Here we provided the weights for a trained DeepAccess model, an ensemble of neural networks predicting accessibility from DNA sequence. We trained this model on ATAC-seq data from ten mouse cell types: stem cell, fibroblast, hepatocyte, endoderm, beta pancreatic cell, alpha pancreatic cell, cardiomyocyte, skeletal muscle, dopaminergic midbrain neuron, and spinal motor neuron using binary cross-entropy loss (multi-task classification) with 4,220,507 genomic regions for training: 3,220,507 regions were open in at least 1 cell type, and 1,000,000 regions were closed in all cell types (randomly sampled from the genome). Chromosome 18 and chromosome 19 were held out for validation and testing, respectively. To define training regions for DeepAccess, we generate 100bp genomic windows across the entire mouse genome. We define a region as accessible in a given cell type if more that 50% of the 100bp region overlaps a MACS2(43) accessible region from that cell type.