📄 genetic and genomic medicine

Additive Multilocus Burden and Epistatic Interactions Improves Genetic Risk Predictions for Complex Diseases

This study introduces an extended polygenic risk score (ePRS) framework that incorporates non-additive multilocus interactions and gene-environment effects to significantly improve genetic risk prediction for complex diseases like type 2 diabetes and celiac disease by capturing high-risk individuals missed by traditional additive models.

Original authors: Multerer, K., Atkinson, P., Woods, L., Tanigawa, Y., Kellis, M., Munkacsi, A.

Published 2026-08-14

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Multerer, K., Atkinson, P., Woods, L., Tanigawa, Y., Kellis, M., Munkacsi, A.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 your DNA as a massive, ancient instruction manual written in a four-letter code. For decades, scientists have tried to predict who might get sick by reading this manual line-by-line, assuming that every single letter adds a tiny, independent amount of risk. This is like trying to guess the flavor of a soup by tasting each ingredient separately and adding up the flavors. It works okay, but it misses the magic that happens when ingredients mix. In the real world, ingredients interact: garlic changes how onions taste, and salt changes how tomatoes behave. In genetics, these interactions are called "epistasis" (when genes talk to each other) and "gene-environment interactions" (when your genes react to things like your diet or stress). The big question scientists have been asking is: if we only look at the ingredients one by one, how much of the "flavor" of disease are we missing?

This paper, titled "Additive Multilocus Burden and Epistatic Interactions Improves Genetic Risk Predictions for Complex Diseases," dives into that missing flavor. The researchers built a new kind of genetic risk calculator that doesn't just add up the ingredients; it looks at how they bump into each other and how they react to the "kitchen environment" (like blood sugar or body weight). They tested this on two very different diseases: Type 2 Diabetes, which is like a complex stew with hundreds of ingredients interacting, and Celiac Disease, which is more like a dish dominated by one or two very loud spices.

The team, led by Keri Multerer and Andrew B. Munkacsi, used data from 235,000 people in the UK Biobank to build five different versions of their new calculator. They found that the old way of adding up risks (the "additive" method) was like listening to a choir where everyone sings the same note. Their new method listened for the harmonies and the dissonance. They discovered that different versions of their new calculator caught different groups of people who were at high risk. In fact, for Type 2 Diabetes, over half of the high-risk individuals caught by the new models were missed entirely by the old method. It's as if the old calculator was looking for people with red hats, while the new ones were also looking for people with blue scarves, green shoes, and yellow gloves—finding many people the first one ignored.

When they combined all these different "ears" into one super-calculator, they could spot even more high-risk individuals, including people who looked perfectly healthy by standard medical checks like BMI or blood sugar levels. However, the paper also notes that this new approach works best for diseases with complex, mixed-up genetics. For Celiac Disease, which is driven mostly by a few very strong genetic signals (like a single, overpowering spice), the new method didn't add much extra value. The researchers suggest that while this approach isn't a magic cure-all yet, it opens a door to understanding the hidden, interactive parts of our genetic risk that we've been ignoring for too long. They emphasize that these findings need to be tested in more diverse groups and in real-world clinical settings before doctors can use them, but the results suggest that looking at how genes interact with each other and our environment is a crucial next step in predicting who might get sick.

Technical Summary: Additive Multilocus Burden and Epistatic Interactions Improves Genetic Risk Predictions for Complex Diseases

Problem Statement
Traditional Polygenic Risk Scores (PRS) estimate genetic risk by aggregating additive effects from genome-wide association studies (GWAS). However, complex disease risk reflects broad-sense heritability ( $H^2$ ), which includes gene-gene (epistatic) and gene-environment interactions that standard additive models systematically exclude. While numerous methods have attempted to model higher-order genetic effects, many either fail to explicitly retain interaction terms or sacrifice interpretability for predictive performance. Consequently, the extent to which interaction-aware feature representations can capture complementary, non-overlapping components of polygenic risk remains unclear.

Methodology
The authors developed an interpretable machine-learning framework, termed extended PRS (ePRS), to explicitly incorporate complementary multi-locus feature encodings while retaining the contributing loci and environmental variables. The study utilized data from approximately 235,000 unrelated white British participants in the UK Biobank, split into training (70%), validation (20%), and holdout (10%) sets.

The framework focused on two diseases with contrasting genetic architectures: Type 2 Diabetes (T2D), which is highly polygenic, and Celiac Disease (CD), which is dominated by large Human Leukocyte Antigen (HLA) effects.

The core methodology involved:

Feature Discovery: Identification of statistically significant SNP pairs using PLINK's fast-epistasis screening ( $p < 5 \times 10^{-6}$ ).
Dual Encoding Strategies: The same significant SNP pairs were transformed into two distinct feature representations:
- Multilocus Additive Burden (G+G): Summing allele counts across paired loci to capture cumulative allelic burden.
- Statistical Epistasis (G×G): Multiplying allele counts to model departures from additive independence.
Gene-Environment Integration: Both G+G and G×G features were combined with 72 cardiometabolic clinical variables (E) to generate interaction models: G+G×E and G×G×E.
Model Training: Seven L1-penalized (Lasso) logistic regression models were trained on these feature sets (G, G+G, G×G, G+G×E, G×G×E, and combined variants).
Redundancy Filtering: A rigorous statistical protocol (using Cohen's $\kappa$ , Jaccard index, and DeLong's test) was applied to identify and retain only statistically distinct models, ensuring that retained models captured unique high-risk cohorts rather than redundant signals.
Composite Scoring: A composite score (ePRSComp) was constructed by classifying individuals as high-risk if they exceeded the threshold in any of the retained distinct models.

Key Results

Distinct High-Risk Cohorts: In T2D, five statistically distinct ePRS models were retained. These models identified largely non-overlapping high-risk individuals. Specifically, 78% of T2D cases in the validation set were identified, with 53% of these cases being exclusive to a single model. The G+G and G×G models were non-redundant despite being derived from the same SNP pairs, indicating they capture complementary biological signals.
Superior Case Detection: Integrating these complementary models into a composite score (ePRSComp) improved case detection beyond standard additive PRS (PRSG) and conventional clinical predictors. In the holdout set, ePRSComp identified 77% of T2D cases (2,077/2,696), including 1,323 additional cases missed by PRSG alone.
Clinical Relevance: The composite score successfully reclassified individuals within clinically normal ranges (e.g., normal BMI, glucose, and HbA1c) into high-risk categories. For glucose, the Net Reclassification Improvement (NRI) was positive (0.08), indicating that ePRSComp captures genetic risk components complementary to standard clinical thresholds.
Disease Architecture Dependence: The benefits of interaction modeling were disease-specific. In T2D, interaction models captured substantial unique risk. In CD, gains were more modest due to the disease's HLA-dominated additive architecture, though interaction models still identified unique cases, particularly those driven by HLA-environment interactions.
Biological Interpretability: Model-specific features mapped to distinct biological pathways and established T2D subtypes (SAID, SIDD, SIRD, MOD, MARD). For instance, G×G×E models were enriched for liver/lipid metabolism and severe insulin-resistant diabetes (SIRD), while G+G models showed broader representation across $\beta$ -cell dysfunction and metabolic pathways.

Significance and Claims
The paper claims that complementary multi-locus feature representations capture distinct components of polygenic risk that are overlooked by conventional additive PRS. The authors argue that aggregate discrimination metrics like AUC can underestimate the value of interaction models when orthogonal feature encodings identify distinct high-risk individuals.

Key contributions include:

Demonstrating Complementarity: Showing that additive burden (G+G) and statistical epistasis (G×G) capture different subsets of high-risk individuals, even when derived from the same underlying genetic data.
Addressing Missing Heritability: Providing evidence that non-additive genetic effects and gene-environment interactions contribute significantly to broad-sense heritability in metabolic diseases like T2D.
Clinical Utility: Showing that a composite approach can identify high-risk individuals who would otherwise be missed by standard clinical screening, particularly those with normal metabolic profiles.
Interpretability: Maintaining biological interpretability by explicitly modeling interaction terms, allowing for the generation of testable hypotheses regarding specific locus pairs and environmental contexts.

The authors conclude that the benefit of integrating multiple feature encodings scales with genetic heterogeneity, offering a pathway to improve risk prediction for highly polygenic diseases while remaining biologically interpretable. They note that prospective validation is required before clinical application.

More like this