MUST-CNN: A Multilayer Shift-and-Stitch Deep Convolutional Architecture for Sequence-Based Protein Structure Prediction (2016-02-12T00:00:00.000000Z)

TL;DR

This work proposes a deep convolutional neural network architecture, MUST-CNN, that uses a novel multilayer shift-and-stitch technique to generate fully dense per-position predictions on protein sequences and beats the state-of-the-art performance on two large protein property prediction datasets.

Abstract

Predicting protein properties such as solvent accessibility and secondary structure from its primary amino acid sequence is an important task in bioinformatics. Recently, a few deep learning models have surpassed the traditional window based multilayer perceptron. Taking inspiration from the image classification domain we propose a deep convolutional neural network architecture, MUST-CNN, to predict protein properties. This architecture uses a novel multilayer shift-and-stitch (MUST) technique to generate fully dense per-position predictions on protein sequences. Our model is significantly simpler than the state-of-the-art, yet achieves better results. By combining MUST and the efficient convolution operation, we can consider far more parameters while retaining very fast prediction speeds. We beat the state-of-the-art performance on two large protein property prediction datasets.

Authors

Zeming Lin

4 papers

Jack Lanchantin

4 papers

Yanjun Qi

3 papers

TL;DR

Abstract

Authors

References40 items

JPred4: a protein secondary structure prediction server

Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification

Protein Secondary Structure Prediction with Long Short Term Memory Networks

Fully convolutional networks for semantic segmentation

A series of PDB-related databanks for everyday needs

Going deeper with convolutions

SSpro/ACCpro 5: almost perfect prediction of protein secondary structure and relative solvent accessibility using profiles, machine learning and structural similarity

Sequence to Sequence Learning with Neural Networks

Recurrent Neural Network Regularization

Neural Machine Translation by Jointly Learning to Align and Translate

Convolutional Neural Networks for Sentence Classification

Recurrent Convolutional Neural Networks for Scene Labeling

Deep Supervised and Convolutional Generative Stochastic Network for Protein Secondary Structure Prediction

OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks

Recurrent Convolutional Neural Networks for Scene Parsing

Practical Bayesian Optimization of Machine Learning Algorithms

A Unified Multitask Architecture for Predicting Local Protein Properties

Deep Sparse Rectifier Neural Networks

Natural Language Processing (Almost) from Scratch

Protein 8-class secondary structure prediction using Conditional Neural Fields

A series of PDB related databases for everyday needs

Evaluation of Pooling Operations in Convolutional Architectures for Object Recognition

A unified architecture for natural language processing: deep neural networks with multitask learning

Application of multiple sequence alignment profiles to improve protein secondary structure prediction

Protein secondary structure prediction based on position-specific scoring matrices.

Gapped BLAST and PSI-BLAST: a new generation of protein database search programs.

Dropout: a simple way to prevent neural networks from overfitting

and Baldi

Model Selection (Small model) We use Bayesian Optimization

Torch7: A Matlab-like Environment for Machine Learning

and Weston

Gradient-based learning applied to document recognition

Summary: End-to-End architecture February 16

Time Training of the small model takes 6 hours, while training of the large model takes one day

Propose a multilayer shift-and-stitch technique for deep CNNs, which signiﬁcantly speeds up training and test time and increases the size of the model we can train

ssp A collapsed version of the 8 class prediction task, since many protein secondary structure prediction algorithms use a 3 class approach instead of the 8-class approach given in dssp

sar Relative solvent accessibility

Model Selection (Large model

Beat the state-of-the-art performance on two large datasets of protein property prediction tasks

Pro-ceedings

Field of Study

Venue Information

Name

Type

URL

Alternate Names