FoldHSphere

FoldHSphere produces hyperspherical embeddings to improve protein fold recognition by representing fold classes with maximally separated prototype vectors on a hypersphere for structural classification across family and fold levels.

Key Features:

• Hyperspherical Embedding Space: Represents fold classes using prototype vectors that are maximally separated on a hypersphere to cluster protein embeddings around fold prototypes.
• Two-Stage Training Procedure: Establishes prototype vectors in the hyperspherical space and then trains a neural network using an angular large margin cosine loss to cluster protein embeddings around those prototypes.
• Advanced Network Architectures: Implements ResCNN-GRU and ResCNN-BGRU architectures that process protein sequences through residual-convolutional blocks followed by a gated recurrent unit (GRU/BGRU) recurrent layer.
• Performance Enhancement: Evaluated on the LINDAHL dataset, achieving 81.3% fold-level accuracy and narrowing the performance gap between family-level and fold-level predictions.
• Discriminative Embeddings: Learns embeddings that are discriminative and representative of protein folds, enabling fold identification even with low amino acid sequence similarity.

Scientific Applications:

• Structural classification: Enables precise protein fold recognition to support structural classification in structural biology and bioinformatics.
• Functional inference: Aids inference of protein function and interactions by improving fold-level assignment.
• Method benchmarking: Serves as a fold-level benchmark on the LINDAHL dataset for comparing fold recognition methods.

Methodology:

Uses a two-stage computational procedure that first establishes prototype vectors in a hyperspherical embedding space and then trains ResCNN-GRU and ResCNN-BGRU networks using an angular large margin cosine loss; networks process input sequences via residual-convolutional blocks followed by GRU/BGRU layers and evaluation on the LINDAHL dataset yielded 81.3% fold-level accuracy.