Accounting for data heterogeneity in integrative analysis and prediction methods: An application to Chronic Obstructive Pulmonary Disease

Jessica Butts, Christine Wendt, R.p. Bowler, Craig P. Hersh, Qi Long, Lynn E. Eberly, Sandra E. SafoarXiv 2021 · DOI · arXiv

Introduction

Chronic Obstructive Pulmonary Disease (COPD) is a complex, multi-faceted condition where patient outcomes and disease progression often vary significantly across different demographic subgroups. Biological sex, in particular, has been identified as a major factor in disease disparity. To understand the underlying mechanisms of such diseases, researchers increasingly rely on "integrative analysis"—the simultaneous study of multiple biological "views," such as genomics and proteomics.

However, traditional integrative methods often struggle with a fundamental trade-off. They either ignore subgroup differences, potentially missing sex-specific drivers, or they analyze subgroups in isolation, which reduces statistical power. Furthermore, simply "stacking" different data types into a single analysis fails to account for the intricate relationships between biological layers. To bridge this gap, we present HIP (Heterogeneity in Integration and Prediction), a statistical framework designed to identify molecular signatures that are both shared across and specific to subgroups.

The HIP Framework: Joint Association and Prediction

HIP is a one-step joint model that addresses the limitations of existing tools by leveraging the strengths of each data view to predict clinical outcomes—in this case, airway wall thickness (AWT).

Latent Variable Decomposition

At its core, HIP assumes that the correlation between different data views (e.g., gene expression and protein levels) is driven by underlying latent "scores" ($Z^s$) unique to each subgroup. This allows the model to capture the shared information across different biological scales.

Hierarchical Sparsity and the "Common-Specific" Model

The primary innovation of HIP lies in its hierarchical sparsity penalty. The model decomposes variable effects into two distinct parts:

  1. A Common Component ($G^d$): Features that affect the outcome across all subgroups.
  2. A Subgroup-Specific Component ($\Xi^{d,s}$): Features that are unique to a specific group (e.g., only males or only females).

By employing a block $l_1/l_2$ penalty (similar to a Group Lasso), HIP encourages "row-sparsity." This mathematical approach ensures the model selects the most impactful variables across all dimensions, effectively filtering out biological noise.

Optimization and Efficiency

Unlike two-step methods that first identify patterns and then predict outcomes, HIP performs these tasks simultaneously. This joint optimization ensures that the identified molecular signatures are directly relevant to the clinical outcome. The algorithm is implemented in PyTorch, utilizing modern optimization techniques like FISTA and Adagrad to handle high-dimensional datasets efficiently.

Experimental Validation

The researchers validated HIP using both synthetic simulations and real-world data from the COPDGene Study, a longitudinal study involving 1,376 individuals. The analysis integrated:

  • Genomics: RNA-sequencing data (5,000 genes).
  • Proteomics: Protein levels (2,000 proteins).
  • Clinical Outcome: Airway Wall Thickness (AWT).

Simulation Results

In scenarios ranging from "Full Overlap" (no heterogeneity) to "Partial Overlap" (mixed features), HIP consistently outperformed existing methods like Canonical Variate Regression (CVR) and Joint Lasso. It achieved higher F1 scores and maintained lower False Positive Rates, demonstrating its robustness in identifying true biological signals amidst subgroup heterogeneity.

Key Findings in COPD

The application of HIP to the COPDGene dataset yielded significant biological insights that traditional models might have overlooked.

Predictive Accuracy

HIP demonstrated superior predictive performance, yielding a lower Mean Squared Error (MSE) compared to standard Lasso or Elastic Net models. The protein scores generated by the model were statistically significant predictors of lung damage, even after adjusting for age, smoking history, and BMI.

Shared and Sex-Specific Signatures

The model identified several key biomarkers:

  • Shared Pathways: Iron homeostasis was identified as a critical pathway for both sexes.
  • Sex-Specific Findings: For females, "granulocyte adhesion" emerged as a top pathway, suggesting that inflammatory mechanisms in COPD may be sex-dependent.
  • Specific Biomarkers: The study highlighted genes such as ADIPOR1 (males) and BCL2L1 (females), and proteins such as NPLOC4 (males) and SMAP1 (females) as primary contributors to variations in airway thickness.

Conclusion and Future Directions

HIP represents a significant advancement in the integrative analysis of complex diseases. By accounting for subgroup heterogeneity without sacrificing statistical power, it provides a more nuanced view of the molecular drivers of COPD.

While the current framework is robust, future work will focus on extending HIP to accommodate longitudinal data to track disease progression over time. Additionally, expanding the model to handle binary outcomes (e.g., disease presence vs. absence) will further enhance its utility in clinical diagnostics. Ultimately, HIP offers a powerful tool for moving toward more personalized and effective therapeutic strategies.

Summary updated

Abstract

Epidemiologic and genetic studies in chronic obstructive pulmonary disease (COPD) and many complex diseases suggest subgroup disparities (e.g., by sex). We consider this problem from the standpoint of integrative analysis where we combine information from different views (e.g., genomics, proteomics, clinical data). Existing integrative analysis methods ignore the heterogeneity in subgroups, and stacking the views and accounting for subgroup heterogeneity does not model the association among the views. To address analytical challenges in the problem of our interest, we propose a statistical approach for joint association and prediction that leverages the strengths in each view to identify molecular signatures that are shared by and specific to males and females and that contribute to the variation in COPD, measured by airway wall thickness. HIP (Heterogeneity in Integration and Prediction) accounts for subgroup heterogeneity, allows for sparsity in variable selection, is applicable to multi-class and to univariate or multivariate continuous outcomes, and incorporates covariate adjustment. We develop efficient algorithms in PyTorch. Our COPD findings have identified several proteins, genes, and pathways that are common and specific to males and females, some of which have been implicated in COPD, while others could lead to new insights into sex differences in COPD mechanisms.
BibTeX
@misc{butts2021accounting,
  author = {Jessica Butts and Christine Wendt and R.P. Bowler and Craig P. Hersh and Qi Long and Lynn E. Eberly and Sandra E. Safo},
  title = {{Accounting for data heterogeneity in integrative analysis and prediction methods: An application to Chronic Obstructive Pulmonary Disease}},
  year = {2021},
  eprint = {2111.06962},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.06962},
}
Figure