Metabolite annotation validation

metabolite-annotation-validation

Summary

Apply confidence score filtering to retain only high-quality metabolite annotations from spectral databases (SIRIUS, ISDB) and library matching, using tool-specific thresholds (e.g., ZodiacScore ≥0.9, cosine similarity ≥0.7) to remove low-confidence rows from annotation tables before downstream analysis.

When to use

After running in silico annotation tools (SIRIUS, ISDB) or spectral library matching on your feature table, when you need to retain only annotations meeting a minimum confidence threshold. Apply this skill before calculating annotation-dependent metrics (Feature Component, Class Component) or before exporting annotations for biological interpretation, to prevent false positives from inflating novel compound counts or misleading chemical class assignments.

When NOT to use

• Do not apply if annotation tables have already been pre-filtered by the source tool and you lack parameter metadata; re-filtering risks creating a compound filter that is opaque to downstream users.
• Do not use if your downstream analysis requires uncertain or low-confidence annotations as negative controls or for sensitivity analysis; filtering removes this ground truth.
• Do not apply if you are performing a method comparison study where consistent treatment of all annotation qualities (high and low confidence) is required.

Inputs

• SIRIUS compound_identification.tsv (containing Zodiac score, Cosmic score, and ConfidenceScore columns)
• ISDB/spectral library annotation table (containing cosine score, shared peaks count, ppm error, and final score columns)
• Configuration parameters: min_ZodiacScore, min_ConfidenceScore, min_score_final, cosine threshold, shared_peaks threshold, ppm_error tolerance

Outputs

• Filtered annotation table (TSV format, same schema as input, subset of rows passing all threshold criteria)
• Row count statistics (original vs. retained rows per annotation source)

How to apply

Load the annotation table (TSV format, e.g., compound_identification.tsv from SIRIUS or ISDB output) and configure tool-specific filters: for SIRIUS annotations apply min_ZodiacScore threshold (default 0.9) to retain only rows where Zodiac score ≥0.9; optionally layer min_ConfidenceScore (default 0.0). For ISDB/spectral library annotations, apply min_score_final (default 0.0), cosine score ≥0.7, shared_peaks ≥10, and ppm_error ≤5 ppm to filter low-quality matches. Apply filters sequentially in order (ppm → shared_peaks → cosine → score thresholds) to avoid redundant computation. Write the filtered output to a new TSV, retaining all original columns for passing rows. Validate by verifying output row count is ≤ input row count and spot-checking that all retained rows meet specified thresholds.

Related tools

• SIRIUS (Generates in silico molecular formula and structure annotations with Zodiac and Cosmic confidence scores; output is filtered by min_ZodiacScore threshold to retain only high-confidence identifications) — https://bio.informatik.uni-jena.de/software/sirius/
• ISDB (Provides spectral library matching annotations with cosine similarity scores, shared peak counts, and ppm error estimates; annotations are filtered by cosine, shared_peaks, ppm_error, and min_score_final thresholds)
• timaR (Performs weighted in silico annotation and taxonomically informed reponderation of ISDB results; output is filtered downstream using the same filter parameters applied to ISDB annotations) — https://taxonomicallyinformedannotation.github.io/tima-r/index.html
• Inventa (Consumes filtered annotation tables as input to Feature Component and Class Component calculations; filtering ensures only validated annotations contribute to originality scoring) — https://github.com/luigiquiros/inventa

Examples

# Load SIRIUS annotations and filter by min_ZodiacScore
df = pd.read_csv('compound_identification.tsv', sep='\t')
df_filtered = df[df['ZodiacScore'] >= 0.9]
df_filtered.to_csv('compound_identification_filtered.tsv', sep='\t', index=False)
# OR in Inventa: min_ZodiacScore = 0.9; min_ConfidenceScore = 0.0; cosine = 0.7; shared_peaks = 10; ppm_error = 5

Evaluation signals

• Output file contains only rows where min_ZodiacScore (SIRIUS) or all of {cosine, shared_peaks, ppm_error, min_score_final} (ISDB) criteria are met; spot-check 5–10 rows to confirm
• Output row count is strictly less than or equal to input row count; equality is acceptable only if all input rows already met thresholds
• All original annotation columns are present in output; no columns were dropped or reordered
• Filtering produces ≥1 passing rows (non-empty output table); empty output signals over-aggressive thresholding and should trigger parameter review
• When applied to same input multiple times with identical parameters, output is byte-identical (deterministic filtering)

Limitations

• Threshold choices (e.g., min_ZodiacScore=0.9, cosine=0.7) are heuristic and may require optimization for your specific instrument, ionization mode, or organism type; no principled derivation is provided in the article.
• Sequential filtering (ppm → shared_peaks → cosine → score) may create rank-dependent artifacts if thresholds are poorly calibrated; tight coupling between parameters can amplify filtering stringency unpredictably.
• No adjustment for multiple hypothesis testing or false discovery rate control across annotation sources; filtering operates independently on SIRIUS vs. ISDB, so combined annotation sets may retain systematic biases.
• Low-abundance or rare metabolites with marginal spectral quality may fail thresholds despite biological relevance, leading to false negatives in downstream discovery workflows.

Evidence

• [other] min_ZodiacScore filtering: "The min_ZodiacScore filtering mechanism applies a cut-off threshold of 0.9 to retain only SIRIUS annotations meeting or exceeding this confidence score, removing lower-scoring rows from the"
• [methods] SIRIUS annotation parameters: "Parse the Zodiac score column and apply a minimum threshold filter (min_ZodiacScore = 0.9) to retain only rows where Zodiac score is ≥0.9."
• [other] ISDB/spectral library filters: "ppm_error = 5 # min error in ppm to consider an annotation valable; shared_peaks = 10 # min number of shared peaks; cosine = 0.7 # min cosine score to consider an annotation valable"
• [other] Filter application workflow: "Set annotation cleaning parameters (ppm_error, shared_peaks, cosine, ionisation_mode) For cleaning-up annotations from GNPS ppm_error = 5 # min error in ppm to consider an annotation valable"
• [methods] Output format and validation: "Write the filtered annotation table to a new TSV file, retaining all original columns for retained rows. Validation: verify the output file contains only rows with Zodiac score ≥0.9 and row count is"
• [other] timaR integration: "tima_results_filename: timaR reponderated output format. - for performing in silico annotations and taxonomically informed reponderation."
• [methods] Downstream use in Inventa: "The Feature Component (FC) is a ratio of the number of specific non-annotated features over the total number of features of each extract."