Spectrum Vector Similarity Searching

Spectrum Vector Similarity Searching

Summary

Use nearest neighbor indexes to efficiently retrieve candidate similar spectra for a query spectrum without exhaustive all-versus-all comparison. This skill enables fast similarity searching on low-dimensional hashed spectrum vectors, drastically reducing the computational cost of building sparse pairwise distance matrices for large-scale MS/MS clustering.

When to use

Apply this skill when you have millions of MS/MS spectra represented as low-dimensional vectors (via feature hashing) and need to compute pairwise distances only between similar spectra rather than comparing every spectrum to every other spectrum. Typical trigger: clustering scale >100k spectra where exhaustive O(n²) comparison is prohibitively expensive, or when downstream analysis (e.g., density-based clustering) requires a sparse similarity graph.

When NOT to use

• Input spectra have not yet been converted to low-dimensional vectors; use feature hashing or another vectorization step first.
• Nearest neighbor indexes have not been pre-constructed; you must build and partition the index before querying.
• Analysis requires exhaustive all-versus-all comparison (e.g., building a complete distance matrix); sparse searching will undercount true neighbors and bias downstream clustering.

Inputs

• Nearest neighbor indexes (Voronoi-partitioned inverted index structure, indexed by precursor m/z bucket)
• Low-dimensional spectrum vectors (feature-hashed representations)
• Query spectrum vector(s) with known precursor m/z

Outputs

• Sparse pairwise distance matrix (query spectrum ID vs. candidate neighbor IDs and their cosine distances)
• List of retrieved nearest neighbors per query spectrum (with similarity scores)

How to apply

First, load pre-constructed nearest neighbor indexes that partition spectrum vectors into buckets by precursor m/z and assign each vector to Voronoi cells via an inverted index structure. For each query spectrum vector, probe the nearest neighbor index by inspecting a configurable number of neighboring cells (controlled by n_probe parameter) to retrieve the k nearest neighbors within a user-defined distance threshold (e.g., cosine distance < 0.10–0.15 for peptide clustering). Compute the actual cosine distances only between the query and its retrieved candidates, not against all spectra. Tune the trade-off between speed and recall by adjusting n_probe (fewer lists = faster but risk missing true neighbors) and n_neighbors_ann (number of candidate neighbors to evaluate). The result is a sparse pairwise distance matrix suitable for input to DBSCAN or other density-based clustering algorithms.

Related tools

• falcon (Command-line tool that implements nearest neighbor indexing, querying, and sparse distance matrix construction for spectrum clustering) — https://github.com/bittremieux/falcon
• spectrum-utils (Utility library for spectrum I/O and preprocessing (version 0.3.5 or compatible) used alongside falcon)

Examples

falcon peak/*.mzml --export_representatives --precursor_tol 20 ppm --fragment_tol 0.05 --eps 0.10 --n_probe 10 --n_neighbors 20

Evaluation signals

• Sparse distance matrix contains only non-zero distances between query spectra and their retrieved neighbors; all other entries are absent or zero, reducing memory footprint by orders of magnitude compared to dense matrices.
• Neighbor retrieval is consistent with precursor m/z bucketing: queries return candidates whose precursor m/z differs by ≤ the specified tolerance (e.g., 20 ppm).
• Recall metric: percentage of true neighbors (by cosine similarity threshold) actually retrieved should be ≥ 95–99% depending on n_probe and n_neighbors_ann settings; lower recall indicates insufficient probe depth.
• Downstream clustering purity: spectra clustered together should have cosine distance ≤ eps threshold (e.g., 0.10–0.15) and represent the same peptide; clusters should not contain mixed peptides (evaluate against reference annotations if available).
• Runtime scales linearly or near-linearly with number of spectra, not quadratically; if wall-clock time grows as O(n²), the sparse index is not being used correctly.

Limitations

• Accuracy and speed are governed by n_probe: inspecting fewer lists reduces computation time but increases risk of missing true neighbors in high-dimensional space, leading to fragmentation of clusters.
• Nearest neighbor index construction itself requires upfront cost; for small datasets (<100k spectra) exhaustive comparison may be faster than indexing overhead.
• Performance depends critically on the quality of low-dimensional vector representation (via feature hashing); poor hashing can cause high-dimensional neighbors to collide spuriously, degrading clustering recall.
• The sparse matrix construction is most effective when spectrum similarity is truly sparse (i.e., most spectra are dissimilar); highly uniform datasets may not benefit from the sparse approach.

Evidence

• [intro] Nearest neighbor indexes enable efficient computation without exhaustive comparison: "The nearest neighbor indexes are used to efficiently compute a sparse pairwise distance matrix without having to exhaustively compare all spectra to each other."
• [other] Workflow: query indexes, retrieve candidates, compute distances only for neighbors: "Query each spectrum vector against the nearest neighbor indexes to retrieve candidate similar spectra within a user-defined distance threshold. Compute pairwise distances only between each spectrum"
• [readme] Vectors partitioned by Voronoi diagram and inverted index for fast searching: "The spectrum vectors in each bucket are partitioned into data subspaces to create a Voronoi diagram, and all vectors are assigned to their nearest representative vector in an inverted index."
• [readme] n_probe parameter controls speed vs accuracy trade-off: "The accuracy and speed of similarity searching is governed by the number of neighboring cells to explore during searching: exploring more cells during searching decreases the chance of missing a"
• [readme] Tunable parameters for nearest neighbor indexing: "n_probe: The maximum number of lists in the inverted index to inspect during querying. Inspecting fewer lists will run faster but will give slightly less accurate clustering results. n_neighbors and"