The recent publication and release to public databases of Dr James Watson's sequenced genome,1 with the exception of all gene information about apolipoprotein E (ApoE), provides a pertinent example of the challenges concerning privacy and the complexities of informed consent in the era of personalized genomics.2 Dr Watson requested that his ApoE gene (APOE) information be redacted, citing concerns about the association that has been shown with late onset Alzheimer's disease (LOAD), which is currently incurable and claimed one of his grandmothers.3

In this letter, without any ‘analysis’ of Dr Watson's genome, and thus respecting Dr Watson's wishes for APOE risk status anonymity, we highlight the challenges concerning the privacy and the complexities of informed consent by pointing out that the deletion of the APOE gene information only may not prevent accurate prediction of Dr Watson's risk for LOAD conveyed by APOE risk alleles. Specifically, linkage disequilibrium (LD) between one or multiple polymorphisms and APOE can be used to predict APOE status using advanced computational tools. Therefore, simply blanking out genotypes at known risk factors is generally not sufficient if the aim is to hide genetic information at these loci.

The major APOE risk for LOAD is generally assumed to come from the ɛ2/ɛ3/ɛ4 haplotype system, with the ɛ4 allele increasing risk for the disorder and the ɛ2 allele being protective.4 The ɛ2/ɛ3/ɛ4 haplotype system is defined by two nonsynonymous single nucleotide polymorphisms (SNPs) in APOE exon 4. One is a C/T SNP (rs429358) that encodes either arginine (C) or cysteine (T) in the ApoE at amino acid 112. The second site defining this haplotype system is a C/T SNP (rs7412), which again encodes arginine (C) or cysteine (T) at ApoE amino acid 158. The allelic compositions of the commonly investigated rs429358-rs7412 haplotypes are T-T for ɛ2, T-C for ɛ3, and C-C for ɛ4. The effects of these coding variants on ApoE function are well defined.5 A recent meta-analysis of LOAD risk in Caucasians (clinic/autopsy cohorts) indicated odds ratios (OR) of 15.6 (95% CI, 10.9–22.5) and 4.3 (95% CI, 3.3–5.5) for APOE ɛ4 homozygotes and ɛ4/ɛ3 heterozygotes respectively, compared to ɛ3 homozygotes.6 The meta-analytic odds ratios in population-based Caucasian samples were 11.8 (95% CI, 7.0–19.8) and 2.8 (95% CI, 2.3–3.5), respectively.6 In a large Rotterdam (Netherlands), population-based prospective study of people aged 55 years or above, it was estimated that 17% of the overall risk of AD could be attributed to the ɛ4 allele, with 3% (95% CI, 0–6%) of cases attributed to the ɛ4/ɛ4 genotype, and 14% (95% CI, 7–21%) to the ɛ4/ɛ3 genotype.7

A recent investigation of LD for 50 SNPs in and surrounding APOE in 550 Caucasians identified multiple SNPs in the TOMM40 gene 15 kb upstream of APOE, and at least one SNP in the other surrounding genes LU, PVRL2, APOC1, APOC4 and CLPTM1 were associated with LOAD risk.8 In particular, the C allele of SNP rs157581 in TOMM40 is in strong LD (r2>0.6) with the C allele of rs429358 in APOE, which defines the ɛ4 allele. For an additive (allelic) logit model, the OR for the presence of ɛ4 versus the status of LOAD was estimated to be 4.1, whereas the OR for LOAD status using the alleles of rs157581 was 2.9.8 Furthermore, using data sets such as those of Yu et al8 and SNPs identified in the surrounding regions of APOE in Dr Watson's sequence, haplotype phasing software could be utilized to easily and accurately predict Dr Watson's APOE risk haplotype status.

In addition, even if genotypes for non-APOE SNPs conveying LOAD risk are not listed in Dr Watson's sequence (ie, because of low sequence coverage), as in the case of TOMM40 SNP rs157581, it would be straightforward to predict Dr Watson's APOE risk status by exclusively using publicly available data, such as HapMap data. Specifically, although the LOAD high-risk APOE SNPs rs429358 and rs7412 and TOMM40 SNP rs157581 are not in the HapMap, a recent genome-wide association screen using 502 627 SNPs performed in 1086 histopathologically verified LOAD cases (n=664) and controls (n=442), identified HapMap SNP rs4420638, located in the APOC1 gene 14 kb downstream of the APOE ɛ4 allele, which has a powerful association with LOAD.9 Indeed, the association between LOAD and the G allele of rs4420638 (P=1 × 10−39) is similar to the association with the APOE ɛ4 allele (rs429358 C allele) itself (P=1 × 10−44), with additive allelic ORs of approximately 4 and 5, respectively.9, 10 Coon et al9 report strong LD between rs4420638 and rs429358 at D′=0.86, which implies an r2 of approximately 0.60 based on Caucasian allele frequency estimates for these SNPs listed in dbSNP.

We note that Dr Watson received genetic counseling and after being made aware of the privacy risks associated with public data broadcast, Dr Watson decided to share his personal genome by releasing it into a publicly accessible scientific database (for full details concerning Dr Watson and Protection of human subjects, Returning research results to research participants, and Data release and data flow, see Box 1 of Wheeler et al1). Nevertheless, during the preparation of this Letter, we contacted Dr Watson and colleagues in December 2007 and February 2008 informing them of the possibility of inferring his risk for LOAD conveyed by APOE risk alleles using surrounding SNP data. As a consequence, the online James Watson Genome Browser (JWGB) has nominally removed all data from the 2-Mb region surrounding APOE.

To demonstrate our point that genetic information is hard to hide, without contravening Dr Watson's wishes for APOE risk status anonymity (see Box 1 of Wheeler et al1), we utilized SNP genotypes identified in Dr J Craig Venter's genome sequence.11 Furthermore, Dr Venter's sequence data reports that he is heterozygote for both the LOAD high-risk APOE SNP rs429358 (T/C) and APOC1 SNP rs4420638 (A/G). Briefly, genotype imputation was performed using the MACH (version 1.0.16) computer program,12 HapMap (CEU)-phased haplotype data (encompassing 144 SNPs) and Dr Venter's genotypes listed for the 200-kb region surrounding rs4420638 (encompassing all 144 HapMap SNPs). Following the two-step approach outlined in the MACH online tutorial and after excluding Dr Venter's genotype data for rs4420638 and all APOE SNPs, we were able to correctly impute Dr Venter's rs4420638 genotype as A/G. The posterior probabilities for Dr Venter's rs4420638 genotype being A/A, A/G or G/G were estimated to be 0.008, 0.992 and 0.000, respectively. The high accuracy of Dr Venter's imputed rs4420638 genotype exemplifies the utility of imputing APOE genetic risk for LOAD.

Finally, although the deletion of 2 Mb is likely excessive for the surrounding APOE region (based on reported LD), as more detailed characterization of the human genome comes to light, it will become even more necessary to redact substantial regions surrounding identified genetic risk variants to avoid the indirect, though accurate, estimation of genetic risk such as those we detail above. For example, in a recent study, using gene expression profiling of Epstein–Barr virus-transformed lymphoblastoid cell lines of all 270 individuals genotyped in the HapMap Consortium, Stranger et al13 reported many instances of the most significant SNP associated with gene expression being located often 100 s of kb and up to 1 Mb outside of the gene transcript, with additional, less significant SNPs, although still useful in estimating risk, being located even further from the gene. Moreover, the potential for indirect estimation of risk will further increase as additional and more detailed genome-wide association studies are performed (which identify new risk loci) and individual human genomes are sequenced.

In summary, hiding genetic information in an otherwise fully disclosed genome sequence is not straightforward because of the availability of genomic data in the public domain that can be used to predict the missing data. We believe the potential for such indirect estimation of genetic risk has considerable relevance to concerns about privacy, confidentiality, discriminatory and defamatory use of genetic data, and the complexities of informed consent for both research participants and their close genetic relatives in the era of personalized genomics.