Machine learning–guided identification of determinants of PD-L1 expression and tumor mutational burden in lung squamous cell carcinoma

Abstract

Programmed death-ligand 1 (PD-L1) expression and tumor mutational burden (TMB) are widely used immunotherapy biomarkers, yet their variability and determinants in lung squamous cell carcinoma (LUSC) remain incompletely characterized in real-world testing settings. We analyzed a retrospective real-world LUSC cohort with clinical PD-L1 immunohistochemistry and next-generation sequencing. PD-L1 tumor proportion score (TPS) was modeled as an ordered endpoint using three clinically used categories (< 1%, 1–49%, ≥ 50%) via proportional-odds ordinal regression; TMB was modeled as a continuous outcome after log transformation using multivariable linear regression. To address the practical question of whether imperfect real-world biomarker data can still support hypothesis generation, we additionally applied random forest and XGBoost as exploratory tools for prediction-oriented feature prioritization and evaluated uncertainty using repeated resampling and permutation-based perturbation analyses. Regression models provided an interpretable adjusted effect landscape but yielded limited statistically significant associations after adjustment. In contrast, tree-based models prioritized a small subset of clinicogenomic features with predictive information for PD-L1 TPS and TMB variability. Resampling and permutation analyses provided an internal assessment of uncertainty, rank stability, and sensitivity of these feature-prioritization signals. In small, heterogeneous real-world LUSC cohorts where conventional regression-based inference may provide limited resolution, machine learning–guided feature prioritization combined with resampling- and permutation-based uncertainty assessment may help identify candidate clinicogenomic signals associated with PD-L1 TPS category and TMB variability. These findings should be interpreted as hypothesis-generating and require validation in larger independent cohorts. 1. In a real-world lung squamous cell carcinoma cohort, conventional multivariable regression yielded limited statistically significant determinants of PD-L1 TPS category and TMB after adjustment. 2. Random forest and XGBoost consistently prioritized a small set of clinicogenomic variables as the most informative predictors of PD-L1 TPS and TMB, supported by permutation-based robustness analyses. 3. Predictive information ranking can complement p-value–centric inference for biomarker determinant discovery in small, imbalanced real-world immunogenomic datasets. 4. Stable clinicogenomic predictors identified by machine learning may help contextualize PD-L1 and TMB interpretation in LUSC and guide hypotheses for validation in independent cohorts.