CAGI dataset.

The CAGI5 challenge dataset⁶¹ was used to evaluate the performance of the models on zero-shot single-nucleotide variant effect generalization as following the same procedure as Ref.⁸³. We only considered MPRA experiments in HepG2 (LDLT, SORT1, F9) and K562 (PKLR). We extracted 230 nucleotide sequences from the reference genome centered on each regulatory region of interest. Alternative alleles are then substituted correspondingly to construct the CAGI test sequences. Pearson correlation was calculated between the varient effect scores by the model and experimentally measured effect size per experiment. For HepG2 performances, we report the average Pearson’s r across the three experiments.