This is the repository to the preprint-paper BERTax: taxonomic classification of DNA sequences with Deep Neural Networks and the published paper: Taxonomic classification of DNA sequences beyond sequence similarity using deep neural networks respectively.
The used data can be found under DOI 10.17605/OSF.IO/QG6MV or https://osf.io/qg6mv/
We provide a Docker container to run BERTax. Pull and run with:
docker run -t --rm -v /path/to/input/files:/in fkre/bertax:latest /in/sequences.faThe docker container can also be run with GPU-support, likely resulting in much faster predictions. For this, the nvidia-container-toolkit has to be installed, the bertax image has to be run with the flag --gpus all.
The image can be built locally (after cloning -- see below) with
docker build -t bertax bertaxInstall in new conda environment
conda create -n bertax -c fkretschmer bertaxActivate environment and install necessary pip-dependencies
conda activate bertax
pip install keras-bert==0.86.0First clone this repository
git clone https://github.com/rnajena/bertax.gitThen download the (big) model file into the resources subfolder.
Finally, install with pip
pip install -e bertaxThe script takes a (multi)fasta as input and outputs a list of predicted classes to the console:
bertax sequences.fastaOptions:
| parameter | explanation |
|---|---|
-o --output_file | write output to specified file (tab-separated format) instead of to the output stream (console) |
--conf_matrix_file | output confidences for all classes of all ranks to JSON file |
--sequence_split | how to handle sequences sequence longer than the maximum (window) size: split into equal chunks (equal_chunks, default) or use random sequence window (window) |
-C --maximum_sequence_chunks | maximum number of chunks to use per (long) sequence |
--running_window | if enabled, a running window approach is chosen to go over each sequence to make predictions |
--running_window_stride | stride for running window (default: 1) |
--custom_window_size | allows specifying a custom, smaller window size |
--chunk_predictions | output predictions per chunk, otherwise (by default) chunk predictions are averaged |
--output_ranks | specify which ranks to include in output (default: superkingdom phylum genus) |
--no_confidence | if set, do not include confidence scores in output |
--batch_size | batch size (i.e., how many sequence chunks to predict at the same time); can be lowered to decrease memory usage and increased for better performance (default: 32) |
-t --nr_threads | set the number of threads used (default: determine automatically) |
Note, that "unknown" is a special placeholder class for each prediction rank, meaning the sequence's taxonomy is predicted to be unlike any possible output class.
Default mode, sequences longer than 1500 nt are split into equal chunks, one prediction (average) per sequence
bertax sequences.faOnly use one random chunk per sequence (for sequences longer than 1500 nt)
bertax --sequence_split window sequences.faOnly output the superkingdom
bertax sequences.fa --output_ranks superkingdomPredict with a running window in 300 nt steps and output predictions for all chunks (no threshold for the number of chunks per sequence)
bertax -C -1 --running_window --running_window_stride 300 --chunk_predictions sequences.faIn the directory confusion_matrices you can find confusion matrices from the publication's results which indicate the classification quality. These matrices could not be included directly in the paper due to the vast amount and size of them.
It is possible to get a visualization of the underlying BERT model, based on bertviz for a specific DNA sequence. For this, additional dependencies have to be installed:
- torch
- transformers
- bertviz==1.0.0
An HTML file with interactive visualization can be created with:
bertax-visualize sequence.faAs visualization is quite performance-intensive for big sequences, parameters can be set
to only visualize a specific part (-a $start -n $size). Both an attention-head view and
model-view are available, set with the parameter --mode {head|model}.
The repository with the code used in the development of BERTax is located at https://github.com/rnajena/bertax_training.
Custom models trained with these scripts can be used in BERTax with the parameter --custom_model_file.
Due to the limited amount of samples that can be used for training, we could not train all known phyla and genera. Therefore, we present here the list of compatible phyla and genera. Note: If the taxon of your sample is not included in this list, there is a high probability that phylum/genus will be predicted as "unknown". If you want you can train your own model, that includes the taxa of interest to you.
Note: We recommend using BERTax only for super kingdom and phylum prediction, but genera are possible. For more details see: our paper at pnas.org
phylum
'Actinobacteria', 'Apicomplexa', 'Aquificae',
'Arthropoda', 'Artverviricota', 'Ascomycota', 'Bacillariophyta', 'Bacteroidetes',
'Basidiomycota', 'Candidatus Thermoplasmatota', 'Chlamydiae', 'Chlorobi',
'Chloroflexi', 'Chlorophyta', 'Chordata', 'Crenarchaeota', 'Cyanobacteria',
'Deinococcus-Thermus', 'Euglenozoa', 'Euryarchaeota', 'Evosea', 'Firmicutes',
'Fusobacteria', 'Gemmatimonadetes', 'Kitrinoviricota', 'Lentisphaerae', 'Mollusca',
'Negarnaviricota', 'Nematoda', 'Nitrospirae', 'Peploviricota', 'Pisuviricota',
'Planctomycetes', 'Platyhelminthes', 'Proteobacteria', 'Rhodophyta', 'Spirochaetes',
'Streptophyta', 'Tenericutes', 'Thaumarchaeota', 'Thermotogae', 'Uroviricota',
'Verrucomicrobia'
genus
'Acidilobus', 'Acidithiobacillus',
'Actinomyces', 'Actinopolyspora', 'Acyrthosiphon', 'Aeromonas', 'Akkermansia', 'Anas',
'Apis', 'Aquila', 'Archaeoglobus', 'Asparagus', 'Aspergillus', 'Astyanax', 'Aythya',
'Bdellovibrio', 'Beta', 'Betta', 'Bifidobacterium', 'Botrytis', 'Brachyspira',
'Bradymonas', 'Brassica', 'Caenorhabditis', 'Calypte', 'Candidatus Kuenenia',
'Candidatus Nitrosocaldus', 'Candidatus Promineofilum', 'Carassius', 'Cercospora',
'Chanos', 'Chlamydia', 'Chrysemys', 'Ciona', 'Citrus', 'Clupea', 'Coffea',
'Colletotrichum', 'Cottoperca', 'Crassostrea', 'Cryptococcus', 'Cucumis', 'Cucurbita',
'Cyanidioschyzon', 'Cynara', 'Cynoglossus', 'Daucus', 'Deinococcus', 'Denticeps',
'Desulfovibrio', 'Dictyostelium', 'Drosophila', 'Echeneis', 'Egibacter', 'Egicoccus',
'Elaeis', 'Equus', 'Erpetoichthys', 'Esox', 'Euzebya', 'Fervidicoccus', 'Frankia',
'Fusarium', 'Gadus', 'Gallus', 'Gemmata', 'Gopherus', 'Gossypium', 'Gouania',
'Helianthus', 'Ictalurus', 'Ktedonosporobacter', 'Legionella', 'Leishmania',
'Lepisosteus', 'Leptospira', 'Limnochorda', 'Malassezia', 'Manihot', 'Mariprofundus',
'Methanobacterium', 'Methanobrevibacter', 'Methanocaldococcus', 'Methanocella',
'Methanopyrus', 'Methanosarcina', 'Microcaecilia', 'Modestobacter', 'Monodelphis',
'Mus', 'Musa', 'Myripristis', 'Neisseria', 'Nitrosopumilus', 'Nitrososphaera',
'Nitrospira', 'Nymphaea', 'Octopus', 'Olea', 'Oncorhynchus', 'Ooceraea',
'Ornithorhynchus', 'Oryctolagus', 'Oryzias', 'Ostreococcus', 'Papaver', 'Perca',
'Phaeodactylum', 'Phyllostomus', 'Physcomitrium', 'Plasmodium', 'Podarcis', 'Pomacea',
'Populus', 'Prosthecochloris', 'Pseudomonas', 'Punica', 'Pyricularia', 'Pyrobaculum',
'Quercus', 'Rhinatrema', 'Rhopalosiphum', 'Roseiflexus', 'Rubrobacter', 'Rudivirus',
'Salarias', 'Salinisphaera', 'Sarcophilus', 'Schistosoma', 'Scleropages',
'Sedimentisphaera', 'Sesamum', 'Solanum', 'Sparus', 'Sphaeramia', 'Spodoptera',
'Sporisorium', 'Stanieria', 'Streptomyces', 'Strigops', 'Synechococcus', 'Takifugu',
'Thalassiosira', 'Theileria', 'Thermococcus', 'Thermogutta', 'Thermus', 'Tribolium',
'Trichoplusia', 'Ustilago', 'Vibrio', 'Vitis', 'Xenopus', 'Xiphophorus',
'Zymoseptoria'
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent