Private information leakage from functional genomics data: Quantification with calibration experiments and reduction via data sanitization protocols

Gamze Gürsoy, Prashant S. Emani, Charlotte M. Brannon, Otto Jolanki, Arif Harmanci, J. Seth Strattan, Andrew D. Miranker, Mark GersteinbioRxiv 2018 · DOI

TL;DR

This paper introduces a novel data sanitization protocol for functional genomics data that allows for secure sharing of raw sequencing reads while minimizing privacy leakage. By developing methods to quantify and reduce the risk of re-identification through linkage attacks, the work enables broader access to valuable genomic datasets for research.

What Problem Does This Paper Address?

The explosion of functional genomics data, generated using powerful next-generation sequencing technologies, offers unprecedented insights into gene regulation and disease mechanisms. However, sharing this raw data, crucial for reproducibility and advanced analysis, poses significant privacy risks. Unlike traditional DNA sequencing, functional genomics data, while not directly intended to identify genetic variants, inherently contains them. This allows for "linkage attacks" where an attacker can combine this data with publicly available information to re-identify individuals and infer sensitive phenotypes. Current data sharing models often rely on controlled access, creating bottlenecks. This paper addresses the need for a method to share raw functional genomics data openly while safeguarding individual privacy.

Methods

The core of this work is a data sanitization protocol that transforms raw functional genomics reads into a "privacy-preserving BAM" (pBAM) format. This is achieved by generalizing or removing variant information from the BAM files. The sanitization process is agnostic to the type of variant (SNP, indel, structural variant) and can be tuned to balance privacy and utility.

To quantify privacy leakage, the authors employ linkage attack strategies, specifically adapting methods from Narayanan and Shmatikov's work on de-anonymizing large sparse datasets. They define three types of linkage attacks (P-P, P-N, N-N) based on whether the external information and the attacked dataset are perfect or noisy. A key metric, the "gap value," is introduced to assess the statistical robustness of a successful linkage.

The paper also details experimental protocols for DNA extraction, whole-genome amplification, various sequencing technologies (Illumina, Nanopore), and RNA-Seq, as well as genotyping arrays.

Datasets

The study utilizes several datasets for its experiments and evaluations:

  • gEUVADIS project: Provided RNA-Seq data for 421 individuals.
  • 1,000 Genomes Project: Used for high-coverage Whole Genome Sequencing (WGS) data, serving as the "perfect information" for linkage attacks.
  • ENCODE consortium: Provided cell line functional genomics data (e.g., ChIP-Seq, Hi-C) used as noisy information for linkage attacks.
  • Environmental samples: Used coffee cups were collected from consented individuals to mimic real-world scenarios of surreptitiously obtained DNA.

The paper does not introduce new benchmark datasets but rather uses existing ones to evaluate its proposed sanitization protocol and linkage attack quantification methods.

Key Insights and Ideas

  • Functional genomics data is inherently privacy-sensitive: Even though the primary goal isn't to identify genetic variants, the data contains them, making it vulnerable to re-identification.
  • Linkage attacks are a significant threat: Combining functional genomics data with external information can reveal sensitive phenotypes, not just genotypes.
  • A tunable sanitization protocol is crucial: A one-size-fits-all approach to privacy is insufficient; the protocol allows for flexible trade-offs between privacy and data utility.
  • Environmental DNA samples (like coffee cups) are a viable source for attackers: This highlights the real-world applicability of privacy risks.
  • "Gap value" is a robust metric for linkage attack reliability: It quantifies how well a matched individual is separated from others in the dataset.

Key Results

  • The proposed data sanitization protocol (pBAM) effectively minimizes privacy leakage while maintaining high data utility.
  • Linkage attacks were successfully performed, demonstrating that sensitive phenotypes can be inferred from functional genomics data. For example, linking coffee cup DNA to RNA-Seq data was achieved for as little as $19 in sequencing costs.
  • The study showed that even with noisy data and a significant number of false-positive genotypes (up to 100,000), linkage accuracy remained high.
  • Different functional genomics assays (e.g., ChIP-Seq, RNA-Seq) exhibit varying degrees of privacy leakage, with some being more susceptible to linkage attacks than others.
  • The pBAM format significantly reduces file sizes (e.g., up to 99.99% compression for BAM to .diff files), aiding in data sharing and storage.

Limitations and Future Work

  • Computational cost of sanitization: While the protocol is effective, the computational time for converting BAM to pBAM can be substantial, especially for large datasets.
  • Subtle leakages: The paper acknowledges that extremely subtle leakages might still exist through certain metrics like MAPQ scores, suggesting further investigation.
  • Imputation effects: The study explored imputation for missing genotypes, noting its complex impact on both privacy and utility, which warrants further detailed analysis.
  • Future work could focus on:
    • Quantifying privacy loss from other sources, such as gene expression values derived through eQTLs.
  • Investigating the impact of viral or bacterial reads in raw alignment files on privacy.
  • Developing more efficient sanitization algorithms.
  • Exploring the application of these techniques to other types of omics data.

Summary updated

Abstract

Abstract The generation of functional genomics datasets is surging, as they provide insight into gene regulation and organismal phenotypes (e.g., genes upregulated in cancer). The intention of functional genomics experiments is not necessarily to study genetic variants, yet they pose privacy concerns due to their use of next-generation sequencing. Moreover, there is a great incentive to share raw reads for better analyses and general research reproducibility. Thus, we need new modes of sharing beyond traditional controlled-access models. Here, we develop a data-sanitization procedure allowing raw functional genomics reads to be shared while minimizing privacy leakage, thus enabling principled privacy-utility trade-offs. It works with traditional Illumina-based assays and newer technologies such as 10x single-cell RNA-sequencing. The procedure depends on quantifying the privacy leakage in reads by statistically linking study participants to known individuals. We carried out these linkages using data from highly accurate reference genomes and more realistic environmental samples.
BibTeX
@misc{gursoy2018private,
  author = {Gamze Gürsoy and Prashant S. Emani and Charlotte M. Brannon and Otto Jolanki and Arif Harmanci and J. Seth Strattan and Andrew D. Miranker and Mark Gerstein},
  title = {{Private information leakage from functional genomics data: Quantification with calibration experiments and reduction via data sanitization protocols}},
  year = {2018},
  doi = {10.1101/345074},
  howpublished = {bioRxiv},
}
Figure