AdaNovo

Oxidized methionine and deamidated asparagine appear in a corpus of peptide-spectrum matches far less often than the twenty canonical amino acids, and a decoder trained with a flat cross-entropy loss has little reason to learn them: predicting the common residues well already minimizes the objective. The cost is not local. A peptide prediction counts as correct only when every one of its residues matches, so one missed modification discards the whole sequence.

AdaNovo answers this by deciding, residue by residue, how much training signal each amino acid deserves. It estimates the conditional mutual information (CMI) between the observed mass spectrum and each target amino acid, and re-weights that residue's loss in proportion. The reasoning is chemical rather than statistical: a modification shifts fragment ion masses and so is manifested in the spectrum and nowhere else, making a residue whose identity leans on the spectrum rather than on the residues decoded before it exactly the one a modification-aware model must attend to. Averaging those per-residue values across a peptide yields the mutual information between spectrum and whole peptide, used to down-weight training pairs that agree poorly because of noise or missing peaks.

The model comes from Stan Z. Li's group at Westlake University, with collaborators at UC San Diego and the University of Science and Technology of China. Its encoder-decoder design follows Casanovo, its principal baseline, and it precedes SearchNovo from the same lab.

#Key Features

• Amino acid-level adaptive training: Per-residue CMI values are normalized within each peptide and scaled into a training weight, so residues the spectrum determines — modified ones above all — contribute more gradient than residues the decoder could have guessed from context.
• PSM-level robustness to unreliable labels: The peptide-averaged mutual information becomes a second weight, normalized across each mini-batch, shrinking the contribution of spectrum-peptide pairs that agree badly — the signature of a mislabeled or noise-corrupted match.
• A spectrum-blind second decoder: Peptide Decoder #2 sees only the previously decoded residues, never the spectrum or precursor, and its output supplies the denominator that turns CMI into a computable log ratio against Decoder #1's spectrum-conditioned prediction. Inference drops it, so it costs training time but not sequencing speed.
• Modifications in the vocabulary: The alphabet holds the twenty canonical amino acids, oxidized methionine, deamidated asparagine and glutamine, and a stop token — 24 tokens, so modified residues are decoded directly rather than inferred from a mass offset.

#Technical Details

The mass spectrum encoder and both peptide decoders are 9-layer transformers with 512 feed-forward dimensions. Each peak is a token, its m/z projected by a fixed sinusoidal embedding and its intensity by a linear layer, with no positional encoding, since peak order is meaningless. Decoding is autoregressive, halting at the stop token or 100 residues, with a precursor m/z filter discarding peptides whose total mass contradicts the measurement. Training used one NVIDIA A100 at batch size 32 with AdamW. The model holds 66.31M parameters against Casanovo's 47.35M, a 40% increase attributable to the second decoder.

Evaluation is confined to the nine-species benchmark introduced by DeepNovo: roughly 1.5 million spectra measured on one instrument, labeled by database search at 1% false discovery rate, under leave-one-out cross-validation that trains on eight species and tests on the ninth, so test peptides are essentially unseen. Averaged over the nine folds, AdaNovo reaches 0.499 peptide-level precision against 0.459 for a reproduced Casanovo, 0.439 for PointNovo and 0.376 for DeepNovo, winning 8 of 9 folds. Gains concentrate where the design predicts: identification of modified amino acids improves by 3.6% to 9.4%, and on human spectra removing the amino acid-level objective drops modified-residue precision from 0.483 to 0.314. On clam bacteria spectra corrupted with synthetic noise, removing the PSM-level objective drops peptide precision from 0.397 to 0.336.

#Applications

De novo sequencing is the only option where no search database exists or where the database is the thing being questioned: antibody repertoire sequencing, human leukocyte antigen neoantigen discovery, metaproteomics of environmental and microbiome samples, and proteins absent from reference proteomes. AdaNovo's emphasis on modified residues extends this to workflows where the modification is itself the readout, since PTMs govern protein function and are often the object of study rather than a nuisance to suppress.

#Impact

AdaNovo's contribution is portable in a way benchmark rank is not: the CMI weighting is a training objective over an existing spectrum-to-peptide architecture, so the argument applies to any autoregressive sequencer whose training matches skew toward canonical residues. Two caveats bound the result: every number reported comes from the nine-species benchmark and no other data, and previously unseen modifications, those outside the 24-token alphabet, remain out of reach. The released code is a single training and inference script with one checkpoint trained on that benchmark, and the repository carries no license file.