Using a deep learning CNN+RNN+CTC structure to establish end-to-end basecalling for the nanopore sequencer.
Built with TensorFlow and python 2.7.
If you found Chiron useful, please consider to cite:
Haotian Teng, Minh Duc Cao, Michael B Hall, Tania Duarte, Sheng Wang, Lachlan J M Coin; Chiron: translating nanopore raw signal directly into nucleotide sequence using deep learning, GigaScience, Volume 7, Issue 5, 1 May 2018, giy037, https://doi.org/10.1093/gigascience/giy037
DNA basecall:
python chiron/entry.py call -i <input_fast5_folder> -o <output_folder> -m chiron/model/DNA_default -p dna-pre
RNA basecall:
python chiron/entry.py call -i <input_fast5_folder> -o <output_folder> -m chiron/model/RNA_default -p rna-pre --mode rna
- Install
- Basecall
- Training
- Train on Google Cloud ML engine
- Distributed training on Google CLoud ML Engine
If you currently have TensorFlow installed on your system, we would advise you to create a virtual environment to install Chiron into, this way there is no clash of versions etc.
If you would like to do this, the best options would be virtualenv, the more user-friendly virtualenvwrapper, or through anaconda. After installing one of these and activating the virtual environment you will be installing Chiron into, continue with the rest of the installation instructions as normal.
To install with pip:
pip install chiron
This will install Chiron, and h5py (required for reading in .fast5 files).
Tensorflow need to be install in addition by:
pip install tensorflow==1.15
or GPU version:
pip install tensorflow-gpu==1.15
git clone https://github.com/haotianteng/chiron.git
cd Chiron
You will also need to install dependencies.
python setup.py install
For CPU-version:
pip install tensorflow==1.15
For GPU-version(Nvidia GPU required):
Install CUDA
Install cuDNN
Install tensorflow gpu version
pip install tensorflow-gpu==1.15
And then add the Chiron into PYTHONPATH,for convinience you can add it to the .bashrc
export PYTHONPATH=[Path to Chiron/Chiron]:$PYTHONPATH
For alternate/detailed installation instructions for TensorFlow, see the documentation.
An example call to Chiron to run basecalling is:
chiron call -i <input_fast5_folder> -o <output_folder> -m <model_folder>
All Chiron functionality can be run from entry.py in the Chiron folder. (You might like to also add the path to Chiron into your PATH for ease of running).
python chiron/entry.py call -i <input_fast5_folder> -o <output_folder> -m <model_folder>
We provide 5 sample fast5 files for DNA(courtesy of nanonet) and 5 sample files for RNA in the GitHub repository and two models (DNA_default and RNA_default) which you can run a test on. These are located in chiron/example_data/. From inside the Chiron repository:
python chiron/entry.py call -i chiron/example_folder/ -o <output_folder> -m chiron/model/DNA_default --preset dna-pre
And from v0.5 we have provide a good RNA model for the direct-RNA basecall.
python chiron/entry.py call -i <input_fast5_folder> -o <output_folder> -m chiron/model/RNA_default --mode rna --preset rna-pre
You can reduce the batch size(defualt is 400) if you have limited RAM/GPU-RAM or extend it, bigger batch size has (very)slightly better result and faster inference speed.
python chiron/entry.py call -i <input_fast5_folder> -o <output_folder> -m chiron/model/RNA_default --mode rna --preset rna-pre -b 200
Any arguments provided afterward will override the argument in the preset, preset arguments values are:
DNA preset arguments value(dna-pre):
'start':0,'batch_size':400,'segment_len':400,'jump':390,'threads':0,'beam':30
RNA preset arguments value(rna-pre):
'start':0,'batch_size':300,'segment_len':2000,'jump':1900,'threads':0,'beam':30
For better RNA basecalling result, the fast5 files can be preproceesd using BoostNano, which is a tool developed by me to segment and remove the polyA and adapter tail.
(From v0.3)
Beam search decoder: chiron call -i -o --beam <beam_width>
Greedy decoder: chiron call -i -o --beam 0
Beam Seach decoder give a higher accuracy, and larger beam width can furthur improve the accuracy. Greedy decoder give a faster decoding speed than the beam search decoder:
| Device | Greedy decoder rate(bp/s) | Beam Search decoder rate(bp/s), beam_width=50 |
|---|---|---|
| CPU | 21 | 17 |
| GPU | 1652 | 1204 |
chiron call will create five folders in <output_folder> called raw, result, segments, meta, and reference.
result: fastq/fasta files with the same name as the fast5 file they contain the basecalling result for. To create a single, merged version of these fasta files, try something likepaste --delimiter=\\n --serial result/*.fasta > merged.fastaraw: Contains a file for each fast5 file with it's raw signal. This file format is an list of integers. i.e544 554 556 571 563 472 467 487 482 513 517 521 495 504 500 520 492 506 ...segments: Contains the segments basecalled from each fast5 file.meta: Contains the meta information for each read (read length, basecalling rate etc.). Each file has the same name as it's fast5 file.reference: Contains the reference sequence (if any).
With -e flag to output fastq file(default) with quality score or fasta file.
Example:
chiron call -i <input_fast5_folder> -o <output_folder> -e fastq
chiron call -i <input_fast5_folder> -o <output_folder> -e fasta
The default DNA model trained on R9.4 protocol with a mix of Lambda and E.coli dataset, and the default RNA model is trained on R9.4 direct RNA kit (-200mV configuration). If the basecalling result is not satisfying, you can train a model on your own training data set.
Recommend training on GPU with TensorFlow - usually 8GB RAM (GPU) is required.
Using raw.py script to extract the signal and label from the re-squiggled fast5 file. (For how to re-squiggle fast5 file, check Tombo re-squiggle)
chiron export -i <fast5 folder> -o <output_folder>
or directly use the raw.py script in utils.
python chiron/utils/raw.py --input <fast5 folder> --output <output_folder> --mode dna
This will generate a tfrecord file for training when using the chiron_rcnn_train.py and chiron_input.py pipeline.
python chiron/utils/file_batch.py --input <fast5 folder> --output <output folder> --length 400 --mode dna
This will generate several binary .bin file for training when using the chiron_train.py and chiron_queue_input.py pipeline.
source activate tensorflow
chiron train --data_dir <signal_label folder> --log_dir <model_log_folder> --model_name <saved_model_name>
or run directly by
python chiron/chiron_rcnn_train.py --data_dir <signal_label folder/ tfrecord file> --log_dir <model_log>
Following parameters can be passed to Chiron when training
data_dir(Required): The folder containing your signal and label files.
log_dir(Required): The folder where you want to save the model.
model_name(Required): The name of the model. The record will be stored in the directory log_dir/model_name/
tfrecord: File name of tfrecord. Default is train.tfrecords.
sequence_len: The length of the segment you want to separate the sequence into. Longer length requires larger RAM.
batch_size: The batch size.
step_rate: Learning rate of the optimizer.
max_step: Maximum step of the optimizer.
k_mer: Chiron supports learning based on k-mer instead of a single nucleotide, this should be an odd number, even numbers will cause an error.
retrain: If this is a new model, or you want to load the model you trained before. The model will be loaded from log_dir/model_name/
Before training the model on cloud ml engine, please check if it is working on local machine or not by following commands
gcloud ml-engine local train \
--module-name chiron.utils.raw \
--package-path chiron.utils/ \
-- --input input_fast5_folder \
--output output
gcloud ml-engine local train \
--module-name chiron.chiron_rcnn_train \
--package-path chiron/
If it is working well, please go to next step.
BUCKET_NAME=chiron-ml
REGION=us-central1
gsutil mb -l $REGION gs://$BUCKET_NAME
gsutil cp -r raw_fast_folder gs://$BUCKET_NAME/fast5-data
JOB_NAME=chiron_single_1
OUTPUT_PATH=gs://$BUCKET_NAME/$JOB_NAME
INPUT_PATH=gs://$BUCKET_NAME/train_tfdata
gcloud ml-engine jobs submit training $JOB_NAME \
--staging-bucket gs://chiron-ml \
--module-name chiron.chiron_rcnn_train \
--package-path chiron/ \
--region $REGION \
--config config.yaml \
-- \
--data_dir gs://$BUCKET_NAME/train_tfdata \
--cache_dir gs://$BUCKET_NAME/cache/train.hdf5 \
--log_dir gs://$BUCKET_NAME/GVM_model
Change configure.yaml according to GCloud Docs For example the following configure_multi_gpu.yaml:
trainingInput:
scaleTier: CUSTOM
masterType: standard_p100
workerType: standard_p100
parameterServerType: large_model
workerCount: 3
parameterServerCount: 3
Will enable 3 workers + 1 master worker with one P-100 GPU in each worker.
FAST5_FOLDER=/my/fast5/
OTUPUT_FOLDER=/my/file_batch/
SEGMENT_LEN=512
Transfer fast5 to file batch
python utils/file_batch.py --input $FAST5_FOLDER --output $OUTPUT_FOLDER --length $SEGMENT_LEN
Copy to Google Cloud
gsutil cp -r $OUTPUT_FOLDER gs://$BUCKET_NAME/file_batch
JOB_NAME=chiron_multi_4
DATA_BUCKET=chiron-training-data
MODEL_BUCKET=chiron-model
REGION=us-central1
MODEL_NAME=test_model1
GPU_NUM=4
gcloud ml-engine jobs submit training ${JOB_NAME} \
--runtime-version 1.6 \
--staging-bucket gs://chiron-model/ \
--module-name chiron.chiron_multi_gpu_train \
--package-path chiron \
--region $REGION \
--config config_multi_gpu.yaml \
-- \
-i gs://$DATA_BUCKET/file_batch \
-o gs://$MODEL_BUCKET/ \
-m ${MODEL_NAME} \
-n ${GPU_NUM}