Models Constrained with Transcriptomics or Proteomics Data Generate Discordant Predictions

Modern high-throughput technologies have enabled detailed characterizations of an organism’s phenotype; this large-scale molecular data allows for a systems-level analysis of biological conditions. Moreover, genome-scale metabolic reconstructions (GENREs) can be leveraged to interpret these data by integrating ‘omics into a GENRE, resulting in a condition-specific models and the generation condition-specific predictions. Transcriptomic data have been most frequently used for data integration due to the datasets’ capacity for genome-wide coverage. However, regulation occurs at the translational level as well as transcriptional level, so proteomics-derived models may be used to generate more accurate condition-specific predictions than transcriptomic-derived models.

To investigate the functional differences between expression integration using either transcriptomic and proteomic data, matched datasets from five biological conditions (GSE65209 and PXD001659) were integrated into a GENRE (from Bosi, et al., doi: 10.1073/pnas.1523199113) using GIMME. As transcriptomic datasets have higher coverage than many proteomics datasets, the effect of data coverage was investigated by random sampling subsets of the data. Predictions (specifically, reaction essentiality) from models generated with matched transcriptomic and proteomic data (i.e. collected from the same biological condition) and mismatched datasets (i.e. from different conditions) were compared; surprisingly, these matched predictions were no more similar than predictions generated from mismatched models. Transcriptomic-derived models predicted more essential reactions than proteomics models, even when controlling for scope (i.e. number of enzymes constrained), indicating transcriptomic-derived models are more constrained than proteomics-derived models. These results emphasize the importance of interpreting predictions as condition-, datatype-, and approach-specific results. Future work will explore the effects of alternative integration approaches and datasets.