Negative frequency-dependent selection: a positive outlook with deep learning.

Balancing selection is a mode of natural selection that maintains genetic diversity through various mechanisms, including negative frequency-dependent selection (NFDS). However, distinguishing the genomic signature of NFDS from those of other balancing selection modes, such as overdominance, remains a significant challenge. In this perspective, we outline strategies to improve the modelling of genomic patterns expected under NFDS, with the goal of better differentiating them from signals of neutrality and alternative selection processes. We demonstrate how resource-efficient deep transfer learning, combined with novel data preprocessing and the modelling of genomic autocovariation, can effectively detect and characterize NFDS using either phased or unphased genotypes, and with or without temporal data from ancient DNA. Finally, we offer practical recommendations for both empiricists and method developers on advancing the detection of NFDS in genomic data. This article is part of the theme issue 'Exploring negative frequency dependent selection across levels: from genetics to ecology and back again'.