Abstract
Given the rapidly growing utility of critical health information revealed in the human genome, secure genomic computation is essential to moving forward, especially as genome sequencing becomes commonplace. We devise and implement proof-of-principle computational operations for precisely identifying causal variants in Mendelian patients using secure multiparty computation methods based on Yao’s protocol. We show multiple real scenarios (small patient cohorts, trio analysis, two hospital collaboration) where the causal variant is discovered jointly, while keeping up to 99.7% of all participants’ most sensitive genomic information private. All similar operations performed today to diagnose such cases are done openly, keeping 0% of participants’ genomic information private. Our work will help usher in an era where genomes can be both utilized and truly protected.
Introduction
Genomic data presents a fundamental "serve or protect" dilemma. To "serve" a patient and find the root cause of a rare disease, clinicians must compare the patient's genome against thousands of others. However, to "protect" the patient, their genome should remain private, as it contains sensitive information regarding disease susceptibility, ancestry, and even the health risks of their relatives. Currently, genomic sharing is an all-or-nothing game: researchers either share data "in the clear" (0% privacy) or use obfuscated summary statistics that often lack the resolution needed for precise diagnosis. This research addresses the need for a middle ground: a way to perform exact computational operations on genomes while ensuring that no party learns anything about the inputs other than the final result.
Methodology
Secure Multiparty Computation (SMC)
The framework leverages Secure Multiparty Computation (SMC), specifically Yao’s protocol (Garbled Circuits), to perform genomic analysis. This cryptographic approach allows multiple parties to jointly compute a function over their inputs while keeping those inputs private.
Data Representation and Architecture
- Bit Vector Conversion: Genomes are converted into bit vectors. For variant-level searches, a variant vector (length ~28 million) is used where "True" indicates the presence of a rare mutation. For gene-level searches, a gene vector (length ~20,000) is used where "1" indicates the gene contains at least one rare functional variant.
- The Two-Cloud Model: To scale the computation to many participants, the system utilizes two non-colluding servers (e.g., managed by different government agencies). Participants "split" their data using additive secret sharing between the two clouds. Neither cloud can see the original data, but together they can execute the logic.
Core Diagnostic Operations
Three primary logic gates were implemented to facilitate common diagnostic workflows:
- MAX: Identifies the gene mutated in the highest number of patients in a cohort.
- SETDIFF (Set Difference): Used in "trio analysis" to find de novo mutations in a child that are absent in both parents.
- INTERSECTION (AND): Finds rare variants shared between two different hospitals or datasets.
Experimental Validation
The study validated its approach using real-world patient data from several sources, including dbGaP studies for Freeman Sheldon Syndrome (FSS) and Miller Syndrome (MiS), 1,210 patients from the Centers for Mendelian Genomics, and a clinical trio from Stanford Hospital.
The Protection Quotient (PQ)
A key contribution of this work is the Protection Quotient (PQ), a metric used to quantify privacy. It is defined as the fraction of private variants withheld from the output divided by the total variants input. Across all scenarios, the methods consistently achieved a PQ of 97.1% to 99.7%.
Performance and Scalability
The framework proved both accurate and computationally practical:
- Cohort Analysis (MAX): Successfully identified causal genes (e.g., MYH3, NOTCH2) in four different disease cohorts in just 5–10 seconds with a PQ >99%.
- Trio Analysis (SETDIFF): Identified the causal ACTB variant in a family. While a single-threaded execution took 57 minutes, a 30-node cluster brought this down to under 2 minutes.
- Hospital Collaboration (INTERSECTION): Two hospitals compared ~11,000 variants and found a shortlist of 159 shared variants in 9.4 minutes, maintaining a 97.1% PQ.
Notably, the protocol's execution time and bandwidth grow logarithmically with the number of participants for cohort and trio analyses, suggesting high scalability for larger studies.
Discussion and Limitations
While the framework demonstrates the feasibility of privacy-preserving genomics, several challenges remain. The current implementation assumes an "honest-but-curious" model, where participants follow the protocol but may attempt to learn extra information. Extending this to defend against malicious actors would increase computational overhead. Furthermore, the study focuses on single-gene (Mendelian) disorders; extending this to complex polygenic traits would require more complex Boolean circuits. Finally, the model relies on the existence of two truly non-colluding entities to host the computation, which presents a significant policy and infrastructure challenge.
Conclusion
This work demonstrates that it is possible to identify disease-causing variants with high precision while keeping the vast majority of a patient's genomic data private. By moving away from the 0% privacy model currently used in clinical genomics, this framework helps usher in an era where genomes can be both utilized for life-saving diagnostics and truly protected.
Summary updated