Lung cancer is the leading cause of cancer mortality worldwide, with NSCLC accounting for over 80 percent of cases. Despite available treatments including surgery, chemotherapy, radiotherapy, and targeted therapy, the 5-year survival rate remains only 15 percent. The emergence of immune checkpoint inhibitors targeting the PD-1/PD-L1 axis has transformed the therapeutic landscape, enabling durable responses and long-term survival in a subset of metastatic NSCLC patients.
The majority of patients do not benefit from immunotherapy, creating an urgent need for predictive biomarkers. Significant heterogeneity in the tumor microenvironment, driven by diverse oncogenic driver mutations, produces variable sensitivity to ICI therapy across patients. Over 40 predictive biomarkers including tumor mutation burden, PD-L1 expression, and microsatellite instability have been evaluated, yet no single marker reliably predicts which patients will respond, creating a substantial clinical prediction gap.
Pathway-level signatures may outperform single-gene biomarkers by capturing system-level biology. Signaling pathways represent functional ensembles of interacting molecules that regulate tumor cell growth, invasion, and immune evasion simultaneously. Characterizing multigene pathway activity can capture the complex interactions between tumors, tumor microenvironments, and host immunity that single biomarkers miss, motivating a systematic pathway-based machine learning approach to immunotherapy response prediction.
The study integrated transcriptomic data from 584 NSCLC patients across four independent ICI-treated cohorts. The OAK cohort (344 patients, NCT02008227) served as the training dataset, while Ravi (118 patients), Jung (27 patients), and Poplar (95 patients, NCT01903993) provided independent validation sets. A metacohort was constructed by integrating the three validation sets, with batch effects corrected using the ComBat algorithm from the sva package. All four cohorts provided pre-treatment RNA sequencing data alongside survival and response information.
A three-step feature selection process systematically identified the optimal pathway and algorithm combination. First, 12,025 pathways from MSigDB (including HALLMARK, REACTOME, KEGG, GOBP, GOCC, and GOMF collections) were screened by univariate Cox regression to retain genes significantly associated with progression-free survival (p less than 0.05). Second, 10 machine learning algorithms were applied individually and in combinations, producing 101 distinct algorithmic approaches. Third, each pathway-algorithm combination was evaluated using concordance index (C-index) averaged across all three validation sets to identify the single most robust predictor.
Biological validation used gene set enrichment analysis, immune infiltration profiling, and immunohistochemistry. GSEA was performed on 12,502 gene sets from MSigDB to characterize pathways enriched in high versus low risk groups. Immune infiltration was assessed using three complementary algorithms: CIBERSORT (22 immune cell types), xCell, and ESTIMATE (stromal and immune scores). To validate protein-level expression, immunohistochemistry was performed on tumor tissue from 15 in-house NSCLC patients, comparing samples with partial response versus progressive disease outcomes.
Progression-free survival was chosen as the primary endpoint to isolate direct treatment response. PFS directly measures treatment-specific efficacy without confounding effects from subsequent therapies that obscure overall survival analysis. PFS also generates timely predictions that can influence clinical decision-making and requires shorter follow-up periods, making it particularly suitable for translational bioinformatics applications where rapid feedback between molecular features and treatment outcomes is essential.
The FOXO-mediated transcription pathway combined with Lasso-RSF emerged as the top-ranked predictor across all validation sets. This combination achieved the highest average C-index of 0.6918 across the three validation cohorts, outperforming all other pathway-algorithm combinations. The Lasso component performs feature selection by penalizing the absolute magnitude of regression coefficients, while random survival forest captures non-linear interactions between features, together providing complementary dimensionality reduction and survival prediction capabilities.
FRS stratified patients into high-risk and low-risk groups with significantly different PFS across all four cohorts. Kaplan-Meier survival analysis demonstrated that high-risk FRS patients showed significantly worse PFS in the OAK training set, and across all three validation sets (Ravi, Jung, and Poplar), with all comparisons reaching p less than 0.05. The hazard ratio for the Jung cohort was 0.097 (95% CI 0.013-0.73), representing nearly complete separation between risk groups. The same survival stratification pattern was confirmed in the metacohort and across overall survival endpoints.
Time-dependent ROC analysis confirmed strong predictive performance, with AUC increasing at longer time horizons. In the OAK training cohort, 1-year AUC was 0.748, 3-year AUC was 0.974, and 5-year AUC was 0.991, indicating that FRS becomes increasingly accurate at predicting long-term outcomes. Standard AUC across the full follow-up period was 0.933 in OAK. Predictive performance in the metacohort, which introduced substantial batch and cohort heterogeneity, maintained AUCs of 0.665, 0.663, and 0.645 at 1, 3, and 5 years respectively.
FRS outperformed all clinical variables in independent prognostic prediction. Multivariate Cox analyses across all four cohorts demonstrated that FRS was an independent predictor of PFS, with significantly higher hazard ratios than age, sex, smoking status, histological subtype, and clinical stage. In the OAK cohort, FRS achieved HR 2.226 (95% CI 2.573-5.108, p less than 0.001), while other clinical variables failed to reach statistical significance. The low-risk FRS group consistently showed significantly higher objective remission rates compared to the high-risk group across all cohorts.
Across all four independent cohorts, only PCK1 and IGFBP1 consistently demonstrated significant associations with inferior PFS. Univariate Cox analyses of all FOXO pathway genes in each individual dataset identified these two genes as the only ones maintaining p less than 0.05 across OAK, Ravi, Jung, and Poplar simultaneously. PCK1 encodes phosphoenolpyruvate carboxykinase 1, a gluconeogenesis enzyme, while IGFBP1 encodes insulin-like growth factor binding protein 1, a metabolic regulator with known roles in tumor cell survival signaling.
Immunohistochemistry on tumor tissue from 15 in-house NSCLC patients confirmed protein-level validation. IHC staining with antibodies against PCK1 and IGFBP1 showed that H-scores (quantifying both percentage of positive cells and staining intensity) were significantly higher in samples from patients with progressive disease compared to those with partial response. This protein-level confirmation bridges the gap between transcriptomic model predictions and biological reality, providing direct clinical evidence that these genes mark poor immunotherapy responders at the tissue level.
FRS demonstrated cross-cancer generalizability in melanoma, urothelial carcinoma, and another NSCLC dataset. External validation in non-NSCLC ICI-treated cohorts including Freeman melanoma, IMvigor210 urothelial carcinoma, and Liu NSCLC datasets showed that low FRS scores were consistently associated with better overall survival, with hazard ratios ranging from 0.41 to 0.68. This generalizability across tumor types suggests that the FOXO-mediated immune regulation captured by FRS reflects fundamental biological mechanisms of ICI response rather than lung cancer-specific biology.
FRS outperformed 43 previously published immunotherapy predictive models across all four cohorts. The 43 comparison models represented a broad range of biological features including neutrophil differentiation, pyroptosis, epithelial-mesenchymal transition, inflammation, hypoxia, iron death, epigenetics, N6-methyladenosine modifications, tumor microenvironment features, and endoplasmic reticulum stress signatures. Univariate Cox analysis confirmed that FRS was the only model showing consistent significant associations with inferior PFS across all cohorts simultaneously.
The key differentiator was cross-cohort stability, not peak performance in a single dataset. Several published models including the Wenhao Ouyang model achieved higher C-index than FRS in specific individual datasets such as Poplar, but performed poorly in other cohorts. This inconsistency indicates model overfitting to training data characteristics. FRS consistently ranked near the top across all cohorts with no such instability, demonstrating that its Lasso-RSF dimensionality reduction strategy produced a more generalizable model than simpler approaches used by competing models.
Low-risk FRS patients show enrichment of immune activation pathways by gene set enrichment analysis. GSEA analysis ranked immune pathways by normalized enrichment score and confirmed that the low-FRS group was enriched for T cell receptor signaling, antigen binding, interferon-gamma response, B cell-mediated immunity, immune effector processes, and immune cell activation and differentiation pathways. This pattern reflects a heightened state of both adaptive and innate immune responses that creates an immunologically active microenvironment conducive to ICI efficacy.
CIBERSORT immune infiltration analysis confirmed that 12 of 22 immune cell types differed significantly between risk groups. The low-risk FRS group showed higher abundance of CD8+ T cells, CD4+ T cells, gamma delta T cells, neutrophils, macrophages, mast cells, and memory B cells compared to high-risk patients. This broad increase in immune infiltration explains why low-risk patients respond better to checkpoint blockade, as ICI efficacy requires the presence of immune effector cells within the tumor that can be activated by checkpoint release.
FOXO proteins regulate T cells, B cells, NK cells, macrophages, and dendritic cells in addition to their tumor-intrinsic functions. FOXO transcription factors control apoptosis, cancer cell metabolism, cell cycle progression, and oxidative stress resistance within tumor cells, while simultaneously shaping the homeostasis and development of multiple immune cell populations. FRS-high patients show enrichment of oncogene methylation-related pathways that suppress immune gene expression, providing a mechanistic explanation for their lower immune infiltration and poor response to checkpoint blockade.
Low-risk FRS patients show elevated expression of CTLA4, BTLA, CD27, and LAG3 immune checkpoint genes. These checkpoint gene expression levels indicate that low-risk patients have immune-engaged tumors where T cells are actively present and engaging cancer cells, but held in check by inhibitory receptors that ICIs can release. High-risk patients show elevated CD276 and VTCN1, which are checkpoint molecules associated with immune exclusion rather than immune engagement, explaining their reduced responsiveness to PD-1 and PD-L1 blockade.
A dual pathway for clinical threshold implementation addresses the need for both near-term and long-term deployment. For near-term use, individual institutions can calibrate institution-specific FRS thresholds using internal well-characterized patient cohorts, providing immediate clinical applicability. The long-term goal is a universal threshold established through a large multicenter prospective study, ideally supported by a standardized assay kit that ensures consistency and comparability across diverse healthcare settings and patient populations.
The model was primarily developed in European and American populations, limiting generalizability to other ethnic groups. NSCLC molecular characteristics including the prevalence of EGFR driver mutations, tumor microenvironment composition, and immunotherapy response patterns vary significantly across races and geographical regions. The Jung cohort represents the only Asian population in the study, and its small size (n=27) and single-center Korean origin make it insufficient for validating FRS in Asian patients, where EGFR mutation prevalence is substantially higher than in Western cohorts.
Cohort-specific cutoffs were used for validation, reflecting heterogeneity that would need standardization for clinical use. The surv_cutpoint function was applied independently within each cohort to identify the FRS threshold maximizing the log-rank statistic for that specific population. While this approach demonstrates the inherent prognostic power of FRS across diverse patient groups, it means the absolute FRS risk score threshold cannot be directly transferred between institutions without local recalibration, representing a practical barrier to implementation that requires a dedicated standardization study.
FRS robustly identified patients likely to benefit from ICI therapy across four independent NSCLC cohorts and a metacohort. By applying 101 machine learning algorithm combinations to 12,025 pathways in 584 patients, the study demonstrated that the FOXO-mediated transcription pathway captured by Lasso-RSF consistently stratified patients into high and low risk groups with significantly different PFS and objective remission rates. This performance stability across cohorts treated with different PD-1 and PD-L1 inhibitors confirms the model's generalizability beyond its training context.
FRS serves as a novel pathway-level biomarker that could guide ICI treatment selection in clinical practice. High-FRS patients, who show poor ICI responses and enrichment of methylation-related immunosuppressive pathways, may be better served by alternative treatment strategies or novel combination approaches targeting FOXO-regulated immune pathways. Low-FRS patients with immune-activated tumor microenvironments and elevated checkpoint gene expression represent the subset most likely to achieve durable benefit from PD-1 and PD-L1 blockade.
The comprehensive biological validation through GSEA, immune infiltration profiling, and IHC strengthens confidence in the model's mechanistic validity. The convergence of transcriptomic pathway enrichment analysis, multi-algorithm immune cell infiltration profiling, and protein-level IHC confirmation in an independent clinical cohort provides three independent lines of evidence that FRS captures real biological differences rather than statistical artifacts from the training data. This mechanistic grounding is essential for the broader scientific community to accept the model as a credible tool for precision oncology applications.