config.json

{
  "model_type": "clipnet",
  "library_name": "keras",
  "framework": "tensorflow",
  "repo_id": "adamyhe/clipnet",
  "clipnet_version": "0.2.2",
  "tensorflow_version": ">=2.14.0,<2.15.0",
  "description": "CLIPNET is an ensemble of convolutional neural networks that predicts transcription initiation from 1000 bp DNA sequence at single nucleotide resolution.",
  "architectures": [
    "CLIPNET"
  ],
  "task": "genomic-sequence-to-function",
  "input": {
    "sequence_length": 1000,
    "channels": 4,
    "channel_order": [
      "A",
      "C",
      "G",
      "T"
    ],
    "encoding": "two-hot",
    "accepted_fasta_codes": {
      "A": [
        2,
        0,
        0,
        0
      ],
      "C": [
        0,
        2,
        0,
        0
      ],
      "G": [
        0,
        0,
        2,
        0
      ],
      "T": [
        0,
        0,
        0,
        2
      ],
      "R": [
        1,
        0,
        1,
        0
      ],
      "Y": [
        0,
        1,
        0,
        1
      ],
      "S": [
        0,
        1,
        1,
        0
      ],
      "W": [
        1,
        0,
        0,
        1
      ],
      "K": [
        0,
        0,
        1,
        1
      ],
      "M": [
        1,
        1,
        0,
        0
      ],
      "N": [
        0,
        0,
        0,
        0
      ]
    }
  },
  "outputs": [
    {
      "name": "profile",
      "shape": [
        1000
      ],
      "description": "Predicted initiation profile for the center 500 bp on the plus strand concatenated with the center 500 bp on the minus strand."
    },
    {
      "name": "quantity",
      "shape": [
        1
      ],
      "description": "Predicted total PRO-cap quantity across both strands."
    }
  ],
  "ensemble": {
    "aggregation": "mean",
    "num_folds": 9,
    "files": [
      "fold_1.h5",
      "fold_2.h5",
      "fold_3.h5",
      "fold_4.h5",
      "fold_5.h5",
      "fold_6.h5",
      "fold_7.h5",
      "fold_8.h5",
      "fold_9.h5"
    ]
  },
  "architecture_hyperparameters": {
    "conv_1": {
      "filters": 64,
      "kernel_size": 8,
      "activation": "elu"
    },
    "conv_2": {
      "filters": 128,
      "kernel_size": 4,
      "activation": "relu"
    },
    "num_dilations": 9,
    "dilation_kernel_size": 3,
    "dilated_conv_filters": 64,
    "dropout": 0.3
  },
  "training": {
    "optimizer": "Adam",
    "learning_rate": 0.001,
    "loss": {
      "profile": "CosineSimilarity",
      "quantity": "msle"
    },
    "batch_size": 256,
    "epochs": 200,
    "early_stopping_patience": 10
  },
  "postprocessing": {
    "recommended_profile_scaling": "profile_scaled = (profile / profile.sum(axis=1, keepdims=True)) * quantity"
  },
  "references": {
    "paper": "https://www.biorxiv.org/content/10.1101/2024.03.13.583868",
    "source_code": "https://github.com/Danko-Lab/clipnet",
    "zenodo": "https://zenodo.org/doi/10.5281/zenodo.10408622"
  },
  "license": "mit"
}