Popformer: Learning general signatures of positive selection with a self-supervised transformer

Area of Science:

• Genetics
• Evolutionary Biology
• Bioinformatics

Background:

• Natural selection drives adaptive evolution and population differentiation.
• Detecting genomic signals of selection is challenging with current methods, which often lack generalization.
• Deep learning offers potential for robust selection detection.

Purpose of the Study:

• To develop a novel deep learning model, Popformer, for detecting general patterns of genetic variation and signatures of natural selection.
• To improve the generalization ability of selection detection methods across diverse evolutionary scenarios.
• To provide a new direction for population genetic inference.

Main Methods:

• Proposed Popformer, a transformer-based model with site-wise and haplotype-wise attention.
• Pre-trained Popformer on human genomic data using a masked language modeling analog.
• Fine-tuned Popformer for selection classification on simulated and real human genetic data.

Main Results:

• Popformer learns meaningful embeddings of genomic windows, reflecting population structure.
• The model accurately performs genotype imputation.
• Popformer demonstrates superior accuracy in selection classification compared to existing methods, generalizing from simulations to real data.

Conclusions:

• Popformer offers a powerful and generalizable approach for detecting natural selection signatures.
• The pre-training strategy on real data enhances model representation and robustness.
• This method opens new avenues for population genetic inference, including recombination rate and ancestry analysis.