SearchNovo

A database search engine identifies a peptide with high precision, provided that peptide is already in the FASTA file it was pointed at; it is silent on anything novel or mutated. A de novo model has the opposite profile: it proposes a sequence for any spectrum, but tandem mass spectra routinely drop the one b- or y-ion separating two indistinguishable candidates, and those errors are why de novo precision has trailed database search. SearchNovo treats the two failure modes as complementary rather than competing.

For each query MS2 spectrum, a retriever pulls the single most similar peptide-spectrum match (PSM) out of a reference database, and a fusion layer conditions the peptide decoder on that reference sequence while it generates. The authors' analysis of why this helps is the substance of the paper: retrieved spectra frequently contain signal peaks the query is missing, and retrieved peptides share long fragments with the target — LTANDIFRK against a target of LTANDVFRK. The reference acts as evidence about peaks the instrument failed to record, not merely a prior over plausible sequences.

The model comes from Jun Xia and colleagues in Stan Z. Li's group at Westlake University, with collaborators at USC, Hunan University, and the National University of Singapore. It sits alongside the catalog's other routes to the same problem: Casanovo introduced the transformer encoder–decoder this family builds on, DiNovo recovers absent fragment ions by pairing complementary mirror-protease spectra, and PLMNovo imports priors from a protein language model. SearchNovo's extra evidence comes from spectra somebody has already sequenced.

#Key Features

• Mass-constrained spectral retrieval: The retriever narrows the database to PSMs whose peptide mass falls within ±20 Da of the query precursor mass, then ranks that subset by MatchMS spectral similarity and keeps the top hit, avoiding an exhaustive all-pairs scan.
• Similarity-weighted fusion layer: Two parallel attention blocks run over the generated prefix — one ordinary self-attention, one attending to the retrieved reference peptide — and their outputs are summed and layer-normalized. The reference embedding is scaled by the query–reference spectral similarity, so a weak match contributes proportionally less.
• Pseudo reference PSMs: A dissimilar reference is worse than none, so training adds a second loss term in which the reference is an empty [PAD]/[EOP] placeholder with its similarity fixed at 1. One set of weights then decodes with or without usable retrieval.
• Modification-aware vocabulary: Post-translational modifications are first-class tokens in the amino acid vocabulary, and a retrieved reference often carries the target's modification.

#Technical Details

The spectrum encoder embeds each MS2 peak by summing a sinusoidal encoding of its m/z, spanning 0.001 to 10,000, with a linear projection of its intensity; the 150 most intense peaks are kept. Transformer self-attention layers produce the spectrum representation, and a causally masked decoder cross-attends to it, stopping at an end-of-peptide token or 100 residues. The model has 50.8 million trainable parameters and was trained for 30 epochs with Adam on a single A100.

Evaluation uses the three NovoBench datasets: Seven-species (low resolution, ~317,000 training spectra), Nine-species (high resolution with three PTM classes, ~499,000), and HC-PT (synthetic tryptic peptides covering canonical human proteins and HLA peptides, ~213,000). Against DeepNovo, PointNovo, InstaNovo, AdaNovo, and Casanovo, SearchNovo reports the best peptide-level precision on all three at 0.259, 0.550, and 0.447, and the best amino-acid-level precision at 0.489, 0.748, and 0.652. Removing the similarity weighting drops peptide-level precision on Nine-species from 0.550 to 0.472. Inference runs faster than the 66.3M-parameter AdaNovo and the 92.3M-parameter InstaNovo.

#Applications

De novo sequencing is the only option where a search database is incomplete or absent: monoclonal antibody sequencing, HLA neoantigen discovery for cancer vaccines, metaproteomics, and organisms with no reference proteome. SearchNovo suits the common middle case, where a lab has accumulated confidently sequenced PSMs from earlier runs on similar samples but the peptides of interest are not among them. That evidence becomes the retrieval database, and the model still generates sequences outside it. The constraint is real: inference needs a reference PSM database on hand, which pure de novo models do not.

#Impact

SearchNovo carries retrieval-augmented generation into proteomics in a form that fits the field's existing assets — every mass spectrometry lab already holds a pile of sequenced PSMs. The paper was published at ICLR 2025 and the training and inference code is released under the MIT license, built on the NovoBench evaluation framework from the same group; no pretrained checkpoint is distributed, so the model must be trained before use. Two limits are worth stating plainly: the benchmarks are held-out spectra rather than wet-lab validation, and no head-to-head comparison against database search is reported, because the test peptides are absent from the training sets by construction. The claim of closing the gap rests on the de novo comparisons alone.