DeepBICCN2 Cell Type-Specific Chromatin Accessibility Predictor Container for the Genomic API for Model Evaluation (GAME)

  DeepBICCN2 Cell Type-Specific Chromatin Accessibility Predictor

  This Zenodo record contains a predictor container for the Genomic API for Model Evaluation (GAME). The system provides computational
  predictions of cell type-specific chromatin accessibility in the mouse motor cortex from DNA sequence alone.

  Model Overview

  DeepBICCN2 is a deep learning model trained on single-cell ATAC-seq data from the BRAIN Initiative Cell Census Network (BICCN). The
  model predicts chromatin accessibility patterns across 19 distinct mouse motor cortex cell types directly from genomic sequences.

  Supported Cell Types

  The model provides predictions for 19 mouse motor cortex cell types:

  - Excitatory neurons: L2/3 IT, L5 ET, L5 IT, L5/6 NP, L6 CT, L6 IT, L6b
  - Inhibitory neurons: Lamp5, Pvalb, Sncg, Sst, Sst Chodl, Vip
  - Glial cells: Astrocytes (Astro), Microglia/PVM, Oligodendrocyte Precursor Cells (OPC), Oligodendrocytes (Oligo)
  - Vascular cells: Endothelial cells (Endo), Vascular Leptomeningeal Cells (VLMC)

  Model Specifications

  - Input: DNA sequences of 2114 base pairs (sequences are automatically padded or cropped to this length)
  - Species: Mouse (Mus musculus)
  - Output: Tn5 cut-site counts representing chromatin accessibility
  - Output Scale: Linear (log scale available on request)
  - Readout Type: Point predictions at sequence center
  - Architecture: Convolutional neural network trained on BICCN scATAC-seq data

  API Features

  The predictor implements the GAME REST API specification and supports:
  - /help endpoint: Model metadata and documentation
  - /formats endpoint: Supported request/response formats (JSON, MessagePack)
  - /predict endpoint: Sequence-to-accessibility predictions
  - Batch predictions for multiple sequences and cell types
  - Automatic sequence padding and cropping
  - Flexible output scaling (linear or log)

  Container Contents

  The deepbiccn2_predictor.sif file includes:
  - Pre-trained DeepBICCN2 model weights
  - Cell type-to-output index mapping
  - Flask-based REST API server
  - Sequence processing utilities (padding, one-hot encoding)
  - CREsted framework and all dependencies
  - GPU-accelerated TensorFlow environment

  Model Files

  - deepbiccn2.keras: Pre-trained model in Keras format
  - deepbiccn2_output_classes.tsv: Cell type index mapping

  Execution Command

  apptainer run --nv \
    deepbiccn2_predictor.sif \
    HOST_IP PORT

  The --nv flag enables GPU acceleration for faster predictions. The predictor will listen on the specified host and port for incoming
  prediction requests.

  Documentation

  Full documentation available at: https://crested.readthedocs.io/en/stable/models/BICCN/deepbiccn2.html

  Citations

  This predictor is based on the DeepBICCN2 model described in:

  Kempynck, N., De Winter, S., et al. CREsted: modeling genomic and synthetic cell type-specific enhancers across tissues and species. Zenodo.      https://doi.org/10.5281/zenodo.13918932