Similarity Network Fusion (SNF)

For each available omics data type (e.g., mRNA expression, DNA methylation, miRNA expression), SNF first constructs a sample-similarity network. In this network, nodes represent patients, and weighted edges represent the pairwise similarity between patients based on that single data type. The similarity is typically computed using a scaled exponential similarity kernel, which emphasizes local neighborhoods. This step transforms raw high-dimensional molecular profiles into a graph structure that captures the relational geometry of the data.

The core innovation of SNF is its non-linear, message-passing fusion process. Starting with the individual patient similarity networks, the algorithm iteratively updates each network by making it more similar to the others at each step. Crucially, this is achieved by diffusing information through the local neighborhood structures:

• A global structure is captured via a full kernel matrix.
• A local affinity matrix encodes only the K-nearest neighbors. The fusion step uses the local affinities to propagate information, ensuring that strong similarities present in only one data type are reinforced, while weak or noisy similarities are diminished, converging to a single consensus network.

Unlike early integration methods that simply concatenate features, SNF excels at capturing complementary information. If a strong patient similarity is supported by only one data type (e.g., a mutation pattern) but is absent in another (e.g., gene expression), the iterative diffusion process will propagate that similarity into the other networks. This allows the final fused network to reflect signals that are shared across all data types as well as those that are unique to a single data type, providing a more holistic and robust view of patient relationships for downstream tasks like subtyping.

Once the fused patient similarity network is constructed, SNF applies spectral clustering to identify distinct patient subgroups. This technique partitions the graph by analyzing the eigenvectors of the network's Laplacian matrix, effectively finding clusters of patients who are densely connected in the fused network. The number of clusters is often estimated using an eigengap heuristic or a rotation cost statistic. The resulting subtypes are clinically meaningful, as demonstrated in original studies where SNF identified cancer subtypes with significantly different survival outcomes that were not discovered by single-omics or concatenation-based analyses.

SNF is inherently robust to data heterogeneity and experimental noise. Because the fusion process relies on the local neighborhood structure (K-nearest neighbors) rather than raw feature values, it is less sensitive to differences in data scale, distribution, or measurement platform across omics types. The iterative diffusion acts as a denoising mechanism: spurious similarities that do not have support across multiple data types or within a strong local neighborhood are iteratively filtered out, while true biological signals are amplified. This makes SNF particularly effective for integrating data from different laboratories or technologies.

A key output of SNF-based patient subtyping is the validation of clinical relevance through survival analysis. After clusters are identified, a Kaplan-Meier plot is generated to visualize survival probability over time for each subtype, and a log-rank test is performed to determine if the differences in survival curves are statistically significant. This step is critical for demonstrating that the computationally derived subtypes correspond to biologically and clinically distinct disease trajectories, providing actionable stratification for prognosis and treatment decisions.

A classical statistical method for multi-omics integration that finds linear combinations of variables from two datasets that are maximally correlated. Unlike SNF, CCA assumes linear relationships.

• Mechanism: Projects high-dimensional omics data into a shared low-dimensional subspace where samples are aligned.
• Limitation: Struggles with non-linear biological relationships and requires paired samples across all modalities.
• Contrast with SNF: SNF uses iterative message-passing on k-nearest neighbor graphs, capturing non-linear local geometries without forcing a global linear projection.

An unsupervised statistical framework that infers a set of latent factors capturing the principal sources of variation across multiple omics data types from the same samples.

• Output: Decomposes variation into factors shared across modalities and factors private to a single modality.
• Advantage: Provides interpretable factor weights that link specific genes, proteins, or metabolites to a latent biological process.
• Relationship to SNF: Both are unsupervised fusion methods, but MOFA uses a Bayesian matrix factorization approach, while SNF operates on sample-similarity networks directly.

A deep learning architecture that operates directly on graph-structured data, representing a modern evolution of the network fusion concept pioneered by SNF.

• Mechanism: Nodes (samples or genes) exchange messages with their neighbors through multiple layers, learning complex relational patterns.
• Application: Used to model gene regulatory networks, protein-protein interaction graphs, and patient similarity networks.
• SNF Connection: SNF constructs a fused patient similarity network that can serve as the input graph for a downstream GNN classifier for survival prediction or drug response modeling.

A mathematical prerequisite for visualizing the high-dimensional omics data that SNF integrates. Algorithms like PCA, t-SNE, and UMAP are essential for exploring the fused network.

• PCA: Linear projection maximizing variance; often used to initialize SNF's distance metrics.
• t-SNE: Preserves local neighborhoods; excellent for visualizing patient clusters post-fusion.
• UMAP: Preserves both local and global structure; faster than t-SNE and widely used to validate SNF-derived subtypes.
• Workflow: SNF fuses networks → spectral clustering on the fused network → UMAP for 2D visualization of subtypes.

The clustering algorithm most commonly applied to the fused similarity matrix produced by SNF to identify patient subtypes.

• Mechanism: Performs eigendecomposition on the graph Laplacian of the fused network, embedding nodes in a low-dimensional space where traditional clustering (e.g., k-means) becomes effective.
• Why SNF uses it: The fused network is a weighted graph; spectral clustering naturally handles the non-convex cluster shapes that emerge from multi-omics data.
• Output: Discrete cluster assignments (e.g., Subtype 1, Subtype 2, Subtype 3) with significantly different survival curves.