NEAR

Searching a protein sequence against a large database is one of the oldest and most-run computations in biology, and its cost is dominated by a filtering step: before any expensive alignment happens, a fast pre-filter must discard the overwhelming majority of candidate targets while keeping essentially every true homolog. In HMMER3 that pre-filter is the MSV gapless alignment stage, and it consumes roughly 70% of the tool's runtime. Protein language models have been proposed as a replacement, embedding sequences so that homology becomes a nearest-neighbor lookup, but in practice they have been slower than the alignment tools they were meant to replace and prone to matching unrelated sequences.

NEAR — Neural Embeddings for Amino acid Relationships — is a direct response to that gap, from the Travis Wheeler lab at the University of Arizona with the first author at the University of Montana. Instead of adapting a general-purpose protein language model, NEAR trains a small residual convolutional network from scratch for one job: producing per-residue embeddings in which residues that a trusted alignment would pair sit close together, and everything else does not.

The result is deliberately narrow. NEAR does not predict structure, function, or fitness, and produces no sequence-level representation intended for general transfer. What it offers is a fixed checkpoint that embeds new query and target sequences and ranks candidate matches, at a fraction of the parameter count, memory footprint, and runtime of the language models it is benchmarked against. It was posted as a preprint in January 2024 and published in Bioinformatics as part of the ISMB/ECCB 2025 proceedings.

#Key Features

• Per-residue rather than per-sequence embeddings: Matching happens at the residue level and is then aggregated across neighbors, preserving the local correspondence that whole-sequence mean-pooled embeddings discard.
• Alignment-guided contrastive training: An N-pair contrastive loss is supervised by HMMER3-generated alignments, so the embedding geometry is trained against exactly the relationships downstream search cares about.
• Small enough to be fast: At roughly 7 million parameters, NEAR uses under 2% of the parameters of the protein language models it is compared with, and its 256-dimensional embeddings cut index memory accordingly.
• Drop-in pre-filter role: Designed to occupy the same slot as HMMER3's MSV stage or MMseqs2's k-mer pre-filter, feeding a smaller candidate set into conventional alignment rather than replacing it.
• Checked-in pretrained checkpoint: The repository ships an 88 MB weight file (models/resnet_877_256.pt) loaded directly by the embedding scripts, so the model runs on new sequences with no retraining.

#Technical Details

NEAR's encoder is a residual convolutional network with 8 residual blocks, kernel size 7, ELU activations, and no weight sharing between convolutional layers, emitting a 256-dimensional vector per residue. Training pairs were assembled from Uniclust30 clusters of UniRef sequences between 256 and 1,024 residues, whose members are less than 30% identical across clusters; sampled pairs within each cluster were aligned with HMMER to supply the supervision signal, yielding 6,190,084 sequence pairs and alignments. Search is performed by indexing target residue embeddings with FAISS, running residue-level k-nearest- neighbor lookup, and aggregating matching residue pairs into a per-target score. On a benchmark of 10,000 queries against 40,000 targets, augmented with shuffled decoys and with low-similarity real sequences as an alternative decoy set, NEAR achieved higher recall at high filtration than MMseqs2, HMMER3's MSV filter, TM-Vec, and residue-level embeddings from ProtTrans and ESM Cambrian. It is at least 5x faster than HMMER3's MSV pre-filter, and its embedding step is dramatically cheaper than the language model baselines, which carry roughly 600 million to 3 billion parameters.

#Applications

NEAR is aimed at the people who build and run large-scale protein annotation pipelines: anyone searching millions of predicted proteins from metagenomes or newly sequenced genomes against Pfam-style reference libraries, where filtration cost sets the wall-clock budget. Slotted in ahead of a profile-HMM aligner, it reduces the candidate set faster and more accurately than the existing filter, which translates directly into throughput. The authors note that its sensitivity on remote homologs suggests standalone use for homology detection, though the paper's evaluation is framed around the pre-filter role.

#Impact

NEAR makes a pointed argument against the assumption that better sequence search requires a bigger model. A 7-million-parameter network trained specifically on alignment supervision beats billion-parameter general-purpose protein language models at ranking homologs, while being faster and lighter than both those models and the hand-engineered filter in the most widely used profile-HMM search tool. That result matters for the broader question of when task-specific representation learning beats foundation-model transfer. Adoption is currently limited: the code is BSD-3-Clause licensed and the checkpoint is public, but the repository is small and its maintainers describe it as a working prototype rather than a polished tool, so users should expect an unstable API and reproduction scripts spread across branches.