Semi-supervised detection of natural selection with positive-unlabeled learning.

Identifying genomic regions shaped by natural selection is a central goal in evolutionary genomics. Existing machine learning methods for this task are typically trained using simulated genomic data labeled according to specific evolutionary scenarios. While effective in controlled settings, these models are limited by their reliance on explicit class labels. They can only detect the specific processes they were trained to recognize, making it difficult to interpret predictions for regions influenced by other evolutionary forces. This limitation is especially problematic when analyzing empirical genomes shaped by a mixture of adaptive, demographic, and ecological factors. One-vs.-rest strategies offer a potential alternative, but suffer from the inherent complexity of modeling all other evolutionary and demographic processes as a catch-all "rest" class. Here, we explore positive-unlabeled learning as a more flexible framework for detection of adaptive events. Positive-unlabeled learning is a semi-supervised approach that permits identification of samples of a target class using only positive labels and an unlabeled background, without requiring explicit modeling of negative samples. To assess the utility of this approach, we focus on a binary classification setting for detecting selective sweeps (positive samples) arising from positive natural selection against a mixed background composed of both unlabeled sweeps and neutrally-evolving regions. To accomplish this goal, we introduce , a method that employs only a set of labeled sweep observations for training while treating all remaining data as unlabeled. By avoiding assumptions about the composition of the background, enables robust sweep discovery in realistic genomic landscapes. We systematically evaluate its performance across a range of demographic, adaptive, and confounding contexts, including domain shift arising from misspecified demographic models, and find that delivers high performance and generalizability. To demonstrate a practical application of , we analyzed European and Bengali in Bangladesh human genomes, treating the empirical genomes as unlabeled sets, and recapitulating several previously-identified sweep candidates. Our results show that provides a powerful and versatile alternative for detecting adaptive regions, with the potential to generalize across a range of genomic landscapes.