Cancer cells sometimes carry their most important oncogenes not on chromosomes but on small rings of DNA — extrachromosomal DNA (ecDNA) — that replicate independently, get divvied out unevenly at each cell division like bacterial plasmids, and rewire gene regulation by dragging distant enhancers next to oncogene promoters. In high-risk medulloblastoma, 18% of patients harbor such circles, and their presence alone makes a patient roughly three times as likely to die within five years (Hazard Ratio ≈ 3.0, $P < 0.005$) (Circular extrachromosomal DNA promotes tumor heterogeneity in high-risk medulloblastoma). Yet the statistical machinery we use to infer how mutations cooperate during tumor evolution — the tools that quantify oncogenic synergy — assumes a strictly branching family tree of clones. A segregating, gainable, losable circle is precisely the thing that tree is not. Meanwhile, that same medulloblastoma work profiles only total overexpression of circle-borne genes; it never asks which isoform of those genes the tumor makes, even though alternative splicing — the process by which a single gene's RNA is cut and rejoined into different protein variants, including pro- versus anti-apoptotic versions of the same gene — is a major determinant of oncogenic potency.
The proposal here connects the three directions with unequal weight. Alternative splicing regulation is the mechanism: I claim circular architecture changes isoform output of ecDNA-borne transcripts. Oncogenic synergy is the phenomenon and the inferential target: splicing alterations on circles should interact epistatically with chromosomal drivers. Horizontal gene transfer is the methodological lens and constraint: ecDNA behaves in somatic evolution like a mobile plasmid in microbial populations, so the population genetics of plasmids — and the reticulate, network-shaped rather than tree-shaped histories they produce (Analysis of plasmid genes by phylogenetic profiling and visualization of homology relationships using Blast2Network) — is the right template for fixing the tree models.
Core Claim
Two claims, one mechanistic and one inferential. Mechanistically: ecDNA amplicons are not just copy-number amplifiers; their circular topology, junction-flanked exons, and enhancer-hijacked promoters alter splicing outcomes — for example by internal 3' UTR splicing that excises microRNA target sites, a mechanism already shown to accelerate oncogenic translation on linear chromosomes (Chan et al.) — and these isoform changes contribute fitness. Inferentially: because ecDNA segregates non-Mendelianly, forcing it into bifurcating clonal trees corrupts epistasis estimates in the same way that ignoring tree structure does. GeneAccord showed that naive independence tests on clonal data inflate false positives to nearly 50% while tree-aware tests hold it to the expected 5%; my claim is that ecDNA-containing tumors inflict a structurally identical distortion on the tree-aware tests themselves. No work in the file bridges this: FiTree fits pairwise fitness landscapes to mutation trees but its authors note unmodeled structural jumps distort estimation; the medulloblastoma ecDNA paper never profiles splicing and never fits epistatic models; and SF3B1 in PDAC demonstrates splicing-factor synergy with KRAS/p53 but entirely within the vertical, chromosomal paradigm.
Research Program
The program has three components. First, a generative model — call it Reticulate-FiTree — that extends the Bayesian multi-type branching process of FiTree with a segregating-element layer. Each clone carries chromosomal mutations as irreversible branch events, as before, plus an ecDNA state: absent, present with copy number $c$, or lost, with transitions governed by plasmid-style mitotic partition dynamics rather than fixed mutation rates. Fitness is
$$f(\mathbf{x}, c) = \exp\Big(\sum_i s_i x_i + \sum_{i<j} s_{ij} x_i x_j + \sum_i e_i x_i c + \beta\, \Delta\Psi_{\text{ecDNA}}\Big)$$
where $x_i$ indicates chromosomal mutation $i$ (splicing-factor mutations included as candidates for $s_{ij}$ with mini-drivers), $s_i$ and $s_{ij}$ are the cell-autonomous and pairwise epistatic terms as in FiTree, $e_i$ captures interaction between a chromosomal lesion and ecDNA presence, and $\Delta\Psi_{\text{ecDNA}}$ is an isoform-shift term on circle-borne transcripts. The point of the last term is to make the splicing mechanism falsifiable: if isoform changes on circles carry no fitness signal, its posterior should concentrate near zero.
Second, a splicing atlas of ecDNA-borne transcripts. Using the medulloblastoma cohorts (481 tumors, St. Jude Cloud, ICGC, CBTN) plus long-read RNA sequencing of ecDNA-positive cell lines, we identify reads that cross the circle's unique junction — the signature that a transcript was produced from the amplicon, not the linear locus — and quantify isoform structure per circle. Candidate mechanisms to score explicitly: internal 3' UTR splicing removing microRNA sites, per the 3USP atlas of Chan et al.; dominant-negative truncations, on the model of the truncated NOS isoforms that poison full-length enzymes described by Stasiv et al.; and junction-proximal exon inclusion driven by enhancer hijacking, since splice-site strength dictates whether exons depend on regulatory RNA-binding proteins to be included, as Choi et al. established, and 3D RNA structure recognition matters, per Fish et al.. Precedent for large genetic units moving laterally in eukaryotes exists — whole chromosomes transfer between asexual fungal pathogens (Frequent horizontal chromosome transfer between asexual fungal insect pathogens) — so the mobile-element analogy is not exotic; what is missing is the post-transfer regulatory readout, exactly the gap that fungal study leaves open.
Third, a perturbation arm. In ecDNA-positive medulloblastoma lines we modulate splicing pharmacologically — CLK1 inhibitors, which shift isoforms (Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching) or the PP1-dependent axis unmasked by homoharringtonine (Sun et al.) — and measure whether ecDNA-positive and ecDNA-negative sibling cells differ in isoform response and viability. The in vivo benchmark for a splicing factor acting only as an accelerator, never alone, is the PDAC result: $Sf3b1^{K700E/+}$ by itself produced no disease, but in KPC mice it cut survival from 64 to 57 days and raised pre-cancerous progression in KC mice from 10% to 64% at 43 weeks (Mutant SF3B1 promotes malignancy in PDAC). We test whether ecDNA presence is another such context — echoing the established finding that a gene's effect flips with its genetic neighborhood, as HES5 does between MYC- and AKT-driven liver tumors (Luiken et al.).
Proposed Experiments
The simulation study is the decisive inferential test. We extend the simulation machinery of GeneAccord and FiTree with ecDNA segregation, generate synthetic tumors with known $s_i$, $s_{ij}$, and $e_i$, and ask whether tree-only methods misattribute fitness when circles are present. The reported calibration numbers give the scale to beat: naive tests reach roughly 50% false positives and tree-aware GeneAccord restores 5% in circle-free simulations (Statistical tests for intra-tumour clonal co-occurrence and exclusivity); no reported result found on Lacuna for FiTree's numerical accuracy on real (as opposed to simulated) AML trees, so its published superiority is qualitative — "consistently outperformed" SCIFIL, fitclone, and diffusion approximation. Success for Reticulate-FiTree means false-positive rates within the nominal 5% band in ecDNA-containing simulations where tree-only inference measurably inflates, and recovery of $e_i$ with tighter credible intervals than FiTree's already-wide epistatic intervals.
On real data we fit the model to the medulloblastoma cohorts, asking whether the isoform term $\Delta\Psi_{\text{ecDNA}}$ adds prognostic power beyond ecDNA presence alone, whose reported multivariate Hazard Ratio of ≈ 3.0 is the reference point (Circular extrachromosomal DNA promotes tumor heterogeneity in high-risk medulloblastoma). We additionally re-fit on the 123-patient single-cell AML cohort (Morita et al., via GeneAccord; FiTree) to test whether splicing-factor mutations — the SRSF-family lesions catalogued in Function of serine/arginine-rich splicing factors in hematopoiesis and hematologic malignancies — show epistasis with the weak-selection mini-drivers that CN/CS-calculator showed collectively rival major drivers; no existing pairwise analysis has tested this.
Baselines
The naive reference point is the $2\times2$ Fisher's exact test on cohort mutation co-occurrence, ignoring clonal structure; on clonal data its false-positive rate reaches nearly 50%, versus 5% for the tree-aware GeneAccord (Statistical tests for intra-tumour clonal co-occurrence and exclusivity), so it functions here mainly as the floor. The strongest competitor for the inferential question is FiTree itself (Bayesian inference of fitness landscapes via tree-structured branching processes), which fits exactly the fitness parameters we care about on exactly the kind of mutation trees our data provide; we will run it on identical inputs, with identical priors and MCMC budgets, and beat it only if Reticulate-FiTree stays calibrated in ecDNA-containing simulations where FiTree inflates false positives or biases $s_{ij}$ — the comparison is deliberately stacked against us on circle-free data, where both should agree. For the driver-prediction side, the strongest baseline is GLIMS (GLIMS: A two-stage gradual-learning method for cancer genes prediction using multi-omics data and co-splicing network), which already integrates alternative splicing into driver discovery and reports pan-cancer accuracy 0.858, AUROC 0.904, AUPRC 0.768; we will re-run it on the same TCGA splits with identical tuning, and count ourselves as winning only if ecDNA-augmented features improve AUROC beyond 0.904 out-of-fold. A second driver baseline is the hypergraph random-walk method PDRWH (A novel hypergraph model for identifying and prioritizing personalized drivers in cancer), whose top-ranked precision of 88.4% in breast cancer we would compare against at matched rank cutoffs. The key ablation is Reticulate-FiTree with the ecDNA layer disabled and with the $\Delta\Psi_{\text{ecDNA}}$ term disabled, run on identical data; the first isolates the reticulate correction and the second isolates the splicing mechanism. For isoform prediction, CellSpliceNet's Spearman $\rho = 0.88$ on held-out C. elegans data (CellSpliceNet) is a calibration reference rather than a fair head-to-head, since our mammalian ecDNA isoform benchmark does not yet exist — part of this proposal is creating it, and we will report CellSpliceNet's architecture trained on that new benchmark as the starting point, noting its authors' own caveat that cross-organism transferability is unverified.
Why not just…?
Why not just run GeneAccord or FiTree on the ecDNA tumors as-is? This is the sharpest objection, because it requires no new method. The evidence against it is internal to those methods: GeneAccord's own simulations show that ignoring the true dependence structure of clones nearly decuples the false-positive rate (Statistical tests for intra-tumour clonal co-occurrence and exclusivity), and FiTree's authors flag unmodeled structural jumps as a distortion source (Bayesian inference of fitness landscapes via tree-structured branching processes). ecDNA is a structural jump generator: a circle can be lost in one division and amplified twenty-fold in the next, so a clone "losing" a branch event is incoherent under their models. Our first experiment settles this empirically — if tree-only inference stays calibrated in ecDNA-containing simulations, the objection wins and the reticulate layer is unnecessary; we will report that result either way.
Why not just use AmpliconArchitect and differential expression, as the medulloblastoma study did? That pipeline is proven and cheap, but it measures total transcript abundance, and the mechanisms at issue are invisible at that resolution: internal 3' UTR splicing changes protein output and cytoplasmic export without changing steady-state RNA levels (Chan et al.), and dominant-negative truncated isoforms change function while gross expression looks unchanged (Stasiv et al.). The junction-spanning read analysis in our splicing atlas is the direct test: if circle-derived transcripts show the same exon structure as linear ones at matched copy number, this objection wins and the mechanistic claim dies.
Why not just use sequence-only splicing predictors (SpliceAI and relatives)? Cheaper and well-validated. But sequence is identical across cell types while isoform choice is not; CellSpliceNet's ablations show removing RNA secondary-structure or RNA-binding-protein expression modalities costs at least 0.06 Spearman, and it outperformed all four sequence-only baselines it tested (CellSpliceNet). On circles the problem is worse: the junction creates sequence contexts no linear model has seen, and splicing-factor dependence hinges on splice-site thermodynamic strength (Choi et al.). Still, we treat sequence-only models as a genuine threat and will run them on the new benchmark explicitly.
Why not just characterize isoforms empirically and skip the fitness modeling? Honest answer: this is a real threat to the modeling half, and the empirical atlas is valuable on its own. But the atlas alone cannot say whether isoform changes on circles are selected or incidental, and selection is what connects to oncogenic synergy. The $\Delta\Psi_{\text{ecDNA}}$ posterior — concentrated at zero or not — is precisely the experiment that adjudicates between "interesting description" and "fitness-relevant mechanism."
Decisive Checks
The proposal stands or falls on three numbers. First, in ecDNA-containing simulations, the false-positive rate of tree-only epistasis inference relative to the 5% calibration GeneAccord achieves without circles. Second, whether junction-spanning transcripts in ecDNA-positive tumors show isoform structures (3USP events, truncations, junction-proximal exon changes) at rates exceeding matched linear loci. Third, whether the isoform term improves survival prediction beyond the ecDNA Hazard Ratio of ≈ 3.0 already reported. If all three go the wrong way, both the mechanistic and inferential claims fail cleanly, and the field learns that circles amplify and rewire enhancers but leave splicing — and tree-based inference — untouched.
References
- Circular extrachromosomal DNA promotes tumor heterogeneity in high-risk medulloblastoma
- Analysis of plasmid genes by phylogenetic profiling and visualization of homology relationships using Blast2Network
- Pan-cancer pervasive upregulation of 3′ UTR splicing drives tumourigenesis
- Statistical tests for intra-tumour clonal co-occurrence and exclusivity
- Bayesian inference of fitness landscapes via tree-structured branching processes
- Mutant SF3B1 promotes malignancy in PDAC
- Regulation of multimers via truncated isoforms: a novel mechanism to control nitric-oxide signaling
- Relative strength of 5’ splice-site strength defines functions of SRSF2 and SRSF6 in alternative splicing of Bcl-x pre-mRNA
- A prometastatic splicing program regulated by SNRPA1 interactions with structured RNA elements
- Frequent horizontal chromosome transfer between asexual fungal insect pathogens
- Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching
- Homoharringtonine regulates the alternative splicing of Bcl-x and caspase 9 through a protein phosphatase 1-dependent mechanism
- NOTCH target gene HES5 mediates oncogenic and tumor suppressive functions in hepatocarcinogenesis
- Function of serine/arginine-rich splicing factors in hematopoiesis and hematopoietic malignancies
- Identification of cancer mini-drivers by deciphering selective landscape in the cancer genome
- GLIMS: A two-stage gradual-learning method for cancer genes prediction using multi-omics data and co-splicing network
- A novel hypergraph model for identifying and prioritizing personalized drivers in cancer
- CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons