SemiNovo: Learning Beyond Search-Identified Spectra for De Novo Peptide Sequencing
Authors: Pan Liu, Dongxin Lyu, Yusen Tan, Yifan Li, Hai-tao Yu, Yunhua Zhong, Zikun Wang, Tianze Ling, Jun Xia
Organizations: Information Hub, HKUST(GZ), Guangzhou, China · Institute of Engineering, Westlake University, Hangzhou, China · Department of Chemistry, The University of Hong Kong, Hong Kong, China · School of Life Sciences, Tsinghua University, Beijing, China
Sensitive peptide identification is fundamental to protein characterization and novel protein discovery, for which de novo peptide sequencing (DNPS) from tandem mass spectra provides an important database-independent approach. Although deep learning has substantially advanced de novo sequencing, existing models are predominantly trained on high-confidence peptide-spectrum matches obtained through database search. This search-conditioned supervision excludes a large fraction of unidentified spectra, creating a distribution mismatch between the selectively labeled training data and the complete spectrum stream encountered at inference time, thereby limiting peptide identification performance. To address this challenge, we propose SemiNovo, a semi-supervised framework that learns jointly from identified and unlabeled spectra. SemiNovo first constructs DarkSpec, a large-scale quality-controlled corpus of unlabeled spectra collected from diverse public proteomics projects, thereby extending training beyond search-identified spectra and mitigating the resulting shift in DNPS supervision. It then combines stepwise mass-constrained autoregressive decoding with cumulative path confidence to derive reliable supervision from unlabeled spectra. Moreover, SemiNovo introduces head-wise chemical-evidence gating to improve spectral evidence aggregation under noisy and incomplete fragmentation. Experiments show that SemiNovo performs favorably against state-of-the-art methods on multiple benchmarks. Code and dataset are publicly available.
Figures & tables
Figure 1. Database search induces biased supervision. (a) Tandem mass spectrometry fragments peptides into spectra that must be assigned amino-acid sequences. (b) Candidate-database search and false discovery rate control identify only part of the acquired spectra. (c) Retained peptide–spectrum matches occupy a selective subset of the complete spectral space. (d) Increasing identification stringency separates retained and excluded distributions, leaving informative dark spectra outside training. Four panels introduce tandem mass spectrometry peptide identification and illustrate search-conditioned supervision bias through database filtering, selective coverage of spectral representation space, and the growing distribution separation between identified and dark spectra.
Figure 2. Overview of SemiNovo . Labeled, weak, and strong spectrum views enter vertically aligned exponential moving average (EMA) teacher and student networks with head-wise chemical-evidence gating. Mass-constrained autoregressive decoding retains the longest prefix whose cumulative path confidence exceeds τ . The student jointly optimizes identified peptides and reliable pseudo-prefixes. The figure shows labeled, weak, and strong spectrum views on the left; vertically aligned EMA teacher and student encoder–decoder networks with chemical-evidence gates in the center; and reliable pseudo-label construction and joint supervised and unsupervised learning on the right.
Figure 3. Chemical-evidence gating suppresses non-localizing precursor-query updates. (a–b) Mean decoder self-attention in Casanovo and SemiNovo. (c) Layer-wise SemiNovo gate values for precursor and residue-prefix queries. Smaller precursor-query gates reduce the influence of global precursor context on residue prediction. Lines and shaded regions denote mean and standard deviation. Two decoder self-attention heatmaps compare Casanovo and SemiNovo, followed by a line plot showing that SemiNovo assigns consistently lower gate values to the precursor query than to residue-prefix queries across nine decoder layers.
Peptide level
Amino-acid level
Nine-species
HC-PT
Seven-species
Nine-species
HC-PT
Seven-species
Method
Prec.
AUC
Prec.
AUC
Prec.
AUC
Prec.
Recall
Prec.
Recall
Prec.
Recall
DeepNovo
0.428
0.376
0.313
0.255
0.204
0.136
0.696
0.638
0.531
0.534
0.492
0.433
PointNovo
0.480
0.436
0.419
0.373
0.022
0.007
0.740
0.671
0.623
0.622
0.196
0.169
Casanovo
0.481
0.439
0.211
0.177
0.119
0.084
0.697
0.696
0.442
0.453
0.322
0.327
AdaNovo
0.505
0.469
0.212
0.178
0.174
0.135
0.698
0.709
0.442
0.451
0.379
0.385
Table 1. De Novo peptide sequencing performance under the NovoBench protocol. We report peptide precision and area under the precision–coverage curve (AUC) together with amino-acid precision and recall. Best and second-best results are shown in bold and underlined, respectively.
Method
Nine-species ↑
HC-PT ↑
Seven-species ↑
DeepNovo
0.576
0.626
0.391
PointNovo
0.629
0.676
0.117
Casanovo
0.706
0.501
0.360
AdaNovo
0.652
0.552
0.448
π -HelixNovo
0.680
0.568
0.473
SearchNovo
0.764
0.715
0.472
Table 2. PTM precision on the three NovoBench datasets.
Figure 4. Three central ablations on Nine-species. Grouped bars compare the complete model with removal of stepwise mass constraints, removal of chemical-evidence gating, and replacement of cumulative path confidence by mean token confidence. Three grouped-bar panels compare the full model with variants without mass constraints, chemical-evidence gating, or cumulative path confidence.
Figure 5. Exact pseudo-label error rate as a function of accepted coverage on the complete Nine-species validation split. The lower panel reports the error reduction of cumulative path confidence relative to the alternatives. Error–coverage curves compare cumulative path, mean token, and minimum token confidence, with a lower panel showing the error reduction from cumulative path confidence.
Factor
Value
Pep. prec.
Curve AUC
Path threshold τ
0.85
0.621
0.572
0.90
0.631
0.595
0.95
0.628
0.588
Loss weight λu
0.5
0.629
0.586
1.0
0.631
0.595
1.5
0.621
0.577
Table 3. Hyperparameter and unlabeled-data sensitivity on the Nine-species validation split.
Figure 6. Three qualitative Nine-species cases. Each row places a compact tandem mass spectrum beside its sequence-level predictions. Gray sticks denote unassigned peaks; blue and orange sticks denote peaks matching theoretical b- and y-ions within 0.02 Da. Blue and red highlights mark the corresponding correct and erroneous sequence regions, respectively; lowercase residues denote their modified forms. Three annotated tandem mass spectra compare exact SemiNovo predictions with supervised predictions containing N-terminal, local-order, and modified-residue errors.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Project
Spectra
Project
Spectra
PXD055748
2,059,318
PXD071883
251,927
PXD060117
666,617
PXD072146
933,697
PXD060201
918,864
PXD072226
467,463
PXD060819
591,603
PXD072402
217,986
PXD061438
1,239,989
PXD072579
88,232
PXD061644
3,001,044
PXD072877
983,211
Appendix
Table 4. Public PRIDE projects in the DarkSpec source pool.
Stage
Spectra
Raw source pool
44,963,379
Rejected: fewer than 20 valid peaks
3,700,995
Rejected: invalid precursor charge
166,308
Rejected: malformed peak arrays
7
Quality-controlled pool
41,096,069
Fixed training corpus
4,500,000
Appendix
Table 5. Quality control and experimental corpus size.
Dataset
Split
Spectra
Exact
Near-key (%)
Nine-species
Train
499,402
0
220 (0.044)
Validation
28,572
0
7 (0.025)
Test
27,142
0
10 (0.037)
HC-PT
Train
213,284
0
58 (0.027)
Validation
25,718
0
7 (0.027)
Test
26,536
0
6 (0.023)
Appendix
Table 6. Data-level overlap between DarkSpec and NovoBench. Near-key entries report candidate count and percentage of the benchmark split.
Component
Setting
Encoder / decoder blocks
9 / 9
Hidden / feed-forward width
512 / 1,024
Attention heads
8
Trainable parameters
48.2M
Dropout / weight decay
0.1 / 0.05
Maximum peaks / sequence length
150 / 100
Appendix
Table 7. Model and optimization configuration.
Random seed
Pep. precision
Curve AUC
454
0.635
0.593
2,026
0.622
0.584
1,145
0.639
0.591
Mean ± std.
0.632±0.009
0.589±0.005
Appendix
Table 8. Statistical stability on Nine-species across three random seeds.
Training setting
Pep. precision
Curve AUC
Supervised checkpoint
0.582
0.538
Supervised continuation
0.587
0.544
SemiNovo
0.635
0.593
Appendix
Table 9. Matched-budget training control on Nine-species.
Confidence criterion
AUC
Acc.@20%
Acc.@40%
Acc.@60%
Cumulative path
0.8531
97.65
92.99
85.89
Mean token
0.8493
95.50
92.75
85.80
Minimum token
0.8504
97.34
92.93
85.45
Appendix
Table 10. Pseudo-label ranking quality on the Nine-species validation split.
Figure 7. Training dynamics for the Nine-species semi-supervised run initialized from the best supervised checkpoint. Reliable-prefix length, complete-sequence acceptance, and validation sequence accuracy follow distinct trajectories as the teacher adapts to 4.5 million unlabeled spectra. Aligned training curves show a growing reliable-prefix length, non-monotonic complete-sequence acceptance, and validation sequence accuracy that peaks before slightly declining.
Method
Params
Labeled/s
Total/s
Peak memory
Casanovo
47.35M
3,513.9
3,513.9
14.60 GiB
SemiNovo (supervised)
48.17M
7,390.1
7,390.1
8.69 GiB
SemiNovo (semi-supervised)
48.17M
1,404.5
2,808.9
12.23 GiB
Appendix
Table 11. Single-H200 training efficiency at batch size 512.
Method
Search
Spectra/s
Peak memory
Casanovo
Original
6.79
4.81 GiB
SemiNovo (supervised)
Casanovo-style
93.15
3.04 GiB
SemiNovo (supervised)
Mass-pruned
611.89
3.04 GiB
SemiNovo
Casanovo-style
110.08
3.04 GiB
SemiNovo
Mass-pruned
606.46
3.04 GiB
Appendix
Table 12. Single-H200 inference efficiency at batch size 512.
Pseudo-label search
Pseudo/s
Full step/s
Memory
Fixed-length greedy
1,516.0
2,309.4
12.89 GiB
Confidence early stopping
2,187.7
3,092.0
12.24 GiB
Early stopping + mass pruning
2,160.1
3,046.2
12.24 GiB
Appendix
Table 13. Efficiency of pseudo-label search at batch size 512.
Dataset
Precursor m/z
z≥3 (%)
Peaks
Entropy
DarkSpec
770.9
59.7
150
4.55
Nine-species
650.8
39.7
89
4.32
HC-PT
598.3
29.7
109
4.55
Seven-species
688.8
36.5
150
4.88
Appendix
Table 14. Distribution statistics after canonical preprocessing. Values are computed from 20,000 randomly sampled spectra per dataset.
Mass spectrometry measures ion mass-to-charge patterns to identify and quantify molecules, with AI linking spectra to structures, peptides, and biological signals.
The discovery of novel proteins relies on sensitive protein identification, for which de novo peptide sequencing (DNPS) from mass spectra is a crucial approach. While deep learning has advanced DNPS, existing models inadequately enforce the fundamental mass consistency constraint, that a predicted peptide's mass must match the experimental measured precursor mass. Previous DNPS methods often treat this critical information as a simple input feature or use it in post-processing, leading to numerous implausible predictions that do not adhere to this fundamental physical property. To address this limitation, we introduce DiffuNovo, a novel regressor-guided diffusion model for de novo peptide sequencing that provides explicit peptide-level mass control. Our approach integrates the mass constraint at two critical stages: during training, a novel peptide-level mass loss guides model optimization, while at inference, regressor-based guidance from gradient-based updates in the latent space steers the generation to compel the predicted peptide adheres to the mass constraint. Comprehensive evaluations on established benchmarks demonstrate that DiffuNovo surpasses state-of-the-art methods in DNPS accuracy. Additionally, as the first DNPS model to employ a diffusion model as its core backbone, DiffuNovo leverages the powerful controllability of diffusion architecture and achieves a significant reduction in mass error, thereby producing much more physically plausible peptides. These innovations represent a substantial advancement toward robust and broadly applicable DNPS. The source code is available in the supplementary material.
Shaorong Chen, Jingbo Zhou, Jun Xia
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
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.
Dongxin Lyu, Jingbo Zhou, Hongxin Xiang +2
Westlake University Hangzhou, Zhejiang, China · Hunan University Changsha, Hunan, China · Shanghai Artificial Intelligence Laboratory Shanghai, China +1
De novo peptide sequencing from tandem mass spectra is essential for identifying peptides without relying on reference databases. Despite advances in deep learning, accurate sequencing remains challenging because experimental spectra are often sparse, noisy, and incomplete, leaving informative b- and y-ion fragments unobserved. Existing methods attempt to recover this missing evidence via latent-space imputation before autoregressive decoding. However, they typically treat imputation as a fixed reconstruction task, without considering which missing fragments are most relevant to decoder errors. Moreover, existing peak representations do not explicitly model mass differences between peaks, despite their fundamental importance. We introduce GyroNovo, a framework with two main contributions. First, we use decoder errors observed during training to adapt the imputation objective, prioritizing fragments associated with frequent decoding errors. We further use the decoder error distribution to construct easy and hard augmented views of each spectrum, enabling the decoder to learn under varying degrees of spectral corruption and missing-fragment severity. Second, we introduce a mass-aware inductive bias into self-attention by using rotary embeddings to encode pairwise mass differences between spectral peaks. Together, these components align missing-fragment recovery with decoder behavior while explicitly incorporating the mass relationships that underlie peptide fragmentation. At inference time, GyroNovo retains a standard encoder-imputer-decoder architecture and requires neither additional inputs nor auxiliary search procedures. Experiments on NovoBench show gains of about 9 percentage points in peptide-level precision and 7 percentage points in amino-acid-level precision over the state-of-the-art baseline. Code: https://github.com/UBC-NLP/gyronovo.
Abdellah El Mekki, Laks V. S. Lakshmanan, Muhammad Abdul-Mageed
The University of British Columbia · Canada Research Chair in NLP and ML