Some views on microarray experimental design

Rainer BreitlingMolecular Plant Science Group & Bioinformatics Research Centre

University of Glasgow, Scotland, UK

Personal Background

• University of Glasgow, Scotland, UK

• Molecular Plant Sciences Group

• Bioinformatics Research Centre

• Functional Genomics Facility

Some common questions in microarray experimental design

• How many arrays will I need?

• Should I pool my samples?

• Which arrays should I choose?

• Which samples should I put together on one array?

Why are microarrays special?

• produce large amounts of data instantaneously

• can look for unexpected effects

• are still quite expensive

almost never repeated

careful design necessary before you start

How many replicates?

• as many as possible

Statistics says: The more replicates, the better your estimate of expression (that’s an asymptotic process, so if you add at least a few replicates, the effect will be really strong)

How many replicates?

2

212/1

)/(

)(4

zz

n

•α significance level (probability of detecting FP)

•1-β power to detect differences (probability of detecting TP)

•σ standard deviation of the log-ratios

•δ detectable difference between class mean log-ratios

•z percentile of standard normal distribution

n required number of arrays (reference design)

How many replicates?

• Five

Experience shows: For most common experiments you get a reasonable list of differentially expressed genes with 5 replicates

How many replicates?

• Three

One to convince yourself, one to convince your boss, one just in case...

How many replicates?

• It depends on– the quality of the sample– the magnitude of the expected effect– the experimental design– the method of analysis

The quality of the sample

• smaller samples (single cells) are more noisy than large samples (tissue homogenates)

• cell cultures are less noisy than patient biopsies

• sample pooling can decrease noise – if individual variation is not of interest

The magnitude of the effect

• Microarrays are very sensitive

• To keep effects small:– use early time points, gentle stimuli– never compare dogs and donuts

• if you get a list of 2000 genes that are significantly changed, your experiment failed!

The magnitude of the effect

• some problematic cases– stably transfected cell lines (are they still the

same cells?)– knock-out organisms (even the same tissue

can be a different)– local changes may be diluted cell

isolation will increase noise

The experimental design

• Three major options:– reference design (flexible)

– balanced block design (efficient)

– loop design (elegant)

The experimental design

• loop designs can save samples...

• ...but they can cause interpretation nightmares in less simple cases (use for large studies, if you have a full-time statistician in the team)

A B

CD

A B C D

R R R R

The method of analysis

• Golub et al. (1999) data set

• 38 leukemia patient bone marrow samples, hybridized individually to Affymetrix microarrays

• Differential expression between two leukemia types was examined, using random subsets of the complete dataset

The method of analysis  0h 9.5h 11.5h 13.5h 15.5h 18.5h 20.5h

   6144 - purine base metabolism

6099 - tricarboxylic acid cycle

6099 - tricarboxylic acid cycle

3773 - heat shock protein activity

6099 - tricarboxylic acid cycle

     9277 - cell wall (sensu Fungi)

3773 - heat shock protein activity

5749 - respiratory chain complex II (sensu Eukarya)

6099 - tricarboxylic acid cycle

3773 - heat shock protein activity

     297 - spermine transporter activity

6950 - response to stress

6121 - oxidative phosphorylation, succinate to ubiquinone

5977 - glycogen metabolism

5749 - respiratory chain complex II (sensu Eukarya)

     15846 - polyamine transport

297 - spermine transporter activity

8177 - succinate dehydrogenase (ubiquinone) activity

6950 - response to stress

6121 - oxidative phosphorylation, succinate to ubiquinone

       4373 - glycogen (starch) synthase activity

3773 - heat shock protein activity

4373 - glycogen (starch) synthase activity

8177 - succinate dehydrogenase (ubiquinone) activity

       15846 - polyamine transport

4373 - glycogen (starch) synthase activity

4129 - cytochrome c oxidase activity

6537 - glutamate biosynthesis

       5353 - fructose transporter activity

7039 - vacuolar protein catabolism

5751 - respiratory chain complex IV (sensu Eukarya)

6097 - glyoxylate cycle

       15578 - mannose transporter activity

6950 - response to stress

5749 - respiratory chain complex II (sensu Eukarya)

5750 - respiratory chain complex III (sensu Eukarya)

       7039 - vacuolar protein catabolism

4129 - cytochrome c oxidase activity

6121 - oxidative phosphorylation, succinate to ubiquinone

9060 - aerobic respiration

       8645 - hexose transport

5751 - respiratory chain complex IV (sensu Eukarya)

8177 - succinate dehydrogenase (ubiquinone) activity

4129 - cytochrome c oxidase activity

iterativeGroupAnalysis(iGA)

glyoxylate

cycle

citrate (TCA) cycle

oxidative phosphorylation

(complex V)

respiratory chaincomplex III

respiratory chaincomplex II

Graph-based iterative

GroupAnalysis (GiGA)

What is a good replicate?

The experiment your competitor at the other side of the globe would do to see if your results are reproducible

Vary “all” parameters – challenge your results

Prepare new samples, from new cultures, using new buffers and new graduate students

Remember to produce matched controls

What is a “bad” replicate?

• technical replicates (i.e. hybridizing the same sample repeatedly)

• dye-swapping experiments (usually gene-specific dye bias is not a big issue, and dye balancing is more efficient anyway)

• pooled samples, hybridized repeatedly

• the same preparation, only labelled twice

Should samples be pooled?

• most samples are already pooled – they come from multiple cells

• pool to increase amount of mRNA, but only as much as necessary

• prepare independent pools to assess variation

• problems: bias, “contamination”, outliers, information loss...

Which arrays are the best?

• Standard arrayscompare and exchange data easily

• Whole-genome arraysdetect unexpected effects, increase confidence

• Single-color arrays (Affymetrix GeneChip)for more complex comparisons

• Annotated arrays

Further reading

• Dobbin, Shih & Simon (2003) J. Natl. Cancer Inst. 95: 1362.

• Yang & Speed (2002) Nature Rev. Genet. 3: 579.

• Breitling (2004) http://www.brc.dcs.gla.ac.uk/~rb106x/microarray_tips.htm

Contact

Rainer Breitling

Bioinformatics Research Centre

Davidson Building A416

R.Breitling@bio.gla.ac.uk

http://www.brc.dcs.gla.ac.uk/~rb106x