Using machine learning to predict and analyze complex trait diseases: Lessons from a simple abstract model

The ability to predict individual genetic susceptibility to a complex trait disease is a major challenge in modern medicine. One approach to addressing this challenge utilizes an additive combination of contributions from a large number of single nucleotide polymorphisms (SNPs), with weights derived from Genome Wide Association Studies (GWAS). While this approach is somewhat successful in predicting whether an individual is likely to develop a specific disease, it does not explain why a person is likely to become sick. Here, we designed and utilized abstract disease models to investigate the relationship between disease structure, susceptibility, and predictability. The model consists of a set of interacting pathways, each including several nodes representing loci at which genetic variants can alter the function of the corresponding proteins. Due to the introduction of thresholds for pathway functionality, and the interplay between the pathways, this model is inherently non-additive. We use this "toy model" together with simulated variant data to examine the effect of changing various properties, some of which cannot be easily controlled in a "real-world" scenario. As expected, larger sample sizes improved the performance; the omission of some contributing variants from the dataset was associated with a significant decrease in performance, whereas adding irrelevant variants had little effect. Surprisingly, diseases with a more complex underlying structure were better predicted than those with a simpler structure. In addition, risk prediction was more accurate for diseases with lower prevalence. The algorithm was robust to a reasonable percentage of false negative disease assignments. The largest decrease in performance occurred when two diseases with different genetic etiologies were classified as a single pathology, as often occurs in clinical situations, and apparently confuses the neural network algorithm. Finally, we show that a post-analysis of a neural network using t-SNE can provide biological insights into the underlying disease structure.