RiboDetector - Accurate and rapid RiboRNA sequences Detector based on deep learning

About Ribodetector

RiboDetector is a software developed to accurately yet rapidly detect and remove rRNA sequences from metagenomeic, metatranscriptomic, and ncRNA sequencing data. It was developed based on LSTMs and optimized for both GPU and CPU usage to achieve a 10 times on CPU and 50 times on a consumer GPU faster runtime compared to the current state-of-the-art software. Moreover, it is very accurate, with ~10 times fewer false classifications. Finally, it has a low level of bias towards any GO functional groups.

Prerequirements

1. Create conda env and install Python v3.8

To be able to use RiboDetector, you need to install Python v3.8 or v3.9 (make sure you have version 3.8 because 3.7 cannot serialize a string larger than 4GiB) with conda:

conda create -n ribodetector python=3.8
conda activate ribodetector

2. Install pytorch in the ribodetector env if GPU is available

To install pytorch compatible with your CUDA version, please fellow this instruction: https://pytorch.org/get-started/locally/. Our code was tested with pytorch v1.7, v1.7.1, v1.10.2.

Note: you can skip this step if you don't use GPU

Installation

Using pip

pip install ribodetector

Using conda

conda install -c bioconda ribodetector

Usage

GPU mode

Example

ribodetector -t 20 \
  -l 100 \
  -i inputs/reads.1.fq.gz inputs/reads.2.fq.gz \
  -m 10 \
  -e rrna \
  --chunk_size 256 \
  -o outputs/reads.nonrrna.1.fq outputs/reads.nonrrna.2.fq

The command lind above excutes ribodetector for paired-end reads with mean length 100 using GPU and 20 CPU cores. The input reads do not need to be same length. RiboDetector supports reads with variable length. Setting -l to the mean read length is recommended.

Full help

usage: ribodetector [-h] [-c CONFIG] [-d DEVICEID] -l LEN -i [INPUT [INPUT ...]]
  -o [OUTPUT [OUTPUT ...]] [-r [RRNA [RRNA ...]]] [-e {rrna,norrna,both,none}] 
  [-t THREADS] [-m MEMORY] [--chunk_size CHUNK_SIZE] [-v]

rRNA sequence detector

optional arguments:
  -h, --help            show this help message and exit
  -c CONFIG, --config CONFIG
                        Path of config file
  -d DEVICEID, --deviceid DEVICEID
                        Indices of GPUs to enable. Quotated comma-separated device ID numbers. (default: all)
  -l LEN, --len LEN     Sequencing read length (mean length). Note: the accuracy reduces for reads shorter than 40.
  -i [INPUT [INPUT ...]], --input [INPUT [INPUT ...]]
                        Path of input sequence files (fasta and fastq), the second file will be considered 
                        as second end if two files given.
  -o [OUTPUT [OUTPUT ...]], --output [OUTPUT [OUTPUT ...]]
                        Path of the output sequence files after rRNAs removal (same number of files as input).
                        (Note: 2 times slower to write gz files)
  -r [RRNA [RRNA ...]], --rrna [RRNA [RRNA ...]]
                        Path of the output sequence file of detected rRNAs (same number of files as input)
  -e {rrna,norrna,both,none}, --ensure {rrna,norrna,both,none}
                        Ensure which classificaion has high confidence for paired end reads.
                        norrna: output only high confident non-rRNAs, the rest are clasified as rRNAs;
                        rrna: vice versa, only high confident rRNAs are classified as rRNA and the rest output as non-rRNAs;
                        both: both non-rRNA and rRNA prediction with high confidence;
                        none: give label based on the mean probability of read pair.
                              (Only applicable for paired end reads, discard the read pair when their predicitons are discordant)
  -t THREADS, --threads THREADS
                        number of threads to use. (default: 10)
  -m MEMORY, --memory MEMORY
                        Amount (GB) of GPU RAM. (default: 12)
  --chunk_size CHUNK_SIZE
                        Use this parameter when having low memory. Parsing the file in chunks.
                        Not needed when free RAM >=5 * your_file_size (uncompressed, sum of paired ends).
                        When chunk_size=256, memory=16 it will load 256 * 16 * 1024 reads each chunk (use ~20 GB for 100bp paired end).
  --log LOG             Log file name
  -v, --version         Show program's version number and exit

CPU mode

Example

ribodetector_cpu -t 20 \
  -l 100 \
  -i inputs/reads.1.fq.gz inputs/reads.2.fq.gz \
  -e rrna \
  --chunk_size 256 \
  -o outputs/reads.nonrrna.1.fq outputs/reads.nonrrna.2.fq

The above command line excutes ribodetector for paired-end reads with mean length 100 using 20 CPU cores. The input reads do not need to be same length. RiboDetector supports reads with variable length. Setting -l to the mean read length is recommended. If you need to save the log into a file, you can specify it with --log <logfile>

Note: when using SLURM job submission system, you need to specify --cpus-per-task to the number you CPU cores you need and set --threads-per-core to 1.

Full help

usage: ribodetector_cpu [-h] [-c CONFIG] -l LEN -i [INPUT [INPUT ...]] 
  -o [OUTPUT [OUTPUT ...]] [-r [RRNA [RRNA ...]]] [-e {rrna,norrna,both,none}] 
  [-t THREADS] [--chunk_size CHUNK_SIZE] [-v]

rRNA sequence detector

optional arguments:
  -h, --help            show this help message and exit
  -c CONFIG, --config CONFIG
                        Path of config file
  -l LEN, --len LEN     Sequencing read length (mean length). Note: the accuracy reduces for reads shorter than 40.
  -i [INPUT [INPUT ...]], --input [INPUT [INPUT ...]]
                        Path of input sequence files (fasta and fastq), the second file will be considered as 
                        second end if two files given.
  -o [OUTPUT [OUTPUT ...]], --output [OUTPUT [OUTPUT ...]]
                        Path of the output sequence files after rRNAs removal (same number of files as input).
                        (Note: 2 times slower to write gz files)
  -r [RRNA [RRNA ...]], --rrna [RRNA [RRNA ...]]
                        Path of the output sequence file of detected rRNAs (same number of files as input)
  -e {rrna,norrna,both,none}, --ensure {rrna,norrna,both,none}
                        Ensure which classificaion has high confidence for paired end reads.
                        norrna: output only high confident non-rRNAs, the rest are clasified as rRNAs;
                        rrna: vice versa, only high confident rRNAs are classified as rRNA and the rest output as non-rRNAs;
                        both: both non-rRNA and rRNA prediction with high confidence;
                        none: give label based on the mean probability of read pair.
                              (Only applicable for paired end reads, discard the read pair when their predicitons are discordant)
  -t THREADS, --threads THREADS
                        number of threads to use. (default: 20)
  --chunk_size CHUNK_SIZE
                        chunk_size * 1024 reads to load each time.
                        When chunk_size=1000 and threads=20, consumming ~20G memory, better to be multiples of the number of threads..
  --log LOG             Log file name
  -v, --version         Show program's version number and exit

Note: RiboDetector uses multiprocessing with shared memory, thus the memory use of a single process indicated in htop or top is actually the total memory used by RiboDector. Some job submission system like SGE mis-calculated the total memory use by adding up the memory use of all process. If you see this do not worry it will cause out of memory issue.

FAQ

1. What should I set for -l when I have reads with variable length?

You can set the -l parameter to the mean read length if you have reads with variable length. The mean read length can be computed with seqkit stats. This parameter tells how many bases will be used to capture the sequences patterns for classification.

2. How does -e parameter work? What should I set (rrna, norrna, none, both)?

This parameter is only necessary for paired end reads. When setting to rrna, the paired read ends will be predicted as rRNA only if both ends were classified as rRNA. If you want to identify or remove rRNAs with high confidence, you should set it to rrna. Conversely, norrna will predict the read pair as nonrRNA only if both ends were classified as nonrRNA. This setting will only output nonrRNAs with high confidence. both will discard the read pairs with two ends classified inconsistently, only pairs with concordant prediction will be reported in the corresponding output. none will take the mean of the probabilities of both ends and decide the final prediction. This is also the default setting.

3. I have very large input file but limited memory, what should I do?

You can set the --chunk_size parameter which specifies how many reads the software load into memory once.

4. What should I do if RiboDetector hangs with SLURM?

The most likely cause is that the requested computational resource is not sufficient for the input file. You need to make sure you specified --cpus-per-task to the number you CPU cores you want to use and set --threads-per-core to 1 in the SLURM submission script or command. If the issue remains, you can try to reduce the memory use by setting --chunk_size parameter in ribodetector or ribodetector_cpu command.

Citation

Deng ZL, Münch PC, Mreches R, McHardy AC. Rapid and accurate detection of ribosomal RNA sequences using deep learning. Nucleic Acids Research. 2022. (https://doi.org/10.1093/nar/gkac112)

Acknowledgements

The scripts from the base dir were from the template pytorch-template by Victor Huang and other contributors.