ecPath detects ecDNA in tumors from histopathology images

ABSTRACT

Circular extrachromosomal DNA (ecDNA) can drive tumor initiation, progression and resistance in some of the most aggressive cancers and is emerging as a promising anti-cancer target. However, detection currently requires costly whole-genome sequencing (WGS) or labor-intensive cytogenetic or FISH imaging, limiting its application in routine clinical diagnosis. To overcome this, we developed ecPath (ecDNA from histopathology), a computational method for predicting ecDNA status from routinely available hematoxylin and eosin (H&E) images. ecPath implements a deep-learning method we call transcriptomics-guided learning, which utilizes both transcriptomics and H&E images during the training phase to enable successful ecDNA prediction from H&E images alone, a task not achievable with models trained on H&E images only. It is trained on more than 6,000 tumor whole-slide images from the TCGA cohort with the best performance in predicting ecDNA status in brain and stomach tumors (average AUC=0.78). ecPath revealed that ecDNA-positive tumors are enriched with pleomorphic, larger and high-density nuclei. Testing in an independent cohort, ecPath predicted ecDNA status of 985 pediatric brain tumor patients with an AUC of 0.72. Finally, we applied ecPath to identify ecDNA-positive tumors in the TCGA cohort for which no WGS data were available. Like WGS-based ecDNA-positive labels, the predicted ecDNA-positive status also identify poor prognoses for low grade glioma patients. These results demonstrate that ecPath enables the detection of ecDNA from routinely available H&E imaging alone and help nominate aggressive tumors with ecDNA to study and target it.