Feature Clustering Intensity Based

feature-clustering-intensity-based

Summary

Cluster LC-MS features that likely derive from the same metabolite by leveraging intensity patterns across samples, a preprocessing step in mWISE that prepares candidates for downstream filtering and diffusion-based prioritization. This skill reduces annotation redundancy by grouping related peaks before network-based ranking.

When to use

Apply this skill after matching mass-to-charge ratios to a KEGG database and obtaining multiple candidate metabolites per feature, but before filtering quasi-molecular adducts. Use it when your feature table includes intensity values across samples and you suspect many peaks represent isotopologues, adducts, or in-source fragments of the same metabolite rather than distinct compounds.

When NOT to use

• Input lacks or has insufficient intensity information across replicates/conditions (clustering cannot identify co-varying features without intensity variance).
• All features are already known to represent distinct metabolites (clustering would be redundant).
• Target is targeted metabolomics with a small, predefined set of peaks (intensity-based clustering designed for untargeted discovery where candidate explosion is common).

Inputs

• Feature table with m/z values, retention time (if available), and intensity matrix across samples
• KEGG candidate assignments (output from matching stage with multiple candidates per feature)

Outputs

• Clustered feature table with cluster membership annotations
• Cluster representatives ready for filtering and diffusion-based prioritization

How to apply

Execute the mWISE featuresClustering function on the feature table populated with m/z values and intensity profiles across samples. The function groups features that co-vary in intensity, implicitly assuming features with similar abundance patterns across replicates or conditions likely originate from a single metabolite. The clustering is intensity-driven, meaning it uses the correlation or co-elution of intensity signals rather than m/z alone. After clustering, each cluster receives a single representative rank in downstream steps, reducing false positives. The output is a clustered candidate table that feeds directly into clusterBased.filter, which then applies quasi-molecular adduct and frequency thresholds to remove spurious members.

Related tools

• mWISE (Provides the featuresClustering function and wraps clustering into the full annotation workflow) — https://dev.b2s.club/b2slab/mWISE
• CAMERA (Informs default adduct and fragment definitions used to refine clusters after intensity-based grouping)
• cliqueMS (Referenced as source for adduct and fragment knowledge integrated into cluster refinement)
• R (Execution environment for featuresClustering function)

Examples

# Load Trypanosoma dataset and apply feature clustering
data("sample.dataset")
data("sample.keggDB")
clustered_features <- featuresClustering(sample.dataset$featureTable)

Evaluation signals

• Cluster size and composition: verify that cluster members have correlated intensity patterns across samples and plausible m/z differences (e.g., ±1.008 Da for isotopes, adduct mass shifts).
• Reduction in candidate redundancy: confirm that the number of unique clusters is substantially less than the number of raw matched candidates, indicating successful grouping.
• Downstream filter stability: after clusterBased.filter is applied to clusters, check that quasi-molecular adducts dominate and that minimum observed frequency threshold removes singletons as intended.
• Performance metric improvement: verify that top-3 precision, recall, and F1-score metrics (computed by performanceEvaluation against df.Ref) improve or remain stable compared to unfiltered candidates.
• Peak recovery validation: confirm that recoveringPeaks function successfully recovers any completely filtered clusters, indicating clusters were appropriately formed.

Limitations

• Clustering quality depends on sample heterogeneity and intensity variance; samples with uniform abundance profiles or low dynamic range may yield uninformative clusters.
• Intensity-based clustering assumes co-variance implies common origin; features with similar abundance by chance (unrelated metabolites with identical sample distribution) may be incorrectly grouped.
• No explicit retention time information is mentioned in the article; if RT is available but not used in clustering, temporally separated isomers may be grouped.
• The article does not specify clustering algorithm parameters (e.g., distance metric, linkage criterion, or correlation threshold); users must rely on mWISE defaults.

Evidence

• [intro] the features that may come from the same metabolite are clustered using the featuresClustering function: "the features that may come from the same metabolite are clustered using the featuresClustering function"
• [intro] mWISE integrates several strategies to provide a fast annotation of peak-intensity tables. It consists of three main steps aimed at i) matching mass-to-charge ratio values to KEGG database, ii) clustering and filtering the potential KEGG candidates: "mWISE integrates several strategies to provide a fast annotation of peak-intensity tables. It consists of three main steps aimed at i) matching mass-to-charge ratio values to KEGG database, ii)"
• [intro] the clusterBased.filter function: "the clusterBased.filter function"
• [intro] the diffusion prioritized table is built using the finalResults function: "the diffusion prioritized table is built using the finalResults function"