arxiv: 2604.06569 · v1 · submitted 2026-04-08 · 🧬 q-bio.GN

Recognition: no theorem link

ECLIPSE: A Composable Pipeline for Predicting ecDNA Formation, Evolution, and Therapeutic Vulnerabilities in Cancer

Bryan Cheng, Jasper Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:28 UTC · model grok-4.3

classification 🧬 q-bio.GN

keywords ecDNAextrachromosomal DNAcancer genomicsmachine learningcausal inferencestochastic differential equationstherapeutic vulnerabilities

0 comments

The pith

Existing ecDNA benchmarks are inflated by circular reasoning, and a new pipeline predicts status from standard genomic features at AUROC 0.812 while modeling dynamics and causal vulnerabilities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that computational detection of extrachromosomal DNA in cancer has relied on flawed benchmarks where models use features that already encode knowledge of ecDNA presence, which inflates reported performance. It introduces the ECLIPSE framework with three modules that correct this by restricting prediction to independent genomic data, embedding physical constraints into dynamic simulations, and using causal methods to isolate real therapeutic targets rather than correlations. A sympathetic reader would care because ecDNA amplifies oncogenes and drives evolution in roughly 30 percent of aggressive cancers, so more reliable prediction and targeting could improve treatment selection without requiring specialized sequencing. The work shows that removing leakage, encoding domain physics, and handling confounding produce better results than complex architectures alone.

Core claim

Existing benchmarks suffer from circular reasoning where models trained on features that already require knowing ecDNA status artificially inflate performance from AUROC 0.724 to 0.967. ECLIPSE is introduced as the first methodologically sound framework comprising three modules: ecDNA-Former achieves AUROC 0.812 using only standard genomic features and demonstrates that ecDNA status is predictable without specialized sequencing; CircularODE captures ecDNA's stochastic dynamics through physics-constrained neural SDEs and reaches r greater than 0.997 on experimental data via zero-shot transfer; VulnCausal applies causal inference to identify therapeutic vulnerabilities with 80x enrichment over

What carries the argument

ECLIPSE composable pipeline whose three modules perform careful feature curation to avoid leakage in prediction, physics-constrained neural stochastic differential equations to model unique ecDNA dynamics, and causal inference to filter spurious correlations when identifying vulnerabilities.

If this is right

ecDNA status becomes predictable from routine genomic assays without specialized sequencing
Stochastic evolution of ecDNA can be simulated accurately enough for zero-shot application to new data
Therapeutic target lists can be enriched 80-fold over random by removing spurious correlations
Methodological fixes for leakage and confounding outperform added model complexity in this domain

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same leakage patterns likely appear in other cancer genomics prediction tasks that mix clinical and molecular features
The physics-constrained SDE approach could transfer to modeling other extrachromosomal or circular nucleic acid elements
Causal filtering of vulnerabilities may generalize to non-ecDNA oncogene amplifications once similar feature sets are defined

Load-bearing premise

That the standard genomic features used for prediction have no hidden dependence on ecDNA status and that causal inference correctly isolates vulnerabilities without unmeasured confounders or non-generalizable zero-shot transfer.

What would settle it

An independent test set of tumors with known ecDNA status where a model trained only on the claimed standard genomic features yields AUROC significantly below 0.812, or controlled experiments showing that the top vulnerabilities from causal analysis fail to show expected effects once potential confounding variables are measured.

Figures

Figures reproduced from arXiv: 2604.06569 by Bryan Cheng, Jasper Zhang.

**Figure 1.** Figure 1: The disconnected ecDNA analysis problem. (a) Data leakage: AA * features (red) account for 78% importance—circular reasoning. (b) Physics mismatch: Standard SDEs learn incorrect variance (0.41 vs. theory 0.25). (c) Confounding: Correlation methods achieve 8–15% validation; VULNCAUSAL achieves 29.8% [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: ECDNA-FORMER results. (a) ROC curves: ECDNA-FORMER AUROC 0.729; removing dosage improves to 0.812. (b) PR curves under 8.9% positive rate. (c) Calibration (ECE=0.131). (d) Ablation: removing dosage improves performance. lineage-varying effects yield high penalty; only genes with invariant ecDNA-specific effects achieve low penalty. Limitation: IRM assumes lineages are valid environments (Rosenfeld et al., … view at source ↗

**Figure 3.** Figure 3: CIRCULARODE and VULNCAUSAL results. (a) Trajectory prediction (MSE = 0.014). (b) Physics validation: variance tracks theory (r = 0.993). (c) External validation on Lange et al. data (r > 0.997). (d) VULNCAUSAL achieves 29.8% validation, 3.7× higher than baselines. r ∼ 0.61 regardless of λphys—physics constraints ensure biological validity rather than improving accuracy (see Appendix U). External Validation… view at source ↗

**Figure 4.** Figure 4: Per-lineage performance analysis. (a) AUROC by lineage: Performance varies substantially across cancer types, with highest AUROC in blood (0.939) and bone (0.912) lineages. Lower performance in skin (0.528) and soft tissue (0.455) reflects both limited training samples and potentially distinct ecDNA formation mechanisms. Dashed orange line indicates overall 5-fold CV performance (0.729). (b) Class distrib… view at source ↗

**Figure 5.** Figure 5: Training dynamics for all ECLIPSE modules. (a) ECDNA-FORMER: Training (orange) and validation (blue) AUROC over 100 epochs. Early stopping prevents overfitting; validation AUROC plateaus at ∼0.73. Best epochs vary by fold (4–110). (b) CIRCULARODE: MSE on trajectory reconstruction decreases rapidly, converging by epoch 30. Low gap between train/val indicates good generalization. (c) VULNCAUSAL: Prediction l… view at source ↗

read the original abstract

Extrachromosomal DNA (ecDNA) represents one of the most pressing challenges in cancer biology: circular DNA structures that amplify oncogenes, evade targeted therapies, and drive tumor evolution in ~30% of aggressive cancers. Despite its clinical importance, computational ecDNA research has been built on broken foundations. We discover that existing benchmarks suffer from circular reasoning -- models trained on features that already require knowing ecDNA status -- artificially inflating performance from AUROC 0.724 to 0.967. We introduce ECLIPSE, the first methodologically sound framework for ecDNA analysis, comprising three modules that transform how we predict, model, and target these structures. ecDNA-Former achieves AUROC 0.812 using only standard genomic features, demonstrating for the first time that ecDNA status is predictable without specialized sequencing, and that careful feature curation matters more than complex architectures. CircularODE captures ecDNA's unique stochastic dynamics through physics-constrained neural SDEs, achieving r > 0.997 on experimental data via zero-shot transfer. VulnCausal applies causal inference to identify therapeutic vulnerabilities, achieving 80x enrichment over chance and 3.7x higher validation than standard approaches by filtering spurious correlations. Together, these modules establish rigorous baselines for an emerging application area and reveal a broader lesson: in high-stakes biomedical ML, methodological rigor -- eliminating leakage, encoding domain physics, addressing confounding -- outweighs architectural innovation. ECLIPSE provides both the tools and the template for principled computational oncology.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ECLIPSE flags real circularity in prior ecDNA benchmarks and offers a three-module pipeline, but the claim of leakage-free features still needs direct verification.

read the letter

The paper's central contribution is spotting that earlier ecDNA classifiers were trained on features already tied to knowing the circular status, which pushed reported AUROCs artificially high. It then builds ECLIPSE around three named pieces: ecDNA-Former for prediction from ordinary genomic inputs, CircularODE for modeling stochastic dynamics with physics-constrained SDEs, and VulnCausal for spotting therapeutic targets via causal methods. That structure is new in this subfield and gives a clearer way to separate the tasks than previous single-model attempts. The reported numbers—0.812 AUROC on standard features, r > 0.997 on zero-shot transfer, and 80x enrichment in the causal module—look stronger than the baselines cited, and the emphasis on avoiding leakage and encoding domain constraints is a fair point for high-stakes biomedical work. The abstract makes a reasonable case that careful curation can matter more than fancier architectures here. The main soft spot is exactly the one the stress-test note raises: the abstract asserts that ecDNA-Former uses only standard features with no implicit ecDNA knowledge, yet supplies no explicit feature list, provenance check, or ablation that would let a reader confirm every input can be computed from raw WGS without any circular-aware caller. If even one CNV or amplicon-derived statistic slipped through from an ecDNA-enriched pipeline, the performance lift is partly leakage rather than genuine signal. The other modules get less scrutiny in the abstract, but the same demand for transparent equations and controls applies. This is aimed at computational oncologists who already work on ecDNA or structural variant modeling; a methods-focused reader will find the critique useful even if they end up re-running the feature checks themselves. It deserves peer review because the methodological concern is substantive and the proposed split into modules is concrete enough to test and improve.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces ECLIPSE, a composable three-module pipeline for ecDNA analysis in cancer. ecDNA-Former predicts ecDNA status using only standard genomic features and achieves AUROC 0.812, claiming to correct circular reasoning that inflated prior benchmarks from 0.724 to 0.967. CircularODE models ecDNA stochastic dynamics via physics-constrained neural SDEs and reports r > 0.997 on experimental data through zero-shot transfer. VulnCausal applies causal inference to identify therapeutic vulnerabilities, reporting 80x enrichment over chance and 3.7x higher validation rates than standard methods. The central thesis is that methodological rigor (leakage elimination, domain physics, confounding control) outweighs architectural complexity.

Significance. If the core claims hold—particularly that ecDNA-Former features are fully independent of ecDNA status and that VulnCausal correctly isolates causal vulnerabilities without unmeasured confounders—this work would supply much-needed rigorous baselines for an emerging area of computational oncology. The explicit critique of circularity in existing benchmarks and the emphasis on physics-constrained modeling represent a constructive contribution to high-stakes biomedical ML.

major comments (1)

[ecDNA-Former module (Methods and Results)] The claim that ecDNA-Former uses only leakage-free standard genomic features (AUROC 0.812) is load-bearing for the paper's central methodological contribution. No explicit feature list, ablation study, or provenance verification is supplied showing that every retained input (e.g., CNV segmentation, amplicon patterns, or mutation features) can be computed from whole-genome sequencing without any ecDNA-aware caller or circular-structure flag. If any feature implicitly encodes ecDNA status, the reported improvement over the 0.724 baseline remains consistent with residual circularity rather than genuine predictability.

minor comments (2)

[Abstract] The abstract states strong performance numbers (AUROC 0.812, r > 0.997, 80x enrichment) without accompanying details on cross-validation strategy, cohort sizes, or negative-control experiments; these should be summarized in the abstract for readers who encounter only the front matter.
[CircularODE module] The notation 'r > 0.997' for CircularODE should specify the exact correlation coefficient (Pearson, Spearman, etc.) and whether it is computed on held-out experimental time series or on simulated trajectories.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful and constructive review. The concern regarding explicit documentation of leakage-free features in ecDNA-Former is well-taken and directly relevant to the paper's central claim. We address this point below and will revise the manuscript to provide the requested transparency.

read point-by-point responses

Referee: [ecDNA-Former module (Methods and Results)] The claim that ecDNA-Former uses only leakage-free standard genomic features (AUROC 0.812) is load-bearing for the paper's central methodological contribution. No explicit feature list, ablation study, or provenance verification is supplied showing that every retained input (e.g., CNV segmentation, amplicon patterns, or mutation features) can be computed from whole-genome sequencing without any ecDNA-aware caller or circular-structure flag. If any feature implicitly encodes ecDNA status, the reported improvement over the 0.724 baseline remains consistent with residual circularity rather than genuine predictability.

Authors: We agree that an explicit feature inventory and ablation analysis are necessary to substantiate the leakage-free claim. The manuscript states that ecDNA-Former relies exclusively on standard genomic features obtainable from whole-genome sequencing pipelines (CNV segmentation via CNVkit or equivalent, mutation profiles from GATK or MuTect, and amplicon patterns from read-depth and breakpoint analysis using DELLY or Lumpy without any circular-structure or ecDNA-specific detection steps). These features are computed prior to and independently of any ecDNA status annotation. However, we acknowledge that the original submission did not include a consolidated feature table or ablation study. In the revised manuscript we will add: (1) a supplementary table enumerating every input feature, its exact computation method, and the standard tool/version used; (2) an ablation study that systematically removes each feature class (including CNV segmentation and amplicon patterns) and reports the resulting AUROC to demonstrate that performance does not collapse when potentially ambiguous signals are excluded; and (3) explicit provenance verification confirming that no retained feature requires an ecDNA-aware caller. These additions will show that the AUROC of 0.812 reflects genuine predictability from non-circular inputs, in contrast to the 0.967 figures obtained by prior work that incorporated ecDNA-dependent features. We believe this revision will fully address the referee's concern while preserving the paper's methodological contribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity in ECLIPSE derivation chain

full rationale

The paper explicitly identifies circular reasoning in prior benchmarks (features that already encode ecDNA status, inflating AUROC from 0.724 to 0.967) and positions its own modules as avoiding this by using only standard genomic features (ecDNA-Former), physics-constrained neural SDEs with zero-shot transfer (CircularODE), and causal inference to filter confounders (VulnCausal). No equations, feature lists, or self-citation chains are supplied that reduce any claimed prediction or result to a fitted input or prior self-result by construction. The reported metrics are framed as independent validations on experimental data, so the derivation remains self-contained without load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract does not specify any free parameters, axioms, or invented entities; full text would be required to identify any fitted values in the neural SDEs or assumptions in the causal model.

pith-pipeline@v0.9.0 · 5577 in / 1304 out tokens · 80465 ms · 2026-05-10T18:28:01.504136+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

doi: 10.1158/1535-7163.MCT-15-0897. King L. Hung, Kathryn E. Yost, Liangqi Xie, Quanming Shi, Konstantin Helmsauer, Jens Luebeck, Robert Sch¨opflin, Joshua T. Lange, Roc´ıo Chamorro Gonz´alez, Natasha E. Weiser, Celine Chen, Maria E. Valieva, Ivy Tsz-Lo Wong, Sihan Wu, Siavash R. Dehkordi, Connor V . Duffy, Katerina Kraft, Jun Tang, Julia A. Belk, John C....

work page doi:10.1158/1535-7163.mct-15-0897 2021
[2]

Xuechen Li, Ting-Kam Leonard Wong, Ricky T

doi: 10.1038/s41588-022-01177-x. Xuechen Li, Ting-Kam Leonard Wong, Ricky T. Q. Chen, and David Duvenaud. Scalable gradients for stochastic differential equations. InProceedings of the International Conference on Artificial Intelligence and Statistics, pp. 3870–3882, 2020. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll ´ar. Focal los...

work page doi:10.1038/s41588-022-01177-x 2020
[3]

Kristen M

doi: 10.1016/j.cell.2017.06.010. Kristen M. Turner, Viraj Deshpande, Doruk Beyter, Tomoyuki Koga, Jessica Rusert, Catherine Lee, Bin Li, Karen Arden, Bing Ren, David A. Nathanson, Harley I. Kornblum, Michael D. Taylor, Sharmeela Kaushal, Webster K. Cavenee, Robert Wechsler-Reya, Frank B. Furnari, Scott R. Van- denberg, P. Nagesh Rao, Geoffrey M. Wahl, Vin...

work page doi:10.1016/j.cell.2017.06.010 2017
[4]

validation

reference Hi-C (GM12878). Fragile Site Proxim- ity 32 Binary indicators and distances to 32 common fragile sites. Source: NCBI RefSeq (O’Leary et al., 2016). Input to fusion112 + Hi-C CNV(40) + Expr(40) + Fragile(32) = 112 scalar features; Hi-C processed separately through graph encoder. The complete list:MYC, MYCN, MYCL, EGFR, ERBB2, CDK4, CDK6, MDM2, MD...

work page arXiv 2016