TLS-Tractor: A transfer learning framework for incorporating summary-statistics into local ancestry-aware GWAS in admixed populations

TLS-Tractor is a new transfer learning method designed for genetic association analysis in recently admixed populations. The method combines external GWAS summary statistics with individual-level data using the generalized method of moments, overcoming limitations of the original Tractor method. In simulations, TLS-Tractor correctly controlled type I error, accurately estimated ancestry-specific effects, and increased statistical power compared to Tractor using only internal data. Analyses with data from African-European participants from the All of Us project combined with Million Veteran Program statistics confirmed these advantages. Local ancestry adjustment improved calibration, localization, and interpretation of results. The new tlstractor R package is 200 times faster at extracting local ancestry segments and 4 to 32 times faster at association testing compared to the original implementation.