GODoc: high-throughput protein function prediction using novel k-nearest-neighbor and voting algorithms.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Biological data generation, particularly from next-generation sequencing, has rapidly increased.
• Experimental annotation of protein function is time-intensive, necessitating computational approaches for hypothesis generation and experimental prioritization.
• Gene Ontology (GO) provides a standardized hierarchical framework for describing protein functions.

Purpose of the Study:

• To develop a general computational framework, GODoc, for predicting protein function using sequence information.
• To address the challenge of multiple GO term prediction within the Gene Ontology.
• To improve the efficiency and accuracy of protein function annotation.

Main Methods:

• Developed GODoc, a framework integrating feature engineering, feature reduction, and a novel k-nearest neighbor algorithm.
• The k-nearest neighbor algorithm was adapted with a training procedure to handle multiple GO term predictions.
• The framework utilizes protein sequence information as input for function prediction.

Main Results:

• GODoc demonstrated superior performance compared to baseline models in the CAFA2 evaluation.
• In the CAFA3 competition, GODoc achieved 10th rank in Cellular Component Ontology prediction.
• The method showed strong performance in species-specific and term-centric tasks, ranking highly for eukaryotic and prokaryotic predictions, and specific functions like biofilm formation and long-term memory.

Conclusions:

• A novel and effective strategy was developed to integrate a training procedure into the k-nearest neighbor algorithm for instance-based learning.
• The developed method successfully addresses the Gene Ontology multiple-label prediction problem, even with thousands of GO terms.
• GODoc offers an effective computational solution for predicting protein function, aiding biological research.