Transfer transcriptomic signatures for infectious diseases (preprint)

ABSTRACT

The modulation of the transcriptome is among the earliest responses to infection, and vaccination. However, defining transcriptome signatures of disease is challenging because logistic, technical and cost factors limit the size and representativeness of samples in clinical studies. These limitations lead to poor performance of signatures when applied to new datasets or varying study settings. Using a novel approach, we leverage existing transcriptomic signatures as classifiers in unseen datasets from prospective studies, with the goal of predicting individual outcomes. Machine learning allowed the identification of sets of genes, which we name transfer transcriptomic signatures, that are predictive across diverse datasets and/or species (rhesus to humans) and that are also suggestive of activated pathways and cell type composition. We demonstrate the usefulness of transfer signatures in two use cases progression of latent to active tuberculosis, and severity of COVID-19 and influenza A H1N1 infection. The broad significance of our work lies in the concept that a small set of archetypal human immunophenotypes, captured by transfer signatures, can explain a larger set of responses to diverse diseases.