SNF integrates omics in patient-similarity space: it builds one patient-by-patient similarity network per omic and fuses them by cross-network diffusion, so feature count buys no votes and the n-much-smaller-than-p curse is dissolved into an n-by-n object. The fused graph is then spectral-clustered into candidate subtypes. The central caution is that spectral clustering always returns the requested number of clusters, so a subtype is a claim to defend with stability, survival separation, and replication - not something the algorithm discovers.
The skill owns SNF and the integrative-clustering landscape (NEMO, PINS, iCluster, CIMLR, intNMF, consensus) plus the validation discipline. Feature-space integration (which molecular axes or features matter) is the sibling mofa-integration or mixomics-analysis; survival mechanics live in clinical-biostatistics/survival-analysis; single-cell graph clustering is a different object in single-cell/clustering.
install.packages(c('SNFtool', 'igraph', 'pheatmap'))
Conceptual prerequisites and notes:
- SNF requires complete data: every patient must have every omic, because the cross-diffusion multiplies aligned n-by-n matrices. For a mosaic cohort, use NEMO instead of dropping patients.
- Standardize each continuous omic per feature before computing distances, or a high-variance feature dominates the distance.
- The SNFtool API is easy to misread: dist2 returns SQUARED Euclidean distance (take the square root), affinityMatrix's width argument is sigma (not alpha), and spectralClustering's K is the number of clusters, colliding with the K-neighbors argument elsewhere.
- SNF produces clusters but no feature-level model; feature attribution is post-hoc via rankFeaturesByNMI.
Tell your AI agent what you want to do: - "Identify patient subtypes by fusing my RNA-seq and methylation data" - "How many subtypes does my fused network actually support?" - "Show me whether the fusion beat my best single omic" - "Check that my subtypes are stable when I resample patients" - "Some patients are missing an omic - what should I use instead of SNF?"
"Fuse my RNA, protein, and methylation networks with SNF and propose patient subtypes. Then defend the subtype count: show me the eigengap estimates, whether the clusters are stable when I resample patients, and whether the fusion actually separates outcome better than my best single omic."
"Estimate the number of clusters from my fused network, but treat that as a plausibility check, not a truth. Show me how the clustering changes if I move K from 20 to 25, and tell me whether my headline subtypes survive."
"About a third of my patients are missing proteomics. Should I still use SNF, or switch to a method built for partial data, and what does the intersection cost me if I force complete cases?"
"Quantify whether my fused clustering agrees more with one omic than the others, and tell me honestly whether integrating helped or whether one omic was carrying the signal."
rankFeaturesByNMI and route survival mechanics and any biomarker claim out for validation.| Parameter | SNFtool argument | Typical range |
|---|---|---|
| K (neighbors) | affinityMatrix(diff, K, sigma), SNF(Wall, K, t) |
10-30 |
| sigma (kernel width) | affinityMatrix(diff, K, sigma) |
0.3-0.8 |
| t (fusion iterations) | SNF(Wall, K, t) |
10-20 |
| C (clusters) | spectralClustering(W, C) |
the claim, defended |
dist2 returns squared distance, so take the square root.sigma, and remember spectralClustering's K is the cluster count, not the neighbor count.rankFeaturesByNMI output as post-hoc characterization, not a model that generated the clusters.