Introduction
Imaging genomics is a rapidly evolving interdisciplinary field that integrates high-dimensional "omics" data—such as genomics, transcriptomics, and proteomics—with complex medical imaging. This synthesis aims to uncover the underlying biological mechanisms of disease. This session summary, derived from the 2017 Pacific Symposium on Biocomputing, highlights how emerging machine learning and network analysis tools are addressing the "missing heritability" of complex conditions like Alzheimer’s disease, schizophrenia, and lung cancer.
The Challenge: The Curse of Dimensionality
For decades, medical researchers have grappled with the "curse of dimensionality." While modern technology allows us to sequence entire genomes (millions of data points) and capture high-resolution brain scans (millions of voxels), connecting a specific genetic mutation to a specific structural change in the brain remains a monumental task.
The core statistical challenge is that the number of variables ($p$) vastly outweighs the number of patients ($n$). Traditional statistical methods often falter in this $p \gg n$ environment, leading to either a high rate of false positives or the failure to detect subtle but significant biological connections. The research presented here moves beyond simple "one gene, one trait" associations, favoring a "systems biology" approach that models the interactive web of genes, proteins, and imaging features.
Computational Strategies and Methodology
To navigate these massive datasets, researchers have developed several sophisticated computational strategies:
- Multivariate and Sparse Models: Techniques such as Discriminative Sparse Canonical Correlation Analysis (DSCCA) and MT-CoReg are employed to identify the strongest correlations between disparate datasets (e.g., thousands of proteins vs. hundreds of brain regions) while filtering out background noise.
- Network Analysis: Rather than analyzing genes in isolation, researchers utilize Weighted Gene Co-expression Network Analysis (WGCNA) and Spatial Expression Networks (SEN) to group genes into functional "modules."
- Adaptive Statistical Testing: New frameworks, such as the adaptive test based on the Proportional Odds Model (POM), allow for robust association testing even when genetic variants exceed the sample size.
- Morphological Feature Extraction: In oncology, automated systems now analyze pathology slides to calculate hundreds of morphological features—including shape, texture, and density—to correlate physical tumor characteristics with underlying genetic mutations.
Data Repositories and Open Science
The advancement of imaging genomics relies heavily on large-scale, open-science repositories that provide the necessary data for cross-validation and discovery:
- ADNI (Alzheimer’s Disease Neuroimaging Initiative): Facilitates the study of brain connectivity and proteomics in the context of aging.
- TCGA (The Cancer Genome Atlas): Enables researchers to link lung adenocarcinoma morphology with mRNA expression.
- MCIC (Mind Clinical Imaging Consortium): A specialized dataset for schizophrenia research, combining fMRI and genotyping.
- AHBA (Allen Human Brain Atlas): Provides a foundation for whole-genome, whole-brain transcriptomic mapping.
Key Insights: Imaging as a Quantitative Phenotype
A central theme of this research is the redefinition of medical imaging. Images are no longer viewed merely as diagnostic snapshots but as "quantitative phenotypes." By converting a brain scan or a pathology slide into a set of discrete numerical values (morphometry or connectivity), researchers can treat the image as a biological measurement comparable to a gene expression level.
Furthermore, by investigating epistatic effects (gene-gene interactions) and enrichment (how groups of genes affect specific brain regions), researchers are beginning to account for the "missing heritability" that single-variant studies could not explain. Notably, the integration of proteomics—measuring protein levels in blood and spinal fluid—has proven to be more discriminative for conditions like Alzheimer's than DNA analysis alone.
Results and Clinical Implications
- Lung Cancer Survival: Analysis identified 283 morphological features from tumor images that correlate strongly with patient survival and specific gene co-expression clusters.
- Alzheimer’s Discrimination: The DSCCA model successfully identified imaging-proteomic patterns capable of distinguishing healthy controls from patients with mild cognitive impairment.
- Brain Modules: Network analysis identified three stable gene clusters in the human brain, mapping the distribution of molecular machinery across anatomical regions.
Limitations and Future Directions
Despite these advancements, the primary hurdle remains sample size. Even with large-scale repositories, studies often involve only 100 to 800 participants—a small number relative to the millions of genetic markers being tested. This disparity continues to limit statistical power.
The future of the field lies in a holistic Systems Biology approach, modeling the entire biological hierarchy from DNA to proteins, cells, and organs simultaneously. The ultimate goal is to transition from identifying simple associations to understanding the causal mechanisms of disease, paving the way for more precise diagnostic and preventive medicine.
Summary updated