Genomic Pretraining Learns to Read Designed Enhancer Motifs, but Not Across Assay Domains

Genomic foundation models are increasingly used to predict regulatory activity from DNA sequence, yet it remains unclear whether their gains reflect transferable regulatory knowledge or interpolation within a particular sequence library and assay. We treat enhancer sequence-to-function prediction as a verification case study for genomic pretraining, testing it at three levels: transfer under library-level distribution shift, the sequence features underlying this transfer using known designed motif locations, and generalization to an independent human primary-cell endpoint. Under leave-one-library-out evaluation, fine-tuned HyenaDNA consistently outperforms its architecture-matched random initialization and a 4-mer ridge baseline. This effect replicates with Nucleotide Transformer and in two additional independent MPRA studies, with a cross-study pattern consistent with a larger advantage in lower-data settings. The gain over a validation-selected, matched-capacity LegNet, however, is small. In-silico mutagenesis shows that pretraining improves localization of the designed motifs from an AUROC of $0.55$ to $0.66$. Across the tested sequence-model families, performance on a human erythroid primary endpoint remains low (Spearman $\rho\leq0.19$), whereas the measured discovery-assay readout predicts the same endpoint ($\rho=0.76$). Genomic pretraining therefore learns useful motif-sensitive features that transfer across designed sequence libraries, but these gains do not justify treating sequence-only models as stand-alone predictors of primary functional activity. A model can generalize across sequences while failing to generalize across the experimental measurement process.