Positive-Unlabeled Learning for Predicting Small Molecule MS2 Identifiability from MS1 Context and Acquisition Parameters

Technical Summary: Positive-Unlabeled Learning for Predicting Small Molecule MS2 Identifiability

Problem Statement
In untargeted metabolomics, the quality of tandem mass spectra (MS2) is the primary determinant of metabolite identifiability. However, optimizing MS2 acquisition parameters through experimental trial-and-error is costly, time-consuming, and impractical given the vast diversity of samples and instruments. A critical barrier to developing computational solutions for quality assessment is the absence of reliable negative labels. Most MS2 spectra remain unannotated not because they are of low quality, but because the corresponding compounds are absent from reference libraries. This "label ambiguity" renders standard supervised learning approaches ineffective, as the model cannot distinguish between low-quality spectra and high-quality spectra of unknown compounds. Consequently, making real-time decisions during data acquisition regarding which scans to prioritize is currently compromised.

Methodology
To address the lack of negative labels, the authors formulate the problem as Positive-Unlabeled (PU) learning. The proposed framework is a deep learning model designed to predict the probability that an MS2 scan will be identifiable using only the preceding MS1 spectrum and the instrument acquisition parameters, without inspecting the MS2 spectrum itself.

The core technical innovation lies in the use of a non-negative positive-unlabeled objective. This specific loss function enables robust training despite the missing negative labels by treating the unlabeled data as a mixture of positive and negative instances, rather than assuming all unlabeled data are negative. The model was trained on a massive dataset comprising over eight million MS2 scans derived from public Orbitrap metabolomics datasets.

Key Contributions

1. Novel Problem Formulation: The paper reframes MS2 identifiability prediction as a PU learning task, directly addressing the fundamental limitation of missing negative labels in metabolomics.
2. Context-Aware Prediction: The model successfully predicts identifiability based solely on precursor context (MS1) and acquisition settings, eliminating the need for post-acquisition spectral inspection.
3. Robust Training Objective: The implementation of a non-negative PU objective allows for effective model training in the absence of ground-truth negative examples.
4. Open Resources: The authors provide the code, model weights, and processed datasets to facilitate reproducibility and further research.

Results
The model was evaluated on laboratory-disjoint benchmarks to ensure generalizability. Key findings include:

• High Recovery Rate: The model recovered 90% of known identifiable spectra in a held-out test set.
• Generalization: Predicted probabilities generalized effectively to unseen laboratories, demonstrating robustness across different experimental settings.
• Physicochemical Stratification: The model successfully stratified unlabeled spectra in a manner consistent with physicochemical properties.
• Independent Validation: High predicted probabilities were independently correlated with:
  • Increased structural explainability.
  • Richer fragmentation patterns.
  • Canonical precursor charge states.
  • Reduced spectral interference.

Significance
The paper claims that MS2 identifiability can be anticipated solely from precursor context and acquisition settings. By overcoming the label ambiguity that has historically limited supervised learning in this domain, this framework provides a computationally efficient method to assess spectral quality. This capability has the potential to inform acquisition-time decision-making, allowing researchers to optimize instrument parameters or prioritize specific scans for identification without the prohibitive costs of exhaustive experimental optimization.