An implementation for Deep Supervised and Convolutional Generative Stochastic Network for Protein Secondary Structure Prediction[1]. Convolutional Generative Stochastic Network is a particular variant of Generative Stochastic Network (Y.Bengio 2013)[2].
The model is evaluated on CullPDB_profile_6133[3] and CB513[4] and are avaiable here: http://www.princeton.edu/~jzthree/datasets/ICML2014/
Model is structured by:
- 2 convolutional layers (activaction functions: tanh) and 1 mean pooling layer.
- 2 convolutional transposed layers, and 1 upsampling layer.
Model generates for every input 12 samples, where the last is the prediction. Moreover generates 12 latent state(H0...H11), in the figure there are 3 latent state for convenience.
Weights are initializated according to Xavier Inizializer [5] and are shared tvhrough the network.
To regularize is used L2 regularization.
Loss function is Binary cross entropy, and Adam the optimizer.
| Q8 Accuracy | Loss |
|---|---|
 |
 |
| Q8 Accuracy | Loss |
|---|---|
 |
 |
[1] Jian Zhou, Olga G. Troyanskaya (2014) - Deep Supervised and Convolutional Generative Stochastic Network for Protein Secondary Structure Prediction
[2] Yoshua Bengio, Eric Thibodeau-Laufer, Guillaume Alain, and Jason Yosinski (2013) - Deep Generative Stochastic Networks Trainable by Backprop
[3] Guoli Wang and Roland L. Dunbrack Jr (2003) - PISCES: a protein sequence culling server
[4] Cuff JA, Barton GJ (1999) - Evaluation and Improvement of Multiple Sequence Methods for Protein Secondary Structure Prediction
[5] Xavier Glorot and Yoshua Bengio (2010) - Understanding the difficulty of training deep feedforward neural networks. International conference on artificial intelligence and statistics.