A multi-context feature learning approach to identify disease-specific gene neighborhoods

A network-based representation learning framework which uses both co-expressed and functional gene pairs to learn continuous gene representations.

Training the model

Our model requires two sets of input gene pairs (co-expressed and functionally related). Both the input files should be space-seperated format as shown below:

SERPINB3 SERPINB4
ARMC3 CFAP52
C9orf24 CFAP52
SNTN CFAP52

Then our feature learning methodology can be implemented using the below syntax:

python main_handler.py --infile co_expressed_pairs.txt --fun_infile functional_pairs.txt

We include both sets of inputs used in our manuscript in the data folder of the repository. To see the entire list of parameters/options:

python main_handler.py --help

Using pre-trained embeddings

Our model can also be trained using a set of pre-trained gene embeddings using the below command.

python pre_trained_handler.py --infile co_expressed_pairs.txt --fun_infile functional_pairs.txt 
                              --init_emb embeddings.txt

The input gene embeddings should be provided in a space-delimited file with the first column containing the gene identified or a symbol. The remaining N columns in each row represent the N-dimensional gene representation.

These pre-trained embeddings could be from an earlier training iteration or representations of genes learned within the same disease context.

Code dependencies

Our method was tested in Python 3.7. The required dependencies or packages include PyTorch, numpy, pandas.