Source-linked AI summary
Joint analysis of functional genomic data and genome-wide association studies of 18 human traits
Joseph K. Pickrell
TL;DR
Choosing which genomic annotations are relevant for interpreting GWAS associations remains challenging because multifactorial-trait loci often lie outside protein-coding exons. This paper develops a hierarchical model that jointly analyzes GWAS and genomic annotations, applying it across 18 traits and 450 annotations to identify biologically interpretable enrichment patterns.
Problem
GWAS loci for multifactorial traits often have small effects and lie outside protein-coding exons, motivating methods that incorporate other genomic elements.
Method
A hierarchical model jointly analyzes GWAS and genomic annotations, allowing association probabilities for genomic blocks and SNPs to depend on regional and SNP-level annotations.
Results
Across 18 traits, the model scanned 450 annotations and identified sparse, biologically interpretable sets of annotations relevant to each phenotype without prior biological knowledge.
Takeaways & Limitations
The model provides a way to identify genomic annotations most relevant to the biology of a phenotype from genome-wide association patterns.
Takeaways & Limitations
The model assumes a genomic region contains at most one causal SNP, an assumption that may become untenable as GWAS sample sizes increase.
Abstract
from arXiv · showhide
Annotations of gene structures and regulatory elements can inform genome-wide association studies (GWAS). However, choosing the relevant annotations for interpreting an association study of a given trait remains challenging. We describe a statistical model that uses association statistics computed across the genome to identify classes of genomic element that are enriched or depleted for loci that influence a trait. The model naturally incorporates multiple types of annotations. We applied the model to GWAS of 18 human traits, including red blood cell traits, platelet traits, glucose levels, lipid levels, height, BMI, and Crohn's disease. For each trait, we evaluated the relevance of 450 different genomic annotations, including protein-coding genes, enhancers, and DNase-I hypersensitive sites in over a hundred tissues and cell lines. We show that the fraction of phenotype-associated SNPs that influence protein sequence ranges from around 2% (for platelet volume) up to around 20% (for LDL cholesterol); that repressed chromatin is significantly depleted for SNPs associated with several traits; and that cell type-specific DNase-I hypersensitive sites are enriched for SNPs associated with several traits (for example, the spleen in platelet volume). Finally, by re-weighting each GWAS using information from functional genomics, we increase the number of loci with high-confidence associations by around 5%.
1 Introduction
GWAS have identified many associated genomic regions, but interpreting their molecular mechanisms remains difficult because multifactorial-trait loci are usually small-effect and outside protein-coding exons. This paper develops a hierarchical model to jointly use GWAS and diverse genomic annotations to identify relevant genomic information.
- GWAS have identified thousands of reproducible regions associated with diverse human phenotypes and diseases.
- Multifactorial-trait GWAS loci generally have small effects and lie outside protein-coding exons.
- Existing enrichment analyses examine whether strongly associated SNPs cluster in genomic regions such as exons, promoters, UTRs, regulatory regions, or cell-type-associated loci.
- These approaches generally cannot analyze more than one annotation simultaneously and do not directly estimate which of two similarly associated SNPs is more likely to influence a trait.
- Hierarchical models can learn shared properties of associated loci, and larger GWAS meta-analyses now provide tens to hundreds of loci for revisiting this strategy.
- The paper presents a hierarchical model that jointly analyzes GWAS and genomic annotations across 18 diseases and traits using 450 genome annotations for each trait.
2 Results
The hierarchical model learned trait-relevant genomic annotations from GWAS across 18 traits, combining regional and SNP-level information to identify enrichment, depletion, and candidate associations.
- 18 GWAS and 450 genomic annotations were analyzed, including annotations from many cell types.The GWAS included traits such as blood-cell traits, Crohn’s disease, BMI, lipid levels, height, bone mineral density, and fasting glucose.
- The model assigned association priors to genomic blocks and SNPs, allowing regional and SNP annotations to explain enrichment patterns across the genome.Blocks averaged 2.5 Mb, and both block-level and SNP-level probabilities were estimated from genome-wide enrichment patterns.
- HDL-associated loci were enriched in HepG2 enhancers and depleted in HepG2 repressed regions; the joint model also included coding exons and K562 repressed chromatin.Because annotations were correlated, the combined model selected representative annotations using cross-validation and conditional analysis.
- Around 2% to around 20% of trait-associated signals were estimated to be driven by nonsynonymous polymorphisms, averaging 10% across traits.Nonsynonymous variants were enriched for every trait, although enrichment was not statistically significant for all traits; synonymous variants were generally not enriched.
- Repressed chromatin was depleted for several traits, while cell type-specific open chromatin was enriched, including CD34+ cells for platelet traits and spleen for platelet density.Examples included log2 enrichment of -3.91 for red blood cell volume in repressed erythroblast-derived chromatin and 1.93 for platelet density in spleen open chromatin.
- A regional posterior probability threshold of 0.9 performed similarly to a stringent P-value threshold, and combining approaches increased identified loci by approximately 5% with false-positive rates close to zero.Functional annotations also highlighted the NR0B2 region near HDL-associated variants, including a nonsynonymous candidate with P-value 1.5 × 10^-6.
3 Discussion
The model identifies sparse, biologically interpretable genomic annotations relevant to phenotypes, while highlighting unresolved links between non-coding variants and gene expression. Its assumptions also constrain applications such as fine-mapping, especially when variants are unobserved or multiple causal variants occur in a region.
- The model scans hundreds of genomic annotations to identify a sparse set of biologically interpretable annotations without prior biological knowledge.
- Repressed chromatin is often depleted for trait-influencing SNPs, potentially reducing the SNPs considered during GWAS fine-mapping.The paper reports that approximately 60–70% of the genome falls in this annotation in any given cell type.
- Connecting non-coding regulatory variants to gene-expression changes remains an important next step for understanding how they affect phenotypes.
- The model assumes association probability depends on SNP density rather than physical region size and permits only one causal SNP per genomic region.The authors note that the single-causal-SNP assumption may become untenable as GWAS sample sizes increase.
- Fine-mapping is limited because around 20% of common SNPs are neither genotyped nor successfully imputed.The paper identifies denser genotyping or sequencing as a reason to revisit this application.
Methods
The hierarchical model estimates annotation-informed priors for associations across LD-defined genomic blocks, using summary-statistic Bayes factors and empirical Bayes learning. Penalized likelihood, cross-validation, and forward selection support a sparse model of relevant annotations.
- Hierarchical model: The genome is divided into blocks larger than LD, each modeled as containing either one causal polymorphism or none.The model distinguishes the probability that a block contains an association from the conditional probability that a SNP within it is causal.
- Bayes-factor model: The single-SNP building block uses an additive linear regression and compares null and alternative association models with a Bayes factor.The approximate Bayes factor can be computed from regression summary statistics rather than individual genotype vectors.
- Hierarchical model: The model is not a multiple-regression model jointly estimating effects of multiple SNPs on a trait.Instead, each block permits at most one associated SNP under the stated model.
- Hierarchical model: The model allows block- and SNP-level causal priors to depend on genomic annotations learned from the data.Region-level annotations include gene density, while SNP-level annotations can include nonsynonymous variants.
- Model fitting: Annotation parameters are estimated by maximizing likelihood, while model comparison shrinks them toward zero using a ridge-regression penalty.The penalty parameter is tuned by 10-fold cross-validation over chromosomal segments.
- Model choice: Forward selection adds the annotation that most improves likelihood, tests jointly significant annotations, and repeats while significance remains.This procedure is intended to produce a relatively sparse model from hundreds of annotations.
Approximating Vi
The analysis estimates SNP-effect standard errors from available summary information when direct standard errors are unavailable, adapting the variance approximation for quantitative and case-control traits.
- Variance approximation: For each SNP, the variance estimate uses external minor allele frequency, sample size, and association information when the regression standard error is rarely reported.Minor allele frequencies were computed from an external sample of the same ancestry using 1000 Genomes data.
- Variance approximation: The variance approximation is independent of measurement scale because the corresponding Z-scores are also scale-independent.
- Case-control studies: For case-control studies, the model remains otherwise identical but approximates variance using the numbers of cases and controls.The resulting variance is expressed on a log-odds scale.
- Posterior probabilities: After fitting, the model estimates regional and SNP-level posterior probabilities of association and causality from Bayes factors.The SNP-level posterior is conditional on the region containing an association, and the overall SNP posterior combines regional and conditional probabilities.
- Conditional analysis: Conditional analysis tests whether an annotation adds information beyond an annotation already included in the combined model.The procedure compares included annotations with other annotations that had significant marginal associations.
arXiv:1311.4843v4 [q-bio.GN] 25 Feb 2014
The supplementary material is identified as concerning the joint analysis of functional genomic data and genome-wide association studies of 18 human traits.
- Supplementary material: The supplementary material accompanies the study titled “Joint analysis of functional genomic data and genome-wide association studies of 18 human traits.”
- Affiliations: The listed affiliations include the New York Genome Center and Columbia University’s Department of Biological Sciences.
1 GWAS data
The study assembled publicly available GWAS summary statistics for 18 traits, including anthropometric, metabolic, lipid, blood, platelet, bone-density, and inflammatory phenotypes.
- Trait datasets: The GWAS collection included height, BMI, bone mineral density, Crohn’s disease, fasting glucose, four lipid traits, six red blood cell traits, and two platelet traits.
- Trait datasets: Height summary statistics covered 2,469,635 SNPs in an average of 129,945 individuals after removing SNPs with sample sizes below 120,000.
- Trait datasets: The bone-density datasets measured femoral-neck and lumbar-spine density across more than 2.4 million SNPs.
- Trait datasets: Crohn’s disease summary statistics included 953,242 SNPs and used 6,299 cases and 15,148 controls as per-SNP sample-size approximations.
- Trait datasets: Fasting-glucose summary statistics included 2,628,880 SNPs and used an overall sample size of 58,074 to approximate per-SNP sample sizes.
- Trait datasets: The lipid datasets covered LDL cholesterol, HDL cholesterol, triglycerides, and total cholesterol.
- Trait datasets: The blood-trait datasets included six red blood cell traits and two platelet traits with millions of genotyped or imputed SNPs.
2 Functional genomic data
The study assembled functional-genomic annotations spanning DNase-I hypersensitivity, ENCODE genome segmentations, and gene-related features. These annotations differed in genomic coverage and included data from many cell types and tissues.
- DNase-I hypersensitivity: DNase-I hypersensitivity data came from 349 Maurano samples and 53 Crawford-lab samples.The samples represented cell lines, sorted blood cells, primary fetal tissues, and specific cell types.
- DNase-I hypersensitivity: Duke DNase-I sites were exactly 150 bases long and covered approximately 1% of the genome.Coverage ranged from 0.4% to 1.9%.
- DNase-I hypersensitivity: Maurano DNase-I sites averaged 514 bases and covered approximately 2.7% of the genome.Coverage ranged from 0.9% to 5.1%.
- Genome segmentations: ENCODE genome segmentations classified non-overlapping regions as CTCF sites, enhancers, promoter-flanking regions, repressed chromatin, transcribed regions, and transcription start sites.The segmentation combined results from two algorithms.
- Genome segmentations: Repressed chromatin covered 66% of the genome on average, whereas enhancer, TSS, and promoter-flanking marks covered 0.9%, 0.7%, and 0.2%, respectively.Repressed-chromatin coverage ranged from 60% to 70% across cell lines.
- Gene annotations: Ensembl gene annotations included nonsynonymous and synonymous SNP status, with coding exons covering about 3% of the genome.3′ UTRs and 5′ UTRs covered 2% and 0.6%, respectively.
3 Imputation of summary statistics
Summary statistics were imputed for common European-population SNPs using ImpG and a phase 1 1000 Genomes reference panel. The procedure successfully recovered association statistics for around 80% of common SNPs.
- Imputation procedure: ImpG v1.0 imputed GWAS summary statistics using all European haplotypes from phase 1 of the 1000 Genomes Project.Only SNPs with minor allele frequency greater than 2% were used, under default settings.
- Imputation coverage: Around 80% of common SNP association statistics were successfully imputed overall.The reference panel comprised European individuals from 1000 Genomes phase 1.
4 Details of application of the hierarchical model
The application combined simulations, jackknife estimation, interaction analyses, and empirical calibration to assess the hierarchical model and identify high-confidence loci. The calibrated PPA threshold produced additional lipid loci while requiring an independent-study comparison for definitive evaluation.
- 4.1 Simulations: Simulations showed that model power increased with enrichment, effect size, and the number of loci.The simulated enrichment factors had log-enrichments of 0.98 and 1.80, with 100 annotations simulated for each parameter set.
- 4.2 Robustness: The model’s robustness was empirically tested by varying the user-set prior variance W and the genomic block window size in Crohn’s disease data.The supplied passage identifies these as the two user-set parameters.
- 4.3 Coding versus non-coding changes: The nonsynonymous analysis used promoter and gene-density annotations, then estimated each SNP’s posterior probability of being causal and used a 20-block jackknife for uncertainty.The corresponding standard-error calculation was based on leave-one-block-out fractions.
- 4.4 Interaction effects: Two traits showed annotation-effect sign reversals between single-annotation and combined models because of statistical interactions.For red blood cell count, fetal renal-pelvis and fetal-stomach DNase annotations were correlated, producing a combined log2 enrichment of 2.41.
- 4.5 Calibrating a significance threshold: The calibration compared P-values and posterior probabilities using lipid GWAS regions and a larger follow-up study, while noting that the datasets were not independent.The comparison was intended to match false-positive reduction, not to establish relative performance.
- 4.5 Calibrating a significance threshold: A PPA threshold of 0.9 yielded 48 HDL, 44 LDL, 51 total-cholesterol, and 30 triglyceride loci when combined with P-value discoveries.These corresponded to increases of 5, 4, 0, and 1 loci over the P-value threshold alone, averaging a 6% increase; the authors call this likely a lower bound.