Session introduction

Heng Huang, Li Shen, Paul M. Thompson, Kun Huang, Junzhou Huang, Lin Yang

Introduction

This paper serves as the official introduction to the "Imaging Genomics" session at the 2018 Pacific Symposium on Biocomputing. It highlights a significant shift in biomedical research: the integration of high-dimensional medical imaging—such as MRI, PET, and histopathology—with "omics" data (genomics, proteomics, and transcriptomics). This convergence is essential for the next generation of precision medicine, providing a more holistic view of disease characterization by linking genetic variations (genotypes) to their physical manifestations (phenotypes).

The Computational Bottleneck

Historically, medical imaging and genomic research have operated in silos. However, understanding complex diseases like Alzheimer’s or various cancers requires a unified approach. The primary obstacle is a "computational bottleneck" caused by the sheer scale and heterogeneity of these massive datasets. To address this, the session introduced seven studies focused on bridging the gap using advanced data science tools, moving the field from simple observations toward actionable clinical insights.

Methodological Innovations

The session featured a variety of cutting-edge computational approaches designed to handle multi-modal data:

  • Deep Learning & CNNs: Utilized for survival prediction and automated nuclear segmentation in histopathology images.
  • Bi-directional Convolutional Recurrent Neural Networks (CRNNs): Employed to predict epigenetic states, such as DNA methylation, directly from non-invasive brain MRI scans.
  • Low-Rank Multi-Task Learning: A method used to identify associations between thousands of Single Nucleotide Polymorphisms (SNPs) and quantitative imaging measures while accounting for the complex underlying correlation structure of the data.
  • Discriminative Bag-of-Nuclear-Words (DBoNW): A novel representation technique that summarizes complex tissue images by treating cell nuclei as "codewords" to predict genomic markers.
  • Harmonized Pipelines: The introduction of the ENIGMA-rsfMRI pipeline, which standardizes the processing of resting-state functional MRI data across global research sites to ensure reproducibility.

Data Ecosystem

The research presented leveraged several of the most significant "big data" repositories in modern biomedicine, including the Alzheimer’s Disease Neuroimaging Initiative (ADNI) for whole-genome sequencing, The Cancer Genome Atlas (TCGA) for molecular profiling, and The Cancer Imaging Archive (TCIA) for corresponding medical imagery. Additionally, the ENIGMA Consortium provided the framework for large-scale meta-analysis of neuroimaging genetics.

Key Insights and Results

The studies yielded several surprising and impactful findings:

  • The Influence of "Silent" Mutations: Research by Miller et al. suggested that synonymous variants—genetic changes that do not alter protein sequences—may still influence Alzheimer’s Disease through "codon bias," which affects protein synthesis efficiency.
  • Imaging vs. Molecular Features: In a "trans-omics" comparison by Srivastava et al., histology images actually outperformed molecular features when predicting cancer stages. However, molecular data remained superior for identifying specific biological subtypes (e.g., PAM50).
  • Cross-Modality Prediction: The session demonstrated that "hidden" molecular states, such as the methylation of the MGMT promoter in brain tumors, can now be predicted with reasonable accuracy using non-invasive imaging like MRI.
  • Survival and Genetic Influence: Huang et al. successfully integrated TCGA imaging and molecular data into a deep learning pipeline that improved survival outcome predictions. Meanwhile, Adhikari et al. used the ENIGMA pipeline to replicate findings showing a moderately strong genetic influence on the brain's resting-state signals.

Limitations and Future Directions

While the integrated approach is promising, the field remains in its early stages. The authors identified three critical areas for future growth:

  1. Computational Scaling: Algorithms must evolve to handle the "unprecedented scale" of imaging-genomics data.
  2. Open Exchange: Progress depends on the willingness of researchers to share not only raw data but also the complex software and innovative technologies required to process it.
  3. Causal Discovery: There is a pressing need to move beyond simple association studies toward rigorous, data-driven discovery that can prove causal links between specific genes and imaging phenotypes.

Summary updated

BibTeX
@inproceedings{huang2018session,
  author = {Heng Huang and Li Shen and Paul M. Thompson and Kun Huang and Junzhou Huang and Lin Yang},
  title = {{Session introduction}},
  booktitle = {PSB},
  year = {2018},
}
Figure