GenoBERT: A Language Model for Accurate Genotype Imputation.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Conventional genotype imputation methods are limited by ancestry bias and reduced accuracy for rare variants.
• Reference-panel based imputation struggles with diverse populations and rare variants.

Purpose of the Study:

• To introduce GenoBERT, a novel transformer-based, reference-free framework for genotype imputation.
• To evaluate GenoBERT's performance against existing methods across diverse datasets and missingness levels.

Main Methods:

• GenoBERT tokenizes phased genotypes and employs a self-attention mechanism to model linkage disequilibrium (LD).
• The framework was benchmarked on the Louisiana Osteoporosis Study (LOS) and 1000 Genomes Project (1KGP) datasets.
• Performance was assessed across various ancestry groups and genotype missingness levels (5-50%).

Main Results:

• GenoBERT achieved the highest overall imputation accuracy compared to four baseline methods.
• High imputation accuracy ($r^2 \approx 0.98$) was maintained at up to 25% missingness and robust performance ($r^2 > 0.90$) at 50% missingness.
• Consistent accuracy gains were observed across different ancestries, with resilience to small sample sizes and weak LD.

Conclusions:

• GenoBERT offers a scalable and robust solution for genotype imputation, eliminating reference-panel dependence.
• The framework provides a strong foundation for downstream genomic modeling, including genome-wide association and risk-prediction studies.
• GenoBERT demonstrates superior performance and broad applicability in genomic data analysis.