TL;DR
This paper introduces PLIGHT, a computational tool that quantifies privacy risks from sparse, noisy genetic data by aligning query SNPs to reference haplotype databases. It demonstrates that even a small number of SNPs can reveal an individual's identity or relatedness, and offers a sanitization tool to mitigate these risks.
What Problem Does This Paper Address?
In the era of big data and widespread genetic information, re-identifying individuals and their relatives from genetic data is a growing concern. While it's known that a large number of Single Nucleotide Polymorphisms (SNPs) can identify individuals, the availability of small, noisy genotype datasets (e.g., from environmental DNA or functional genomics) raises questions about the privacy risks associated with even limited genetic information. This paper aims to quantify the informativeness of such sparse and noisy SNP sets and develop methods to mitigate the associated privacy risks.
Methods
The core of this work is the PLIGHT (Privacy Leakage by Inference across Genotypic HMM Trajectories) computational tool. PLIGHT employs a Hidden Markov Model (HMM) framework, specifically based on the Li-Stephens model, to model recombination and mutation processes. It aligns sparse, noisy query SNP sets to a reference haplotype database by finding "piecewise" alignments. This allows PLIGHT to identify segments of reference haplotypes that best match the query SNPs, forming a "genotypic trajectory" through the reference haplotype space.
Key methodological aspects include:
- Haplotype Inference: PLIGHT infers best-fit haplotype pairs for different genomic regions based on a reference database.
- Mosaic Genome Construction: It can construct "mosaic genomes" by combining segments from different source individuals.
- Privacy Attack Simulation: PLIGHT is used to simulate various privacy attacks, including membership inference and relational inference.
- Phenotype Inference: The tool can also be used to explore potential phenotypic information leakage through coarse-grained GWAS SNP imputation and Polygenic Risk Score (PRS) calculations.
- Sanitization Tool: A complementary tool, PLIGHT_SanitizeGenotypes, is provided to remove the most identifying SNPs from a dataset.
To handle large reference databases, PLIGHT offers three algorithms:
- PLIGHT_Exact: Performs exact HMM inference.
- PLIGHT_Truncated: Truncates possible trajectory extensions to reduce memory.
- PLIGHT_Iterative: Subdivides the reference space into smaller cohorts for more manageable processing, offering better scalability.
Datasets
The primary reference database used in this study is the 1000 Genomes Phase 3 database, containing 2,504 phased genotypes from 26 populations. For some analyses, a subset of the CDX subpopulation was also used.
For simulating noisy and sparse data, the authors utilized:
- Simulated genotype "mosaics": Composites of genotypes from two source individuals.
- Genotypes from coffee cup swabs: These represent real-world noisy and potentially contaminated samples.
The study also references the GWAS catalog for phenotype association studies.
Key Insights and Ideas
- Minimal SNPs, Maximum Risk: Even a small number of common, noise-free SNPs (as few as 10) can be sufficient to identify an individual within a large database. This challenges the assumption that only large SNP sets pose a significant re-identification risk.
- Beyond Direct Identification: Privacy leakage extends beyond direct identification. Partial matches of genomic segments can reveal relatedness and even allow for coarse-grained phenotypic inference.
- Noisy Data is Still Informative: Even noisy SNPs, such as those from environmental samples (coffee cups), can still carry enough information to identify individuals or infer phenotypic traits.
- HMMs as a Privacy Tool: The HMM framework, traditionally used for genetic analysis, is repurposed here as a powerful tool for quantifying and mitigating privacy risks.
- Sanitization is Crucial: The development of a sanitization tool highlights the practical need to selectively remove identifying SNPs to balance data utility with privacy protection.
Key Results
- Individual Identification: With ~5,000 haplotypes, 10 noise-free SNPs were sufficient for individual identification.
- Mosaic Identification: ~20 SNPs could identify components in two-individual simulated mosaics.
- Kinship Inference: 20-30 SNPs were sufficient to identify first-order relatives (parents, children, siblings).
- Noisy Data Identification: ~30 noisy SNPs from coffee cups identified an individual within a database.
- Phenotypic Leakage: Local genotype matches allowed for some phenotypic information leakage through GWAS SNP imputation and PRS calculations, even when the individual was not directly in the database.
- PLIGHT_Iterative Performance: This algorithm showed significantly better scaling properties for large reference databases compared to PLIGHT_Exact and PLIGHT_Truncated.
Limitations and Future Work
- Computational Cost: While PLIGHT_Iterative improves scalability, exact HMM inference can still be computationally intensive, especially for very large databases.
- Assumptions: The current model assumes a linear reference genome and uniform transition probabilities, which might not fully capture complex genomic structures or population stratification.
- Sanitization Strategy: The paper presents one strategy for sanitization (removing SNPs with fewest parallel best-fit trajectories) but acknowledges that other strategies might be more suitable depending on the use case.
- Future Work: The authors suggest exploring more sophisticated privacy risk assessment platforms, integrating PLIGHT with graph genomes for more comprehensive variation representation, and applying the underlying model to other problems like analyzing contaminated samples or viral mutations. They also highlight the need for careful utility-privacy trade-offs based on intended data usage.
Summary updated