🧬 biology

Integrated Polygenic Risk Scores: Composite and Ensemble Learning Approaches for Precision Medicine

This paper proposes and validates novel composite and adaptive ensemble learning methods for integrating disease and pharmacogenomics GWAS data to overcome existing limitations in polygenic risk score development, thereby significantly improving drug response prediction and patient stratification in precision medicine.

Original authors: Judong Shen, Junming Guan, Song Zhai, Wujuan Zhong, Devan Mehrotra

Published 2026-09-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Judong Shen, Junming Guan, Song Zhai, Wujuan Zhong, Devan Mehrotra

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to predict how a person will react to a new medicine. For decades, doctors have relied on a one-size-fits-all approach, but the future of healthcare lies in precision medicine: tailoring treatments to the unique genetic makeup of each patient. To do this, scientists use a tool called a polygenic risk score. Think of this score as a way to add up thousands of tiny genetic instructions scattered across a person's DNA to see if they are naturally prone to a certain disease or likely to respond well to a specific drug. However, building these scores has been like trying to solve a puzzle with missing pieces. Scientists have long struggled to separate two types of genetic influences: those that determine how sick a person is before they start treatment, and those that determine how their body specifically reacts to the medicine itself. If you cannot tell these two apart, you cannot accurately predict who will benefit from a drug and who might not.

A team of researchers at Merck & Co. and the University of Chicago has developed a new way to solve this puzzle. They created a method that combines two different types of genetic data to build a much sharper picture of drug response. Traditionally, scientists have tried to build these scores using data from large studies of people with a disease, or from smaller studies of people already taking a drug. The first approach is good at predicting who is sick, but it misses how the drug works. The second approach is direct but often lacks enough data to be reliable. The new method, which the researchers call an "ensemble" approach, acts like a smart filter that learns how to blend these two sources of information perfectly. Instead of guessing how to mix the data, the computer tests many different ways of combining them and selects the one that works best for the specific situation.

In their work, the researchers first proposed a simpler way to mix the data, which they call a composite method. This approach assumes that the genetic factors making a person sick are the same as the factors that make them respond to a drug, and it tries to subtract one from the other to find the drug-specific signal. While this worked better than old methods, the researchers realized that the relationship between disease and drug response is not always simple. Sometimes the genetic drivers of a disease help the drug work; other times, they might work against it. To handle this uncertainty, they built a more advanced system that learns the best way to combine the data on its own. This system tests a wide range of possibilities, weighing the disease data and the drug data differently until it finds the perfect balance to predict outcomes.

To test if this new system actually works, the team ran thousands of computer simulations. They created virtual populations with known genetic traits and simulated how they would respond to treatment under different conditions. In these tests, their new method consistently outperformed the standard approaches used today. It was better at predicting the overall outcome and, crucially, much better at identifying the specific genetic signals that indicate a drug will work. The simulations showed that while older methods often missed the mark or got confused by the noise in the data, the new ensemble method could adapt to different scenarios and find the true signal. It was particularly effective when the genetic factors for disease and drug response were not perfectly aligned, a common real-world situation that stumped previous tools.

The researchers then took their method to the real world, applying it to data from the IMPROVE-IT clinical trial, a major study involving over 5,000 patients who were given a cholesterol-lowering drug. The goal was to predict how much each patient's bad cholesterol would drop after one month of treatment. The standard methods, which rely on just disease data or just drug data, failed to find a strong pattern. They could not clearly separate the patients who would get a big benefit from those who would get a small one. In contrast, the new method successfully identified a clear genetic signal. It could distinguish between patients who would see a significant drop in cholesterol and those who would not. The method was so precise that it could sort patients into groups where the treatment effect was obvious, a capability that is essential for doctors who want to prescribe the right drug to the right person.

One of the most important findings was that the new method could identify specific genes that drive the drug's effect. The analysis pointed to genes known to be involved in how the body processes cholesterol, confirming that the method was finding biologically real connections rather than random noise. This ability to pinpoint the right genetic markers means that in the future, doctors might be able to look at a patient's DNA and know with greater confidence whether a specific medication will lower their cholesterol effectively. The study suggests that by combining different types of genetic information, scientists can overcome the limitations of small sample sizes and messy data that have held back progress in this field.

The researchers acknowledge that their work is not a final solution. The success of their method depends on the quality of the data available, and they noted that the specific drug trial they used had some limitations, such as a relatively small number of participants compared to massive disease studies. They also pointed out that the relationship between disease and drug response can vary between different populations and different medicines, meaning the method would need to be tested and refined for other conditions. However, the results from this study provide a strong proof of concept. By showing that an adaptive, learning-based approach can integrate diverse genetic data to improve prediction, the team has offered a new path forward for precision medicine. Their work demonstrates that when we stop treating genetic data as separate islands and start connecting them intelligently, we can unlock a clearer understanding of how our genes shape our response to the medicines that save our lives.

Technical Summary: Integrated Polygenic Risk Scores: Composite and Ensemble Learning Approaches for Precision Medicine

Problem Statement
Pharmacogenomics (PGx) aims to predict drug response and stratify patients using genetic information. Polygenic risk scores (PRS) are a primary tool for this, aggregating effects across multiple genetic variants. However, existing PRS development strategies face significant limitations in the context of drug response prediction:

Disease-based PRS (PRS-Dis): These utilize Genome-Wide Association Study (GWAS) summary statistics from disease cohorts. While they offer large sample sizes and precise effect estimates, they primarily capture prognostic effects (genetic influence on disease risk independent of treatment) and fail to capture predictive effects (genotype-by-treatment interactions). Consequently, they often explain only a small fraction of drug response heritability.
PGx-Treatment-based PRS (PRS-PGx-T): These are derived from GWAS data of treated individuals. While they directly target drug response, they are often hindered by small sample sizes. Furthermore, in treated-only cohorts, prognostic and predictive effects are confounded (only the total genetic effect is identifiable), making it difficult to disentangle the specific predictive signals required for patient stratification.

The core challenge is integrating these two complementary data sources—disease GWAS and PGx-T GWAS—to construct PRS that accurately estimate both prognostic and predictive genetic effects, thereby improving drug response prediction and enabling effective patient stratification.

Methodology
The authors propose an integrative framework, PRS-DisPGx, which leverages summary statistics from both disease and PGx-T GWAS. The framework is built upon a linear modeling assumption where drug response ( $\Delta Y$ ) is a function of prognostic effects ( $\beta$ ) and predictive effects ( $\alpha$ ) interacting with treatment ( $T$ ).

The paper introduces three primary methodological strategies:

Naïve Composite Approaches (PRS-DisPGx-C):
These methods integrate the two data sources using fixed assumptions about the relationship between disease genetic effects and prognostic drug response effects.
- Assumption: Disease genetic effects ( $\mu$ ) are correlated with prognostic drug effects ( $b$ ).
- Implementation: Two versions are proposed based on the direction of this correlation:
  - PRS-DisPGx-C: Assumes $\mu \approx b$ . Prognostic effects are estimated from disease GWAS ( $\hat{\beta} = \hat{\mu}_{dis}$ ), and predictive effects are estimated by subtracting the disease effects from the total PGx-T effects ( $\hat{\alpha} = \hat{c}_{PGx} - \hat{\mu}_{dis}$ ).
  - PRS-DisPGx-C1: Assumes $\mu \approx -b$ . Prognostic effects are estimated as $-\hat{\mu}_{dis}$ , and predictive effects as $\hat{c}_{PGx} + \hat{\mu}_{dis}$ .
- Levels of Integration: Both approaches are implemented at two stages:
  - PRS-level (C-p): Shrinkage (SNP selection/weight refinement) is applied to disease and PGx-T summary statistics separately before integration.
  - Variant-level (C-v): Raw marginal effect sizes are integrated first to create composite estimates, followed by a single shrinkage step.
Adaptive Ensemble Learning Approach (PRS-DisPGx-E):
Recognizing that the optimal direction of integration (sign of correlation) and the optimal weighting are unknown, the authors propose an ensemble method.
- Mechanism: Instead of fixed weights, the method generates a list of $K$ composite effect size estimates using a grid of weights ( $w_k$ ) ranging from -1 to 1.
- Optimization: Using an external validation set, the method learns optimal ensemble weights ( $b_{k}$ ) via Linear Regression (LR) or Elastic Net (EN) to linearly combine the $K$ composite estimates for both prognostic and predictive PRS.
- Versions: Similar to the composite methods, this is implemented as PRS-DisPGx-E-p (shrinkage before ensemble) and PRS-DisPGx-E-v (shrinkage after ensemble).
Theoretical Framework:
The authors derive analytical expressions for the theoretical $R^2$ of these methods. They identify three key correlations governing performance:
- $\rho_Y$ : Cross-phenotype correlation between disease and drug-response genetic components.
- $\rho_C$ : Cross-population correlation between base and target cohorts.
- $\rho_E$ : Correlation between prognostic and predictive effects.
  The analysis demonstrates that the optimal method depends on the specific values of these correlations, particularly the sign and magnitude of $\rho_Y$ and $\rho_E$ .

Key Results

Simulation Studies:
- Overall Prediction Accuracy ( $R^2$ ): The ensemble approach, specifically the variant-level version (PRS-DisPGx-E-v), consistently achieved the highest overall $R^2$ across various scenarios, outperforming PRS-Dis, PRS-PGx-T, and the fixed composite methods.
- Predictive Effect Detection: While PRS-PGx-T often showed high overall accuracy, it struggled to isolate predictive effects due to the confounding of prognostic and predictive signals. The composite and ensemble methods, which explicitly separate these effects, demonstrated significantly higher statistical significance in detecting treatment-by-PRS interactions (lower predictive p-values).
- Robustness: The ensemble methods were robust to variations in the ratio of prognostic to predictive effect magnitudes and polygenicity levels. The variant-level ensemble (E-v) showed superior adaptivity compared to the PRS-level version.
Real-World Application (IMPROVE-IT Trial):
- Dataset: Applied to PGx GWAS data from the IMPROVE-IT trial (predicting LDL-C reduction) using summary statistics from the Global Lipids Genetics Consortium (disease) and Sadler et al. (PGx-T).
- Performance: The variant-level ensemble method (PRS-DisPGx-E-v) achieved an overall $R^2$ of ~~0.075, outperforming the baseline PRS-Dis (~~0.049) by >50%.
- Stratification: The methods successfully stratified patients. The variant-level composite method (PRS-DisPGx-C-v) showed the strongest predictive signal (p-value $2.89 \times 10^{-9}$ ), and the ensemble methods showed pronounced positive trends in treatment effects across PRS quantiles.
- Biological Plausibility: Top predictive SNPs identified by the proposed methods mapped to genes known to be associated with LDL-C (e.g., APOE, APOC1, LDLR), validating their biological relevance.
- Anomaly: PRS-PGx-T performed poorly in this specific dataset (lowest $R^2$ ), likely due to small sample size, signal cancellation (negative correlation between prognostic and predictive effects), and treatment heterogeneity between the base and target cohorts.

Significance and Claims
The paper claims three main contributions:

Methodological Innovation: It introduces novel composite and ensemble learning frameworks that effectively integrate disease and PGx-T GWAS summary statistics. The adaptive ensemble approach (PRS-DisPGx-E) is highlighted as a robust solution that does not require prior knowledge of the underlying genetic correlations.
Theoretical Insight: The study provides a theoretical framework elucidating how prediction performance is bounded by trait heritability and determined by the interplay of three specific correlations ( $\rho_Y, \rho_C, \rho_E$ ). This explains why certain methods fail under specific genetic architectures (e.g., when prognostic and predictive effects cancel out).
Practical Utility: Through simulations and the IMPROVE-IT application, the authors demonstrate that their methods significantly improve drug response prediction accuracy and patient stratification capabilities compared to conventional single-source PRS approaches.

The authors conclude that these integrative strategies offer a practical path forward for precision medicine, enabling more accurate genetic prediction of drug response. They note limitations regarding the dependency on the quality of summary statistics and the need for larger PGx-T datasets, and suggest future work could focus on optimizing objective functions specifically for predictive effect detection and incorporating multi-ancestry data.

More like this