DMVFL-RSA

DMVFL-RSA predicts protein relative solvent accessibility (RSA) by applying a deep multi-view feature learning framework that integrates multiple neural network architectures and sequence-based features to improve RSA prediction accuracy.

Key Features:

• Deep Multi-View Feature Learning Framework: Employs a DMVFL framework to learn complementary representations from multiple views of sequence information.
• Neural Network Units: Integrates Bidirectional Long Short-Term Memory Recurrent Neural Network (BiLSTM RNN), Squeeze-and-Excitation Networks, and fully-connected hidden layers.
• Sequence-Based Single-View Features: Incorporates Position-Specific Scoring Matrix (PSSM), Position-Specific Frequency Matrix (PSFM), predicted secondary structure, and roughly predicted three-state RSA probability as inputs.
• Multiple Feedback Mechanism: Uses a customized feedback mechanism to extract and enhance discriminative information from single-view features.
• Prediction Modes and Thresholds: Supports two-state (25%), three-state (9% and 36%), and four-state (4%, 25%, 50%) RSA classification thresholds.
• Benchmark Evaluation and Metrics: Evaluated on TEST524 and CASP14-derived (CASP14set) datasets and reports high Pearson correlation coefficient values for real-valued RSA predictions.

Scientific Applications:

• Two-State Prediction: Classifies residues as exposed or buried using an exposure threshold of 25%.
• Three-State Prediction: Classifies residues into three exposure states using exposure thresholds of 9% and 36%.
• Four-State Prediction: Classifies residues into four exposure states using exposure thresholds of 4%, 25%, and 50%.
• Real-Valued RSA Prediction and Correlation Analysis: Produces real-valued RSA predictions assessed by Pearson correlation coefficient to quantify agreement with native RSA.
• Benchmarking and Method Comparison: Enables performance comparison against state-of-the-art predictors on TEST524 and CASP14set datasets.

Methodology:

Integrates PSSM, PSFM, predicted secondary structure, and roughly predicted three-state RSA probability as single-view inputs into a DMVFL architecture combining BiLSTM RNN, Squeeze-and-Excitation networks, and fully-connected hidden layers, with a customized feedback mechanism; evaluated on TEST524 and CASP14set using Pearson correlation coefficient for real-valued RSA.