Learning protein sequence embeddings using information from structure.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein structure prediction from amino acid sequences is crucial for understanding protein function.
• Current sequence-based methods struggle with detecting structural similarity in distantly related proteins.
• This limitation hinders knowledge transfer between structurally similar proteins.

Purpose of the Study:

• To develop a novel representation learning framework for inferring protein structural properties from sequences.
• To improve the prediction of structural similarity between proteins, even with significant sequence divergence.
• To create transferable protein sequence embeddings that encode structural information.

Main Methods:

• Utilized bidirectional long short-term memory (LSTM) models to generate vector embeddings for protein sequences.
• Implemented a two-part feedback mechanism incorporating global structural similarity and pairwise residue contact maps.
• Introduced a novel soft symmetric alignment (SSA) measure for comparing embedding sequences of arbitrary lengths.

Main Results:

• The framework successfully learned position-specific embeddings encoding structural information without direct position correspondence.
• Empirical results show the multi-task framework outperforms existing sequence-based and structure-based alignment methods in predicting structural similarity.
• Learned embeddings improved state-of-the-art performance in transmembrane domain prediction, demonstrating transferability.

Conclusions:

• Representation learning offers a powerful approach to protein structure inference from sequence.
• The proposed framework and similarity measure advance the field of sequence-based protein structure analysis.
• The learned embeddings have broad applicability, enhancing performance on related protein sequence tasks.