Sparse Parallel Independent Component Analysis and Its Application to Identify Stable and Replicable Imaging-genomic Association Patterns in UK Biobank

Kuaikuai Duan, Jiayu Chen, Zening Fu, Rogers F. Silva, Vince D. Calhoun, Michela Dell’Orco, Nora I. Perrone‐Bizzozero, Yuhui Du, Wenhao Jiang, Jingyu LiumedRxiv 2022 · DOI

Abstract

Data fusion analyses of brain imaging and genomics enable the linking of genomic factors to brain patterns. Due to the small to modest effect sizes of common genetic variants, it is usually challenging to reliably identify relevant genetic factors from the rest of the genome with the typical sample size in neuroimaging studies. To alleviate this problem, we propose sparse parallel independent component analysis (spICA) to leverage the sparsity of individual genomic sources, building upon the existing parallel independent component analysis (pICA) algorithm. Sparsity is enforced by performing Hoyer projection on the estimated independent sources. Simulation results demonstrate that spICA yields improved recovery of imaging-genomic associations and sources compared to pICA. We applied spICA to whole-brain gray matter volume (GMV) and whole-genome single nucleotide polymorphisms (SNPs) data of five different sets of 24,985 discovery samples in the UK Biobank. We identified three GMV sources significantly and stably associated with one SNP source and replicated these associations. GMV sources highlighted frontal, parietal, and temporal regions. Their corresponding loadings on individuals were related to multiple cognitive measures, and the temporal region interacts with age influencing cognition. The SNP component underscored SNPs in chromosome 17 that were enriched in the inflammation response pathway and in regulation effect in the prefrontal cortex via gene expression, methylation, transcription expression, and isoform percentage.

Introduction

The field of imaging genomics aims to map how genetic variations influence the structure and function of the human brain. However, this task is akin to finding a needle in a haystack; common genetic variants, such as single nucleotide polymorphisms (SNPs), typically exert very small individual effects. When millions of these markers are analyzed alongside high-resolution neuroimaging data, standard statistical methods often suffer from overfitting—identifying patterns in one cohort that fail to replicate in another.

While existing tools like parallel Independent Component Analysis (pICA) are effective at identifying independent patterns across different data modalities simultaneously, they often struggle to isolate specific genomic signals from background noise. To address this, researchers have developed sparse parallel independent component analysis (spICA), a robust framework designed to handle high-dimensional genomic data while ensuring results are biologically meaningful and replicable.

Methodology

The spICA Algorithm

Building upon the traditional pICA framework, spICA utilizes the "Infomax" principle to identify components that are maximally independent within each data type (e.g., brain regions and SNPs) while maximizing the correlation between them. The critical innovation in spICA is the enforcement of sparsity through the integration of the Hoyer projection.

Sparsity and Denoising

Unlike common sparsity constraints such as $L_1$ regularization, the Hoyer index is scale-invariant. This makes it uniquely suited for ICA, where the scale of components is arbitrary. By mathematically forcing genomic components to be sparse, the algorithm effectively "denoises" the data. It focuses on the most influential SNPs and suppresses the background genomic profile, allowing for the recovery of clearer biological signals.

Experimental Setup

Simulations

The researchers benchmarked spICA against standard pICA across various simulated scenarios, testing different sample sizes, noise levels, and signal strengths. These tests evaluated the algorithm's accuracy in recovering "ground truth" brain-genome associations.

UK Biobank Analysis

The method was applied to a large-scale population study involving 35,692 European Caucasians from the UK Biobank (UKB). The analysis integrated:

  • Whole-brain Gray Matter Volume (GMV): Derived from T1-weighted MRI.
  • Whole-genome SNPs: Over 1.15 million genetic markers.
  • Validation: The dataset was split into five sets of discovery (70%, approx. 24,985 samples) and replication (30%) samples to rigorously test the stability of the findings.

Key Results and Biological Insights

Superior Performance and Stability

In simulations, spICA achieved a correlation of 0.96 with ground-truth brain components, significantly outperforming the 0.73 achieved by standard pICA. In the UKB data, the algorithm identified three GMV sources (highlighting the superior temporal gyrus, parietal lobule, and medial frontal gyri) that were stably associated with one SNP source across all discovery sets.

The Chromosome 17 Inflammation Pathway

The identified SNP component consistently pointed to a "hotspot" on Chromosome 17. These SNPs were found to be enriched in pathways related to the inflammation response. Biological validation confirmed that the top SNPs identified by spICA act as cis-eQTLs, regulating gene expression, methylation, and protein isoforms within the prefrontal cortex.

The loadings of the identified brain regions were significantly correlated with cognitive measures, such as fluid intelligence and memory. Interestingly, the study revealed that the relationship between the temporal brain region and cognitive scores changed with age, suggesting that these genetic factors may influence how the brain ages and how that aging impacts cognitive function.

Discussion and Future Directions

The development of spICA represents a significant step forward in imaging genomics, providing a tool that cuts through genomic noise to find replicable brain-gene links. While the current study focused on European populations to avoid population stratification, future research must validate these patterns in more diverse global cohorts.

Furthermore, while the Hoyer projection simplifies the enforcement of sparsity, selecting the optimal level of sparsity still requires careful tuning. The authors suggest that spICA could be extended to "N-way" fusion, potentially linking brain imaging, genomics, and proteomics in a single, unified model to further unravel the complexities of human biology.

Summary updated

Abstract

Abstract Data fusion analyses of brain imaging and genomics enable the linking of genomic factors to brain patterns. Due to the small to modest effect sizes of common genetic variants, it is usually challenging to reliably identify relevant genetic factors from the rest of the genome with the typical sample size in neuroimaging studies. To alleviate this problem, we propose sparse parallel independent component analysis (spICA) to leverage the sparsity of individual genomic sources, building upon the existing parallel independent component analysis (pICA) algorithm. Sparsity is enforced by performing Hoyer projection on the estimated independent sources. Simulation results demonstrate that spICA yields improved recovery of imaging-genomic associations and sources compared to pICA. We applied spICA to whole-brain gray matter volume (GMV) and whole-genome single nucleotide polymorphisms (SNPs) data of five different sets of 24,985 discovery samples in the UK Biobank. We identified three GMV sources significantly and stably associated with one SNP source and replicated these associations. GMV sources highlighted frontal, parietal, and temporal regions. Their corresponding loadings on individuals were related to multiple cognitive measures, and the temporal region interacts with age influencing cognition. The SNP component underscored SNPs in chromosome 17 that were enriched in the inflammation response pathway and in regulation effect in the prefrontal cortex via gene expression, methylation, transcription expression, and isoform percentage.
BibTeX
@misc{duan2022sparse,
  author = {Kuaikuai Duan and Jiayu Chen and Zening Fu and Rogers F. Silva and Vince D. Calhoun and Michela Dell’Orco and Nora I. Perrone‐Bizzozero and Yuhui Du and Wenhao Jiang and Jingyu Liu},
  title = {{Sparse Parallel Independent Component Analysis and Its Application to Identify Stable and Replicable Imaging-genomic Association Patterns in UK Biobank}},
  year = {2022},
  doi = {10.1101/2022.06.27.22276981},
  howpublished = {medRxiv},
}
Figure