Deep learning captures the effect of epistasis in multifactorial diseases

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Disease Risk Prediction

Background:

• Polygenic risk scores (PRS) traditionally use linear models assuming independent SNP contributions.
• Complex diseases involve non-linear gene-gene interactions (epistasis) often missed by linear models.
• Accurate prediction of multifactorial diseases requires methods that capture these complex genetic architectures.

Purpose of the Study:

• To evaluate the performance of non-linear machine learning and deep learning models in predicting multifactorial disease risk.
• To investigate the impact of epistasis on disease prediction accuracy.
• To compare non-linear approaches against traditional linear models for genetic risk assessment.

Main Methods:

• Generated simulated genetic data with varying epistasis models (additive, multiplicative, threshold) using GAMETES.
• Employed machine learning algorithms including multilayer perceptron (MLP), convolutional neural network (CNN), recurrent neural network (RNN), Lasso regression, random forest, and gradient boosting.
• Assessed model performance using metrics such as accuracy, AUC-ROC, AUC-PR, recall, precision, and F1 score on both simulated and real genetic data.

Main Results:

• Non-linear models significantly outperformed linear regression (LASSO) as epistasis strength increased.
• Gradient boosting showed superior performance for obesity and psoriasis prediction.
• Deep learning models demonstrated significantly better performance than linear approaches for type 1 diabetes prediction.

Conclusions:

• Non-linear machine learning and deep learning models are more effective than linear models for disease risk prediction when epistasis is present.
• These advanced computational approaches enhance the accuracy of genetic risk assessment for complex multifactorial diseases.
• The study highlights the importance of considering non-linear genetic interactions in genomic prediction models.