Redundancy-weighting the PDB for detailed secondary structure prediction using deep-learning models.

Area of Science:

• Structural biology
• Bioinformatics
• Machine learning

Background:

• The Protein Data Bank (PDB) contains imbalanced data, often leading to the use of nonredundant (NR) subsets.
• Redundancy-weighting (RW) offers an alternative by down-weighting, not discarding, redundant entries.
• Machine learning (ML) methods for secondary structure prediction (SSP) can benefit from improved data handling.

Purpose of the Study:

• To compare the performance of deep-learning models for SSP when trained on RW versus NR datasets.
• To evaluate the impact of RW datasets on predicting different classes of secondary structures.

Main Methods:

• Deep-learning models were trained using both RW and NR datasets derived from the PDB.
• The prediction accuracy of these models was evaluated for 3-class (HCE), 8-class (DSSP), and 13-class (STR2) secondary structures.

Main Results:

• Models trained on RW datasets consistently outperformed those trained on NR datasets.
• Improved prediction accuracy was observed across all evaluated secondary structure classification schemes (HCE, DSSP, STR2).

Conclusions:

• Training deep-learning models on RW datasets enhances secondary structure prediction accuracy.
• RW datasets provide a more informative data source for ML in structural biology, potentially enabling more detailed secondary structure classifications.