Deep Learning and Classic method in GWAS

From perplexity.ai

Deep learning approaches to genome-wide association studies (GWAS) offer several advantages and disadvantages compared to classic GWAS methods. Here's a comparison of deep GWAS and classic GWAS:

Pros of deep GWAS:

1. Improved detection of complex genetic associations: Deep learning models can capture non-linear relationships and interactions between genetic variants, potentially uncovering associations that classic GWAS might miss[4].

2. Reduced multiple testing burden: Methods like DeepWAS prioritize SNPs based on functional roles, reducing the number of tests performed and potentially increasing statistical power[4].

3. Integration of functional genomic data: Deep learning approaches can incorporate additional biological information, such as regulatory data or cell-type-specific effects, to improve the interpretation of results[4].

4. Ability to handle high-dimensional data: Deep learning models are well-suited for analyzing large-scale genomic datasets with millions of variants[1].

5. Potential for improved prediction accuracy: Some deep learning methods have shown promise in improving the prediction of complex traits compared to traditional approaches[4].

Cons of deep GWAS:

1. Increased computational complexity: Deep learning models often require significant computational resources and time to train and evaluate, especially when dealing with large genomic datasets[3][4].

2. Risk of overfitting: Deep learning models with many parameters may be prone to overfitting, particularly when sample sizes are limited[1].

3. Reduced interpretability: The complex nature of deep learning models can make it challenging to interpret the biological significance of identified associations[1].

4. Dependence on training data quality: The performance of deep learning models can be sensitive to the quality and representativeness of the training data, as seen with methods like GWANN[4].

5. Potential for missing important associations: Some deep learning approaches that prioritize specific SNPs or functional units may overlook other biologically relevant associations[4].

Pros of classic GWAS:

1. Well-established methodology: Classic GWAS has a long history of use and validation in identifying genetic associations for complex traits[1][2].

2. Straightforward interpretation: Results from classic GWAS are generally easier to interpret and communicate to non-experts[1].

3. Unbiased genome-wide scanning: Classic GWAS examines associations across the entire genome without prior assumptions, potentially uncovering unexpected relationships[2].

4. Robust statistical framework: Classic GWAS relies on well-understood statistical methods for hypothesis testing and multiple comparison correction[3].

Cons of classic GWAS:

1. Limited ability to detect complex interactions: Traditional GWAS may miss non-linear relationships or epistatic interactions between genetic variants[1][4].

2. Multiple testing burden: The large number of statistical tests performed in classic GWAS requires stringent significance thresholds, potentially missing true associations with smaller effect sizes[3][4].

3. Difficulty in identifying causal variants: Classic GWAS often identifies tag SNPs in linkage disequilibrium with causal variants, requiring additional follow-up studies to pinpoint functional variants[1].

4. Limited incorporation of functional genomic information: Classic GWAS typically does not directly integrate additional biological data in the primary analysis[4].

In conclusion, while deep learning approaches to GWAS offer promising advantages in detecting complex genetic associations and integrating functional genomic data, they also come with increased computational complexity and potential interpretability challenges. Classic GWAS remains a valuable and well-established method for identifying genetic associations, but may miss some complex relationships that deep learning approaches can uncover. The choice between deep GWAS and classic GWAS depends on the specific research question, available data, and computational resources.

Citations:

[1] https://www.nature.com/articles/s43586-021-00056-9

[2] https://gwern.net/doc/genetics/heritable/2019-tam.pdf

[3] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2801749/

[4] https://www.mdpi.com/2073-4425/15/1/34

[5] https://www.youtube.com/watch?v=G2d5qaJ_g3w

E coli GWAS

 E. coli GWAS with additional data sets

output.gd seems to be files that HG used grep to extract SNP 

GWAS

 

https://twitter.com/Jazlyn_Mooney/status/1403009120152612868?s=20

trans effect on gene expression, heritability

 

https://www.cell.com/cell/fulltext/S0092-8674(19)30400-3?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS0092867419304003%3Fshowall%3Dtrue

GWAS of three molecular traits highlights core genes and pathways alongside a highly polygenic background

https://elifesciences.org/articles/58615

These GWAS show that GWAS hits are often found in pathway, though only count for ~10% heritability. 

This is related to Hong's network aging model on lifespan 

COVID19-hg GWAS meta-analyses round 2

COVID19-hg GWAS meta-analyses round 2

https://www.covid19hg.org/results/

statistical power, GWAS, quantitative genetics

Genet Epidemiol. 2015 Oct 10. doi: 10.1002/gepi.21935. Evaluating the Calibration and Power of Three Gene-Based Association Tests of Rare Variants for the X Chromosome. Ma C1, Boehnke M1, Lee S1; GoT2D Investigators.

Calibrating the Performance of SNP Arrays for WholeGenome Association Studies Ke Hao, Eric E. Schadt, John D. Storey, 2008, PLOS Genetics


Family-based association tests for sequence data, and comparisons with population-based association tests, Iuliana Ionita-Laza1, Seunggeun Lee, Vladimir Makarov, Joseph D. Buxbaum, and Xihong Lin

GWAS meta analysis

Gilman et al, Neuron, 2011. p898-907. NetBag on Autism
NetBag is a greedy approaches. The clustering methods started with one or two genes in CNV as ‘seeds’.

Gilman11 generated a weighted background human gene network for their study.

Gilman11 compared the cluster raw pvalue, called local pvalue to the p-values from random networks. The adjusted p-value is called global p-value.



-----------------------------------------------------------------------------------------------------------------------

AIS13 categorize pathway association methods into canonical and de nov pathway methods.

For de novo pathway discovery, integer linear program (ILP) is used in Leiserson , Blokh, Plos Comput Biol. Simultaneous identification of multiple driver pathway in cancer.

Steiner tree problem where one seeks the lowest cost pathway that connect the associated genes. See Liu et al, BMC Sys Biol 2012, Gene, pathway and nework frameworks to identify epistatic interactions of single nucleotide polymorphisms derived from GWAS data.

-----------------------------------------------------------------------------------------------------------------------
LERR13 review the method on protein-protein and protein-DNA networks to identify 'causal' genetic variant. (Their 'causal' definition is a narrowly defined one).

LERR13 argues that GWAS mostly find SNP that are LD with the actual 'causal' gene. This problem is of less concern to CNV analysis. One solution is to use network to 'rank' genes in the same haplotype known to be associate to the phenotype of interest or similar phenotypes. (This method is in the spirit of our recent CNV paper).

The green square represents the 'known' 'causal' gene. So, this is largely a traversal-measures based method.

LERR13 argues that networks contribute to 'missing heritability'.

LERR13 seems to suggest that protein-DNA networks are better suited for expression QTL (eQTL).

LERR13 shows that OMIM is the source of 'causal' gene information for most network based GWAS (table 1). Only one paper use GeneCards as an alternative source.

LERR13 cited several pathway enrichment analysis of GWAS. It argues that interaction are treated equally in these enrichment analysis. (This can be cited in our CNV replies). The authors then show several method use weighted networks to identify network modules using iterative 'seed and extend' method. (For comparison, our CNV paper did not use seed explicitly, and avoid some 'prejudice').

LERR13 also discussed subnetwork modules with mutation hotspots in cancer genomes.

----------------------------------------------------------------------------------------------------------

BGTF12: GWAS use meta-analysis of multiple data sets to reduce false positives and increase statistical power.

A major concern of GWAS meta analysis is the heterogeneity in the data sets, such LD difference among data sets, chip differences.  However, Lin and Zeng 2010 (Gene epidemil) show heterogeneity is not a significant factor using simulation studies.

Combination across data sets is the frequentist approach, cumulative studies is the Bayesian approach.

In R, GWAS meta-analysis package: Metrafor, rmeta, and CATMAP.

BGTF12 argues that GWAS data should be 'cleaned' and imputed before meta-analysis.

Reference:
[AIS13] Atias, Istrail, Sharan 2013, Current Opinion in Genetics and Development. Pathway-based analysis of genomic variation data.

[LERR13] Leiserson, Eldrige, Ramachandran, Raphael, 2013, Current Opinion in Genetics and Development. Network analysis of GWAS data.

[BGTF12], begum, ghosh, tseng, feigold, 2012 NAS, comprehensive literature review and statistical consideration for GWAS meta analysis

meta analysis of GWAS

NGR review by  Evangelos Evangelou and John P. A. Ioannidis, 2013

EI13 wrote that FDR has been rarely used in GWAS, though popular in omics. 

When phenotype is hard to define, GWAS often has problems. 

DNA content, rDNA cluster size detection by flow cytometer

Marcus Nordberg used flow cytometer to detect genome size variation. He thought genome variation is primarily due to rDNA cluster size changes.  In yeast, rDNA cluster size -> translational speed and accuracy OR extra-chromosomal rDNA circles OR genomic instability.  rDNA and ribosomal function is related to lifespan. So, flow cytometer can be used to quickly phenotype rDNA sizes.

Alternatively, qPCR, short-reads mapping, and southern hybridization can be used. In this case, standardized rDNA clusters will be used to generate a standard curve. The rDNA cluster reference strains are available in  David Bedwell's lab at UAB.