CAFA3
University of Turku CAFA3 project
Files are in the new machine in address: /home/sukaew/CAFA3
CNN experiment can be run with python train.py You'll need to copy the data folder from /home/kahaka/CAFA3/
Running preprocessing and sequence analysis
All preprocessing steps and sequence analyses can be run within the directory 'sequence_features' using the following command line.
python3 target_process.py -o [out_folder] -s [ori_seq]
The program needs two inputs, the out_folder and the ori_seq. The out_folder should be absolute path ended with '/' where the input ori_seq file/directory and the output features directory reside. The input ori_seq can be one of these four formats (folder of non-compressed fasta files, tar.gz, gz and zip). The sequence analysese include Blast Protein, DeltaBlast, Interproscan5, NetAcet, predGPI, nucPred and Taxonomy hierarchy. All analysis results are in folder called feature.
Running the Feature Generation, Classification and Analysis
All experiments can be run using the program run.py. The experimental code uses a three-step system. One or more of these actions can be performed using the command line option --action or --a. By default, all three actions (build, classify and statistics) are performed.
The run.py program can be called like this:
python run.py -e [TASK] -o [OUTPUT] --targets external
The [TASK] value can be one of cafa3, cafa3hpo or cafapi. Depending on task, different input files are used. The --targets option defines how to handle CAFA targets.
Making predictions with the neural model
cd neural
Download and extract data (data.tar.gz) and model files (features_only.tar.gz) from https://github.com/TurkuNLP/CAFA3/releases/tag/v0.0
python3 predict_new.py ./features_only/ ./data/devel_sequences.fasta.gz ./data/examples.json.gz ./devel_predictions.tsv.gz
This will use the trained model from ./features_only/ directory and make predictions for the target sequences. The input fasta file should not contain linebreaks within the sequences. examples.json.gz contains the pre-generated features. The last parameter is the output path.
Cross-validation
By default, the scikit-learn classification will use the train/devel/test split for the learning data. To use n-fold cross-validation instead, use the --fold option of run.py. To do 10-fold cross-validation, the program can be run 10 times using a script like this:
for FOLD in 0 1 2 3 4 5 6 7 8 9; do python run.py -o /tmp/CAFA10fold/fold$FOLD --fold $FOLD; done
Ensemble
The program ensemble.py can be used to combine predictions from different systems and the BLAST fallback baseline. To run the ensemble, use a command like:
python ensemble.py -a [PRED1_DIR] -b [PRED2_DIR] -o [OUTPUT] --baseline 4 --simple --terms 1000000 --write --cafa --clear