Multi-omics analysis of adipogenesis

Understanding adipocyte development, also known as adipogenesis, is crit- ical for improving metabolic health. Additionally, adipose tissue depot, sex, age, body-mass index, Type 2 diabetes status, and cell size are known to play important roles in adipogenesis. Here, two classes of unsupervised learning methods were used to investigate which of these phenotypes were most strongly associated with differences in omics data between human adipocytes. Data consisted of two omics data types, including RNA- Sequencing (RNA-Seq) and Adipocyte Profiling (AP), as well as nine pa- tient phenotypes: day, depot, sex, age, body-mass index, Type 2 diabetes status, mean cell area, cell area variance, and % small cells. Data was clustered with a systematic pipeline using either single-view learning or multi-view learning methods, where single-view learning was performed using Principal Component Analysis (PCA) and multi-view learning was performed using either Multi-Omics Factor Analysis (MOFA+) or Mul- tiple Canonical Correlation Analysis (MCCA). Chi-square permutation tests were used to test for association between K-Means clusters and phe- notype labels and an extra layer of FDR-Correction was performed to account for multiple testing across all clustering analyses. Day (RNA-Seq PCA P-Value = 0.0092; AP PCA P-Value = 0.4224; MOFA+ P-Value = 0.0078; MCCA P-Value = 0.0092) and depot (RNA-Seq PCA P-Value = 0.0352; AP PCA P-Value = 0.0736; MOFA+ P-Value = 0.0450; MCCA P-Value = 0.0092) were identified as the primary drivers of clustering. However, none of the other phenotypes drove clustering, potentially due to unbalanced clusters or reduced statistical power from stratification. This multi-omics analysis of adipogenesis demonstrates the power of unsuper- vised clustering methods for the generation of robust biological insights that are consistent with the current literature surrounding day and depot. Nonetheless, since neither single-view nor multi-view learning methods de- tected associations for any of the other phenotypes, the inclusion of more samples and more omics is preferred for future studies