Spectral Library Matching With Cosine Similarity

Spectral Library Matching with Cosine Similarity

Summary

Cosine similarity scoring is used within a cascade search strategy to confidently match query mass spectra to unmodified peptides in a spectral library after approximate nearest neighbor filtering. This scoring method forms the first-pass identification layer before attempting open modification searches on spectra without confident matches.

When to use

Apply this skill when you have a query spectrum and a reduced candidate set of library spectra (via approximate nearest neighbor indexing), and need to search for exact unmodified peptide matches with high confidence before proceeding to variable-mass-shift modification searches. Use this as the first stage of a cascade search to maximize unmodified identifications while controlling false discovery rate.

When NOT to use

• Input is already a curated list of peptide identifications with validated FDR — re-scoring is redundant.
• Query spectra are from a different instrument class or mass range than the spectral library, causing systematic score distribution shifts.
• Modification-rich proteomes where the majority of spectra have post-translational modifications that violate the 'unmodified match' assumption in the first cascade pass.

Inputs

• Query mass spectrum (m/z and intensity pairs)
• Approximate nearest neighbor filtered candidate library spectra (m/z and intensity pairs)
• Library spectrum annotations (peptide sequences)

Outputs

• Ranked list of library matches with cosine similarity scores
• Assigned spectrum-level or peptide-level FDR values
• Binary classification: spectra with confident unmodified matches vs. spectra requiring modification search

How to apply

After retrieving a limited candidate set via approximate nearest neighbor search, apply cosine similarity scoring to compare each query spectrum against candidate library spectra. Rank candidates by cosine similarity score and apply spectrum-level or peptide-level false discovery rate filtering to assign confidence thresholds. Spectra with confident unmodified matches are identified and removed from downstream processing; only spectra without confident matches proceed to open modification search using the shifted dot product score. This two-stage approach maximizes sensitivity for unmodified identifications while maintaining strict FDR control across the full pipeline.

Related tools

• ANN-SoLo (Spectral library search engine implementing cascade search with cosine similarity scoring for unmodified peptide identification, followed by shifted dot product scoring for open modification search) — https://github.com/bittremieux/ANN-SoLo

Evaluation signals

• Cosine similarity scores follow a bimodal distribution with clear separation between true matches and false matches, enabling FDR filtering at a meaningful threshold.
• Spectrum-level or peptide-level FDR values assigned to all identifications are ≤ the specified control threshold (e.g., ≤ 1% or ≤ 5%).
• Spectra identified via cosine similarity in the cascade pass show minimal overlap with spectra subsequently identified via shifted dot product scoring in the modification search pass, confirming non-redundant cascade stages.
• Total number of identified spectra (unmodified + modified) exceeds what would be achieved by either method alone, validating the cascade strategy's sensitivity gain.
• When the same query spectrum is searched against alternative libraries with different composition, rank order of top candidates by cosine similarity remains consistent for true matches.

Limitations

• Cosine similarity scoring assumes query and library spectra are comparable in instrument resolution, ionization method, and dynamic range; cross-instrument searches may require score recalibration.
• The method is designed for spectral library matching and does not perform de novo sequencing; it requires a curated library and does not identify novel peptides.
• Python 3.6 to 3.9 are required; Python 3.10 and newer are not currently supported by the ANN-SoLo implementation.
• FDR control is performed post-hoc via statistical filtering; the confidence threshold depends on library quality and may vary across datasets.

Evidence

• [other] first search candidates for exact unmodified peptide matches using cosine similarity scoring: "first search candidates for exact unmodified peptide matches using cosine similarity scoring"
• [readme] cascade search strategy to maximize the number of identified unmodified and modified spectra while strictly controlling the false discovery rate: "This is combined with a cascade search strategy to maximize the number of identified unmodified and modified spectra while strictly controlling the false discovery rate"
• [other] Rank all identified matches (unmodified and modified) by score and apply false discovery rate control via spectrum-level or peptide-level FDR filtering: "Rank all identified matches (unmodified and modified) by score and apply false discovery rate control via spectrum-level or peptide-level FDR filtering"
• [other] For spectra without confident unmodified matches, perform open modification search on the same candidates, allowing variable mass shifts: "For spectra without confident unmodified matches, perform open modification search on the same candidates, allowing variable mass shifts"