This forked version of Gustav Madslund and Mikkel Møller Brusen's work implements prediction of multi-labelled datasets of the subcellular localization (subcel). All installation is done as previous. The original branch take the subcellular localization encoded as an n x 1 array of integers (1-10), where n is the number of proteins. This fork takes its input as an n x 10 array of integers (0,1) where 0/1 indicate the presents/absences of a particular location. E.g. Old format: 2 -> new format: [0,1,0,0,0,0,0,0,0,0]. Additionally error function was changed to accommodate this new option for multi-labelling, and the output metrics was changed to output F1 and Exact match as these are more meaningful metrics when multi-labels are possible.
secpred part of the code have been left untoched and untested. forked by MortenSkovsted
The following is left (almost) unchanged from the original readme: This is the code repository that accompany the master thesis by Gustav Madslund and Mikkel Møller Brusen.
The goal of the project was to evaluate pre-trained amino acid embeddings in protein prediction tasks
The software is coded in Python 3.6 using the Pytorch 1.1 version. To run the software smoothly, it is recommended to use those versions.
The code was made to be run on a CUDA GPU but can run on CPU too, although this will take forever... In order to run configurations that utilize the bi-direction pre-trained embeddings, at least 17GB RAM.
In order to run all configurations the following datasets are needed:
- Deeploc which is the deeploc dataset encoded as profiles
- Deeploc_raw which is the deeploc dataset without encoding (raw sequences)
- SecPred which is the filtered CullPDB dataset encoded as profiles. Files with _no_x have X replaced by A.
- SecPred_raw which is the CB513 dataset without encoding (raw sequences). X has been replaced by A.
all of which can be downloaded here {Edit:See comment in the top!!!:MortenSkovsted}
The datasets should then be positioned in the data/ directory similarly to the already included Deeploc_raw dataset.
Model architecture and other settings are controlled by config files in the configs/{task} directory. Each config is task specific, such that subcellular localization configurations can be found in configs/subcel directory and secondary structure prediction configurations can be found in configs/secpred directory.
To start training a model, we need to first give the task as argument when running main.py e.g. subcel and then choose a configuration with --config. For example, if we want to train with the configuration configs/subcel/deeploc_raw, we should use the following command:
python3 main.py subcel --config deeploc_raw
All configurations are created such that no hyperparameters needs to be specified, although they are possible if you want to do experiments with a specific configuration. For a list of all avaiable arguments the following commands are usefull:
python3 main.py --help
python3 main.py subcel --help
python3 main.py secpred --help
The best models based on validation performance will be saved under save/{task}/{config_name}/ where task can be subcel or secpred and config_name is the configuration that is training.