ecPICK: A deep learning-enabled spatial diagnostic platform for direct ecDNA identification and clinical prognosis across pan-cancer histopathology

Extrachromosomal DNA (ecDNA) is an important driver of oncogene amplification and drug resistance; however, its clinical assessment is constrained by the high costs of sequencing and lack of spatial resolution in conventional assays. Thus, a cost-effective, clinically translatable platform is required for ecDNA quantification and localization using routine pathological samples.

We developed the deep learning framework ecPICK that identifies and localizes ecDNA in routine H&E-stained whole-slide images. The model was trained and tested using 4,280 images representing 20 different cancers. Its diagnostic efficacy was evaluated by area under the curve (AUC) analysis, and its spatial accuracy was verified via fluorescent in situ hybridization (FISH). In addition, the tumor microenvironment associated with ecDNA was examined by combining ecPICK with spatial transcriptomics.

ecPICK showed strong agreement with FISH-validated ecDNA levels (R² = 0.85), with a strong pan-cancer AUC of 0.789. Among the clinical cohorts, ecPICK identified ecDNA as an independent prognostic predictor beyond detection. Based on spatial research, ecDNA-rich areas preserve a unique microenvironment marked by suppressed immune cell function, dense collagen deposition, and alterations in mitochondrial metabolism.

ecPICK provides a scalable, budget-conscious platform for ecDNA mapping without the need for high-cost sequencing. By revealing the spatial remodeling of the tumor landscape, it represents a powerful tool for rapid patient stratification and novel insights into ecDNA-mediated malignant progression.