Background and ObjectiveThe identification of actionable variants is fundamental for guiding critical therapeutic decisions in the management of an increasing number of cancer types, including non-small cell lung cancer. However, conventional DNA-based genetic testing methods are expensive and time-consuming. In this study, we test the hypothesis that morphological heterogeneity in tumor cell populations reflects the level of genetic variation, which in turn provides information about the presence of mutations. This study aims to harness tumor heterogeneity derived from Quantitative Pathology Features (QPFs) for the classification of known (mutation or fusion) and unknown drivers in lung cancer patients.MethodsThe analysis employed a comprehensive dataset from Foundation Medicine Inc., consisting of 2422 whole slide image (WSI) slides. This dataset was divided into a training set (60%) and two distinct testing sets (20% each). Random Field-Of-View (FOV) images were extracted from tumor regions in the WSI slides. These FOV images were further refined using Reinhard color normalization technique prior to feature extraction. QPFs characterizing individual tumor cell nuclei were then extracted and aggregated at the WSI level, providing a comprehensive tumor morphology information for each slide. The WSI level features were used to compute four different heterogeneity metrics for each QPF, namely, Standard Deviation, Kolmogorov-Smirnov statistics, Quadratic Entropy and Outlier Percentage. These morphological heterogeneity metrics for all QPFs were concatenated and used to train and test the XGBoost model for mutation detection. Model optimization was achieved by performing a five-fold cross-validation technique on the training set, while the model's stability was ascertained using two separate testing sets. The model results were cross-referenced with ground-truth labels detected through an orthogonal mutation detection method, next-generation sequencing (NGS). Model performance was assessed using the area under the precision-recall curve (PR-AUC) due to class imbalance in the data distribution.ResultsFindings revealed heterogeneous patterns indicative of different genetic drivers, with driver oncogenes (KRAS, EGFR, BRAF mutations, and ALK, ROS1, RET fusions) showing distinctive heterogeneity profiles compared to tumor suppressors. Notably, remarkable predictive performance was achieved, with a PR-AUC of 0.88 and 0.85 respectively on two separate test sets. This signifies the high precision-recall capacity of our model in distinguishing classes of genetic drivers based on tumor heterogeneity. ConclusionsOur findings demonstrate the potential of quantitative pathology-based tumor heterogeneity to classify genomic drivers. It reveals that the heterogeneity inherent in tumors provides a valuable source of information for characterizing the molecular landscape of cancer. Importantly, this method can potentially enable more accurate and efficient diagnosis of cancer, as well as guide treatment strategies based on the detected mutations. This work sheds light on the underexplored connection between tumor heterogeneity and genomic alterations, underscoring the potential of morphological heterogeneity to serve as a novel predictive biomarker.

Learning Objective

1. Quantitative pathology-based tumor heterogeneity has potential to serve as a novel predictive biomarker for mutation detection.