Data-driven inference of local behavioural rules predicts emergent properties of Phytophthora zoospore dispersal

Relating local behavioural rules to emergent population-level dynamics is a central challenge in quantitative biology. Data-driven agent-based models offer a powerful route to infer such rules directly from experimental observations, but their predictive value ultimately depends on whether the inferred behaviour remains valid beyond the conditions used for model construction. Here, we address this question using biflagellate zoospores of plant-pathogenic Phytophthora as an experimentally tractable model system of microswimmers. From nearly 60,000 freely swimming P. nicotianae trajectories, we inferred local behavioural rules governing speed and turning, including the temporal persistence of distinct movement states, and implemented them in generative agent-based simulations. An empirical two-state model reproduced the principal emergent properties of experimental dispersal, while independent hidden Markov modelling recovered a closely related behavioural organisation. We then challenged these inferred swimming rules in a distinct biological and physical context, using P. palmivora zoospores navigating an obstacle-filled environment. Without recalibrating the underlying swimming behaviour, the model reproduced emergent dispersal when combined with simple collision rules and identified transient post-collision slowdown as a key response required to account for the observed trajectories. These results show that local behavioural rules inferred from individual motion can retain predictive power beyond the conditions from which they were learned. More broadly, our framework provides a way to distinguish models that merely reproduce their training observations from those that capture transferable features of biological movement.