NovoCRF: Globally Normalized De Novo Peptide Sequencing on a Prefix-Mass Lattice

Rui Qiao; Lei Xin; Baozhen Shan

cs.LGNov 3, 2026

NovoCRF: Globally Normalized De Novo Peptide Sequencing on a Prefix-Mass Lattice

Authors: Rui Qiao, Lei Xin, Baozhen Shan

Organizations: Bioinformatics Solutions Inc.

Abstract

De novo peptide sequencing has moved from dynamic programming over hand-designed or generative scores to large transformers that learn the task end to end. We present NovoCRF, a conditional random field (CRF) whose states are prefix masses, so that its paths are exactly the peptides within +/-0.06 Da of the precursor mass. A small network scores each edge from physics-derived ion-match features and is trained with its scores globally normalized over all mass-feasible peptides. To our knowledge, NovoCRF is the first de novo method trained with exact normalization over all mass-feasible peptides. Our software has three modules, each with exactly one job: a precursor mass finder recovers the accurate precursor mass from MS1, the CRF finds the peptides that fit the precursor mass by exact k-best decoding, and two BiGRU rerankers add sequence context and internal-fragment evidence. The model has 310k parameters and trains in about 17 GPU-hours. On seven held-out data-dependent acquisition datasets published in 2026 (126,262 labeled spectra), NovoCRF's mean peptide recall is 42.4%, against 39.1% for Casanovo v5 with five beams, and its amino-acid recall 64.0% against 53.1%, at 1,746 spectra/s on 18 CPU threads, 55x Casanovo's throughput on a GPU. Split by fragment evidence, NovoCRF leads on labels with up to three peptide bonds without a b or y ion, and Casanovo only beyond. NovoCRF can also search modifications it was never trained on at their exact mass: it finds 32.1% of the hydroxyproline peptides of a lion bone sample.

Figures & tables

Panel of Figure 1 — Figure 1: Top: the prefix-mass lattice. The true peptide (SLAVEK, orange) is one path from bin 0 to the terminal window, normalized against all mass-feasible paths (gray); bars show the evidence at active bins. Bottom: the pipeline, whose output is a peptide with a calibrated confidence.

Component	Shape	Parameters
Scorer MLP $f_{\theta}$	$86{\to}64{\to}64{\to}64{\to}8$	14,408
Residue prior $h_{\phi}$ ; biases $\gamma_{k}$	$31{\to}64{\to}24$ ; 24	3,632
Edge embeddings $\bm{e}_{k},\bm{g}_{k}$	$2\times 24\times 8$	384
Start edges $\bm{e}^{\mathrm{st}}_{j},\gamma^{\mathrm{st}}_{j}$ ; $\bm{v}_{0}$ ; $\kappa$	$3\times 9$ ; 8; 1	36
CRF total		18,460
First-stage reranker; calibrator	BiGRU-64; logistic	70,639

Table 1: NovoCRF parameters, counted from the model’s checkpoints.

Dataset	Sample	Enzyme; instrument	MS2	Labels
PXD076585 [ 29 ]	human A549	trypsin; Exploris	131,789	31,275
PXD071561 [ 30 ]	squid optic lobe	Lys-C; Exploris	59,389	24,701
PXD073929 [ 31 ]	frog toe	Glu-C; Exploris	103,345	2,276
PXD072115 [ 32 ]	chickpea peptides	none; timsTOF Pro	32,981 a	18,992 b
PXD074750 [ 33 ]	Candida auris	prot. K; Exploris	62,389	22,488
PXD071022 [ 34 ]	HLA-I, colorectal	none; Ascend	85,732	21,887

Table 2: The seven held-out test datasets (one raw file each). Labels: PEAKS DB search, 1% PSM FDR. Exploris, Ascend, Fusion: Orbitrap Exploris 480, Ascend and Fusion; prot. K: proteinase K. a PASEF precursors. b Without 275 labels on precursors of unknown charge.

Engine		human, trypsin	squid, Lys-C	frog, Glu-C	chickpea	C. auris , prot. K	HLA-I	HLA-I	Mean
		PXD076585	PXD071561	PXD073929	PXD072115	PXD074750	PXD071022	PXD082505
NovoCRF	Pep	34.2	69.2	36.7	15.6	25.3	61.1	54.5	42.4
NovoCRF	AA	58.8	78.6	61.0	44.7	49.6	79.9	75.5	64.0
CasanovoV5 beam 5	Pep	35.0	65.4	40.9	8.5	24.1	50.7	49.4	39.1
CasanovoV5 beam 5	AA	50.9	68.9	59.1	25.1	37.9	65.4	64.3	53.1
CasanovoV5 beam 1	Pep	30.9	60.5	37.3	6.8	21.8	44.9	44.9	35.3

Table 3: Top-1 peptide recall (Pep) and amino-acid recall (AA) in % on the seven held-out datasets, and their means. Best per column in bold.

Engine	recall	prec.	recall	prec.	AUC	/s
	Peptide		Amino acid		PR	Spectra
NovoCRF (CPU)	42.4	42.4	64.0	65.0	0.381	1,746
CasanovoV5 beam 5 (GPU)	39.1	40.9	53.1	56.2	0.366	32
CasanovoV5 beam 1 (GPU)	35.3	36.9	46.6	49.3	0.337	251
PEAKS de novo (CPU)	22.0	22.1	48.5	48.2	0.174	145

Table 4: Means over the seven datasets (%; PR AUC as a fraction) and inference throughput. Casanovo and PEAKS de novo speeds are from runs on msconvert MGFs (683,418 spectra) on a quiet machine; NovoCRF’s from 683,821 PrecursorMassFinder spectra. Best in bold.

			CasanovoV5
MGF	Labels fit	NovoCRF	beam 5	beam 1
msconvert	88.5	40.1	38.9	35.2
PrecursorMassFinder	97.9	42.4	39.1	35.3

Table 5: Precursor input. Share of labels whose mass is within the precursor tolerance of the MGF precursor, and mean peptide recall (%) with msconvert or PrecursorMassFinder MGFs.

	b/y rule					extended rule
Missing bonds	Labels	NovoCRF	CasanovoV5 beam 5	CasanovoV5 beam 1	PEAKS de novo	Labels	NovoCRF	CasanovoV5 beam 5
0	21,870	89.9	80.6	75.3	70.3	30,984	80.4	71.3
1	24,183	73.1	64.8	59.0	35.8	27,917	61.8	54.8
2	20,724	50.7	45.1	39.6	16.9	21,643	37.8	35.3
3	14,740	25.1	23.4	19.6	4.1	13,207	15.6	16.1
4–5	18,495	8.0	9.9	7.5	0.7	14,624	4.9	6.6

Table 6: Peptide recall (%) by the number of the label’s bonds without an observed fragment ion, under the b/y rule and the extended rule, pooled over all 126,262 labels. Best per rule in bold.

Figure 2: HLA-B*15:02 ligand (PXD082505, scan 82428, 2+). Top: fragment ions of the label FLWAGGRASY, which is also CasanovoV5 beam 5’s answer; bottom: ions of NovoCRF’s answer FLWAGGSRAY. Dark bars: b/y ions; light bars: a ions and neutral losses (“

|

”: bond with any matched ion; “

\cdot

”: none); only the stronger b/y peaks are labeled. Both answers explain the same seven bonds and fit the precursor at

-2.7

ppm; they differ only in the order of R, A and S inside the unsupported stretch R

\cdot

\cdot

Figure 3: The reverse case: a squid peptide (PXD071561, scan 24916, 2+). Top: fragment ions of the label RYAMVVEDGVVK (oxidized M, shown as M(+16)), which is also NovoCRF’s answer; bottom: ions of CasanovoV5 beam 5’s answer RFSMVVEDGVVK, reported with confidence 0.83. Marks as in Fig. 2 . YA and FS have the same mass, so the two answers share every bond but the second: there the label has a 2 and y

{}_{10}-

H 2 O (labeled), CasanovoV5 beam 5’s answer no ion between F and S.

	Hyp labels		Labels without Hyp
Peptide length	Labels	Recall	Labels	Recall
$\leq$ 12	478	71.3	492	71.3
13–16	607	47.4	261	44.8
17–20	682	27.3	88	27.3
$>$ 20	999	7.3	67	14.9
All, P+15.995 given	2,766	32.1	908	55.3

Table 7: Lion bone (PXD065297, elastase): peptide recall (%) on the authors’ MaxQuant labels, with residues matched within 0.01

\,\mathrm{Da}

, so that a hydroxyproline (Hyp) site must be called as P+15.995 rather than L. The rows by length give Hyp as a user modification; the last two rows cover all labels, with and without it.

Comparison	Training spectra	Recall
Global CRF	0.86M (MassIVE-KB)	48.9
Local (MEMM), N $\to$ C		38.9
Local (MEMM), C $\to$ N		40.9
With source term	5.2M (MassIVE-KB)	52.6
Without source term	5.2M (MassIVE-KB)	49.1

Table 8: Development ablations on MassIVE-KB: peptide recall (%) of the CRF’s top-1 on 2k validation spectra, earlier feature set. Each group differs only in the listed component. Best in bold.

Mass Spectrometry

Mass spectrometry measures ion mass-to-charge patterns to identify and quantify molecules, with AI linking spectra to structures, peptides, and biological signals.

See all 38 papers

MemNovo: Look Back at the Spectrum for Balanced De Novo Peptide Sequencing from Mass Spectrometry
Jun 10, 2026Dongxin Lyu, Jingbo Zhou, Hongxin Xiang +2Memory-Augmented Neural Networks Transformer Inference
Westlake University Hangzhou, Zhejiang, China · Hunan University Changsha, Hunan, China · Shanghai Artificial Intelligence Laboratory Shanghai, China +1
Regressor-guided Diffusion Model for De Novo Peptide Sequencing with Explicit Mass Control
Feb 23, 2026Shaorong Chen, Jingbo Zhou, Jun XiaMass Spectrometry Diffusion Model Guidance
Zhejiang University, Hangzhou, China, 310058 · AI Lab, Westlake University, Hangzhou, China, 310030 · AIMS Lab, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China, 511453 +1
GyroNovo: Error-Guided Fragment Imputation with Mass-Aware Attention for De Novo Peptide Sequencing
Sep 24, 2026Abdellah El Mekki, Laks V. S. Lakshmanan, Muhammad Abdul-MageedMass Spectrometry Attention Mechanisms
The University of British Columbia · Canada Research Chair in NLP and ML

Jun 10, 2026J/K move · Enter open · S save

MemNovo: Look Back at the Spectrum for Balanced De Novo Peptide Sequencing from Mass Spectrometry

Dongxin Lyu, Jingbo Zhou, Hongxin Xiang, Yuqiang Li+1

Westlake University Hangzhou, Zhejiang, China · Hunan University Changsha, Hunan, China · Shanghai Artificial Intelligence Laboratory Shanghai, China+1

Memory-Augmented Neural Networks Transformer Inference Mass Spectrometry Autoregressive Decoding

Open paper AskPDF arXiv

De novo peptide sequencing from tandem mass spectrometry is pivotal in proteomics, enabling identification of novel peptides without reference databases. While recent Transformer-based encoder-decoder models have achieved remarkable performance, we uncover a critical pathology in their inference dynamics. Through comprehensive feature scaling experiments, we demonstrate that existing auto-regressive peptide decoders tend to over-rely on generated-sequence priors while progressively under-utilizing fine-grained physical evidence from the input mass spectrum. This phenomenon leads to suboptimal results, where generated peptide sequences are biologically plausible yet not faithful to the input spectrum. To rectify this, we propose MemNovo, a training-free and plug-and-play mechanism that re-balances peptide and spectral contributions at inference time. MemNovo alleviates the information bottleneck by establishing a persistent spectral memory bank and injecting retrieved features directly into the final decoding stage via an ultra-conservative residual connection. Theoretical analysis confirms that this mechanism restores the mutual information between the decoder state and the raw spectrum. Extensive experiments on the Nine Species benchmark with two representative baselines, Casanovo and InstaNovo, demonstrate that MemNovo consistently improves both amino acid precision and peptide precision, achieving up to 39.1% relative improvement in peptide precision for Casanovo and up to 3.9% for InstaNovo, with negligible computational overhead.

Abstract

Mass Spectrometry

Explore similar work

MemNovo: Look Back at the Spectrum for Balanced De Novo Peptide Sequencing from Mass Spectrometry

Regressor-guided Diffusion Model for De Novo Peptide Sequencing with Explicit Mass Control

GyroNovo: Error-Guided Fragment Imputation with Mass-Aware Attention for De Novo Peptide Sequencing