AllenInstitute/BG-human-dilatedCNN

Dilated CNN trained on human (Homo sapiens) basal ganglia chromatin accessibility across 59 cell-type tracks. Part of the DNA sequence-modeling resources for the basal ganglia cell atlas. Published in Johansen, Fu et al., 2025.

Background

Learns cis-regulatory sequence logic underlying basal ganglia cell-type specialization from snATAC-seq accessibility profiles. Use it to score candidate enhancers, run in silico mutagenesis, or guide cell-type-specific sequence design.

Availability

Currently private. RegRex (formerly enhancer-designer) is an internal Allen Institute tool with a public release planned. For early access contact Kasia Kedzierska (kasia.kedzierska@alleninstitute.org).

Model Details

• Architecture: dilated_cnn
• Species: human
• Number of tracks: 59
• Framework: PyTorch Lightning
• Checkpoint format: Lightning .ckpt

Target Tracks

• AMY-SLEA-BNST_D1_GABA
• AMY-SLEA-BNST_GABA
• Astrocyte
• BAM
• BF_SKOR1_Glut
• B_cells
• COP
• Endo
• Ependymal
• GPe-NDB-SI_LHX6-LHX8-GBX1_GABA
• ... (49 more)

Usage

from regrex.models import load_model

model = load_model("AllenInstitute/BG-human-dilatedCNN")