fcgr-cnn

Frequence Chaos Game Representation with Deep Learning

Create a virtual environment and install packages

python -m venv env
source env/bin/activate
pip install -r requirements.txt

Set parameters for the experiment in parameters.py

• See (and include) preprocessing functions at preprocessing.py

Run codes in this order

1. undersample_sequences.py
2. extract_sequences.py
3. split_data.py will create a file datasets.json with train, validation and test sets
4. train.py
5. test.py
6. classification_metrics
7. clustering_metrics

A folder data/ will be created to save all intermediate results:

• <SPECIE>/ with all sequences extracted individually in the fasta file, in separated folders by label (Clade)
• train/ will contain
  • undersample_by_clade.csv
  • available_by_clade.csv a summary of the available sequences by clade, subject to the restrictions made in undersample_sequences.py(remove duplicates and empty rows)
  • selected_by_clade.csv a summary of the selected sequences by clade
  • checkpoints/ will save the best weights during training.
  • preprocessing.json will save a list with the preprocessing applied to each FCGR during training.
  • training_log.csv: accuracy and loss and learning rate per epoch for train and validation sets.
  • test/ will save all the metrics (classification and clustering) resulting from the evaluation of the best model on the test set.
  • plots accuracy and loss plots during training, confusion matrix
  • saliency_maps/ representative FCGR by clade, saliency map and relevant k-mers for that representative.

A folder fcgr-<KMER>-mer/ will contain all the FCGR created from the sequences in data/<SPECIE>