Deep Recovery of Effective Migration Surfaces

The isolation-by-distance principle suggests that genetic relatedness decreases as geographic distance increases. While its existence has been confirmed in natural populations, recent studies have found that the strength of this relationship can vary due to barriers, habitat compatibility or other factors. Effective migration surfaces, which serve as a visual representation of spatial patterns in isolation-by-distance, are popular for analyzing these patterns.

However, fitting likelihood-based EMS models can be computationally intensive, necessitating repeated large-scale matrix calculations. The authors introduce Deep Recovery of Effective Migration Surfaces (dreems), a graph neural network designed to efficiently estimate effective migration surfaces. This innovative approach maps genetic covariance and sampling locations directly to an effective migration surface, providing predictive uncertainty estimates simultaneously.

Once trained, the network can adapt to different graph sizes and sampling designs without the need for retraining. When applied to independent coalescent simulations not seen during training, dreems demonstrated roughly 10-30% lower estimation error compared to a leading likelihood-based method, with the greatest improvements observed in sparse sampling designs frequently encountered in real-world scenarios.

Furthermore, a reanalysis of several published datasets using dreems recovered spatial patterns that were broadly consistent with earlier likelihood-based estimates, all while significantly reducing computation time by orders of magnitude.