, tumor purity19, neutrophil-to-lymphocyte ratio [NLR]9,20, PD-L1 level21, and TMB22) have shown potential association with progression-free survival (PFS) or overall survival (OS) in previous studies. However, conventional modelling frameworks typically rely on machine learning or deep learning–based prediction models23,24; clinicians must either impute absent values, which potentially risks calibration drift, or abandon risk estimation25.

In this work, we aim to develop a clinical attribute-based risk estimation matrix for immune checkpoint inhibitors in NSCLC (CLARITY-ICI Matrix) even when baseline data are incomplete. Toward this goal, we compile a large individual-patient immunotherapy dataset for NSCLC and devise the CLARITY-ICI Matrix with three objectives: to systematically quantify the prognostic value of 17 routinely captured variables on PFS and OS in 17,051 patients; to remain operable when any subset of baseline attributes is missing, thereby reflecting real-world clinical constraints; and to provide transparent, evidence-based counselling for prognostic risk stratification of immunotherapy-treated patients across different geographic populations. We describe here the construction, external validation, and practical application of the matrix and explore the context-dependent heterogeneity of common clinical determinants to predict ICI benefits in patients with NSCLC.

Results

Patient population

A total of 17,051 individual patients were eligible for inclusion. The study flowchart is shown in Fig. 1. The dataset comprised 14 998 patients from 84 previously published cohorts and 2 053 patients from seven tertiary hospitals in mainland China. Flowcharts detailing study-level and patient-level inclusion and exclusion procedures for both data sources are provided in Supplementary Figs. S1, S2. The median patient age was 65 years (interquartile range [IQR], 57–72 years). Among patients with available sex information, 59.5% (9645/16,198) of patients were male, and 40.6% (6553/16 198) were female. The study population included 5 706 (33.5%) East Asian patients, 1339 (7.9%) European patients, 9949 (58.3%) North American patients, and 57 (0.3%) patients from New Zealand. PFS data were available for 10 130 patients, with a median PFS of 5.4 months (IQR, 2.0–12.1 months). OS data were available for 13,539 patients, with a median OS of 15.2 months (IQR, 6.7–26.7 months). Baseline demographic, clinical, and molecular characteristics are summarized in Supplementary Table S1.

Fig. 1. Flowchart of the overall study design and analytical workflow.

Fig. 1

Open in a new tab

P_Treat previous treatment history, IO_Type treatment regimen, FGA fraction genome altered, Purity tumor purity, NLR neutrophil-to-lymphocyte ratio, RWD: data source (real-world or not), and TMB tumor mutation burden.

For each clinical variable evaluated as an index stratifier, Stage 1 Kaplan–Meier curves for PFS and OS and the corresponding Stage 2 multivariable Cox regression results are presented in Supplementary Figs. S3–S46. Results for TMB are shown in the main Figs. 2–4. These analyses were visualized consistently across the Discovery cohort, External Test 1 cohort (Western population), and External Test 2 cohort (East Asian population). We performed survival analyses for strata that met established guidelines for multivariable analysis26 and the sample size requirements by the PASS estimation. Based on the prognostic performance of each variable across the endpoints and cohorts, the CLARITY-ICI Matrix was subsequently constructed.

Fig. 3. Kaplan–Meier curves of progression-free survival and overall survival by tumor mutation burden (TMB) strata, along with the multivariable Cox regression analysis of the 16 Stage 2 variables within each stratum in the External Test 1 cohort.

Fig. 3

Open in a new tab

For progression-free survival, patients in the TMB High group (n = 959 patients) and TMB Low group (n = 1 339 patients) were included. For overall survival, patients in the TMB High group (n = 2 058 patients) and TMB Low group (n = 4 112 patients) were included. For Kaplan-Meier survival curves, P values were calculated using a two-sided log-rank test. For boxplots, P values were derived from two-sided multivariable Cox proportional hazards regression models. No adjustment for multiple comparisons was performed. Source data are provided as a Source Data file. Stage 2 variables: IO_Type = treatment regimen (Mono monotherapy, Chemo immunotherapy combined with chemotherapy, and Other: immunotherapy combined with other therapy); FGA fraction genome altered; LTD longest tumor diameter; Patho pathological type (LUAD adenocarcinoma and LUSC squamous-cell carcinoma), RWD data source from real world = Yes or No, EAS East Asian, NAM North American, EUR European, Purity Tumor purity, NLR neutrophil-to-lymphocyte ratio; P_treat = prior treatment history. Data are presented as median values with interquartile range. Box plots show the median (center line), interquartile range (box limits), and whiskers extending to 1.5 × the interquartile range.

Fig. 2. Kaplan–Meier curves of progression-free survival and overall survival by tumor mutation burden (TMB) strata, along with the multivariable Cox regression analysis of the 16 Stage 2 variables within each stratum in the Discovery cohort.

Fig. 2

Open in a new tab

For progression-free survival, patients in the TMB High group (n = 2338 patients) and TMB Low group (n = 1778 patients) were included. For overall survival, patients in the TMB High group (n = 2 306 patients) and TMB Low group (n = 2265 patients) were included. For Kaplan-Meier survival curves, P values were calculated using a two-sided log-rank test. For boxplots, P values were derived from two-sided multivariable Cox proportional hazards regression models. No adjustment for multiple comparisons was performed. Source data are provided as a Source Data file. Stage 2 variables: IO_Type = treatment regimen (Mono monotherapy, Chemo immunotherapy combined with chemotherapy, and Other: immunotherapy combined with other therapy); FGA fraction genome altered, LTD longest tumor diameter; Patho pathological type (LUAD adenocarcinoma and LUSC squamous-cell carcinoma); RWD data source from real world = Yes or No; EAS: East Asian; NAM North American, EUR European; Purity = Tumor purity; NLR = neutrophil-to-lymphocyte ratio; P_treat = prior treatment history. Data are presented as median values with interquartile range. Box plots show the median (center line), interquartile range (box limits), and whiskers extending to 1.5 × the interquartile range.

Fig. 4. Kaplan–Meier curves of progression-free survival and overall survival by tumor mutation burden (TMB) strata, along with the multivariable Cox regression analysis of the 16 Stage 2 variables within each stratum in the External Test 2 cohort.

Fig. 4

Open in a new tab

For progression-free survival, patients in the TMB High group (n = 1293 patients) and TMB Low group (n = 2 401 patients) were included. For overall survival, patients in the TMB High group (n = 1426 patients) and TMB Low group (n = 1126 patients) were included. For Kaplan-Meier survival curves, P values were calculated using a two-sided log-rank test. For boxplots, P values were derived from two-sided multivariable Cox proportional hazards regression models. No adjustment for multiple comparisons was performed. Source data are provided as a Source Data file. Stage 2 variables: IO_Type = treatment regimen (Mono monotherapy, Chemo immunotherapy combined with chemotherapy, and Other: immunotherapy combined with other therapy), FGA fraction genome altered, LTD longest tumor diameter, Patho pathological type (LUAD adenocarcinoma and LUSC: squamous-cell carcinoma), RWD data source from real world = Yes or No, EAS East Asian; NAM North American; EUR European, Purity Tumor purity, NLR neutrophil-to-lymphocyte ratio, P_treat prior treatment history. Data are presented as median values with interquartile range. Box plots show the median (center line), interquartile range (box limits), and whiskers extending to 1.5 × the interquartile range.

All CLARITY-ICI Matrix

Across the Discovery cohort and two external validation cohorts, evaluation of the 17 routinely collected clinical variables in Stage 1 yielded a total of 31 index strata for construction of the All CLARITY-ICI Matrix. These strata defined clinically interpretable entry points under the “start-anywhere” framework. Specifically, age, ECOG performance status, treatment regimen, treatment line, prior therapy history, histological subtype, data source, sex, smoking status, clinical stage, and TMB were retained as index stratifiers.

The number of strata retained in the All CLARITY-ICI Matrix varied substantially across index stratifiers. Strata defined by age, treatment line, prior treatment history, sex, and smoking status, together with their corresponding Stage 2 variables, constituted the core risk stratification matrix cells for both short-term and long-term survival outcomes, whereas other index stratifiers, such as treatment regimen and data source, contributed only a subset of retained strata.

As shown in Fig. 5A, multivariable Cox regression analyses in Stage 2 demonstrated marked context dependence, whereby the prognostic relevance of individual clinical variables varied across index-defined strata. In line with this pattern, prognostic stratification for PFS and OS was not uniformly concordant across index strata in the All CLARITY-ICI Matrix. Notably, ECOG performance status contributed to PFS and OS risk stratification across 24 index-defined strata in all three cohorts, indicating a broad and stable prognostic role. By contrast, clinical stage primarily informed OS risk stratification across 16 index-defined strata, whereas its contribution to PFS risk stratification was limited to two strata (TMB = Low and Sex = Male).

Fig. 5. The CLARITY-ICI Matrices for survival risk stratification constructed using the two-stage framework.

Fig. 5

Open in a new tab

(A), (B), and (C) denote the All, Western, and East Asian CLARITY-ICI Matrix, respectively. Source data are provided as a Source Data file. IO_Type: treatment regimen (Mono: immunotherapy monotherapy, Chemo: immunotherapy combined with chemotherapy, and Other: immunotherapy combined with other treatment), P_Treat: previous treatment history, NLR neutrophil-to-lymphocyte ratio, EAS East Asian, NAM North American, EUR European, Patho pathological type (LUAD adenocarcinoma and LUSC squamous cell lung carcinoma), and TMB tumor mutation burden.

Western CLARITY-ICI Matrix

Using the same two-stage analytic framework, we constructed a Western-specific CLARITY-ICI Matrix based on the Discovery cohort and External Test 1 cohort. Stage 1 evaluation of the 17 routinely collected clinical variables yielded 38 index strata retained in the Western CLARITY-ICI Matrix (Fig. 5B).

In the Western CLARITY-ICI Matrix, tumor purity was additionally retained as a population-specific index stratifier beyond those shared with the All CLARITY-ICI Matrix. The number of retained strata varied markedly across index attributes. Strata defined by age, treatment line, prior treatment history, sex, smoking status, clinical stage, and TMB formed the core units of survival risk stratification for both short- and long-term outcomes, whereas other attributes, such as treatment regimen and data source, contributed fewer retained strata.

Stage 2 multivariable analyses demonstrated similar context-dependent patterns (Fig. 5B). ECOG performance status remained the most stable determinant of survival risk, contributing to both PFS and OS stratification across a wide range of index strata. Clinical stage primarily informed OS risk, contributing across 19 of 21 strata, while its contribution to PFS risk was limited to two strata. Among other prognostic variables, TMB and histological subtype contributed to OS risk stratification in five strata, whereas smoking status and TMB contributed to PFS stratification within the subgroup of patients younger than 65 years.

East Asian CLARITY-ICI Matrix

Using the same analytic framework, we constructed an East Asian–specific CLARITY-ICI Matrix based on the Discovery cohort and External Test 2 cohort. Stage 1 evaluation yielded 43 index strata retained in the East Asian CLARITY-ICI Matrix (Fig. 5C).

In this matrix, longest tumor diameter, NLR, and PD-L1 expression were retained as population-specific index stratifiers. As in the All and Western matrices, the number of retained strata varied substantially across index attributes. Strata defined by age, longest tumor diameter, treatment line, prior treatment history, histological subtype, sex, and smoking status formed the core units of survival risk stratification for both short- and long-term outcomes, whereas NLR and treatment regimen contributed a more limited set of strata.

Stage 2 analyses again demonstrated context-dependent patterns (Fig. 5C). ECOG performance status remained a stable determinant of survival risk, contributing to both PFS and OS stratification across multiple strata. Clinical stage primarily informed OS risk, contributing across 20 of 23 strata, while its contribution to PFS stratification was restricted to three strata. Distinct from the Western matrix, data source, prior treatment history, and treatment regimen emerged as Stage 2 prognostic factors in the East Asian population, each contributing to survival risk stratification within specific strata. For example, within the stratum defined by longest tumor diameter >50 mm, treatment regimen informed OS risk: immunotherapy monotherapy was associated with the least favorable prognosis (median OS, 9.5 months), followed by immunotherapy combined with chemotherapy (median OS, 15.8 months), whereas other combination regimens showed the most favorable outcomes (median OS, 19.1 months; P < 0.05 in both the Discovery and External Test 2 cohorts).

Sensitivity analyses and robustness assessment

When alternative cohort partitions using a 2:1:1 ratio for the Discovery, External Test 1, and External Test 2 cohorts were applied, sensitivity analyses consistently retained ECOG performance status and clinical stage as key Stage 2 prognostic factors in the All CLARITY-ICI Matrix. Core variables for risk stratification identified in the primary analysis were preserved in both the Western and East Asian matrices (Supplementary Fig. S47).

Variance inflation factor analysis identified marked collinearity between prior treatment history and treatment line (Supplementary Table S2). After removal of prior treatment history, sensitivity analyses showed that treatment line was retained as a significant Stage 2 prognostic factor across the revised matrices (Supplementary Fig. S48). In addition, after removal of PD-L1 and TMB, the All CLARITY-ICI Matrix retained a highly similar structure, with ECOG performance status and clinical stage remaining the dominant prognostic factors and no meaningful changes in risk stratification or survival estimates compared with the main analysis (Supplementary Fig. S49). Compared with complete-case analyses, multivariate imputation by chained equation (MICE)-based matrices retained all prognostic variables contributing to survival risk stratification across populations, except for P_Treat in the Western cohort (Supplementary Fig. S50), supporting the validity of the imputation strategy.

Restricted cubic spline analyses showed no significant non-linearity for five of seven continuous variables (Supplementary Fig. S51 and Supplementary Table S3). Although PD-L1 expression showed statistically significant non-linearity, its spline curve demonstrated a monotonic association consistent with clinical expectations, with increasing PD-L1 levels associated with lower hazard; the spline-derived reference point (50%) aligned with commonly used clinical thresholds. In contrast, TMB exhibited a non-monotonic pattern. Based on spline analyses for OS and PFS, a three-level TMB categorization (<5, 5–15, >15) was evaluated. Matrices constructed using this three-level TMB definition showed patterns nearly identical to the primary analysis when TMB was used as a Stage 1 index stratifier (Supplementary Fig. S52). However, when TMB functioned as a Stage 2 variable, survival patterns were not strictly monotonic in the Western and East Asian matrices, with OS in the low-TMB group exceeding that of the intermediate group in the Western Matrix (Fig. S52B) and PFS in the intermediate group exceeding that of the high-TMB group in the East Asian Matrix (Fig. S52C).

Propensity score matching analyses showed that “region” exhibited the highest baseline imbalance for both OS and PFS (Supplementary Tables S4, S5). After matching, covariate balance improved substantially, and Kaplan–Meier curves for OS and PFS remained highly consistent with the primary analyses (Figs. S53, S54), demonstrating the robustness of the survival patterns captured by the CLARITY-ICI Matrix.

Finally, benchmark comparisons showed that the proposed framework achieved C-indices ranging from 0.65 to 0.71 (Supplementary Table S6) across the three cohorts, comparable to the recently published LORIS model in NSCLC immunotherapy9, while requiring approximately half the number of input variables. Time-dependent ROC analyses (Supplementary Table S7) further showed consistent performance between the two approaches, with both demonstrating improved predictive accuracy over longer follow-up durations. In addition, using 2018 as the cutoff, patients treated with immunotherapy before versus after this time point showed significant differences in survival across populations from North America, Europe, and East Asia (all P  <  0.05; Supplementary Table S8 and Fig. S55).

Discussion

This study introduces the CLARITY-ICI Matrix, a clinically deployable, missing data-tolerant framework that integrates 17 routinely collected variables to support prognostic risk stratification for immunotherapy in patients with NSCLC. By analyzing 17,051 individual patients, we addressed three persistent challenges in the field: (1) the limited ability of existing multivariable prognostic tools to remain operable in the presence of incomplete data; (2) the insufficient characterization of regional heterogeneity in immunotherapy-related prognostic risk stratification across distinct NSCLC populations; and (3) the lack of a clinically interpretable framework that transparently explains how and why individual clinical attributes should inform prognostic counselling.

Using a unified two-stage analytic framework, we constructed an All CLARITY-ICI Matrix that integrates prognostic risk features mainly from North American, European, and East Asian NSCLC populations, alongside Western- and East Asian–specific CLARITY-ICI Matrices tailored to population-specific characteristics. Across these matrices, this study delineated both the index stratifiers shared across different regional populations and those that were population-specific. Importantly, the number of retained strata varied markedly across index attributes, with survival risk stratification effects confined to specific strata and endpoints, while the same variables remained neutral in others. This pattern underscores the heterogeneous stratification capacity of commonly used clinical variables in the context of immunotherapy27,28. These findings show that reliance on individual prognostic variables to predict both short-term treatment response and long-term survival should shift toward more granular, population-aware, and endpoint-specific risk stratification strategies.

Multivariable Cox regression analyses identified ECOG performance status and clinical stage as core prognostic determinants across all matrices. However, further examination revealed distinct roles for these variables: ECOG performance status functioned as a global determinant of both short-term disease control (PFS) and long-term survival (OS), whereas clinical stage primarily informed long-term OS risk stratification, with limited relevance for PFS. These observations are concordant with prior reports13,28–30. Notably, the East Asian CLARITY-ICI Matrix captured population-specific prognostic structures that differed from those observed in Western cohorts. For histological subtype, the Western CLARITY-ICI Matrix indicated the most favorable OS risk stratification in patients with lung adenocarcinoma, followed by those with LUSC and other pathological types, whereas the East Asian CLARITY-ICI Matrix revealed an inverse pattern for PFS in patients collected from the real world. This divergence highlights that the prognostic implications of clinical variables such as histology are population- and endpoint specific, likely reflecting differences in disease biology, treatment patterns, and post-progression management across regions.

In addition, while previous studies suggested poorer outcomes for real-world European NSCLC patients compared with non–real-world cohorts31, our study extends this observation by demonstrating a similar pattern in East Asian populations, with a nearly two-month difference in median PFS (7.0 months vs 8.9 months, P < 0.05). Moreover, the prognostic role of smoking status has remained controversial in immunotherapy studies32–35. Our results show that its impact is context-dependent and modified by clinical factors such as functional status, disease stage, and regions. The CLARITY-ICI Matrix clarifies this heterogeneity by allowing smoking status to contribute to risk stratification only in settings where its prognostic signal is consistently supported, while avoiding overgeneralization.

Sensitivity analyses further reinforced the robustness of the framework. After removing strongly collinear variables (prior treatment history) or clinically less accessible biomarkers (PD-L1 and TMB), the core architecture of the CLARITY-ICI Matrix remained largely unchanged, while the prognostic relevance of other clinically established variables became more prominent36. Despite relying on a maximum of three routinely collected clinical variables within each stratum, the CLARITY-ICI Matrix demonstrated stable prognostic performance across multiple independent validation cohorts. When benchmarked against the contemporary LORIS model, which incorporates six or five clinical variables9, the CLARITY-ICI Matrix achieved comparable performance in terms of C-index and time-dependent area under the receiver operating characteristic curve. Although the CLARITY-ICI Matrix is not intended to propose a biomarker or guide treatment selection, these comparisons highlight its methodological strengths and practical utility for survival outcomes stratification.

The CLARITY-ICI Matrix was designed for daily clinical use, analogous to familiar TNM staging tables. In practice, clinicians may use available clinical attributes as the index variable (Stage 1) on the vertical axis, identify the patient’s corresponding stratum, and then scan horizontally across the available attributes (Stage 2) to obtain the corresponding survival risk stratification estimates. Because all variables are processed using identical classification criteria, results are consistent regardless of the chosen starting point. To facilitate clinical adoption, we developed an interactive online tool (http://www.ciitool.com/#/clarityICI), and a demonstration video of the CLARITY-ICI Matrix (Supplementary Video 1).

Although prior systematic reviews and meta-analyses have examined selected prognostic variables for immunotherapy, no previous study has comprehensively evaluated commonly used clinical attributes at the patient level in a cohort exceeding 17 000 NSCLC patients, nor produced a population-specific, missing data–tolerant framework for survival risk stratification34,35,37,38. Propensity score–matched analyses demonstrated that region exhibited the greatest baseline imbalance for both OS and PFS, supporting the necessity of region-specific external validation and the construction of distinct Western and East Asian matrices. Given that differences between PFS and OS may be driven in part by post-progression treatment rather than intrinsic prognosis39,40, we evaluated these endpoints independently. Although the “start-anywhere” property allows clinicians to estimate risk using available information, not all routinely collected variables demonstrated prognostic value. For example, despite interest in FGA, its limited availability in current practice suggests that its clinical utility for immunotherapy requires further evaluation.

The strengths of this study include its large sample size, racial and geographic diversity, and a systematic two-stage approach that filtered out contradictory or non-reproducible signals. Nevertheless, this study has limitations. More granular classification of treatment regimens may further clarify survival heterogeneity arising from regional practice differences. In particular, the relatively low proportion of Western patients receiving immunotherapy combined with chemotherapy may partly explain why it did not emerge as a prognostic factor in Western cohorts. Future studies focusing on Western populations treated with combination regimens are warranted to further evaluate this association. In addition, given that TMB measurement is influenced by heterogeneity in sequencing platforms, gene panels, and bioinformatic pipelines41,42, the sensitivity analyses show that the prognostic implications of TMB categorization should be interpreted with caution to reduce potential heterogeneity-driven inconsistencies in prognostic assessment. The CLARITY-ICI Matrix developed in this study enables prognostic risk stratification for patients with NSCLC receiving immunotherapy; however, comparative evidence with other treatment modalities, such as chemotherapy and targeted therapy, remains limited compared with the latest study43. Future work will therefore seek to extend this framework and further explore its potential role in supporting clinical decision-making. Finally, prospective validation is a critical prerequisite for the future clinical application of the CLARITY-ICI Matrix. We have registered a prospective clinical study (researchregistry11749) and plan to collect immunotherapy-treated NSCLC patients over two years to evaluate the prognostic risk stratification performance of the CLARITY-ICI Matrix.

In conclusion, the CLARITY-ICI Matrix provides a pragmatic and statistically validated approach for stratifying survival risk in immunotherapy-treated NSCLC using routinely available clinical attributes, complementing PD-L1, TMB, and emerging biomarkers. By explicitly delineating regional heterogeneity and endpoint-specific prognostic evidence for individual variables, the framework facilitates transparent patient counselling and may inform trial stratification and health-economic modelling. Prospective evaluation in diverse populations and integration into clinical workflows are the next steps toward maximizing translational impact.

Methods

Ethics approval

The retrospective collection and analysis of de-identified clinical data were approved by the Institutional Review Boards of the participating institutions: The First Affiliated Hospital of China Medical University (AF-SOP-07-1), The First Affiliated Hospital of University of Science and Technology of China (2023-RE-113), Guangdong Provincial People’s Hospital (KY2025-473-02), Nanfang Hospital of Southern Medical University (NFEC-2024-423), Shengjing Hospital of China Medical University (2026PS026K), Zhuhai People’s Hospital (2025-KT-145), and The First Affiliated Hospital of Zhengzhou University (2024-KY-2262-002). China Medical University approved the data integration and CLARITY-ICI model construction using de-identified datasets and the data collected from the other 84 cohorts (CMU-2025-126). This study was conducted in accordance with the 1964 Declaration of Helsinki. The requirement for written informed consent was waived because all data were retrospectively collected and de-identified.

Study design and data sources

A pooled, individual-patient data analysis integrating two data streams was conducted: (1) publicly available immunotherapy cohorts identified through a systematic search of PubMed, Web of Science, The Cancer Imaging Archive, and cBioPortal databases; and (2) an institutional data series collected from the above tertiary hospitals in mainland China. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for observational studies. For detailed patient numbers and sex composition per cohort, see Source_Data_Study_Sex_summary_Table S1.

Systematic search and external cohort selection

PubMed and Web of Science were searched from database inception to December 1, 2025 using the following strategy “(non-small-cell lung cancer OR NSCLC) AND (checkpoint inhibitor OR pembrolizumab OR nivolumab OR atezolizumab OR durvalumab OR ipilimumab OR tremelimumab OR immunotherapy)” without language restrictions. The Cancer Imaging Archive and cBioPortal databases were screened in parallel.

A study was included if (1) a downloadable spreadsheet (xlsx, csv, pdf, or txt) containing patient-level entries was available and (2) the file included survival data (follow-up time and censoring status) along with ≥1 of 17 prespecified baseline variables (see below). All previously published studies, whether prospective or retrospective, that meet the above criteria were included. Datasets were excluded if (1) only aggregate statistics were reported, (2) patient identifiers overlapped with those of another included cohort, (3) clinical variables with outliers may affect survival analysis, or (4) patient-level files had restricted access. The reference lists of the eligible articles were also manually searched.

Patient enrollment

Patients from seven Chinese tertiary centers were eligible if they had histologically confirmed NSCLC, received at least one cycle of ICI monotherapy or combination therapy, and had evaluable PFS or OS. Data collection for patients receiving immunotherapy spans from January 1, 2014 to December 31, 2024. Data were extracted retrospectively from electronic medical records using standardized case report forms. Patients were excluded if all the required baseline variables were missing or if survival endpoints could not be determined. All treatments for the patients included in this study were performed in accordance with clinical guidelines, and the decision to continue, discontinue, or terminate immunotherapy was made with the patient’s consent.

Harmonized baseline variables

Seventeen attributes were standardized across sources. These were as follows: demography and functional status: (1) age (<65 years vs ≥65 years)37, (2) sex, (3) region (North American, European, East Asian, or Other), (4) smoking status (ever/current vs never)32, (5) ECOG score (0 vs ≥1)16, and (6) data source (retrospective real-world data vs prospective non–real-world data); tumor burden: (7) clinical stage (I–III vs IV, American Joint Committee on Cancer 7th or 8th edition) and (8) longest tumor diameter (≤50 mm vs >50 mm44; centrally measured where image was available); treatment history including (9) therapy before ICI (yes vs no)9, (10) planned treatment regimen (monotherapy [Mono], combined with chemotherapy [Chemo], or combined with other therapies [Other]), and (11) treatment line (1 L vs ≥2 L); and pathology/laboratory: (12) histology (adenocarcinoma [LUAD], squamous-cell carcinoma [LUSC], and Others), (13) FGA (<0.20 vs ≥0.20)18, (14) tumor purity (<50% vs ≥50%)45, (15) NLR (low vs high, dichotomized at pooled median)20, (16) PD-L1 expression (<50% vs ≥50%)21, and (17) TMB (≤10 vs >10 mut/Mb)22.

The variables were extracted directly from the source files and re-coded according to the definitions above. Thus, the final pooled dataset comprised data of a regionally diverse, individual-level cohort suitable for the systematic evaluation of the determinants of survival risk in patients with NSCLC who received immunotherapy.

Endpoints

PFS was measured from ICI initiation to radiographic or clinical progression or death, whichever occurred first. OS was measured from the start of treatment until any-cause death. Patients who were alive without disease progression were censored at the last follow-up (institutional series, May 1, 2025; external cohorts per original report).

Discovery and validation cohorts

To ensure that dataset partitioning balanced regional population heterogeneity, sample size comparability, and the rigor of independent validation, we randomly divided all data into a Discovery cohort, External Test 1 cohort, and External Test 2 cohort in an approximate 1:1:1 ratio. The Discovery cohort included patients from multiple geographic regions, whereas the External Test 1 cohort comprised exclusively Western populations and the External Test 2 cohort comprised exclusively East Asian populations. This design enabled direct assessment of regional differences in clinical characteristics and survival risk stratification, while also supporting the construction of region-specific CLARITY-ICI Matrices. To demonstrate the robustness of the proposed framework, we additionally performed multiple sensitivity analyses using alternative 2:1:1 partitions based on the same rules. Across all partitions, data from any individual study were assigned to only one cohort to prevent data leakage. The complete two-stage pipeline remained unchanged across all cohorts. Reproducibility of variable effects was assessed by verifying that statistical significance and the direction of effects for each endpoint were consistent across the cohorts.

Two-stage analytic framework

Stage 1: Index-attribute stratification. Each of the 17 clinical variables was sequentially treated as an index stratifier. Patients were partitioned according to the predefined categories of the index attribute (as mentioned above), and Kaplan–Meier curves with log-rank tests were used to assess global differences in PFS and OS across strata.

Stage 2: Within-stratum multivariable prognostic risk evaluation. For each stratum in Stage 1, the remaining clinical variables were simultaneously entered into multivariable Cox proportional hazards regression models to identify independent prognostic factors for PFS and OS, which were evaluated as separate endpoints. Hazard ratios (HRs) with 95% confidence intervals (CIs) and Wald P values were estimated while accounting for inter-variable dependence. An attribute was retained within a given stratum if it demonstrated concordant statistical significance (P < 0.05 with a consistent direction of effect) for the same endpoint across the relevant cohorts, according to the matrix being constructed (All, Western, or East Asian). Attributes that failed to meet this concordance criterion were excluded. This multivariable, endpoint-specific, multi-cohort concordance strategy reduced confounding from collinearity, treatment sequencing, and regional practice patterns, while preserving the interpretability and flexibility of the CLARITY-ICI Matrix.

To further characterize regional heterogeneity in immunotherapy-related prognostic risk stratification, we constructed multiple CLARITY-ICI Matrices based on different population groupings. For the All CLARITY-ICI Matrix, an attribute was retained if it demonstrated concordant statistical significance for the given endpoint within the same stratum across the Discovery, External Test 1, and External Test 2 cohorts; for the Western CLARITY-ICI Matrix, retention required concordant significance across the Discovery and External Test 1 cohorts; and for the East Asian CLARITY-ICI Matrix, retention required concordant significance across the Discovery and External Test 2 cohorts.

Based on this two-stage procedure, the final CLARITY-ICI Matrices provide stratified, region-specific, and evidence-based prognostic risk estimates for immunotherapy-treated patients with NSCLC, while remaining operable under conditions of incomplete baseline clinical variables. A detailed build process is shown in Fig. 1.

Statistics & Reproducibility

Continuous variables are summarized as medians with interquartile ranges (IQRs), and categorical variables are presented as counts with percentages. Survival outcomes, including PFS and OS, were estimated using the Kaplan–Meier method and compared using log-rank tests. The proportional hazards assumption was assessed using Schoenfeld residuals. Model discrimination was quantified using Harrell’s concordance index (C-index) with corresponding 95% CI. Sample size was not prospectively determined because this study was based on retrospectively collected and publicly available datasets. All eligible patients meeting the predefined inclusion criteria were included in the analysis. Sample-size adequacy was assessed using PASS v15.0 (NCSS, USA). In addition, multivariable Cox regression analyses were performed in accordance with established methodological recommendations, with up to 17 variables included and a total sample size of 17,051 patients, ensuring an adequate sample-to-parameter ratio and sufficient statistical power across all analyses26. No additional data exclusions were performed after cohort assembly, except for the predefined eligibility criteria and variable-specific missing-data handling described in the Methods. Patients were subsequently assigned to the Discovery, External Test 1, and External Test 2 datasets according to the predefined study design. Because this was a retrospective observational study based on existing datasets, no experimental randomization or investigator blinding was performed.

To evaluate potential collinearity, all 17 clinical variables were assessed using variance inflation factors. Variables exhibiting marked collinearity were removed, and the remaining variables were re-entered into the CLARITY-ICI Matrix for sensitivity analyses.

Missing baseline variables were handled using Multiple Imputation by Chained Equations (MICE), and the imputed data were used for the primary analyses. Sensitivity analyses using complete-case data were conducted by restricting analyses to patients with complete information for all Stage 2 variables, thereby eliminating the influence of multiple imputation. Results from complete-case analyses were compared with those obtained from MICE-imputed data to assess consistency in retained prognostic variables.

Restricted cubic spline analyses were used to assess potential non-linear effects of continuous variables, including age, longest tumor diameter, NLR, FGA, tumor purity, TMB, and PD-L1 expression. Sensitivity analyses with alternative variable discretization strategies were also performed. In addition, PD-L1 expression and TMB were excluded as an ablation analysis, and the CLARITY-ICI Matrix was reconstructed using the remaining 15 routinely available clinical variables to assess the robustness of survival risk stratification in the absence of these biomarkers.

For cohort-level sensitivity analyses, studies were re-partitioned at the study level using an alternative 2:1:1 ratio. The complete two-stage analytic pipeline remained unchanged across all partitions, and reproducibility was assessed by verifying consistency in statistical significance and direction of effects across cohorts.

For regional sensitivity analyses, population-specific CLARITY-ICI Matrices were constructed for Western and East Asian cohorts in parallel with the All CLARITY-ICI Matrix. Attributes were retained based on endpoint-specific, multi-cohort concordance criteria appropriate to each matrix.

In addition, propensity score matching was applied to further evaluate the robustness of the CLARITY-ICI Matrix. The clinical variable with the lowest missing rate was selected as the exposure variable, with the remaining 16 variables included as covariates. One-to-one propensity score matching was performed in the full cohort. Propensity scores were estimated using multivariable logistic regression, and nearest-neighbor matching without replacement was applied with a caliper width of 0.05.

Furthermore, within each index stratum defined by the CLARITY-ICI Matrix, we constructed prognostic models using the retained clinical variables identified by multivariable Cox regression analysis. Model performance was assessed by calculating Harrell’s C-indices with corresponding 95% CIs. To further evaluate predictive stability over time, we additionally computed time-dependent area under the receiver operating characteristic curve values for OS at 0.5, 1, 2, and 3 years. All statistical analyses were conducted using R software (version 4.3.1; R Foundation for Statistical Computing, Vienna, Austria). All tests were two-sided, and P values < 0.05 were considered statistically significant.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Supplementary information

Supplementary Information (6.5MB, pdf)
41467_2026_77375_MOESM2_ESM.pdf (113.8KB, pdf)

Description of Additional Supplementary Files

Supplementary Video 1 (22.1MB, mp4)
Reporting Summary (5MB, pdf)
Transparent Peer Review file (5.4MB, pdf)

Source data

Source Data (5MB, zip)

Acknowledgements

J.D.S. discloses support for the research of this work from the National Natural Science Foundation of China [Grant number 92259104], the Scientific Research Fund of Liaoning Provincial Education Department [grant number LJ242410159058] and Xingliao Talent Program of Liaoning Province [Grant number XLYC2403102].

Author contributions

Study concepts and design: L.W., S.C., H.W., N.X, J.Z., B.C., T.A., G.Y., JN.S., Z.Z., Q.H., B.L., L.Y., and JD.S. Literature research: L.W., Q.S., J.Z., B.C., N.X, JN.S., Z.Z., L.Y., and JD.S. Clinical studies: L.W., S.C., J.Z., Q.H., B.L., L.Y., and JD.S. Experimental studies/data analysis: L.W., S.C., H.W., J.Z., Z.L., B.C., L.Y., and JD.S. Statistical analysis: L.W., N.X, JN.S., Z.Z., and JD.S. Manuscript preparation: All authors Manuscript editing: All authors All authors have read and approved the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

J.D.S. discloses support for the research of this work from the National Natural Science Foundation of China [grant number 92259104], the Scientific Research Fund of Liaoning Provincial Education Department [grant number LJ242410159058], and Xingliao Talent Program of Liaoning Province [grant number XLYC2403102].

Data availability

The source code and the source data for each experiment are openly available at the following figshare repository: https://doi.org/10.6084/m9.figshare.31015861. The data that support the findings of this study are available at: https://github.com/JD910/CLARITY-ICI-Matrix. Source Data are provided with this paper. The data underlying each figure and table are provided in a single Excel file entitled “Source Data,” with each figure and table corresponding to a separate worksheet. Source data are provided with this paper.

Code availability

All code used in this study for CLARITY-ICI Matrix building has been provided at: https://github.com/JD910/CLARITY-ICI-Matrix (https://doi.org/10.5281/zenodo.18973268), and figshare: https://doi.org/10.6084/m9.figshare.31015861. An interactive online CLARITY-ICI Matrix tool is publicly accessible at http://www.ciitool.com/#/clarityICI.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Lu Wang, Nan Xu, Hairui Wang, Bosinan Chen, Qiuju Sheng, Jie Zhang.

Contributor Information

Song Chen, Email: chensongchina@163.com.

Li Yu, Email: yulimail369@163.com.

Jiangdian Song, Email: song.jd0910@gmail.com.

Supplementary information

The online version contains show [QJ][#,63]supplementary material available at https://doi.org/10.1038/s41467-026-77375-6.

References

  • 1.Doroshow, D. B. et al. Immunotherapy in Non-small Cell Lung Cancer: Facts And Hopes. Clin. Cancer Res. J. Am. Assoc. Cancer Res.25, 4592–4602 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Sacher, A. G. & Gandhi, L. Biomarkers for the clinical use of PD-1/PD-L1 inhibitors in non-small-cell lung cancer: a review. JAMA Oncol.2, 1217–1222 (2016). [DOI] [PubMed] [Google Scholar]
  • 3.Ricciuti, B. et al. Association of high tumor mutation burden in non-small cell lung cancers with increased immune infiltration and improved clinical outcomes of PD-L1 Blockade across PD-L1 expression levels. JAMA Oncol.8, 1160–1168 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Sholl, L. M. et al. The promises and challenges of tumor mutation burden as an immunotherapy biomarker: a perspective from the International Association for the Study of Lung Cancer Pathology Committee. J. Thorac. Oncol.: Publ. Int. Assoc. Study Lung Cancer15, 1409–1424 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Moutafi, M. K. et al. High-throughput transcriptome profiling indicates ribosomal RNAs to be associated with resistance to immunotherapy in non-small cell lung cancer (NSCLC). J. Immunother. Cancer12, e009039 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Ravi, A. et al. Genomic and transcriptomic analysis of checkpoint blockade response in advanced non-small cell lung cancer. Nat. Genet.55, 807–819 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Captier, N. et al. Integration of clinical, pathological, radiological, and transcriptomic data improves prediction for first-line immunotherapy outcome in metastatic non-small cell lung cancer. Nat. Commun.16, 614 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Lobefaro, R. et al. Immunotherapy in advanced Non-Small Cell Lung Cancer patients with poor performance status: The role of clinical-pathological variables and inflammatory biomarkers. Lung Cancer152, 165–173 (2021). [DOI] [PubMed] [Google Scholar]
  • 9.Chang, T.-G. et al. LORIS robustly predicts patient outcomes with immune checkpoint blockade therapy using common clinical, pathologic and genomic features. Nat. Cancer5, 1158–1175 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Conforti, F. et al. Sex-based heterogeneity in response to lung cancer immunotherapy: a systematic review and meta-analysis. J. Natl. Cancer Inst.111, 772–781 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Lee, W.-C. et al. Multiregion gene expression profiling reveals heterogeneity in molecular subtypes and immunotherapy response signatures in lung cancer. Mod. Pathol.: J. U. S. Can. Acad. Pathol., Inc.31, 947–955 (2018). [DOI] [PubMed] [Google Scholar]
  • 12.Apavaloaei, A. et al. Tumor antigens preferentially derive from unmutated genomic sequences in melanoma and non-small cell lung cancer. Nat. Cancer8, 1419–1437 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Sehgal, K. et al. Association of Performance Status With Survival in Patients With Advanced Non-Small Cell Lung Cancer Treated With Pembrolizumab Monotherapy. JAMA Netw. Open4, e2037120 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Herbst, R. S. et al. Pembrolizumab versus docetaxel for previously treated, PD-L1-positive, advanced non-small-cell lung cancer (KEYNOTE-010): a randomised controlled trial. Lancet387, 1540–1550 (2016). [DOI] [PubMed] [Google Scholar]
  • 15.Hakozaki, T., Kitadai, R., Kitagawa, S. & Hosomi, Y. P52.04 efficacy of immune checkpoint inhibitor monotherapy for patients with massive non-small-cell lung cancer. J. Thorac. Oncol.16, S520–S521 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Paz-Ares, L. G. et al. First-Line Nivolumab Plus Ipilimumab with chemotherapy versus chemotherapy alone for metastatic NSCLC in CheckMate 9LA: 3-year clinical update and outcomes in patients with brain metastases or select somatic mutations. J. Thorac. Oncol.18, 204–222 (2023). [Google Scholar]
  • 17.Shen, Y., Chen, J.-Q. & Li, X.-P. Differences between lung adenocarcinoma and lung squamous cell carcinoma: Driver genes, therapeutic targets, and clinical efficacy. Genes Dis.12, 101374 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Chang, T.-G. et al. Optimizing cancer immunotherapy response prediction by tumor aneuploidy score and fraction of copy number alterations. NPJ Precis Oncol.7, 54 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Feng, H. et al. Identification of Signature Genes and Characterizations of Tumor Immune Microenvironment and Tumor Purity in Lung Adenocarcinoma Based on Machine Learning. Front. Med.9, 843749 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Diem, S. et al. Neutrophil-to-Lymphocyte ratio (NLR) and Platelet-to-Lymphocyte ratio (PLR) as prognostic markers in patients with non-small cell lung cancer (NSCLC) treated with nivolumab. Lung Cancer111, 176–181 (2017). [DOI] [PubMed] [Google Scholar]
  • 21.Reck, M. et al. Updated Analysis of KEYNOTE-024: Pembrolizumab Versus Platinum-Based Chemotherapy for Advanced Non-Small-Cell Lung Cancer With PD-L1 Tumor Proportion Score of 50% or Greater. J. Clin. Oncol.:J. Am. Soc. Clin. Oncol.37, 537–546 (2019). [DOI] [PubMed] [Google Scholar]
  • 22.van de Haar, J. et al. Combining Genomic Biomarkers to Guide Immunotherapy in Non-Small Cell Lung Cancer. Clin. Cancer Res. J. Am. Assoc. Cancer Res.30, 1307–1318 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Wang, L. et al. Role of artificial intelligence in medical image analysis. Chin. Med J.138, 2879–2894 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Joseph, S. et al. GA-AGN: A generative adversarial network and attention gated network model for enhanced lung cancer detection using chest CT scans. EngMedicine2, 100077 (2025). [Google Scholar]
  • 25.Austin, P. C., White, I. R., Lee, D. S. & van Buuren, S. Missing Data in Clinical Research: A Tutorial on Multiple Imputation. Can. J. Cardiol.37, 1322–1331 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Balachandran, V. P., Gonen, M., Smith, J. J. & DeMatteo, R. P. Nomograms in oncology: more than meets the eye. Lancet Oncol.16, e173–e180 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Liang, S. et al. The prognostic biological markers of immunotherapy for non-small cell lung cancer: current landscape and future perspective. Front. Immunol.14, 1249980 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Lobefaro, R. et al. Immunotherapy in advanced Non-Small Cell Lung Cancer patients with poor performance status: The role of clinical-pathological variables and inflammatory biomarkers. Lung Cancer152, 165–173 (2020). [DOI] [PubMed] [Google Scholar]
  • 29.Tomasik, B., Bieńkowski, M., Braun, M., Popat, S. & Dziadziuszko, R. Effectiveness and safety of immunotherapy in NSCLC patients with ECOG PS score ≥2 - Systematic review and meta-analysis. Lung Cancer158, 97–106 (2021). [DOI] [PubMed] [Google Scholar]
  • 30.Gao, J., Zhang, C., Wei, Z. & Ye, X. Immunotherapy for early-stage non-small cell lung cancer: A system review. J. Cancer Res Ther.19, 849–865 (2023). [DOI] [PubMed] [Google Scholar]
  • 31.Cramer-van der Welle, C. M. et al. Real-world outcomes versus clinical trial results of immunotherapy in stage IV non-small cell lung cancer (NSCLC) in the Netherlands. Sci. Rep.11, 6306 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Popat, S. et al. Association Between Smoking History and Overall Survival in Patients Receiving Pembrolizumab for First-Line Treatment of Advanced Non-Small Cell Lung Cancer. JAMA Netw. Open5, e2214046 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Norum, J. & Nieder, C. Tobacco smoking and cessation and PD-L1 inhibitors in non-small cell lung cancer (NSCLC): a review of the literature. ESMO Open3, e000406 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Zhao, W. et al. Impact of smoking history on response to immunotherapy in non-small-cell lung cancer: a systematic review and meta-analysis. Front. Oncol.11, 703143 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Luo, D. et al. Effect of smoking status on immunotherapy for lung cancer: a systematic review and meta-analysis. Front. Oncol.14, 1422160 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Xu, Y. et al. The association of PD-L1 expression with the efficacy of anti-PD-1/PD-L1 immunotherapy and survival of non-small cell lung cancer patients: a meta-analysis of randomized controlled trials. Transl. Lung Cancer Res.8, 413–428 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Yan, X., Tian, X., Wu, Z. & Han, W. Impact of age on the efficacy of immune checkpoint inhibitor-based combination therapy for non-small-cell lung cancer: a systematic review and meta-analysis. Front. Oncol.10, 1671 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Wang, Y. et al. Comparison between immunotherapy efficacy in early non-small cell lung cancer and advanced non-small cell lung cancer: a systematic review. BMC Med.20, 426 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Petrelli, F. & Barni, S. Correlation of progression-free and post-progression survival with overall survival in advanced colorectal cancer. Ann. Oncol.: J. Eur. Soc. Med. Oncol.24, 186–192 (2012). [DOI] [PubMed] [Google Scholar]
  • 40.Rutkowski, J., Saad, E. D., Burzykowski, T., Buyse, M. & Jassem, J. Chronological trends in progression-free, overall, and post-progression survival in first-line therapy for advanced NSCLC. J. Thorac. Oncol.: Off. Publ. Int. Assoc. Study Lung Cancer14, 1619–1627 (2019). [DOI] [PubMed] [Google Scholar]
  • 41.Fancello, L., Gandini, S., Pelicci, P. G. & Mazzarella, L. Tumor mutational burden quantification from targeted gene panels: major advancements and challenges. J. Immunother. Cancer7, 183 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Vega, D. M. et al. Aligning tumor mutational burden (TMB) quantification across diagnostic platforms: phase II of the Friends of Cancer Research TMB Harmonization Project. Ann. Oncol.: J. Eur. Soc. Med. Oncol.32, 1626–1636 (2021). [DOI] [PubMed] [Google Scholar]
  • 43.Arango-Argoty, G. et al. AI-driven predictive biomarker discovery with contrastive learning to improve clinical trial outcomes. Cancer Cell43, 875–890.e8 (2025). [DOI] [PubMed] [Google Scholar]
  • 44.Uehara, Y. et al. Association between the baseline tumor size and outcomes of patients with non-small cell lung cancer treated with first-line immune checkpoint inhibitor monotherapy or in combination with chemotherapy. Transl. Lung Cancer Res.11, 135–149 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Anagnostou, V. et al. Multimodal genomic features predict outcome of immune checkpoint blockade in non-small-cell lung cancer. Nat. Cancer1, 99–111 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information (6.5MB, pdf)
41467_2026_77375_MOESM2_ESM.pdf (113.8KB, pdf)

Description of Additional Supplementary Files

Supplementary Video 1 (22.1MB, mp4)
Reporting Summary (5MB, pdf)
Transparent Peer Review file (5.4MB, pdf)
Source Data (5MB, zip)

Data Availability Statement

The source code and the source data for each experiment are openly available at the following figshare repository: https://doi.org/10.6084/m9.figshare.31015861. The data that support the findings of this study are available at: https://github.com/JD910/CLARITY-ICI-Matrix. Source Data are provided with this paper. The data underlying each figure and table are provided in a single Excel file entitled “Source Data,” with each figure and table corresponding to a separate worksheet. Source data are provided with this paper.

All code used in this study for CLARITY-ICI Matrix building has been provided at: https://github.com/JD910/CLARITY-ICI-Matrix (https://doi.org/10.5281/zenodo.18973268), and figshare: https://doi.org/10.6084/m9.figshare.31015861. An interactive online CLARITY-ICI Matrix tool is publicly accessible at http://www.ciitool.com/#/clarityICI.

Articles from Nature Communications are provided here courtesy of Nature Publishing Group

RESOURCES