1. Some slides adapted from J. Fridlyand
Analysis of Array CGH Data
by Hanni Willenbrock
Some slides adapted from J. Fridlyand
BioSys course: DNA Microarray Analysis – Lecture, 2007

2. Outline
Introduction to comparative genomic hybridization (CGH) and array CGH
Data analysis approaches
Breakpoint detection
Loss and gain analysis
Application of segmentation to testing
Real data example 1: Application to a primary tumor dataset
Real data example 2: comparative genomic profiling of bacterial strains

3. Comparative Genomic Hybridization
Study types :
Gain or loss of genetic material
To find variations in the genetic material
Purposes:
Study of chromosomal aberrations often found in cancer and developmental abnormalities.
Study of variations in the baseline sequence in a microbial population (microbial comparative genomics).
As will be obvious in a few slides, the physical dependency in this type of data facilitates this (that only 1 sample is necessary to draw conclusions).
PhD defense, October 27th 2006

4. A Variety of Genetic Alterations Underlie Developmental Abnormalities and Disease
Inappropriate gene activation or inactivation can be caused by:
Mutation
Epigenetic gene silencing (e.g. addition of methyl groups)
Reciprocal translocation (exchange of fragments between two non-homologous chromosomes)
Gain or loss of genetic material
Any of the above may lead to an oncogene activation or
to inactivation of a tumor suppressor.

5. Existing techniques for detecting structural abnormalities
Albertson and Pinkel, Human Molecular Genetics, 2003

6. Some microarray platforms for copy number analysis
BAC arrays
Affymetrix SNP chip (500 K)
Representational oligonucleotide microarray analysis (ROMA) in
Whole genome tiling arrays
Own design (NimbleGen/NimbleExpress)

7. Array CGH: BAC arrays HumArray3.1
12 mm
HumArray3.1
2464 human BAC clones spotted in triplicates
kbp

8. Array CGH Maps DNA Copy Number Alterations to Positions in the Genome
Test Genomic DNA
Reference Genomic DNA
Cot-1 DNA
Gain of DNA copies in tumor
Loss of DNA copies in tumor
Ratio
Position on Sequence

9. Example: Detection of DiGeorge region
(A) Detection of deletion in the DiGeorge region by FISH.
A chromosome 22 subtelomere probe (green) and the TUPLE1 probe for the DiGeorge region (red) were hybridized to metaphase chromosomes from a normal individual and an individual with the deletion. The arrow indicates the missing red FISH signal on the deleted chromosome.
(B) Array CGH copy number profile of chromosome 22 showing deletion in the DiGeorge region (arrow).
Albertson and Pinkel, Human Molecular Genetics, 2003

10. Structural abnormalities*
*HSR: homogeneously staining region
Albertson and Pinkel, Human Molecular Genetics, 2003

11. Tumor Genomes are Stable Copy Number Profiles of a Tumor & Recurrence

12. Analysis of array CGH
Goal: To partition the clones into sets with the same copy number and to characterize the genomic segments in terms of copy number.
Biological model: genomic rearrangements lead to gains or losses of sizable contiguous parts of the genome, possibly spanning entire
chromosomes, or, alternatively, to focal high-level amplifications.

13. Varying genomic complexity
Breakpoints

14. Exercise Part I: Plot and view array CGH data
DNA Microarray Analysis Course, 2007

15. Observed clone value and spatial coherence? ?
Useful to make use of the physical dependence of the nearby clones,
which translates into copy number dependence.
DNA Microarray Analysis Course, 2006

16. Expected log2 ratio as a function of copy number change, normal cell contamination and ploidy
Reference ploidy=2
Reference ploidy=3
2.58
100%
2.0
50%
0.58
0.42
10%
0.0
0.58
0.38
0.07

17. Many algorithms to choose from
Simulation Study
Many algorithms to choose from
Mainly evaluated only on limited examples
Few comparisons between algorithm performance
Choice of evaluation criteria:
False breakpoint detection vs. missed breakpoints
Sample type preferences (size of segments, noise, etc)

18. Methods for Segmentation
HMM: Hidden Markov Model (aCGH package)
Fit HMMs in which any state is reachable from any other state (Fridlyand et al, JMVA, 2004).
CBS: Circular binary segmentation (DNAcopy package)
Tertiary splits of the chromosomes into contiguous regions of equal copy number and assesses significance of the proposed splits by using a permutation reference distribution (Olshen et al, Biostatistics, 2004).
GLAD: Gain and Loss Analysis of DNA (GLAD package)
Detects chromosomal breakpoints by estimating a piecewise constant function that is based on adaptive weights smoothing (Hupe et al, Bioinformatics, 2004).

19. Comparison Scheme Use of simulated data, where the truth is known
The noise is controlled (see later slide)
One segment
True breakpoint
false predicted breakpoint

20. Breakpoint Detection Accuracy

21. Exercise Part II: Segmentation and breakpoint prediction
DNA Microarray Analysis Course, 2007

22. DNA Microarray Analysis Course, 2006
Merging segments
Note: that all procedures operate on individual
chromosomes, therefore resulting in a large number of
segments with mean values close to each other.
Additional Challenge: reduce number of segments by
merging the ones that are likely to correspond to the same
copy number.
This will facilitate inference of altered regions.
DNA Microarray Analysis Course, 2006

23. DNA Microarray Analysis Course, 2006
Merging
For estimating actual copy number levels from segmentations
DNA Microarray Analysis Course, 2006

24. Segmentation and Merging
DNA Microarray Analysis Course, 2006

25. ROC Curve: Identification of copy number alterations for varying thresholds

26. Exercise Part III: Estimate copy number gain and losses
DNA Microarray Analysis Course, 2007

27. Using segmentation for testing (phenotype association studies)
Example case:
Find clones (or whole segments) that are significantly differing in copy number between two cancer subtypes.
Task:
Investigate whether incorporating spatial information (segmentation) into testing for differential copy number increases detection power.
Data type:
Samples with either of 2 different phenotypes (e.g. 2 different cancer subtypes)
How:
Comparison of sensitivity and specificity using:
Original test statistic (no use of spatial information)
Segmented T-statistic derived from original log2 ratios
T-statistic computed from segmented log2 ratios

28. Simulation of Array CGH Data
Real biological variation considered:
Breast cancer data used as model data
Segment length and copy number is taken from the empirical distribution observed in breast cancer data (DNAcopy segmentation).
Mixture of cells (sample is not pure)
Each sample was assigned a value, Pt: proportion of tumor cells, between 0.3 and 0.7 from a uniform distribution.
Experimental noise is Gaussian
Standard deviations drawn from a uniform distribution between 0.1 and 0.2 to imitate real data where the noise may vary between experiments.
Cancer subtypes are heterogeneous
Certain aberrations characteristic for a cancer subtype may only exist in a percentage of the patients with that cancer subtype. Thus, in each sample, segments with copy number alterations (copy number not 2) was removed at random with probability 30%.

29. Testing samples (original values)
37.5% x9
57.0%
x11
20 samples from either of 2 classes, red is true copy number, black dots are simulated values, circles around example of heterogeneity

30. Testing samples (original values)
Red:
True different clones

31. Testing: why is multiple testing necessary?
standard
p-value cutoff for alpha=0.05
=> Many false positives

32. Testing: why is multiple testing necessary?
Significance with random class assignments?
By chance, many test statistics are below/above standard significance thresholds
2.93
5.29
-3.99 (maximum deviating value)

33. The maxT Multiple Testing Correction
By repeating random class assigningment and testing, e.g. 100 times, the following ”permutation reference distribution” of maximum absolute test statistic is obtained (maxT distribution):
We wish to control the family wise error rate (FWER) at alpha=0.05 (5% chance of 1 false positive).
Therefore, the cut-off should be such that only in 5% of the random cases, we will get one false positive (95 percentile): cutoff = 5
standard significance
threshold
MaxT multiple testing corrected threshold

34. Testing samples (original values)
maxT p-value cutoff for alpha = 0.05
standard
p-value cutoff for alpha=0.05

35. Testing: Segmenting test statistics
Reference

36. Testing segmented samples
1. Segmentation of individual samples...

37. Testing segmented samples
Reference
2. T-statistic from segmented individual samples...

38. Detecting regions with differential copy number
Willenbrock and Fridlyand. Bioinformatics 2005; 21(22):

39. Variation of Simulation Parameters
Signal2noise
CBS consistently the best performance
HMM has the highest FDR
GLAD is least sensitive
Alternative empirical distributions of segment lengths
HMM has highest sensitivity for segment sizes below 10
CBS has highest sensitivity for segment sizes 10 or larger
GLAD consistently performes the worst
Outlier detection

40. Real Data Example 1: Primary Tumor Data
75 oral squamous cell carcinomas (SCCs)
TP53 mutational status of all samples was determined using sequence information (Snijders et al., 2005)
Tasks:
Characterize wild-type and mutant samples with respect to their genomic alterations
Build a classifier to predict TP53 mutational status

41. Frequency of Gain/Loss Comparisons
Threshold-based
Merge-based
5% altered
33% altered

42. Why such a difference in alteration frequency?
+ 2.5x MAD
- 2.5x MAD
Willenbrock and Fridlyand. Bioinformatics 2005; 21(22):
High threshold-based cut-off is due to the high experimental noise of the paraffin-embedded tumors

43. Classification results
Willenbrock and Fridlyand. Bioinformatics 2005; 21(22):

44. Real Data Example 2: Comparative genomic profiling of several Escherichia coli strains
The microarray design included probes for:
7 known E. coli strains
39 known E. coli bacteriophages
104 known E. coli virulence genes
Experimentally:
2 sequenced control strains (W3110 and EDL933), 3 replicates
2 non-sequenced strains (D1 and 3538), 3 replicates
Bacteriophage: 3538 (stx2::cat), 2 replicates

45. Comparative Genomic Profiling: challenges
Ratio problems: some genes might be present on query strain but not on the known reference strain.
Single channel microarrays or dual channel microarrays?
In this case, we used an Affymetrix single channel custom made array (NimbleExpress)
Partly present genes versus similar but different genes.

46. Homology between the 7 E. coli strains included on the microarray
Very high similarity between the two K-12 strains and between the two O157:H7 strains.
Percentage of homologues for E. coli genomes in columns found in E. coli genomes in rows.
Willenbrock et al. Journal of Bacteriology Nov;188(22):

47. BLAST Atlas
Willenbrock et al. Journal of Bacteriology Nov;188(22):

48. Hybridization Atlases
Willenbrock et al. Journal of Bacteriology Nov;188(22):
Probe hybridizations for experiments (samples) result in a similar pattern as expected from the BLAST atlas.

49. Mapping the phage Φ3538 (stx2::cat)
Willenbrock et al. Journal of Bacteriology Nov;188(22):

50. Zoom of phage Φ3538 (stx2::cat)
Willenbrock et al. Journal of Bacteriology Nov;188(22):
The hybridization pattern is very similar for the phage, strain 3538 and strain D1.

51. Hierarchical Cluster Analysis
K-12
D1 is very similar to the K-12 type strains (W MG1655).

52. E. coli virulence genes
Willenbrock et al. Journal of Bacteriology Nov;188(22):
D1 is probably still a commensal strain (An organism participating in a symbiotic relationship from which it benefits while the other is unaffected).

53. Comparative genomic profiling of two E. coli strains
Summary
Comparative genomic profiling of two E. coli strains
0175:H16 D1
0157:H7 3538
Identification of virulence genes and phage elements
Conclusions:
D1 is similar to the K-12 type strains
Characterization of D1 and 3538 genes:
Identification of a number of genes involved in DNA transfer and recombination

54. Advantages over Conventional Expression Arrays
Hybridization of DNA to microarray (DNA is much more stable)
Little normalization is necessary
Use of spatial coherence in the analysis
Only 1 sample is necessary to draw conclusions (it is still necessary with biological replicates to be able to draw general conclusions regarding a certain biological subtype)
Results may be easier interpretable and correlated with sample phenotypes (e.g. loss of oncogene repressor -> certain cancer subtype)

55. Summary
Numerous methods have been introduced for segmentation of DNA copy number data and breakpoint identification. It is important to benchmark them against existing methods (however, only feasible if the software is publicly available)
Currently, CBS (DNAcopy package) has the best overall performance
Use of spatial dependency in the analysis improves testing power on clone-by-clone basis
Merging of segmentation results improves copy number phenotype characterization
Study types:
Study of copy number in cancer samples
Study of samples from patients with mental diseases
Comparison of bacterial strains

56. Questions? Exercise Part IV + Bonus exercise: Real data analysis
DNA Microarray Analysis Course, 2007