Dynamic evolution of oncogene amplification across 80,000 cancer cell genomes

Copy-number (CN) amplification is a major mechanism of oncogene activation and a therapeutic target, yet its evolution in human cancers is incompletely understood. We analyzed single-cell whole-genome sequencing (scWGS) data from >80,000 cancer cells in 100 tumors from major cancer types (ovary, breast, lung, and brain) and experimental models to resolve mechanisms and evolution of oncogene amplification. CN distributions across cells revealed two patterns: narrow, uniform peaks from symmetric segregation, consistent with intrachromosomal amplifications (ICamps), and broad, heavy-tailed variation with extreme outliers (>100 copies/cell) indicative of extrachromosomal circular DNA (ecDNA). A probabilistic mixture model of these distributions classified 74 (15%) of 503 amplified regions as ecDNA. These ecDNAs most frequently involved MYC , EGFR , and MDM2 , and were enriched in glioblastoma and lung cancers, whereas ovarian and triple-negative breast cancers predominantly showed ICamps. Notably, ICamp events showed significant subclonal specificity, with 227 (53%) of 429 events displaying multiple modes in the CN distribution. These modes were congruent with other genomic features from phylogenetic analysis, suggesting lineage inheritance patterns of symmetric division and clonal expansion. Diversification arose via numeric mechanisms (aneuploidy or genome doubling) and subclone-specific structural variants (e.g., breakage-fusion-bridge cycles or chromothripsis). CN modulation was bidirectional, including loss of amplified derivative chromosome via subclone-specific aneuploidy. Joint analysis of copy-number and structural variants at single-cell resolution uncovered several mechanisms of ecDNA-driven cancer evolution. First, remodeling of ecDNAs through internal rearrangements and recombination between distinct species was often observed, leading to oncogene co-selection. Second, convergent evolution via acquisition of distinct EGFR -containing ecDNAs was observed in a glioblastoma, suggesting the potential role of ecDNA loss in shaping subsequent evolution. Third, scWGS of isogenic cell lines distinguished present ecDNAs from historical ecDNAs that underwent genomic rearrangements resulting in chromosomal re-integration. Finally, we show that comparison between circular genome graph-based prediction versus the CN distribution-based prediction of ecDNAs revealed substantial discrepancy with notable tissue-type specificity. In conclusion, this study reveals interpretable distributions of oncogene amplifications consistent with distinct ICamp and ecDNA generative processes. This approach, refining ecDNA identification beyond standard genome graphs, further elucidates how distinct mechanisms of oncogene amplification diversify cancer cell populations. We suggest these new insights will inform patient selection for emerging ecDNA-directed therapies.