Abstract
Cancer-driver discovery equates importance with recurrence, and a recent generation of spatial generative models predicts perturbation responses without identifying the causal effect of the endogenous events a tumour actually carries. We reframe a glioma driver as an event with identifiable counterfactual control over the tumour’s multi-scale ecosystem state, and show that—precisely because endogenous events are not randomised—this effect can be identified from the tumour’s own clonal mosaicism, regional heterogeneity, and longitudinal recurrence as natural experiments. We implement a single doubly-robust (AIPW) identification spine reused without modification across six biological scales—from molecular events and cell states through spatial niches, potential-field geometry, macro imaging geometry, to longitudinal recurrence—each emitting a positivity-aware contrast that abstains off-support (grade i0), a sensitivity bound, and a permutation control. The multi-scale design enables two identification-validity tools no single-scale method can have: a cross-scale over-identification test that turns channel agreement into an internal falsification of confound- ing, and a negative-control calibration of the discovery null. Every nomination passes a fail-closed evidence-tier × identifiability-grade firewall; functional anchoring (CRISPR dependency enrichment and perturbation-effect correlation) is validated on planted panels—real-data anchoring is pre-registered future work. On real glioma corpora the framework recovers the textbook cell-state architecture and the cell-state continuum; on the field-standard DLPFC benchmark its niche module beats the non-spatial standard. On planted ground truth it recovers the true effect within its interval while a confounding-blind contrast is biased, and an ablation matrix shows the causal machinery—not model capacity—does the work. We concede that a linear model predicts as well as ours: the contribution is identification, not prediction, and a calibrated refusal to answer is a first-class output.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MISO_003.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=M3qaVatTCv
BibTex
@InProceedings{AmbMur_Identifiable_MICCAISAT2026,
author = { Ambati, Murari},
title = { { Identifiable Counterfactual Control Defines Glioma Drivers: A Multi-Scale Causal Framework that Abstains Off-Support } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}