prot2vec_contextualvec

This repository is for training prot2vec models and generating contextual vectors for S/T/Y sites in protein sequences.

Basic requirement

To run the scripts in this repository, you need to install

• Python >= 2.7.x

• Numpy >= 1.3

• SciPy >= 0.7

• Gensim >= 2.3.0

The detailed description for the word2vec and doc2vec model can be found in Distributed representations of words and phrases and their compositionality and Distributed representation of Sentences and Documents.

Training your own prot2vec

To train your own prot2vec model, you will need to prepare a database of protein sequences in the format of FASTA.

If the path to your protein sequence database is $path_to_database, you can train a prot2vec model by running the following script,

> ./run_prot2vec.sh $PATH_TO_DATABASE 

This training process could take several days depending on the size of your protein sequence database.

The trained prot2vec model will be under the directory DATA.

Hyper-parameters for training

The hyper-parameters for training the prot2vec models is the same with the one for training a doc2vec model.

You can change the hyper-parameters at line 9 in the script run_prot2vec.sh. These parameters include,

• size, which indicates the size of the vectors generated. The default values is set to 100.

• window, which indicate the maximum distance between the target words and its context words used for prediction. The default value is set to 25.

• train_alpha, which is the initial learning rate. The default value is set to 0.005.

• iter, which is the number of iteractions the training performed on the corpus. The default value is set to 400.

• negative, which indicates the number of examples sampled for negative sampling. The default value is set to 5.

• hs, which indicates whether hierarchical softmax is performed. The default value is set to 0.

The above hyper-parameters are derived from the doc2vec model, for which more detailed hyper-parameter descriptions can be found here.

Obtain the contextual vectors for protein sequences

After the training of the prot2vec model is completed, you can obtain the contextual feature vectors for S/T/Y sites from your own prot2vec model.

You should prepare the protein sequences for query in the format of FASTA.

If the path to this query file is $PATH_TO_QUERY, you can obtain the contextual vectors (of window size = 2*3+1) for Y sites by running the following script,

> get_context.sh $PATH_TO_QUERY $PATH_TO_PROT2VEC 3 Y

Here, the $PATH_TO_PROT2VEC refers to the path where your prot2vec model is saved. You can find it in directory OUTPUT after training. The $PATH_TO_PROT2VEC could be something like OUTPUT/uniprot-all_s100_w25_lr0.005_itr400_ns5_hs0.

The window size can be adjusted to any size that you would like to test and the site type can be one of ('S', 'T', 'Y' and '*') where * means all three types of sites.

By combining the generated contextual vectors with other residue-level vector of potential phosphorylation sites, you can train your own models for predicting phosphorylation sites.