Detecting adaptive and ghost introgression in human populations using machine learning approaches

It is now widely recognized that introgression from archaic humans left a lasting impact on the modern human genome, yet many questions remain unanswered regarding how introgressed variants contributed to adaptation and whether additional, unsampled archaic hominins also contributed to the modern human gene pool. In this talk, I will present recent works from my lab to study both adaptive and ghost introgression in humans. First, I will discuss a work on adaptive introgression in modern Peruvian populations. We developed a multi-layered validation framework and identified 20 high confidence adaptive introgression candidates inherited from both Neanderthals and Denisovans. Reconstruction of selection history revealed at least two temporally distinct waves of adaptive introgression, one overlapping with the peopling of the Americas and another occurring around the Andean agricultural transition. Candidate loci were enriched in pathways related to immune signaling, hypoxia, and lipid metabolism, with CXCR4 emerging as a compelling candidate of recent immune-related adaptive introgression. In the second half of the talk, I will present a scalable machine learning framework for detecting “ghost” or super-archaic introgression. Instead of explicitly reconstructing genealogies, our approach leverages statistical proxies of local coalescence and neural network models to distinguish ghost introgression from alternative demographic processes. Application of this framework to Oceanian and Tibetan genomes revealed widespread shared signals of deep archaic ancestry, including enrichment near the HLA locus and suggestive signals overlapping the EPAS1 gene.