Results: AlphaGenome as a zero-shot imputer of individual-level gene expression — benchmark in a multi-ancestry cohort

Per-gene Spearman correlation results, predictions, and gene panel for the preprint: Astudillo, P. "AlphaGenome as a zero-shot imputer of individual-level gene expression: benchmark in a multi-ancestry cohort".

This deposit contains the derived results needed to reproduce all figures and tables in the preprint without rerunning the full ~30-hour AlphaGenome pipeline. Specifically:

The 100-gene panel definition (60 high-PrediXcan-R² + 20 mid-R² + 20 large-aFC), 99 of which yielded valid predictions.

The 731 1000G individual identifiers from the MAGE cohort.

Per-(gene, individual) predictions from PrediXcan Elastic Net (GTEx v8 EBV-transformed lymphocytes) and AlphaGenome (zero-shot, exon-weighted aggregation, EFO:0000572 ontology).

Per-gene Spearman correlations vs measured MAGE expression (overall and stratified by continental ancestry group).

The AlphaGenome ontology snapshot used to select the LCL term.

Raw inputs (1000 Genomes phased VCFs, MAGE expression matrices, PrediXcan model database, GRCh38 reference FASTA) are not redistributed but can be downloaded from their original sources using the scripts in the accompanying code archive.

Analysis code: 10.5281/zenodo.19854340.