Development and validation of a circulating tumor DNA-based machine learning model for predicting immunotherapy response in non-small cell lung cancer.

Introduction
The advent of immune checkpoint inhibitors (ICIs) has transformed the treatment landscape for advanced non-small cell lung cancer (NSCLC) by producing durable responses in a subset of patients (1-3). Existing clinical biomarkers—such as programmed death-ligand 1 (PD-L1) expression, tumor mutational burden (TMB), and gene-expression signatures—show statistical associations with ICI benefit but lack sufficient accuracy for reliable patient stratification (4). Long-term trial data (for example, the PACIFIC study) have revealed substantial heterogeneity in survival outcomes across prognostic subgroups, underscoring the urgent need for improved predictive strategies (5).
Assessing immunotherapy efficacy poses unique challenges. Atypical response patterns (including delayed responses and pseudoprogression) complicate timely evaluation and can confound endpoint interpretation (6). Moreover, in immunotherapy trials for NSCLC, the relationship between progression-free survival (PFS) and overall survival (OS) is often discordant, which limits the utility of conventional endpoints for early decision-making (7). These factors highlight the necessity of predictive models that combine baseline molecular features with tumor-biology dynamics to guide earlier and more accurate therapeutic choices (8).
Circulating tumor DNA (ctDNA) offers a minimally invasive window into tumor genomics and has emerged as a promising source of predictive biomarkers (9-11). ctDNA captures somatic mutation and epigenetic landscapes and can be sampled serially from peripheral blood. Higher blood-based TMB (bTMB) measured from ctDNA has been associated with improved ICI response, likely reflecting increased neoantigen burden and enhanced immunogenicity (12). Specific driver alterations (for example, STK11, KEAP1, and lesions affecting JAK-STAT signaling) can reshape the tumor microenvironment and influence sensitivity to ICIs (13). Specific driver gene mutations influence immunotherapy responsiveness by modulating the tumor microenvironment—particularly through alterations in PD-L1 expression and immune cell infiltration—thereby offering molecular indicators for personalized therapeutic strategies (14). Empirical studies further demonstrate that ctDNA kinetic profiles function as early biomarkers of response in NSCLC immunotherapy (15). Patients achieving post-treatment ctDNA clearance, defined by mutant allele frequencies below detection thresholds, exhibited markedly prolonged median PFS (mPFS), with outcomes showing strong concordance with radiographic evaluations (16). Additionally, ctDNA profiling can reveal acquired resistance mechanisms (e.g., JAK1/2 loss, B2M inactivation, promoter methylation) that facilitate immune escape (17), and minimal-residual-disease (MRD) monitoring by ctDNA can identify patients at higher risk of recurrence who may benefit from adjuvant strategies (18).
Machine learning and deep-learning approaches have shown substantial promise for integrating high-dimensional genomic and clinical data to improve outcome prediction in oncology (19,20). For instance, deep networks trained on whole-exome sequencing (WES) or combined radiogenomic data can capture complex, non-linear relationships among TMB, neoantigen load, and the immune microenvironment, producing more accurate predictions of ICI response than many conventional methods (21,22).
Building on these observations, we developed a support vector machine (SVM) model using pretreatment ctDNA WES features to predict durable clinical benefit (DCB) from ICIs in advanced NSCLC. The model was trained on the OAK trial cohort externally validated in the POPLAR trial and a regional multicenter cohort to assess predictive performance and potential clinical utility. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1182/rc).

Methods

Public clinical trial cohorts (OAK and POPLAR)
Individual patient-level data (IPD) were obtained from two multicenter randomized trials of atezolizumab in advanced/metastatic NSCLC (OAK and POPLAR), and the overall study design and cohort integration workflow are illustrated in Figure 1. After excluding patients who did not receive atezolizumab, specimens failing quality control, or records with incomplete clinical metadata or missing ctDNA/WES data, 303 patients from OAK were used as the training cohort and 97 patients from POPLAR served as an independent validation cohort (Figure 1). For reference on the clinical trials of atezolizumab used here, see Socinski et al. (23). For clarity, we defined blood TMB as the number of somatic mutations per megabase of coding sequence identified from ctDNA WES data.

Local validation cohort
A prospective multicenter regional cohort was assembled between April 2023 and June 2024. Eligible patients had pathologically confirmed stage III–IV NSCLC and were scheduled to receive programmed death-1 (PD-1)/PD-L1 inhibitor therapy (monotherapy or combination). Of 52 initially enrolled patients, 2 discontinued before completing two cycles and 3 were lost to follow-up, leaving 47 evaluable cases; 6 were ctDNA-negative (exploratory) and 41 were ctDNA-positive and included in the primary validation analyses (Figure 1). Peripheral blood (20 mL) was collected 24 h before treatment in Streck cfDNA tubes, shipped at 2–8 °C, and processed at a central laboratory (iGeneTech, Wuhu, China) within 72h. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Affiliated Hospital of Guizhou Medical University (No. 2023-LS-02), and all participating hospitals were informed and agreed to the study. All participants provided written informed consent.

Sample processing and sequencing
Plasma ctDNA was extracted using a magnetic bead-based protocol. DNA integrity and quantity were assessed by agarose gel electrophoresis and Qubit fluorometry. WES libraries were prepared with hybrid-capture exome probes following fragmentation (150–250 bp), end repair, adapter ligation, and polymerase chain reaction (PCR) enrichment. Libraries passing quality control (QC) (≥25 ng/µL, expected fragment distribution 220–450 bp) were sequenced on an Illumina NovaSeq 6000 platform with 2×150 bp paired reads to achieve median exome coverage adequate for variant calling. Somatic single-nucleotide variants (SNVs) and small insertions/deletions (InDels) were called using standard pipelines, filtered against population germline databases, and compared with matched normals when available (Figure 2).

Feature selection and model development
The analytic workflow is shown in Figure 3. In the OAK training set, we performed high-dimensional feature reduction with the least absolute shrinkage and selection operator (LASSO) using five-fold cross-validation to select a parsimonious set of predictive somatic features (final panel: 41 features; Figure 4). These binary mutation features were then used to train an SVM classifier implemented in R (version 4.3.x; e1071 package). We used a radial basis function (RBF) kernel and optimized hyperparameters (penalty C and kernel width σ) via grid search on the training folds. An ε-insensitive loss (ε =0.015) was applied during SVM fitting. To avoid information leakage between feature selection and model evaluation, we employed nested five-fold cross-validation during model development (i.e., the inner loop for hyperparameter tuning/feature selection and the outer loop for performance estimation). The final model produced an SVM score for each patient (Table S1), interpreted as a continuous measure of predicted probability of DCB. For general machine-learning context on similar clinical prediction tasks (24).

Clinical endpoints
DCB was defined as complete response (CR), partial response (PR), or stable disease (SD) sustained for ≥6 months from treatment initiation; non-DCB was defined as progressive disease (PD) or SD <6 months. PFS was measured from the first ICI dose to radiographic progression (per RECIST v1.1) or death; OS was measured from treatment start to death from any cause. Tumor responses were assessed by computed tomography (CT) or magnetic resonance imaging (MRI) every 6–8 weeks and adjudicated using RECIST v1.1, with confirmation imaging used to limit misclassification from pseudoprogression.

Data sharing statement
The datasets analyzed in this study were derived from publicly available clinical trials and a prospectively collected regional cohort. IPD from the OAK (ClinicalTrials.gov identifier: NCT02008227) and POPLAR (NCT01903993) studies were obtained through authorized data-sharing agreements with the data-holding institution (F. Hoffmann-La Roche Ltd.). The regional validation cohort data were collected at our institution following ethical approval and patient-informed consent. Deidentified patient-level data supporting the findings of this study are available from the corresponding author upon reasonable request. Requests will be reviewed in accordance with institutional and data protection policies to ensure appropriate data use and privacy protection.

Statistical analysis
Model discrimination was evaluated by receiver operating characteristic (ROC) curve analysis and reported as area under the curve (AUC) with 95% confidence intervals (CI). The optimal threshold to dichotomize the SVM score into SVM-high and SVM-low groups was selected using the Youden index on the training ROC. Continuous variables are presented as medians with interquartile ranges (IQR) and compared using Mann-Whitney U tests; categorical variables are presented as counts (%) and compared using χ2 or Fisher’s exact tests as appropriate. Kaplan-Meier methods and log-rank tests were used to compare PFS and OS between groups. Functional enrichment of the 41-gene signature was performed using Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway annotation (via standard R packages); enrichment results were assessed with appropriate multiple-testing correction and presented for pathways with P<0.05 and false discovery rate (FDR) <0.25. For comparable ctDNA-based enrichment workflows and pathway interpretation in similar studies, see Peng et al. (25).

Results

Baseline characteristics of study cohorts
A total of 303 patients from the OAK trial were included in the training cohort, 97 from the POPLAR trial in the independent validation cohort, and 41 ctDNA-positive patients from the prospective regional cohort. Clinical and demographic features of the three cohorts are summarized in Tables 1-3. In all cohorts, DCB was observed in approximately one-third of patients. No significant associations were identified between baseline clinicopathologic variables (age, sex, smoking status, histology) and DCB.
Baseline blood TMB did not significantly differ between DCB and non-DCB groups in either the OAK (median 7.0 vs. 8.5 mut/Mb, P=0.25) or POPLAR cohort (median 6.5 vs. 8.0 mut/Mb, P=0.65), confirming its modest predictive capacity compared with integrated ctDNA signatures. Similarly, in the regional cohort, median bTMB values were not significantly associated with outcome (P=0.70).

SVM model predictive performance
Using 41 ctDNA-derived genomic features selected by LASSO regression, the SVM model demonstrated strong discriminative performance in the OAK training cohort. The median SVM score was 0.05 in non-DCB patients vs. 0.95 in DCB patients (P<0.001), producing an AUC of 0.87 (95% CI: 0.82–0.91) (Figure 5A).
External validation in the POPLAR cohort yielded comparable results: the SVM model distinguished outcomes with an AUC of 0.92 (95% CI: 0.86–0.97) (Figure 5B). Similarly, in the regional cohort, the model achieved an AUC of 0.83 (95% CI: 0.69–0.98) (Figure 5C). The model achieved high predictive accuracy across cohorts, with ROC curves shown in Figure 5. These results confirm the generalizability of the model across independent datasets, despite differences in treatment regimens.
Of note, the distribution of SVM scores was highly bimodal across cohorts, with most non-DCB patients clustering at the lower extreme (~0.05) and DCB patients at the higher extreme (~0.95). This distribution underscores the model’s ability to act as a near-binary classifier, facilitating patient stratification in a clinically interpretable manner.

Survival stratification by SVM score
The SVM score threshold of 0.55 was determined by the Youden index from the ROC curve in the OAK training cohort, stratifying patients into SVM-low (>0.55, low-risk group) and SVM-high (≤0.55, high-risk group; reference group). Across all cohorts, SVM-low patients achieved significantly longer PFS and OS compared with SVM-high patients (Figures 6,7).
In OAK, mPFS was 11.6 vs. 1.7 months [hazard ratio (HR) =0.35; P<0.001] (Figure 6A), and median OS (mOS) was 19.5 vs. 9.3 months (HR =0.62, P<0.001) (Figure 6B). In POPLAR, mPFS was 11.5 vs. 1.9 months (HR =0.36, P<0.001) (Figure 6C), and mOS was 20.7 vs. 9.7 months (HR =0.475, P=0.003) (Figure 6D). In the regional cohort, which included 47 patients who completed the treatment and had adequate follow-up (Table S2), the mPFS was 12.8 vs. 3.0 months (HR =0.191, P=0.002) (Figure 7A), and the mOS was 17.3 vs. 6.4 months (HR =0.159, P=0.002) (Figure 7B).
Importantly, among patients with initial SD at the first radiologic evaluation, the SVM model stratified long-term outcomes: SVM-low patients had markedly longer survival than SVM-high patients [mPFS 12.8 vs. 4.2 months, P=0.04; mOS 15.6 vs. 13.4 months, P=0.03] (Figure 7C,7D). This suggests that the model may help distinguish pseudoprogression from true non-response, a recognized clinical challenge in immunotherapy (26).

Gene enrichment analysis
Functional annotation of the 41 selected genes revealed significant enrichment in biologically relevant pathways (Table S3). KEGG analysis identified VEGF signaling, Ras signaling, cell cycle regulation, JAK-STAT signaling, and neurotrophin signaling as key enriched pathways (Figure 8A,8B). These pathways are known to influence tumor immune evasion and therapeutic resistance.
Notably, genes such as STK11, KEAP1, TP53, RB1, EGFR, MTOR, PIK3R1, RAF1, FLT1, and CDKN2A were included, many of which have been implicated in immune suppression and poor outcomes with ICIs in NSCLC (27). The enrichment results provide mechanistic support for the predictive value of the identified ctDNA signature.

Discussion
In this study, we developed and validated a ctDNA-based machine learning model that predicts immunotherapy benefit in advanced NSCLC. Using baseline ctDNA WES data, we trained an SVM model in the OAK trial cohort and externally validated it in both the POPLAR trial and a regional prospective cohort. Across all datasets, the model demonstrated robust and reproducible predictive performance, with AUC exceeding 0.80. Patients categorized as SVM-low exhibited significantly longer PFS and OS compared with SVM-high patients.
These findings highlight the potential of ctDNA-derived mutational profiles as a noninvasive, pretreatment biomarker for predicting DCB from ICIs. The use of baseline ctDNA signatures offers a practical advantage over dynamic monitoring approaches, which require serial sampling and longitudinal follow-up (28). By leveraging a single timepoint sample, the model facilitates early patient stratification and may assist clinicians in optimizing ICI selection and avoiding ineffective therapy.
Traditional biomarkers such as PD-L1 expression and TMB have limited accuracy in predicting ICI response. The modest predictive power of blood TMB observed in the OAK and POPLAR studies was consistent with our findings (29). In contrast, our model integrates multiple somatic genomic features using a data-driven approach, thereby capturing a broader range of molecular determinants related to immune sensitivity and resistance.
Previous studies have suggested that certain driver mutations—such as STK11, KEAP1, PTEN, and EGFR—are associated with a non-inflamed tumor microenvironment and poor outcomes with ICIs (30). Our model incorporated several of these genes, which likely contributed to its predictive strength. Our model incorporated several of these genes, which likely contributed to its predictive strength. Emerging data suggest that NTRK1-related signaling may also modulate immune responsiveness and checkpoint inhibitor efficacy (31). Moreover, the enriched signaling pathways identified in this study, including VEGF, JAK-STAT, Ras, and cell-cycle regulation, are consistent with known immunosuppressive and oncogenic mechanisms. For instance, VEGF signaling can inhibit dendritic cell maturation and promote T-cell exhaustion, whereas JAK-STAT alterations may impair interferon-γ signaling and antigen presentation (32). These convergent findings reinforce the biological relevance of the ctDNA-derived gene set.
In addition, the bimodal distribution of SVM scores observed in our model suggests clear separation between responders and non-responders, which aligns with the clinical dichotomy frequently observed in ICI-treated NSCLC populations (33). The reproducibility of this distribution across three independent cohorts underscores the model’s stability and the potential to generalize beyond specific datasets or treatment regimens.
The clinical utility of the proposed SVM model lies in its ability to identify patients likely to derive benefit from ICIs prior to treatment initiation. This approach is particularly valuable given the heterogeneous and sometimes unpredictable responses observed with immunotherapy. By using pretreatment ctDNA data, our model could complement or even surpass traditional biomarkers such as PD-L1 and TMB in predictive accuracy.
Furthermore, among patients who initially achieved SD after two cycles of ICI therapy, the model successfully stratified long-term outcomes, distinguishing those with true disease control from those with eventual progression. This capability could help resolve the diagnostic ambiguity of pseudoprogression, a phenomenon that complicates early treatment decisions in immunotherapy. Consequently, the model may serve as a decision-support tool, helping clinicians determine whether to continue, adjust, or discontinue treatment.
Given that the regional validation cohort included patients receiving combination regimens (immunotherapy plus chemotherapy and/or antiangiogenic therapy), the model’s ability to retain discriminative power in this heterogeneous clinical context suggests potential applicability across therapeutic settings. Nevertheless, further validation in large-scale, prospective, and ethnically diverse populations is warranted to confirm its clinical robustness.
Several limitations should be acknowledged. First, the OAK and POPLAR cohorts represent Western populations, while our local validation cohort consisted primarily of East Asian patients. Ethnic and genomic differences may influence ctDNA mutational profiles and immune microenvironment characteristics (34). Therefore, future studies should include multiethnic and geographically diverse cohorts to enhance the generalizability of the model. Second, although the regional cohort introduced real-world heterogeneity in treatment regimens, including immunochemotherapy and anti-VEGF therapy, this variability might have confounded certain associations. Stratified analyses by treatment type in larger cohorts will be necessary to confirm the robustness of the predictive model under different therapeutic contexts. Third, the SVM score distribution appeared bimodal, indicating that the model primarily behaves as a binary classifier. While this enhances interpretability, it may limit its use for probabilistic predictions in intermediate cases. Future model iterations could incorporate calibration techniques or probabilistic frameworks to improve granularity. Fourth, the current study relied exclusively on baseline ctDNA features without integrating other potential biomarkers such as PD-L1 expression, tumor-infiltrating lymphocytes, or radiomic features. A multimodal approach combining ctDNA, imaging, and immune profiling might further improve predictive accuracy. Finally, the sample size of the local validation cohort was modest. Larger, prospective trials with serial sampling would allow the integration of dynamic ctDNA kinetics, providing a comprehensive framework for assessing both baseline prediction and early response monitoring (35).
In summary, our ctDNA-based SVM model provides a novel, noninvasive strategy for predicting ICI benefit in advanced NSCLC using a single pretreatment blood sample. The model demonstrated consistent predictive accuracy across multiple independent cohorts and revealed biologically relevant gene signatures associated with immune responsiveness. By improving pretreatment stratification, this approach could facilitate personalized immunotherapy and reduce unnecessary exposure to ineffective treatments. Further multicenter validation and integration with complementary biomarkers are ongoing to accelerate its clinical translation.
In this multi-cohort study, we developed and validated a ctDNA-based machine learning model capable of predicting immunotherapy benefit in patients with advanced NSCLC. By integrating pretreatment ctDNA mutational features using an SVM algorithm, the model accurately stratified patients into high- and low-benefit subgroups across three independent cohorts, achieving high reproducibility and robust predictive power. Unlike conventional biomarkers such as PD-L1 expression and TMB, which show limited predictive value, our model leverages the complexity of the ctDNA mutational landscape to capture multiple mechanisms influencing immune responsiveness. The identified 41-gene signature was enriched in key oncogenic and immune-regulatory pathways—including VEGF, JAK-STAT, and Ras signaling—consistent with prior studies on immune resistance in NSCLC. The main strength of this approach lies in its reliance on a single baseline liquid biopsy, enabling early patient selection without the need for serial sampling or invasive tissue biopsies. This feature makes the model practical for real-world applications. In addition, its predictive validity across both clinical trial and regional real-world cohorts supports potential generalizability. However, further work remains necessary. Prospective validation in larger, ethnically diverse populations is essential to confirm the model’s robustness and transferability. Future studies should also explore multimodal integration—combining ctDNA data with PD-L1 expression, radiomic features, and tumor microenvironment markers—to further enhance prediction accuracy.

Conclusions
In conclusion, this ctDNA-based SVM model offers a powerful and noninvasive tool for personalizing immunotherapy in NSCLC. With continued validation, it may support precision treatment decisions, improve clinical outcomes, and help optimize the deployment of ICIs in oncology practice.

Supplementary
The article’s supplementary files as