Convolutional Bidirectional LSTM with attention mechanism for predicting protein subcellular localization. The model was trained using the MultiLoc dataset, which counts with 5959 proteins.
Jose Juan Almagro Armenteros, DTU Bioinformatics
There are two files in the data folder: 1) "test.npz" independent set to calculate the final performance of the model 2) "train.npz" training set.
Each file includes a numpy array with the proteins sequences already encoded in profiles, a numpy array with the masks of each sequence and a numpy vector with the target of each protein.
The training is performed running the script "train.py". This is a minimal example:
python train.py -i train.npz -t test.npz
The default options are the optimals one, but the training will be really slow on CPU.
To run it on GPU use these flags before the command
THEANO_FLAGS=device=gpu0,dnn.conv.algo_fwd=time_once,dnn.conv.algo_bwd_filter=time_once,dnn.conv.algo_bwd_data=time_once,warn_float64=warn python train.py -i train.npz -t test.npz
Ole Winther, DTU Compute Henrik Nielsen, DTU Bioinformatics Søren and Casper Kaae Sønderby, University of Copenhagen