EpiCast: leveraging virtual epigenomic features to predict episomal regulatory activity across cell types

Data and trained models accompanying the EpiCast manuscript. epicast_data.zip mirrors the layout of the code repository, https://github.com/maplecai/EpiCast, so unzipping it at the repository root puts every file where the analysis scripts already look for it: data/gosai_mpra/: the 760,679-CRE MPRA label table of Gosai et al. and the virtual epigenomic feature (VEF) matrices derived from Sei, Enformer, Borzoi and AlphaGenome predictions of the same CREs. data/castillo_mpra/: the independent MPRA of the zero-shot evaluation and its AlphaGenome VEF matrix. data/DHS64/: predictions of the DHS64 baseline on those CREs. saved/: the four trained runs, each with its training configuration, its best checkpoint and the held-out predictions the figures are computed from. The raw ENCODE files behind the label table are not included; the download link of every one of them is in data/Gosai_MPRA/metadata.csv in the repository.