A transparent and generalizable deep-learning framework for genomic ancestry prediction.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate genetic ancestry characterization is vital for genomic study reproducibility and fairness.
• Deep learning offers potential for advanced genetic data analysis.
• Existing methods may lack generalizability and interpretability.

Purpose of the Study:

• To develop a generalizable and interpretable deep-learning framework for genetic ancestry prediction.
• To assess model performance across diverse populations and with missing genetic data.
• To enhance transparency in ancestry prediction models.

Main Methods:

• Adaptation of the Diet Network deep-learning architecture for single-nucleotide polymorphism (SNP) data.
• Training on the Thousand Genomes Project dataset.
• Validation on CARTaGENE, Montreal Heart Institute, and All of Us biobanks.
• Application of attribution techniques (Saliency Maps, DeepLift, GradientShap, Integrated Gradients) for interpretability.

Main Results:

• The deep-learning model demonstrated strong generalizability across diverse biobanks.
• Predictions remained robust even with high proportions of missing SNPs.
• The model successfully inferred population structure for underrepresented groups (e.g., North Africans).
• Attribution techniques confirmed predictions were driven by biologically relevant SNPs.

Conclusions:

• A generalizable and interpretable deep-learning framework for genetic ancestry inference in large biobanks has been presented.
• This framework facilitates broader genetic ancestry characterization, supporting inclusive biomedical applications.
• The study provides practical tools for integrating genetic data into healthcare solutions.