Source-linked AI summary
Competitive Fragmentation Modeling of ESI-MS/MS spectra for putative metabolite identification
F. Allen, R. Greiner, D. Wishart
TL;DR
Interpreting ESI-MS/MS spectra for metabolite identification is difficult, and reference databases cover only a limited portion of possible compounds. The paper introduces Competitive Fragmentation Modeling, a learned generative model of CID fragmentation, and evaluates it for spectrum prediction and putative identification. CFM outperforms full peak enumeration and existing identification methods across the reported datasets, while its candidate-list setting does not support de novo identification of unknown molecules.
Problem
Limited ESI-MS/MS reference databases make it difficult to identify metabolites for which no reference spectra are available.
Method
Competitive Fragmentation Modeling learns a generative model of CID fragmentation from MS/MS data, including a combined-energy extension for spectra measured at multiple energies.
Results
CFM significantly outperforms full enumeration for spectrum prediction and existing methods MetFrag and FingerID for putative metabolite identification across the reported datasets and conditions.
Takeaways & Limitations
CFM provides a structure-based route to predicting spectra and ranking candidate metabolites when reference spectra are unavailable.
Takeaways & Limitations
The predictive approach requires a candidate-molecule list and therefore cannot identify molecules absent from that list; combined-energy CFM also showed minimal difference from single-energy CFM.
Abstract
from arXiv · showhide
Electrospray tandem mass spectrometry (ESI-MS/MS) is commonly used in high throughput metabolomics. One of the key obstacles to the effective use of this technology is the difficulty in interpreting measured spectra to accurately and efficiently identify metabolites. Traditional methods for automated metabolite identification compare the target MS or MS/MS spectrum to the spectra in a reference database, ranking candidates based on the closeness of the match. However the limited coverage of available databases has led to an interest in computational methods for predicting reference MS/MS spectra from chemical structures. This work proposes a probabilistic generative model for the MS/MS fragmentation process, which we call Competitive Fragmentation Modeling (CFM), and a machine learning approach for learning parameters for this model from MS/MS data. We show that CFM can be used in both a MS/MS spectrum prediction task (ie, predicting the mass spectrum from a chemical structure), and in a putative metabolite identification task (ranking possible structures for a target MS/MS spectrum). In the MS/MS spectrum prediction task, CFM shows significantly improved performance when compared to a full enumeration of all peaks corresponding to substructures of the molecule. In the metabolite identification task, CFM obtains substantially better rankings for the correct candidate than existing methods (MetFrag and FingerID) on tripeptide and metabolite data, when querying PubChem or KEGG for candidate structures of similar mass.
1 Introduction
ESI-MS/MS is widely used in metabolomics, but interpreting spectra for metabolite identification remains difficult because reference databases have limited coverage. The paper reviews existing approaches and proposes learning a generative fragmentation model to improve the precision of predicted spectra.
- ESI-MS/MS uses collision-induced dissociation to fragment molecules, but interpreting the resulting spectra to identify metabolites is a major bottleneck.
- Reference-spectrum methods are constrained by limited ESI-MS/MS database coverage relative to the number of known chemical structures and metabolites.HMDB contained ESI-MS/MS data for around 800 compounds, while MassBank contained approximately 2,000 unique compounds against more than 19 million PubChem structures.
- Predicting spectra from candidate chemical structures can identify metabolites without complete reference spectra, but cannot identify molecules absent from the candidate list.
- Rule-based methods use manually curated fragmentation rules, whereas combinatorial methods enumerate possible fragments and compare their peaks with the measured spectrum.
- Combinatorial methods generally achieve high recall but poor precision because they predict many more peaks than are observed.
- The paper investigates a machine-learning approach that learns a generative CID fragmentation model, estimating event likelihoods to predict the peaks most likely to be observed.
2 Methods
CFM models ESI-MS/MS fragmentation as probabilistic transitions among charged fragments, learned from spectra and extended to combine multiple collision energies. Fragmentation graphs enumerate feasible charged products, while competing break probabilities and energy-specific models support tractable spectrum prediction.
- 2 Methods: CFM models ESI-MS/MS CID fragmentation as a stochastic Markov process of state transitions between charged fragments.SE-CFM uses latent fragment states, with the initial molecule and output peak observed; repeated runs produce a complete predicted spectrum.
- 2 Methods: The model assumes singly positively charged, common-isotope input molecules and predominantly two-way fragmentations without mass or charge loss.These assumptions reduce the branching factor and make fragment enumeration computationally feasible, although some do not always hold in real-world spectra.
- 2 Methods: Fragments are recursively enumerated to a maximum depth, producing a directed acyclic graph of possible charged fragments while recording neutral-loss information on edges.Neutral fragments are omitted because mass spectrometers cannot detect them, and their possible breaks are not recursively explored.
- 2 Methods: Transition probabilities are assigned only to feasible child fragments and self-transitions, with all outgoing probabilities from a fragment summing to one.The model avoids a separate parameter for every possible fragmentation by generalizing across similar molecular breaks.
- 2 Methods: A softmax converts break tendencies into competing probabilities, so the likelihood of one break depends on the other possible breaks from the same parent.For similar H2O losses, a competing NH3 break can lower the probability assigned to the corresponding fragmentation.
- 2 Methods: SE-CFM learns separate models for low, medium, and high collision energies, whereas CE-CFM combines multiple energy spectra with separate transition likelihoods for each energy block.CE-CFM retains the SE-CFM fragment states, transition rules, and observation model while relaxing the homogeneity assumption across energy blocks.
3 Experimental Results
CFM was evaluated for spectrum prediction and metabolite identification across tripeptide, Metlin, and MassBank data. It improved precision-oriented spectrum metrics and candidate rankings, while performance varied by collision energy and database.
- Spectrum Prediction: CFM methods predicted more than 75% of total tripeptide peak intensity and approximately 60% of total metabolite peak intensity.These weighted-recall results indicate stronger coverage of higher-intensity peaks than unweighted recall alone captures.
- Spectrum Prediction: CFM obtained several orders of magnitude better precision and Jaccard scores than full peak enumeration, with a corresponding loss of recall.The comparison held across tripeptide, Metlin metabolite, and MassBank metabolite data sets.
- Spectrum Prediction: Low-energy SE-CFM weighted recall was 78%, 73%, and 81% for tripeptide, Metlin, and MassBank data, versus 73%, 29%, and 37% at high energy.Results were averaged across energy levels in the main comparison, while these values separate low- and high-energy spectra.
- Spectrum Prediction: On tripeptides, CFM achieved higher recall and similar precision to a heuristic fragmentation model, producing improved Jaccard scores.The result demonstrates that the learned model extracted fragmentation patterns comparably to expert-designed mechanisms in this data set.
- Spectrum Prediction: For metabolite spectra, weighted recall remained above 60%, while CFM precision and Jaccard scores exceeded full enumeration.MassBank weighted recall and precision were fairly comparable to Metlin results, suggesting learned trends transferred across laboratories.
- Metabolite Identification: CFM ranked candidates substantially better than MetFrag and FingerID across all three data sets for both PubChem and KEGG queries.For metabolite identification, KEGG results exceeded 70% top-1 accuracy for both metabolite sets, while the correct candidate was almost always above rank 5 (>95%).
4 Conclusion
CFM models ESI-MS/MS fragmentation and learns its parameters from data, with evaluations indicating performance useful for metabolomics experimentalists.
- CFM models the ESI-MS/MS fragmentation process and provides a method for training model parameters from MS/MS data.
- Across multiple datasets and conditions, head-to-head comparisons found that CFM significantly outperformed existing state-of-the-art methods.
- The reported performance reached a level that could be useful to experimentalists performing metabolomics studies.