TL;DR
This paper introduces novel integrative analysis methods for mental disorder omics data, combining horizontal (across diseases) and vertical (across data types like gene expression and CNVs) approaches. By integrating data from bipolar disorder and schizophrenia, the study demonstrates that these advanced techniques yield more biologically sensible findings than traditional single-disease, single-omics analyses, paving the way for deeper insights into mental health research.
What Problem Does This Paper Address?
Mental disorders like bipolar disorder and schizophrenia are complex, multifactorial conditions. While omics technologies (like gene expression and copy number variation - CNV) have generated vast amounts of data, most research analyzes these data in isolation: either focusing on a single disorder or a single omics type. This siloed approach limits our understanding of the intricate biological mechanisms underlying these diseases, potentially missing crucial cross-disease connections and interactions between different molecular layers. The paper addresses this gap by proposing and demonstrating integrative analysis methods that leverage both horizontal integration (combining data from related disorders) and vertical integration (combining different omics data types for the same disorder) to uncover more comprehensive and biologically meaningful insights.
Methods
The study proposes and applies three main types of integrative analysis:
- Vertical Integrative Analysis of Multi-omics Data (A1): This approach integrates gene expression and CNV data for a single disease. It involves:
- Sparse Principal Component Analysis (SPCA): Applied to gene expression data to reduce dimensionality and handle high correlations, focusing on important variables.
- "Reversed Regression": CNV data is regressed onto the sparse principal components of gene expression. This step aims to decompose CNV information into that overlapping with gene expression and that which is independent.
-
Logistic Regression: The independent CNV component, along with the gene expression PCs, is used to build an outcome model for disease prediction. This method accounts for potential collinearity and aims for interpretability. Alternative approaches (A2, A3, A4) are also discussed for comparison, which perform SPCA separately on gene expression and CNVs without explicit integration of overlapping information.
-
Horizontal Integrative Analysis for Disease Marker Identification (B1, B2): This method integrates data from multiple related diseases (bipolar disorder and schizophrenia) to identify common disease markers. It uses penalized logistic regression, incorporating a novel penalty term that encourages similarity between the models for the two diseases. Two types of penalties are explored:
- Magnitude-based shrinkage penalty (B1): Promotes similar magnitudes of coefficients for the same marker across diseases.
-
Sign-based shrinkage penalty (B2): Promotes similar signs of coefficients for the same marker across diseases. A benchmark approach (B3) analyzing each disease separately with Lasso is used for comparison.
-
Horizontal Integrative Analysis of Gene Expression-CNV Regulations (C1, C2): This approach integrates gene expression and CNV data across related diseases to identify regulatory relationships. It uses penalized regression to model gene expression as a function of CNVs, again incorporating penalties that promote similarity in the estimated regulatory coefficients between diseases. Similar magnitude-based and sign-based penalties are used. A benchmark approach (C3) analyzing each disease separately with Lasso is used for comparison.
Datasets
The study utilizes data from the Stanley Medical Research Institute (SMRI), which includes:
- Gene expression measurements (Affy Hgu133A chips) for 86 subjects.
- Single nucleotide polymorphism (SNP) data (Affy SNP5.0 chip) for 153 subjects, from which CNV measurements were extracted using PennCNV.
After data matching, a total of 71 subjects (23 bipolar disorder, 24 schizophrenics, and 24 normal controls) had both gene expression and CNV measurements for 19,053 genes. The authors also conducted simulations using these real data characteristics to evaluate their methods.
Key Insights and Ideas
- Integration is Key: The core idea is that combining information across related diseases and across different omics layers can reveal more robust and biologically relevant findings than analyzing them in isolation.
- Beyond "Benchmark" Analysis: The integrative approaches consistently produced different results compared to traditional, non-integrative methods, suggesting they capture distinct biological signals.
- Borrowing Strength: Horizontal integration allows "borrowing strength" from related diseases, which is particularly useful when sample sizes are limited, leading to more stable and reliable identification of disease markers and regulatory relationships.
- Penalties for Similarity: The novel penalty terms in horizontal integration are designed to explicitly encourage similarity in model parameters (magnitudes and/or signs) between related diseases, reflecting their known biological connections.
- Decomposing CNV Information: The vertical integration method's approach to decompose CNV data into "overlapping" and "independent" components offers a nuanced way to understand how CNVs contribute to disease, both in conjunction with and independently of gene expression.
Key Results
- Vertical Integration (A1): Identified different sets of important genes compared to alternative approaches. Preliminary literature searches supported the biological relevance of some identified genes (e.g., MAPK1, YWHAE, TPH1, NGF, AKT1). Prediction performance (measured by Corrected Prediction Ratio - CPR) showed a slight advantage for the integrative approach A1 over alternatives, though CPRs were moderate due to small sample size and limited omics data.
- Horizontal Integration for Markers (B1, B2): Approaches B1 and B2 identified more overlapping gene expressions associated with both bipolar disorder and schizophrenia compared to the benchmark B3. Simulation studies showed that B1 and B2 generally outperformed B3 in identifying true associations with lower false positive rates, especially when diseases shared similarity.
- Horizontal Integration for Regulations (C1, C2): Integrative analyses C1 and C2 identified different gene expression-CNV regulations than the benchmark C3. Distances between coefficient matrices for C1/C2 were smaller between diseases than between disease and normal, supporting the biological relatedness of bipolar disorder and schizophrenia. Simulation studies demonstrated that C1 and C2 achieved better performance (higher TPR, lower FPR) in identifying gene expression-CNV regulations.
Limitations and Future Work
- Sample Size: The authors acknowledge that the sample size is limited, which impacts the robustness of findings and prediction performance. Larger datasets would allow for more definitive conclusions and potentially the analysis of all genes without prescreening.
- Data Types: The study primarily focuses on gene expression and CNVs. Future work could incorporate other omics data types (e.g., DNA methylation, proteomics) and clinical, environmental, and socioeconomic factors for a more comprehensive model.
- Inference Challenges: High-dimensional inference for penalized models remains a challenging area, and the current study focuses on estimation and prediction rather than formal statistical inference.
- Generalizability: While the methods are presented for bipolar disorder and schizophrenia, the authors suggest they are applicable to other mental disorders and omics data types.
- Validation: The authors emphasize that functional/experimental validation and additional independent data are needed to confirm the findings.
The paper provides a strong foundation and a prototype for applying integrative analysis techniques to mental disorder research, highlighting its potential to uncover novel insights into disease etiology, biomarkers, and treatment.
Summary updated