An Integrated Approach for the Assessment of Chromosomal Abnormalities

Transcription

1 An Integrated Approach for the Assessment of Chromosomal Abnormalities Department of Biostatistics Johns Hopkins Bloomberg School of Public Health June 6, 2007

2 Karyotypes

3 Mitosis and Meiosis

4 Meiosis

5 Meiosis Errors

6 Trisomy

7 Trisomy

8 Disomies

9 DNA changes

10 DNA changes

11 The data

12 Deletion

13 FISH

14 Amplification

15 Uniparental Isodisomy

16 Cancer samples

17 Mosaicism

18 SNPchip S4 classes and methods

19 Estimation 1 By SNP: Estimate genotype and copy number for each SNP. 2 Within a sample: Borrow strength between SNPs to infer regions of LOH and copy number changes. 3 Between samples: Comparison between normal and disease populations to find chromosomal alterations associated with disease.

20 More information The confidence in genotype calls can differ substantially between SNPs! 4 2 sense antisense

21 The structure of the data we observe At each SNP, we observe a noisy measure of the true copy number and genotype (and possibly also measures of confidence in those estimates).

22 Another HMM...? Novel (and we believe, important) HMM features: 1 Model the observation sequence of genotype calls and copy number jointly (Vanilla) 2 Integrate confidence estimates of the genotype calls and copy number estimates (ICE)

23 The Vanilla HMM components Observations ĈN and ĜT Hidden states Initial state probability distribution Transition probabilities Emission probabilities

24 Hidden states

25 Transition probabilities Following suggestions in the literature, we model the transition probabilities as a function of the distance d between SNPs. Specifically, let θ(d) 1 e 2d denote the probability that SNP i is not informative (I c ) for SNP at i + 1. For example: τ =P { (d) = P { = P { P i+1 i, d } i+1, I i, d } + P i+1, I c i, d i+1 I, i, d } P { I i, d } + i+1 I c, i, d P I c i, d = P { } θ(d).

26 Emission probabilities We assume conditional independence between copy number estimates and the genotype calls. For example: f( c CN, c GT ) = f( c CN ) f( c GT ) n o n o = f ccn f cgt = β n ccn o β n cgt o.

27 Integrating confidence estimates for genotype calls Let S cgt be the confidence score for the genotype estimate. We can estimate from Hapmap the following densities: n o n o n f SĤOM ĤOM, HOM, f SĤOM ĤOM, HET, f S HET d HET, d o n HOM, f S HET d HET, d o HET. Note: n o f SĤOM ĤOM, n f S HET d HET, d o n o f SĤOM ĤOM, HOM n f S HET d HET, d o HOM.

28 Emission Probabilities - Loss Recall that f( c CN, c GT ) = f( c CN ) f( c GT ) n o n o = f ccn f cgt = β n ccn o β n cgt o. If the state for a particular SNP is Loss, we have β n cgt, S cgt o = f n o n cgt f S cgt GT, c o.

29 Emission Probabilities - Retention For retention, the true genotype can be HET or HOM: β n cgt, Sd GT o n o n = f cgt f S GT d GT, c o n o n = f cgt f S GT d, HOM GT, c o n + f S d GT, HET GT, c o n o n = f cgt f S GT d HOM, GT, c o n f HOM GT, c o n + f S GT d HET, GT, c o n f HET GT, c o n o n = f cgt f S GT d HOM, GT c o n f HOM GT, c o n + f S GT d HET, GT c o n f HET GT, c o

30 Vanilla ICE comparison 5.0 A D B C E Amp LOH vanilla Norm Del Amp LOH ICE Norm Del Bioconductor package: ICE Amp LOH A D vanilla B vanilla E Norm Del Amp LOH A vanilla D vanilla ICE B ICE E Norm Del ICE ICE

31 A HapMap sample LOH 2 normal AA/BB AB Mb

32 Many HapMap samples

33 SNP Trio

34 SNP Trio

35 SNP Trio

36 SNP Trio

37 Probability Given the number of SNPs for a particular event region type (M), among all uploaded informative autosomal SNPs (X) of that trio, what is the probability to have the observed number (N) or more informative non-bpi SNPs of that type clustered together solely by chance? In other words, if the M SNPs of a particular event region type were randomly dispersed among the X informative autosomal SNPs, how probable is it to observe an event of the same or larger magnitude (defined by the number of consecutive SNPs of that event region)?

38 Probability Let p be the probability of a positive outcome in a Bernoulli trial (e.g. a 1 for 0/1 outcomes). Assume that we have Z trials. Let LS be the length of the largest block of consecutive ones in those Z trials. Let P Z (LS < N) be the probability that LS is smaller than some number N.

39 Probability If Z < N, then P Z (LS < N) = 1. If Z = N, then P Z (LS < N) = 1 p N. If Z > N, then P Z (LS < N) = N 1 k=0 p k (1 p)p Z k 1 (LS < N)

40 Probability Tabulate the probabilities P 1 (LS < N),..., P N (LS < N). Calculate P N+1 (LS < N), P N+2 (LS < N),..., P X (LS < N), iteratively. Using p = M/X, the probability of having N or more informative SNPs of a particular non-bpi type clustered together solely by chance is equal to 1 P X (LS < N).

41 HMM for SNP Trio chromosome 10 chromosome 22 BPI BPI UPI F UPI F UPI M UPI M MI D MI D MI S non BPI MI S non BPI BPI BPI position (Mb) position (Mb)

42 Acknowledgments Rob Scharpf Giovanni Parmigiani Rafael Irizarry & Gang Benilton Carvalho, Wenyi Wang Jonathan Pevsner & Lab Nate Miller, Eli Roberson, Jason Ting

43 References Carvalho B, Bengtsson H, Speed TP, Irizarry RA (2007) Exploration, normalization, and genotype calls of high-density oligonucleotide SNP array data. Biostatistics, 8(2): Scharpf RB, Ting JC, Pevsner J, Ruczinski I (2007). SNPchip: R classes and methods for SNP array data. Bioinformatics, 23(5): Scharpf RB, Parmigiani G, Ruczinski I (2007). A hidden markov model for joint estimation of genotype and copy number in high-throughput SNP chips. JHU Biostatistics Working papers, #136. Ting JC, Ye Y, Thomas GH, Ruczinski I, Pevsner J (2006). Analysis and visualization of chromosomal abnormalities in SNP data with SNPscan. BMC Bioinformatics, 7(1):25. Ting JC, Roberson ED, Miller N, et al, Ruczinski I, Thomas GH, Pevsner J (2007). Visualization of uniparental inheritance, Mendelian inconsistencies, deletions and parent of origin effects in single nucleotide polymorphism trio data with SNPtrio. Human Mutation, (in press). Wang W, Caravalho B, Miller N, Pevsner J, Chakravarti A, Irizarry RA (2006) Estimating genome-wide copy number using allele specific mixture models. JHU Biostatistics Working papers, #122.