Rockfish

Update: R10.4.1 code can be found here (branch r10.4.1).

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

Find the small example here.

Requirements

• Linux (tested on Ubuntu 20.04)
• ONT Guppy >= 5 (sup model; tested on 5.0.14 - download)
• Python >= 3.9
• CUDA (for GPU inference; tested on 11.3)

Python Requirements

Python requirements can be found in setup.cfg

Installation

0. a) Create new environment (e.g. Conda):

   conda create --name rockfish python=3.9

   b) Activate the environment

   conda activate rockfish

1. Clone the repository

   git clone https://github.com/lbcb-sci/rockfish.git rockfish && cd rockfish

2. Run installation

   pip install --extra-index-url https://download.pytorch.org/whl/cu113 .

   Note: "cu113" installs PyTorch for CUDA 11.3. If you want to install PyTorch for other CUDA version, replace "cu113" with appropriate version (e.g. for CUDA 10.2 "cu102"). For CPU version replace "cu113" with "cpu".

   Installation should take a few minutes on a desktop computer with reasonable network bandwidth.

3. Download models Available models: base, small (or both with all)

   rockfish download -m {all, base, small} -s <save_path>

Inference

1. Guppy basecalling

   guppy_basecaller -i <fast5_folder> -r -s <save_folder> --config <config_file> --fast5_out --device <cpu_or_gpus>

   Note: --fast5_out will output fast5 files with fastq and move_table groups needed for inference. This parameter is mandatory.

2. Run inference

   rockfish inference -i <saved_fast5_files> --reference <reference_path> --model_path <model_path> -r -t <n_workers> -b <batch_size> -d <devices>

   • Number of workers -t sets number of processes for generating the data.
   • Batch size -b is an optional parameter with default value of $4096$. However, for some GPUS (like V100 or A100 with 32GB/42 GB VRAM), it's appropriate to set it to a higher value (e.g. $n_{gpu} \times 8192$ or $n_{gpu} \times 16384$ for base model).
   • Examples of device parameter -d:
     • CPU: No parameter
     • 1 GPU: -d 0
     • 2 GPUs: -d 0,1
     • 2 GPUs (3rd and 4th GPU): -d 2,3

Models

Model	Encoder layers	Decoder layers	Features	Feedforward	Guppy config
Base	12	12	384	2048	dna_r9.4.1_450bps_sup.cfg
Small	6	6	128	1024	dna_r9.4.1_450bps_sup.cfg

Example

Run the example script on 1000 pre-basecalled (Guppy 5.0.14. sup) fast5 files (sampled from Nanopolish data):

GPUS=<devices> ./scripts/example.sh

Note: Use GPUS var to set GPUs for the inference. Omit GPUS if the inference is run on CPU.

Result of the inference is predictions.tsv file. It is tab-delimited text file with four fileds:

1. Read-id
2. Contig name
3. Position in the contig
4. 5mC probability

Note: The predictions can have a slightly different order compared to the predictions in the expected.tsv file due to multiprocessing. The running time should be about 5 minutes for CPU mode. GPU should be significantly faster.

Acknowledgement

This work has been supported in part by Croatian Science Foundation under the project Single genome and metagenome assembly (IP-2018-01-5886), by Epigenomics and Epitranscriptomics Research seed grant from Genome Institute of Singapore (GIS), by Career Development Fund (C210812037) from ASTAR, and by the ASTAR Computational Resource Centre through the use of its high-performance computing facilities.