Fine Tuning Thresholds to Facilitate Integration of Transcriptomics Data

Integration of omics datatypes is becoming increasingly commonplace in the field of metabolic modeling. Among all of the omics datatypes, transcriptomics is the most chosen for integration with genome-scale metabolic models (GEMs) resulting in a wide array of algorithms. These algorithms are often preceded by preprocessing the transcriptomics data to generate list of high confidence genes or reactions. Multiple studies in the past have identified parameters related to preprocessing to be important, specifically thresholding of gene expression profile. However, it is not clear how thresholds should be decided. Here, in our analysis we look at tissue-specific transcriptomic data from HPA and GTex and provided novel insights on what factors could determine thresholds and what are the largest sources of false positive and false negative predictions. Our analysis also provides novel insights on gene expression patterns of tissue-specific and housekeeping genes. Further, we also suggest improvements, validate using proteomics data and metabolic tasks. We also applied our thresholding method to cancer cell-line transcriptomics data sets. Thus, we demonstrate better applicability of thresholds across different data sets.