ProteinUnetLM

ProteinUnetLM predicts protein secondary structures (SS) in an eight-class (SS8) scheme using a convolutional Attention U-Net to provide accurate secondary-structure annotations for protein characterization.

Key Features:

• Architecture: The model employs a convolutional Attention U-Net architecture for SS8 prediction, combining convolutional feature extraction with attention-based contextual integration.
• Performance Metrics: Training incorporates the Matthews correlation coefficient into the loss function to address class imbalance and evaluation uses the adjusted geometric mean (AGM).
• Comparative Advantage: ProteinUnetLM outperforms LSTM-based predictors such as SPOT-1D-LM and NetSurfP-3.0 for rare classes including 310-helix (G), beta-bridge (B), and high curvature loop (S), and on datasets with sequences without homologs, free-modeling targets, and chameleon sequences.

Scientific Applications:

• Protein characterization: Accurate SS8 predictions support protein characterization and downstream tasks including drug discovery, enzyme design, and elucidation of disease mechanisms.

Methodology:

The method uses a convolutional Attention U-Net architecture, a loss function incorporating the Matthews correlation coefficient to handle class imbalance, and evaluation with the adjusted geometric mean (AGM), with performance compared against LSTM-based predictors (SPOT-1D-LM, NetSurfP-3.0).