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

Area of Science:

• Population genetics
• Computational biology
• Genomics

Background:

• Detecting natural selection is crucial for understanding adaptive evolution.
• Deep learning methods for identifying selection signatures often lack generalizability across diverse evolutionary scenarios.
• Large-scale human genetic variation data necessitates advanced computational approaches.

Purpose of the Study:

• To develop a novel transformer-based model, Popformer, for learning general patterns of genetic variation.
• To improve the accuracy and generalizability of detecting natural selection signatures.
• To provide a new direction for population genetic inference methods.

Main Methods:

• Developed Popformer, a transformer-based model with site-wise and haplotype-wise attention.
• Incorporated relative positional embeddings for inter-SNP distances.
• Pre-trained the model using a masked language modeling objective on human genomic data, similar to genetic imputation.

Main Results:

• Popformer learns meaningful genomic embeddings that correlate with population structure.
• The model demonstrates high accuracy in genotype imputation.
• Fine-tuned Popformer shows superior accuracy in selection classification compared to existing methods, generalizing from simulations to real human data.

Conclusions:

• Popformer offers a powerful new approach for population genetic inference.
• The model's ability to generalize across diverse evolutionary scenarios and real data is a significant advancement.
• Future applications include inferring recombination rates, introgression, and local ancestry.