Multi‐omics integration via similarity network fusion to detect subtypes of aging

Abstract

AbstractBackgroundMolecular subtyping of brain tissue provides insights into the heterogeneity of common neurodegenerative conditions, such as Alzheimer’s disease (AD). However, existing subtyping studies have mostly focused on single data modalities, such as RNA sequencing (RNAseq), which provide incomplete neurobiological information on pathological processes. To remedy this, we applied similarity Network Fusion (SNF), a method capable of integrating multiple high‐dimensional multi‐‘omics data modalities simultaneously.MethodWe analyzed human frontal cortex brain tissue samples characterized by five ‘omic modalities bulk RNAseq (18,629 genes), DNA methylation (53,932 cpg sites), histone H3K9 acetylation (26,384 peaks), tandem mass tag proteomics (7,737 proteins), and metabolomics (654 metabolites). SNF followed by spectral clustering was used for subtype detection, with subtype numbers determined by eigen‐gaps and dip‐test statistics. Normalized Mutual Information (NMI) was calculated to determine the contribution of each modality and feature to the fused network. Resulting subtypes were characterized by associations with 12 age‐related neuropathologies and cognitive performance.ResultFusion of all five data modalities (overlapping n = 111) yielded four molecular subtypes (nS1 = 32, nS2 = 26, nS3 = 31, nS4 = 22); S1 exhibited lower episodic memory performance than other subtypes proximal to death (t = 4.7, p = 1.5×10‐4). Histone acetylation (NMI = 0.32) and RNAseq (NMI = 0.16) contributed most strongly to this fused network; the top individual features were an acetylation peak in the promoter of CD200 and RNA abundance of PARP4. Secondary analysis fusing only RNAseq and histone acetylation (n = 520) yielded five subtypes which were correlated with the fully integrated subtypes (Fisher’s p = 5.0×10‐4) and strongly associated with AD neuropathology and episodic and semantic memory. Sensitivity analyses of all modality combinations and sample subsets found substantial influences of sample size and subtype number, but reinforced the importance of histone acetylation, RNAseq, and DNA methylation.ConclusionWe identified highly integrative molecular subtypes of aging derived from up to five multi‐‘omics data modalities simultaneously. These subtypes recapitulate some features of previous subtyping work in AD using single modalities, but also provide new molecular targets and shed light on the benefits and challenges of multi‐omic integration and individual subtyping in this field.