Uncertainty-Aware Personalised Gene Expression Prediction Improves Cell-Type-Specific Association Testing

Abstract

Linking genetic variation to disease requires accurate prediction of its effects on gene expression in relevant cell types. Sequence-to-function models offer a way to predict these effects and, with training on personal genomes, have begun to capture inter-individual expression differences. However, effect sign accuracy remains limited, and predictive uncertainty is unquantified. Here we introduce ATTICUS, which combines personalised sequence representations with uncertainty-aware expression prediction. ATTICUS transfers regulatory information learned in bulk expression pretraining to cell-type-specific expression prediction without requiring genotypes paired with single-cell expression data. Across blood and brain cell types in OneK1K and ROSMAP, ATTICUS approximately doubles expression prediction correlations relative to scPrediXcan and substantially improves sign accuracy. Moreover, ATTICUS enables uncertainty-aware transcriptome-wide association studies (TWAS). Applied to systemic lupus erythematosus and Alzheimer's disease, ATTICUS identifies more associations, with greater cell-type specificity and stronger directional concordance, and reduced statistical inflation. Together, our results demonstrate that personalised training makes predictions more directionally consistent, and uncertainty propagation improves the calibration of TWAS and the prioritisation of disease genes. ATTICUS provides a framework for both, enabling the use of single-cell expression datasets without matched genotypes to inform uncertainty-aware TWAS, expanding their use in prioritising disease-associated genes and cell types.

Mendel, a foundation model of human genetic variation, prioritizes regulatory variants and improves gene expression prediction
A. Salman, H. Feng, P. Wei +4
bioRxiv28 Sep 2026PreprintCC BYFull textGenetics & genomicsbioRxiv
Mendel, a foundation model of human genetic variation, prioritizes regulatory variants and improves gene expression prediction
Chong Wu, Arham Salman, Haonan Feng +4
Research Square26 Oct 2026PreprintCC BY 4.0Full textGenetics & genomicsPMC
Personalised prediction of anomalous bone regeneration in multiple myeloma patients
F. Gibson, R. E. Andrews, M. A. Lawson +10
bioRxiv26 Oct 2026PreprintCC BYFull textSystems & synthetic biologybioRxiv
Perturbation-based propagation of bacterial composition uncertainty improves robustness and reproducibility of microbiome differential abundance analysis.
M. E. Andrews, T. Vincennie, M. Koprivica +6
bioRxiv23 Oct 2026PreprintCC BYFull textBioinformatics & computational biologybioRxiv
Neural representation of the subjective uncertainty of prediction in the human brain
Xiaoyue Sun, Xizi Gong, Huan Luo +2
Communications Biology30 Sep 2026Peer-reviewedCC BY-NC-ND 4.0Full textNeuroscienceCrossref

Cite

BibTeX RIS

BibTeX

@misc{styll2026uncertaintyaware,
  title = {{Uncertainty-Aware Personalised Gene Expression Prediction Improves Cell-Type-Specific Association Testing}},
  author = {P. Styll and Y. Di and N. Cai},
  howpublished = {bioRxiv},
  year = {2026},
  month = nov,
  doi = {10.64898/2026.11.06.764604},
  url = {https://www.biorxiv.org/content/10.64898/2026.11.06.764604v1},
  note = {CC BY-NC},
}

RIS

TY  - UNPB
TI  - Uncertainty-Aware Personalised Gene Expression Prediction Improves Cell-Type-Specific Association Testing
AU  - P. Styll
AU  - Y. Di
AU  - N. Cai
PY  - 2026
DA  - 2026/11/06
PB  - bioRxiv
DO  - 10.64898/2026.11.06.764604
UR  - https://www.biorxiv.org/content/10.64898/2026.11.06.764604v1
AB  - Linking genetic variation to disease requires accurate prediction of its effects on gene expression in relevant cell types. Sequence-to-function models offer a way to predict these effects and, with training on personal genomes, have begun to capture inter-individual expression differences. However, effect sign accuracy remains limited, and predictive uncertainty is unquantified. Here we introduce ATTICUS, which combines personalised sequence representations with uncertainty-aware expression prediction. ATTICUS transfers regulatory information learned in bulk expression pretraining to cell-type-specific expression prediction without requiring genotypes paired with single-cell expression data. Across blood and brain cell types in OneK1K and ROSMAP, ATTICUS approximately doubles expression prediction correlations relative to scPrediXcan and substantially improves sign accuracy. Moreover, ATTICUS enables uncertainty-aware transcriptome-wide association studies (TWAS). Applied to systemic lupus erythematosus and Alzheimer's disease, ATTICUS identifies more associations, with greater cell-type specificity and stronger directional concordance, and reduced statistical inflation. Together, our results demonstrate that personalised training makes predictions more directionally consistent, and uncertainty propagation improves the calibration of TWAS and the prioritisation of disease genes. ATTICUS provides a framework for both, enabling the use of single-cell expression datasets without matched genotypes to inform uncertainty-aware TWAS, expanding their use in prioritising disease-associated genes and cell types.
ER  -

Abstract

Related

Cite