TL;DR
This paper introduces a novel method, Relative Genome Information (RGI), to quantify the "disorder" or information content of an individual's genome relative to a reference population. They demonstrate that this RGI is significantly elevated in early-onset breast cancer cases compared to controls, suggesting a strong polygenic component to the disease driven by the cumulative effect of thousands of genetic variants.
What Problem Does This Paper Address?
Many common diseases, like breast cancer, are known to have a complex genetic basis, meaning they arise from the interplay of numerous genetic variations and environmental factors. While Genome-Wide Association Studies (GWAS) have identified many genetic variants associated with disease risk, their individual contributions are often very small (low penetrance). This makes it difficult to predict an individual's susceptibility based on these known variants, leaving a significant portion of the "missing heritability" unexplained. The challenge is to find a way to quantify the cumulative effect of these numerous, low-impact genetic variations across the entire genome, rather than focusing on individual loci.
Methods
The core innovation is the development of Relative Genome Information (RGI). This measure is inspired by information theory and quantifies how "unusual" an individual's genome is compared to a reference population. Here's a breakdown:
- Reference Population: A large control dataset (the UK 1958 Birth Cohort) is used to estimate the probability of observing specific allele pairs at each genetic marker (SNP). This establishes a baseline of "normal" genomic variation.
- Relative Local Information (RLI): For each individual and each SNP, the RLI measures the "surprisal" of observing their specific allele pair, given the probability estimated from the reference population. A higher RLI means the observed allele pair is less common in the reference population.
- Relative Genome Information (RGI): The RGI for an entire genome is the sum of the RLIs across all considered SNPs. Informally, it's a measure of the overall "disorder" or deviation from the reference genome's structure.
- Expected Information Per Locus (EIL): RGI is often normalized by the number of SNPs analyzed to get EIL, allowing for comparisons across datasets of different sizes.
- Statistical Analysis: The association between EIL and disease odds was assessed using logistic generalized additive models. Differences between groups were tested using Wilcoxon rank-sum tests, with P-values adjusted for multiple comparisons using the Benjamini-Hochberg procedure.
Datasets
The study utilized several large-scale genetic datasets:
- POSH (Prospective study of Outcomes in Sporadic versus Hereditary breast cancer) cohort: ~574 early-onset breast cancer cases.
- ABCFS (Australian Breast Cancer Family Study): ~204 early-onset breast cancer cases and ~287 unaffected controls.
- WTCCC (Wellcome Trust Case Control Consortium): Two sets of controls:
- 1958 British Birth Cohort: ~2699 controls.
- UK National Blood Service (NBS) Collection: ~2501 controls.
These datasets were combined and filtered for quality control, resulting in approximately 475,000 shared SNPs across all five datasets. The HapMap2 project's reference populations (African, Asian, Caucasian) were used for initial ethnic admixture checks.
Key Insights and Ideas
- Genome-wide "disorder" as a risk factor: The central idea is that deviations from a "typical" genomic structure, when aggregated across thousands of variants, can be a significant indicator of disease risk, even if individual variants have small effects.
- Beyond individual SNPs: This approach moves beyond the limitations of GWAS, which focus on identifying individual significant associations, to capture the collective impact of many variants.
- Polygenic nature of breast cancer: The findings strongly support the notion that early-onset breast cancer is a highly polygenic disease, with its genetic basis distributed across a vast number of genomic loci.
- Information theory as a tool for genomics: The successful application of information theory concepts (like entropy and surprisal) to quantify genomic variation opens up new avenues for analyzing complex genetic traits.
Key Results
- Elevated RGI in cases: Both independent sets of early-onset breast cancer cases showed significantly higher RGI (and EIL) compared to all three control sets.
- Strong association with risk: When cases and controls were amalgamated, a strong positive association was found between EIL and breast cancer odds. The odds ratio increased sharply for EIL values above 1.175, with the highest percentiles of EIL conferring an odds ratio greater than 12 compared to the lowest 99%.
- Thousands of loci involved: The observed differences in EIL between cases and controls were attributed to the cumulative effect of an estimated 327 distinct loci on average, but this number increased dramatically to an estimated 4954 loci for the highest risk percentiles. This confirms the polygenic nature.
- Genome-wide distribution: Elevated EIL was observed across 19 out of 22 chromosomes, and significantly so on 12, indicating that the genetic contribution is distributed throughout the genome, not localized to specific regions.
- Annotation patterns: EIL was higher in intergenic regions compared to 5'/3' UTRs and exonic regions, consistent with stronger purifying selection acting on functionally important regions.
Limitations and Future Work
- Ethnic variation: The authors acknowledge that significant ethnic variation within case or control populations could confound the results. They excluded individuals showing evidence of mixed ethnicity.
- Reference population choice: While they used a large control cohort (WTCCC 1958), the accuracy of the RGI measure depends on the quality and representativeness of this reference population.
- Future Work: The paper suggests that this approach could be applied to investigate the polygenic basis of other complex diseases. Further analysis of the relationship between global genome structure and disease risk is warranted. The authors also note that while their method is computationally feasible on a desktop PC for moderate sample sizes, scaling to extremely large datasets might require further optimization.
Summary updated