Peak Mass Intensity Feature Encoding

peak-mass-intensity-feature-encoding

Summary

Converts raw mass spectra into discrete token pairs (peak-mass and peak-intensity tokens) to enable downstream neural embedding and similarity matching. This tokenization step standardizes spectral representations for Word2vec-based embedding in high-throughput spectrum matching workflows.

When to use

Apply this skill when you have raw mass spectra (e.g., from NIST 2017 or MassBank) and need to prepare them for Word2vec embedding or other token-based neural models. Use it as a preprocessing step before training spectral embedding models, especially when scale and accuracy of spectrum matching are priorities.

When NOT to use

• Spectra are already represented as fixed-length vectors or dense feature matrices; tokenization would be redundant.
• Your downstream model does not benefit from sequence-based representations (e.g., if you are using direct kernel methods or simple dot-product similarity).
• Peak-level noise or baseline subtraction has not been performed; raw, uncleaned spectra may produce uninformative tokens.

Inputs

• Raw mass spectra (e.g., from NIST 2017 or MassBank experimental libraries)
• Peak m/z values and intensity measurements per spectrum

Outputs

• Tokenized spectra: sequences of peak-mass and peak-intensity token pairs
• Token vocabulary: discrete bins for mass and intensity ranges

How to apply

Extract peak mass-to-charge ratios (m/z) and their corresponding intensities from each mass spectrum in your reference library. Tokenize each peak as a pair of discrete tokens—one representing the peak-mass bin and one representing the peak-intensity magnitude. Aggregate all tokenized peaks from a single spectrum into a sequence (analogous to a sentence in NLP) that preserves the spectral structure. These sequences become the input training data for Word2vec models trained with gensim, configured to learn fixed-length embedding vectors that capture spectral relationships. Validation follows by checking that cosine similarity between known similar spectra exceeds a threshold of 0.7.

Related tools

• gensim (Word2vec model training on tokenized spectra sequences to learn fixed-length embedding vectors)
• Python 3.7 (Runtime environment for tokenization, token sequence assembly, and model training)

Evaluation signals

• Token vocabulary size and frequency distribution are reasonable (no extreme sparsity or collapse to a few dominant tokens).
• Tokenized sequences maintain spectrum-level uniqueness: known similar spectra produce similar token sequences.
• Post-embedding cosine similarity between known spectral pairs exceeds the 0.7 threshold, confirming that embeddings preserve spectral relationship structure.
• No information loss in the token representation: round-trip conversion (spectrum → tokens → embedding) preserves rank order of similarity judgments.

Limitations

• Tokenization introduces quantization loss: fine-grained mass and intensity information is binned into discrete tokens, potentially losing precision in boundary regions.
• Token vocabulary design (bin size, granularity) is not detailed in the source material; optimal tokenization parameters may be task- and library-dependent.
• FastEI installation currently supports Windows 64-bit only (Windows 7, Windows 10), limiting portability of the full system.
• The article does not provide explicit guidance on handling spectra with missing or sparse peaks, or on standardizing peak intensity normalization before tokenization.

Evidence

• [other] Prepare mass spectra data by tokenizing peaks into peak-mass and peak-intensity tokens from NIST 2017 and MassBank experimental libraries.: "Prepare mass spectra data by tokenizing peaks into peak-mass and peak-intensity tokens from NIST 2017 and MassBank experimental libraries."
• [other] Train a Word2vec model using gensim with the tokenized spectra as input sentences, configured to learn fixed-length embedding vectors that capture spectral relationships.: "Train a Word2vec model using gensim with the tokenized spectra as input sentences, configured to learn fixed-length embedding vectors that capture spectral relationships."
• [other] Validate that embeddings preserve spectral similarity by checking cosine similarity correlations between known similar spectra exceed a threshold of 0.7.: "Validate that embeddings preserve spectral similarity by checking cosine similarity correlations between known similar spectra exceed a threshold of 0.7."
• [other] FastEI implements Word2vec-based spectrum embedding as a mechanism to improve accuracy in spectrum matching by converting spectra into embedded vector representations using gensim.: "FastEI implements Word2vec-based spectrum embedding as a mechanism to improve accuracy in spectrum matching by converting spectra into embedded vector representations using gensim."
• [readme] experimental libraries (e.g., NIST 2017 and MassBank libraries): "experimental libraries (e.g., NIST 2017 and MassBank libraries)"