Coalescence and Translation: A Language Model for Population Genetics.

Area of Science:

• Population genetics
• Computational biology
• Machine learning

Background:

• Probabilistic models like the sequentially Markovian coalescent (SMC) are used for population genetic inference but have limitations in scalability and predefined assumptions.
• Recent advances in deep learning and simulation offer a new approach to infer evolutionary processes from synthetic genetic data.

Purpose of the Study:

• To reframe the inference of coalescence times as a translation problem between genomic mutation patterns and the ancestral recombination graph (ARG).
• To develop and evaluate a deep learning model for scalable and robust population genetic inference.

Main Methods:

• Developed cxt, a decoder-only transformer model inspired by large language models.
• Trained cxt on synthetic genetic data from the stdpopsim catalog.
• Evaluated cxt's performance against state-of-the-art MCMC-based likelihood models.

Main Results:

• cxt performs comparably to state-of-the-art methods across diverse demographic scenarios, including out-of-distribution settings.
• The model demonstrates robust generalization and enables efficient, large-scale inference, generating millions of predictions rapidly.
• cxt provides well-calibrated approximate posterior distributions for uncertainty quantification.

Conclusions:

• cxt offers a flexible and scalable deep learning approach for inferring genealogical history from genomic data.
• This work bridges deep learning and coalescent theory, moving towards a foundation model for population genetics.
• The model's ability to handle diverse scenarios and provide uncertainty estimates enhances its utility in population genetic studies.