[Paper Review] MUFold-SS: Protein Secondary Structure Prediction Using Deep Inception-Inside-Inception Networks

Chao Fang, Yi Shang+1 more|arXiv (Cornell University)|Sep 12, 2017

Advanced Malware Detection Techniques27 references5 citations

TL;DR

MUFold-SS introduces Deep3I, a deep inception-inside-inception neural network that predicts protein secondary structure using both amino acid sequences and PSI-BLAST profiles. It achieves state-of-the-art performance with 82.8% Q3 and 71.1% Q8 accuracy on the CB513 benchmark, outperforming existing methods in both accuracy and speed.

ABSTRACT

Motivation: Protein secondary structure prediction can provide important information for protein 3D structure prediction and protein functions. Deep learning, which has been successfully applied to various research fields such as image classification and voice recognition, provides a new opportunity to significantly improve the secondary structure prediction accuracy. Although several deep-learning methods have been developed for secondary structure prediction, there is room for improvement. MUFold-SS was developed to address these issues. Results: Here, a very deep neural network, the deep inception-inside-inception networks (Deep3I), is proposed for protein secondary structure prediction and a software tool was implemented using this network. This network takes two inputs: a protein sequence and a profile generated by PSI-BLAST. The output is the predicted eight states (Q8) or three states (Q3) of secondary structures. The proposed Deep3I not only achieves the state-of-the-art performance but also runs faster than other tools. Deep3I achieves Q3 82.8% and Q8 71.1% accuracies on the CB513 benchmark.

Motivation & Objective

• To improve the accuracy of protein secondary structure prediction using deep learning.
• To develop a faster and more accurate deep neural network architecture for secondary structure prediction.
• To integrate sequence and profile information effectively using a novel network design.
• To achieve state-of-the-art performance on standard benchmarks like CB513.
• To provide a publicly available tool for high-accuracy secondary structure prediction.

Proposed method

• The method employs a deep inception-inside-inception network (Deep3I) with multiple inception modules stacked in a hierarchical manner.
• The network takes two inputs: the primary amino acid sequence and a position-specific scoring matrix (PSSM) generated by PSI-BLAST.
• The architecture uses 1x1, 3x3, and 5x5 convolutional filters within each inception module to capture multi-scale features.
• Skip connections and batch normalization are used to stabilize training and improve gradient flow.
• The final layer outputs probabilities for eight secondary structure states (Q8) or three states (Q3).
• The model is trained end-to-end using stochastic gradient descent with cross-entropy loss.

Experimental results

Research questions

• RQ1 Can a deep inception-inside-inception network architecture improve secondary structure prediction accuracy beyond existing deep learning methods?
• RQ2 How does combining sequence and profile information through a deep CNN enhance prediction performance?
• RQ3 Does the proposed Deep3I architecture achieve faster inference times while maintaining high accuracy?
• RQ4 To what extent does the model generalize across diverse protein families on standard benchmarks?
• RQ5 Can the network design reduce overfitting and improve training stability in secondary structure prediction?

Key findings

• MUFold-SS achieves a Q3 accuracy of 82.8% on the CB513 benchmark, representing state-of-the-art performance.
• The model attains a Q8 accuracy of 71.1% on the same benchmark, significantly outperforming previous methods.
• The Deep3I architecture enables faster inference compared to other deep learning-based secondary structure predictors.
• The integration of PSI-BLAST profiles with sequence data improves prediction robustness and accuracy.
• The hierarchical inception design effectively captures multi-scale sequence and evolutionary features.
• The model demonstrates strong generalization across diverse protein families in the benchmark dataset.