VADEr: Vision Transformer-Inspired Framework for Polygenic Risk Reveals Underlying Genetic Heterogeneity in Prostate...

Area of Science:

• Genomics
• Artificial Intelligence
• Computational Biology

Background:

• Polygenic risk scores (PRSs) traditionally assess genetic liability but often assume feature independence.
• Complex interactions among genetic variants are crucial for accurate disease risk prediction.
• Deep learning architectures, particularly Transformers, excel at capturing feature dependencies.

Purpose of the Study:

• To introduce VADEr, a Vision Transformer-inspired architecture for genotype-to-phenotype prediction.
• To model local and global interactions within genetic data for enhanced disease risk assessment.
• To develop DARTH scores for interpretable identification of genetic risk drivers.

Main Methods:

• Developed VADEr, a deep learning model integrating NLP and computer vision techniques for genetic data analysis.
• Applied VADEr to predict prostate cancer (PCa) risk using genetic variant data.
• Formulated DARTH scores, an attention-based metric, to attribute risk contributions to specific genomic regions.

Main Results:

• VADEr significantly outperformed benchmark methods in predicting PCa risk across multiple metrics (accuracy, average precision, MCC).
• DARTH scores identified key PCa risk loci (e.g., HOXB13, TMPRSS2, MSMB) and revealed genetic heterogeneity.
• DARTH scores highlighted associations between specific loci (e.g., LMTK2) and PCa molecular subtypes (e.g., SPOP).

Conclusions:

• VADEr effectively captures genetic variant dependencies, advancing genotype-to-phenotype prediction for complex diseases.
• DARTH scores provide interpretable insights into personalized genetic risk factors.
• The VADEr-DARTH framework offers a promising direction for precision medicine in complex diseases like PCa.