🧬 genomics

Covariate-aware genomic prediction of blood metabolite profiles using multi-task neural networks

This study introduces a multi-task neural network framework that effectively predicts blood metabolite profiles by separating genetic, covariate, and joint contributions, revealing that nonlinear modeling of covariates—particularly age—is the primary driver of improved predictive performance over traditional linear methods.

Original authors: Guler, M. N., Alver, M., Haller, T., Jay, F., Pagani, L., Milani, L., Yelmen, B.

Published 2026-09-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guler, M. N., Alver, M., Haller, T., Jay, F., Pagani, L., Milani, L., Yelmen, B.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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, humming factory of chemistry. Inside our blood, tiny molecules called metabolites act as immediate snapshots of our physiological state, reflecting everything from the food we eat to the genes we inherited. These chemical signals bridge the gap between our DNA and how we actually feel and function, offering a window into risks for heart disease, diabetes, and other conditions. For decades, scientists have mapped the genetic roots of these molecules, identifying which parts of our genetic code influence them. However, knowing which genes are involved is different from being able to predict the exact chemical profile of a person's blood just by looking at their DNA. The challenge lies in the sheer complexity: these molecules do not act in isolation but influence one another in intricate, non-linear ways, and their levels are also swayed by factors like age, sex, and body weight. The question remains whether advanced computer models can untangle this web to predict a person's metabolic health more accurately than traditional methods.

A team of researchers from the University of Tartu in Estonia set out to answer this by building a new kind of predictive engine. They turned to the Estonian Biobank, a massive repository containing genetic data and blood samples from over 200,000 people. From this pool, they focused on nearly 110 different metabolites measured in the blood, including various fats and cholesterol types. The researchers trained a sophisticated artificial intelligence system, specifically a multi-task neural network, to learn the relationship between a person's genetic code and their blood chemistry. Unlike older models that treat each chemical trait as a separate, isolated problem, this new system was designed to learn all the traits simultaneously, allowing it to spot shared patterns across the entire metabolic landscape. To ensure they understood exactly where the improvements came from, they broke the model down into three distinct parts: one part that learned from non-genetic factors like age and body mass index, a second part that learned from the genetic data alone, and a third part that tried to capture any complex interactions between the two.

The results showed that this new approach worked better than the standard tools currently used in the field. When tested on a group of people the model had never seen before, the multi-task neural network achieved a higher level of accuracy than linear models, which assume relationships are simple and straight lines. On average, the new model explained about 22 percent of the variation in the metabolite profiles, a modest but meaningful improvement over the best linear methods, which explained about 21 percent. The researchers found that this gain was not primarily due to the model discovering new, hidden genetic interactions. Instead, the improvement came mostly from the model's ability to handle non-linear relationships in the non-genetic data. In simpler terms, the model was better at understanding how factors like age affect the body in complex, curved ways rather than just straight lines. For instance, the effect of aging on blood chemistry is not a simple, steady climb; it accelerates and shifts, and the new model captured these nuances far better than older, rigid formulas.

While the model excelled at interpreting environmental and lifestyle factors, its ability to predict the genetic component of the metabolites was more limited. The genetic part of the model added a small amount of predictive power, but it did not consistently outperform simpler linear methods when looking at genes alone. This suggests that while the new system is excellent at processing the complex, shifting influence of a person's life and environment, the genetic signals themselves are still difficult to decode with high precision using current data. The model did show particular strength in predicting lipid-related traits, such as different types of cholesterol and fatty acids, which make sense given that these molecules are tightly linked in the body's transport systems. However, for many other metabolites, the difference between the new complex model and the older, simpler ones was negligible.

The study concludes that while advanced neural networks offer a path forward, they are not a magic bullet that instantly solves the puzzle of genetic prediction. The primary value of this new framework is not just a slight bump in accuracy, but the ability to dissect exactly why a model succeeds or fails. By separating the contributions of genetics, environment, and their interplay, the researchers demonstrated that the biggest gains in predicting blood chemistry come from better modeling of how age and body composition affect the body, rather than from uncovering new genetic secrets. This distinction is crucial for future medical applications; it suggests that to improve predictions of metabolic health, scientists may need to focus more on capturing the complex, non-linear ways our environment shapes our biology, rather than solely on finding more genetic markers. The work provides a clear roadmap for future studies, showing that the most powerful tools are those that can carefully distinguish between the different sources of biological variation.

Technical Summary: Covariate-aware genomic prediction of blood metabolite profiles using multi-task neural networks

Problem Statement
While circulating metabolites serve as critical proximal readouts of physiological state and bridges between the genome and phenotype, their genomic prediction has largely relied on linear models treating metabolites as independent outcomes. Although genome-wide association studies (GWAS) have characterized the genetic architecture of these traits, it remains unclear how accurately multi-metabolite profiles can be predicted from genomic data and whether nonlinear models can improve performance by capturing complex interactions. Furthermore, existing predictive models often fail to distinguish whether performance gains arise from covariate effects, genetic effects, or joint covariate-genetic interactions. The study addresses the need for a framework that simultaneously models shared metabolite structure, accounts for nonlinear covariate effects, and decomposes predictive contributions to identify the sources of performance differences.

Methodology
The authors utilized genotype and Nightingale NMR metabolomics data from the Estonian Biobank (EstBB), comprising 211,259 individuals. After filtering for European ancestry, relatedness, and missing data, the dataset was split into training, validation, model-selection, and held-out test sets. The analysis focused on 109 target traits (107 non-derived metabolites plus HDL-C and LDL-C).

Genetic Predictor Construction: Metabolite-specific GWAS were performed in the training set. Variants associated with at least one biomarker ( $P < 1 \times 10^{-5}$ ) were LD-pruned and merged into a union predictor set of 13,049 variants.
Model Architecture: The core innovation is a three-stage residual decomposition framework implemented via neural networks (NN):
1. Covariate Module: Predicts metabolite levels using age, sex, BMI, and the first ten genetic principal components (PCs).
2. Genetic Module: Predicts the residuals from the first stage using the genetic predictor set.
3. Joint Module: Predicts the remaining residuals using both covariates and genotypes to capture interaction effects.
Model Variants: The study compared several architectures:
- Multi-task NN: A nonlinear model with shared hidden layers predicting all 109 metabolites simultaneously.
- Single-task NN: Nonlinear models trained independently for each metabolite.
- Activation-free Multi-task NN: A linear comparator retaining the multi-task architecture but removing nonlinear activation functions.
- Linear Baselines: Elastic Net, Ridge Regression, and Ordinary Least Squares (OLS).
Evaluation: Performance was assessed on the held-out test set using $R^2$ , MSE, MAE, and Pearson's $r$ . Uncertainty was estimated via 1,000 paired bootstrap replicates. Module-wise contributions were evaluated by calculating the $R^2$ of each component against the observed phenotype.

Key Results

Overall Performance: The nonlinear multi-task NN achieved the highest mean held-out $R^2$ across all metabolites ($0.219$), outperforming the single-task NN ($0.211$), Elastic Net ($0.207$), and the activation-free multi-task model ($0.191$).
Metabolite-Level Comparisons: The multi-task NN outperformed Elastic Net for 52.3% of metabolites (57/109), while Elastic Net was superior for only 8.3% (9/109). The advantage over the single-task NN was less decisive, with the multi-task model favored for 40.4% of traits.
Source of Gains (Decomposition):
- Covariate Module: The primary source of improvement. The nonlinear multi-task NN significantly outperformed the activation-free model in the covariate module (mean $R^2$ of $0.162$ vs. $0.130$), indicating that nonlinear modeling of covariates (e.g., age trajectories, sex-specific effects) drives the performance gain.
- Genetic Module: Contributions were smaller and heterogeneous. The activation-free multi-task model actually showed slightly higher average genetic module $R^2$ ($0.061$) than the nonlinear multi-task NN ($0.055$), suggesting that shared linear structure is sufficient for much of the genetic signal.
- Joint Module: The contribution of modeling joint covariate-genotype interactions was minimal and unstable across models (mean $R^2 \approx 0.009$ ).
Trait Specificity: The largest gains of the nonlinear multi-task NN over linear baselines were concentrated in lipid and lipoprotein-related traits (e.g., lipoprotein subclasses, fatty acids), which exhibit high biological correlation and shared structure.
Heritability Context: Observed genotype-based correlations remained below the theoretical limits implied by SNP-heritability estimates, suggesting that current models do not fully recover additive genetic signals, likely due to finite sample sizes and imperfect variant tagging.

Significance and Claims
The paper claims to provide a framework for identifying the specific sources of predictive differences between models in genomic prediction. The authors conclude that while nonlinear multi-task neural networks offer the strongest average performance for predicting circulating metabolite profiles, the improvement over strong linear baselines (like Elastic Net) is modest and metabolite-dependent.

Crucially, the study demonstrates that the observed gains are primarily driven by the nonlinear modeling of covariate features (such as age and BMI) rather than by capturing complex higher-order genetic interactions or genotype-covariate interactions. The multi-task approach offers statistical and computational efficiency by allowing a single model to be tuned for multiple traits, but the paper cautions that strong linear baselines remain difficult to beat in genomic prediction. The authors position their work not as a definitive solution for all genomic prediction tasks, but as a methodological tool to disentangle whether performance improvements stem from shared trait structure, nonlinear covariate effects, or complex genetic interactions. They note that future applications, such as generating genetic scores for metabolite profiles in cohorts without metabolomic assays, require external validation and calibration.

More like this