Evaluating the impact of topological protein features on the negative examples selection

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Supervised machine learning for automated protein-function prediction (AFP) requires both positive and negative examples.
• Public databases like Gene Ontology often lack explicit negative functional annotations for proteins.
• Identifying informative negative examples (negative selection) is a critical challenge in AFP.

Purpose of the Study:

• To analyze the impact of various protein features on the selection of negative proteins for Gene Ontology (GO) terms.
• To identify which protein features are most relevant for discriminating reliable negative examples in AFP.
• To evaluate the utility of these features for both negative selection and protein function prediction.

Main Methods:

• Network-based analysis of protein relationships (protein-protein and genetic interactions).
• Utilized term-aware and term-unaware protein features, including graph centrality measures and protein multifunctionality.
• Validated feature informativeness through temporal holdout experiments on yeast, mouse, and human proteomes.

Main Results:

• Term-aware features were generally more informative for negative selection.
• Node betweenness emerged as the most relevant term-unaware feature.
• The protein's positive neighborhood was the most predictive feature for the AFP task itself.

Conclusions:

• The proposed features enhance the effectiveness of negative selection algorithms, particularly in a temporal holdout setting.
• Exploiting nonlinear combinations of features further improved the performance of negative selection.
• This work provides insights into crucial protein features for improving AFP accuracy.