COSSMO

Competitive Splice Site Model

This is the code used to train the models described in the paper

Hannes Bretschneider, Shreshth Gandhi, Khalid Zuberi, Amit G Deshwar, and Brendan J Frey
COSSMO: Predicting Competitive Alternative Splice Site Selection using Deep Learning
Bioinformatics, Volume 34, Issue 13, 1 July 2018, Pages i429–i437,
https://doi.org/10.1093/bioinformatics/bty244

Installation

Dependencies

COSSMO requires Python 2.7 and TensorFlow 1.8. We provide a tested Conda environment.

Anaconda Python

We recommend using COSSMO with the conda package manager from Anaconda Python.

To create a Conda environment for running COSSMO, install either the environment.yml or environment_gpu.yml (includes TensorFlow with GPU support) file:

conda env create -f environment.yml

or

conda env create -f environment_gpu.yml  

Activate your environment with

conda activate cossmo

Pip

Alternatively you can install the dependencies via the requirements.txt file for pip:

pip install -f requirements.txt

Install COSSMO

After creating the dependencies you can install the COSSMO package. The following command will symlink the COSSMO package into your environment.

pip install -e .

Tests

To run the tests, first install a test runner in your environment. We recommend py.test:

conda install pytest

Running py.test from the root directory will discover all the test automatically:

pytest

Training

Training script

The training script is at bin/train_cossmo.py. You can call the script with the --help option to receive details of the parameters:

$ python bin/train_cossmo.py --help
usage: train_cossmo.py [-h] [--configuration-file CONFIGURATION_FILE]
                       [--gpu GPU] [--intra-op-threads INTRA_OP_THREADS]
                       [--inter-op-threads INTER_OP_THREADS] [--test-only]
                       [--fold FOLD]

optional arguments:
  -h, --help            show this help message and exit
  --configuration-file CONFIGURATION_FILE
                        Path to a configuration file, containing all
                        hyperparameters, model definitions, etc. See the
                        provided examples for details.
  --gpu GPU             GPU device ID to use for training. This is equivalent
                        to setting the CUDA_VISIBLE_DEVICES environment
                        variable. It is recommend to set this option when you
                        have more than one GPU device in your system to
                        prevent TensorFlow from claiming all devices.
  --intra-op-threads INTRA_OP_THREADS
                        See https://github.com/tensorflow/tensorflow/blob/26b4
                        dfa65d360f2793ad75083c797d57f8661b93/tensorflow/core/p
                        rotobuf/config.proto#L165 for the meaning of this
                        parameter.
  --inter-op-threads INTER_OP_THREADS
                        See https://github.com/tensorflow/tensorflow/blob/26b4
                        dfa65d360f2793ad75083c797d57f8661b93/tensorflow/core/p
                        rotobuf/config.proto#L165 for the meaning of this
                        parameter.
  --test-only           Don't train, only evaluate test set.
  --fold FOLD           Cross-validation fold to train on. When set, this
                        overrides the`cv_fold` key in the configuration file.

Configuration files

Configuration files to replicate the models described in the paper are available in configuration_files/. Before you can use these configuration files, you must edit them and provide the correct file paths for the dataset path and output path for your system.

COSSMO with BERT (https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertModel)

To run the new COSSMO model.py from Colab notebook, we will need to install Tensorflow 1.13 for the tensorflow.contrib to work.

pip install tensorflow==1.13.2

The next step would be to load the configuration file and store it in a variable config1, that will be passed as a parameter while calling the model to run.

The model can be called with

c.main(configuration = config1, continue_from=None, intra_op_threads=2, inter_op_threads=6, test_only=False)

An example ipynb file is attached to this repository with the code for these steps.