Learning supervised embeddings for large scale sequence comparisons.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Similarity-based search in large sequence collections is crucial for bioinformatics.
• Existing alignment-based methods like BLAST struggle to scale with emerging datasets.
• Alignment-free methods leveraging sequence lexical structure offer alternatives.

Purpose of the Study:

• Introduce supervised methods (SuperVec, SuperVecX) for learning sequence embeddings.
• Enhance representation learning by incorporating class-related information.
• Improve sequence retrieval and classification performance.

Main Methods:

• Developed SuperVec and SuperVecX, supervised approaches extending Representation Learning (RepL).
• Incorporated class information during training to ensure proximal representations for related sequences.
• Proposed hierarchical tree-based methods (H-SuperVec, H-SuperVecX) for sequence retrieval.

Main Results:

• SuperVec and SuperVecX embeddings demonstrated superior performance in sequence retrieval and classification tasks compared to unsupervised RepL methods.
• The hierarchical methods (H-SuperVec, H-SuperVecX) showed improved retrieval and classification accuracy.
• New methods were an order of magnitude faster than BLAST for database retrieval.

Conclusions:

• Supervised sequence embedding methods (SuperVec, SuperVecX) offer significant advantages over existing approaches.
• Hierarchical methods enhance retrieval efficiency and accuracy in large-scale sequence analysis.
• These methods enable faster, more precise bioinformatics searches, facilitating hybrid search strategies.