ASFold-DNN

ASFold-DNN enhances protein fold recognition by extracting evolutionary features and applying a fully connected neural network to improve structure-based inference for protein function, drug development, and gene therapy.

Key Features:

• Evolutionary Feature Extraction: Extracts four distinct groups of evolutionary features from protein primary structures and applies a preliminary selection to optimize parameters for ACC_HMM and SXG_HMM.
• Feature Selection Optimization: Uses multiple feature selection algorithms and adjusts internal threshold values to identify the most relevant feature scheme for model training.
• Neural Network Architecture: Implements a Full Connected Neural Network (FCNN) with fully optimized hyper-parameters for protein fold prediction.

Scientific Applications:

• Protein Function Prediction: Enables inference of protein function from predicted folds derived from evolutionary feature representations.
• Drug Development: Supports structure-based identification of potential drug targets through improved fold recognition.
• Gene Therapy: Provides structural insights relevant to the design and interpretation of gene therapy strategies.
• Benchmarking and Generalization: Validated on additional datasets including ASTRAL186 and LE and reported to outperform existing state-of-the-art methods on those benchmarks.

Methodology:

Performs evolutionary feature extraction followed by rigorous feature selection; constructs and optimizes a Full Connected Neural Network (FCNN) with tuned hyper-parameters; evaluates performance on low sequence similarity datasets DD, EDD, and TG (reported accuracies 85.28%, 95.00%, and 88.84%, respectively) and on ASTRAL186 and LE for generalization assessment.