NEAR: Neural Embeddings for Amino acid Relationships.

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Bioinformatics

Background:

• Protein language models (PLMs) show promise in replacing traditional sequence alignment methods for protein database searches.
• Current PLMs are often slower and produce more false positives than alignment-based tools.
• Efficient and accurate homology detection is crucial for understanding protein function and evolution.

Purpose of the Study:

• To introduce NEAR, a neural network-based method designed to enhance the speed and accuracy of searching for homologous proteins in large databases.
• To evaluate NEAR's performance against state-of-the-art PLMs and existing pre-filtering methods.

Main Methods:

• NEAR utilizes a ResNet embedding model trained with contrastive learning guided by trusted sequence alignments.
• It computes per-residue embeddings for protein sequences.
• Homolog candidates are identified using a pipeline involving residue-level k-nearest neighbors (k-NN) search and neighbor aggregation.

Main Results:

• NEAR demonstrates substantially improved accuracy compared to state-of-the-art PLMs on a benchmark of remote homologs and decoys.
• The method exhibits lower memory requirements and faster embedding and search speeds.
• As a pre-filter for profile hidden Markov model (pHMM) searches, NEAR is at least 5x faster than HMMER3's pre-filter and outperforms the pre-filter in the nail tool.

Conclusions:

• NEAR offers a significant advancement in protein homology search, improving both speed and accuracy.
• Its effectiveness as a high-speed pre-filter enhances sensitive annotation pipelines, outperforming current methods.
• NEAR presents a valuable alternative for standalone homology detection with potentially higher sensitivity than standard alignment methods.