Federated Analysis for Privacy-Preserving Data Sharing: A Technical and Legal Primer.

Continued advances in precision medicine rely on the widespread sharing of data that relate human genetic variation to disease. However, data sharing is severely limited by legal, regulatory, and ethical restrictions that safeguard patient privacy. Federated analysis addresses this problem by transferring the code to the data-providing the technical and legal capability to analyze the data within their secure home environment rather than transferring the data to another institution for analysis. This allows researchers to gain new insights from data that cannot be moved, while respecting patient privacy and the data stewards' legal obligations. Because federated analysis is a technical solution to the legal challenges inherent in data sharing, the technology and policy implications must be evaluated together. Here, we summarize the technical approaches to federated analysis and provide a legal analysis of their policy implications.

Modeling the impact of data sharing on variant classification.

OBJECTIVE: Many genetic variants are classified, but many more are variants of uncertain significance (VUS). Clinical observations of patients and their families may provide sufficient evidence to classify VUS. Understanding how long it takes to accumulate sufficient patient data to classify VUS can inform decisions in data sharing, disease management, and functional assay development. MATERIALS AND METHODS: Our software models the accumulation of clinical evidence (and excludes all other types of evidence) to measure their unique impact on variant interpretation. We illustrate the time and probability for VUS classification when laboratories share evidence, when they silo evidence, and when they share only variant interpretations. RESULTS: Using conservative assumptions for frequencies of observed clinical evidence, our models show the probability of classifying rare pathogenic variants with an allele frequency of 1/100 000 increases from less than 25% with no data sharing to nearly 80% after one year when labs share data, with nearly 100% classification after 5 years. Conversely, our models found that extremely rare (1/1 000 000) variants have a low probability of classification using only clinical data. DISCUSSION: These results quantify the utility of data sharing and demonstrate the importance of alternative lines of evidence for interpreting rare variants. Understanding variant classification circumstances and timelines provides valuable insight for data owners, patients, and service providers. While our modeling parameters are based on our own assumptions of the rate of accumulation of clinical observations, users may download the software and run simulations with updated parameters. CONCLUSIONS: The modeling software is available at https://github.com/BRCAChallenge/classification-timelines.

Federated analysis of BRCA1 and BRCA2 variation in a Japanese cohort.

More than 40% of the germline variants in ClinVar today are variants of uncertain significance (VUSs). These variants remain unclassified in part because the patient-level data needed for their interpretation is siloed. Federated analysis can overcome this problem by "bringing the code to the data": analyzing the sensitive patient-level data computationally within its secure home institution and providing researchers with valuable insights from data that would not otherwise be accessible. We tested this principle with a federated analysis of breast cancer clinical data at RIKEN, derived from the BioBank Japan repository. We were able to analyze these data within RIKEN's secure computational framework without the need to transfer the data, gathering evidence for the interpretation of several variants. This exercise represents an approach to help realize the core charter of the Global Alliance for Genomics and Health (GA4GH): to responsibly share genomic data for the benefit of human health.

A database of metazoan cytochrome c oxidase subunit I gene sequences derived from GenBank with CO-ARBitrator.

The Cytochrome C Oxidase subunit I gene ("COI") is the de facto standard for animal DNA barcoding. Organism identification based on COI requires an accurate and extensive annotated database of COI sequences. Such a database can also be of value in reconstructing evolutionary history and in diversity studies. Two COI databases are currently available: BOLD and Midori. BOLD's submissions conform to stringent sequence and metadata requirements; BOLD is specific to COI but makes no attempt to be comprehensive. Midori, derived from GenBank, has more sequences but less stringent standards than BOLD, resulting in higher error rates. To address the need for a comprehensive and accurate COI database, we adapted the ARBitrator algorithm, which classifies based only on sequence properties and has successfully auto-curated bacterial genes mined from GenBank. The adapted algorithm, which we call CO-ARBitrator, built a database of over a million metazoan COI sequences. Sensitivity and specificity are significantly higher than Midori. Specificity is comparable to what BOLD achieves with data quality prerequisites. Results and software are publicly available.