HolobiomicsLab recursive-propagation-algorithm-development
Summary
Implement a recursive annotation propagation algorithm that iteratively refines metabolite identities by leveraging confidence scores and relationships across both knowledge-driven (biochemical ontology) and data-driven (MS2 similarity) network layers. This skill combines topology-based traversal with confidence-weighted evidence aggregation to enhance metabolite annotation coverage and accuracy in untargeted metabolomics.
When to use
When you have constructed a two-layer metabolite annotation network (knowledge-driven and data-driven) and need to propagate initial seed annotations (e.g., from database matches or high-confidence MS2 similarity) across the network to infer identities for unannotated or ambiguous metabolite nodes. Apply this skill when annotation confidence from a single layer is insufficient or when you wish to leverage complementary evidence from both biochemical relationships and experimental co-occurrence patterns simultaneously.
When NOT to use
- Input metabolite network is single-layer or unimodal (only knowledge-driven OR only data-driven) — the propagation algorithm's advantage derives from leveraging both layers; single-layer networks offer no comparative benefit.
- Initial seed annotations have uniformly low confidence or are sparse (< 5–10% of nodes); insufficient starting evidence will propagate noise rather than refine identities.
- Metabolite identities are already comprehensively annotated from a single authoritative database; propagation is unnecessary and may introduce conflicts.
Inputs
- Two-layer interactive metabolite network with knowledge-driven nodes/edges (biochemical pathways, metabolic reaction networks) and data-driven nodes/edges (MS2 similarity networks from experimental mass spectrometry data)
- Seed metabolite annotations with initial confidence scores for each layer
- Edge weight matrices for both layers (e.g., reaction reliability scores, MS2 cosine similarity values)
Outputs
- Recursively refined metabolite annotations with aggregated confidence scores across iterations
- Annotation propagation trajectories documenting which layer(s) contributed to each metabolite identity inference
- Final annotation table with coverage metrics (proportion of metabolites annotated before and after propagation)
How to apply
Begin by defining confidence scores for initial metabolite annotations from both layers (e.g., database match reliability from knowledge-driven layer, MS2 cosine similarity from data-driven layer). Implement a recursive traversal algorithm that, for each annotated metabolite node, visits neighboring nodes in both network layers and aggregates confidence evidence from both paths. At each recursion step, compute a refined confidence score by combining parent node confidence with edge weights (e.g., reaction certainty, feature similarity quantiles) and layer-specific priors. Terminate recursion when confidence increments fall below a threshold (e.g., < 0.01) or maximum iteration depth is reached. Validate the algorithm by comparing final annotation assignments for a test metabolite set against known reference annotations, checking for improved coverage and maintained consistency across iterations.
Related tools
- MrnAnnoAlgo3 (MetDNA3) (Core algorithm module that implements the two-layer interactive networking topology and recursive annotation propagation for metabolite annotation in untargeted metabolomics) — https://github.com/ZhuMetLab/MrnAnnoAlgo3
Evaluation signals
- Annotation coverage increases monotonically across recursion iterations (or reaches stable plateau) with no decrease in already-assigned confidences.
- Aggregated confidence scores for each metabolite reflect contributions from both layers (i.e., annotated nodes show non-zero evidence from knowledge-driven AND data-driven pathways when applicable).
- Recursively annotated metabolites match reference annotations with precision ≥ 90% when compared against a test set with known identities; novel propagated annotations are chemically plausible (e.g., belong to same pathway or biochemical class as seed annotation).
- Recursion terminates within predetermined iteration count (e.g., < 50 iterations) and execution time is < 1 hour for typical untargeted metabolomics datasets (consistent with '10-fold faster' claim).
- Cross-validation on held-out seed annotations demonstrates that algorithm recovers true identities with precision comparable to or exceeding single-layer annotation methods.
Limitations
- Algorithm performance is sensitive to the quality and completeness of the underlying two-layer network; sparse or incomplete knowledge-driven or data-driven layers will limit effective propagation.
- Confidence score aggregation is heuristic-dependent (choice of weighting scheme, layer priors, convergence threshold); different parameterizations may yield different final annotations.
- Propagation can amplify errors from initial seed annotations if incorrect high-confidence annotations are recursively reinforced across iterations; robust filtering of seed quality is critical.
- Algorithm assumes that both network layers are reasonably well-aligned (i.e., that knowledge-driven and data-driven relationships are largely congruent); highly conflicting layer evidence may produce ambiguous or contradictory annotations.
Evidence
- [other] implement-recursive-propagation: "Implement the recursive annotation propagation algorithm that iteratively refines metabolite identities by leveraging confidence scores and relationships from both layers."
- [readme] two-layer-topology: "designed to annotate metabolites through a two-layer interactive networking topology (knowledge-driven and data-driven) and recursive annotation propagation algorithms"
- [readme] performance-claim: "Processes a typical untargeted metabolomics dataset in just one hour—over 10-fold faster than previous versions."
- [readme] layer-definitions: "Integrates knowledge-driven (biochemical pathways, metabolic reaction networks) and data-driven (experimental MS2 similarity networks) layers for comprehensive and accurate metabolite annotation."