csb5/metagenn

MetageNN is a proof of concept neural network used to obtain a memory-efficient taxonomic classifier robust to sequencing errors and missing genomes

MetageNN

MetageNN is a proof of concept memory-efficient long-read taxonomic classifier that is robust to sequencing errors and missing genomes. MetageNN is based on a neural network model that uses short k-mer profiles of sequences to reduce the impact of “distribution shift” when extrapolating from training on genome sequences to testing on error-prone long reads. MetageNN can be used on sequences not classified by conventional methods and offers an alternative approach for memory-efficient classifiers that can be optimized further.

Requirements

bash install_requirements.sh

Data

You can find the list of genomes used to train (either small or the main database) as well as the list of isolates used to test at /data. A link is also provided to download the "small database" training dataset (1x coverage) that can be used to train MetageNN.

Counting k-mers

MetageNN can be trained using any sequence length. For our proof of concept, we sampled sequences of 1kbp from genomes. To count the k-mers of these sequences, we used the Phylopythia k-mer counting algorithm [1] by using the following command:

save_file = 'path_to_file/genome_segments.fasta'
fasta2kmers2 -i save_file -j 6 -k 6 -s 0 -l 0 -n 1 -f k_mers_counted.6mer        

In the example above, we counted canonical 6mers given an input file in fasta format.

Settings

In the /settings folder you will find JSON files containing the best hyperparameters for MetageNN for both databases (the small database and the main database). MetageNN can load these files during training.

Training

To train MetageNN on the small database of genomes you can run (please download the "small database" training dataset first found at /data):

python code/MetageNN_train.py -s settings/MetageNN_settings_small_database.json

Cite

Preprint to be released.

References

[1] https://github.com/algbioi/kmer_counting

Contact

For additional information, help and bug reports please email Rafael Peres da Silva ([email protected]).