NEAR: neural embeddings for amino acid relationships

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Genomics

Background:

• Protein language models (PLMs) show promise but face speed and accuracy limitations.
• Classical sequence alignment methods are established but can be outperformed.
• Efficient homology search is crucial for large-scale protein annotation.

Purpose of the Study:

• Introduce Neural Embeddings for Amino acid Relationships (NEAR) for improved protein homology search.
• Enhance both the speed and accuracy of identifying homologous protein sequences.
• Evaluate NEAR's performance against state-of-the-art methods and its utility as a pre-filter.

Main Methods:

• Developed NEAR, a method using neural representation learning and a ResNet embedding model.
• Trained the model using contrastive learning guided by trusted sequence alignments.
• Employed a pipeline of residue-level k-NN search and neighbor aggregation for identifying candidates.

Main Results:

• NEAR demonstrates substantially improved accuracy compared to state-of-the-art PLMs on benchmark datasets.
• NEAR exhibits lower memory requirements and faster embedding and search speeds.
• NEAR functions as a high-speed pre-filter, outperforming existing methods like HMMER3's pre-filter.

Conclusions:

• NEAR offers a significant advancement in protein homology detection, balancing speed and accuracy.
• The method shows potential for standalone homology detection with enhanced sensitivity.
• NEAR serves as an effective and rapid pre-filter for sensitive annotation pipelines.