Uncertainty quantification in variable selection for genetic fine-mapping using bayesian neural networks.

Area of Science:

• Statistical genetics
• Machine learning
• Computational biology

Background:

• Variable selection is crucial for understanding complex traits.
• Existing methods struggle with nonlinear genetic effects and discrete phenotypes.
• Quantifying uncertainty in variable selection remains a challenge.

Purpose of the Study:

• To develop a novel Bayesian neural network framework for robust variable selection.
• To generalize existing regression frameworks by incorporating nonlinear genotypic data and discrete phenotypes.
• To improve the identification of true effect variables in complex traits.

Main Methods:

• Developed an "ensemble of single-effect neural networks" (ESNN) framework.
• Utilized Bayesian neural networks to model nonlinear genetic structures (e.g., dominance effects).
• Extended capabilities to model discrete phenotypes (e.g., case-control studies).

Main Results:

• ESNN produces calibrated posterior summaries, including credible sets and posterior inclusion probabilities.
• Demonstrated superior performance in simulations, especially for traits with non-additive variation from correlated variants.
• Real-data application showed ESNN improves upon state-of-the-art methods for identifying true effect variables.

Conclusions:

• The ESNN framework offers a powerful and flexible approach for variable selection in statistical genetics.
• ESNN effectively handles complex genetic architectures, including non-additive effects and discrete outcomes.
• This method advances the identification of genetic variants underlying complex traits.