Temperature dependence of mosquitoes: comparing mechanistic and machine learning approaches

Mosquitoes that can transmit viral pathogens of humans (e.g., Aedes, Anopheles, and Culex genera) are of increasing concern for global public health. These vectors not only cause a rapidly growing infectious disease burden (e.g., malaria, dengue, Zika, chikungunya, yellow fever), but are also spreading in geographic range under climate change. As small-bodied ectotherms, mosquitoes are strongly affected by temperature, and prior laboratory research has demonstrated nonlinear and unimodal relationships between temperature and mosquito life history traits, with upper and lower thermal limits and intermediate thermal optima as predicted by first principles of enzyme kinetics and the metabolic theory of ecology. However, it remains unknown how mosquito thermal responses measured in laboratory experiments relate to the realized thermal responses of mosquitoes in the field.

To address these questions, we leverage thousands of global mosquito occurrences and geospatial satellite rasters at high temporal and spatial resolution to construct machine-learning based species distribution models. We apply new methods to restrict models to the temporally-relevant mosquito activity season and to ecologically-relevant spatial background sampling for comparison to mosquito occurrence records. Thermal response signatures from these ‘top-down’ statistical models supported the nonlinear temperature responses apparent in the laboratory-derived mosquito abundance estimates. Further, we found that, across mosquitoes, thermal minima estimated from laboratory studies were highly correlated with those estimated from field-based models. However, we could not estimate thermal maxima from field occurrence records due to inherent limitations in background sampling, and comparisons of thermal optima were inconclusive. Together, the current evidence suggests that laboratory studies have the potential to be highly transportable to real-world mosquito inferences for thermal minima, while other key temperature dependence metrics merit further investigation.