Domain-adaptive neural networks improve supervised machine learning based on simulated population genetic data.

Area of Science:

• Population genetics
• Machine learning
• Computational biology

Background:

• Supervised machine learning methods are powerful for population genetic inference using simulated data.
• These methods can fail when simulated training data does not match real-world data (simulation mis-specification).

Purpose of the Study:

• To frame simulation mis-specification as a domain adaptation problem.
• To apply domain adaptation techniques to mitigate the effects of simulation mis-specification in population genetic models.
• To improve the accuracy of population genetic inference methods.

Main Methods:

• Framed simulation mis-specification as a domain adaptation problem.
• Applied a gradient reversal layer (GRL) based domain adaptation technique.
• Tested the approach on two deep-learning methods: SIA (for positive selection inference) and ReLERNN (for recombination rate inference).

Main Results:

• Domain adaptation substantially mitigated the effects of simulation mis-specification.
• The domain-adaptive framework compensated for ancestral recombination graph (ARG) inference error in SIA.
• The domain-adaptive SIA (dadaSIA) model yielded improved selection coefficient estimates in the 1000 Genomes CEU population.

Conclusions:

• Domain adaptation is an effective strategy to address simulation mis-specification in population genetics.
• The developed domain-adaptive framework enhances the reliability of machine learning models in population genetics.
• This approach is expected to be widely applicable in the growing field of machine learning for population genetics.