PD-L1 alone is insufficient to select first-line immunotherapy. Pembrolizumab is approved as first-line therapy for advanced non-small cell lung cancer with PD-L1 tumor proportion score of 50% or above. Yet in the landmark KEYNOTE-024 trial, only 45% of these highly selected patients responded to pembrolizumab alone, revealing a critical need for additional biomarkers.
The KEAP1/NFE2L2 pathway regulates cellular resistance to oxidative stress. KEAP1 normally targets the transcription factor NFE2L2 for degradation, but somatic mutations in either gene lead to constitutive activation of antioxidant and metabolic reprogramming pathways. These mutations are found in 3.5-15% of NSCLC cases for NFE2L2 and 12-17% for KEAP1.
KEAP1/NFE2L2 mutations have been shown to predict radiation resistance in localized NSCLC treated with radiotherapy but not surgery, linked to enhanced free radical scavenging and reduced DNA damage. There is also evidence suggesting these mutations influence response to immune checkpoint inhibitors, though results across studies have been conflicting.
Standard genetic testing for KEAP1/NFE2L2 mutations is rarely performed in clinical practice due to cost and limited availability. A non-invasive imaging-based surrogate for mutation status would therefore be clinically valuable. A radiomics model called MUTPET was previously developed and externally validated using FDG-PET/CT data, and this study tests it in the immunotherapy setting for the first time.
A retrospective validation of a pre-existing radiomics model. This study enrolled 94 consecutive patients with advanced or metastatic NSCLC and PD-L1 TPS of 50% or more who received first-line pembrolizumab monotherapy at four institutions. All patients underwent pre-treatment FDG-PET/CT at a single center. Seventeen patients who experienced hyperprogression within 60 days of treatment were excluded, leaving 77 patients for analysis.
The MUTPET model was developed prior to this study using multiple public datasets and externally validated on 151 patients treated with radiotherapy. It combines five radiomics features: four derived from PET images (including gray-level size zone matrix features with wavelet filters and gray-level co-occurrence matrix features) and one maximum pixel value feature from CT. A predicted mutation risk above 20% classified patients as harboring a KEAP1 or NFE2L2 mutation.
To ensure transferability across imaging devices and institutional differences, harmonic normalization was performed using the neuroCombat procedure, pooling the original training cohort with the present cohort. The same model architecture, features, coefficients, and probability threshold from the original publication were applied without retraining.
Primary endpoint was progression-free survival, defined as time from immunotherapy start to disease progression or death. Overall survival was the secondary endpoint. Robustness of the radiomics model was assessed by having two physicians independently segment all primary lesions, with agreement measured by the DICE coefficient.
Predicted KEAP1/NFE2L2 mutation associated with longer PFS. In univariable analysis, the MUTPET model was a statistically significant predictor of improved progression-free survival (HR = 0.51, 95% CI 0.30-0.91, p = 0.02) when patients were classified as predicted mutated. Median PFS was 18.8 months for predicted mutated patients versus 9.2 months for predicted non-mutated patients.
In multivariable analysis, the association trended toward significance but did not reach the threshold (HR = 0.60, 95% CI 0.34-1.06, p = 0.08), likely reflecting the small sample size and the dominant effect of liver metastases in the multivariate model.
For overall survival, the MUTPET model was not statistically significant in either univariable (HR = 0.61, p = 0.10) or multivariable analysis, though median OS numerically favored predicted mutated patients at 27.9 months versus 18.0 months for predicted non-mutated patients.
During follow-up, 56 of 77 patients (72.2%) experienced disease progression with a median PFS of 11.8 months, and 51 patients (66.2%) died, with a median OS of 22.9 months. The cohort was predominantly male (78%), current or former smokers (88%), and stage IV at diagnosis (85.7%).
Liver metastases dominated multivariable survival analysis. Liver metastases were the strongest independent predictor of both shortened PFS (HR = 7.67, p less than 0.0001) and shortened OS (HR = 8.96, p less than 0.0001) in multivariable analysis. Squamous cell carcinoma histology tended toward unfavorable PFS (HR = 1.75, p = 0.08) but did not reach significance.
Two nomograms were developed: a clinical nomogram using histology and liver metastases status, and a combined nomogram that also incorporated the MUTPET radiomics prediction score. The combined nomogram stratified patients into four PFS groups with a stronger chi-test score (39.43, p less than 0.0001) compared to the clinical nomogram (35.17, p less than 0.0001).
The combined nomogram assigned: 1 point for predicted non-mutated status, 1 point for squamous cell carcinoma histology, and 7 points for liver metastases. The four resulting risk groups showed a clear stepwise separation in PFS Kaplan-Meier curves, with statistical significance confirmed by both independent physician delineations.
When evaluated in the full non-selected cohort including hyperprogressors, the MUTPET model alone lost statistical significance. However, the combined radiomics nomogram still outperformed the clinical-only nomogram, suggesting the radiomic mutation score retains additive value when integrated with clinical context even in unfiltered populations.
Why mutated tumors may respond better to immunotherapy. Studies by Xu et al. showed that KEAP1/NFE2L2 mutations correlate with higher tumor mutational burden and PD-L1 expression, both established biomarkers for immunotherapy response. These mutations appeared to be associated with improved survival with immunotherapy compared to chemotherapy in that cohort.
Conflicting evidence comes from Zhu et al., who analyzed 853 patients in the OAK and POPLAR trial cohorts and found KEAP1/NFE2L2 mutations associated with poorer OS and PFS on immunotherapy, particularly in adenocarcinoma. The current study's findings align with the Xu et al. interpretation, supporting a potentially beneficial role of these mutations in the PD-L1 high subgroup.
Liver metastases were confirmed as a major independent adverse prognostic factor for both PFS and OS, consistent with literature showing reduced immunotherapy benefit in patients with hepatic involvement. This is biologically plausible given the immunosuppressive liver microenvironment and immune tolerance mechanisms associated with hepatic metastases.
The brain metastases subgroup showed no significant survival impact, which aligns with accumulating evidence that immune checkpoint inhibitors retain activity as single agents against intracranial metastases in NSCLC, making brain involvement less prognostically distinct than liver involvement in this setting.
High inter-reader reproducibility of the radiomics model. Only 1 of 77 patients differed in KEAP1/NFE2L2 mutation prediction between the two independent physician delineations, confirming strong model robustness. The mean DICE coefficient between readers was 0.89 for PET and 0.85 for CT delineations, reflecting high segmentation agreement.
Key limitations include the retrospective single-center design, small sample size, and the fact that actual KEAP1/NFE2L2 mutational status was not genetically confirmed in this cohort -- the study evaluates a predicted mutation status from imaging. This limits the direct interpretability of the findings, though the model was previously validated against confirmed mutation status in a separate 151-patient cohort.
The exclusion of hyperprogressors (18% of the initial cohort) improves the biological signal but limits external validity. Hyperprogression under immunotherapy is a distinct phenomenon associated with worse survival, and including these patients diluted the radiomics model's prognostic signal in the unfiltered cohort.
This study underlines the potential for radiomics to non-invasively characterize tumor genomics, supporting future strategies of therapeutic intensification or de-escalation. The MUTPET model could guide the decision between pembrolizumab monotherapy versus chemo-immunotherapy combination as first-line treatment for PD-L1-high advanced NSCLC patients.