Similarity Network Fusion (SNF)

The foundational step where each data type (e.g., mRNA expression, DNA methylation) is used to build a separate patient similarity network. In this graph, nodes represent patients, and weighted edges represent the similarity between them. Crucially, SNF employs a scaled exponential similarity kernel, which calculates the Euclidean distance between patients and then exponentiates it. This non-linear transformation effectively amplifies strong similarities and dampens weak, noisy ones, creating a sparse and robust network structure that captures the local neighborhood of each patient.

This is the core innovation of SNF. Rather than a one-time merge, the individual patient similarity networks are fused through a non-linear, iterative message-passing process. In each iteration, a network's structure is updated by diffusing information from the other networks, but only where it is consistent. The mathematical update step uses a global structure matrix (the full similarity network) and a local affinity matrix (K-nearest neighbors). By exchanging information only along strong local affinities, the process systematically reduces noise and converges to a single fused network that captures shared and complementary signals from all data types.

Once the final fused patient similarity network is generated, it serves as the input for spectral clustering to identify clinically meaningful patient subtypes. This technique uses the eigenvectors of the network's Laplacian matrix to perform dimensionality reduction, projecting the complex graph structure into a lower-dimensional space where traditional clustering (like K-means) becomes highly effective. This step directly addresses the challenge of finding cohesive patient groups within the integrated, high-dimensional data, producing the final patient stratification that can be correlated with clinical outcomes like survival.

A defining advantage of SNF is its inherent robustness to noise and data heterogeneity. Because the fusion process relies on the cross-validation of local network structures, a strong signal present in only one data type will be incorporated, while random noise that does not have support across multiple networks is naturally filtered out. This makes SNF exceptionally powerful for integrating data types with very different statistical properties, scales, and noise distributions—a common challenge in multi-omics studies—without requiring complex per-data-type normalization or feature selection.

SNF excels at capturing complementary information that would be lost by simple data concatenation. For example, one data type might clearly separate patients into two major groups, while another data type reveals a subtle subdivision within one of those groups. The iterative fusion process preserves both the strong global signal and the weaker local signal, allowing the final fused network to reflect a more complete and nuanced patient landscape. This ability to amplify weak but consistent signals across data types is key to discovering novel disease subtypes.

The SNF algorithm is computationally efficient, scaling quadratically with the number of patients but remaining fast for typical cohort sizes. The most intensive step is the iterative network diffusion, which involves matrix multiplication. However, the use of sparse K-nearest neighbor graphs for the local affinity matrix significantly reduces the computational burden. Implementations in R and Python, often leveraging optimized linear algebra libraries, allow SNF to process hundreds of patients across multiple omics data types in minutes on standard hardware, making it a practical tool for biomedical research.

The fundamental data structure SNF operates on. A graph-based representation where nodes are individual patients and weighted edges quantify the clinical or molecular similarity between them. These networks are constructed from individual data types (e.g., mRNA expression, DNA methylation) before fusion. The edge weight is typically defined using a scaled exponential similarity kernel, which emphasizes local neighborhoods and robustly captures shared biological signals while filtering out spurious distant correlations.

A complementary statistical framework for unsupervised integration of heterogeneous omics data. Unlike SNF's network-centric approach, MOFA infers a set of latent factors that capture the principal sources of biological variation across data modalities. It is particularly effective for identifying drivers of patient heterogeneity when the underlying biological processes are expected to manifest across multiple data layers simultaneously.

A resampling-based methodology used to validate the stability of patient subgroups discovered by SNF. It aggregates results from multiple clustering runs on perturbed data to calculate a consensus matrix, quantifying the robustness of each patient assignment. This is a critical step for ensuring that the identified molecular subtypes are not artifacts of a specific algorithmic run but represent genuine, reproducible biological structures.

A greedy optimization method for community detection in large networks, frequently applied to the fused SNF network to identify patient clusters. The algorithm maximizes a metric called modularity, which measures the density of links inside communities compared to links between communities. Its computational efficiency makes it suitable for large patient cohorts, iteratively merging nodes to find the partition with the highest modularity score.

An internal cluster validation metric used to evaluate the quality of patient stratification post-SNF. It measures how similar a patient is to its own cluster compared to other clusters, producing a score from -1 to 1.

• +1: Patient is well-matched to its own cluster.
• 0: Patient is on the boundary between clusters.
• -1: Patient may be assigned to the wrong cluster. A high average silhouette width indicates dense, well-separated patient subgroups.

A survival analysis technique used to clinically validate SNF-derived patient subtypes. After identifying clusters, this model tests whether the subgroups exhibit statistically significant differences in survival outcomes. A significant log-rank test and a high hazard ratio between clusters confirm that the computationally derived stratification has prognostic clinical utility, linking molecular patterns to tangible patient endpoints like progression-free or overall survival.