Skip to content

Predicting pathogenic potentials of short DNA reads with reverse-complement deep neural networks.

License

Notifications You must be signed in to change notification settings

JakubBartoszewicz/DeePaC

Repository files navigation

DeePaC

DeePaC is a python package and a CLI tool for predicting labels (e.g. pathogenic potentials) from short DNA sequences (e.g. Illumina reads) with interpretable reverse-complement neural networks. For details, see the original paper in Bioinformatics: https://doi.org/10.1093/bioinformatics/btz541. For details regarding the interpretability functionalities of DeePaC, see the paper in NAR Genomics and Bioinformatics: https://doi.org/10.1093/nargab/lqab004. For the description of the most current models, see papers in Briefings in Bioinformatics (https://doi.org/10.1093/bib/bbab269) and Bioinformatics (https://doi.org/10.1093/bioinformatics/btac495).

Documentation can be found here: https://rki_bioinformatics.gitlab.io/DeePaC/. See also the main repo here: https://gitlab.com/dacs-hpi/deepac and the legacy URL https://gitlab.com/rki_bioinformatics/DeePaC.

Plug-ins

DeePaC-strain

Basic version of DeePaC comes with built-in models trained to predict pathogenic potentials of NGS reads originating from novel bacteral species. If you want to predict pathogenicity of novel strains of known species, try the DeePaC-strain plugin available here: https://gitlab.com/dacs-hpi/DeePaC-strain.

DeePaC-vir

If you want to detect novel human viruses, try the DeePaC-vir plugin: https://gitlab.com/dacs-hpi/DeePaC-vir.

DeePaC-Live

If you want to run the predictions in real-time during an Illumina sequencing run, try DeePaC-Live: https://gitlab.com/dacs-hpi/deepac-live.

Installation

We recommend using Bioconda (based on the conda package manager) or custom Docker images based on official Tensorflow images. Alternatively, a pip installation is possible as well. For installation on IBM Power Systems (e.g. AC992), see separate installation instructions (experimental).

With Bioconda (recommended)

install with bioconda

You can install DeePaC with bioconda. Set up the bioconda channel first (channel ordering is important):

conda config --add channels defaults
conda config --add channels bioconda
conda config --add channels conda-forge
conda config --set channel_priority strict

We recommend setting up an isolated conda environment:

# python 3.7-3.9 are supported
conda create -n my_env python=3.9
conda activate my_env

and then:

# For the latest version (0.14.1), this should work both for GPU-equipped and CPU-only systems
conda install deepac

If conda is slow, consider installing mamba.

Optional: download and compile the latest deepac-live custom models DOI:

deepac getmodels --fetch

If you want to install the plugins as well, use:

conda install deepacvir deepacstrain

With Docker (also recommended)

Requirements:

See TF Docker installation guide and the NVIDIA Docker support installation guide for details. The guide below assumes you have Docker 19.03 or above.

You can then pull the desired image:

# Basic installation - CPU only
docker pull dacshpi/deepac:0.14.1

# For GPU support
docker pull dacshpi/deepac:0.14.1-gpu

And run it:

# Basic installation - CPU only
docker run -v $(pwd):/deepac -u $(id -u):$(id -g) --rm dacshpi/deepac:0.14.1 deepac --help
docker run -v $(pwd):/deepac -u $(id -u):$(id -g) --rm dacshpi/deepac:0.14.1 deepac test -q

# With GPU support
docker run -v $(pwd):/deepac -u $(id -u):$(id -g) --rm --gpus all dacshpi/deepac:0.14.1-gpu deepac test

# If you want to use the shell inside the container
docker run -it -v $(pwd):/deepac -u $(id -u):$(id -g) --rm --gpus all dacshpi/deepac:0.14.1-gpu bash

The image ships the main deepac package along with the deepac-vir and deepac-strain plugins. See the basic usage guide below for more deepac commands.

Optional: download and compile the latest deepac-live custom models DOI:

docker run -v $(pwd):/deepac -u $(id -u):$(id -g) --rm --gpus all dacshpi/deepac:0.14.1 deepac --fetch

For more information about the usage of the NVIDIA container toolkit (e.g. selecting the GPUs to use), consult the User Guide.

The dacshpi/deepac:latest corresponds to the latest version of the CPU build. We recommend using explicit version tags instead.

With pip

We recommend setting up an isolated conda environment (see above). Alternatively, you can use a virtualenv virtual environment (note that deepac requires python 3):

# use -p to use the desired python interpreter (python 3.6 or higher required)
virtualenv -p /usr/bin/python3 my_env
source my_env/bin/activate

You can then install DeePaC with pip. For GPU support, you need to install CUDA and CuDNN manually first (see TensorFlow installation guide for details). Then you can do the same as above:

pip install deepac

Optional: download and compile the latest deepac-live custom models DOI:

deepac getmodels --fetch

If you want to install the plugins, use:

pip install deepacvir deepacstrain

Optional: run tests

Optionally, you can run explicit tests of your installation. Note that it may take some time on a CPU.

# Run standard tests
deepac test
# Run quick tests (eg. on CPUs)
deepac test -q
# Test using specific GPUs (here: /device:GPU:0 and /device:GPU:1) 
deepac test -g 0 1
# Test explainability and gwpa workflows
deepac test -xp
# Full tests
deepac test -a
# Full quick tests (eg. on GPUs with limited memory)
deepac test -aq

Help

To see help, just use

deepac --help
deepac predict --help
deepac train --help
# Etc.

Basic use: prediction

You can predict pathogenic potentials with one of the built-in models out of the box:

# A rapid CNN (trained on IMG/M data)
deepac predict -r input.fasta
# A sensitive LSTM (trained on IMG/M data)
deepac predict -s input.fasta

The rapid and the sensitive models are trained to predict pathogenic potentials of novel bacterial species. For details, see https://doi.org/10.1093/bioinformatics/btz541 or https://www.biorxiv.org/content/10.1101/535286v3.

To quickly filter your data according to predicted pathogenic potentials, you can use:

deepac predict -r input.fasta
deepac filter input.fasta input_predictions.npy -t 0.5

Note that after running predict, you can use the input_predictions.npy to filter your fasta file with different thresholds. You can also add pathogenic potentials to the fasta headers in the output files:

deepac filter input.fasta input_predictions.npy -t 0.75 -p -o output-75.fasta
deepac filter input.fasta input_predictions.npy -t 0.9 -p -o output-90.fasta

Advanced use

Config templates

To get the config templates in the current working directory, simply use:

deepac templates

Preprocessing

For more complex analyzes, it can be useful to preprocess the fasta files by converting them to binary numpy arrays. Use:

deepac preproc preproc_config.ini

See the config_templates directory of the GitLab repository (https://gitlab.com/rki_bioinformatics/DeePaC/) for a sample configuration file.

Training

You can use the built-in architectures to train a new model:

deepac train -r -T train_data.npy -t train_labels.npy -V val_data.npy -v val_labels.npy
deepac train -s -T train_data.npy -t train_labels.npy -V val_data.npy -v val_labels.npy

To train a new model based on you custom configuration, use

deepac train -c nn_train_config.ini

If you train an LSTM on a GPU, a CUDNNLSTM implementation will be used. To convert the resulting model to be CPU-compatible, use deepac convert. You can also use it to save the weights of a model, or recompile a model from a set of weights:

# Save model weights and convert the model to an equivalent with the same architecture and weights.
# Other config parameters can be adjusted
deepac convert model_config.ini saved_model.h5
# Recompile the model
deepac convert saved_model_config.ini saved_model_weights.h5 -w

Evaluation

To evaluate a trained model, use

# Read-by-read performance
deepac eval -r eval_config.ini
# Species-by-species performance
deepac eval -s eval_species_config.ini
# Ensemble performance
deepac eval -e eval_ens_config.ini

See the configs directory for sample configuration files. Note that deepac eval -s requires precomputed predictions and a csv file with a number of DNA reads for each species in each of the classes.

TPU (experimental)

If you want to use a TPU, run DeePaC with the --tpu parameter:

# Test a TPU
deepac --tpu colab test

Intepretability workflows

Filter visualization

To find the most relevant filters and visualize them, use the following minimum workflow:

# Calculate filter and nucleotide contibutions (partial Shapley values) for the first convolutional layer
# using mean-centered weight matrices and "easy" calculation mode
deepac explain fcontribs -m model.h5 -eb -t test_data.npy -N test_nonpatho.fasta -P test_patho.fasta -o fcontribs 

# Create filter ranking
deepac explain franking -f fcontribs/filter_scores -y test_labels.npy -p test_predictions.npy -o franking

# Prepare transfac files for filter visualization (weighted by filter contribution)
deepac explain fa2transfac -i fcontribs/fasta -o fcontribs/transfac -w -W fcontribs/filter_scores

# Visualize nucleotide contribution sequence logos
deepac explain xlogos -i fcontribs/fasta -s fcontribs/nuc_scores -I fcontribs/transfac -t train_data.npy -o xlogos

You can browse through other supplementary functionalities and parameters by checking the help:

deepac explain -h
deepac explain fcontribs -h
deepac explain xlogos -h
# etc.

Genome-wide phenotype potential analysis (GWPA)

To find interesting regions of a whole genome, use this workflow to generate nucleotide-resolution maps of predicted phenotype potentials and nucleotide contributions:

# Fragment the genomes into pseudoreads
deepac gwpa fragment -g genomes_fasta -o fragmented_genomes

# Create the .genome annotation files from .gff3 files corresponding to your .fasta files
deepac gwpa gff2genome genomes_gff3 genomes_genome

# Predict the pathogenic potential of each pseudoread
deepac predict -r -a fragmented_genomes/sample1_fragmented_genomes.npy -o predictions/sample1_pred.npy

# Create bedgraphs of mean pathogenic potential at each position of the genome
# Can be visualized in IGV
deepac gwpa genomemap -f fragmented_genomes -p predictions -g genomes_genome -o bedgraph

# Rank genes by mean pathogenic potential
deepac gwpa granking -p bedgraph -g genomes_gff -o granking

# Create bedgraphs of mean nuclotide contribution at each position of the genome
# Can be visualized in IGV
deepac gwpa ntcontribs -m model.h5 -f fragmented_genomes -g genomes_genome -o bedgraph_nt

You can browse through other supplementary functionalities and parameters by checking the help:

deepac gwpa -h
deepac gwpa genomemap -h
deepac gwpa ntcontribs -h
# etc.

Filter enrichment analysis

Finally, you can check for filter enrichment in annotated genes or other genomic features:

# Get filter activations, genome-wide
deepac gwpa factiv -m model.h5 -t fragmented_genomes/sample1_fragmented_genomes.npy -f fragmented_genomes/sample1_fragmented_genomes.fasta -o factiv

# Check for enrichment within annotated genomic features
deepac gwpa fenrichment -i factiv -g genomes_gff/sample1.gff -o fenrichment

Supplementary data and scripts

Datasets are available here: DOI (bacteria) and here: DOI (viruses). In the supplement_paper directory you can find the R scripts and data files used in the papers for dataset preprocessing and benchmarking.

Erratum

The second sentence in section 2.2.3 of the bacterial DeePaC paper (https://doi.org/10.1093/bioinformatics/btz541) is partially incomplete.

Published text: “All were initialized with He weight initialization (He et al, 2015) and trained…”

Should be: “All were initialized with He weight initialization (He et al, 2015) or Glorot initialization (Glorot & Bengio, 2010) for recurrent and feedforward layers respectively and trained…

Known issues

Unfortunately, the following issues are independent of the DeePaC codebase:

  • pip installation of pybedtools (a deepac dependency) requires libz-dev and will fail if it is not present on your system. To solve this, install libz-dev or use the bioconda installation.
  • A bug in TF 2.2 may cause training to hang when using Keras Sequence input (i.e. if your training config contains Use_TFData = False and LoadTrainingByBatch = True). To solve this, use TF 2.1 or TF 2.3+, pre-load your data into memory (LoadTrainingByBatch = False) or use TFDataset input (Use_TFData = True).
  • A bug in TF 2.1 resets the optimizer state when continuing interrupted training. DeePaC will notice that and warn you, but to solve this, upgrade to TF 2.2+.
  • h5py>=3.0 is not compatible with Tensorflow at the moment and will cause errors when loading Keras (and DeePaC) models (hence, deepac tests will fail as well). Conda installation takes care of it automatically, but the pip Tensorflow installation does not. To solve it, use conda installation or install h5py<3.0. This issue should be resolved in a future version of Tensorflow.
  • shap 0.38 requires IPython but the pip installer does not install it. Manual installation solves the problem.

Cite us

If you find DeePaC useful, please cite:

@article{10.1093/bioinformatics/btz541,
    author = {Bartoszewicz, Jakub M and Seidel, Anja and Rentzsch, Robert and Renard, Bernhard Y},
    title = "{DeePaC: predicting pathogenic potential of novel DNA with reverse-complement neural networks}",
    journal = {Bioinformatics},
    volume = {36},
    number = {1},
    pages = {81-89},
    year = {2020},
    month = {01},
    issn = {1367-4803},
    doi = {10.1093/bioinformatics/btz541},
    url = {https://doi.org/10.1093/bioinformatics/btz541},
    eprint = {https://academic.oup.com/bioinformatics/article-pdf/36/1/81/31813920/btz541.pdf},
}

@article{10.1093/nargab/lqab004,
    author = {Bartoszewicz, Jakub M and Seidel, Anja and Renard, Bernhard Y},
    title = "{Interpretable detection of novel human viruses from genome sequencing data}",
    journal = {NAR Genomics and Bioinformatics},
    volume = {3},
    number = {1},
    year = {2021},
    month = {02},
    issn = {2631-9268},
    doi = {10.1093/nargab/lqab004},
    url = {https://doi.org/10.1093/nargab/lqab004},
    note = {lqab004},
    eprint = {https://academic.oup.com/nargab/article-pdf/3/1/lqab004/36165658/lqab004.pdf},
}

@article{10.1093/bib/bbab269,
    author = {Bartoszewicz, Jakub M and Genske, Ulrich and Renard, Bernhard Y},
    title = "{Deep learning-based real-time detection of novel pathogens during sequencing}",
    journal = {Briefings in Bioinformatics},
    volume = {22},
    number = {6},
    year = {2021},
    month = {07},
    issn = {1477-4054},
    doi = {10.1093/bib/bbab269},
    url = {https://doi.org/10.1093/bib/bbab269},
    note = {bbab269},
    eprint = {https://academic.oup.com/bib/article-pdf/22/6/bbab269/41088711/bbab269.pdf},
}

@article{10.1093/bioinformatics/btac495,
    author = {Bartoszewicz, Jakub M and Nasri, Ferdous and Nowicka, Melania and Renard, Bernhard Y},
    title = "{Detecting DNA of novel fungal pathogens using ResNets and a curated fungi-hosts data collection}",
    journal = {Bioinformatics},
    volume = {38},
    number = {Supplement_2},
    pages = {ii168-ii174},
    year = {2022},
    month = {09},
    issn = {1367-4803},
    doi = {10.1093/bioinformatics/btac495},
    url = {https://doi.org/10.1093/bioinformatics/btac495},
    eprint = {https://academic.oup.com/bioinformatics/article-pdf/38/Supplement\_2/ii168/45884230/btac495.pdf},
}

About

Predicting pathogenic potentials of short DNA reads with reverse-complement deep neural networks.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published