Hybrid autoencoder with orthogonal latent space for robust population structure inference.

Area of Science:

• Human Genetics
• Bioinformatics
• Computational Biology

Background:

• Accurate population structure and genomic ancestry analysis are crucial in human genetics.
• Existing methods often struggle with real-world data issues like laboratory artifacts, outliers, genotyping errors, missing data, and related individuals.
• High-quality genotype data is typically required for reliable inference.

Purpose of the Study:

• To introduce a novel hybrid method, SAE-IBS, for robust genomic ancestry inference.
• To address limitations of existing methods in handling poor quality data and relatedness.
• To provide an open-source program for practical application of the proposed method.

Main Methods:

• A hybrid approach combining matrix decomposition (e.g., PCA) and neural network (autoencoder) techniques.
• Development of SAE-IBS, yielding an orthogonal latent space for enhanced dimensionality selection and non-linear transformations.
• Creation of an accompanying open-source program for robust ancestry inference.

Main Results:

• SAE-IBS achieves higher accuracy than existing methods for projecting low-quality target samples onto a reference ancestry space.
• The method generates a robust ancestry space, even in the presence of related individuals.
• The resulting ancestry space allows for non-linear projections and exhibits clear separation between population groups.

Conclusions:

• SAE-IBS offers a significant advancement in robust genomic ancestry inference, particularly for challenging datasets.
• The method effectively handles missing data, genotyping errors, and relatedness, improving upon traditional approaches.
• The developed open-source tool facilitates wider application of accurate and robust population structure analysis.