Bonito

Bonito is an open source research basecaller for Oxford Nanopore reads.

For anything other than basecaller training or method development please use dorado.

$ pip install --upgrade pip
$ pip install ont-bonito
$ bonito basecaller [email protected] /data/reads > basecalls.bam

Bonito supports writing aligned/unaligned {fastq, sam, bam, cram}.

$ bonito basecaller [email protected] --reference reference.mmi /data/reads > basecalls.bam

Bonito will download and cache the basecalling model automatically on first use but all models can be downloaded with -

$ bonito download --models --show  # show all available models
$ bonito download --models         # download all available models

Transformer Models

The bonito.transformer package requires flash-attn.

This must be manually installed as the flash-attn packaging system prevents it from being listed as a normal dependency.

Setting CUDA_HOME to the relevant library directory will help avoid CUDA version mismatches between packages.

Modified Bases

Modified base calling is handled by Remora.

$ bonito basecaller [email protected] /data/reads --modified-bases 5mC --reference ref.mmi > basecalls_with_mods.bam

See available modified base models with the remora model list_pretrained command.

Training your own model

To train a model using your own reads, first basecall the reads with the additional --save-ctc flag and use the output directory as the input directory for training.

$ bonito basecaller [email protected] --save-ctc --reference reference.mmi /data/reads > /data/training/ctc-data/basecalls.sam
$ bonito train --directory /data/training/ctc-data /data/training/model-dir

In addition to training a new model from scratch you can also easily fine tune one of the pretrained models.

bonito train --epochs 1 --lr 5e-4 --pretrained [email protected] --directory /data/training/ctc-data /data/training/fine-tuned-model

If you are interested in method development and don't have you own set of reads then a pre-prepared set is provide.

$ bonito download --training
$ bonito train /data/training/model-dir

All training calls use Automatic Mixed Precision to speed up training. To disable this, set the --no-amp flag to True.

Developer Quickstart

$ git clone https://github.com/nanoporetech/bonito.git  # or fork first and clone that
$ cd bonito
$ python3 -m venv venv3
$ source venv3/bin/activate
(venv3) $ pip install --upgrade pip
(venv3) $ pip install -e .[cu118] --extra-index-url https://download.pytorch.org/whl/cu118

The ont-bonito[cu118] and ont-bonito[cu121] optional dependencies can be used, along with the corresponding --extra-index-url, to ensure the PyTorch package matches the local CUDA setup.

Interface

• bonito view - view a model architecture for a given .toml file and the number of parameters in the network.
• bonito train - train a bonito model.
• bonito evaluate - evaluate a model performance.
• bonito download - download pretrained models and training datasets.
• bonito basecaller - basecaller (.fast5 -> .bam).

References

• Sequence Modeling With CTC
• Quartznet: Deep Automatic Speech Recognition With 1D Time-Channel Separable Convolutions
• Pair consensus decoding improves accuracy of neural network basecallers for nanopore sequencing

Licence and Copyright

(c) 2019 Oxford Nanopore Technologies Ltd.

Bonito is distributed under the terms of the Oxford Nanopore Technologies, Ltd. Public License, v. 1.0. If a copy of the License was not distributed with this file, You can obtain one at http://nanoporetech.com

Research Release

Research releases are provided as technology demonstrators to provide early access to features or stimulate Community development of tools. Support for this software will be minimal and is only provided directly by the developers. Feature requests, improvements, and discussions are welcome and can be implemented by forking and pull requests. However much as we would like to rectify every issue and piece of feedback users may have, the developers may have limited resource for support of this software. Research releases may be unstable and subject to rapid iteration by Oxford Nanopore Technologies.

Citation

@software{bonito,
  title = {Bonito: A PyTorch Basecaller for Oxford Nanopore Reads},
  author = {{Chris Seymour, Oxford Nanopore Technologies Ltd.}},
  year = {2019},
  url = {https://github.com/nanoporetech/bonito},
  note = {Oxford Nanopore Technologies, Ltd. Public License, v. 1.0},
  abstract = {Bonito is an open source research basecaller for Oxford Nanopore reads. It provides a flexible platform for training and developing basecalling models using PyTorch.}
}