Causalcall

Code for the paper Causalcall: nanopore basecalling using a temporal convolutional network.

Environment

• Ubuntu 14.04
• python 3.6
• tensorflow 1.8
  Dependencies:
  numpy, collections, threading, tempfile, h5py, statsmodels, difflib, argparse, tqdm, multiprocessing, psutil

Usages

Preparing the training data:

Extract training and validation data (.tfrecords) from resquiggled fast5 files:

python raw.py -i training/validation_fast5_files_folder -o tfrecords_files_folder -f train.tfrecords/validate.tfrecords

Training:

Train the model with default parameters:

CUDA_VISIBLE_DEVICES=cuda_id python train.py -i tfrecords_files_folder -o model_folder -m model_name

Make sure that train.tfrecords and validate.tfrecords are both in the tfrecords_directory.
If you want to change parameters, use train.py -h for more details.

Basecalling:

CUDA_VISIBLE_DEVICES=cuda_id python basecall.py -i fast5_files_folder -o results_folder -m path_to_model

Path to default model: ./model/DNAmodel/

Acknowledgement

We thank Chiron for providing the source code. Causalcall is developed on the basic framework of Chiron's code. (The parts of preprocessing input data and converting outputs of the TCN-based model to base sequences are revised based on Chiron's code following MPL 2.0.)