Integrative Multi-Omic Analysis of CCLE Cell Lines

We used DRAGON to compute partial correlations between multi-omic data of CCLE cell lines. In particular, we computed partial correlations between the four following data type pairs across all CCLE cell lines: (1) miRNA levels and gene knockout screens, (2) protein levels and metabolite levels, (3) cell viability assays after drug exposure and gene knockout screens, and (4) TF targeting and metabolite levels. For each association, the final number of cell line samples is the intersection of the cell lines for each modality. DRAGON builds a GGM that implements covariance shrinkage with tuning parameters specific to each biological layer or “ome,” represented by a different data type, a novel addition to covariance shrinkage that enables DRAGON to account for varying data structures and sparsity of different multi-omic layers [52 (link)]. The magnitude of DRAGON partial correlation values may not be always interpretable without a reference because they are derived from a regularized, shrunken covariance matrix [98 (link)]. All variables were standardized to have a mean of 0 and a standard deviation of 1 before running DRAGON.
To compute associations between protein levels and metabolite concentrations, we averaged protein isoform levels to reduce the set of 12,755 measured proteins to 12,197 unique proteins. The final number of samples used to compute this association represented 258 cells shared between the 375 cells for proteomics data and 928 cells for metabolomic data. To compute associations between LDH levels and its substrate lactate, and because the LDH isozymes (LDHA and LDHB) catalyze opposite biochemical reactions, we created two new variables in the DRAGON network accounting for the ratio between isozymes: $${\displaystyle \begin{array}{c}{\text{LDHA}}_{\text{normalized}}=1_{\left[\frac{\text{LDHA}}{\text{LDHB}},>,1\right]}.\frac{\text{LDHA}}{\text{LDHB}}\\ {\text{LDHB}}_{\text{normalized}}=1_{\left[\frac{\text{LDHB}}{\text{LDHA}},>,1\right]}.\frac{\text{LDHB}}{\text{LDHA}}\end{array}}$$ where LDHA and LDHB represent protein levels of LDH isozymes. This normalization reflects our understanding of the nonlinear relation between the ratio of LDHA/LDHB and lactate concentrations: when LDHA is dominant, LDH produces lactate; therefore, we expect a positive correlation with lactate levels, and conversely, when LDHB is dominant, lactate is a substrate for LDH and the correlation should be negative. We did not include pyruvate concentrations because it was not among the measured metabolites in CCLE.