Abstract
Motivated by imaging genetics, this paper introduces a semi-nonparametric varying coefficients modeling framework to reveal varying associations between genetic markers and imaging responses. We aim to conduct a comprehensive theoretical analysis of estimation and inference procedures applicable to these models. By employing the kernel machine method, we estimate unknown varying coefficient functions and derive their representer theorem. We also establish the theoretical properties of these estimated functions, including their rate of convergence, Bahadur representation, point-wise limit distributions, and confidence intervals. Additionally, we propose test statistics under a linear mixed effects model framework to assess the significance of all varying coefficients, taking into account within-subject dependence. The efficacy of our proposed methodology is demonstrated through simulation studies and an application to data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).
Introduction
In the rapidly evolving field of neuroimaging, researchers are constantly looking for better ways to link our genetic code to the physical structure of our brains. Imaging genetics aims to identify the "missing links" between high-dimensional genetic markers, such as Single Nucleotide Polymorphisms (SNPs), and complex imaging phenotypes like the shape of the hippocampus. However, this field faces two major hurdles: the functional nature of imaging data (which varies across surfaces or volumes) and the mismatch of traditional models. Linear models are often too rigid, while fully non-parametric models can be computational "black boxes" that require massive datasets.
To address these challenges, we propose a Semi-nonparametric Varying Coefficient (SVC) model. This framework combines the flexibility of non-parametric kernel machines with the interpretability of varying-coefficient models, allowing for the mapping of complex, non-linear genetic effects onto functional imaging data while rigorously controlling for clinical covariates.
Methodology: The SVC Framework
The core of our approach is the SVC model defined as:
$$y_i(s) = h(x_i, s) + z_i^T \gamma(s) + \epsilon_i(s)$$
where:
- $y_i(s)$: Represents the imaging trait at a specific location $s$ (e.g., a point on a brain surface).
- $h(x_i, s)$: A non-parametric function capturing the effect of genetic markers $x_i$.
- $z_i^T \gamma(s)$: A varying-coefficient term for clinical variables $z_i$, where the effect $\gamma(s)$ varies across different locations on the brain surface.
Estimation and Inference
We utilize the Least Squares Kernel Machine (LSKM) method within a Reproducing Kernel Hilbert Space (RKHS). This allows us to model non-linear genetic effects without pre-specifying the exact functional form. A key theoretical contribution is the derivation of a Representer Theorem, which proves that the optimal solution in an infinite-dimensional space can be found using finite-dimensional matrix calculations, ensuring computational feasibility. Furthermore, we developed score tests to assess significance by treating the non-parametric components as random effects within a linear mixed-model framework.
Theoretical Insights: The Phase Transition
A significant theoretical finding in this work is the "phase transition" phenomenon in convergence rates. We discovered that once the number of measurement points on a brain surface (design points) reaches a certain threshold, increasing the resolution further does not improve the model's convergence rate. At this stage, the primary bottleneck becomes the number of subjects ($n$) rather than the spatial resolution of the scan ($m$).
Experimental Results and Application
The methodology was validated using simulation studies and real-world data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), involving 606 subjects. We focused on hippocampal morphometry and 1,703 Linkage Disequilibrium (LD) blocks across the genome.
Key Findings
- Demographic Impact: Clinical variables such as age and gender, along with their interactions, were highly significant ($p < 10^{-16}$) for hippocampal structure.
- Genetic Architecture: After Bonferroni correction, the model identified 109 significant genetic blocks for the left hippocampus and 245 for the right.
- Asymmetry: Interestingly, only 27 blocks were significant for both sides, suggesting that the left and right hippocampi have distinct genetic architectures.
- Alzheimer's Links: The overlapping significant regions included the 19q13.32 region, which contains the APOE gene, the most well-known genetic risk factor for Alzheimer’s Disease. Other identified blocks were linked to amyloid-beta levels and cognitive decline.
Conclusion
The SVC model provides a robust bridge between machine learning flexibility and statistical rigor. By grouping SNPs into LD blocks, the model captures collective genetic influences that are more biologically meaningful than single-mutation analysis. Future work will focus on scaling these methods to higher-dimensional 3D volumes and longitudinal 4D data to track brain changes over time.
Summary updated