Abstract
Functional genomics experiments, such as RNA-seq, provide non-individual specific information about gene expression under different conditions such as disease and normal. There is great desire to share these data. However, privacy concerns often preclude sharing of the raw reads. To enable safe sharing, aggregated summaries such as read-depth signal profiles and levels of gene expression are used. Projects such as GTEx and ENCODE share these because they ostensibly do not leak much identifying information. Here, we attempt to quantify the validity of this statement, measuring the leakage of genomic deletions from signal profiles. We present information theoretic measures for the degree to which one can genotype these deletions. We then develop practical genotyping approaches and demonstrate how to use these to identify an individual within a large cohort in the context of linking attacks. Finally, we present an anonymization method removing much of the leakage from signal profiles.
Introduction
In the era of precision medicine, sharing genomic data is vital for research, but privacy concerns remain a significant bottleneck. To mitigate the risks associated with sharing raw DNA sequences, researchers frequently distribute "aggregated summaries" or "signal profiles." These profiles represent read-depth—the frequency with which specific genomic regions are sequenced—rather than the raw nucleotide sequence itself.
Large-scale consortia, such as ENCODE and GTEx, publicly share these profiles under the assumption that they do not leak sufficient identifying information to compromise participant privacy. This study challenges that assumption by investigating whether structural variants, specifically genomic deletions, leave a detectable "fingerprint" in signal profiles that can be used to link an anonymous profile to a known individual in a restricted-access database.
Methodology
The research framework quantifies and exploits information leakage through a two-step "linking attack" involving genotyping and profile matching.
Genotyping from Signal Profiles
The authors developed methods to predict an individual's deletion genotypes directly from functional genomics signal profiles. The approach varies by data type:
- RNA-seq: Small deletions are identified by sudden "dips" in signal, quantified using a self-to-neighbor signal ratio.
- ChIP-seq: Large deletions are detected using average signal levels across broader genomic regions.
- Hi-C: 3D interaction matrices are collapsed into 1D signal profiles to identify structural variants.
The fundamental mechanism relies on the fact that genomic deletions result in a characteristic loss of signal in the corresponding regions of the functional genomics profile.
Linking Attacks and Metrics
Predicted genotypes are compared against a restricted genotype dataset (e.g., the 1000 Genomes Project). A successful match between the predicted deletions from a public profile and the known deletions of an individual in a private database constitutes a re-identification. To quantify this threat, two metrics were introduced:
- Genome-wide Predictability ($\pi_{GW}$): Measures the accuracy with which a variant can be inferred from a profile.
- Individual Characterizing Information (ICI): Measures the extent to which a specific variant helps uniquely identify an individual within a cohort.
Key Results
The study validated the attack using datasets from GEUVADIS, GTEx, ENCODE, and the 1000 Genomes Project.
- High Re-identification Accuracy: In the GEUVADIS dataset, the authors achieved 100% linking accuracy using approximately 40 variants predicted from RNA-seq profiles.
- Cross-Dataset Robustness: The attack remained effective even when the deletion discovery occurred in one population and the attack was executed on another.
- Vulnerability of 3D Data: Hi-C contact matrices were found to be vulnerable at resolutions of 10 kb or smaller.
- Effective Defense Mechanism: The authors proposed a median filter anonymization method to smooth out the signal "dips" that leak information. Applying this filter to GEUVADIS RNA-seq profiles reduced the percentage of vulnerable individuals from nearly 100% to just 1%, while preserving the scientific utility of the data.
Discussion and Conclusion
Functional genomics datasets present a "Yin-Yang" challenge: they are designed to capture dynamic gene activity but inadvertently record permanent, identifying genomic structures. While RNA-seq provides high depth for small deletions, ChIP-seq offers breadth for larger variants. Crucially, only a small number of correctly genotyped deletions (30-40) are required to uniquely identify an individual among thousands.
While this study focused on deletions, other structural variants like insertions or inversions may also leak information. Future work should explore the potential for machine learning to predict genotypes from higher-level data, such as metabolic pathway activity, ensuring that privacy protections evolve alongside genomic data sharing practices.
Summary updated