Synthetic data for pharmacogenetics: enabling scalable and secure research.

Area of Science:

• Pharmacogenetics
• Bioinformatics
• Data Science

Background:

• Synthetic data generation (SDG) is crucial for pharmacogenetics (PGx) research, especially with limited or sensitive patient data.
• Evaluating various SDG methods is essential to determine their suitability for complex PGx datasets.
• Assessing both data utility and privacy is critical for responsible data sharing and research.

Purpose of the Study:

• To evaluate the performance of seven synthetic data generation (SDG) methods for pharmacogenetics (PGx) research.
• To compare traditional and deep learning-based SDG approaches on high-dimensional genotype and phenotype PGx data.
• To assess SDG methods based on broad utility, specific utility, and privacy risk.

Main Methods:

• Seven SDG methods (synthpop, avatar, copula, copulagan, ctgan, tvae, tabula) were evaluated.
• PGx profiles from 142 patients were used, with scenarios including high-dimensional genotype (104 variables) and phenotype (24 variables) data.
• Performance was assessed using propensity score mean squared error (pMSE) for broad utility, weighted F1 score for specific utility, and ε-identifiability for privacy risk.

Main Results:

• Copula and synthpop demonstrated consistent strong performance, balancing low privacy risk (ε-identifiability: 0.25-0.35) with competitive utility.
• Deep learning models (tabula, tvae) achieved lower pMSE but had higher privacy risks (>0.4) and limited predictive gains.
• Specific utility (F1 score) was weakly correlated with broad utility (pMSE), indicating distributional fidelity does not guarantee predictive relevance.

Conclusions:

• No single SDG method excelled across all evaluation criteria.
• For privacy-sensitive PGx research, copula and synthpop offer a reliable trade-off between utility and privacy, particularly for high-dimensional, limited-sample datasets.
• Multimetric evaluation is essential, as general utility metrics do not always predict specific predictive utility.