PP2 Exercise

We provide two different prediction models.

1. Word2Vec Embeddings with a feed-forward neural network in PyTorch
2. Google Seq2Seq Prediction

Running

To run the project, you need to use Python 3.6+. The required dependencies are listed below:

torch 
torchvision
numpy 
matplotlib
biovec 
sklearn 

To install the dependencies run

    pip install -r requirements.txt 

Feed-Forward Neural Network

Run Prediction

To run the protein-protein binding predictions

    python predict.py -i input_file.fasta -o output_file.fasta

Example input and output files can be found under: data/test_input.fasta and data/test_output.fasta. The script uses a pre-trained model that can be found under trained_models/ffnn_model.ckpt

Parse and Split the Data

The data needs to be split into a training and test set for later cross-valdiation and bootstrapping.

python scripts/preprocessing/split_data.py \
		--ppi_protvecs=scripts/preprocessing/ppi_as_vec.npy \
		--train_set=ppi_vec_train.npy \
		--test_set=ppi_vec_test.npy

Train Model

python scripts/ffnn/train_ffnn_w2v.py \
		--training_set=scripts/preprocessing/ppi_vec_train.npy \
		--model=trained_models/ffnn_model.ckpt \
		--num_epochs=100 \
		--batch_size=100

Validation

python scripts/ffnn/validation/cross_validation.py \
		--ppi_protvecs=scripts/preprocessing/ppi_vec_train.npy \
		--num_epochs=100 \
		--num_split=5

Bootstrapping

python scripts/ffnn/validation/bootstrapping.py \
		--test_set=scripts/preprocessing/ppi_vec_test.npy \
		--model=trained_models/ffnn_model.ckpt \
		--num_boot=1000

Visualisation

The results of cross-validation and bootstrapping are summarised in scripts/postprocessing/final_model.html