Mendelian Randomization

A deconfounding method

Dr. Jørn OlsenEpi 200B

February 23 & 25, 2010

Stanner SA, Hughes J, Kelly CNM, Buttriss J. A review of the epidemiological evidence for the ‘antioxidant hypothesis’. Public Health Nutrition: 2003;7(3):407-422.

Violent death/cancer

Se-cholesterol

Cause

Confounding

Reverse causation?

Katan. Lancet. 1986

Given that ‘the gradient in serum cholesterol levels in the population is associated with a gradient in APOE [genotype]’, under the causal hypothesis we would expect to see a corresponding association between APOE and cancer. The absence of such a genetic association ‘would suggest that the association between low cholesterol and cancer is spurious’.

This argument assumes that the APOE gene is randomly allocated according to the laws of Mendel

Mendelian Laws, published 1866

The law of segregation: Hereditary traits are determined by pairs of factors, which segregate (separate) during gamete formation and reunite in the zygote.

B = yellow b = green BB/bb

P-generation B/B b/b homozygotes

Gametes B b

F1 generation B/b Random process

B/b B/b Heterozygotes

Gametes B b B b

F2 generation B/B, B/b, b/B b/b Random process

Yellow Green

3 1

Independent assortment: The alleles of one pair of genes segregate independently of other pairs of genes during gametogenesis

P generation B/B A/AYellow, roundB = yellow b = green

b/b a/aGreen, wrinkledA = round a = wrinkled

Gametes B A b a

F1 B/b A/aYellow/round

B/b A/aYellow/round

Gametes BA Ba bA ba BA Ba bA ba

F2 9/16 yellow/round (¾ x ¾)3/16 yellow/wrinkled (¾ x ¼)3/16 green/round (¼ x ¾)1/16 green/wrinkled (¼ x ¼)

Gametes BA Ba bA ba

BA B/B A/A B/B A/a B/b A/A B/b A/a

Ba B/B A/a B/B a/a B/b A/a B/b a/a

bA B/b A/A B/b A/a b/b A/A b/b A/a

ba B/b A/a B/b a/a b/b A/a b/b a/a

9 : 3 : 3 :1

Mendel’s Experiment - visit to home page

APOE + High Se-cholAPOE - Low Se-chol

If the APOE gene is randomized then we will not expect an association between the genotype and cancer under Ho

G = geneIP = intermediate phenotypeD = disease

G IP D

Genotype Low cholesterol Cancer

If IP D is confounded or due to reverse causation D IP, then there should be no G D association.

Low cholesterol cancer

confounders

Cancer low cholesterol

IP

G

DG --- Dnot present inpopulation at risk

Observations Folate NTD

Confounding? Folate NTD

C C

Randomized trial

MRC-trial + trial in Hungary

Folate Vitamins High risk women NTD

0 0 N00 D00

0 + N0+ D0+

+ 0 N+0 D+0

+ + N++ D++

PR

Could we have avoided the NTDs caused by low folate intake?

( ) / ( )( ) / ( )

.D D N ND D N N

0 +0

00 0 00 0

0 3

Mechanism Folate Homocystein NTD

Genotype MTHFR (TT) Homocystein NTD ?

If TT, Tt, tt are associated with different levels of homocystein, one would expect these genotypes to carry different risks for NTD

MTHFR (TT) NTD

Homocystein

Since these genotypes are randomized at birth, other correlates of homocystein are not expected to be associated with the genotypes

Important use of MR could be to eliminate confounding as a cause of associations between E and D in situations where randomization is not an option.

Mendelian deconfounding.

Alcohol and CVD

Most studies show the following:

CVD

0 6alcohol drinks/day

Observation Alcohol CVD

Confounding Alcohol CVD

DietSmokingExerciseEtc.

All populations have slow and fast metabolizers of alcohol guided by genetic factors

ALCOHOL ACETALDEHYDE ACETIC ACID

Slow metabolizers at step 1 = High level of alcoholSlow metabolizers at step 2 = High level of acetaldehyde

1 2

Genotype Slow metabolisers CVD

But Slow metabolisers AlcoholCVD ?

Genotype influence phenotype

Fixed alcohol intake: slow metabolisers CVD

Alcohol metabolism

ADH1 produces two different enzymes: 1 – fast, 2 – slow.

If alcohol protects against MI 2 should have low risk given the same intake.

Relative risks of myocardial infarction according to the genotypes

Variable ADH3 Genotype P value

11 1 2 2 2

No. Of subjects (%)

Patients 161 (51) 184 (46) 51 (13)

Controls 279 (36) 361 (47) 130 (17)

Relative risk (95% CI)

Matched 1.0 0.90 (0.69-1.17) 0.72 (0.50-1.05) 0.09

Multivariate 1.0 0.81 (0.61-1.09) 0.64 (0.43-0.98) 0.03

Multivariate, with adjustment for

1.0 0.83 (0.62-1.11) 0.65 (0.43-0.99) 0.04

alcohol consumption

Multivariate relative risk of myocardial infarction according to the genotypes and the level of daily alcohol consumption

0

0.2

0.4

0.6

0.8

1

1.2

Relative risk of myocardial infarction

< 1 Drink/day > 1 Drink/day

y1

y2

y3

P = 0.04

P = 0.04

P = 0.001

ALDH2Alcohol Acetaldehyde Acetic acid

ALDH2•1•1 fast

ALDH2•2•2 slow

Alcohol esophageal cancer

Alcohol (not carcinogenic in animal models)Confounding, modification of carcinogens – tissue damageMetabolite – acetaldehyde?

Elimination of acetaldehyde – ALDH2 enzymeA point mutation in ALDH2 ALDH2 x 2 x 2Allele and inability to metabolize acetaldehyde

Alcohol ALDH2 x 2 x 2: 18 times higher peakALDH2 x 1 x 2: 5 times higher peak

thanALDH2 x 1 x 1

Lewis, Sarah J. and Smith, George Davey. Alcohol, ALDH2, and Esophageal Cancer: A Meta-analysis which Illustrates the Potentials and Limitations of a Mendelian Randomization Approach. Cancer Epidemiol Biomarkers Prev. 2005;14(8):1967-1971

Lewis, Sarah J. and Smith, George Davey. Alcohol, ALDH2, and Esophageal Cancer: A Meta-analysis which Illustrates the Potentials and Limitations of a Mendelian Randomization Approach. Cancer Epidemiol Biomarkers Prev. 2005;14(8):1967-1971

Lewis, Sarah J. and Smith, George Davey. Alcohol, ALDH2, and Esophageal Cancer: A Meta-analysis which Illustrates the Potentials and Limitations of a Mendelian Randomization Approach. Cancer Epidemiol Biomarkers Prev. 2005;14(8):1967-1971

Plasma fibrinogen correlates with MI

A common mutation in G A the -fibrinogen gene is associated with high fibrinogen levels.

The G/A genotype correlated with plasma fibrinogen in men and in pre-menopausal women or post-menopausal women treated with HRT

ANOVA ANOVAp < 0.001 p < 0.001

2,4

2,6

2,8

3

3,2

3,4

3,6

3,8

G/G

G/A

A/A

****

****

*

Pla

sma

fi brin

og

en

(g

/ L)

Women Menn = 4,889 n = 3,972

GENOTYPE

If fibrinogen is a cause of MI, the G/A or A/A genotype should have high risk of MI.

Without ischemic heart disease

With ischemic heart disease

No. of women 3937 174

Genotype

G/G 0.63 (2467) 0.58 (101)

G/A 0.34 (1323) 0.40 (70)

A/A 0.04 (147) 0.02 (3)

A-allele 0.21 0.22

No. of men 2915 315

Genotype

G/G 0.64 (1877) 0.65 (204)

G/A 0.31 (913) 0.31 (98)

A/A 0.04 (124) 0.04 (13)

A-allele 0.20

C-reactive protein (CRP) is a maker of systemic inflammation.

CRP High Blood Pressure (HBP) - many studies show association

but could be

HBP CRP – reverse causation

Or

CRP HBP

C1

C2

Most studies haveMeasurement errors of confounders (obesity, smoking, social factors)Most studies rely on cross-sectional data; inversecausation

GG, GC, CC; G associated with high levels of CRP

Smith GC, et al. Association of C-Reactive Protein with Blood Pressure and Hypertension: Life Course Confounding and Mendelain Randomization Tests of Causality. Arterioscler Thromb Vasc Biol. 2005;25:1051-1056.

Smith GC, et al. Association of C-Reactive Protein with Blood Pressure and Hypertension: Life Course Confounding and Mendelain Randomization Tests of Causality. Arterioscler Thromb Vasc Biol. 2005;25:1051-1056.

Some studies show that coffee correlates with stillbirth. If caffeine is the culprit we can examine this by studying slow and fast metabolism of caffeine. Slow metabolism should have the highest risk at fixed levels of coffee intake.

No studies are large enough to demonstrate this.

Alcohol Acetaldehyde Eliminated

Alcohol Dehydrogenates ALDH2

ALDH 2x2x2 Homozygotes 18 times higher peak acealdehyde levels ALDH 2x1x2 Heterozygoter 5 timesADH 2x1x1 Homozygotes

Acetaldehyde cause nausea, headache etc. that makes drinking alcohol unpleasant-lowers the exposure

Is alcohol in high levels causing hypertension? Alcohol BP

Diet Obesity Stress Physical activity

ALDH2 Alcohol BP

CI..CN

Chen et al. PLOS Medicine 2008 5 (3): e52Meteanalysis: search word in PubMed and ISI ALDH2, hypertension, blood pressure, cardiovascular disease, heart disease-studies published before 2007Reference listMeta-analysis guidelines

JAMA 2000;283:2008-12

Alcohol intake in matter for males:

ALDH 2x1x1 20-30g /dayALDH 2x1x2 10-15g /day ALDH 2x2x2 0-2g /day

Or for ALDH 2x2x2 versus ALDH 2x1x1 for hypertension in males

Difference in systolicand diastolic bloodpressure for ALHD2x2x2versus ALDHD 2x1x1

Hypothesis

A high level of intrauterine exposure totestosterone increases the risk of autism / ADHD(Simon –Cohen)

Twin girls with a twin brother have a higher prenatal exposure than twin girls with a twin sister.We expect autism/ADHD to have genetic as well as environmental causesHow can we study it the intrauterine level of androgens play a role

Take Dz twins, compare females in FM sets with females in FF sets. FM females should have higher risksHow do we use mendelian randomization in this study?

Nitsch D., et al. Limits to Causal Inference based on Mendelian Randomization: A Comparison with Randomized Controlled Trails. Am J Epidemiol. 2006;163:397-403.

Comparison of RCT (intention to treat)

And MR

Nitsch D., et al. Limits to Causal Inference based on Mendelian Randomization: A Comparison with Randomized Controlled Trails. Am J Epidemiol. 2006;163:397-403.

Problems

Population stratification/genetic confounding: Different ethnic groups may have different genotypes and different disease risks.

Linkage disequilibrium – association between genes because they are located close together on the chromosome.

Gene 1 and Gene 2 associated-are in linkage disequilibrium:

Gen 2 Conf

Gene1 Exp Disease

Is Gene 1 a useful instrumental variable for Exp?

Gene 2 Cont

Gene 1 Exp Disease

Is Gene1 still a useful instrumental variable forExp?

ConfMaternal Maternal obesity offspring obesityFTO

Offspring FTO

Can maternal FTO be used as an instrumental variable for Maternal obesity to study offspring obesity?

Canalization: environmental adaptation to aspecific genotype like for ALDH2

Limitations

Requires well defined and rather strong genetic risk factors with high penetrance.

Should not interfere with behaviour of relevance to the exposure under study – unless this exposure could be controlled in the design or analysis.

Slow metabolisers of alcohol tend to drink less.

Genes involved in detoxification only play a role when the toxic exposure is present (gene-environment interaction). Without the exposure the gene has no function.

Mendelian Randomization

First described as such by Katan (Lancet 1986; i: 507-81).

Is now being explored in situations where we have a genotype that is causally linked to a phenotype.

Smith GD, Ebrahim S. Mendelian randomization: prospects, potentials, and limitations. Int J Epidemiol 2004; 33: 30-42.

Katan MB. Commentary: Mendelian randomization, 18 years on. Int J Epidemiol 2004; 33: 10-11.

Chen et al. PLOS medicine 2008; 5: 461-470

OffspringGenotypea´a´

OffspringGenotype

aa

OffspringGenotypeaa´

OffspringGenotypea´a

MotherGenotypeaa´

FatherGenotypeaa´

Allele that is inheritedfrom each parent israndomly determined

Mendelian randomization in case-parent design

The case-parent triad design

aa’ aa Parents

aa’ Cases

a’ more frequent in cases than expected from Mendel laws (probability in this case 0.50)

Will a gene with a specific polymorphism (GX) cause disease?

If 4 parents look like this

GX/G GX/G : ¾ of offspring expected to have the SNP

GX/G G/G : ½ of offspring expected to have the SNP

GX/ GX G/G : All will have the SNP

G/G G/G : None will have the SNP

2.25 of their 4 affected children are expected to have the GX if the SNP correlate with the disease.

This expectation be tested, but a larger sample is needed.