FANCY: Fast Estimation of Privacy Risk in Functional Genomics Data

Gamze Gürsoy, Charlotte M. Brannon, Fábio C. P. Navarro, Mark Gerstein

TL;DR

This paper introduces FANCY, a novel computational tool that estimates the privacy risk in functional genomics data (RNA-Seq, ATAC-Seq, ChIP-Seq) by predicting the number of "leaking variants" without requiring explicit genotyping. This allows researchers to quickly assess data privacy before release, accelerating scientific discovery.

What Problem Does This Paper Address?

Functional genomics data, generated from techniques like RNA-Seq, ATAC-Seq, and ChIP-Seq, is becoming increasingly valuable for clinical applications and scientific advancement. However, this data also contains information that could potentially be used to re-identify individuals, raising significant privacy concerns. Traditional methods for assessing privacy risk often rely on genotyping, which is computationally intensive, resource-heavy, and requires access to gold-standard genotype data that may not always be available. This creates a bottleneck for data sharing and can delay the release of crucial scientific findings. The paper aims to bridge this gap by providing a fast, accurate, and accessible method for quantifying privacy leakage directly from raw functional genomics reads.

Methods

The core of FANCY is a two-step supervised learning approach.

  1. Feature Extraction: The method utilizes readily available sequencing statistics derived from raw reads. These include:
  2. Average depth per base pair (d): How many reads cover each base on average.
  3. Breadth of coverage (b): The fraction of the genome covered by at least one read.
  4. Statistical properties of depth distribution: Standard deviation (σ), skewness (±s), and kurtosis (k) of the read depth across the genome.
  5. Type of experiment: Categorical information indicating whether the data is from RNA-Seq, ATAC-Seq, or ChIP-Seq.
  6. Privacy Risk Estimation:
    • Gaussian Process Regression (GPR): A GPR model with a Matern kernel is employed to predict the cumulative number of single nucleotide variants (SNVs) that could leak privacy. GPR is chosen for its ability to provide uncertainty estimates.
  7. Rare vs. Common Variant Estimation: A second step estimates the number of rare and common leaking variants separately, providing a more nuanced risk assessment.
  8. FANCYlow: A specialized version of the model trained on data with fewer than 1000 leaking variants is developed to improve accuracy in low-leakage scenarios, which are critical for privacy assessment.

Additionally, the authors developed a Random Forest classifier as a plug-in to FANCY. This classifier uses the same sequencing statistics to predict the type of experimental assay (RNA-Seq, ATAC-Seq, ChIP-Seq), which can be useful for datasets with missing metadata.

Datasets

The study utilized several publicly available datasets for training and validation:

  • RNA-Seq data: From the gEUVADIS project (432 individuals) and a study by Kilpinen et al. (used for validation).
  • H3K27ac ChIP-Seq data: From the PsychENCODE Consortium (100 individuals).
  • ATAC-Seq data: From the BrainGVEX project (344 individuals) and the PsychENCODE Consortium (288 individuals).

The researchers also used data from the 1,000 Genomes Project to categorize variants into rare and common categories. For determining the "true number of leaking SNVs" during model development, they employed GATK Best Practices on RNA-Seq and DNA data, and overlapped these with gold-standard SNVs from whole-genome sequencing.

Key Insights and Ideas

  • Sequencing statistics as privacy proxies: The study demonstrates that fundamental sequencing statistics, which are easily obtainable from raw reads, are surprisingly good indicators of privacy leakage. This bypasses the need for complex genotyping.
  • Assay type matters: Different functional genomics assays (RNA-Seq, ATAC-Seq, ChIP-Seq) probe different genomic regions and have distinct coverage profiles, leading to varying privacy risks. FANCY accounts for this by including assay type as a feature.
  • Low-leakage scenarios are critical: The authors highlight that the privacy risk assessment is most valuable when the number of leaking variants is low, as even a few additional leaked variants can drastically change the privacy status of a dataset. This motivated the development of FANCYlow.
  • Feature importance: Breadth of coverage and mean depth were found to be the most influential features for predicting leaking variants, though using all features yielded the most accurate model.

Key Results

  • High prediction accuracy: FANCY achieved an average R² of 0.99 for training and an R² of 0.90-0.99 on independent test sets for different assay types.
  • Low prediction error: On average, FANCY exhibited an 8% prediction error across all independent test sets.
  • FANCYlow performance: The specialized FANCYlow model achieved RMSEs of 75.64 (RNA-Seq), 74.8 (ATAC-Seq), and 74.0 (ChIP-Seq) for predicting leakage below 1000 variants.
  • Qualitative risk assessment: FANCY provides a simple color-coded warning message (green, yellow, red) for data sharing risk, making it accessible to a broad range of users.
  • Random Forest classifier: The assay type prediction classifier achieved an average accuracy of 96.8%, precision of 94.9%, recall of 90.2%, and F1 score of 93.3%.

Limitations and Future Work

  • Computational complexity of GPR: While GPR offers uncertainty estimation, it can be computationally intensive for very large datasets.
  • Under-prediction in low-leakage scenarios: Although FANCYlow was developed to address this, minimizing under-predictions remains a critical challenge when an individual's privacy is at stake.
  • Future Work: The authors suggest that FANCY can be used to plan functional genomics experiments by back-calculating required sequencing statistics for a desired number of variants. The Random Forest classifier also offers utility for data utility maximization in cases of missing metadata. The paper also mentions the potential for FANCY to accelerate data sharing by providing rapid privacy risk assessments.

Summary updated

Abstract

Abstract Functional genomics data is becoming clinically actionable, raising privacy concerns. However, quantifying the privacy leakage by genotyping is difficult due to the heterogeneous nature of sequencing techniques. Thus, we present FANCY, a tool that rapidly estimates number of leaking variants from raw RNA-Seq, ATAC-Seq and ChIP-Seq reads, without explicit genotyping. FANCY employs supervised regression using overall sequencing statistics as features and provides an estimate of the overall privacy risk before data release. FANCY can predict the cumulative number of leaking SNVs with a 0.95 average R 2 for all independent test sets. We acknowledged the importance of accurate prediction even when the number of leaked variants is low, so we developed a special version of model, which can make predictions with higher accuracy for only a few leaking variants. A python and MATLAB implementation of FANCY, as well as custom scripts to generate the features can be found at https://github.com/gersteinlab/FANCY . We also provide jupyter notebooks so that users can optimize the parameters in the regression model based on their own data. An easy-to-use webserver that takes inputs and displays results can be found at fancy.gersteinlab.org .
BibTeX
@article{gursoy2021fancy,
  author = {Gamze Gürsoy and Charlotte M. Brannon and Fábio C. P. Navarro and Mark Gerstein},
  title = {{FANCY: Fast Estimation of Privacy Risk in Functional Genomics Data}},
  journal = {Bioinformatics},
  year = {2021},
  doi = {10.1093/bioinformatics/btaa661},
}
Figure