granek / dorado

A LibTorch Basecaller for Oxford Nanopore Reads

Dorado

Dorado is a high-performance, easy-to-use, open source basecaller for Oxford Nanopore reads.

Features

• One executable with sensible defaults, automatic hardware detection and configuration.
• Runs on Apple silicon (M1/2 family) and Nvidia GPUs including multi-GPU with linear scaling.
• Modified basecalling (Remora models).
• Duplex basecalling.
• POD5 support for highest basecalling performance.
• Based on libtorch, the C++ API for pytorch.
• Multiple custom optimisations in CUDA and Metal for maximising inference performance.

If you encounter any problems building or running Dorado please report an issue.

Installation

• dorado-0.1.1-linux-x64
• dorado-0.1.1-osx-arm64
• dorado-0.1.1-win64

Running

To run Dorado, download a model and point it to POD5 files (Fast5 files are supported but will not be as performant).

$ dorado download --model [email protected]
$ dorado basecaller [email protected] pod5s/ > calls.sam

To call modifications simply add --modified-bases.

$ dorado basecaller [email protected] pod5s/ --modified-bases 5mCG_5hmCG > calls.sam

For unaligned BAM output, dorado output can be piped to BAM using samtoools:

$ dorado basecaller [email protected] pod5s/ | samtools view -Sh > calls.bam

Stereo Duplex Calling:

$ dorado duplex [email protected] pod5s/ --pairs pairs.txt > duplex.sam

Platforms

Dorado has been tested on the following systems:

Platform	GPU/CPU
Windows	(G)V100, A100
Apple	M1, M1 Pro, M1 Max, M1 Ultra
Linux	(G)V100, A100

Systems not listed above but which have Nvidia GPUs with >=8GB VRAM and architecture from Volta onwards have not been widely tested but are expected to work. If you encounter problems with running on your system please report an issue

Roadmap

Dorado is still in alpha stage and not feature-complete, the following features form the core of our roadmap:

1. DNA Barcode multiplexing
2. Alignment (output aligned BAMs).
3. Python API

Performance tips

1. For optimal performance Dorado requires POD5 file input. Please convert your Fast5 files before basecalling.
2. Dorado will automatically detect your GPUs' free memory and select an appropriate batch size.
3. Dorado will automatically run in multi-GPU ('cuda:all') mode. If you have a hetrogenous collection of GPUs select the faster GPUs using the --device flag (e.g --device "cuda:0,2). Not doing this will have a detrimental impact on performance.

Available basecalling models

To download all available dorado models run:

$ dorado download --model all

The following models are currently available:

• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]
• [email protected]

Developer quickstart

Linux dependencies

The following packages are necessary to build dorado in a barebones environment (e.g. the official ubuntu:jammy docker image)

apt-get update && \
    apt-get install -y --no-install-recommends \
        curl \
        git \
        ca-certificates \
        build-essential \
        nvidia-cuda-toolkit \
        libhdf5-dev \
        libssl-dev \
        libzstd-dev \
        cmake \
        autoconf \
        automake

Clone and build

The commands below will build dorado and install it in /opt. NUM_THREADS controls the number of threads that cmake uses to compile dorado. It can be set to a value higher than "1", but using too many threads can use all available RAM and cause compilation to fail. Peak memory usage seems to be 1-2GB per thread.

export NUM_THREADS=1
git clone https://github.com/nanoporetech/dorado.git /dorado
cd /dorado
cmake -S . -B cmake-build -DCMAKE_CUDA_COMPILER=nvcc
cmake --build cmake-build --config Release --parallel $NUM_THREADS
ctest --test-dir cmake-build
cmake --install cmake-build --prefix /opt
rm -rf /dorado
cd /

Pre commit

The project uses pre-commit to ensure code is consistently formatted, you can set this up using pip:

$ pip install pre-commit
$ pre-commit install

Licence and Copyright

(c) 2022 Oxford Nanopore Technologies Ltd.

Dorado 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