Abstract
DNA-methylation classiers assign each brain tumour to one of a xed set of methylation classes, discarding where it sits within and between those classes the structure most relevant to progres- sion. We instead t the methylation manifold as an epigenetic poten- tial eld U (z) = − log p(z), a Gaussian-mixture log-density whose basins are the established methylation regimes (supervised from Ceccarelli la- bels), and derive three quantities a class label cannot express: basin stability (curvature, depth, escape barriers), a counterfactual perturba- tion response, and centrally a continuous erosion coordinate ε along the documented G-CIMP-high→low demethylation axis. The construc- tion is a deterministic, cross-language parity-proven kernel (TypeScript ≡ Python to 10−6) behind a candidate-tier governance rewall that abstains o-support and emits a nomination, not a diagnosis. On 685 TCGA-GBMLGG tumours (450K methylation, RNA-seq, copy-number, survival) the eld recovers the regimes at ARI 0.914, matching a super- vised classier (0.911). Within IDH-mutant non-codel glioma, adding ε to the discrete class and to age+grade lifts genuinely out-of-sample survival concordance 0.661 → 0.717 (optimism-corrected ∆C +0.041; adjusted HR 5.1, 95% CI 2.311.3); a model-free KaplanMeier split conrms it (log-rank p=5.4 × 10−4), and the eect survives a 203-setting abla- tion (79.8% preserved), a permutation null, and half-split replication. A cross-platform expression proxy does not replicate an honest neg- ative localising the signal to methylation topology. Read as a Langevin system, G-CIMP-high faces only a 0.91-nat barrier into the deep Codel basin that 1p/19q codeletion genetically forbids a falsiable account of why the observed erosion is slow.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMMCA2026_009.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=XZMYC2iZmY
BibTex
@InProceedings{AmbMur_The_MICCAISAT2026,
author = { Ambati, Murari},
title = { { The Glioma Epigenome as a Potential Landscape: A Continuous G-CIMP Erosion Coordinate that the Methylation Class Discards } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}
