@anshulkundaje TraitGym authors show this disadvantages Borzoi in supp. figs, but only slightly. Also not a major critique, but something I think about when I see Borzoi-type models compared to genomic language models in this way
@anshulkundaje that are more tolerant to gene expression changes than GWAS hits. In principle, Borzoi may predict control variants cause larger changes to expression than their matched GWAS hit for this reason (disadvantaging it for AUPRC), while likelihood predictors aren’t affected by this
@anshulkundaje because this evaluation assumes causal variants cause larger changes to expression than controls, which is only explicitly true when matched by gene (and cell type, but you already made that point). When not matched by gene, control variants may very well be near genes