TL;DR
This paper reveals a significant privacy risk: a simple list of allele-specific gene expression (ASE) names from an individual can be used to recover their genome or infer their phenotype. This highlights a critical gap in current data anonymization practices for functional genomics data, even when explicit variant information is not shared.
What Problem Does This Paper Address?
The increasing availability of large-scale genomic and functional genomics datasets, such as RNA sequencing data, has fueled biomedical research. However, this surge in data also raises concerns about individual privacy. Previous work has shown that gene expression data can be linked to sensitive phenotypes. This paper specifically addresses the privacy implications of Allele-Specific Expression (ASE), a phenomenon where only one copy of a gene (maternal or paternal allele) is expressed. While ASE is biologically interesting and linked to disease predisposition, the authors demonstrate that even without explicit genetic variant information, a list of ASE genes can be exploited to de-anonymize individuals. This creates a "privacy conundrum" because researchers might assume that sharing ASE gene lists is safe, even if the data is derived from individuals who did not provide broad consent for their genomic data to be shared.
Methods
The authors propose two main attack scenarios:
- Recovering the genome of an individual: Given a list of ASE genes for a known individual, the goal is to link this list to a publicly available anonymized genome dataset to recover the individual's genome.
- Compiling Candidate SNPs: They first map gene names to exon locations using the GENCODE annotation. Then, for each gene, they identify Single Nucleotide Variants (SNVs) within its exons from a VCF file of the 1,000 Genomes Project. They filter these SNVs based on heterozygous genotyping frequency, creating a candidate SNP list for each ASE gene.
- Linking Attacks: They employ two statistical approaches to link the candidate SNP list to anonymized genomes:
- Best Match Approach: This method calculates a linking score based on the intersection of candidate heterozygous genotypes and the genotypes in the anonymized database. The genome with the highest score is considered the match.
- Entropy Approach: This method assigns a probability of correctly linking the ASE list to each genome, providing a distribution of probabilities.
- Inferring the phenotype of an individual: This scenario involves linking a known individual's genome to an anonymized database containing ASE gene expression and sensitive phenotypes. The adversary uses the known genome to mine ASE gene lists and then links these to phenotypes.
They also investigated the impact of auxiliary data, such as biological sex and ancestry, on linking accuracy.
Datasets
- 1,000 Genomes Project: Used for obtaining VCF files to identify candidate SNPs and for providing a database of anonymized genomes for linking attacks.
- AlleleDB: A database of allele-specific gene expression data from 382 individuals was used as the query set for linking attacks.
- GTEx (Genotype-Tissue Expression) and PsychENCODE projects: Mentioned as examples of publicly available anonymized databases that contain ASE and sensitive phenotypes, relevant to Scenario 2.
No new benchmarks were introduced; instead, existing public datasets were utilized to demonstrate the privacy risks.
Key Insights and Ideas
- ASE Gene Lists are Not Anonymized: The most striking insight is that a list of ASE gene names, devoid of explicit variant information, is sufficient to de-anonymize individuals. This challenges the assumption that such lists are inherently private.
- Recovering Genomes is Feasible: The study demonstrates a practical method for recovering an individual's genome by linking their ASE gene list to public genome databases.
- Phenotype Inference is Possible: The research also shows how ASE data can be used in conjunction with known genomes to infer sensitive phenotypes from anonymized datasets.
- HLA Genes are Informative but Not Solely Responsible: While Human Leukocyte Antigen (HLA) genes are highly polymorphic and appear frequently in ASE lists, their removal did not eliminate the linking ability, suggesting that other genes also contribute significantly.
- Auxiliary Data Enhances Linking: Incorporating simple demographic information like biological sex and ancestry significantly improves the accuracy of linking, underscoring the cumulative privacy risk when multiple data types are combined.
Key Results
- Using only ASE gene lists, 55% of individuals could be linked to their genomes in a database of 2,504 individuals. This increased to 80% when relaxing the matching criteria to the top 20 best matches.
- Removing highly polymorphic HLA genes from the ASE lists increased the linking percentage from 55% to 66%.
- Adding biological sex or ancestry information individually increased the linking percentage to 60%. Combining both biological sex and ancestry increased the linking percentage by approximately 10% with high statistical significance.
- The study successfully demonstrated the feasibility of these privacy breaches using data from individuals who provided broad consent.
Limitations and Future Work
- Limitations:
- The study focused on specific linking attack scenarios and datasets. The generalizability to all types of functional genomics data and all possible attack vectors may require further investigation.
- While the paper demonstrates feasibility, the exact rate of successful attacks in real-world, unconstrained scenarios might vary.
- Future Work:
- The authors suggest a three-pronged approach to mediate genomic privacy issues: (1) develop clear informed consent policies, (2) educate participants on risks and benefits, and (3) establish laws and legislation to prevent malicious use of genetic information.
- Further research could explore more sophisticated anonymization techniques for functional genomics data.
- Investigating the imputation of other genetic markers through linkage disequilibrium, as mentioned in the paper, could reveal even broader privacy implications.
Summary updated