Abstract
Group-aggregated responses to tuberculosis (TB) have been well characterized on a molecular level. However, human beings differ and individual responses to infection vary. We have combined a novel approach to individual gene set analysis (GSA) with the clustering of transcriptomic profiles of TB patients from seven datasets in order to identify individual molecular endotypes of transcriptomic responses to TB. We found that TB patients differ with respect to the intensity of their hallmark interferon (IFN) responses, but they also show variability in their complement system, metabolic responses and multiple other pathways. This variability cannot be sufficiently explained with covariates such as gender or age, and the molecular endotypes are found across studies and populations. Using datasets from a Cynomolgus macaque model of TB, we revealed that transcriptional signatures of different molecular TB endotypes did not depend on TB progression post-infection. Moreover, we provide evidence that patients with molecular endotypes characterized by high levels of IFN responses (IFN-rich), suffered from more severe lung pathology than those with lower levels of IFN responses (IFN-low). Harnessing machine learning (ML) models, we derived gene signatures classifying IFN-rich and IFN-low TB endotypes and revealed that the IFN-low signature allowed slightly more reliable overall classification of TB patients from non-TB patients than the IFN-rich one. Using the paradigm of molecular endotypes and the ML-based predictions allows more precisely tailored treatment regimens, predicting treatment-outcome with higher accuracy and therefore bridging the gap between conventional treatment and precision medicine.