PROSPECT is a massive, standardized, and expertly labeled dataset of over 61 million mass spectra designed to benchmark machine learning models in proteomics. By providing a high-quality "ground truth" for peptide properties, it lowers the entry barrier for AI researchers and enables fair comparison across models for the first time.
The Missing "North Star" in Proteomics
In the world of proteomics—the large-scale study of proteins—scientists use Tandem Mass Spectrometry (MS/MS) to identify the building blocks of life. The process involves breaking proteins into smaller chains called peptides, fragmenting them, and measuring the resulting "fingerprint" (a spectrum).
While deep learning has shown incredible promise in predicting these fingerprints from amino acid sequences, the field has historically lacked a standardized benchmark. Most researchers train models on private, in-house datasets or poorly curated public repositories. This fragmentation makes it nearly impossible to compare two different models fairly or to know if a new architecture is genuinely superior or merely overfitted to a specific lab's noise. PROSPECT addresses this critical gap by providing a massive, standardized reference dataset.
Methods & Dataset Construction
To build PROSPECT, the authors processed raw data from ProteomeTools, a project that utilizes synthetic peptides (manually created sequences) rather than complex, noisy human samples.
The dataset construction followed a rigorous pipeline:
- Identification: The team used MaxQuant software to match raw mass spectrometry data to known peptide sequences.
- Annotation: This is the dataset's "secret sauce." An expert system was deployed to label not just the standard fragment ions (b and y ions) but also 26 different types of "neutral losses"—chemical groups that detach during fragmentation.
- Scale: The resulting dataset, hosted on Zenodo, spans 12 data packages containing 838,000 unique peptides, 61.7 million spectra, and 5.7 billion annotated peaks.
Unlike previous datasets used for models like Prosit or DeepLC, PROSPECT is significantly larger and includes far more detailed annotations for charge states and neutral losses.
Task Formulation & Applications
PROSPECT frames the complex biological data into clear, standardized machine learning tasks, making it highly accessible for AI researchers.
The dataset is primarily designed for two supervised ML tasks:
- Retention Time Prediction: A regression task predicting how long a peptide takes to travel through a liquid chromatography column.
- MS/MS Spectrum Prediction: A multi-output regression task predicting the intensity of various fragment ions across a spectrum.
Key Insights and Results
By establishing a standardized dataset, the authors uncovered several important insights for the field:
- Synthetic Data is a Feature, Not a Bug: By using synthetic peptides, the authors eliminated the overlapping "noise" inherent in biological samples. This provides a cleaner signal, allowing ML models to learn the fundamental physics of peptide fragmentation.
- Neutral Losses are Underutilized: Most traditional models ignore neutral losses. PROSPECT's inclusion of these features suggests that these "minor" peaks contain valuable chemical information that deep learning models can exploit to improve identification accuracy.
- Standardized Evaluation: The paper moves beyond simple "accuracy" by recommending specific, robust metrics. They propose using Spectral Angle (SA) for spectra and $\Delta t_{95\%}$ (the time window containing 95% of prediction errors) for retention time, providing a common evaluation language for the community.
To prove the dataset's viability, the authors benchmarked it using Prosit, a state-of-the-art deep learning model. For retention time, the model achieved a near-perfect Pearson correlation ($R = 1.00$) with a $\Delta t_{95\%}$ of 85 seconds. For spectrum prediction, they demonstrated that model performance varies significantly based on the fragmentation method (e.g., HCD vs. CID) and the precursor charge, highlighting that a "one size fits all" model is insufficient for modern proteomics.
Limitations and Future Work
While PROSPECT is a massive leap forward, the authors note a few limitations:
- Limited Modifications: The current dataset only includes two common chemical modifications (carbamidomethylation and oxidation). Real-world biology involves hundreds of Post-Translational Modifications (PTMs), such as phosphorylation, which are not yet covered.
- Synthetic vs. Biological Gap: While synthetic data is clean and excellent for foundational training, models trained on it may still require fine-tuning to handle the "messiness" of real human tissue or blood samples.
Looking ahead, PROSPECT opens the door for advanced ML applications beyond simple prediction. The authors suggest the dataset could power de novo sequencing (identifying a peptide directly from a spectrum without a database) or self-supervised learning to create rich amino acid embeddings that capture complex chemical properties far better than simple one-hot encoding.
Summary updated