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

Motile microorganisms explore complex environments in search of nutrients, hosts and favourable ecological niches. Plant-pathogenic oomycetes, for instance, undergo such an exploratory phase through biflagellate zoospores that actively swim through water-filled soil pores before infecting host tissues. Linking individual zoospore swimming behaviour to emergent dispersal remains challenging. Here, we present an end-to-end, data-driven framework that transforms time-lapse microscopy image sequences into generative agent-based simulations of zoospore dispersal by inferring local behavioural rules directly from experimental trajectories. Using Phytophthora nicotianae as a model system, we isolated nearly 60,000 zoospore trajectories and quantified both local behavioural descriptors and emergent trajectory properties. Local behavioural measurements were first used to infer an empirical two-state model distinguishing SLOW and FAST swimming regimes while capturing temporal memory and the coupling between speed and turning. Implemented within an agent-based cellular automaton, this model reproduced the principal emergent properties of experimental dispersal. We then independently inferred the behavioural organisation of zoospore swimming using hidden Markov models. The most parsimonious two-state HMM recovered a closely related behavioural organisation, while revealing that the inferred states jointly reflected swimming speed, turning dynamics and directional persistence rather than speed alone. Finally, we challenged the inferred behavioural rules in an independent obstacle-filled environment. Combined with simple collision hypotheses, the model reproduced emergent dispersal without recalibrating the swimming rules and identified transient post-collision slowdown as a key response required to account for the experimental trajectories. Together, these results demonstrate that experimentally inferred local behavioural rules possess predictive power beyond the conditions used for their calibration. More broadly, this work establishes a predictive framework linking quantitative microscopy, behavioural-rule inference and generative modelling of microbial dispersal.