Ascertain whether per-gene models outperform a single multi-output model for transcriptome prediction from WSIs

Determine whether training separate regression models for each individual gene yields better or comparable prediction accuracy than training a single multi-output model that simultaneously regresses all N = 20,530 genes when predicting mRNA gene-expression profiles from hematoxylin and eosin-stained whole-slide images.

Background

Because the task involves predicting thousands of gene-expression values simultaneously, a key modeling decision is whether to train individual models per gene or to train a single multi-output model across all genes. The paper highlights this design choice as explicitly unclear at the outset and investigates it empirically.

The uncertainty is important for both accuracy and computational cost considerations, since per-gene training dramatically increases the number of models required.