False discovery rate estimation untargeted

false-discovery-rate-estimation-untargeted

Summary

The Passatutto method enables false discovery rate (FDR) estimation in large-scale untargeted metabolomics annotation by generating synthetic decoy spectra and databases in parallel with reference entries. This skill is essential for controlling annotation error rates when matching observed mass spectra against compound libraries without predefined target lists.

When to use

Apply this skill when performing untargeted metabolomics annotation (i.e., matching observed spectra to a compound database without a pre-defined target list) and you need to assign statistical significance or confidence to candidate metabolite identifications. Use it specifically when the scale of the annotation task makes manual validation infeasible and you require FDR estimates to distinguish true compound matches from false positives.

When NOT to use

• Input spectra are already matched to a predefined target list (targeted metabolomics) — use targeted statistical methods instead.
• The analysis goal is to rank or prioritize candidate identifications without requiring formal significance testing — decoy generation adds computational overhead not justified by the analysis question.
• Spectral features have already been aggregated into a feature table; decoy generation requires raw or minimally processed spectral data.

Inputs

• metabolomics spectral dataset (mass spectrometry spectra with m/z and intensity values)
• compound reference database (structured collection of known metabolite spectra or spectral properties)

Outputs

• decoy spectral database (synthetic spectra generated by permutation)
• unified search space (combined reference and decoy compound entries)
• FDR-annotated annotation results (candidate metabolite matches with per-match FDR values and significance estimates)

How to apply

Load both the metabolomics spectral dataset (mass spectrometry spectral measurements) and a compound reference database as inputs. Apply the Passatutto decoy generation algorithm to create synthetic decoy spectra by scrambling or permuting spectral features while preserving key statistical properties (e.g., peak intensity distribution, spectral density). Construct a parallel decoy compound database using these generated decoy spectra. Combine the reference and decoy databases into a unified search space. Score candidate annotations against both reference and decoy entries using the same scoring metric to compute FDR statistics. The FDR is derived from the ratio of decoy matches to reference matches; outputs include FDR-annotated results with significance estimates for each metabolite match, enabling practitioners to filter annotations by a chosen FDR threshold.

Related tools

• commons-math3 (Statistical computations and FDR calculations from scoring distributions)
• jfreechart (Visualization of decoy and reference scoring distributions for FDR diagnostics)
• jopt-simple (Command-line argument parsing for workflow configuration)
• trove4j (Memory-efficient data structures for large spectral and decoy databases)
• Passatutto (Reference implementation of the decoy generation and FDR estimation workflow) — https://github.com/boecker-lab/passatuto

Evaluation signals

• Decoy spectra statistics (peak count, intensity distribution, feature permutation patterns) match reference spectra marginal distributions, confirming decoy generation preserved key properties.
• FDR values are monotonically non-decreasing as candidate score thresholds are relaxed, confirming consistency of FDR estimation.
• The ratio of decoy matches to reference matches is stable across multiple runs or cross-validation folds, confirming reproducibility of the decoy search.
• Comparison of FDR-filtered annotations against independent validation data (e.g., standards, orthogonal techniques) shows expected sensitivity–specificity trade-off at chosen FDR thresholds.
• The unified search space contains equal numbers of reference and decoy entries (or a documented ratio) and no overlap between reference and decoy compound identifiers.

Limitations

• Decoy generation by feature permutation assumes that scrambled spectra preserve the null distribution of random matches; if spectral features are highly structured or skewed, decoy generation may not accurately calibrate FDR.
• The method requires a sufficiently large reference database to generate meaningful decoy candidates; performance on very small or specialized databases is not established.
• Computational cost scales with database size and spectrum complexity; no guidance is provided in the article on runtime or memory limits.
• The README notes 'No changelog found', indicating the project may lack version tracking or backward compatibility documentation.

Evidence

• [other] The paper describes a significance estimation method called Passatutto that is designed to enable large-scale untargeted metabolomics annotations, addressing the need for false discovery rate control in this domain.: "The paper describes a significance estimation method called Passatutto that is designed to enable large-scale untargeted metabolomics annotations, addressing the need for false discovery rate control"
• [other] Apply the Passatutto decoy generation algorithm to create synthetic decoy spectra by scrambling or permuting spectral features while preserving key statistical properties.: "Apply the Passatutto decoy generation algorithm to create synthetic decoy spectra by scrambling or permuting spectral features while preserving key statistical properties."
• [other] Score candidate annotations against both reference and decoy entries to compute false discovery rate statistics.: "Score candidate annotations against both reference and decoy entries to compute false discovery rate statistics."
• [other] Output the FDR-annotated results with significance estimates for each metabolite match.: "Output the FDR-annotated results with significance estimates for each metabolite match."
• [readme] commons-math3-3.4.1, hamcrest-core-1.3, jcommon-1.0.21, jfreechart-1.0.17-experimental: "commons-math3-3.4.1, hamcrest-core-1.3, jcommon-1.0.21, jfreechart-1.0.17-experimental"