Deep learning for protein secondary structure prediction: Pre and post-AlphaFold

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Deep neural networks (DNNs) are the leading method for protein secondary structure prediction (PSSP).
• DNNs have improved PSSP accuracy to over 80% for three-state prediction.
• Various DNN architectures and techniques are employed in PSSP.

Purpose of the Study:

• To comprehensively review trends and challenges in DNNs for PSSP.
• To highlight advancements and future directions in the field.
• To identify areas for further accuracy improvement in PSSP.

Main Methods:

• Review of deep learning methods including CNNs, RNNs, Inception, and GNNs.
• Application of NLP and computer vision techniques like attention mechanisms, ResNet, and U-Net.
• Exploration of protein language models and evolutionary information for PSSP input.

Main Results:

• DNNs have significantly enhanced PSSP accuracy.
• Integration of NLP-inspired methods and advanced architectures are key trends.
• Pre-trained language models offer a promising new input strategy for PSSP.

Conclusions:

• The field of PSSP is rapidly evolving with DNNs.
• Opportunities exist to enhance evolutionary information and leverage protein language models.
• Further improvements are needed to reach the theoretical accuracy limits in PSSP.