Abstract Cancer cells often carry large gene amplifications that can arise through very different structural mechanisms. In some tumors, amplified oncogenes reside on circular extrachromosomal DNA, or ecDNA. In others, the same oncogenes reside within chromosomes in homogeneously staining regions, or HSRs. While HSRs propagate through conventional chromosomal segregation, ecDNA replicates and divides unevenly, creating extreme variation in oncogene dosage from cell to cell. These architectural differences are increasingly linked to aggressive tumor behavior, therapeutic resistance, and rapid evolution, yet their functional and molecular consequences remain poorly understood. Traditional bulk sequencing cannot distinguish ecDNA from HSRs, and imaging based detection has relied on manual, low throughput workflows, limiting our ability to assess amplification architecture across cell populations. To address this challenge, we developed AI based imaging tools that automatically identify ecDNA and HSRs in fluorescence microscopy images and classify amplification architecture across hundreds of nuclei per experiment. In parallel, we created computational approaches that infer amplification architecture directly from single cell sequencing data, using the characteristic copy number patterns generated by ecDNA versus HSR based amplification. These automated tools provide scalable, reproducible structural profiling of cancer cells. We integrated these structural assignments with 10x Genomics single cell multiome data and BioSkryb ResolveOME, which provides whole genome and whole transcriptome profiles from the same individual cells. We applied this framework to six human cancer cell lines, three dominated by ecDNA and three dominated by HSR amplifications. Across these models, the frequently amplified oncogenic locus PVT1 offered a shared point of comparison. The data revealed striking differences between the two amplification types: Genes amplified on ecDNA showed broad spreads of expression and distinct isoform usage, including a consistent PVT1 isoform whereas HSR amplifications produced tighter, more predictable transcriptional profiles. These findings suggest that ecDNA enables tumors to access a wider range of transcriptional states that may support bet hedging and drug resistant phenotypes. HSRs may stabilize expression programs, emerging through ecDNA reintegration under selective pressure. Together, these results demonstrate how AI enabled imaging and computational inference, combined with single cell multiomics, can uncover the hidden architecture of oncogene amplification in cancer and link it directly to transcriptional output. This framework provides a scalable path for understanding how genome organization drives tumor evolution, therapeutic resistance, and cancer aggressiveness. Citation Format: Yue Wang, Jingting Chen, Oliver Cope, Aarav Mehta, Dalia Fleifel, Christina Gutierrez-Ford, Poorya Behnamie, Santiago Haase, Saygin Gulec, Timothy C. Elston, Philip M. Spanheimer, Caroline Tomblin, Alison Rojas, Tia Tate, Jeremy E. Purvis, Jeremy Wang, Joseph M. Dahl, Sam Wolff, Jean Cook, Elizabeth C. Brunk. AI enabled imaging and single cell multiomics reveal how gene amplification architecture shapes gene expression abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6796.
Wang et al. (Fri,) studied this question.