NSFDEB-NERC: Machine learning tools to discover balancing selection in genomes from spatial and temporal autocorrelations

Overview Balancing selection is a key driver of adaptive evolution that maintains variation within and across species. Yet, despite a few notable examples, little is known about recent or fleeting balancing selection, likely because its genomic footprints are difficult to distinguish from those of neutral evolution. Whereas

the renaissance of artificial intelligence has transformed how we search for adaptive genomic regions, the characterization of recent balancing selection from contemporary genomic data remains difficult. Moreover, data obtained from historical samples and nonmodel organisms are often fraught with technical

hurdles that available methods are ill-equipped to navigate. Hence, the objective of this proposal is to develop a suite of deep learning tools for studying recent and transient balancing selection from temporally and spatially sampled genomic data, while ensuring that these methods account for the many

technical challenges encountered in nonmodel study systems. Specifically, we will design predictors for detecting signals and learning parameters of balancing selection from incomplete, low-quality, and unphased ancient samples (Aim 1), using approaches that circumvent the uncertainty in genetic and

demographic parameters (Aim 2), and that extend to data generated by cost-efficient pooled sequencing strategies (Aim 3). We will deploy these new tools to three empirical datasets, which respectively encompass a set of recent balancing selection case studies for which the technical issues tackled by

Aims 1, 2, and 3 are designed to overcome. Preliminary findings support the promise and feasibility of the proposed aims, and we expect these tools to provide the evolution community with a powerful framework to address currently unanswerable questions about adaptation in both model and nonmodel systems.

Intellectual Merit Elucidating the processes underlying adaptive maintenance of variation within species is a fundamental problem in evolutionary biology, and one for which available tools are ill-equipped to address from the vast, often non-ideal, data that exist for nonmodel study systems. The PIs have demonstrated success in

designing statistical and machine learning methods for uncovering footprints of adaptation and addressing targeted hypotheses about balancing selection across several study systems. Thus, they are well-poised to develop the proposed deep learning approaches for studying balancing selection, by leveraging genomic, spatial, and temporal autocorrelations across a variety of data types characteristic of

those from nonmodel organisms. Availability of these methods will facilitate studies of balancing selection when data are incomplete, low-quality, and unphased (Aim 1), when genetic and demographic parameters are uncertain (Aim 2), and when genotype information at the individual level is unavailable

(Aim 3). Moreover, our proposed applications of these methods to a diversity of study systems will address questions regarding the roles and specific modes of balancing selection at different temporal and geographic scales. Finally, the developed tools will be applicable to a wide range of data types common

across model and nonmodel organisms, empowering future studies of adaptation by removing barriers imposed by limitations of data quality and current knowledge of demographic history.