PLMNovo

De novo peptide sequencing infers the amino acid sequence of a peptide directly from its tandem mass spectrometry (MS/MS) spectrum, without matching against a reference database. It is essential when the peptide of interest is absent from existing databases, such as in immunopeptidomics, antibody characterization, and metaproteomics. Deep learning approaches, most prominently the transformer-based Casanovo, have become the standard for this task, but they learn the mapping from spectra to sequence purely from paired spectrum-peptide data.

PLMNovo, developed by researchers at Duke University and released as a preprint in October 2025, takes a different route by injecting prior biological knowledge from pre-trained protein language models (PLMs). Rather than treating sequence prediction as an isolated decoding problem, it casts peptide-spectrum matching as a constrained optimization that forces the representation of a spectrum to align with the PLM representation of its corresponding peptide. This grounds the spectrum encoder in the sequence statistics a PLM has already learned from millions of natural proteins.

#Key Features

• PLM-guided alignment: A spectrum encoder and a protein language model are trained so that spectrum embeddings and peptide embeddings occupy a shared, aligned space, transferring learned sequence priors into the sequencing model.
• Constrained learning formulation: Peptide-spectrum matching is posed as a constrained optimization problem with an explicit alignment constraint, rather than a single unconstrained loss.
• Primal-dual training: A Lagrangian primal-dual algorithm jointly trains the spectrum encoder and peptide decoder, with the option to fine-tune the pre-trained PLM during optimization.
• Benchmark gains: On established de novo sequencing benchmarks, PLMNovo outperforms several state-of-the-art deep-learning baselines.

#Technical Details

PLMNovo couples a spectrum encoder, which ingests fragment peaks and precursor information, with a peptide decoder that emits the amino acid sequence. A pre-trained protein language model provides embeddings of candidate peptides, and the training objective enforces alignment between the spectrum encoder's output and these PLM embeddings. The constrained learning problem is solved with a Lagrangian primal-dual method that adaptively balances the sequencing loss against the alignment constraint, and the PLM can optionally be fine-tuned rather than kept frozen. Evaluated on standard tandem mass spectrometry benchmarks, the framework reports improved sequencing accuracy relative to prior deep-learning de novo methods.

#Applications

PLMNovo targets proteomics workflows where reference-database search is insufficient, including identification of novel or mutated peptides, antibody and T-cell-receptor sequencing, immunopeptidomics of MHC-presented antigens, and metaproteomics of uncharacterized organisms. By anchoring spectrum interpretation to protein language model priors, it is aimed at mass spectrometry practitioners seeking higher-confidence calls on peptides that lie outside curated databases.

#Impact

PLMNovo illustrates a broader trend of coupling task-specific encoders to frozen or fine-tuned foundation models via representation alignment, here bringing protein language model knowledge into mass-spectrometry-based sequencing. As a preprint awaiting peer review, its results are reported on in-silico benchmarks, and no public code or model weights have been released to date, which currently limits independent reproduction and deployment.