Feature pairing confidence scoring

feature-pairing-confidence-scoring

Summary

Compute and assign confidence scores to candidate feature pairs from two LC-MS metabolomic datasets by evaluating similarity across m/z and retention time dimensions, then resolve one-to-one correspondences by selecting highest-confidence matches. This skill is essential for confidently aligning untargeted metabolomic features across replicate or comparative LC-MS runs.

When to use

You have two LC-MS feature tables (each with m/z, retention time, and intensity columns) and need to establish reliable correspondence between features across datasets. Specifically, apply this skill when: (1) you have computed initial pairwise similarity scores across m/z and retention time dimensions, (2) you have candidate feature pairs that may contain false positives, and (3) your analysis goal requires unambiguous one-to-one feature mappings rather than ambiguous many-to-many associations.

When NOT to use

• Input datasets are already manually curated or contain pre-validated feature correspondences; scoring is redundant.
• You require many-to-many feature associations or fuzzy matching rather than strict one-to-one correspondence.
• LC-MS datasets are from fundamentally different analytical methods (e.g., different ionization, column chemistry) where m/z and retention time alone are insufficient predictors of true correspondence.

Inputs

• LC-MS feature table 1 (with m/z, retention time, and intensity columns)
• LC-MS feature table 2 (with m/z, retention time, and intensity columns)
• mass-to-charge tolerance (in ppm or Da)
• retention time tolerance (in seconds or minutes)

Outputs

• Matched feature pairs table (original feature identifiers, m/z values, retention times, match confidence scores)
• Confidence score matrix (pairwise similarity scores)

How to apply

Load both LC-MS feature tables into Matlab and apply pairwise feature matching by computing similarity scores using the M2S matching algorithm, which evaluates both m/z and retention time dimensions simultaneously. Filter candidate matches by imposing mass-to-charge and temporal tolerances to eliminate false positives—these tolerances should reflect your instrument's typical m/z accuracy and retention time stability. Resolve competing one-to-one feature correspondences by selecting the highest-confidence match for each feature pair; this ensures each input feature maps to at most one output feature and vice versa, avoiding redundant or contradictory assignments. Compile matched pairs into a structured output table that preserves original feature identifiers, m/z values, retention times, and match confidence scores for downstream interpretation and troubleshooting.

Related tools

• M2S (Matlab package that implements pairwise feature matching and confidence scoring for untargeted metabolomic LC-MS datasets) — https://github.com/rjdossan/M2S

Evaluation signals

• Matched feature pairs preserve one-to-one cardinality: each input feature from dataset 1 pairs to exactly one feature in dataset 2 and vice versa (no unmapped or multiply-assigned features).
• Confidence scores are monotonically correlated with match accuracy: true-positive pairs should exhibit higher scores than false-positive pairs, validated by independent verification (e.g., standards, replicate analysis, or orthogonal methods).
• Output table is non-empty and contains all required columns (original feature identifiers, m/z, retention time, confidence scores) with no NaN or missing values in scored pairs.
• m/z and retention time differences between matched pairs fall within the imposed tolerances; pairs violating tolerance thresholds should be filtered or flagged.
• Confidence score distribution shows clear separation between high-confidence matches and filtered-out candidates, indicating effective discrimination between true and false correspondences.

Limitations

• Performance depends critically on the choice of mass-to-charge and retention time tolerances; overly loose tolerances yield false positives, overly strict tolerances yield false negatives. Tolerances should be instrument- and method-specific.
• M2S was developed for untargeted metabolomic datasets and may not generalize to targeted methods, data-dependent acquisition, or fundamentally different analytical workflows.
• No changelog or version history is documented, limiting ability to assess reproducibility across package updates or to understand which algorithm variants have been tested.
• Confidence scoring relies only on m/z and retention time similarity; it does not incorporate intensity, spectral fragmentation, or chemical structure information, which may improve resolution of ambiguous pairs.

Evidence

• [other] Apply pairwise feature matching by computing similarity scores across m/z and retention time dimensions using the M2S matching algorithm.: "Apply pairwise feature matching by computing similarity scores across m/z and retention time dimensions using the M2S matching algorithm."
• [other] Filter candidate matches based on mass-to-charge and temporal tolerances to eliminate false positives.: "Filter candidate matches based on mass-to-charge and temporal tolerances to eliminate false positives."
• [other] Resolve one-to-one feature correspondences by selecting the highest-confidence match for each feature pair.: "Resolve one-to-one feature correspondences by selecting the highest-confidence match for each feature pair."
• [readme] Matlab package to match untargeted metabolomic features of two LC-MS datasets: "Matlab package to match untargeted metabolomic features of two LC-MS datasets"
• [other] Compile matched pairs into a structured output table with original feature identifiers, m/z values, retention times, and match confidence scores.: "Compile matched pairs into a structured output table with original feature identifiers, m/z values, retention times, and match confidence scores."