📊 statistics

Graph-Adaptive Horseshoe for Compositional Regression

The paper proposes GRACE, a fully Bayesian framework that addresses the challenges of compositional regression by integrating a novel linear reparameterization with an adaptive, outcome-driven shrinkage graph to simultaneously perform variable selection, improve predictive accuracy, and recover feature relationships that differ from traditional fixed phylogenetic or ecological networks.

Original authors: Satabdi Saha, Christine B. Peterson

Published 2026-08-19

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Satabdi Saha, Christine B. Peterson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the human body, particularly within the mouth, trillions of microscopic organisms live together in complex communities. Scientists call this collection the microbiome. When researchers study these communities, they do not count the absolute number of bacteria, because that number changes constantly depending on how much saliva or plaque is collected. Instead, they measure the relative abundance, or the proportion of each type of bacteria compared to the whole group. This creates a unique mathematical puzzle: if the proportion of one type of bacteria goes up, the proportions of all the others must go down to keep the total at one hundred percent. This interdependence makes it difficult to use standard statistical tools to find out which specific bacteria are linked to a health condition. For years, scientists have tried to solve this by assuming that bacteria that are closely related on the evolutionary tree of life would behave similarly. They built models based on these fixed family trees, hoping that evolutionary closeness would predict how bacteria interact with human health. However, this assumption often fails because bacteria that are distant relatives might still act alike in the human body, while close relatives might have very different effects.

A team of researchers has developed a new method to untangle these relationships without relying on pre-made assumptions about bacterial family trees. They created a system called GRACE, which stands for Graph-Adaptive Horseshoe for Compositional Regression. Instead of forcing the data to fit a static map of evolutionary relationships, this new approach lets the data itself draw the map. The researchers applied this method to a study of adults who were at risk for diabetes but had not yet developed the disease. They analyzed samples of plaque from below the gumline to see which bacteria were associated with insulin resistance, a condition where the body struggles to regulate blood sugar. By using a flexible, learning-based model, the team discovered that the bacteria linked to insulin resistance formed groups based on their actual effect on the body, not just their evolutionary history.

The researchers tested their new method using simulated data that mimicked the messy, zero-filled nature of real microbiome samples. In these tests, they compared GRACE against existing methods that rely on fixed graphs derived from phylogenetic trees or co-occurrence patterns. The results showed that when the fixed graphs did not match the true biological relationships, the older methods struggled to find the correct bacteria or to reconstruct the network of connections. GRACE, however, adapted to the data. It successfully identified the correct groups of bacteria and recovered the underlying structure of their relationships, even when the initial assumptions about how the bacteria were related were wrong. The method proved robust even when the data contained many missing values, a common issue in microbiome studies where some bacteria are too rare to be detected in every sample.

When the team applied GRACE to the real data from the ORIGINS study, which involved 152 adults, the findings offered a clearer picture of the oral microbiome's role in metabolic health. The model identified specific bacteria associated with insulin resistance, including Tannerella forsythia and Selenomonas artemidis. These bacteria are known to be involved in gum disease, but the new analysis showed how they clustered together based on their shared impact on blood sugar regulation. Crucially, the network of connections that GRACE built between these bacteria looked very different from the traditional evolutionary tree. The bacteria that the model grouped together were not necessarily the closest relatives in the tree of life. Instead, they were grouped because they moved in the same direction regarding the health outcome: some bacteria were linked to higher insulin resistance, while others were linked to lower resistance.

The study also revealed that the way bacteria are connected in the body is not a simple reflection of their evolutionary past. The researchers found that the "distance" between bacteria in terms of their effect on insulin resistance did not always match their distance on the evolutionary tree. For instance, some bacteria that were evolutionarily close were far apart in their relationship to the disease, while others that were distant relatives acted in unison. This suggests that relying solely on evolutionary trees to understand microbiome data can lead researchers to miss important connections. The new method allows the data to speak for itself, creating a map of relationships that is driven by the actual health outcome rather than by a pre-existing theory.

The researchers emphasized that the graph produced by GRACE should not be viewed as a map of physical interactions between bacteria, such as who eats whom or who helps whom survive. Instead, it is a summary of how different bacteria behave in relation to a specific health condition. It is a hypothesis-generating tool that highlights which groups of bacteria share a common influence on the body. In the case of the ORIGINS study, the model grouped bacteria into modules that aligned with known biological functions, such as those involved in periodontal disease and metabolic dysfunction. This alignment with existing medical knowledge gave the researchers confidence that the model was capturing real biological signals rather than random noise.

One of the key strengths of this approach is its ability to handle uncertainty. The model does not just produce a single, rigid answer; it provides a range of possibilities and indicates how confident it is in each connection. This is particularly important in complex biological systems where data can be sparse and noisy. By allowing the connections between bacteria to be learned from the data, the method avoids the pitfalls of forcing the data into a box that might not fit. The researchers noted that when they were unsure about the initial relationships between bacteria, using a weaker prior assumption allowed the model to learn the structure directly from the observations, leading to better results.

The application of this method to the oral microbiome data provided a more nuanced understanding of the link between gum health and diabetes. The study confirmed that certain bacteria, like Tannerella forsythia, are consistently associated with higher levels of insulin resistance. It also highlighted that these bacteria do not act in isolation but are part of a larger community structure that influences the outcome. The model showed that the bacteria associated with insulin resistance formed distinct clusters that were separate from those associated with healthy glucose regulation. This level of detail helps researchers move beyond simply listing which bacteria are present and start understanding how the community as a whole functions.

The researchers concluded that while evolutionary trees are useful for understanding the history of life, they are not always the best guide for understanding how organisms interact with human health. The fixed graphs used in previous studies often failed to capture the true relationships that drive disease outcomes. By developing a method that adapts to the data, the team has provided a new tool for researchers to explore the microbiome. This tool can help identify which bacteria are truly important for a specific condition and how they relate to one another in that context. The findings suggest that the future of microbiome research lies in flexible, outcome-driven approaches that let the data reveal its own structure, rather than forcing it to conform to pre-existing maps.

The study also addressed the practical challenges of working with microbiome data, such as the presence of many zero values when bacteria are not detected. The new method proved to be stable and accurate even in these difficult conditions, outperforming other techniques that struggled with the noise. This robustness makes it a promising candidate for future studies where data quality might vary. The researchers plan to extend their work to model the raw counts of bacteria directly and to incorporate multiple sources of network information, which could further improve the accuracy and reliability of the results.

Ultimately, this work represents a shift in how scientists approach the complexity of the microbiome. It moves away from the idea that there is a single, fixed map of relationships that applies to all situations. Instead, it embraces the idea that the relationships between bacteria are fluid and depend on the specific context, such as the health condition being studied. By using a method that learns these relationships from the data, researchers can gain a deeper and more accurate understanding of the role the microbiome plays in human health. The study of the oral microbiome and its link to diabetes is just one example of how this approach can be applied to other areas of medicine, potentially uncovering new insights into the complex web of life within us.

Technical Summary: Graph-Adaptive Horseshoe for Compositional Regression (GRACE)

Problem Statement
Compositional data, characterized by a unit-sum constraint (e.g., microbiome abundances), present unique challenges for variable selection in regression. Traditional methods fail because features are inherently dependent. While existing approaches attempt to incorporate structural information—such as phylogenetic trees or ecological distances—via fixed similarity matrices, these static graphs often fail to reflect outcome-relevant relationships. As demonstrated in the paper's motivating example using oral microbiome data, evolutionary distance (phylogeny) and co-occurrence patterns (Bray–Curtis) frequently diverge from the actual associations with clinical outcomes like insulin resistance. Furthermore, existing Bayesian methods that utilize fixed graphs (e.g., BAZE, BCGLM) lack the ability to adapt the shrinkage structure to the specific outcome data, making it difficult to distinguish between similarity induced by the prior graph and similarity supported by the observed data.

Methodology: GRACE
The authors propose GRACE (GRaph-Adaptive horseshoe for Compositional rEgression), a fully Bayesian hierarchical framework designed to simultaneously perform variable selection, enforce compositional constraints, and learn an outcome-driven shrinkage graph.

Compositional Reparameterization: To address the constraint $\sum \beta_j = 0$ , the model employs a novel linear reparameterization of the regression coefficients $\beta$ . By defining a transformation matrix $T$ and utilizing a Cholesky decomposition, the prior on $\beta$ is constructed such that the compositionality constraint is satisfied in probability as a scaling parameter $c \to \infty$ .
Structured Horseshoe Prior: The model places a horseshoe prior on the regression coefficients to induce sparsity. Unlike standard horseshoe priors, the local shrinkage parameters $\lambda_j$ are not independent. Instead, the log-shrinkage parameters $\psi_j = \log(\lambda_j)$ are modeled using an intrinsic autoregressive (CAR) prior. This encourages taxa connected in the graph to have similar shrinkage levels, thereby smoothing regression effects across related features.
Adaptive Graph Learning: The core innovation is the treatment of the graph edge weights $\omega_{jk}$ $ω_{j k}$ as unknown parameters rather than fixed inputs. The authors place a scaled beta2 (SBeta2) prior on each edge weight. This distribution allows for:
- Outcome-specific adaptation: Edge weights adjust based on the data to reflect true effect-sharing relationships.
- Uncertainty quantification: The posterior distribution of $\omega_{jk}$ provides credible intervals, distinguishing data-supported connections from those driven solely by the prior.
- Flexibility: The prior can be centered at a reference graph (e.g., phylogenetic) if available, but the hyperparameters allow the model to downweight the prior if the data suggests a different structure.
Posterior Inference: To handle high-dimensional inference, the authors developed a customized Gibbs sampler. For parameters lacking closed-form conditionals (specifically the log-shrinkage parameters $\psi$ ), they utilize Elliptical Slice Sampling (ESS) combined with coordinate-wise slice updates to ensure efficient mixing and scalability.

Key Contributions

Joint Learning: GRACE is the first method to jointly enforce compositional constraints, perform sparse variable selection, and learn an outcome-informed shrinkage graph within a unified Bayesian framework.
Adaptive Shrinkage: By learning edge weights via SBeta2 priors, the method moves beyond fixed similarity matrices, allowing the graph structure to adapt to the specific outcome being predicted.
Robustness to Misspecification: The framework provides a mechanism to quantify uncertainty in the graph structure, offering a safeguard against the use of misspecified external graphs (e.g., phylogenetic trees that do not align with functional outcomes).
Computational Efficiency: The development of a hybrid Gibbs sampler incorporating ESS enables the application of these complex hierarchical models to high-dimensional microbiome datasets.

Results

Simulation Studies: Extensive simulations demonstrate that GRACE achieves competitive predictive accuracy and improved graph recovery compared to existing methods (including GCLR, BAZE, and BCGLM). Notably, GRACE maintains robustness under graph misspecification (e.g., low-overlap or hub-structure perturbations) and varying levels of zero-inflation, which are common in microbiome data. While graph recovery degrades when the prior graph is severely misaligned with the outcome, the method's predictive performance remains stable.
ORIGINS Study Application: Applied to oral microbiome data from the ORIGINS study (predicting insulin resistance via HOMA-IR), GRACE identified taxa associated with the outcome, including Tannerella forsythia and Selenomonas artemidis. Crucially, the learned graph revealed that taxa associated with insulin resistance formed modules that differed substantially from phylogenetic or co-occurrence networks. The resulting graph organized taxa into "effect-sharing modules" (e.g., a positive-effect module containing Selenomonas and Tannerella) that were enriched for periodontal and metabolic relevance, providing a hypothesis-generating summary of outcome-driven relationships.

Significance and Claims
The paper claims that fixed predictor graphs, while useful for regularization, may not faithfully represent the relationships between features that are relevant to a specific outcome. GRACE addresses this by shifting from static, pre-defined structures to adaptive, outcome-informed approaches. The authors emphasize that the learned graph should be interpreted not as a microbial interaction network, but as a summary of how taxa relate to the outcome through shared regression effects. The study underscores the necessity of methods that can distinguish between structural similarity (e.g., evolutionary distance) and functional similarity (outcome relevance), particularly in high-dimensional compositional settings where such distinctions are critical for biological insight.

More like this