Similarity Network Fusion for Aggregating Data Types on a Genomic Scale

Link: https://github.com/yik-cyber/SNFtool — Summary: Similarity Network Fusion aggregates different genomic data types. Inspired by multi-view methods in computer vision, the paper designs a graph fusion network to address the difficulty of jointly processing genomic data. Patients with the same cancer form a cohort; different genomic measurements per patient build separate graphs, then a fusion method merges them into one graph containing all genomic information. Clustering on the fused graph yields cancer subtypes; regression predicts patient survival risk. Framework and model structure: Problem statement Existing methods combine all genomic data types but have issues: Concatenation: low signal-to-noise; - Independent analysis then merge: separate outputs, hard to unify; - Prior gene selection: focuses on shared signal, loses complementary information; - ML clustering: very sensitive to number of pre-selected genes.

This paper proposes SNF (Similarity Network Fusion): build a similarity network for each data type;

Fuse all similarity networks nonlinearly into a single output network.

SNF advantages: captures both shared and complementary information across genomic types for fuller extraction;

Handles multiple genomic data types, robust to noise, works with few samples;

Iterative fusion removes weak edges and strengthens strong ones.

Method Below fuses two genomic types as an example. Raw data: patients with the same cancer; DNA methylation and mRNA per patient. (a) two genomic profiles per patient; rows patients, columns genes. (b) pairwise similarity matrices. (c) graphs from similarity; nodes patients, edge weights similarity. (c) single-color edges—each edge from one data type. (d) fusion iteration; two graphs learn from each other until convergence → (e) fused graph with edges from both types. A glioma dataset illustrates SNF with three genomic types: build three graphs, fuse, cluster with spectral clustering into subtypes, use the network as regularization in regression for survival risk. Similarity formula Graph fusion iteration Coding Split data in two parts

Compute similarity matrix from raw data

Fuse similarity matrices

Cluster fused matrix into two clusters

Visualize similarity networks

Binarize similarity matrix

On the model

Related article https://www.sciencedirect.com/science/article/abs/pii/S016926071930327X?via%3Dihub