Skip to content

SKILL.md

Latest commit

 

History

History
235 lines (168 loc) · 7.31 KB

SKILL.md

File metadata and controls

235 lines (168 loc) · 7.31 KB
name bio-multi-omics-similarity-network
description Similarity Network Fusion (SNF) for patient stratification using multi-omics data. Integrates multiple data types into a unified patient similarity network. Use when performing patient stratification or integrating multi-omics data into unified similarity networks.
tool_type r
primary_tool SNFtool

Version Compatibility

Reference examples tested with: scanpy 1.10+

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

  • R: packageVersion('<pkg>') 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.

Similarity Network Fusion

"Stratify patients using multi-omics data" → Fuse omics-specific patient similarity networks into a unified network for subtype discovery and clustering.

  • R: SNFtool::SNF() to fuse networks, spectralClustering() for subtyping

Basic SNF Workflow

Goal: Fuse multiple omics-specific patient similarity networks into a single unified network.

Approach: Compute per-omics distance and affinity matrices, then iteratively fuse with SNF.

library(SNFtool)

# Load omics data (samples x features)
data1 <- as.matrix(read.csv('rnaseq.csv', row.names = 1))
data2 <- as.matrix(read.csv('methylation.csv', row.names = 1))
data3 <- as.matrix(read.csv('mirna.csv', row.names = 1))

# Ensure matching samples
common <- Reduce(intersect, list(rownames(data1), rownames(data2), rownames(data3)))
data1 <- data1[common, ]
data2 <- data2[common, ]
data3 <- data3[common, ]

# Compute distance matrices
dist1 <- dist2(as.matrix(data1), as.matrix(data1))
dist2 <- dist2(as.matrix(data2), as.matrix(data2))
dist3 <- dist2(as.matrix(data3), as.matrix(data3))

# Construct affinity matrices
# K = number of neighbors, alpha = hyperparameter
K <- 20
alpha <- 0.5

aff1 <- affinityMatrix(dist1, K, alpha)
aff2 <- affinityMatrix(dist2, K, alpha)
aff3 <- affinityMatrix(dist3, K, alpha)

# Fuse networks
# T = number of iterations
fused <- SNF(list(aff1, aff2, aff3), K = K, t = 20)

Cluster Patients

Goal: Identify patient subtypes from the fused similarity network using spectral clustering.

Approach: Estimate optimal cluster count from the fused graph, then apply spectral clustering.

# Determine optimal number of clusters
estimateNumberOfClustersGivenGraph(fused, NUMC = 2:10)

# Spectral clustering
num_clusters <- 3
clusters <- spectralClustering(fused, num_clusters)

# Add to sample metadata
sample_info <- data.frame(
    Sample = rownames(data1),
    Cluster = factor(clusters)
)

Visualize Network

Goal: Display the fused patient network as a graph and heatmap with cluster annotations.

Approach: Convert the fused matrix to an igraph object, filter weak edges, and render with cluster coloring.

library(igraph)

# Convert to igraph
g <- graph_from_adjacency_matrix(fused, mode = 'undirected', weighted = TRUE, diag = FALSE)

# Remove weak edges
threshold <- quantile(E(g)$weight, 0.9)
g_filtered <- delete_edges(g, E(g)[weight < threshold])

# Plot
V(g_filtered)$color <- clusters
plot(g_filtered, vertex.size = 5, vertex.label = NA,
     edge.width = E(g_filtered)$weight * 2,
     main = 'SNF Patient Network')

# Heatmap
library(pheatmap)
pheatmap(fused, cluster_rows = TRUE, cluster_cols = TRUE,
         annotation_row = sample_info['Cluster'],
         show_rownames = FALSE, show_colnames = FALSE)

Normalized Mutual Information

Goal: Evaluate clustering quality by comparing SNF clusters against known subtypes and single-omics baselines.

Approach: Compute NMI between predicted clusters and true labels for fused vs individual affinity networks.

# Compare with known labels
true_labels <- read.csv('phenotype.csv')$Subtype

# NMI score
nmi <- calNMI(clusters, true_labels)
cat('NMI:', nmi, '\n')

# Compare individual vs fused
nmi_rna <- calNMI(spectralClustering(aff1, num_clusters), true_labels)
nmi_meth <- calNMI(spectralClustering(aff2, num_clusters), true_labels)
nmi_mirna <- calNMI(spectralClustering(aff3, num_clusters), true_labels)

cat('NMI RNA only:', nmi_rna, '\n')
cat('NMI Methylation only:', nmi_meth, '\n')
cat('NMI miRNA only:', nmi_mirna, '\n')
cat('NMI Fused:', nmi, '\n')

Feature Ranking with SNF

Goal: Rank features by their contribution to the SNF-derived patient clusters.

Approach: Perform ANOVA per feature across cluster assignments, ranking by F-statistic p-value.

# Rank features by their contribution to clustering
# Using network-based method

# For each omics layer
rank_features <- function(data, clusters) {
    # Calculate feature importance based on cluster separation
    f_values <- apply(data, 2, function(x) {
        summary(aov(x ~ factor(clusters)))[[1]][1, 4]
    })
    f_values[is.na(f_values)] <- 1
    names(sort(f_values))
}

top_rna <- rank_features(data1, clusters)
top_meth <- rank_features(data2, clusters)

Survival Analysis with Clusters

Goal: Assess clinical relevance of SNF clusters by comparing survival outcomes between subtypes.

Approach: Fit Kaplan-Meier curves per cluster and test significance with the log-rank test.

library(survival)
library(survminer)

# Load survival data
surv_data <- read.csv('survival.csv')
surv_data$Cluster <- clusters[match(surv_data$Sample, rownames(data1))]

# Kaplan-Meier
fit <- survfit(Surv(Time, Event) ~ Cluster, data = surv_data)

ggsurvplot(fit, data = surv_data, pval = TRUE,
           risk.table = TRUE, palette = 'jco',
           title = 'SNF Cluster Survival')

# Log-rank test
survdiff(Surv(Time, Event) ~ Cluster, data = surv_data)

Parameter Tuning

Goal: Optimize SNF hyperparameters (K neighbors, alpha) for best clustering performance.

Approach: Grid search over K and alpha values, evaluating each combination by NMI against known labels.

# Grid search over K and alpha
K_range <- c(10, 20, 30)
alpha_range <- c(0.3, 0.5, 0.8)

results <- expand.grid(K = K_range, alpha = alpha_range, NMI = NA)

for (i in 1:nrow(results)) {
    aff1 <- affinityMatrix(dist1, results$K[i], results$alpha[i])
    aff2 <- affinityMatrix(dist2, results$K[i], results$alpha[i])
    aff3 <- affinityMatrix(dist3, results$K[i], results$alpha[i])

    fused <- SNF(list(aff1, aff2, aff3), K = results$K[i], t = 20)
    clusters <- spectralClustering(fused, num_clusters)
    results$NMI[i] <- calNMI(clusters, true_labels)
}

best <- results[which.max(results$NMI), ]
cat('Best parameters: K =', best$K, ', alpha =', best$alpha, '\n')

Integration with Clinical Features

Goal: Incorporate clinical variables as an additional data view in the SNF fusion.

Approach: Encode clinical features numerically, compute a clinical affinity matrix, and include it in the SNF fusion step.

# Add clinical features as another view
clinical <- read.csv('clinical.csv', row.names = 1)
clinical_numeric <- model.matrix(~ . - 1, data = clinical)

dist_clinical <- dist2(clinical_numeric, clinical_numeric)
aff_clinical <- affinityMatrix(dist_clinical, K, alpha)

# Fuse all including clinical
fused_with_clinical <- SNF(list(aff1, aff2, aff3, aff_clinical), K = K, t = 20)

Related Skills

  • mofa-integration - Factor-based integration
  • mixomics-analysis - Supervised integration
  • single-cell/clustering - Single-cell clustering methods