Emergent populations derived with unsupervised learning of human whole genomes

Abstract

Artificial intelligence (AI) holds great promise to precisely classify human ancestry and the genetic causes of complex diseases. I have constructed an unsupervised machine learning paradigm that examines the whole genome as a hyper-dense, nonlinear, multidimensional feature space. The AI system culminates in 26 neural network neurons each sensitive to a specific heritage that can identify an individual’s component genetic heritages with a top-5 error of <0.5%. Importantly, I observed some populations previously thought to belong to single stratum are composed of multiple strata – for instance Japan is defined as a uniform population using previous methods. I found that the Japanese individuals segregate to two very distinct populations. This work represents an essential step towards understanding the genetic background of patients to enable precision medicine causal disease gene identification.