Network Diffusion Prioritization

network-diffusion-prioritization

Summary

Applies diffusion-based prioritization on a metabolite network graph to rank candidate annotations by propagating probability scores through network topology. This skill prioritizes likely metabolite identities by leveraging context-aware network information rather than raw matching scores alone.

When to use

After clustering and filtering KEGG candidates for LC-MS features, when you have a ranked set of candidate metabolites per feature and access to a metabolite interaction network (e.g., from FELLA). Use this skill to re-rank candidates by their network distance and topological context, especially when you want to suppress low-confidence matches and elevate contextually coherent identifications.

When NOT to use

• Input candidate set is empty or has only one candidate per feature (diffusion will not improve ranking when there is no neighborhood context).
• Metabolite network graph is disconnected or highly fragmented (diffusion effectiveness degrades when most candidates are in isolated components).
• Upstream matching or clustering has failed to produce reliable candidate sets (diffusion amplifies noise from poor initial annotations).

Inputs

• Clustered and filtered KEGG candidates table (output from clusterBased.filter)
• Feature intensity matrix or probability scores per candidate
• Undirected metabolite network graph (igraph object or adjacency structure)
• Quasi-molecular adduct and frequency threshold metadata

Outputs

• Diffusion-prioritized ranked candidates table
• Final ranked results with z-score normalized diffusion scores
• Recovered peak assignments for completely filtered-out features

How to apply

Compute diffusion input scores using the diffusion.input function, selecting either 'probability' (soft weights based on matching confidence) or 'binary' (hard presence/absence) as the input.type. Apply diffusion scoring via network propagation on an igraph-compatible undirected metabolite graph, optionally normalizing scores using z-score transformation to account for graph topology (which suppresses candidates in low-connectivity regions). Recover completely removed peaks using recoveringPeaks, then compile final ranked results using finalResults. The z-score normalization is strongly recommended to normalize diffusion scores by taking into account the topology of the graph, preventing high-degree hub nodes from artificially inflating candidate scores.

Related tools

• mWISE (Executes diffusion.input and diffusion scoring functions; integrates network diffusion into the full annotation pipeline) — https://dev.b2s.club/b2slab/mWISE
• FELLA (Provides sample metabolite network graph (sample.graph) used for diffusion propagation)
• igraph (Converts and manipulates network graph structure (e.g., as.undirected) for diffusion computation)
• R (Runtime environment for diffusion functions and network operations)

Examples

diff.input <- diffusion.input(candidates, input.type='probability'); diff.scores <- diffusion(diff.input, g.metab, scores='z'); final <- finalResults(diff.scores); perf <- performanceEvaluation(final, df.Ref, top.cmps=3)

Evaluation signals

• Verify that diffusion scores are normalized (z-scores have mean ≈ 0 and SD ≈ 1) when z-score normalization is enabled, confirming topology correction is applied.
• Check that top-ranked candidates per feature have higher network connectivity or shorter paths to other high-confidence hits in the graph.
• Confirm that recoveringPeaks function successfully restores features that were removed by cluster-based filtering, with non-zero diffusion scores.
• Compare precision, recall, and F1-score against reference peaks (df.Ref) using performanceEvaluation with top.cmps argument; expect improvement over pre-diffusion rankings.
• Validate that final ranked results table contains no null or missing diffusion scores for retained candidates, and that rank order is monotonically consistent with diffusion score values.

Limitations

• Diffusion effectiveness depends on network coverage: candidates not in the metabolite graph or in isolated components will not benefit from prioritization.
• Raw diffusion scores (without z-score normalization) can be biased by network hub nodes, artificially elevating candidates near high-degree metabolites.
• Performance is sensitive to the choice of diffusion input type (probability vs. binary); selecting the wrong input.type can worsen ranking quality.
• Recovery of completely filtered peaks via recoveringPeaks may reintroduce false positives if the original cluster-based filter removed them for valid reasons.

Evidence

• [intro] building a final prioritized list using diffusion in networks: "building a final prioritized list using diffusion in networks"
• [intro] The different diffusion inputs can be computed using the diffusion.input function with probability or binary input types: "The input.type argument can be set to probability or binary."
• [intro] Z-score normalization accounts for graph topology in diffusion scoring: "The z score normalizes the diffusion scores by taking into account the topology of the graph."
• [intro] Recovery of completely removed peaks via recoveringPeaks function: "The recoveringPeaks function recovers the peaks that have been completely removed by the cluster-based filter."
• [intro] Final results compilation using finalResults function: "the diffusion prioritized table is built using the finalResults function"