TripletProt: Deep Representation Learning of Proteins Based On Siamese Networks.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Pretrained representations are valuable in machine learning but computationally expensive.
• Representation learning for biological entities can simplify supervised learning challenges.
• Existing methods often require substantial computational resources.

Purpose of the Study:

• Introduce TripletProt, a new protein representation learning approach.
• Leverage protein-protein interaction (PPI) networks for efficient representation.
• Reduce computational costs and representation length for bioinformatics tasks.

Main Methods:

• Developed TripletProt using Siamese neural networks.
• Utilized protein-protein interaction (PPI) networks as the primary data source.
• Compared TripletProt against UniRep and DeepGO using functional annotation tasks.

Main Results:

• TripletProt demonstrated improved F1 scores in sub-cellular localization and gene ontology prediction.
• The method achieved high performance solely relying on PPI network data.
• Generated representations are computationally less expensive and shorter than comparable methods.

Conclusions:

• TripletProt is an effective and efficient method for protein representation learning.
• The approach shows significant potential for various protein informatics tasks.
• TripletProt offers a computationally advantageous alternative for biological data analysis.