Features extracted using tensor decomposition reflect the biological features of the temporal patterns of human blood multimodal metabolome

Abstract

AbstractHigh-throughput omics technologies have enabled the profiling of entire biological systems. For the biological interpretation of such omics data, two analyses, hypothesis- and data-driven analyses including tensor decomposition, have been used. Both analyses have their own advantages and disadvantages and are mutually complementary; however, a direct comparison of these two analyses for omics data is poorly examined.We applied tensor decomposition (TD) to a dataset representing changes in the concentrations of 562 blood molecules at 14 time points in 20 healthy human subjects after ingestion of 75 g oral glucose. We characterized each molecule by individual dependence (constant/variable) and time dependence (sustained/transient). Three of the four features extracted by TD were characterized by our previous hypothesis-driven study, indicating that TD can extract some of the same features obtained by hypothesis-driven analysis in a non-biased manner. In contrast to the years taken for our previous hypothesis-driven analysis, the data-driven analysis in this study took days, indicating that TD can extract biological features in a non-biased manner without the time-consuming process of hypothesis generation.Author SummaryFor biological interpretation of lage-scale omics data, two analyses, hypothesis-driven analysis and data-driven analysis including tensor decomposition, have been used. These two analyses have their own advantages and disadvantages, and are mutually complementary. However, the direct comparison between these two analyses for omic data is poorly examined. In this study, we applied tensor decomposition to a dataset representing temporal changes in the human 562 blood molecules as data-driven analysis and extracted three features. We have previously analyzed the same data by hypothesis-driven analysis (Fujita et al., 2022). The three features extracted by the tensor decomposition are the same features extracted by the hypothesis-driven analysis, indicating that the tensor decomposition can extract the features in an unbiased manner. Although the same features can be extracted by the tensor decomposition and hypothesis-driven analysis, hypothesis-driven analysis in our earlier study took years (Fujita et al., 2022), while feature extraction by tensor decomposition took only days in this study. Thus, tensor decomposition can extract biological features in a non-biased manner without time-consuming process of hypothesis generation. We propose that tensor decomposition can be the first choice for analysis of omic data.