Post-processing enhances protein secondary structure prediction with second order deep learning and embeddings.

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Protein Secondary Structure Prediction (PSSP) is crucial for understanding protein function.
• Deep learning approaches offer promising advancements in PSSP accuracy.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for accurate PSSP.
• To assess the impact of input representations and post-processing techniques on prediction performance.

Main Methods:

• A Convolutional Neural Network (CNN) trained with the Subsampled Hessian Newton (SHN) method was employed.
• A two-dimensional input representation using embeddings from a pre-trained protein language model was utilized.
• Ensemble and filtering techniques were applied for post-processing.

Main Results:

• The CNN model achieved Q3 accuracies of 79.96% (CB513) and 81.45% (PISCES) without post-processing.
• Post-processing improved Q3 accuracy to 93.65% (CB513) and 87.13% (PISCES).
• On CASP13, Q3 accuracy reached 98.12% and SOV score reached 96.98% with optimized post-processing.

Conclusions:

• The proposed CNN model with SHN optimization and language model embeddings demonstrates high efficacy for PSSP.
• Post-processing techniques significantly enhance prediction accuracy.
• Embeddings can be as effective as traditional multiple sequence alignments for input representation.