Transgenic fishes are nearing commercialization for aquaculture around the world. Farmed transgenic fish would likely escape from typical production facilities and interbreed with wild relatives. Models could help risk assessors predict the likelihood and consequences of transgene flow; however, predictions from models, such as the net fitness model, have not been confirmed using real populations of transgenic fish. This requires experiments in confined environments with conditions relevant to those fish might encounter in nature. We introduced transgenic Japanese medaka Oryzias latipes, bearing a salmonid growth hormone gene construct, and wild-type medaka into four environments: (A) high food availability, no predation; (B) high food availability, simulated predation; (C) low food availability, no predation; and (D) low food availability, simulated predation. We maintained 24 populations of wild-type and transgenic fish under these environments for approximately three generations and measured population size and transgene frequency at 210 days. We created deterministic and stochastic versions of a demographic simulation model, parameterized with fitness trait values collected under the same environmental conditions, to predict transgene frequency under each environment, and compared observed results to model predictions.

Results/Conclusions

In experimental populations, final transgenic population size was greater in Environment A than in all other Environments. The final transgene frequency in Environment A (0.332) was greater than that in Environment C or D. Both models predicted that transgene frequency in Environment A would be the highest, but also overestimated transgene frequency compared to observed results. Fecundity, fertility, juvenile viability, and mating advantage, when varied randomly over ranges observed in experiments, led to predicted transgene frequencies that overlapped with observations in Environments B and C but not in the more extreme Environments A and D. Our results illustrate the danger of measuring fitness traits in one sterile environment to parameterize a deterministic model, and use model outcomes to inform risk assessment decisions. We strongly recommend building uncertainty analysis into gene flow models, at least by incorporating parameter variability, and confirming model predictions with data collected under relevant environmental conditions before using such models to inform ecological risk assessments.