Imran Noorani

Simulating multi-ecDNA dynamics in vivo.A, DNA FISH and nascent RNAscope images from the tumor of patient A5, showing the coexistence of both EGFR-ecDNA and PDGFRA-ecDNA. B, Circular structures of three distinct oncogenic ecDNAs detected in the tumor of patient A5. C, GBMs in the GB-UK and PCAWG cohorts containing more than one oncogenic species and their respective copy numbers. D, In a multiple-ecDNA application of SPECIES, the clone-initiating cell begins with k1 and k2 copies of ecDNA 1 and ecDNA 2, respectively. E, Representation of the impacts of ecDNA cosegregation and coselection on the number of inherited ecDNA copies during cell division. Cosegregation drives correlated inheritance of both ecDNA types, whereas coselection favors tumor cells carrying at least one copy of each type. F, Application of SPECIES to simulate two separate ecDNA species. Simulated tumors were initialized with a single cell, carrying k1 and k2 copies of each ecDNA species. Representative model images show examples of low and high cosegregation and coselection. G and H, Parameter inference summary for the tumor of patient A5, for which we measured both the EGFR-ecDNA and PDGFRA-ecDNA copy-number distributions using DNA FISH, using multiple-ecDNA SPECIES. Parameters sp and sm represent selection coefficients for tumor cells with 1 (pure) or 2 (mixed) ecDNA species, respectively. I, Comparison of inferred k and k1 values for all patient tumors confirmed by WGS to harbor two or more ecDNA species. J, The presence of ecDNA amplifications may be used to aid stratification, given that oncogenes on ecDNA have an inherent resistance mechanism through the ability of ecDNA to dynamically adjust copy number in response to targeted agents. Earlier monitoring and intervention are recommended for those patient tumors with ecDNA that will receive targeted therapies.

EcDNA accumulation in genetically engineered in vivo and ex vivo murine models. A, Neural stem cells were propagated within the SVZ of genetically engineered mice for 4 months. B,Myc-ecDNAs were induced in adult murine neural stem cells, which were then cultured ex vivo for 5 weeks. C, DNA FISH analysis of neural stem cells within the SVZ of genetically engineered mice (i) without Myc-ecDNA (Myc+/+; P53fl/fl; Nestin-Cre, n = 2) and (ii) with Myc-ecDNA (Mycec/+; p53fl/fl; Nestin-Cre, n = 2), both on a background of homozygous Trp53 loss. D, ecDNA copy-number dynamics in murine adult neural stem cells (data from ref. 33) with corresponding simulated dynamics, assuming either neutral or positively selected ecDNA. (A and B, created using BioRender assets. https://BioRender.com/a1hpt9n.)

Selection dynamics of EGFR SVs on ecDNA. A, A subclonal EGFRvIII variant on ecDNA in the tumor of patient A23. B, Variant heteroplasmy in the core region of all samples across the GB-UK and PCAWG cohorts, which contained EGFR variant–bearing ecDNA. C, Effects of (i) pre- and (ii) postclonal expansion mutation of EGFR-ecDNA on its resulting spatial pattern in the expanded tumor. D, SPECIES was applied to simulate the emergence of an advantageous EGFRvIII mutation on ecDNA, either prior to or after the onset of clonal expansion of the tumor. The example shown is for a pre-expansion mutation, with parameters (kwt=20, kvar=1, Vvar=0). The SD of ecDNA heteroplasmy is shown by the shaded region in the lower plot. E, Model predictions for resulting variant heteroplasmy in a tumor for a range of wild type- and variant-bearing ecDNA copy numbers in the clone-initiating tumor cell (for pre-expansion EGFR mutation) and mutation times (for post-expansion EGFR mutation). The red-shaded region denotes parameter combinations inconsistent with observed patient core heteroplasmy values. Mean ± variance is derived from 1,000 simulated tumors. F, SPECIES model predictions for the probability of observing EGFRvIII ecDNA across tumor core and margin, for a pair of core and random margin regions. Mean values are derived from 1,000 simulated tumors. G, (Top) IHC slide for the tumor of patient A20, showing the presence of EGFRvIII in the tumor core and infiltrating margin (blue = hematoxylin; brown = EGFRvIII). (Bottom) Summary of EGFRvIII in core and margins of GB-UK cohort samples, determined by IHC. H, Variant allele frequencies of point mutations within EGFR in PCAWG samples harboring EGFR-ecDNA. Red lines represent the frequency expected for a mutation carried by all ecDNAs within a sample. I, Distributions of EGFRwt-ecDNA abundance in the cell that acquires the first mutation, for a range of ecDNA mutation rates. Patient tumor values are derived from their respective inferred k values, assuming that the arrival of a mutation on ecDNA initiates clonal expansion, implying that the number of EGFRwt-ecDNA in the mutating cell is thus equal to k − 1 (with the remaining one ecDNA carrying the mutated allele). J, Timeline of EGFR-ecDNA tumorigenesis. (i) Initial EGFRwt-ecDNA is generated and accumulates in precancerous cells, conferring a moderately positive selective advantage. In time, the EGFRvIII mutation occurs on one of the accumulated EGFRwt-ecDNAs, conferring a strong positive advantage. (ii) Reconstructed ecDNA structures from the tumor of GB-UK patient A5 confirm that EGFRwt and EGFRvIII ecDNAs share common breakpoints, indicating that EGFRvIII arose from the mutation of an existing EGFRwt-ecDNA. (iii) EGFRvIII-ecDNA further promotes clonal expansion, driving tumor growth.

SPECIES spatial modeling of ecDNA evolution. A and B, DNA FISH staining of two representative GBM samples from the GB-UK cohort, revealing (A) EGFR- and (B) PDGFRA-ecDNA. EGFR-ecDNA seem to congregate into “hubs” (white arrows). Single-cell ecDNA copy-number distributions are derived from DNA FISH images using unbiased image analysis. C, (i) SPECIES is initiated with a single tumor cell containing k copies of ecDNA. (ii) Model of ecDNA selection. Cells carrying one or more copies of ecDNA divide at a rate s faster than cells lacking ecDNA. (iii) Cells push neighbors within a radius q on the lattice to make space required for cell division. D, Spatial computational model of ecDNA-driven tumors. ABC was used to infer optimal model parameters for the initial number of ecDNA, k; ecDNA-conferred selection, s; and cell pushing strength, q, for a given set of patient tumor measurements. E and F, Example output of the ABC model fitting algorithm using (E) 2 and (F) 3 spatial regions. (Left to right) Inferred k, s, and q and model best-fit distributions for each tumor region. The sum of the Wasserstein distance between patient and simulated distributions for each tumor region, representing closeness of fit, is denoted by σ. G, Comparison of inferred k, s, and q when analyzing 2 (core and margin) vs. 3 (core, margin, and leading edge) spatial regions. H, Summary of inferred model parameters for all sampled human GBM tumors. The highlighted color represents ecDNA-amplified oncogenes (EGFR, PDGFRA, or CDK4) in each tumor. I, Inferred model parameters for all patient tumors (excluding the tumor of patient A5, which contained separate EGFR- and PDGFRA-ecDNA species), stratified by ecDNA-amplified oncogene (EGFR or PDGFRA). (D, created using BioRender assets. https://BioRender.com/7f2pume)

Workflow of the study analysis. A, A total of 94 IDHwt GBMs were analyzed (35 from the PCAWG cohort and 59 from the GB-UK cohort). B, The core region of all patient tumors was analyzed using WGS, and GB-UK patient tumors were further subjected to multiregion DNA FISH and nascent RNAscope analysis to characterize spatial patterns of oncogenic ecDNAs. C, The ecDNA genetic landscape and fine structure were analyzed using WGS information, and mathematical modeling of single-cell copy number was employed to reconstruct evolutionary histories in GB-UK patient tumors. (B, created using BioRender assets. https://BioRender.com/kgae4q5; https://BioRender.com/7f2pume)

Supplementary Methods, providing details about methodology used throughout the study.