Time and memory usage by SCNIC
To evaluate the memory resources needed by SCNIC, we ran the SCNIC modules step locally on a 2015 MacBook Pro with 16 GB RAM with a 2.5 GHz Quad-Core Intel Core i7 processor for both the Great Lakes dataset and an integrated microbiome-metabolome dataset with 1,301 features, which will be referred to as the GT dataset. The runtime was recorded across 3 runs per method (SMD vs LMM) for each dataset using GNU Time, and memory was profiled using memory-profiler 0.60.0. The “within” step, which calculates correlations between features and creates the correlation network was not tested because it depends greatly on the correlation metric used, and the runtime and memory usage of FastSpar (likely the most computationally intensive correlation metric to be used in this step) have already been profiled[39]. The modules step only utilizes a correlation matrix and as such does not scale with the number of samples, only the number of features, except when the values of the count table are being summed, which is a generally trivial calculation compared to the module generation step.
RESULTS