Coalescence and translation: A language model for population genetics

Area of Science:

• Population genetics
• Computational biology
• Machine learning

Background:

• Traditional population genetic inference relies on specialized models like the sequentially Markovian coalescent, which have limitations in scalability and predefined assumptions.
• Recent advancements in deep learning and genomic simulation offer a novel approach to inferring evolutionary processes directly from data.

Purpose of the Study:

• To develop a novel deep learning framework, cxt, for inferring coalescence times and population evolutionary history.
• To reframe coalescence time inference as a translation problem between genomic mutation patterns and the ancestral recombination graph.

Main Methods:

• Developed cxt, a decoder-only transformer model inspired by large language models.
• cxt autoregressively predicts coalescent events conditioned on local mutational context.
• Trained the model on extensive simulations from the stdpopsim catalog.

Main Results:

• cxt demonstrates competitive performance against state-of-the-art Markov Chain Monte Carlo (MCMC) methods in population genetic inference.
• The model achieves high accuracy on simulated data and shows robust generalization to out-of-distribution scenarios.
• cxt enables efficient, large-scale inference, generating millions of predictions rapidly and providing well-calibrated approximate posteriors for uncertainty quantification.

Conclusions:

• cxt offers a powerful, scalable, and accurate alternative to traditional methods for population genetic inference.
• The model successfully applies to complex empirical population genomic data from humans and mosquitoes.
• This deep learning approach holds potential for further advancements through fine-tuning and broader applications in evolutionary biology.