Research map / Probabilistic and scientific modeling / Molecules and proteins

PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein interface and docking prediction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces PepSpecBench, a unified, backbone-disjoint, cross-species benchmark for peptide MS/MS fragment-intensity prediction that projects six baselines into a shared 234-dim ion space and probes robustness to NCE and charge perturbations.

Problem

Peptide MS/MS spectrum prediction lacks comparable evaluation: heterogeneous preprocessing and output spaces, leakage from random or modified-sequence splits, and little cross-species or metadata-sensitivity testing.

Approach

Harmonizes PROSPECT and MassIVE-KB under shared filters, enforces backbone-disjoint splits by hashing the naked peptide sequence, projects all model outputs into one canonical b/y ion space, and adds zero-shot cross-species OOD sets and NCE/charge perturbation probes.

Key results

• Prosit Transformer has the best in-domain SA on MassIVE-KB (0.0977) while Prosit leads on PROSPECT (SA 0.1376), so top rankings are source-dependent.
• Naive random splitting lowers Prosit's median SA on PROSPECT from 0.1376 to 0.1254 and raises PCC from 0.8921 to 0.9099, showing inflated accuracy.
• Under MassIVE-KB training, PredFull outperforms Prosit on all five OOD testbeds, and Prosit reaches negative median PCC on several subsets; AlphaPeptDeep collapses on 98.0% of charge-perturbed spectra.

Surprising findings

• In-domain leaderboard ordering fails to predict OOD robustness; the MassIVE-KB-trained Prosit models rank last-but-one OOD despite top in-domain scores.
• AlphaPeptDeep ignores the precursor charge input, with 98.0% collapsed predictions under a z=2 to 3 shift.

Limitations

• Backbone-disjoint splitting still permits higher-order sequence homology (13% of PROSPECT test backbones within edit distance 1 of train).
• Scope limited to three PTMs (acetylation, carbamidomethylation, oxidation), peptide lengths 6-40, and charges 1-6.
• Cross-species OOD comparisons are confounded by differences in NCE, PTM rate and instrument type across sources.

Experimental setup

Datasets: PROSPECT MassIVE-KB PXD014877 PXD027742

Benchmarks: PepSpecBench in-domain (MassIVE-KB, PROSPECT) PepSpecBench cross-species OOD suite (7 subsets)

Metrics: spectral angle (SA) spectral angle similarity (SAS) Pearson correlation (PCC)

Baselines: Prosit Prosit Transformer AlphaPeptDeep PredFull UniSpec FastSpel

Builds on: Prosit AlphaPeptDeep PredFull UniSpec FastSpel

Closest papers by meaning

• NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.91)
• MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.89)
• PROSPECT PTMs: Rich Labeled Tandem Mass Spectrometry Dataset of Modified Peptides for Machine Learning in Proteomics (NeurIPS 2024 · similarity 0.88)
• Latent Imputation before Prediction: A New Computational Paradigm for De Novo Peptide Sequencing (ICML 2025 · similarity 0.87)
• GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem (NeurIPS 2026 · similarity 0.87)
• AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases (NeurIPS 2024 · similarity 0.87)