An integrative multimodal machine learning signature of primary resistance to immunotherapy in advanced non-small cell lung cancer: biomarker analysis from the PIONeeR study

BackgroundImmune checkpoint inhibitors (ICIs) have transformed the treatment landscape for advanced non-small cell lung cancer (NSCLC), yet primary resistance remains common, with only [~]50% of patients responding to first-line chemo-immunotherapy and 20-30% to monotherapy. Existing biomarkers such as PD-L1 expression and Tumor Mutational Burden (TMB) demonstrate limited predictive accuracy, underscoring the need for more comprehensive, integrative approaches. MethodsWe conducted a prospective, multicenter study involving 439 patients with advanced NSCLC treated with anti-PD-(L)1 ICI across first-line combo with chemotherapy and later-line monotherapy settings. A total of 443 pre-treatment tumor and blood-derived biomarkers--including genomic alterations, immune cell phenotypes, proteic markers, and routine laboratory tests--were profiled. Extensive biostatistics adjusted for PD-L1 expression were conducted. A rigorously benchmarked machine learning (ML) pipeline including 36 feature selection methods embedded into an optimism-correction framework was applied to identify predictors of primary resistance (PrR). ResultsSingle biomarkers showed limited predictive utility, with PD-L1 (AUC 0.62, positive predictive value (PPV) 49.6%), TMB (AUC 0.55, PPV 43.1%), and key gene mutations (e.g., STK11, KEAP1) failing to achieve significance after multiple testing correction. A gradient boosting ML model integrating 18 selected features yielded a corrected AUC of 0.69 and a Positive Predictive Value (PPV) of 60% for PrR, outperforming standard biomarkers. In first-line patients, the model achieved a PPV of 51% and Negative Predictive Value (NPV) of 79% (baseline PrR rate: 29.9%); in subsequent-line patients, PPV reached 64% (PrR rate: 55.1%). Importantly, the signature also stratified Progression-Free Survival (PFS): high-risk patients had a median PFS of 3.9 vs. 14.6 months in low-risk patients (HR 0.307, p < 0.0001). Features from routine blood tests--such as serum chloride, albumin, CRP, and monocyte-to-lymphocyte ratio (MLR)--accounted for half of the final model and demonstrated independent associations with both PrR and PFS (e.g., chloride: OR 0.616, AUC 0.626; HR 0.685, C-index 0.61). SHAP-based individual-level model explainability revealed heterogeneous and nonlinear biomarker contributions, including cases where high CRP, low albumin, or elevated MLR overrode favorable PD-L1 or Treg profiles. A biomarker dashboard including interactive visualizations is available at https://compo.inria.fr/pioneer-website/. ConclusionsMultimodal machine learning integration of clinical, genomic, immune, and laboratory data enables improved prediction of ICI resistance in NSCLC beyond current biomarkers. This approach not only captures the multifaceted nature of tumour-host interactions but also highlights the underrecognized predictive value of accessible blood-based markers, offering a path toward individualized immunotherapy decision-making. HighlightsO_LIWe conducted a prospective, multicenter biomarker study of 439 patients with advanced NSCLC treated with anti-PD-(L)1 therapies across first- and later-line settings, profiling a total of 443 clinical, routine laboratory parameters, tumor and blood-derived biomarkers prior to treatment initiation. C_LIO_LIA machine learning-derived 18-feature multimodal signature predicted primary resistance (PrR) to anti-PD-(L)1 therapy with higher accuracy than standard biomarkers (AUC 0.69, PPV 60%). Routine blood test features, including serum chloride, albumin, C-reactive protein, and monocyte-to-lymphocyte ratio, emerged as dominant contributors to the predictive model. C_LIO_LIThe model demonstrated robust prognostic stratification for PrR and progression-free survival. C_LIO_LIStandard-of-care immunotherapy biomarkers PD-L1 expression and TMB showed limited predictive utility (AUC [~]0.60) and were consistently outperformed by the integrative multivariable model. C_LIO_LIAn interactive digital tool using SHAP values enabled individualized prediction interpretation and illustrated heterogeneous, patient-specific resistance drivers. C_LIO_LIKeywords: multimodal biomarkers; machine learning; immunotherapy resistance; biomarker discovery C_LI