Input
- precursor m/z
- precursor charge
- fragment peak m/z
- fragment peak intensity
- original spectrum ID (prediction alignment only)
Competition Overview
Translate tandem mass spectra into peptide sequences
Proteins are major functional molecules in living systems, and their functions are closely related to amino acid sequence and structure. Mass spectrometry measures peptides at high throughput; the challenge is to recover sequence from the measured signal.
Competition Goals
Central goal: develop AI models that accurately translate MS/MS spectra into peptide sequences.
For every unknown spectrum, the model predicts the complete peptide sequence that generated it and reports a confidence score.
Challenge Task
Input an unknown spectrum; output a sequence and confidence
This is a spectrum-to-sequence generation task: directly recover a peptide sequence from fragment-peak information. Dataset scale, files and splits are described on the Dataset page.
AI Model
Any architecture. Learn fragmentation, mass constraints, intensity patterns and sequence context from spectrum–peptide pairs.
Output
- predicted peptide sequence
- prediction confidence score
Database search
Like multiple choice
Generate candidates from a predefined protein database, then select the sequence that best matches the spectrum.
VS
de novo sequencing
Like an open-ended question
No candidate answers are provided; the model must generate the complete amino acid sequence directly from the spectrum.
Research directions
Accurate sequencing, generalization and efficient inference
Accurate sequencing
Can AI reconstruct complete sequences from complex, sparse and noisy spectra?
Generalization
Will performance hold on unseen peptides and independent experimental projects?
Reproducibility
Can frozen code and weights reproduce submitted predictions in a shared evaluation environment?
Data scaling
How does performance change from millions to hundreds of millions of training pairs?
Efficient AI
Can accuracy be balanced with model size, inference time and compute?
Why de novo sequencing?
From searching a database to discovering unknown sequences
Database search can only match candidates in a predefined protein space. If a real peptide is absent, it cannot be returned. de novo generation opens discovery space for novel peptides, variants, non-canonical products, immunopeptides and poorly annotated organisms. These are potential applications, not separate competition tracks.
An AI perspective
A spectrum-to-sequence generation problem with physical constraints
From an AI perspective, proteomics maps directly onto familiar problems in signals, sequences and domain generalization.
Proteomics
AI Perspective
MS/MS spectrum
Sparse continuous signal
Spectrum peak
Observation / token
m/z
Continuous position
Peak intensity
Signal strength
Peptide
Target sequence
Amino acid
Vocabulary
Precursor mass
Global physical constraint
Fragmentation rule
Physical inductive bias
Instrument
Technical domain
Species
Biological domain
de novo sequencing
Spectrum-to-sequence generation
Whether your background is in Transformers, LLMs, sequence modeling, computer vision, diffusion models, representation learning, graph neural networks or foundation models, you can contribute to this problem.
Why is this a challenging AI problem?
Sparse, continuous and noisy input
Variable-length structured generation
Physical mass constraints
Instrument, species and project shifts
OOD generalization and leakage control
Reproducible inference with frozen models
accuracy × generalization × reproducibility × efficiency
de novo peptide sequencing
Understand de novo sequencing in four steps
You do not need a full proteomics course. Start with how a peptide becomes a spectrum, then see how AI solves the process in reverse.
Protein becomes peptides
Proteins consist of amino acids and are commonly broken into shorter peptides. PEPTIDE has seven residues.
The peptide is fragmented
A selected peptide ion is fragmented, producing characteristic fragment-ion series such as b and y ions.
Fragments form an MS/MS spectrum
The instrument measures fragment m/z and intensity. Mass differences between related peaks carry sequence information.
AI generates the sequence
The model learns fragmentation and mass relationships from spectrum–peptide pairs, then generates a sequence for an unseen spectrum.
This is an inverse problem
Forward experiment
Peptide sequence→Fragmentation + MS measurement→MS/MS spectrum
AI inference task
MS/MS spectrum→AI model→Peptide sequence + confidence
Keep learning
From fundamentals to current models
These public resources are sequenced for participants without a proteomics background.
Start with the basics
EMBL-EBI · Introduction to Proteomics
EMBL-EBI · MS-based Proteomics
Read representative models
DeepNovo · PubMed
Casanovo · ICML 2022
NovoBench · benchmark survey
Try it yourself
Casanovo · code & documentation
InstaNovo · code & model
Then consult the competition baseline, evaluation guide and submission example.
100 million training PSMs, consistent evaluation splits
Training
100M100 million PSMs in 417 Parquet files, from 527 projects and 14,208 source files, covering 38 standardized species/taxonomic labels and 10 instrument labels.
Public test
3M3 million PSMs. Spectra and reference labels support development and local evaluation. Species are represented in training, but projects are disjoint.
Hidden test
Spectra are released after model freezing and package verification. Reference labels remain private; species and project composition are not disclosed.
Isolation and generalization
Training, public test and hidden test are pairwise disjoint by project, source file and normalized peptide sequence (I/L equivalent, modification annotations removed). Both test sets use instrument conditions represented in training; there is no unseen-instrument condition. Hidden evaluation tests generalization to independent projects and unseen peptides.
Files and downloads
Parquet data are distributed through Google Drive and support batch reading. Inputs include fragment m/z and intensity, precursor m/z and charge. Original spectrum_id values align predictions and must not be used to retrieve answers. Training and public data open at launch; hidden spectra follow the staged timeline.
Data guide and Google Drive downloadsEvaluate complete peptide predictions
The primary metric is species-macro peptide recall; higher is better. Public scores provide development feedback. Hidden-test Private scores determine final ranking, subject to reproduction and compliance review.
Primary metric: species-macro peptide recall
For each species, divide the number of spectrum records with a correctly predicted complete peptide by all reference records for that species, then take the arithmetic mean across species. Every spectrum counts; peptide sequences are not deduplicated. NO_PREDICTION counts as incorrect and remains in the denominator.
Complete backbone matching
Predicted and reference sequences undergo identical normalization and must match exactly across the complete backbone. I/L are equivalent. Supported numeric and UNIMOD modification annotations are removed; modification type and position are not scored. Unsupported notation is rejected.
Confidence and runtime
A finite numeric confidence score is required, but it does not weight the primary metric or impose a confidence or 1% FDR cutoff. Runtime is an award-eligibility requirement, not a score weight or a near-tie ranking criterion. See Kaggle for the runtime limit and testing conditions.
Full evaluation detailsCompetition rules
Eligibility
Individuals and teams of up to five may enter, including students, researchers, engineers and industry teams, subject to applicable platform participation terms.
External data and models
External training data and pretrained models are allowed with disclosure of sources, versions, licenses and known data overlaps. Any model architecture is allowed; there is no separate restricted-data or research track.
De novo inference and integrity
Inference must not use protein/peptide databases, spectral libraries or database search to generate, retrieve, filter or rerank candidate answers. Training on hidden test data, manual test annotation, answer-retrieval components keyed to test identity or inferred species, exchanging private predictions and bypassing submission limits through multiple accounts are prohibited. Scores obtained in violation of the rules will not be recognized.
Model freezing and reproduction
Teams entering final evaluation must submit and freeze complete code, weights, configuration and runtime environment before hidden spectra are released. Approved teams generate final predictions only with the frozen pipeline. Organizers reproduce predictions and audit training-data sources; non-reproducible or non-compliant results are ineligible for awards.
Submission format and leaderboards
Public and final evaluation use the same Kaggle competition. Public CSVs retain original spectrum_id values and cover all 3 million public records, without padding hidden rows. Use the columns below; enter NO_PREDICTION and confidence 0 for abstentions. After hidden release, submit complete predictions for that stage. Daily submission limits are specified on Kaggle.
spectrum_idpredicted_peptide_sequenceconfidence_score
Format and public submission exampleCNY 170,000 in prizes, a six-month challenge
Registration and competition open on 26 September 2026 at 10:00 and close on 27 March 2027 at 00:00 (the end of 26 March). The competition lasts six months. All times are Beijing time (UTC+8).
Total prize pool: CNY 170,000 for six teams (including individual entrants). All amounts are before tax and paid in CNY.
Second prize¥30,0002 team(s) · amount per team
First prize¥50,0001 team(s) · amount per team
Third prize¥20,0003 team(s) · amount per team
Awards are subject to reproduction and compliance review. The competition also seeks to facilitate internship referrals, research exchange and collaboration; any opportunities will be announced separately.
Who can participate?
Students, researchers, engineers and industry teams in AI, machine learning, data science, bioinformatics, proteomics and related fields are welcome. Enter individually or as a team of up to five, subject to platform terms.
Kaggle · registration opens 26 September, 10:00Baseline and development support
The XuanjiNovo 100M pretrained baseline achieves 58.03% species-macro peptide backbone recall on all 3 million public spectra. Code and pretrained weights are open source. Teams may improve permitted pretrained models or develop other models. Evaluation documentation and CSV examples are provided on Kaggle.
XuanjiNovo repositoryEvaluation and submission guideTimeline · Beijing time UTC+8
Registration and development
2026-09-26 10:00
Training and public data open via Google Drive.
Model collection
≈ 2027-02-10
Development continues; updates are allowed until freezing.
Model freeze
≈ 2027-03-10
Final packages freeze; approximately two days for verification.
Hidden spectra release
2027-03-12
Approved teams infer with frozen pipelines for approximately two weeks.
Final prediction deadline
2027-03-26 24:00
Equivalent to 27 March 00:00; reproduction follows.
Post-competition reproduction depends on submission volume and workload. Results and awards follow review and any applicable appeals; dates will be announced separately. Schedule changes will be communicated in official competition announcements.
Registration and questions
Registration opens on 26 September 2026 at 10:00 Beijing time. Public development and final evaluation share one Kaggle page. Contact: · . See competition announcements and discussions for updates.
Announcements and discussionsOrganizers
Westlake University
Bruker
Frequently asked questions
When can I register?
Registration opens on Kaggle on 26 September 2026 at 10:00 Beijing time. Final predictions are due at the end of 26 March 2027.
Are external data and pretrained models allowed?
Yes, with source, version, license and known-overlap disclosure. Hidden-test training and answer-retrieval inference are prohibited.
Will GPUs be provided?
Participant compute support is not currently committed. Any arrangements will be announced separately; teams may use their own resources.
Does confidence affect ranking?
The primary score uses neither confidence weighting nor a 1% FDR cutoff. A finite numeric confidence field remains required.
Is hidden data available at launch?
No. Model packages are collected, frozen and verified before hidden spectra are released. Hidden labels remain private, and final results undergo reproduction review.
Do public submissions need hidden rows?
No. Submit predictions for all 3 million public spectra with original IDs, without rows from other stages.