Sequence Representations and Their Utility for Predicting Protein-Protein Interactions.

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Protein-protein interactions (PPIs) are fundamental to cellular function.
• Machine learning methods are widely used for PPI prediction, but rely heavily on feature engineering.
• Existing methods often use physico-chemical properties derived from protein sequences, with few directly utilizing sequence information.

Purpose of the Study:

• To explore the utility of sequence embeddings for predicting protein-protein interactions.
• To introduce and evaluate a novel feature construction method, SuperVecNW, for generating sequence embeddings.
• To compare the performance of sequence embeddings against traditional feature representations.

Main Methods:

• Constructed protein pair feature vectors by concatenating sequence embeddings.
• Utilized established Word2Vec-based methods (Seq2Vec, BioVec) and a novel method (SuperVecNW) for learning sequence embeddings.
• Employed a binary classifier for PPI prediction using the generated feature vectors.
• Tested the approach on human and yeast PPI datasets and specific biological networks (CD9, Ras-Raf-Mek-Erk-Elk-Srf, Wnt-related).

Main Results:

• Low-dimensional sequence embeddings achieved superior prediction results compared to most physico-chemical property-based representations.
• The proposed SuperVecNW method incorporates network information alongside contextual sequence information.
• The approach demonstrated effectiveness across diverse PPI datasets and networks.

Conclusions:

• Sequence embeddings provide a powerful and simpler alternative for feature vector construction in PPI prediction.
• SuperVecNW offers a promising method for generating informative sequence embeddings by integrating network context.
• This study highlights the potential of sequence-based features for advancing PPI prediction methodologies.