Enformer: Regulatory Activity from DNA
Key Facts
- Input: one-hot DNA (196,608 bp × 4). Centered on region of interest
- Output: 896 bins × 128 bp (~115 kb predicted region) across ~5,313 tracks (CAGE, DNase, ATAC, ChIP-seq, 100+ cell types)
- Architecture: Stem conv → 6 dilated conv blocks → 11 transformer layers (8 heads, relative pos encodings) → pointwise conv
- Training: Poisson log-likelihood loss, softplus output, hg38 + mm10
- Borzoi extends to 524 kb input, 32
[Description truncada. Veja o README completo no GitHub.]