Predicting ROS1 and ALK fusions in NSCLC from H&E slides with a two-step vision transformer approach.

Non-small cell lung cancer (NSCLC) is one of the deadliest and most prevalent cancers worldwide, with 5-year survival rates of ~28%. The molecular heterogeneity within NSCLC encompasses several types of genetic alterations, such as mutations, amplifications, and rearrangements, and can drive aggressive tumor behavior and poor response to therapy. Among these genetic alterations are ALK and ROS1 fusions. Though these fusion events are relatively rare, their identification is crucial for selecting effective targeted treatments and avoiding therapies with significant side-effects. Fluorescent in situ hybridization (FISH), immunohistochemistry (IHC), and sequencing of DNA and RNA are standard methods to detect ALK and ROS1 fusions, but they are costly, time-consuming, and require adequate tumor tissue. Here we employ deep learning models using whole slide images (WSIs) of hematoxylin and eosin (H&E)-stained formalin-fixed paraffin embedded (FFPE) NSCLC tumor specimens to identify tumors most likely to harbor ALK and ROS1 fusions in a cohort of 33,014 patients, out of which 306 and 697 patients are positive for ROS1 or ALK fusions, respectively. A vision transformer model (MoCo-V3) was trained as a feature extractor, followed by training transformer-based models to predict the presence of ROS1 and ALK fusions. Due to the limited positive sample size for ROS1, a two-step specialized training procedure was implemented to enhance prediction performance during cross-validation. Our approach achieved receiver-operating characteristic areas under the curves (ROC AUCs) of 0.85 for ROS1 and 0.84 for ALK on a holdout dataset, demonstrating the effectiveness of this method. This framework holds significant potential for clinical application by offering a scalable, accurate, and cost-efficient method for detecting ALK and ROS1 fusions. Furthermore, it may serve as a pre-screening tool to identify candidates for confirmatory diagnostic testing and clinical trials, ultimately improving the efficiency of selecting appropriately targeted therapies for NSCLC patients.