Human Islet Transcriptomics in Type 2 Diabetes Reveals Conserved Dysfunction Axes and a Candidate Gene Panel
- 1Universidad Autónoma Metropolitana Unidad Cuajimalpa
Abstract
Background.Type 2 diabetes mellitus (T2D) involves progressive pancreatic beta-cell dysfunction, but conserved islet programs require cross-cohort and single-cell validation.
Methods.We integrated four human islet transcriptomic cohorts using differential expression, random-effects meta-analysis, Hallmark scoring, and module scoring. To test generalizability without reusing held-out outcomes, modules were re-derived from three cohorts and scored in the fourth in rotating study-level cross-fitting, with expression-and detection-matched random gene-set nulls. Machine-learning performance in GSE164416 was estimated by nested stratified 5-fold cross-validation with DEG calling and feature selection repeated within each outer training fold. An annotation-compatible score was tested in GSE50244, and beta-cell validation used GSE221156.
Results.Cross-fitted IsletDysfunctionScore effects were positive in all four held-out cohorts (d = 0.39-1.36; pooled Hedges g = 0.905, 95% CI 0.501-1.309, p = 1.14 x 10^-5, I2 = 5.5%). Matched gene-set null testing supported effects beyond random expectation in two cohorts. Nested machine learning yielded AUC = 0.960 (bootstrap 95% CI 0.900-0.997); GABRA2 and ENSG00000284653 were selected in all outer folds. The external 8-gene consensus-pool score achieved AUC = 0.907 (DeLong 95% CI 0.818-0.996), but its fixed threshold did not transfer. In GSE221156, the single-cell dysfunction-score analog increased and BetaIdentitySecretion decreased in T2D; ImmuneStress was nominally increased but did not survive false-discovery correction across the single-cell score family.
Conclusions.Human T2D islets show a reproducible stress/identity-loss framework that generalizes directionally under study-level cross-fitting and beta-cell-resolved validation. The candidate panel is a research transcriptomic signature requiring further validation.
Related
- Neurovascular uncoupling in type 2 diabetes mellitus: multimodal MRI reveals the association between hippocampal and subcortical structures dysfunction and cognitive decline
- Cross-dataset transcriptomics–proteomics analysis identifies Tlr2, Anxa5, and Grin2a as candidate mouse-brain signatures associated with chronic cerebral hypoperfusion
- Spatial transcriptomics reveals interferon pathway enrichment in early relapsing serous carcinoma
- Single-nucleus transcriptomics reveals multicellular remodeling beyond the aganglionic segment in Hirschsprung disease
- High-resolution spatial transcriptomics reveals fibroblast and neuroimmune microenvironments in endometriosis lesions
BibTeX
@misc{romero2026human,
title = {{Human Islet Transcriptomics in Type 2 Diabetes Reveals Conserved Dysfunction Axes and a Candidate Gene Panel}},
author = {R. Romero},
howpublished = {medRxiv},
year = {2026},
month = jun,
doi = {10.64898/2026.06.08.26355184},
url = {https://www.medrxiv.org/content/10.64898/2026.06.08.26355184v2},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Human Islet Transcriptomics in Type 2 Diabetes Reveals Conserved Dysfunction Axes and a Candidate Gene Panel
AU - R. Romero
PY - 2026
DA - 2026/06/08
PB - medRxiv
DO - 10.64898/2026.06.08.26355184
UR - https://www.medrxiv.org/content/10.64898/2026.06.08.26355184v2
AB - Background.Type 2 diabetes mellitus (T2D) involves progressive pancreatic beta-cell dysfunction, but conserved islet programs require cross-cohort and single-cell validation. Methods.We integrated four human islet transcriptomic cohorts using differential expression, random-effects meta-analysis, Hallmark scoring, and module scoring. To test generalizability without reusing held-out outcomes, modules were re-derived from three cohorts and scored in the fourth in rotating study-level cross-fitting, with expression-and detection-matched random gene-set nulls. Machine-learning performance in GSE164416 was estimated by nested stratified 5-fold cross-validation with DEG calling and feature selection repeated within each outer training fold. An annotation-compatible score was tested in GSE50244, and beta-cell validation used GSE221156. Results.Cross-fitted IsletDysfunctionScore effects were positive in all four held-out cohorts (d = 0.39-1.36; pooled Hedges g = 0.905, 95% CI 0.501-1.309, p = 1.14 x 10^-5, I2 = 5.5%). Matched gene-set null testing supported effects beyond random expectation in two cohorts. Nested machine learning yielded AUC = 0.960 (bootstrap 95% CI 0.900-0.997); GABRA2 and ENSG00000284653 were selected in all outer folds. The external 8-gene consensus-pool score achieved AUC = 0.907 (DeLong 95% CI 0.818-0.996), but its fixed threshold did not transfer. In GSE221156, the single-cell dysfunction-score analog increased and BetaIdentitySecretion decreased in T2D; ImmuneStress was nominally increased but did not survive false-discovery correction across the single-cell score family. Conclusions.Human T2D islets show a reproducible stress/identity-loss framework that generalizes directionally under study-level cross-fitting and beta-cell-resolved validation. The candidate panel is a research transcriptomic signature requiring further validation.
ER -