AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

matchms

Processes, cleans, compares, and searches tandem mass spectra with matchms. Use for MS/MS file I/O, metadata harmonization, peak filtering, spectral similarity, library matching, score matrices, and molecular-similarity networks. Use pyopenms instead for LC-MS feature detection or proteomics pipelines.

Original source
K-Dense-AI
License
MIT
Pinned revision
5f2bc0341354fd5590aef76a533ada1e1e5b4e62

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read k-dense-ai--matchms at revision 5f2bc0341354fd5590aef76a533ada1e1e5b4e62, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 7

references/similarity.md

Similarity and Scores (matchms 0.33.1)

Use this reference to select a similarity class, interpret its output, extract top hits, or scale a comparison. The authoritative API is matchms.similarity.

Core Calculation

from matchms import calculate_scores
from matchms.similarity import CosineGreedy

metric = CosineGreedy(tolerance=0.02)
scores = calculate_scores(
    references=reference_spectra,
    queries=query_spectra,
    similarity_function=metric,
    array_type="numpy",
    is_symmetric=False,
)

Set is_symmetric=True only when references and queries are the same collection in the same order and the metric is symmetric.

Understand the Output Before Indexing

Scores.scores is a sparsestack.StackedSparseArray, not a plain two-dimensional NumPy array. Its shape is (n_references, n_queries, n_score_fields).

Cosine-family methods produce two fields:

print(scores.score_names)
# ('CosineGreedy_score', 'CosineGreedy_matches')

similarities = scores.to_array("CosineGreedy_score")
matched_peaks = scores.to_array("CosineGreedy_matches")

Scalar methods such as FlashSimilarity, PrecursorMzMatch, and FingerprintSimilarity produce one field named after the class.

Top hits for one query

score_name = "CosineGreedy_score"
matches_name = "CosineGreedy_matches"

ranked = scores.scores_by_query(query_spectrum, name=score_name, sort=True)
for reference, value in ranked[:10]:
    print(
        reference.get("spectrum_id"),
        float(value[score_name]),
        int(value[matches_name]),
    )

The first tuple item is the actual reference Spectrum, not an integer index. The second item can be a structured NumPy record containing every score field.

Iterate stored pairs

for reference, query, values in scores:
    print(reference.get("id"), query.get("id"), values)

Iteration covers stored coordinates. After sparse filtering, that may be only a subset of the Cartesian product.

Peak-Based Cosine Methods

All cosine classes below use:

tolerance=0.1, mz_power=0.0, intensity_power=1.0

unless noted otherwise. Tolerance is in daltons for these classes.

CosineGreedy

Greedily assigns candidate peak pairs within tolerance.

Use for:

  • routine spectral-library comparisons;
  • a transparent baseline;
  • moderate collections where exact assignment is not essential.

Output fields: CosineGreedy_score, CosineGreedy_matches.

CosineHungarian

Uses optimal assignment rather than greedy assignment.

Use for:

  • benchmarking peak-assignment effects;
  • smaller datasets;
  • cases where greedy ambiguity materially affects results.

It is computationally more expensive than CosineGreedy.

CosineLinear

Added in 0.33.0. It first greedily merges nearby peaks using a 2 * tolerance separation criterion, then performs linear-time matching on the merged spectra. The score and matched-peak count therefore need not equal CosineGreedy on the original peaks, even with identical tolerance and intensity powers.

Use when:

  • cosine is the intended metric;
  • matrix size makes assignment cost important; and
  • you have benchmarked agreement and runtime on representative spectra.

Output fields: CosineLinear_score, CosineLinear_matches.

Modified Cosine

Modified cosine allows unshifted peak matches and matches shifted by the difference in precursor m/z. Both spectra need valid precursor_mz.

ModifiedCosineGreedy

from matchms.similarity import ModifiedCosineGreedy

metric = ModifiedCosineGreedy(tolerance=0.02)

This is the current name for the implementation formerly called ModifiedCosine. It uses greedy assignment.

ModifiedCosineHungarian

from matchms.similarity import ModifiedCosineHungarian

metric = ModifiedCosineHungarian(tolerance=0.02)

Use for exact modified-cosine assignment in benchmarks or method development. It is slower than the greedy variant.

Do not infer that a shifted match proves a specific chemical transformation. Inspect precursor delta, adduct/charge compatibility, shifted peaks, and orthogonal annotations.

Neutral-Loss Cosine

from matchms.similarity import NeutralLossesCosine

metric = NeutralLossesCosine(
    tolerance=0.02,
    ignore_peaks_above_precursor=True,
)

NeutralLossesCosine computes losses from precursor_mz - fragment_mz. Both spectra require precursor m/z. Do not call the removed add_losses() filter; losses are computed on demand in current matchms.

Output fields: NeutralLossesCosine_score, NeutralLossesCosine_matches.

Fast Matrix-Oriented Methods

BlinkCosine

BLINK-style approximate cosine:

from matchms.similarity import BlinkCosine

metric = BlinkCosine(
    tolerance=0.01,
    bin_width=0.001,
    min_relative_intensity=0.01,
    top_k=None,
    batch_size=1024,
    sparse_score_min=0.0,
)

Important parameters include peak preprocessing, precursor cropping, batch size, and sparse score minimum. By default crop_above_precursor=True crops where a precursor is available; record precursor coverage or disable this option when that extra step is inappropriate. Output contains a float32 score and matched-peak count. Validate approximation behavior against CosineGreedy on a subset before changing production workflows.

FlashSimilarity

Fast matrix scoring based on the Flash Entropy approach:

from matchms.similarity import FlashSimilarity

entropy = FlashSimilarity(
    score_type="spectral_entropy",
    matching_mode="fragment",
    tolerance=0.02,
)

fast_modified_cosine = FlashSimilarity(
    score_type="cosine",
    matching_mode="hybrid",
    tolerance=0.02,
)

Current choices and behavior:

  • score_type: spectral_entropy or cosine
  • matching_mode: fragment, neutral_loss, or hybrid
  • use_ppm=True switches fragment tolerance to symmetric ppm; default is Da
  • remove_precursor=False, noise_cutoff=0.01, merge_within=0, and normalize_to_half=True are defaults; entropy weighting is internal
  • neutral-loss/hybrid matching and an identity precursor gate need valid precursor metadata
  • identity_precursor_tolerance and identity_use_ppm control a separate precursor gate

pair() exists but emits a warning because it is not the optimized use. Call calculate_scores() for multi-spectrum matrices. A 1-by-1 Scores.calculate call still uses pair(). Flash output is a scalar field named FlashSimilarity, not a score/matches pair. Hybrid matching gives unshifted fragment matches priority; do not assume equality with greedy or Hungarian modified cosine.

In 0.33.1, matrix() allocates a dense output even for array_type="sparse", and is_symmetric=True does not reduce Flash runtime. Its default worker count uses available CPUs on Unix; call metric.matrix(..., n_jobs=1) for an explicit serial calculation (this returns a raw array, not Scores). Do not assume a sparse output is a bounded-memory or bounded-CPU search.

Verified 0.33.1 defect: score_type="spectral_entropy" with matching_mode="neutral_loss" adds fragment and loss contributions, producing a self-score near 2 on the tested three-peak spectrum. Do not use that combination as a normalized entropy score; choose fragment entropy or NeutralLossesCosine as appropriate. Do not silently clip or rescale it.

BinnedEmbeddingSimilarity

from matchms.similarity import BinnedEmbeddingSimilarity

metric = BinnedEmbeddingSimilarity(
    similarity="cosine",
    max_mz=1005,
    bin_width=1.0,
    intensity_power=1.0,
)

This creates fixed-width binned spectral embeddings. Bin width controls both resolution and dimensionality. Reject empty/all-zero embeddings, including spectra with no retained peak below max_mz, before scoring. The class also supports approximate-neighbor indexing through the current PyNNDescent backend; use it only after checking recall against exact neighbors. In 0.33.1 the matrix method accepts only dense output and is_symmetric=True; use it for all-vs-all input. A normal rectangular calculate_scores call with its default is_symmetric=False raises an error. For rectangular embedding comparisons, compute the embeddings explicitly and compare those with the intended vector metric.

Candidate and Metadata Matches

These methods are useful as gates or additional evidence. They are not substitutes for peak-pattern similarity.

PrecursorMzMatch

from matchms.similarity import PrecursorMzMatch

absolute = PrecursorMzMatch(tolerance=0.02, tolerance_type="Dalton")
relative = PrecursorMzMatch(tolerance=10, tolerance_type="ppm")

Output field: PrecursorMzMatch (boolean).

ParentMassMatch

from matchms.similarity import ParentMassMatch

metric = ParentMassMatch(tolerance=0.02)

Requires parent_mass. Output field: ParentMassMatch (boolean). The current constructor does not expose a ppm mode.

MetadataMatch

from matchms.similarity import MetadataMatch

same_mode = MetadataMatch(field="ionmode", matching_type="equal_match")
near_rt = MetadataMatch(
    field="retention_time",
    matching_type="difference",
    tolerance=0.2,
)

Current matching types are equal_match and difference. Older examples that use matching_type="exact" are invalid.

IntersectMz

IntersectMz(scaling=1.0) is a simple m/z-intersection score useful in tests or specialized workflows. It is not a drop-in replacement for tolerance-aware, intensity-weighted spectral scoring.

Molecular Fingerprint Similarity

FingerprintSimilarity compares molecular fingerprints derived from known structures. It therefore measures structure similarity, not spectral similarity.

from matchms import Fingerprints, calculate_scores
from matchms.similarity import FingerprintSimilarity

fp_store = Fingerprints(
    fingerprint_algorithm="morgan2",
    fingerprint_method="bit",
    nbits=2048,
)
fp_store.compute_fingerprints(spectra)

usable = []
for spectrum in spectra:
    fingerprint = fp_store.get_fingerprint_by_spectrum(spectrum)
    if fingerprint is not None:
        spectrum.set("fingerprint", fingerprint)
        usable.append(spectrum)

scores = calculate_scores(
    usable,
    usable,
    FingerprintSimilarity(similarity_measure="jaccard"),
    is_symmetric=True,
)

Similarity measures are jaccard, dice, and cosine.

The old add_fingerprint() filter still satisfies the 0.33.1 FingerprintSimilarity interface, but it is marked for removal in matchms 1.0. The Fingerprints bridge above avoids calling that deprecated filter while remaining compatible with the current similarity class.

Efficient Precursor-Gated Search

Filter score coordinates before calculating an expensive spectral metric:

from matchms import calculate_scores
from matchms.similarity import ModifiedCosineGreedy, PrecursorMzMatch

scores = calculate_scores(
    references,
    queries,
    PrecursorMzMatch(tolerance=10, tolerance_type="ppm"),
    array_type="sparse",
)
scores.filter_by_range(name="PrecursorMzMatch", low=0.5)
if len(scores.scores.row) == 0:
    raise ValueError("No candidates passed the precursor gate")
scores.calculate(
    ModifiedCosineGreedy(tolerance=0.02),
    array_type="sparse",
    join_type="left",
)

In 0.33.1, Scores.calculate selects its coordinate-only sparse_array path only with an existing score layer, join_type="left" or "inner", and fewer than half of all coordinates retained. Otherwise it computes the full matrix then joins; setting array_type="sparse" alone does not change this. A 1-by-1 input always uses pair(), even after a rejecting gate. Stop explicitly when no candidate coordinates remain. Confirm score names and coordinate counts after each stage. An overly narrow precursor gate removes analogs or different adducts; this example is for identity-oriented candidate search.

The Pipeline class formalizes the same pattern and can persist a YAML workflow. See workflows.md.

Filtering and Exporting Scores

print(scores.score_names)

scores.filter_by_range(
    name="ModifiedCosineGreedy_score",
    low=0.6,
    above_operator=">=",
)

dense = scores.to_array("ModifiedCosineGreedy_score")
coo = scores.to_coo("ModifiedCosineGreedy_score")
scores.to_json("scores.json")

filter_by_range() mutates the stored score coordinates. Preserve an unfiltered copy or serialize first when alternative thresholds must be compared.

Dense arrays require memory proportional to n_references * n_queries. A sparse container is most useful only when a mask or score threshold leaves relatively few stored pairs.

Combining Metrics

Do not directly index scores.scores[j, i] and assume a scalar. Extract named layers:

cosine = scores.to_array("CosineGreedy_score")
matches = scores.to_array("CosineGreedy_matches")

If combining metrics:

  • define whether missing/filtered pairs are zero, missing, or excluded;
  • normalize only when score semantics justify it;
  • fit or justify weights on an appropriate validation set;
  • avoid leaking query identities into tuning;
  • retain each component score in the output; and
  • report the exact formula and package version.

Interpretation Checklist

  • Was identical preprocessing applied to references and queries?
  • Is the tolerance appropriate for the instrument and calibration?
  • Are precursor m/z, charge, ion mode, and adduct metadata valid?
  • How many peaks matched, and what fraction of each spectrum do they represent?
  • Is the hit driven by one dominant/common fragment?
  • Are collision energy and acquisition conditions comparable?
  • Is the query actually present in the searched library?
  • Was the threshold validated on representative positives and negatives?
  • Has the top hit been inspected with a mirror plot?

A high score ranks a candidate under a chosen metric. It does not, by itself, establish compound identity.