Evaluating a variety of text-mined features for automatic protein function prediction with GOstruct

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Current protein function prediction methods often overlook valuable data within the biomedical literature.
• Integrating textual information can enhance the accuracy and interpretability of computational models.

Purpose of the Study:

• To evaluate the effectiveness of literature-derived features for predicting human protein function using the GOstruct model.
• To assess the impact of ontology term co-mention and bag-of-words features on prediction accuracy.

Main Methods:

• Utilized a structured output support vector machine model (GOstruct).
• Extracted features from biomedical literature, including ontology term co-mentions and bag-of-words.
• Developed a medium-throughput pipeline for annotating co-mention data.

Main Results:

• Literature-based features significantly improved protein function prediction accuracy (F-max: MF=0.408, BP=0.461, CC=0.608).
• Manual review indicated that some false positives could be biologically valid, supported by literature evidence.
• The annotation pipeline efficiently processed a large dataset of co-mentions.

Conclusions:

• Biomedical literature is a valuable resource for enhancing computational protein function prediction.
• Literature features provide a basis for verifying automated predictions and uncovering novel biological insights.
• The developed pipeline can accelerate protein curation by leveraging literature data.