Machine learning tools for deciphering the regulatory logic of enhancers in health and disease

Transcriptional enhancers are DNA regulatory elements that control the levels and spatiotemporal patterns of gene expression during development, homeostasis, and pathophysiological processes. Enhancer identification and characterization at the genome-wide scale rely on their structural characteristics, such as chromatin accessibility, binding of transcription factors and cofactors, activating histone modifications, 3D interactions with other regulatory elements, as well as functional characteristics measured by massively parallel reporter assays and sequence conservation approaches. Recently, machine learning approaches and particularly deep learning models (Enformer, BPNet, DeepSTARR, etc.) allow the prediction of enhancers, the impact of variants on their activity and the inference of transcription factor binding sites, leading, among others, to the construction of the first completely synthetic enhancers. We present the above computational tools and discuss their diverse applications towards cracking the enhancer regulatory code, which could have far-reaching ramifications for uncovering essential regulatory mechanisms and diagnosing and treating diseases. With an emphasis on sepsis, a leading cause of morbidity and mortality in hospitalized patients, we discuss computational approaches to identify sepsis-associated endotypes, circuits, and immune cell states and signatures characteristic of this condition, which could aid in developing novel therapies.