Word embedding based spectrum representation

word-embedding-based-spectrum-representation

Summary

Convert mass spectra into document-like representations where peaks and neutral losses are encoded as learned word embeddings, enabling structural similarity inference through natural language processing models such as Word2Vec. This approach yields superior true-positive-to-false-positive ratios compared to cosine-based spectral similarity during library matching.

When to use

When comparing large numbers of MS/MS spectra against spectral libraries or in molecular networking, particularly when molecules differ by multiple structural modifications and cosine-based scores produce excessive false positives. Apply when you have access to a large, chemically diverse training corpus (e.g., ≥76,000 spectra) to train or retrain Word2Vec embeddings on fragment ion and neutral loss vocabularies specific to your experimental context (e.g., positive or negative ionization mode).

When NOT to use

• If input spectra are from GC-MS; Spec2Vec relies on neutral losses which are usually not measured in GC-MS data.
• If you have only a small dataset (<10,000 spectra) with limited chemical diversity and no access to a large pretrained model; Word2Vec requires substantial training data to learn robust embeddings.
• If spectra contain fragments and losses far outside those represented in the training set and you cannot or will not retrain the model; unknown peaks will inflate the missing-fraction threshold and reduce reliability.

Inputs

• MS/MS spectra in standard formats (mzML, mzXML, msp, MGF, JSON)
• Spectrum metadata (precursor m/z, parent mass, ionization mode)
• InChIKey annotations or other chemical structure ground truth (for evaluation)
• Trained Word2Vec model or large training corpus of spectra for Word2Vec fitting

Outputs

• Spectrum documents (text files with peak and neutral loss tokens)
• Trained Word2Vec embedding model (gensim format)
• Spec2Vec similarity score matrix (sparse or dense format)
• Receiver-operator-characteristic curves and true-positive/false-positive rates
• Library matches ranked by Spec2Vec similarity with confidence scores

How to apply

Convert each processed spectrum into a text document by representing each peak as '[redacted-email]' (m/z to defined decimal precision) and adding neutral losses as '[redacted-email]' (calculated as precursor − peak for losses between 5.0 and 200.0 Da). Train a Word2Vec model (or use a pretrained model) on a large set of such documents to learn embeddings that capture co-occurrence relationships between fragments and losses across structurally similar compounds. Apply parent-mass-scaled peak filtering (max_peaks = 0.5 × parent mass) before embedding to reduce noise and computational cost. Compute Spec2Vec similarity scores by comparing the learned embeddings of query and library spectra; this yields scores that correlate more strongly with structural similarity than cosine or modified cosine. Use a missing-fraction threshold (e.g., <0.05) to flag spectra with unknown peaks outside the learned vocabulary. Evaluate performance using receiver-operator-characteristic curves and true-positive/false-positive rates on a held-out test set with known InChIKey ground truth (first 14 characters / planar structure).

Related tools

• Spec2Vec (Compute spectral similarity scores based on learned Word2Vec embeddings of peaks and neutral losses) — https://github.com/iomega/spec2vec
• Word2Vec (gensim) (Train embeddings of fragment ions and neutral losses to capture fragmentation relationships)
• matchms (Import, filter, and process mass spectra; compute cosine and modified cosine similarity for comparison) — https://github.com/matchms/matchms
• Numpy, Scipy, Pandas (Numerical operations, similarity matrix computation, and data frame handling)
• Numba (JIT compilation of similarity score computations for performance optimization)

Examples

from spec2vec import Spec2Vec; from gensim.models import Word2Vec; model = Word2Vec.load('word2vec_model.model'); spec2vec = Spec2Vec(model); scores = spec2vec.pair_iterator(query_spectrum, library_spectra)

Evaluation signals

• Spec2Vec similarity scores should correlate more strongly with structural similarity (measured via InChIKey planar structure matches) than cosine or modified cosine scores, visible as higher R² in correlation plots.
• True-positive rate at a fixed false-positive rate threshold should exceed that of cosine/modified cosine; ROC curves should show Spec2Vec with larger area-under-curve (AUC ≥ 0.88 reported in Huber et al. 2021).
• Missing-fraction estimates (uncovered weighted part of spectra with unknown peaks) should remain below the set threshold (e.g., <0.05) for the majority of test spectra, indicating sufficient vocabulary overlap with the trained model.
• Rank-order of library hits should place true-positive matches (same planar InChIKey) higher than false positives when sorting by descending Spec2Vec similarity.
• Reproducibility check: similarity score matrices computed from the same spectra and model should be identical (bit-for-bit); pre-trained models from Zenodo should be loadable and produce consistent scores.

Limitations

• Spec2Vec requires training data and may underperform on spectral features (peaks or losses) not well-represented in the training set; retraining on new experimental conditions is often necessary.
• Not yet validated on GC-MS data, as neutral losses are typically not measured in GC-MS workflows.
• Model pre-training on a large dataset reduces but does not eliminate the need for retraining when applying to data with significantly different ionization modes or instrumental parameters.
• Unknown peaks not present in the trained Word2Vec vocabulary cannot be directly embedded; missing-fraction estimates provide only indirect uncertainty quantification.
• Computational cost of Word2Vec training scales with corpus size and vector dimensionality; model inference is faster than training but slower than simple cosine similarity for very large databases.

Evidence

• [methods] After processing, spectra are converted to documents. For this, every peak is represented by a word that contains its position up to a defined decimal precision ("[redacted-email]"): "After processing, spectra are converted to documents. For this, every peak is represented by a word that contains its position up to a defined decimal precision ('[redacted-email]')"
• [methods] In addition to all peaks of a spectrum, neutral losses between 5.0 and 200.0 Da were added as "[redacted-email]". Neutral losses are calculated as precursor − peak: "In addition to all peaks of a spectrum, neutral losses between 5.0 and 200.0 Da were added as '[redacted-email]'. Neutral losses are calculated as precursor − peak"
• [abstract] Spec2Vec scores correlate better with structural similarity than cosine-based scores: "Spec2Vec scores correlate better with structural similarity than cosine-based scores"
• [results] high Spec2Vec spectrum similarity score correlates stronger with structural similarity than the cosine or modified cosine scores: "high Spec2Vec spectrum similarity score correlates stronger with structural similarity than the cosine or modified cosine scores"
• [methods] the maximum number of kept peaks per spectrum was set to scale linearly with the estimated parent mass: max(n_peaks) = 0.5 × parentmass: "the maximum number of kept peaks per spectrum was set to scale linearly with the estimated parent mass: max(n_peaks) = 0.5 × parentmass"
• [methods] In those instances some words (= peaks) of a given spectra might be unknown to the model. In those instances we can estimate the impact of the missing words by assessing the uncovered weighted part: "In those instances some words (= peaks) of a given spectra might be unknown to the model. In those instances we can estimate the impact of the missing words by assessing the uncovered weighted part"
• [intro] Spec2Vec is a novel spectral similarity score based upon learnt embeddings of spectra. Instead inspired by the success of algorithms from the field of natural language: "Spec2Vec is a novel spectral similarity score based upon learnt embeddings of spectra"
• [discussion] For GC-MS, neutral losses are usually not measured.: "For GC-MS, neutral losses are usually not measured."
• [readme] Where Word2Vec learns relationships between words in sentences, spec2vec does so for mass fragments and neutral losses in MS/MS spectra.: "Where Word2Vec learns relationships between words in sentences, spec2vec does so for mass fragments and neutral losses in MS/MS spectra."
• [abstract] Spec2Vec is computationally more scalable allowing structural analogue searches in large databases within seconds: "Spec2Vec is computationally more scalable allowing structural analogue searches in large databases within seconds"