Continuous Distributed Representation of Biological Sequences for Deep Proteomics and Genomics

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics
• Proteomics

Background:

• Biological sequences like proteins and genes require efficient representation for deep learning applications.
• Existing methods for analyzing protein sequences can be computationally intensive and may not capture complex relationships.

Purpose of the Study:

• To introduce a novel representation and feature extraction method for biological sequences, termed bio-vectors (BioVec).
• To specifically develop and evaluate protein-vectors (ProtVec) for applications in proteomics, including protein family classification and structure prediction.

Main Methods:

• Utilized artificial neural network approaches to represent protein sequences as dense n-dimensional vectors (ProtVec).
• Applied ProtVec to classify 324,018 protein sequences from Swiss-Prot into 7,027 families.
• Employed support vector machine classifiers with ProtVec to distinguish disordered proteins from structured proteins using DisProt and FG-Nup databases.

Main Results:

• Achieved an average protein family classification accuracy of 93%±0.06%, outperforming existing methods.
• Demonstrated high accuracy in predicting disordered proteins: 99.8% for FG-Nups and 100.0% for DisProt sequences.
• Showcased that sequence data alone, when processed by ProtVec, can yield accurate insights into protein structure.

Conclusions:

• ProtVec offers a powerful and versatile representation for biological sequences, particularly proteins.
• This method serves as effective pre-training for various deep learning tasks in bioinformatics.
• ProtVec enables accurate classification and structure prediction with a single training instance, streamlining bioinformatics investigations.