Deep Recovery of Effective Migration Surfaces

dreems, a graph neural network for effective migration surfaces, achieves roughly 10–30% lower estimation error than a likelihood-based method on unseen coalescent simulations, especially with sparse sampling.

• Why it matters: Likelihood-based EMS fitting is computationally demanding, limiting exploratory use of spatial IBD patterns when barriers or habitat differences vary across landscapes.
• What they did: The method maps genetic covariance and sampling locations to an EMS and outputs predictive uncertainty, using amortized inference so it can handle different graph sizes and sampling designs without re-training.
• The result: In independent simulations not used for training, dreems reduces estimation error by 10–30% versus a state-of-the-art likelihood approach, and reanalyzing published datasets matches earlier spatial patterns while cutting computation time by orders of magnitude.

AI summary of the abstract · may contain errors · not medical advice · read the paper.