Gene Ontology Capsule GAN: an improved architecture for protein function prediction

Area of Science:

• Structural biology
• Computational biology
• Machine learning in bioinformatics

Background:

• Proteins, essential for life, are amino acid chains folded into 3D structures.
• Convolutional neural networks (CNNs) are used for protein function prediction from sequences.
• CNNs face limitations in detecting spatial hierarchies due to computation cost and translational invariance.

Purpose of the Study:

• To explore capsule networks for protein function prediction, overcoming CNN limitations.
• To leverage capsule networks' ability to capture hierarchical spatial information.
• To improve the accuracy of predicting protein functions using advanced deep learning models.

Main Methods:

• Utilized capsule networks, a deep learning architecture focusing on hierarchical links.
• Developed Gene Ontology Capsule GAN (GOCAPGAN) for protein function prediction.
• Compared the performance of capsule networks against standard CNNs.

Main Results:

• Capsule networks demonstrated improved accuracy in predicting protein functions.
• GOCAPGAN achieved an F1 score of 82.6%.
• GOCAPGAN obtained a precision score of 90.4% and a recall score of 76.1%.

Conclusions:

• Capsule networks offer significant potential for addressing challenges in structural biology.
• GOCAPGAN represents an advancement in deep learning for protein function prediction.
• The study highlights the advantage of capsule networks in capturing spatial relationships crucial for biological functions.