Abstract
In an age where commercial entities collect and profit from large amounts of private information, science must strive to offer society viable ways to preserve privacy while benefiting from the power of data sharing. Patient phenotypes and genotypes are critical for building groups of phenotypically-similar patients, identifying the gene that best explains their common phenotypes, and ultimately diagnosing Mendelian diseases. However, direct computation over these quantities requires sensitive data to be shared openly, compromising patient privacy. While existing protocols address secure computation over genotype data, they often ignore the initial step of identifying phenotypically-similar patients. In this work, we develop secure protocols for COHORT DISCOVERY and GENE PRIORITIZATION, enabling a complete, secure end-to-end disease diagnosis pipeline.
Introduction
Diagnosing rare Mendelian diseases—disorders caused by mutations in a single gene—is a complex process that relies heavily on data aggregation. Clinicians must compare a patient’s observable symptoms (phenotypes) and genetic variants (genotypes) against vast databases of known cases. This necessity creates a fundamental tension: while sharing data is essential for a life-saving diagnosis, it exposes patients to significant privacy risks, including genetic discrimination and data breaches.
Current cryptographic solutions have largely focused on the final stages of the diagnostic pipeline, such as secure genotype comparison. However, the critical first step—finding a cohort of similar patients based on clinical symptoms—has remained a privacy bottleneck. This research introduces a secure, end-to-end pipeline that protects sensitive patient data from the initial clinical encounter through to the final genetic ranking.
Methodology: Secure Multiparty Computation
The proposed system utilizes Secure Multiparty Computation (MPC), a subfield of cryptography that allows multiple parties to jointly compute a function over their inputs while keeping those inputs private. Specifically, the authors employ the ABY approach, a hybrid framework that switches between three types of secret sharing:
- Arithmetic Sharing: Optimized for additions and multiplications, used for calculating similarity scores.
- Boolean Sharing & Yao’s Garbled Circuits: Optimized for non-linear operations like comparisons and finding maximum values.
The Two-Cloud Model
The architecture follows a "two-cloud" model involving two non-colluding servers (Node 1 and Node 2). Patient data is split into "shares"—randomized components that are meaningless in isolation. Each node receives one share. By performing mathematical operations on these shares, the nodes can compute the final diagnostic result without ever "seeing" the underlying patient data. To ensure real-world viability, the system uses a preprocessing phase where a third-party "dealer" generates precomputed random values (Beaver triples), shifting the heaviest computational burdens to an offline stage.
Technical Innovations
The researchers introduced several key adaptations to make secure diagnosis both accurate and efficient:
Adapting Phrank for MPC
The core of the diagnostic logic is based on Phrank (Phenotype RANKing), an algorithm that treats symptoms as nodes in a weighted graph based on the Human Phenotype Ontology (HPO). To make Phrank compatible with MPC, the authors reformulated the graph-based similarity calculations into a series of matrix-vector multiplications. This allows the system to compute complex phenotypic similarities using efficient arithmetic sharing.
Binary Projections and Provider Privacy
To accelerate gene prioritization, the protocol uses "projection matrices." Instead of scanning all 20,000+ protein-coding genes, the system securely narrows the search to approximately 3,400 known disease-causing genes. Furthermore, the protocol protects the service provider’s proprietary knowledge base; the patient only learns the top-ranked candidate genes, while the provider’s full database remains encrypted.
Experimental Results
The protocols were validated using real-world datasets, including the Stanford STARR database and the Deciphering Developmental Disorders (DDD) cohort.
Accuracy
In gene prioritization tests involving 169 diagnosed patients, the secure protocol successfully placed the correct causative gene in the top 10 list for nearly 90% of cases (150 patients). This demonstrates that the cryptographic overhead does not degrade the clinical utility of the diagnostic algorithm.
Performance and Scalability
The system proved efficient enough for clinical deployment:
- Cohort Discovery: Identifying two distinct disease groups within a pool of 1,067 patients took approximately 10 seconds.
- Gene Prioritization: Ranking thousands of genes for an individual patient was completed in 84 seconds.
While gene prioritization scales linearly with the size of the reference database, cohort discovery currently scales quadratically. This means that while the system is highly effective for groups of 1,000 patients, further optimization is required to handle populations in the hundreds of thousands.
Conclusion and Future Work
This work demonstrates that privacy and medical progress are not mutually exclusive. By securing the entire diagnostic pipeline, from phenotype-based cohort discovery to genotype-based gene prioritization, the researchers have provided a blueprint for a global, secure diagnostic network. Future developments may include integrating these protocols into mobile applications and expanding the system to include automated literature reviews, further empowering patients and clinicians in the search for answers to rare genetic diseases.
Summary updated