Bio Multi Omics Similarity Network

Version Compatibility

Reference examples tested with: SNFtool 2.3+, igraph 2.0+, pheatmap 1.0+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('SNFtool') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

SNFtool argument names are easy to misread: affinityMatrix's width argument is sigma (not alpha or mu), dist2 returns SQUARED Euclidean distance (take the square root), and spectralClustering's K is the number of CLUSTERS, a name collision with the K-neighbors argument of affinityMatrix/SNF.

Similarity Network Fusion

"Stratify my patients using multi-omics data" -> Fuse per-omic patient-similarity networks into one graph and spectral-cluster it - because the clustering always returns the number of subtypes requested, so the subtype is a claim to defend, not a discovery.

• R: SNF() to fuse networks, spectralClustering() to partition, then validate

Scope: patient-similarity-space integration - per-omic affinity networks, SNF fusion, spectral clustering into candidate subtypes, cluster-number choice, and the stability/survival/replication defense, plus the integrative-clustering landscape (NEMO, PINS, iCluster, CIMLR, intNMF, consensus). Feature-space latent factors -> mofa-integration. Supervised feature signatures -> mixomics-analysis. Survival mechanics -> clinical-biostatistics/survival-analysis. Single-cell graph clustering -> single-cell/clustering.

The Single Most Important Modern Insight -- Spectral Clustering Always Returns the Requested Number of Clusters, So a Subtype Is a Claim to Defend, Not a Thing Discovered

Ask the spectral step for four clusters and it returns four, whether or not four real patient groups exist; the eigengap suggesting "four" means only that the fused graph has a roughly four-block shape, not that the disease has four subtypes. SNF's fusion is genuinely powerful - its cross-network diffusion reinforces patient pairs that multiple omics agree on and erodes omic-specific noise - but power to draw a boundary is not evidence the boundary is real. Three defenses, none of which the algorithm supplies:

1. Stability. Resample the patients, re-cluster, and measure agreement (consensus index / NMI across subsamples). A real subtype recurs; an artifact does not. Report stability across a fixed (K, sigma) grid rather than tuning those hyperparameters to a target.
2. Survival separation after adjustment. Subtypes should differ in outcome in a Cox model adjusted for known prognostic covariates (stage, age, grade), so a "subtype" is not just re-encoded stage. The discovery-cohort p is supporting evidence, not proof - and a permutation p is safer than the chi-square approximation at small arm sizes.
3. Replication and the fused-versus-best-single-omic check. Multi-omics does not consistently beat the best single omic (Rappoport and Shamir 2018), and SNF's diffusion can dilute a signal carried by one omic. Benchmark the fused clustering against each single-omic clustering, and replicate the subtypes in an independent cohort.

The honest report names the hyperparameters, shows the clusters are not knife-edge-sensitive to them, and treats the discovery survival p as supporting evidence, not the finding.

Tool Taxonomy (Integrative Clustering)

Decision Tree by Scenario

Build the Per-Omic Affinity Networks

Goal: Turn each omic into a patient-by-patient similarity network on a common scale, collapsing the high-dimensional feature space into an n-by-n object so feature count buys no votes.

Approach: Standardize each continuous omic per feature, compute the (square-rooted) Euclidean distance, then apply the local-scaled Gaussian kernel. SNF requires every patient to have every omic, so intersect to common samples first and report how many that drops.

library(SNFtool)

K <- 20        # neighbors defining the local kernel bandwidth; SNFtool guidance 10-30; changes cluster count
sigma <- 0.5   # kernel width multiplier (affinityMatrix's third arg, named sigma not alpha); guidance 0.3-0.8
t_iter <- 20   # cross-diffusion iterations; converges by ~10-20

norm_views <- lapply(list(rna=rna, meth=meth, mirna=mirna), standardNormalization)   # per-feature z-score before distance
dists <- lapply(norm_views, function(x) dist2(x, x)^(1/2))                            # dist2 returns SQUARED distance
affinities <- lapply(dists, function(d) affinityMatrix(d, K, sigma))

Fuse and Cluster

Goal: Fuse the per-omic networks into one graph and partition it, treating the cluster number as the central claim rather than a nuisance parameter.

Approach: Run SNF's cross-diffusion, read the four cluster-number estimates the package returns (not as truth but as plausibility), then spectral-cluster. The package itself warns the estimates cannot guarantee accuracy.

fused <- SNF(affinities, K, t_iter)
estimateNumberOfClustersGivenGraph(fused, NUMC=2:8)   # returns FOUR estimates: K1/K12 (eigengap), K2/K22 (rotation cost)
clusters <- spectralClustering(fused, K=4, type=3)    # here K is the CLUSTER COUNT (not neighbors); type 3 = Ng-Jordan-Weiss default

Defend the Subtypes

Goal: Show the clusters are stable, separate outcome, and are not just the best single omic before calling them subtypes.

Approach: Compare the fused clustering against each single-omic clustering, assess stability under resampling, and rank the post-hoc feature attribution with the package function rather than a hand-rolled test. Survival mechanics are routed out.

concordanceNetworkNMI(c(affinities, list(fused)), C=4)   # NMI among per-omic and fused clusterings: did fusion beat the best single omic?
feat_rank <- rankFeaturesByNMI(norm_views, fused)        # POST-HOC attribution: features that track the clusters, not a model that made them

Validate survival separation in a covariate-adjusted Cox model and report events per arm (clinical-biostatistics/survival-analysis owns the mechanics); assess stability by resampling patients and re-clustering, reporting agreement across subsamples. To assign a new patient to an existing subtype without re-clustering, use groupPredict(train_views, test_views, groups, K=20, method=1) (label propagation). For a mosaic cohort, switch to NEMO rather than dropping the incomplete patients.

Per-Method Failure Modes

Subtype count reported as a discovery

Trigger: "I found 4 subtypes" with no defense. Mechanism: spectral clustering returns the requested C regardless of structure. Symptom: a clean-looking partition that does not replicate. Fix: require eigengap plausibility, resampling stability, covariate-adjusted survival, and external replication before claiming a subtype count.

Hyperparameters tuned to a label, then claimed

Trigger: grid-searching K and sigma to maximize NMI against known labels. Mechanism: in real discovery there are no labels, so tuning to NMI is circular. Symptom: a result that only holds at the chosen (K, sigma). Fix: fix or pre-register a small (K, sigma) grid and show the clustering is stable across it; report sensitivity as a result.