Leveraging hierarchical structures for genetic block interaction studies using the hierarchical transformer.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Epistasis, or genetic variant interaction, describes how one genetic variant can mask the effect of another.
• Understanding epistasis is crucial for complex diseases, as combined genetic effects within blocks offer greater analytical power than single markers.
• Current methods struggle to integrate gene structure modeling with interaction learning for efficient gene interaction searching.

Purpose of the Study:

• To develop an end-to-end neural model for effective gene interaction searching.
• To enhance the discovery of epistasis by modeling genetic blocks and their interactions.
• To improve the computational power and reduce noise in genetic association studies.

Main Methods:

• Developed a neural genetic block interaction searching model.
• Augmented a hierarchical transformer architecture to model genetic blocks.
• Employed a hierarchical attention mechanism for cross-block relationship mapping.

Main Results:

• The model effectively processes large SNP chip inputs and outputs genetic block interaction heatmaps.
• Demonstrated substantial improvements over traditional exhaustive searching and existing neural network methods.
• Validated performance on both simulation data and UK Biobank studies.

Conclusions:

• The proposed neural model offers a powerful approach for discovering complex genetic interactions.
• This method advances the field of genetic association studies by effectively leveraging genetic block information.
• The model provides a significant improvement in identifying epistasis relevant to complex diseases.