Predicting migratory survival in a songbird hybrid zone using machine learning.

Extrinsic postzygotic isolation-selection against hybrids between populations that have undergone divergent ecological adaptation-is hypothesized to play a central role in speciation but is notoriously difficult to demonstrate, as it requires evidence that ecologically relevant traits influence hybrid fitness. We addressed this challenge using individual radio tracking and machine learning in a hybrid zone between two songbirds where differences in seasonal migration are thought to serve as extrinsic isolating barriers. Using detection rates as a proxy for survival, we built a random forest classification model to predict survival from a set of genetic, morphological, and behavioural traits. The model achieved moderate overall accuracy (62%), with class-specific performance metrics that exceeded null expectations for birds that survived migration. Although no traits had consistent univariate effects on model prediction, the traits that contributed most to classification were migration year, sex, wing shape, and fall migratory orientation. When considering pairwise combinations of traits, the strongest effects were driven by interactions with migration year, suggesting that temporal variation strongly modulates the predictive contribution of all other variables. These results underscore the challenges of predicting complex and inherently stochastic ecological outcomes such as survival. Nonetheless, the traits most consistently associated with survival provide promising targets for future investigation, and our approach offers a roadmap for similar efforts to link phenotypic variation with fitness in natural populations.