Curated neurodegeneration axis signals are sensitive to cell composition and analytic choices: an adversarial re-analysis of a transcriptomic scoring pipeline
- 1Maulana Azad College
Abstract
Background: Curated gene-set ("axis") scores are widely used to summarise disease biology in postmortem brain transcriptomes. In bulk tissue such scores can also change with cell-type composition, with broad transcriptome shifts in diseased tissue, or with analysis choices made after the results were seen. We tested these explanations on our own published pipeline, NeuroFate, whose preprint reported an Alzheimer's disease (AD) "neuronal vulnerability" axis and a Parkinson's disease (PD) "divergent" synuclein-mitochondrial axis. Results: A code audit found that the AD discovery axis averaged Sst, Pvalb and Lamp5 interneuron cell-type proportions together with its gene-expression features; that 90 endpoint x axis tests were run before the endpoints were locked; that axes were defined from the cohorts later cited as discovery evidence; and that the PD-divergent label came from an axis-specific rule defined after the result. Their contribution to the AD association could not be quantified because the donor-level data were not archived. In GSE7621 (16 PD, 9 controls) the synuclein-mitochondrial association reproduced (exact p = 0.0011) but depended on the scoring method (exact p from 0.001 to 0.057; with per-gene standardisation no axis reached Benjamini-Hochberg significance), lay at the 2.4th percentile of 10,000 expression-matched random probe sets (two-sided p = 0.06), and was attenuated by 76% (95% interval 26% to >100%) after adjustment for a dopaminergic-marker expression proxy (PD coefficient -1.13 to -0.27 SD). A small laser-captured cohort (GSE20141; 10 PD, 8 controls) was inconclusive: no axis reached p < 0.05, and the synuclein-mitochondrial effect had the same sign. In 1,000 simulated cohorts, a naive test called a purely composition-driven gene set significant in 82% of cohorts, whereas the proposed battery classified it as composition-independent in 5% while recovering a planted within-cell effect in 98% (moderate effect) and 50% (weak effect) of cohorts. Conclusions: The original axis associations do not establish composition-independent disease programs; their interpretation is sensitive to feature construction, scoring and adjustment for cell composition. Curated axis scores in bulk brain tissue should be reported with competitive null distributions, composition adjustment, exact small-sample inference, and endpoints and scoring fixed before analysis. We provide code for these checks.
Related
- Integrated Transcriptomic and Structural Prioritization of Sulforaphane Targets in Neurodegeneration: A Multi-Dataset In Silico Study
- TMEM55B downregulation promotes neurodegeneration in Alzheimer disease models.
- Progressive neurodegeneration of the locus coeruleus drives rapid eye movement sleep fragmentation
- 5-HT1A Antagonist WAY-100635 Prevents Chronic Neurodegeneration and Immune Activation After TBI
- The human-specific lncRNA HD-LINC drives lipid droplet-associated neurodegeneration in Huntington's disease
BibTeX
@misc{ghosh2026curated,
title = {{Curated neurodegeneration axis signals are sensitive to cell composition and analytic choices: an adversarial re-analysis of a transcriptomic scoring pipeline}},
author = {N. Ghosh and K. Sinha},
howpublished = {bioRxiv},
year = {2026},
month = may,
doi = {10.64898/2026.05.23.727380},
url = {https://www.biorxiv.org/content/10.64898/2026.05.23.727380v2},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Curated neurodegeneration axis signals are sensitive to cell composition and analytic choices: an adversarial re-analysis of a transcriptomic scoring pipeline
AU - N. Ghosh
AU - K. Sinha
PY - 2026
DA - 2026/05/23
PB - bioRxiv
DO - 10.64898/2026.05.23.727380
UR - https://www.biorxiv.org/content/10.64898/2026.05.23.727380v2
AB - Background: Curated gene-set ("axis") scores are widely used to summarise disease biology in postmortem brain transcriptomes. In bulk tissue such scores can also change with cell-type composition, with broad transcriptome shifts in diseased tissue, or with analysis choices made after the results were seen. We tested these explanations on our own published pipeline, NeuroFate, whose preprint reported an Alzheimer's disease (AD) "neuronal vulnerability" axis and a Parkinson's disease (PD) "divergent" synuclein-mitochondrial axis. Results: A code audit found that the AD discovery axis averaged Sst, Pvalb and Lamp5 interneuron cell-type proportions together with its gene-expression features; that 90 endpoint x axis tests were run before the endpoints were locked; that axes were defined from the cohorts later cited as discovery evidence; and that the PD-divergent label came from an axis-specific rule defined after the result. Their contribution to the AD association could not be quantified because the donor-level data were not archived. In GSE7621 (16 PD, 9 controls) the synuclein-mitochondrial association reproduced (exact p = 0.0011) but depended on the scoring method (exact p from 0.001 to 0.057; with per-gene standardisation no axis reached Benjamini-Hochberg significance), lay at the 2.4th percentile of 10,000 expression-matched random probe sets (two-sided p = 0.06), and was attenuated by 76% (95% interval 26% to >100%) after adjustment for a dopaminergic-marker expression proxy (PD coefficient -1.13 to -0.27 SD). A small laser-captured cohort (GSE20141; 10 PD, 8 controls) was inconclusive: no axis reached p < 0.05, and the synuclein-mitochondrial effect had the same sign. In 1,000 simulated cohorts, a naive test called a purely composition-driven gene set significant in 82% of cohorts, whereas the proposed battery classified it as composition-independent in 5% while recovering a planted within-cell effect in 98% (moderate effect) and 50% (weak effect) of cohorts. Conclusions: The original axis associations do not establish composition-independent disease programs; their interpretation is sensitive to feature construction, scoring and adjustment for cell composition. Curated axis scores in bulk brain tissue should be reported with competitive null distributions, composition adjustment, exact small-sample inference, and endpoints and scoring fixed before analysis. We provide code for these checks.
ER -