GOlite

description

GOlite is a python package that trains different Neural Network models for protein function prediction. It adapts to different encodings by building dynamic structures. It's designed to train on encoded data produced by protEncoder.

Available Structures

1. Convolutional Neural Network (CNN)

In this package, 1D and 2D are available to choose from based on the input shape. The network structure uses the following elements:

• binary cross-entropy loss
• Adam optimizer
• ReLU activation function in the hidden layers
• Sigmoid activation function in the final classification layer
• filters size changes in each layer and it's selected by the user and follows this pattern: start,end,step
• filters count is selected by the user
• one fully connected layer in the end
• MaxPooling layer is added after each filter in order to avoid overfitting

2. Dense Convolutional Network (DenseNet)

figures from [1]

DenseNet implementation in Keras is not as flexible as CNN, therefore, we had to use the provided structures without adjustments. There are three prebuilt designs with different depths: DenseNet121, DenseNet169, and DenseNet201. DenseNet is considere superior to CNN in the image classification field in terms of efficiency and accuracy.

Input

for more input options see the package documentation:

GOlite --help

Output

Training

• Model folder for each epoch.
  • Folder name pattern: outPrefix_modelType_epochNumber
• Five Images plotting the following measurements across training and validation subsets in all epochs so far.
  • AUC
  • Loss
  • Mean Squared Error (MSE)
  • Categorical Accuracy
  • Categorical Cross Entropy
• Five npy files storing the model performance measurements in all the previous epochs

Prediction

• One npy file for each batch storing encoded labels vectors.
  • File name pattern: inputFile_prdcts.npy
• One npy file for each batch storing the probabilities for each label and for all proteins.
  • File name pattern: inputFile_prdctsCert.npy

Getting started

Installing the package

• Download the latest release: releases

• In your command line environment:

pip install path/to/GOlite-x.x.x-py2.py3-none-any.whl

Run an example

• In your Command Line environment:

GOlite -d m1000_s50_n1000_part*_onehot.npy" -l m1000_s50_n1000_part*_FGOA.npy" -o oneHot -e 100 -b 500 -f 64 -s 8,64,8

Performance

Fmax scores of GOlite models

These scores were acquired following CAFA3 instructions

Fmax scores ot the top GOlite models agains the top models in CAFA3 and published Deep-Learning-Based models

These scores are acquired using NK, and full evaluation modes

For more details on the experiment methodology please refere to my thesis or contact me through the email