DescFold

DescFold predicts protein fold types from amino acid sequences by integrating multiple sequence- and structure-based descriptors with machine learning for template-based fold recognition.

Key Features:

• Machine Learning Integration: Uses Support Vector Machines (SVMs) to integrate multiple sequence- and structure-based descriptors for fold recognition.
• Profile-Sequence-Alignment-Based Descriptor: Utilizes Psi-blast e-values and bit scores to capture sequence alignment information.
• Sequence-Profile-Alignment-Based Descriptor: Employs Rps-blast e-values and bit scores to enhance sensitivity to sequence variations.
• Secondary Structure Element Alignment (SSEA): Aligns secondary structure elements to capture structural motif information.
• PROSITE Functional Motifs: Incorporates PROSITE motif occurrences as predictive features.
• Profile-Profile Alignment Score (PPA): Generates profile-profile alignment scores using the COMPASS algorithm treating one profile as a template with known 3D structure.
• Structure-Seeded Profile: Creates structure-seeded profiles via TM-align structural alignment and refines template profiles by searching structural neighbors.
• Profile-Structural-Profile-Alignment-Based Descriptor (PSPA): Uses COMPASS to align profiles and structural information for improved discrimination.
• Training and Testing: Trained on 1,835 diverse proteins from SCOP 1.73 and tested on 1,866 proteins from SCOP 1.75.
• Performance Metrics: At a false positive rate below 5%, recognizes structural homologs at the fold level for approximately 46% of test proteins and was benchmarked against LiveBench targets and the Lindahl dataset.

Scientific Applications:

• Structural annotation: Predicts protein folds from sequences to support structural annotation of proteins.
• Functional inference: Provides structural predictions that aid inference of protein function and interactions.
• Evolutionary analysis: Detects fold-level homology to assess evolutionary relationships.
• Drug discovery and enzyme design: Supplies structure-based insights without experimental structures to support drug discovery and enzyme design efforts.
• Genome annotation: Assists annotation of newly sequenced proteins by assigning probable folds.

Methodology:

SVMs are trained to integrate descriptors computed from Psi-blast, Rps-blast, SSEA, PROSITE motif occurrences, COMPASS-derived PPA and PSPA scores, and TM-align–derived structure-seeded profiles (including searches of structural neighbors); training used SCOP 1.73 (1,835 proteins) and testing used SCOP 1.75 (1,866 proteins) with evaluation including false positive rate thresholds and benchmarks such as LiveBench and the Lindahl dataset.