PS4: a next-generation dataset for protein single-sequence secondary structure prediction.

Area of Science:

• Biochemistry and Structural Biology
• Computational Biology
• Bioinformatics

Background:

• Protein secondary structure prediction is crucial for understanding protein folding and tertiary structure determination.
• Current prediction models often rely on computationally intensive multiple sequence alignments.
• Limited availability of large, non-redundant datasets hinders the development of efficient secondary structure prediction algorithms.

Purpose of the Study:

• To introduce PS4, a novel and comprehensive dataset for protein secondary structure prediction.
• To facilitate the development of lightweight algorithms for secondary structure prediction.
• To overcome the data bottleneck in training accurate prediction models.

Main Methods:

• Compilation of 18,731 non-redundant protein chains with associated secondary structure labels to create the PS4 dataset.
• Ensuring dataset non-redundancy against commonly used secondary structure datasets in the literature.
• Performing ablation studies using PS4 for training secondary structure prediction models.

Main Results:

• The PS4 dataset contains 18,731 non-redundant protein chains and their secondary structure labels.
• Training on the PS4 dataset enabled secondary structure prediction algorithms to achieve state-of-the-art accuracy.
• The models demonstrated high performance on the CB513 test set in zero-shot evaluations.

Conclusions:

• The PS4 dataset is a valuable resource for advancing protein secondary structure prediction.
• Lightweight algorithms trained on PS4 can achieve high accuracy, reducing reliance on multiple sequence alignments.
• This work provides a foundation for improved tertiary structure prediction and protein folding studies.