Patient Similarity Networks

The fundamental architecture of a PSN defines patients as nodes and similarity scores as edges. Edges are derived from a chosen distance metric (e.g., Euclidean distance for continuous variables, Jaccard index for binary mutations) applied to a multi-dimensional feature vector. A critical design choice is the similarity threshold; edges are often sparsified using a k-nearest neighbors (k-NN) approach to retain only the strongest connections and reduce computational noise.

SNF is a robust computational method for integrating heterogeneous data types (e.g., mRNA expression, DNA methylation, clinical labs) into a single comprehensive patient network. The process involves:

• Constructing a separate patient similarity network for each available data type.
• Iteratively updating each network using a message-passing theory, making them more similar to each other with each step.
• Converging on a fused network that captures both shared and complementary information, providing a holistic view of patient biology.

Once a PSN is constructed, community detection algorithms are applied to partition the graph into clusters of densely connected nodes. The Louvain algorithm is a popular choice, as it greedily optimizes a metric called modularity to identify natural groupings without pre-specifying the number of clusters. These resulting communities represent computationally derived patient subtypes that can be correlated with clinical outcomes like drug response or disease progression.

PSNs can be extended to model disease progression over time. A dynamic PSN incorporates longitudinal data by creating a series of networks at discrete time points. Analyzing how a patient's position within the network topology shifts—for example, moving from a mild to a severe disease community—provides a powerful framework for trajectory inference and predicting future clinical events, moving beyond a static snapshot of disease.

Modern PSN analysis often leverages Graph Neural Networks (GNNs) for supervised learning tasks. Instead of manual feature engineering, a GNN can operate directly on the graph structure. A Graph Convolutional Network (GCN) aggregates feature information from a patient's neighbors to generate a context-aware embedding. This allows for highly accurate node classification tasks, such as predicting a patient's response to a specific therapy based on their position and connections within the similarity network.

A critical characteristic of a clinically useful PSN is its interpretability. After clusters are identified, they must be characterized by their defining features (e.g., a specific mutation, elevated biomarker). Statistical enrichment analysis is used to link network communities to known biological pathways. The ultimate validation is a robust association with a clinical endpoint, such as overall survival or treatment response, demonstrating that the network-derived stratification captures a clinically actionable signal.

A computational method for integrating diverse patient data types by constructing and fusing patient similarity networks into a single comprehensive view. SNF iteratively updates each network to make it more similar to the others, converging on a fused network that captures shared and complementary information.

• Handles non-linear cross-modal relationships without assuming data distributions
• Robust to noise and heterogeneity across data types (genomic, proteomic, clinical)
• Outperforms early integration methods by avoiding the 'curse of dimensionality'

A greedy optimization method for community detection in large networks that maximizes modularity to identify densely connected patient groups. Applied to patient similarity networks, it partitions patients into clinically meaningful subgroups without pre-specifying cluster count.

• Modularity optimization: Quantifies the strength of division into communities
• Two-phase iterative process: Local node movement followed by network aggregation
• Scales to networks with millions of nodes and edges

A statistical framework for the unsupervised integration of multiple omics data types to discover the principal sources of biological variation driving patient subgroups. MOFA infers latent factors that capture the global sources of variability across data modalities.

• Decomposes variation into factors shared across omics and factors specific to individual data types
• Handles missing data natively without imputation
• Provides interpretable feature weights linking factors to specific genes, proteins, or metabolites

A method for studying the shape of complex patient data using persistent homology to detect high-dimensional voids and connectivity patterns. TDA constructs simplicial complexes from patient similarity networks to reveal continuous disease trajectories.

• The Mapper algorithm creates a graph representation preserving topological structure
• Identifies disease subpopulations that form loops or flares rather than discrete clusters
• Particularly effective for visualizing continuous phenotypic spectra like cancer evolution

A method that simultaneously learns feature representations with an autoencoder and assigns clusters in the latent space to improve patient stratification. DEC jointly optimizes the embedding and clustering objectives.

• Uses a Kullback-Leibler divergence loss to refine cluster assignments iteratively
• Learns a non-linear mapping from high-dimensional patient features to a low-dimensional latent space
• Outperforms two-stage approaches (reduce then cluster) on complex biomedical data

A class of models, such as the Dirichlet Process Mixture, that allow the number of patient clusters to grow with the data rather than being fixed a priori. This avoids the need to pre-specify the number of subgroups.

• The Chinese Restaurant Process prior defines a distribution over infinite clusterings
• Automatically infers the optimal number of patient subtypes from the data
• Provides full posterior uncertainty over cluster assignments, not just point estimates