Mass Fragment Neutral Loss Representation

mass-fragment-neutral-loss-representation

Summary

Convert tandem mass spectrometry (MS/MS) spectra into a bag-of-fragments representation with extracted neutral losses and noise filtering, producing a structured document-term matrix suitable for probabilistic topic modeling. This preprocessing step transforms raw spectral peaks into interpretable fragmentation vocabularies that enable discovery of recurring substructures.

When to use

When you have raw or minimally processed MS/MS spectra (in positive or negative ion mode) and aim to infer recurring fragmentation patterns (Mass2Motifs) using topic modeling. Apply this skill before LDA to convert peaks into a standardized bag-of-fragments format with neutral losses explicitly represented and noise removed.

When NOT to use

• When input spectra are already library-matched compounds with known structures; MS2LDA is designed for unsupervised discovery, not confirmation of known compounds.
• When spectra have very low signal-to-noise ratio or minimal fragmentation (e.g., single dominant peak); bag-of-fragments will be too sparse to infer meaningful motifs.
• When you require quantitative abundance estimation rather than presence/absence or relative intensity patterns; bag-of-fragments treats fragments as discrete vocabulary terms, losing fine-grained intensity dynamics.

Inputs

• raw MS/MS spectra (mzML, mzXML, or similar format)
• precursor m/z and mass values
• spectrum intensity profiles
• ion mode specification (positive/negative)

Outputs

• bag-of-fragments document-term matrix
• neutral-loss vocabulary
• noise-filtered spectral corpus
• fragment-loss probability distributions (input to LDA)
• JSON-serialized spectral representation

How to apply

Load preprocessed spectra and convert each spectrum into a bag-of-fragments representation by extracting individual m/z peaks and computing neutral losses (differences between precursor mass and fragment peaks). Filter out low-intensity noise peaks according to signal-to-noise thresholds and ion-mode-specific cutoffs. Construct a document-term matrix where each spectrum is a document and fragments/losses are terms with their intensities or occurrence counts as weights. This bag-of-fragments corpus is then passed to LDA, which learns which fragments and losses co-occur across spectra to infer motif topics. The representation succeeds when fragments and losses are discretized into a vocabulary, noise is suppressed below a usable threshold, and the resulting matrix is sparse and suitable for convergence under LDA iteration.

Related tools

• MS2LDA (orchestrates the full workflow including preprocessing, LDA modeling, and annotation of bag-of-fragments spectra) — https://github.com/vdhooftcompmet/MS2LDA
• Latent Dirichlet Allocation (LDA) (probabilistic topic model applied to the bag-of-fragments matrix to infer Mass2Motifs)
• Python (runtime environment for loading, filtering, and serializing spectral corpus)

Examples

from ms2lda.preprocessing import prepare_spectra; corpus = prepare_spectra('spectra.mzML', ion_mode='positive', noise_threshold=0.01, min_fragment_mz=50); print(f'Corpus size: {len(corpus)} spectra, {len(corpus.vocabulary)} fragments/losses')

Evaluation signals

• Bag-of-fragments matrix is sparse, non-empty, and compatible with LDA input schema (document × term with non-negative weights).
• Neutral loss vocabulary is complete: all losses are >= 1 m/z and <= precursor m/z; no impossible or duplicate loss entries.
• Noise filtering reduces the number of low-intensity peaks by >50% while retaining significant fragment peaks (e.g., base peak and top N peaks per spectrum remain).
• Serialized JSON output contains valid JSON with consistent schema: each spectrum has 'fragments', 'losses', 'precursor_mz', and 'intensity' fields populated.
• LDA convergence monitoring shows decreasing perplexity or increasing likelihood over iterations; if convergence is flat or diverges, the fragment representation may be malformed or the corpus too sparse.

Limitations

• Bag-of-fragments discards peak order and intensity fine structure; temporal or sequential fragmentation information is lost.
• Neutral loss computation depends on accurate precursor mass; errors in precursor assignment propagate into the loss vocabulary.
• Noise filtering thresholds are dataset- and ion-mode-dependent; aggressive filtering may discard rare but informative low-abundance losses.
• Only a fraction of available mass spectrometry information is traditionally utilized; this representation focuses on fragment and loss patterns but does not encode charge states, isotope patterns, or other spectral features.
• The bag-of-fragments approach assumes fragments and losses are exchangeable within a spectrum; it cannot encode positional or hierarchical fragmentation pathways.

Evidence

• [methods] Convert MS/MS spectra into a bag-of-fragments format, extract neutral losses, and filter out noise: "Convert MS/MS spectra into a bag-of-fragments format and extract neutral losses, filter out noise"
• [other] Load the preprocessed spectral corpus (bag-of-fragments representation with neutral losses extracted and noise filtered) into memory using Python: "Load the preprocessed spectral corpus (bag-of-fragments representation with neutral losses extracted and noise filtered) into memory using Python"
• [other] MS2LDA applies Latent Dirichlet Allocation to infer which motifs are most likely to explain the observed fragmentation patterns in mass spectrometry data: "MS2LDA applies Latent Dirichlet Allocation to infer which motifs are most likely to explain the observed fragmentation patterns in mass spectrometry data"
• [readme] Preprocessing → filter & clean your spectra (positive/negative ion mode): "Preprocessing → filter & clean your spectra (positive/negative ion mode)"
• [readme] MS2LDA identifies recurring substructures (motifs) across spectral datasets without relying on prior compound identification: "MS2LDA identifies recurring substructures (motifs) across spectral datasets without relying on prior compound identification"