We end our study by speculating that the same principles - over-dispersion being predictive of ecDNA - may extend beyond genomics assays. We show that scAmp can accurately detect ecDNA from DNA FISH on fixed patient samples, a routine clinical assays to detect amplifications.

Importantly, scAmp enables retrospective single-cell analyses of ecDNA in patient datasets. We analyzed 73 patient tumors from TCGA, including one GBM sample that showed subclonal evolution of ecDNA, with each subclone showing distinct chromatin accessibility states.

scAmp is accurate on both simulated and cell line data, and also revealed new ecDNA biology: e.g., a class of chromosomal amps that likely derived from ecDNA (e.g., BT474 below). While scAmp accurately classifies these, bulk assays cannot because their sequences resemble ecDNA.

scAmp's algorithm is based on a key feature of ecDNA copy-number distributions: namely that they are over-dispersed compared to chromosomal amplifications. Based on this observation, we trained models to predict ecDNA status directly from single-cell copy-number distributions.