Source-linked AI summary
Fragmentation trees reloaded
Kai Dührkop, Sebastian Böcker
TL;DR
Untargeted metabolomics still contains many unknown compounds, motivating methods that can interpret tandem mass spectra without complete library coverage. This paper scores fragmentation trees with a Bayesian maximum a posteriori estimator and reports significantly better performance on formula identification and structurally similar compound searching. The approach supports automated investigation of unknown metabolites, with performance depending on available information and modeling assumptions.
Problem
The vast majority of metabolites remain unknown, while database-based methods cannot identify compounds absent from available structure databases.
Method
The paper models fragmentation-tree selection as Bayesian maximum a posteriori estimation, using priors for molecular formulas and fragmentation edges and likelihoods for signal and noise peaks.
Results
The new scoring significantly outperforms previous approaches for de novo molecular-formula identification and searching databases for chemically similar compounds.
Takeaways & Limitations
The method can expedite untargeted-metabolomics workflows by enabling automated investigation of unknown compounds from tandem mass-spectrometry data.
Takeaways & Limitations
The likelihood model lacks general models for signal-peak intensities and noise-peak masses, so it uses peak mass for signals and intensity for noise.
Abstract
from arXiv · showhide
Metabolites, small molecules that are involved in cellular reactions, provide a direct functional signature of cellular state. Untargeted metabolomics experiments usually relies on tandem mass spectrometry to identify the thousands of compounds in a biological sample. Today, the vast majority of metabolites remain unknown. Fragmentation trees have become a powerful tool for the interpretation of tandem mass spectrometry data of small molecules. These trees are found by combinatorial optimization, and aim at explaining the experimental data via fragmentation cascades. To obtain biochemically meaningful results requires an elaborate optimization function. We present a new scoring for computing fragmentation trees, transforming the combinatorial optimization into a maximum a posteriori estimator. We demonstrate the superiority of the new scoring for two tasks: Both for the de novo identification of molecular formulas of unknown compounds, and for searching a database for structurally similar compounds, our methods performs significantly better than the previous scoring, as well as other methods for this task. Our method can expedite the workflow for untargeted metabolomics, allowing researchers to investigate unknowns using automated computational methods.
1 Introduction
Untargeted metabolomics must identify changing peaks despite the large fraction of unknown metabolites and limited library coverage. The paper introduces a Bayesian maximum a posteriori scoring approach for fragmentation trees that improves formula identification and structurally similar compound searching.
- 1 Introduction: The vast majority of metabolites remain unknown, making peak identity determination a central challenge in untargeted metabolomics.This challenge is especially pronounced for non-model organisms and secondary metabolites.
- 1 Introduction: Spectral-library and structure-database methods cannot reliably identify compounds absent from those resources.Existing molecular-formula methods often require information beyond tandem mass spectra.
- 1 Introduction: The paper chooses fragmentation trees using Bayesian analysis and maximum a posteriori estimation to explain observed tandem mass-spectrometry data.The new scoring targets the fragmentation tree that best explains the observed data.
- 1 Introduction: The new fragmentation-tree scoring significantly outperforms previous scoring and state-of-the-art methods in formula identification and chemically similar compound searching.These are the two derived evaluation tasks reported for the method.
- 1 Introduction: The method is implemented as a new version of the SIRIUS framework for MS and MS/MS analysis.The paper points readers to a schematic workflow in Fig. 1.
2 Fragmentation Trees
Fragmentation trees represent molecular formulas and their fragmentation cascades as rooted directed trees supported by measured spectrum peaks. Their computation can be reformulated as a weighted directed acyclic graph optimization with at most one formula selected per peak.
- 2 Fragmentation Trees: The measured spectrum consists of peak masses and positive intensities, compared with theoretical formulas using a user-provided relative mass-accuracy parameter.High measurement accuracy and sparse spectra generally allow a theoretical formula to explain at most one measured peak.
- 2 Fragmentation Trees: A fragmentation tree contains molecular-formula nodes for the ion and fragments, with directed edges modeling successive fragmentation events.Each node is reachable from the root through a unique sequence of edges.
- 2 Fragmentation Trees: For an edge from u to v, the molecular-formula loss is u −v, with u ≥v componentwise and a nonnegative loss formula.The theoretical mass function applies to fragment and loss formulas.
- 2 Fragmentation Trees: A tree is supported when each node maps injectively to a unique measured peak within the chosen mass-accuracy threshold.Unassigned spectrum peaks are treated as noise peaks.
- 2 Fragmentation Trees: The fragmentation graph is a directed acyclic graph whose nodes are candidate formulas for measured peaks and whose colors group formulas assigned to the same peak.At most one vertex of each color may be used in a fragmentation tree.
3 Maximum a Posteriori Estimation
The paper formulates fragmentation-tree selection as maximum a posteriori estimation, combining priors on tree structure and chemistry with a likelihood for observed and noise peaks. The resulting posterior ordering is optimized through weighted colorful subtrees.
- Posterior formulation: Maximum a posteriori estimation ranks candidate fragmentation trees by maximizing P(D|T)·P(T), combining data fit with prior biochemical knowledge.Only trees supported by the measured spectrum receive nonzero likelihood.
- Prior probability of the tree: The tree prior combines size, root-formula, fragment, and edge factors, including chemical plausibility, loss distributions, common fragments, and tree-size preferences.These components encode prior information about likely molecular formulas, losses, fragments, and informative tree sizes.
- Prior probability of the tree: Root priors constrain structurally implausible formulas and model valence, radical status, HMOTCR, RDBE, and related chemical characteristics without discarding formulas.The root prior can help identify formulas when compounds fragment poorly, while becoming less important as more peaks are explained.
- Prior probability of the tree: Edge priors penalize implausible or large losses and incorporate common losses and fragments, with common-loss probabilities learned from training data.Implausible-loss penalties rely on expert knowledge because the relevant penalties are difficult to learn from data.
- Likelihood of the tree: The likelihood models assigned peaks through mass deviations and treats unassigned peaks as noise, using independence assumptions across measured and noise peaks.The noise model uses peak intensities, whereas signal-peak assessment uses peak masses because general intensity and noise-mass models are unavailable.
- Optimization: With defined edge weights, the maximum-weight colorful subtree is equivalent to the fragmentation tree with maximum posterior probability.This connects the Bayesian formulation to the combinatorial optimization problem solved on the fragmentation graph.
4 Results
The evaluation fits model parameters on GNPS and Agilent spectra, then tests fragmentation-tree scoring for molecular-formula prediction and chemical-similarity searching. The new scoring improves molecular-formula recovery over prior approaches, while the supplied passages describe the similarity-search setup but not its numerical outcome.
- Datasets: 2,006 GNPS non-peptide compounds were analyzed alongside the Agilent MassHunter Forensics/Toxicology PCDL dataset.GNPS spectra were recorded in positive mode on an Agilent QTOF, with parent-mass accuracy below 10 ppm.
- Datasets: The datasets were split into CHNOPS and contains FClBrI batches, with 1,589 and 417 GNPS compounds respectively, and 1,540 and 580 Agilent compounds.The second batch contains at least one FClBrI atom.
- Parameter estimation: Hyperparameters for priors and likelihood estimation were fitted iteratively, using fragmentation trees from the previous round to determine current parameters.The paper refers readers to its long version for fitting details.
- Parameter estimation: Mass errors followed a normal distribution, while noise-peak frequency decayed exponentially with increasing intensity in both datasets.The fitted settings included MA = 10 ppm and σm = 10 for the mass-error model; noise parameters differed between GNPS and Agilent.
- Parameter estimation: 34 common losses were identified, including 13 previously listed losses and 16 assigned to known structures.After correcting for common losses, the observed loss-mass distribution agreed with a fitted log-normal distribution with mode 55.84 Da.
- Molecular-formula prediction: Molecular-formula prediction was evaluated by ranking fragmentation trees and formulas by posterior probability, using separate element-alphabet assumptions for the two batches.The evaluation compares the new scoring with prior scores paired with ILP and dynamic programming.
- Similarity search: Similarity search used leave-one-out and cross-database evaluations, measuring chemical similarity with Tanimoto coefficients and comparing tree alignments, peak counting, and MACCS fingerprints.MACCS requires known compound structures, whereas spectral and fragmentation-tree alignments use only spectra.
5 Conclusion
The paper presents a maximum a posteriori estimator for fragmentation-tree computation and reports better performance than previous approaches. In CASMI 2013, the method identified 8 of 12 molecular formulas correctly, with 3 additional formulas placed in the TOP2; isotope-pattern information raised this to 10 of 12.
- Conclusion: The maximum a posteriori estimator performs significantly better than previous approaches for computing fragmentation trees.The paper notes that isotope-pattern information can further improve identification, but it was unavailable for the two evaluated datasets.
- Conclusion: 8 of 12 molecular formulas were correctly identified in CASMI 2013, with 3 additional formulas placed in the TOP2.Combining fragmentation-tree analysis with isotope-pattern analysis identified 10 of 12 formulas.
- Conclusion: The new scoring also supported molecular-fingerprint prediction from tandem MS data, producing significantly improved prediction performance for structure-database searching.The improved fingerprints were built using fragmentation trees computed with the new scoring.