Dimensionality reduction of genetic data using contrastive learning.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Dimensionality reduction is crucial for visualizing complex genetic datasets.
• Existing methods like PCA, t-SNE, and UMAP have limitations in preserving both local and global genetic data structures.
• Deep learning offers potential for novel dimensionality reduction approaches in genomics.

Purpose of the Study:

• To introduce a novel contrastive learning framework for dimensionality reduction on genetic data.
• To create PCA-like population visualizations that better preserve data structure and generalize to new data.
• To develop tailored loss functions and data augmentation strategies for SNP genotype datasets.

Main Methods:

• Utilized contrastive learning, a self-supervised deep learning technique.
• Developed a custom loss function outperforming standard contrastive learning losses.
• Implemented a data augmentation scheme specifically for SNP genotype data.
• Compared the framework against Principal Component Analysis (PCA), t-SNE, UMAP, and popvae.

Main Results:

• The proposed method demonstrated strong preservation of global genetic data structure.
• Achieved superior generalization properties compared to t-SNE, UMAP, and popvae.
• Maintained a high degree of relative distance preservation between individuals.
• Successfully performed population classification on dog and human genotype datasets.

Conclusions:

• The contrastive learning framework provides an effective approach for dimensionality reduction and visualization of genetic data.
• The method offers advantages in generalization and structure preservation over contemporary techniques.
• The deep learning framework allows for projecting new samples and fine-tuning models, incorporating domain-specific information.