Learning Your Identity and Disease from Research Papers: Information Leaks in Genome-Wide Association Study

Transcription

1 Learning Your Identity and Disease from Research Papers: Information Leaks in Genome-Wide Association Study Rui Wang, Yong Li, XiaoFeng Wang, Haixu Tang and Xiaoyong Zhou Indiana University at Bloomington

2 Genomic Revolutions Low-cost genotyping Revolutionary applications

3 Genome-Wide Association Study Case Group Control Group Single Nucleotide Polymorphism (SNP)

4 Identification Risk Consequence of identifications Participant protection De-identification Aggregation Is this sufficient?

5 Attack on Aggregated Data Single-allele frequencies Major: 0; Minor: 1 Homer s attack NIH s Reactions

6 The Rest of The Iceberg Other genome data Test statistics Linkage Disequilibrium (LD) Haplotype sequences Other sources Publications

7 Our Scary Findings ID from GWAS publications Test statistics LD statistics Allele Frequencies Statistical Identification Pair-wise allele frequencies SNP Sequences Work on real genome data Conclusion: Urgent needs to thoroughly study the problem

8 Why Doing This? Facilitate Dissemination of Genome Data SAFELY A Lesson From the Internet: Build Protection Into the Core!

9 Terms Alleles Single (0 1) Pair-wise (00, 01, 10, 11) Genotype Combinations of two sets of alleles Haplotype SNP Sequence (phased genotype) Locus Surrounding region of a SNP site

10 GWAS: Backgrounds GWAS Study Quality Control Info Leaks p values: leak joint allele frequencies of case & control Association Detection p values: leak case frequencies & control frequencies LD in Regions of Association r 2 : reveal LD of SNP sequences Replication Disclose cohorts with same frequency distributions

11 Homer s Attack Reference Group (Pop) Case Group (M) Pop j : 0.3 Pop j+1 : 0.6 Pop j+2 : 0.3 Y j : 1 Y j+1 : 0 Y j+2 : 1 Y i Pop i Y i M i M j : 0.8 M j+1 : 0.2 M j+2 : 0.6 H 0 Not in M D(Y i ) = Y i Pop i - Y i M i ΣD

12 What we can do Reverse engineer test statistics To find allele frequencies LD-based statistical identification Recover SNP sequences

13 Allele Frequency (Single) SNP 1 SNP2 SNP3 r 2 (1,3) C r 2 0* (1,2) r 2 (2,3) p 2 p 3

14 Allele Frequencies (Pair-wise) 2 2 (C00N *0C 2 (C N C C 0* ) 00 *0 0* ) L r< = < U C C C C C*1 C C C C 0* 1* *0 *1 = C = C = C = C 0* 0* 1* C + C 1* *0 + C + C *0 *1 (1) (2) (3) (4) (5) Catch: C 00 not unique Integer constraint Inaccurate r-squares Signs

15 Homer-Style Attack Based On LD? Why? Single AF: n LD: n(n-1)/2 But how? Validity of the test statistic D(Y i ) = Y i Pop i - Y i M i r 2 2 (C00C11 C10C01) = C 0* C 1* C *0 C *1

16 Our Statistical Attack We have to use signed r Distribution of T r? Markov model Reference? T T r = ( Y = 00 + Y T 1 i j N 11 ) ( r R + 1) / 2 ( Y 00 + Y 11 ) ( r C + 1) / 2 = ( r C r R )( Y 00 + Y 11 Y 01 Y 10 )

17 Recover SNP Sequences Contingency table problem Studied for decades Very difficult Divide-and-Conquer 1. Construct each haplotype block 2. Connect different blocks

18 Simple Defense Low-precision statistics Correlation among SNPs Thresholds How to determine them? Noises Consistency check Maximum-likelihood approximation

19 Evaluations Data: the HapMap project Locus: FGFR2 174 SNPs Used in a real GWAS study Population Africa backgrounds 200: half cases and half controls

20 Allele Frequencies and Signs

21 Statistical Powers 20 times more powerful than Homer s test (T p )

22 Recover Haplotypes Linear equation solving: rref Integer Programming: bintprog 100 individuals, 10 blocks, 174 SNPs System: 2.80GHz Core 2 Duo, 3GB memory Fully restored within 12 hours

23 Discussion Genotypes vs. Haplotypes Defense Differential privacy

24 Conclusion New attacks and new understanding Many open research problems

25 Contacts Dr. XiaoFeng Wang Web: du/xw7 System Security Lab: sysseclab.informatics.india na.edu Dr. Haixu Tang Web: du/hatang

26 References Good: from the same population Bad: from different populations good reference average reference

27 More In-depth Studies Larger populations: Low-precision statistics (200 cases, 200 references)