DBNX: A Machine Learning Method for Ensembling Polygenic Risk Scores and Non-Genetic Factors.

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Polygenic risk scoring (PRS) evaluates genetic susceptibility using multiple variants for disease prediction.
• Current PRS methods have limitations in disease and population specificity and often neglect non-genetic factors.
• Ensemble methods combining multiple PRS can improve prediction but often require training data.

Purpose of the Study:

• To develop an unsupervised method for aggregating diverse polygenic risk scores (PRS).
• To create a model integrating PRS with non-genetic factors for a comprehensive Composite Risk Score (CRS).
• To enhance disease prediction accuracy by leveraging both genetic and non-genetic risk factors.

Main Methods:

• An unsupervised Deep Belief Network (DBN) was used to aggregate PRS from various methods.
• The DBN approach does not require training data, enabling direct ensembling of existing PRS.
• The DBNX model combines DBN with XGBoost to integrate PRS and non-genetic factors, generating a CRS.

Main Results:

• The DBN achieved performance comparable to supervised ensemble methods and outperformed the Super Learner on small datasets.
• DBNX demonstrated superior performance over other ensemble methods in predicting four diseases using the UK Biobank dataset.
• The DBNX model effectively integrated genetic and non-genetic factors for a more accurate Composite Risk Score.

Conclusions:

• Unsupervised DBN offers an effective alternative for PRS aggregation, outperforming supervised methods in certain scenarios.
• DBNX provides a robust framework for generating a Composite Risk Score by combining polygenic risk and non-genetic factors.
• The DBNX model represents a significant advancement in personalized disease risk prediction by incorporating multi-factorial data.