NovoCRF puts the mass constraint into the state space: 42.4% mean peptide recall on seven held-out sets, at 55 times five-beam Casanovo's throughput on CPU
Synopsis
De novo peptide sequencing has largely moved to large transformers that emit residues step by step, with precursor mass entering only through decoding rules or filters. NovoCRF makes the states prefix masses, so every path is a mass-feasible peptide, scores each edge from physics-derived ion-match features, and normalizes exactly over all mass-feasible peptides. On seven 2026 DDA held-out sets totaling 126,262 labeled spectra, mean peptide recall is 42.4% and amino-acid recall 64.0%, against 39.1% and 53.1% for Casanovo v5 with five beams, at 1,746 spectra/s on 18 CPU threads.
Interpretation
The search space becomes a prefix-mass lattice, so every path the model outputs is already a peptide consistent with the precursor mass, without beam-search filtering or a mass-control unit at decoding. Earlier neural de novo tools generate residues step by step or position by position, bringing the precursor mass in at decoding through knapsack filtering, a mass-control unit or graph search, and normalize the likelihood locally per step or per output position. On MassIVE-KB development data, replacing the global normalizer with a softmax over each state's outgoing or incoming edges, a maximum-entropy Markov model on the same lattice, cost 8 to 10 points of peptide recall at equal training data.
Each edge is scored from physics-derived ion-match features rather than lifting a single m/z into a high-dimensional vector and asking the network to relearn arithmetic. Casanovo and similar models encode each peak's m/z as a 512-dimensional sinusoidal vector, with the mass constraint entering only through decoding rules. The scorer is a three-hidden-layer MLP of 64 ReLU units, the CRF layer itself has no parameters, the CRF totals 18,460 parameters and the full model 309,581, training in about 17 GPU-hours on one RTX 3090.
Precursor-mass correction, exact search and sequence context are split into three modules with one job each, and modifications never seen in training can be searched at their exact mass. A large end-to-end model can partly compensate for a wrong precursor mass but cannot separate correction from search for inspection. Switching from msconvert MGFs to PrecursorMassFinder output raised the share of labels fitting the precursor mass from 88.5% to 97.9% and lifted NovoCRF's mean recall by 2.24 points without retraining; given C+119.004 it found 64.9% of the 578 cysteinylated labels in the HLA-I development set, and none without it.
Perspective
Results cover seven DDA held-out sets made public in 2026, one raw file each, scored once after training; throughput was measured on 18 threads of an Intel i9-13900K, and the phone result on the colorectal HLA-I test set.
The authors report that the 3.3-point peptide-recall lead is not significant (paired t-test and Wilcoxon) while the 10.9-point amino-acid lead is; Casanovo v5 with five beams has higher peptide recall on the human trypsin and frog Glu-C sets, and its remaining advantage concentrates on labels with four or more bonds lacking b or y ion support. Candidates are limited to 48 residues, and the evidence analysis counts only the listed ion types.
