Large sample size and nonlinear sparse models outline epistatic effects in inflammatory bowel disease

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Polygenic diseases exhibit complex nonlinear interactions at the molecular level.
• Current genotype-to-phenotype modeling predominantly uses linear models, despite evidence of nonlinearity.
• Genetic data is inherently underdetermined due to the vast number of variants and challenges in cohort collection.

Purpose of the Study:

• To investigate the performance of nonlinear models, specifically neural networks, in genotype-to-phenotype modeling for polygenic diseases.
• To address the underdetermination issue in genetic data through advanced modeling techniques.
• To explore the potential of neural networks to capture epistatic effects in disease pathogenesis.

Main Methods:

• Development and application of a biologically meaningful sparsified neural network architecture.
• Training and validation using whole exome sequencing data for inflammatory bowel disease case-control prediction.
• Comparison of neural network performance against traditional additive models.

Main Results:

• A neural network model demonstrated superior predictive performance compared to additive approaches in inflammatory bowel disease prediction.
• The study provided empirical evidence supporting the presence of both positive and negative epistatic effects in inflammatory bowel disease.
• Controlled complexity and sufficient training data were key to the neural network's success.

Conclusions:

• Underdetermination in genetic data is a significant factor contributing to the perceived optimality of additive models in clinical genetics.
• Nonlinear models, such as neural networks, offer a more powerful approach for genotype-to-phenotype modeling when properly implemented.
• This work advances the understanding and modeling of complex genetic architectures in polygenic diseases.