Rockfish is the deep learning based tool for detecting 5mC DNA base modifications.

• ONT Guppy (for basecalling)
• ONT Tombo (for re-segmentation)
• Python >=3.7
• Appropriate CUDA version

1. Clone the repository

   git clone ... Rockfish && cd Rockfish

2. Create virtual environment

   • Using venv

     python3.7 -m venv rockfish_venv
     source rockfish_venv/bin/activate

   • or conda

     conda create --name rockfish python=3.7
     conda activate rockfish

3. Install requirements

   pip install -r requirements.txt

4. Install appropriate PyTorch version (built for specific CUDA version) and pytorch-lighting

   https://pytorch.org/get-started/locally/

   E.g. PyTorch 1.8, pip package for CUDA 11.1

   pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
   pip install pytorch-lightning

CUDA_VISIBLE_DEVICES=<cuda_devices> python src/inference.py [options ...] <model_checkpoint> <test_path> <out_path>

  <model_checkpoint>
    trained model checkpoint file
  <test_path>
    file or folder containing re-segmented fast5 files
  <out_path>
    path to the output tsv file

  options:
    -r, --recursive
      default: false
      recursively search for the input files in <test_path>  
    -b, --batch_size <int>
      default: 1024
      batch size for inference
    -t, --n_workers <int>
      default: 0
      number of workers for reading and processing data
    --prefetch_factor <int>
      default: 4
      number of batches fetched before processing, active only if n_workers > 0
    
  extraction options:
    --reseg_path <str>
      default: 'Analyses/RawGenomeCorrected_000/BaseCalled_template'
      path to the re-segmentation group in fast5
    --norm_method <str>
      default: 'standardization'
      function used for signal normalization
    --motif <str>
      default: 'CG'
      motif used to find genomic positions, regex can be used
    --sample_size <int>
      default: 20
      number of signal points for every base in k-mer
    --window <int>
      default: 8
      size of window left and right of central position, total k-mer: K = 2*W+1

1. Extract features

   python src/extract_features.py [options ...] <input_path> <output_path>
   
     <input_path> 
       file or folder containing re-segmented fast5 files
     <output_path>
       folder where the processed data will be stored
   
     options:
       -r, --recursive 
         default: false
         recursively search for the input files in <input_path>
       -t, --workers <int>
         default: 0
         number of workers used for data generation
       --label <int>
         default: None
         label to store for the given examples (0 or 1), not stored if not set
       --reseg_path <str>
         default: 'Analyses/RawGenomeCorrected_000/BaseCalled_template'
         path to the re-segmentation group in fast5
       --norm_method <str>
         default: 'standardization'
         function used for signal normalization
       --motif <str>
         default: 'CG'
         motif used to find genomic positions, regex can be used
       --sample_size <int>
         default: 20
         number of signal points for every base in k-mer
       --window <int>
         default: 8
         size of window left and right of central position, total k-mer: K = 2*W+1
       --bed_path <str>
         default: None
         path to the bedmethyl file used for position filtering
       --bed_filter <str>
         default: None
         filter method if bed_path is set, currently implemented only high_confidence (n_reads >= 10, meth_freq (label) either 0 (0) or 100 (1))

2. Train Rockfish

   CUDA_VISIBLE_DEVICES=<cuda_devices> python src/train.py [options ...] <train_path> <val_path>
   
     <train_path>
       file containing extracted data in binary format used for training
     <val_path>
       file containing extracted data in binary format used for validation
     
     options:
       -b, --train_batch_size <int>
         default: 128
         mini-batch size used for training (effective batch size is n_gpu * train_batch_size)
       --val_batch_size <int>
         default: 1024
         mini-batch size used for validation (effective n_gpu * val_batch_size)
       --iterable
         default: false
         instead of loading data in RAM, iteratively fetch data (not recommended)
   
     model options:
       --dropout <float>
         default: 0.1
         dropout value used for training
       --nhead <int>
         default: 8
         number of heads in multi-head attention
       --dim_ff <int>
         default: 1024
         dimension of hidden layer in feed-forward network (in transformer)
       --nlayers <int>
         default: 6
         number of transformer encoder layers
   
     train options:
       --epochs <int>
         default: 30
         number of training epochs
       --wd <float>
         default: 1e-4
         weight decay used in AdamW
       --lr <float>
         default: 1e-4
         learning rate upper bound for CyclicLR
       --step_size_up <int>
         default: None
         number of iterations for half cycle in CyclicLR, inferred from train dataset size if not set 

This work has been supported in part by AI Singapore under the project the Deep Generative Modeling of Epigenomics Data (AISG-RPKS-2019-001), by Croatian Science Foundation under the project Single genome and metagenome assembly (IP-2018-01-5886), by the A*STAR Computational Resource Centre and by the National Supercomputing Centre, Singapore through the use of their high-performance computing facilities. D.S. and M.Š. have been partially supported by funding from A*STAR, Singapore.