GODoc

GODoc predicts protein functions from sequence data by assigning Gene Ontology (GO) terms using feature engineering, dimensionality reduction, and a k-nearest-neighbor algorithm adapted for multi-label prediction.

Key Features:

• Sequence-based GO prediction: Predicts Gene Ontology terms (biological process, cellular component, molecular function) directly from protein sequence information.
• Feature engineering and reduction: Integrates feature engineering and dimensionality reduction techniques to represent sequence-derived features for classification.
• k-nearest-neighbor for multi-label prediction: Uses a k-nearest-neighbor (k-NN) algorithm tailored to handle multiple simultaneous GO term annotations.
• Adapted training procedure: Incorporates an innovative training procedure into the instance-based k-NN algorithm to improve multi-label prediction performance.
• GO hierarchy handling: Designed to address the hierarchical structure of Gene Ontology across biological processes, cellular components, and molecular functions.
• Benchmark validation: Validated on CAFA benchmark datasets, with reported performance in CAFA2 and specific rankings in CAFA3 (10th overall Cellular Component, 10th eukaryotic Cellular Component, 8th prokaryotic Molecular Function) and term-centric placements (third for biofilm formation in Pseudomonas aeruginosa; tied first for long-term memory in Drosophila melanogaster).

Scientific Applications:

• Protein function annotation: Assigns GO terms to uncharacterized proteins to support functional annotation efforts.
• Cross-ontology prediction: Predicts annotations across GO ontologies (biological process, cellular component, molecular function) for eukaryotic and prokaryotic proteins.
• Hypothesis generation and prioritization: Provides predictions to help formulate experimental hypotheses and prioritize targets for wet-lab validation.
• Term-centric prediction: Supports prediction of specific phenotypes or processes such as biofilm formation in Pseudomonas aeruginosa and long-term memory in Drosophila melanogaster.

Methodology:

Integrates feature engineering and dimensionality reduction with a k-nearest-neighbor algorithm adapted for multi-label (multiple GO term) prediction, incorporates an adapted training procedure within the instance-based k-NN framework, and evaluates performance on CAFA2 and CAFA3 benchmark datasets.