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

ATTICUS, an uncertainty-aware personalised sequence-to-expression model, approximately doubles expression prediction correlations and substantially improves sign accuracy across blood and brain cell types in OneK1K and ROSMAP.

• Why it matters: Genetic variants must be mapped to disease-relevant cell types, but existing sequence-to-function models have limited effect sign accuracy and do not quantify predictive uncertainty for downstream association testing.
• What they did: ATTICUS uses personalised sequence representations and uncertainty-aware expression prediction, transferring regulatory knowledge from bulk pretraining to cell-type-specific prediction without needing genotypes paired with single-cell expression data.
• The result: In uncertainty-aware TWAS for systemic lupus erythematosus and Alzheimer's disease, ATTICUS finds more associations with greater cell-type specificity, stronger directional concordance, and reduced statistical inflation, improving calibration and disease-gene prioritisation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.