Examining applications of Neural Networks in predicting polygenic traits

Abstract

Polygenic risk scores are scores used in precision medicine in order to assess an individual's risk of having a certain quantitative trait based on his or her genetics. Previous works have shown that machine learning, namely Gradient Boosted Regression Trees, can be successfully applied to calibrate the weights of the risk score to improve its predictive power in a target population. Neural networks are a very powerful class of machine learning algorithms that have demonstrated success in various elds of genetics, and in this work, we examined the predictive power of a polygenic risk score that uses neural networks to perform the weight calibration. Using a single neural network, we were able to obtain prediction R2 of 0.234 and 0.074 for height and BMI, respectively. We further experimented with changing the dimension of the input features, using ensembled models, and varying the number of splits used to train the models in order to obtain a nal prediction R2 of 0.242 for height and 0.0804 for BMI, achieving a relative improvement of 1.26% in prediction R2 for height. Furthermore, we performed extensive analysis of the behaviour of the neural network-calibrated weights. In our analysis, we highlighted several potential drawbacks of using neural networks, as well as machine learning algorithms in general when performing the weight calibration, and o er several suggestions for improving the consistency and performance of machine learning-calibrated weights for future research.