SPHERE: Making Sensitive Data Directly Usable and Shareable in the Age of AI

SPHERE is a model-free method that converts sensitive datasets into shareable synthetic twins while the original records stay in the local environment. Compared with sharing raw data or relying on centralized access, it separates usability from data leaving its domain, so AI systems and external collaborators can work without touching the original records. The abstract reports evaluation across 33 datasets in five scientific domains, together with certification of each twin's privacy and fidelity.

For statistical structure, the twin reproduces means, variances and correlations exactly, gives numerically identical effect size and P value in linear statistical analysis, and retains nonlinear machine-learning utility. This offers a concrete fidelity characterization for whether synthetic data can stand in for original data in analysis, rather than a general usability claim. Based on the abstract's statements about reproducing statistical structure and numerical identity in linear analysis; specific numbers, sample sizes and figures are not given in the loaded text.

Frontier AI agents running on the twin reach the same scientific conclusions as on the original records, and analyses reproduce genome- and proteome-wide results at UK Biobank scale and recover landmark studies across three independent cohorts and consortia. It moves validation of synthetic twins from statistical metrics to end-to-end scientific conclusions, and extends to large-scale omics and multi-cohort replication. Abstract-level reporting involving UK Biobank scale and three independent cohorts and consortia; specific replication metrics and statistics are not presented in the loaded text.

The approach extends to deep-learning embeddings across language, vision and time-series with minimal utility loss, and the Stanford Alzheimer's Disease Research Center cohort is made openly available for the first time as a SPHERE twin spanning nine modalities that any registered researcher can analyze without an approval process. It broadens applicability from tabular sensitive data to embedding representations and provides a concrete open-data release case. Abstract statements about the embedding extension and the cohort release; modality details, access mechanics and the magnitude of utility loss are not elaborated in the loaded text.

The work targets researchers, data-holding institutions and collaborative consortia that need to share or let AI use sensitive human data under privacy constraints, in settings where original records must remain local while external parties need to analyze or model them; its claims cover five scientific domains, UK Biobank-scale omics analyses, replication across three independent cohorts and consortia, embeddings across language, vision and time-series, and an open release of a nine-modality Alzheimer's cohort twin.

A careful reader would still watch: the specific setup and strength of the adversarial re-identification attacks, the criteria behind the privacy and fidelity certification, the quantified magnitude of retained nonlinear machine-learning utility, the concrete measure of minimal utility loss in the embedding extension, and the access and governance arrangements for the open nine-modality cohort twin; because the loaded text reflects an incomplete reading scope without figures or tables, these details cannot be confirmed from the available text.