XPRS: A Tool for Interpretable and Explainable Polygenic Risk Score

Naryeong Kim, Seunggeun LeemedRxiv 2024 · DOI

TL;DR

XPRS (eXplainable Polygenic Risk Score) is a new open-source software tool designed to transform "black box" genetic risk predictions into interpretable biological insights. By decomposing complex Polygenic Risk Scores (PRS) into specific gene and variant-level contributions using Shapley values, XPRS provides clinicians and researchers with the "why" behind an individual's disease susceptibility.

What Problem Does This Paper Address?

Polygenic Risk Scores (PRS) have become a powerful tool in personalized medicine, aggregating thousands or millions of tiny genetic variations (SNPs) to predict a person's risk for diseases like Type 2 Diabetes or heart disease. However, their clinical utility is stalled by a major hurdle: interpretability.

Because a PRS is typically a massive weighted sum of variants—many of which have unknown biological functions—it is nearly impossible for a clinician to explain to a patient which specific biological pathways are driving their high-risk score. XPRS tackles this by mapping these millions of variants back to functional units (genes and regions) and quantifying their individual impact on the final score.

Methods

XPRS introduces a structured pipeline to move from raw genomic data to visual explanations:

  1. Three-Phase Variant Mapping: To bridge the gap between SNPs and genes, the tool uses a hierarchical mapping protocol:
  2. Positional mapping: Using genomic coordinates.
  3. Functional mapping (cS2G): Using eQTLs and chromatin interaction data (Hi-C) to link variants to the genes they actually regulate.
  4. Heritability-based mapping: For remaining variants, it uses GWAS p-values or heritability estimates to group them into genomic regions.
  5. SHAP-based Decomposition: The tool calculates Attributed Values ($A_{gene}$) using Shapley Additive Explanations (SHAP). In a linear PRS model, this represents how much a specific gene deviates an individual's risk from the population average.
  6. Regionizing: To handle the high redundancy in genomic data (where multiple genes overlap), XPRS consolidates overlapping profiles into distinct "regions" to streamline the analysis.
  7. Implementation: The backend is built with C++ for high-performance computation and R for visualization, wrapped in a user-friendly Flask web interface.

Datasets

The authors demonstrated XPRS using the following data:

  • Genotype Data: 503 East Asian samples from the 1000 Genomes Project, encompassing over 80 million variants.
  • PRS Model: A Type 2 Diabetes (T2D) scoring file specifically evaluated for the Korean population.
  • Association Data: Curated gene lists from the GWAS Catalog to filter for established disease-associated genes.

Key Insights and Ideas

  • Genes as the Unit of Interpretation: While PRS is calculated at the SNP level, humans think in terms of genes. By aggregating SNP weights into gene-level "contribution scores," the results become immediately more actionable for biological research and clinical counseling.
  • The Power of SHAP: By using SHAP values, XPRS can tell a patient: "Your risk score is 0.028 higher than average specifically because of your variants in the CDKAL1 gene." This moves genetic counseling from abstract statistics to concrete biology.
  • Multilevel Visualization: The tool provides a "drill-down" experience. Users start at the Population Level (which genes matter most for everyone?) and move to the Individual Level (which genes matter for this patient?), finally reaching the SNP Level (which specific mutations within that gene are the culprits?).

Key Results

  • Computational Efficiency: Despite processing 80 million variants, XPRS is remarkably fast. A population-level analysis of 503 samples took approximately 5.72 minutes on an 8-core CPU. Individual-level visualizations are generated in just 2.54 seconds.
  • Biological Validation: In a T2D case study, XPRS successfully highlighted well-known risk genes. At the population level, KCNQ1 and KCNQ1OT1 showed the highest variance. For a specific high-risk individual (HG00464), XPRS identified CDKAL1 as the primary driver of their elevated risk.
  • Clinical Communication: The tool generates LocusZoom-like plots and Manhattan plots that allow clinicians to visually demonstrate genetic risk factors to patients, potentially increasing trust and adherence to preventative measures.

Limitations and Future Work

  • Privacy Constraints: Because genetic data is highly sensitive, XPRS is designed to be run locally or on private servers rather than as a centralized public web service. This may limit accessibility for users without their own server infrastructure.
  • Mapping Accuracy: The tool’s effectiveness is inherently tied to the accuracy of SNP-to-gene mapping. As the field of functional genomics evolves, the mapping protocols within XPRS will need constant updates to reflect the latest understanding of how non-coding variants affect gene expression.
  • Model Scope: While designed for linear PRS models (the current standard), future iterations could potentially expand to non-linear machine learning models used in genomics.

Summary updated

Abstract

Abstract The polygenic risk score (PRS) is an important method for assessing genetic susceptibility to diseases; however, its clinical utility is limited by a lack of interpretability tools. To address this problem, we introduce eXplainable PRS (XPRS), an interpretation and visualization tool that decomposes PRSs into genes/regions and single nucleotide polymorphism (SNP) contribution scores via Shapley additive explanations (SHAPs), which provide insights into specific genes and SNPs that significantly contribute to the PRS of an individual. This software features a multilevel visualization approach, including Manhattan plots, LocusZoom-like plots and tables at the population and individual levels, to highlight important genes and SNPs. By implementing with a user-friendly web interface, XPRS allows for straightforward data input and interpretation. By bridging the gap between complex genetic data and actionable clinical insights, XPRS can improve communication between clinicians and patients.
BibTeX
@misc{kim2024xprs,
  author = {Naryeong Kim and Seunggeun Lee},
  title = {{XPRS: A Tool for Interpretable and Explainable Polygenic Risk Score}},
  year = {2024},
  doi = {10.1101/2024.10.24.24316050},
  howpublished = {medRxiv},
}
Figure