Neural ADMIXTURE for rapid genomic clustering

Abstract Characterizing the genetic structure of large cohorts has become increasingly important as studies extend to massive, diverse biobanks. Popular methods decompose individual genomes into fractional cluster assignments with each representing a vector DNA variant frequencies. However, rapidly increasing biobank sizes, these have computationally intractable. Here we present Neural ADMIXTURE, neural network autoencoder that follows same modeling assumptions current standard algorithm, while reducing compute time by orders magnitude surpassing even fastest alternatives. One month continuous using ADMIXTURE can be reduced just hours ADMIXTURE. A multi-head approach allows offer further acceleration computing multiple numbers in single run. Furthermore, models stored, allowing assignment performed on new data linear without needing share training samples.