Research map / Probabilistic and scientific modeling / Molecules and proteins
PROSPECT PTMs: Rich Labeled Tandem Mass Spectrometry Dataset of Modified Peptides for Machine Learning in Proteomics
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 PROSPECT PTMs, four annotated tandem mass spectrometry datasets of ~1.2M modified peptides (30 amino acid-PTM pairs) from ProteomeTools, and benchmarks Prosit variants, DeepLC, and AlphaPeptDeep on retention time, fragment intensity, and charge state tasks.
Problem
High-quality labeled reference data for MS/MS property prediction (retention time, fragment ion intensity, charge state) of modified peptides is missing, which impedes machine learning model development for PTM-containing peptides.
Approach
Four curated, expert-annotated datasets of modified peptides with ProForma sequence strings, split so that Test-PTM contains unseen PTMs and unmodified counterparts, plus benchmarks comparing PTM encoding strategies (naive embedding, delta mass, delta atoms).
Key results
- Prosit-DeltaAtoms achieved the best retention time Timedelta 95 of 12.7 (unseen PTMs) and 9.9 (seen PTMs) versus 26.1 and 25.8 for the Prosit baseline.
- Prosit-DeltaAtoms reached spectral angle 0.86 on unseen PTMs and 0.89 on seen PTMs, the best fragment intensity results among the Prosit variants.
- The Prosit baseline trained on unmodified peptides performed poorly on modified sequences, with retention time Timedelta 95 of 26.1 on unseen PTMs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Encode PTMs using introduced atom counts instead of ignoring PTMs | Prosit · Test-PTM, unseen PTMs | Timedelta 95: 26.1 -> 12.7 | Helped |
| Encode PTMs using introduced atom counts instead of ignoring PTMs | Prosit · Test-PTM, seen PTMs | Timedelta 95: 25.8 -> 9.9 | Helped |
| Encode PTMs using introduced atom counts instead of ignoring PTMs | Prosit · Test-PTM, unseen PTMs | Spectral Angle: 0.75 -> 0.86 | Helped |
| Encode PTMs using introduced atom counts instead of ignoring PTMs | Prosit · Test-PTM, seen PTMs | Spectral Angle: 0.74 -> 0.89 | Helped |
| Encode PTMs using introduced mass instead of ignoring PTMs | Prosit · Test-PTM, unseen PTMs | Timedelta 95: 26.1 -> 17.2 | Helped |
| Encode PTMs using introduced mass instead of ignoring PTMs | Prosit · Test-PTM, seen PTMs | Timedelta 95: 25.8 -> 14.2 | Helped |
| Encode PTMs using introduced mass instead of ignoring PTMs | Prosit · Test-PTM, unseen PTMs | Spectral Angle: 0.75 -> 0.81 | Helped |
| Encode PTMs using introduced mass instead of ignoring PTMs | Prosit · Test-PTM, seen PTMs | Spectral Angle: 0.74 -> 0.87 | Helped |
| Use naive PTM encoding instead of ignoring PTMs | Prosit · Test-PTM, unseen PTMs | Timedelta 95: 26.1 -> 25.2 | Helped |
| Use naive PTM encoding instead of ignoring PTMs | Prosit · Test-PTM, seen PTMs | Timedelta 95: 25.8 -> 16.4 | Helped |
| Use naive PTM encoding instead of ignoring PTMs | Prosit · Test-PTM, unseen PTMs | Spectral Angle: 0.75 -> 0.79 | Helped |
| Use naive PTM encoding instead of ignoring PTMs | Prosit · Test-PTM, seen PTMs | Spectral Angle: 0.74 -> 0.89 | Helped |
Surprising findings
- Modifications with the same mass often affect retention time and fragment intensities differently, so simple mass-based features are insufficient.
- Some modifications share a Unimod ID and molecular structure yet differ in stereochemistry and impact peptide properties differently.
Limitations
- Annotation pipeline over-annotates peaks and can assign the same label to multiple peaks within m/z tolerance.
- Mass analyzers covered are only Orbitraps and Iontraps; other instruments are not represented.
- Most PROSPECT PTM peptides occur in a single precursor charge state, limiting charge state prediction data.
Experimental setup
Datasets: ProteomeTools PROSPECT 21-PTMs
Benchmarks: PROSPECT PTMs Test-PTM
Metrics: Timedelta 95 Spectral Angle MAE
Baselines: Prosit baseline DeepLC AlphaPeptDeep Prosit naive-encoding Prosit DeltaMass
Builds on: Prosit PROSPECT ProteomeTools AlphaPeptDeep DeepLC
Closest papers by meaning
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction (ICLR 2025 · similarity 0.88)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.88)
- NovoBench: Benchmarking Deep Learning-based \emph{De Novo} Sequencing Methods in Proteomics (NeurIPS 2024 · similarity 0.86)
- AdaNovo: Towards Robust \emph{De Novo} Peptide Sequencing in Proteomics against Data Biases (NeurIPS 2024 · similarity 0.86)
- Improving PTM Site Prediction by Coupling of Multi-Granularity Structure and Multi-Scale Sequence Representation (AAAI 2024 · similarity 0.85)
- MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.