PhosTransfer

Requirements:

• python 2.7
• tensorflow
• Numpy
• matplotlib
• shutil
• sklearn
• PsiBlast
• PSIPRED (optional)
• DISOPRED3 (optional)

Preprocessing: feature extraction

Config #home_path, #uniref90_psi_blast_database, #blast_path, #tool_dir in src/prep_vec/util.py:

• home_path: path to saving generated feature files.
• uniref90_psi_blast_database: path to blast database. We use the uniref90filt database.
• blast_path: path to exculable psiblast. Install from PsiBlast.
• tool_dir: path to feature generation tools. Install from PSIPRED and DISOPRED3.

Pre-genearted features can be downloaded here. To regenerate features from scratch, please configure and run src/prep_vec/residue2vec.py.

Model training and testing

Config and run src/phospho_prediction.py.

• mode=cv: run cross validation. Trained models will be saved to OUTPUT/checkpoints and the best models will be saved to FINALS/models.
• mode=test: run independent test. Test performance will be saved to #output_log that is configured.

Pretrained models for different kinases can be downloaded here. The predicted results for our independent tests can be found in FINALS/predicts.

Benchmark dataset

Here we release the benchmark dataset for hierarchical phosphorylation site prediction. The DATA directory structure is as follows

• Combined_train
  • ST
    • sites: annotated phosphorylation sites
    • chains: amino acid sequences in Fasta format
    • ID.txt: protein identifications
    • sub-directories: children nodes
  • T (same as above)
• Combined_test
  • ST (same as above)
  • T (same as above)

The benchmark dataset can be downloaded here.