TL;DR
This paper introduces a novel framework for conducting Genome-Wide Association Studies (GWAS) across multiple institutions without compromising patient privacy. By leveraging secure multi-party computation (MPC), the system allows for collaborative analysis of genomic data, enabling larger and more robust studies while keeping sensitive information localized.
What Problem Does This Paper Address?
Genome-Wide Association Studies (GWAS) are crucial for understanding the genetic basis of diseases. However, to achieve statistically significant results, GWAS often require a massive number of samples, far exceeding what a single institution can typically gather. Federated learning approaches, where data remains at its source, offer a solution by enabling cross-institutional collaboration. The primary challenge, however, is ensuring patient privacy and data confidentiality when genomic data, which is highly sensitive and immutable, is shared or processed across institutional boundaries. Existing privacy regulations like HIPAA, while important, can make direct sharing of raw genomic data infeasible, creating a significant barrier to collaborative research.
Methods
The core of this work is a privacy-preserving framework for GWAS built upon secure multi-party computation (MPC). Specifically, the authors utilize the Portable Circuit Format (PCF) framework and its associated runtime, BetterYao. The approach involves:
- MPC Layering: GWAS computations are layered on top of secure MPC systems, allowing two or more parties to jointly compute a function without revealing their individual private data.
- PCF Framework: This framework allows for the expression of computations in a high-level C-like language, which is then compiled into a low-level circuit representation for secure execution.
- Automated Workflow: An end-to-end workflow is implemented, involving data preprocessing (using Python), automatic generation of C-like code for MPC, compilation into garbled circuits using PCF/LCCYao, and execution via the BetterYao runtime.
- Specific Computations: The framework is demonstrated for two fundamental GWAS computations: Minor Allele Frequency (MAF) counting and χ² statistic calculation.
- Addressing PCF Limitations: The PCF/BetterYao system has limitations, such as small input size constraints and the requirement for fixed loop iterations. The authors address these by dynamically generating C code based on preprocessed metadata and by partitioning data into smaller chunks for processing.
Datasets
The experimental evaluation uses a sample dataset derived from the Personal Genome Project (PGP) participants and HapMap CEU individuals. This dataset consists of SNPs for 311 genotypes from 400 individuals, evenly split into case and control groups (200 each). For performance testing, this dataset was replicated to simulate larger datasets with up to 9330 genotypes.
Key Insights and Ideas
- MPC as a Privacy Enabler for GWAS: The paper demonstrates that MPC can effectively bridge the gap between the need for large-scale genomic data and stringent privacy requirements, making federated GWAS a practical reality.
- Trade-offs for Performance: The authors acknowledge that certain aspects of their design, like revealing the size of input data, are a deliberate trade-off for significant performance gains, arguing this coarse-grained information is acceptable.
- Parallelism is Key: Despite PCF/BetterYao not natively supporting parallelism, the authors achieve substantial speedups by parallelizing computations across genotypes, which are independent.
- Garbled Circuits for Statistical Computations: The successful implementation of MAF and χ² statistics using garbled circuits highlights the applicability of MPC to complex statistical analyses in bioinformatics.
- Evaluator vs. Generator Communication Costs: A notable finding is that the circuit evaluator incurs significantly higher communication costs than the circuit generator, a factor to consider in system design.
Key Results
- Efficient Computation: The framework demonstrates promising performance, with the χ² statistic computation over thousands of genotypes taking only a few minutes.
- Scalability: Performance scales well with the number of CPU cores, showing a roughly 3.4x speedup when increasing cores from 1 to 4.
- Genotype Scalability: Running times for MAF and χ² computations scaled linearly with the number of genotypes. For example, increasing genotypes from 311 to 3110 increased MAF running time by approximately 9.4x.
- Invariant to Individual Count: Varying the number of individuals in the case/control groups had no significant impact on the online computation time, as this primarily affects the offline preprocessing stage.
- Communication Costs: Data transfer for χ² computations was 3-4 times greater than for MAF computations due to the complexity of the underlying operations (multiplication and division).
Limitations and Future Work
- Two-Party Limitation: The current implementation is limited to two-party computations, restricting collaborative analysis to only two institutions. The authors plan to extend this to support more participants by leveraging other MPC platforms or re-engineering the implementation.
- Computational and Communication Overhead: While promising, the MPC approach still incurs noticeable overhead compared to centralized, non-encrypted environments. This can be a bottleneck for extremely large datasets.
- Further Parallelization: The authors suggest further investigation into improving privacy-preserving federated GWAS analysis by enhancing parallel computing capabilities.
- Security Model: The security guarantees are largely dependent on the underlying MPC runtime protocol, which can support varying security models (semi-honest to malicious). The current design leaks aggregated information about per-party input data size.
Summary updated