Genetic distinction between contiguous urban and rural multimammate mice in Tanzania despite gene flow.

Special conditions are required for genetic differentiation to arise at a local geographical scale in the face of gene flow. The Natal multimammate mouse, Mastomys natalensis, is the most widely distributed and abundant rodent in sub-Saharan Africa. A notorious agricultural pest and a natural host for many zoonotic diseases, it can live in close proximity to humans and appears to compete with other rodents for the synanthropic niche. We surveyed its population genetic structure across a 180-km transect in central Tanzania along which the landscape varied between agricultural land in a rural setting and natural woody vegetation, rivers, roads and a city (Morogoro). We sampled M. natalensis across 10 localities and genotyped 15 microsatellite loci from 515 individuals. Hierarchical STRUCTURE analyses show a K-invariant pattern distinguishing Morogoro suburbs (located in the centre of the transect) from nine surrounding rural localities. Landscape connectivity analyses in Circuitscape and comparison of rainfall patterns suggest that neither geographical isolation nor natural breeding asynchrony could explain the genetic differentiation of the urban population. Using the isolation-with-migration model implemented in IMa2, we inferred that a split between suburban and rural populations would have occurred recently (<150 years ago) with higher urban effective population density consistent with an urban source to rural sink of effective migration. The observed genetic differentiation of urban multimammate mice is striking given the uninterrupted distribution of the animal throughout the landscape and the high estimates of effective migration (2Ne M = 3.0 and 29.7), suggesting a strong selection gradient across the urban boundary.

INDIVIDUAL-BASED MODELS: POWERFUL OR POWER STRUGGLE?

Individual-based models (IBMs) offer endless possibilities to explore various research questions but come with high model complexity and computational burden. Large-scale IBMs have become feasible but the novel hardware architectures require adapted software. The increased model complexity also requires systematic exploration to gain thorough system understanding. We elaborate on the development of IBMs for vaccine-preventable infectious diseases and model exploration with active learning. Investment in IBM simulator code can lead to significant runtime reductions. We found large performance differences due to data locality. Sorting the population once, reduced simulation time by a factor two. Storing person attributes separately instead of using person objects also seemed more efficient. Next, we improved model performance up to 70% by structuring potential contacts based on health status before processing disease transmission. The active learning approach we present is based on iterative surrogate modelling and model-guided experimentation. Symbolic regression is used for nonlinear response surface modelling with automatic feature selection. We illustrate our approach using an IBM for influenza vaccination. After optimizing the parameter spade, we observed an inverse relationship between vaccination coverage and the clinical attack rate reinforced by herd immunity. These insights can be used to focus and optimise research activities, and to reduce both dimensionality and decision uncertainty.