Dissecting AI-based mutation prediction in lung adenocarcinoma: A comprehensive real-world study

Introduction: Molecular profiling of lung cancer is essential to identify genetic alterations that predict response to targeted therapy. While deep learning shows promise for predicting oncogenic mutations from whole tissue images, existing studies often face challenges such as limited sample sizes, a focus on earlier stage patients, and insufficient analysis of robustness and generalizability. Methods: This retrospective study evaluates factors influencing mutation prediction accuracy using the large Heidelberg Lung Adenocarcinoma Cohort (HLCC), a cohort of 2356 late-stage FFPE samples. Validation is performed in the publicly available TCGA-LUAD cohort. Results: Models trained on the larger HLCC cohort generalized well to the TCGA dataset for mutations in EGFR (AUC 0.76), STK11 (AUC 0.71) and TP53 (AUC 0.75), in line with the hypothesis that larger cohort sizes improve model robustness. Variation in performance due to pre-processing and modeling choices, such as mutation variant calling, affected EGFR prediction accuracy by up to 7 %. Discussion: Model explanations suggest that acinar and papillary growth patterns are critical for the detection of EGFR mutations, whereas solid growth patterns and large nuclei are indicative of TP53 mutations. These findings highlight the importance of specific morphological features in mutation detection and the potential of deep learning models to improve mutation prediction accuracy. Conclusion: Although deep learning models trained on larger cohorts show improved robustness and generalizability in predicting oncogenic mutations, they cannot replace comprehensive molecular profiling. However, they may support patient pre-selection for clinical trials and deepen the insight in genotype-phenotype relationships.