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

Area of Science:

• Genetics and Genomics
• Computational Biology
• Medical Informatics

Background:

• Predicting individual genetic susceptibility to complex diseases remains a significant challenge in medicine.
• Current methods often use additive models based on single nucleotide polymorphisms (SNPs) from Genome-Wide Association Studies (GWAS), which have limitations in explaining disease mechanisms.
• Understanding the relationship between disease structure, genetic susceptibility, and predictability is crucial for advancing personalized medicine.

Purpose of the Study:

• To investigate the relationship between disease structure, genetic susceptibility, and predictability using abstract, non-additive disease models.
• To examine how various factors, such as sample size, variant data completeness, disease complexity, and prevalence, affect disease risk prediction.
• To explore the utility of t-distributed Stochastic Neighbor Embedding (t-SNE) for gaining biological insights into disease structures from predictive models.

Main Methods:

• Designed and utilized abstract, non-additive disease models representing interacting pathways with genetic variant effects.
• Employed simulated genetic variant data to test the predictive model under various controlled conditions.
• Assessed the impact of sample size, variant data quality (omission/addition of variants), disease complexity, prevalence, and diagnostic accuracy on prediction performance.

Main Results:

• Larger sample sizes improved prediction performance, while omitting relevant variants significantly decreased it; adding irrelevant variants had minimal impact.
• Diseases with more complex underlying structures and lower prevalence were predicted more accurately.
• The predictive algorithm demonstrated robustness to false negative assignments but struggled when distinct diseases with different genetic etiologies were misclassified as one.

Conclusions:

• Non-additive genetic architectures and disease complexity are critical factors influencing the predictability of complex traits.
• Abstract disease models provide a valuable framework for dissecting the interplay between genetic architecture and disease risk prediction.
• Post-analysis using t-SNE on neural network models can reveal underlying biological insights into disease structures, aiding in understanding disease etiology.