A comparative analysis of methods for predicting clinical outcomes using high-dimensional genomic datasets

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Accurate prediction of disease status from high-dimensional genomic data is crucial for personalized medicine.
• Epistatic interactions, where multiple genes influence a phenotype, pose a significant challenge for traditional prediction models.
• Bayesian network (BN)-based methods have shown promise in learning complex genetic interactions.

Purpose of the Study:

• To evaluate the performance of binary prediction methods for disease status using high-dimensional genomic data.
• To test the hypothesis that the efficient Bayesian multivariate classifier (EBMC), a BN-based method, excels in predicting disease status.
• To compare EBMC against other established methods like logistic regression, SVM, and naive Bayes.

Main Methods:

• Eight binary prediction methods were evaluated: naive Bayes (NB), model averaging NB (MANB), feature selection NB (FSNB), EBMC, logistic regression (LR), support vector machines (SVM), Lasso, and extreme learning machines (ELM).
• The methods were tested on diverse datasets, including 1000-SNP and 10,000-SNP simulated datasets, semi-synthetic sets, and real genome-wide association studies (GWAS) datasets.
• Performance was assessed using fivefold cross-validation and in-sample testing.

Main Results:

• The Support Vector Machine (SVM) performed best on the 1000-SNP dataset.
• Bayesian network (BN)-based methods, particularly EBMC, demonstrated superior performance on larger and more complex datasets.
• In-sample testing revealed that LR, SVM, Lasso, ELM, and NB tended to overfit the data.

Conclusions:

• The efficient Bayesian multivariate classifier (EBMC) effectively predicts disease status in high-dimensional genomic datasets with epistatic-like interactions, supporting the initial hypothesis.
• EBMC outperformed naive Bayes (NB) when strong predictors were present, while NB was more effective with numerous weak predictors.
• The predictive capability of BN-based methods remained stable as data dimensionality increased, highlighting their scalability.