Deconvolution-derived cell-type expression targets for personal genome sequence-to-expression prediction

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Sequence-to-function models predict gene regulation but struggle with individual gene expression differences.
• Cell-type-specific regulatory effects are often masked in bulk RNA sequencing data.
• Limited availability of paired genotype and single-cell expression data hinders accurate prediction.

Purpose of the Study:

• To assess the utility of deconvolution for generating cell-type-specific expression targets from bulk RNA-seq for personal genome expression prediction.
• To compare the performance of various sequence-derived models against genotype-feature baselines.
• To investigate the impact of target data quality and cohort size on model performance.

Main Methods:

• Deconvolution of GTEx v8 bulk RNA-seq data using BayesPrism with single-nucleus reference profiles across 83 tissue-cell-type contexts.
• Comparison of genotype-feature models, Enformer (frozen and fine-tuned), and Borzoi models.
• Evaluation using random and nonlinear-enriched gene sets with coverage-aware sensitivity analysis.

Main Results:

• Deconvolved expression showed good agreement with matched pseudobulked single-nucleus RNA-seq (median donor-level Pearson correlations 0.53-0.73).
• Sequence-derived models, particularly Enformer, outperformed genotype-feature baselines.
• Model performance correlated positively with deconvolution-pseudobulk agreement, indicating target reliability is crucial.

Conclusions:

• Bulk RNA-seq deconvolution is a feasible method for creating cell-type-resolved training targets for gene expression prediction.
• Target data quality and cohort size remain significant limitations for accurate prediction.
• Frozen pretrained sequence representations offer a computationally efficient and competitive baseline for personal sequence-to-expression modeling.