Contrasting within-donor gene covariation identifies disease-associated cells and genes in case-control single-cell studies
- 1Center for Computational Biology and Bioinformatics, Department of Medical and Molecular Genetics, Indiana University School of Medicine
Abstract
In case-control single-cell studies, disease often affects only a subset of cells of a given type within each patient. Comparing all cells from patients with all cells from control donors therefore dilutes disease signals, while donor variation and recorded or unrecorded batch effects can be confounded with disease status. We present IDC (Identification of Disease-associated Cells), which uses this cellular heterogeneity as signal. A mixture of affected and unaffected cells creates coordinated variation among disease-associated genes within each patient, whereas shifts shared by all cells of a donor leave this within-donor covariance unchanged. IDC therefore contrasts within-donor gene-gene covariance between patients and controls to learn disease-associated expression programs without batch labels, then applies positive-unlabeled learning to score each cell's disease-like state and estimate each donor's disease burden. IDC also computes, in closed form, gene contribution scores and leave-gene-out cell scores that exclude each tested gene, enabling within-donor differential expression testing and cross-donor meta-analysis without retraining. In simulations, IDC identified affected cells and estimated donor burden more accurately than deep generative and label-refinement methods, remained robust to strong unrecorded batch effects, and detected differentially expressed genes more sensitively than case-control pseudobulk methods while controlling false discoveries. In human pancreatic islets and prefrontal cortex, disease burden in held-out donors separated type 2 diabetes (T2D) from non-diabetic donors in beta cells, and late-stage Alzheimer's disease (AD) from non-AD donors in astrocytes and inhibitory neurons, while placing prediabetic and early-AD donors at intermediate levels. Genes with the highest IDC contribution scores were more strongly enriched for the KEGG T2D and AD pathways than case-control pseudobulk rankings, and within-donor testing revealed coordinated metabolic, cell-adhesion and synaptic changes along the inferred disease axis.
Related
- Molecular evidence of spotted fever group Rickettsia and Lyme disease– and relapsing fever–associated Borrelia, and the absence of scrub typhus-associated Orientia in ixodid ticks from contrasting habitats in Peninsular Malaysia
- Longitudinal linear-epitope profiling of SLE identifies novel antigens associated with disease and disease activity
- Metabarcoding analysis identifies the phyllosphere fungal communities associated with a newly emerged leaf fall disease in rubber plants in Indonesia
- Cross-platform comparison of diabetic foot ulcer transcriptomes identifies partially concordant disease-associated signals
- TCF21 loss identifies a conserved inflammation-associated fibroblast state across Crohn's disease and ulcerative colitis
BibTeX
@misc{chang2026contrasting,
title = {{Contrasting within-donor gene covariation identifies disease-associated cells and genes in case-control single-cell studies}},
author = {F. Chang and Z. Duren},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.06.764894},
url = {https://www.biorxiv.org/content/10.64898/2026.11.06.764894v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Contrasting within-donor gene covariation identifies disease-associated cells and genes in case-control single-cell studies
AU - F. Chang
AU - Z. Duren
PY - 2026
DA - 2026/11/06
PB - bioRxiv
DO - 10.64898/2026.11.06.764894
UR - https://www.biorxiv.org/content/10.64898/2026.11.06.764894v1
AB - In case-control single-cell studies, disease often affects only a subset of cells of a given type within each patient. Comparing all cells from patients with all cells from control donors therefore dilutes disease signals, while donor variation and recorded or unrecorded batch effects can be confounded with disease status. We present IDC (Identification of Disease-associated Cells), which uses this cellular heterogeneity as signal. A mixture of affected and unaffected cells creates coordinated variation among disease-associated genes within each patient, whereas shifts shared by all cells of a donor leave this within-donor covariance unchanged. IDC therefore contrasts within-donor gene-gene covariance between patients and controls to learn disease-associated expression programs without batch labels, then applies positive-unlabeled learning to score each cell's disease-like state and estimate each donor's disease burden. IDC also computes, in closed form, gene contribution scores and leave-gene-out cell scores that exclude each tested gene, enabling within-donor differential expression testing and cross-donor meta-analysis without retraining. In simulations, IDC identified affected cells and estimated donor burden more accurately than deep generative and label-refinement methods, remained robust to strong unrecorded batch effects, and detected differentially expressed genes more sensitively than case-control pseudobulk methods while controlling false discoveries. In human pancreatic islets and prefrontal cortex, disease burden in held-out donors separated type 2 diabetes (T2D) from non-diabetic donors in beta cells, and late-stage Alzheimer's disease (AD) from non-AD donors in astrocytes and inhibitory neurons, while placing prediabetic and early-AD donors at intermediate levels. Genes with the highest IDC contribution scores were more strongly enriched for the KEGG T2D and AD pathways than case-control pseudobulk rankings, and within-donor testing revealed coordinated metabolic, cell-adhesion and synaptic changes along the inferred disease axis.
ER -