Multiomics-Based Colorectal Cancer Molecular Subtyping Using Local Scaling Network Fusion

Abstract

Colorectal cancer (CRC) is a heterogeneous disease with distinct molecular properties. Tremendous works for CRC molecular subtyping are mainly based on gene expression profiling, which cannot capture the complementary information from other data types. Based on the classical multiomics data integration method similarity network fusion (SNF), which, however, suffers the trivial parameters setting, we developed local scaling SNF (Ls-SNF) that employs the local scaling method to construct patient affinity before network fusion. Local scaling infers the self-tuning of sample-to-sample distance and can eliminate the scaling problem. We have demonstrated the effectiveness of Ls-SNF on other cancer molecular subtyping in our previous study. In this study Ls-SNF applied in CRC molecular subtyping shows clear integrated patterns of gene expression, miRNA expression, and DNA methylation. Compared with the consensus molecular subtypes, subtypes identified by Ls-SNF achieved more significant association with clinical outcomes (p = 9.6 × 10⁻³, log-rank test). Certain mutations showed very significant enrichment in Ls-SNF subtypes, such as Class 3 were enriched for microsatellite instability (MSI) (p < 0.001), BRAF-mutant (p < 0.001), and CIMP high (p < 0.001). Ls-SNF subtypes also revealed better performance than some clinical risk factors in univariate and multivariate analyses (p = 0.002; p = 0.01).