🧬 biology

Gene-environment interactions for plasma proteins improve disease prediction

This study constructs a comprehensive atlas of gene-environment interactions for plasma proteins by identifying novel variance quantitative trait loci and their interactions with diverse exposures, demonstrating that incorporating these interactions significantly improves the prediction accuracy of 46 incident diseases.

Original authors: Chunshui Yu, Ying Wang, Sijia Wang, Yao wang, Mengge Liu, Jinbo Qi, Jia Wang, Jiayuan Xu, Hui Xue, Nana Liu, Hao Ding, Feng Liu, Wen Qin, Meng Liang

Published 2026-09-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chunshui Yu, Ying Wang, Sijia Wang, Yao wang, Mengge Liu, Jinbo Qi, Jia Wang, Jiayuan Xu, Hui Xue, Nana Liu, Hao Ding, Feng Liu, Wen Qin, Meng Liang

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The human body is a vast, interconnected system where our genetic code and the world around us are in constant conversation. We inherit a set of instructions in our DNA, but how those instructions play out often depends on our environment—what we eat, where we live, and the habits we form. Scientists have long known that genes and environment interact, but for a long time, they studied these factors separately, looking at how genes alone affect health or how lifestyle choices alone change our biology. This approach missed the complex reality that a specific genetic makeup might make one person highly sensitive to a particular environmental factor, while another person with different genes remains unaffected. To understand disease risk more deeply, researchers needed to map these specific conversations between our genes and our surroundings, particularly looking at plasma proteins. These proteins circulate in our blood, acting as messengers that reflect the body's current state and its response to both internal biology and external pressures. By understanding how genes and environment combine to shape these proteins, scientists hope to move beyond broad, one-size-fits-all health advice toward predictions that are tailored to an individual's unique biology.

A team of researchers at Tianjin Medical University General Hospital recently took a massive step forward in this field by analyzing data from over 51,000 people in the UK Biobank. They focused on nearly 3,000 different plasma proteins, examining how variations in DNA interacted with hundreds of different environmental exposures, ranging from diet and smoking to blood pressure and body measurements. Instead of looking for simple, direct links where a gene always causes a protein to rise or fall, they searched for a more subtle signal: instances where the effect of an environmental factor on a protein depended entirely on a person's genetic code. They found that for more than 1,000 of these proteins, the genetic code acted like a switch, determining whether an environmental factor would have a strong effect, a weak effect, or even an opposite effect. For example, they discovered that for a protein called pregnancy zone protein, the amount of fat in a person's right leg had a strong positive relationship with protein levels in people with one version of a gene, no relationship in people with another version, and a negative relationship in those with a third version. This kind of nuance was invisible to previous studies that only looked at average effects across all people.

The researchers identified over 1,300 specific locations in the genome where these interactions occurred, a number that nearly doubles the previous record for such discoveries. They then tested how these genetic-environmental interactions held up against 536 different exposures, uncovering nearly 500 independent interactions. Remarkably, almost all of these findings were new, suggesting that the landscape of how our genes and environment mix is far more complex and varied than previously understood. The study also revealed that these interactions are not random; they often cluster around biological processes related to immune response, wound healing, and blood clotting. One striking example involved a protein called GP2, which helps protect the lining of the gut. While smoking showed only a very weak overall link to GP2 levels in the general population, the researchers found that for people with a specific genetic variant, smoking actually increased GP2 levels, while for those with a different variant, it decreased them. This suggests that the reason smoking is harmful to the gut might be much more pronounced for certain individuals, hidden by the fact that it has no effect on others.

Beyond mapping these interactions, the team asked a practical question: does knowing about these gene-environment conversations help predict disease? They built computer models to forecast the risk of 321 different diseases, starting with a baseline model that used only basic demographic information and protein levels. They then added the new gene-environment interaction data to see if it improved the predictions. The results showed that for 46 of the diseases, the new model was significantly better at identifying who was at risk. The improvement was most dramatic for conditions like sepsis, a severe reaction to infection, where the model's accuracy jumped by a substantial margin. For these diseases, the interactions between specific proteins and factors like blood platelet counts or cholesterol levels provided crucial clues that the standard models missed. The study did not claim to have solved the puzzle of disease prediction, but it demonstrated that ignoring the interplay between genes and environment leaves out vital information. By including these interactions, doctors and scientists may eventually be able to offer more precise warnings and prevention strategies, telling a person not just that they are at risk, but exactly why their unique combination of genes and lifestyle puts them there.

Technical Summary: Gene-environment interactions for plasma proteins improve disease prediction

Problem Statement
Plasma proteins serve as critical biomarkers for physiological homeostasis and disease pathology, shaped by both genetic and environmental factors. While large-scale studies have extensively mapped the main effects of genetic variants (protein quantitative trait loci, pQTLs) and environmental exposures on plasma protein levels, they have largely overlooked gene–environment (G × E) interactions. Detecting these interactions across the genome and exposome presents substantial statistical and computational challenges due to the massive multiple-testing burden. Previous work utilizing variance quantitative trait loci (vQTLs) as a screening strategy for G × E interactions has shown promise but was limited to a subset of plasma proteins (1,463 proteins) and a narrower range of environmental exposures. Furthermore, the extent to which G × E interactions contribute to the prediction of incident diseases remains unassessed.

Methodology
This study leveraged data from 51,736 participants of the UK Biobank Pharma Proteomics Project (UKB-PPP) with 2,920 plasma proteins measured via Olink panels. The research employed a multi-stage analytical framework:

vQTL Screening: Genome-wide vQTL analyses were performed in a discovery cohort (n = 32,943) using the median-based Levene's test to identify loci where phenotypic variance differs across genotypes, serving as a proxy for potential G × E interactions. A Bonferroni-corrected threshold ( $P < 1.71 \times 10^{-11}$ ) was applied.
Exposome-wide Interaction Scan: Identified vQTLs were tested against 536 exposures (105 disease and 431 non-disease factors) categorized into internal and external domains. Stepwise conditional analysis was used to identify independent G × E interactions.
Validation and Meta-analysis: Results were validated in an independent cohort (n = 18,793) comprising diverse ethnic groups (White, Asian, Black, Mixed, Other) using meta-analysis.
Functional Annotation: Variants were annotated using FUMA, ANNOVAR, CADD, SIFT, and PolyPhen-2. Protein–protein interaction (PPI) networks and Gene Ontology (GO) enrichment analyses were conducted for affected proteins.
Genotype-Stratified Mediation: Mediation analyses were performed to extend genotype-specific exposure effects to indirect pathways, examining how internal exposures mediate the relationship between external exposures and plasma proteins within specific genotype subgroups.
Disease Prediction: A Light Gradient Boosting Machine (LightGBM) was used to predict 321 incident diseases. Two models were compared: a "basic" model (demographics + top 30 disease-associated proteins) and an "enhanced" model (basic model + G × E interaction terms). Performance was evaluated using Area Under the Curve (AUC) and the DeLong test.

Key Contributions and Results

Expanded vQTL Atlas: The study identified 1,357 independent vQTLs for 1,001 plasma proteins, nearly doubling the number of previously reported vQTLs (677 vQTLs for 568 proteins). These vQTLs were unevenly distributed, with the highest density on chromosome 6.
Novel G × E Interactions: The analysis revealed 498 independent G × E interactions involving 286 variants, 185 exposures, and 211 proteins. Notably, 487 of these interactions (97.79%) were novel, having not been reported in previous studies or the GWAS Catalog.
Genotype-Dependent Effects: The study demonstrated that G × E interactions can reveal genotype-dependent exposure effects that are masked in main-effect analyses. For example, the association between right leg fat percentage and Pregnancy Zone Protein (PZP) was positive in CC carriers, non-significant in CT carriers, and negative in TT carriers. Similarly, smoking status showed opposing effects on Glycoprotein 2 (GP2) levels depending on rs687289 genotype.
Biological Insights:
- Variants: 238 variants with independent interactions were primarily located in intronic and intergenic regions, suggesting regulatory mechanisms. Eight nonsynonymous variants were identified, four of which were predicted to be damaging.
- Pathways: Enrichment analyses indicated that genes and proteins involved in G × E interactions are primarily associated with immune regulation, wound healing, blood coagulation, and responses to external stimuli.
- Mediation: Genotype-stratified mediation analyses identified specific indirect pathways. For instance, the effect of current tobacco smoking on BRSK2 via platelet count was significant only in rs7932863 AG heterozygotes.
Disease Prediction Improvement: Incorporating G × E interaction terms into the prediction model significantly improved performance for 46 out of 321 incident diseases (14.33% at Bonferroni-corrected significance). The most substantial improvements were observed for implicit sepsis (AUC increased from 0.682 to 0.825, $\Delta$ AUC = 0.143), vascular diseases of the intestine, and explicit sepsis. SHAP value analysis identified specific interactions (e.g., platelet crit × rs116947050) as top contributors to these improvements.

Significance
The paper claims that this study provides the most comprehensive atlas to date of vQTLs and G × E interactions in the human plasma proteome. By identifying nearly 500 novel interactions, the work significantly expands the understanding of how genetic backgrounds modulate environmental influences on protein levels. Crucially, the study demonstrates that adding G × E interaction terms to demographic and protein-based models yields statistically significant improvements in disease prediction for a subset of conditions, particularly those involving immune and inflammatory processes like sepsis. This suggests that G × E interactions capture additional disease-risk information beyond main effects, offering a pathway toward more precise, genotype-stratified preventive interventions. The findings highlight that modest main effects of environmental exposures may mask strong, genotype-specific risks, underscoring the necessity of interaction analysis for personalized medicine.

More like this