Research

Generate, federate, audit

A synthetic replica has to be made, it has to be made from data that cannot be pooled, and someone has to check what it gives away. Those are the three threads of the project.

Generate

Synthetic genomic data that keeps the biology

A synthetic dataset comes from a generative model trained on the real one. For gene expression, the real dataset is a table with a row for each gene and a column for each sample (or each cell, in single-cell data), and every entry says how active that gene was. A useful replica keeps the relationships a biologist would look for, such as which genes rise and fall together and which samples resemble each other.

We compared existing generators for bulk and single-cell RNA-seq and built new ones, from graphical models to foundation models. In the CAMDA 2025 Health Privacy Challenge, our joint team with Sage Bionetworks placed first in the bulk RNA-seq track and second in the single-cell track.

Inherited (germline) variation is harder. A person’s variants are stored in a variant call file (VCF), which lists the positions where their genome differs from a reference. With Sage Bionetworks we are generating synthetic VCFs for neurofibromatosis type 1 (NF1). To check them, we annotate the real and the synthetic variants and compare the pathogenicity predictions that come out. We are also testing how far generative foundation models can go in producing germline sequence directly.

Federate

One generator, trained across repositories that share nothing

Data on a rare disease is minimal and fragmented across custodians. A generator trained on all of it would be better than one trained on any single part, but a custodian is not permitted to hand its records to the others, or to a central server that would do the training.

Secure multi-party computation (MPC) removes the need to. Each custodian splits every value into random-looking pieces called shares and sends each piece to a different computing server. One share alone is a uniformly random number and says nothing about the value. The servers run the training computation on the shares, and only the final output is ever reassembled.

MPC protects the inputs. It does nothing about what the output itself gives away, so we pair it with differential privacy, which adds calibrated noise during training to limit how much any one donor can change the result. Together they give input privacy (no custodian discloses data to anyone, including the servers) and output privacy (the synthetic data is limited in what it reveals about any individual).

We have applied this to RNA-seq generation (Journal of Computational Biology, 2026), and we are packaging the algorithms and cryptographic protocols as a software library that custodians can run on NAIRR infrastructure.

Fig. 1

Vaults that never open

Hover a moving share
Every value that leaves a custodian is split into three shares that add up to it (mod 256), and each share goes to a different server. One share on its own is a uniformly random number, so no server learns anything about a record. The servers compute on the shares, and the synthetic replica is the only thing that is ever opened.

Audit

We attack the replica before anyone else can

A synthetic record belongs to no one, but the generator that produced it was trained on real people and can leak them. The standard test is a membership inference attack. The attacker holds one person’s data and the synthetic dataset, and has to decide whether that person was in the training set. In a rare disease cohort the answer is sensitive on its own, because being in the cohort means having the disease.

We build these attacks and run them against synthetic data, including our own. Our attack on synthetic single-cell RNA-seq data succeeds, and it was validated independently in the CAMDA Health Privacy Challenge (NeurIPS 2026). A companion study asks the same question of bulk RNA-seq, “Does Synthetic Bulk RNA-seq Data Protect Donors?”, and took first prize at CAMDA during ISMB 2026.

Single-cell data adds a problem that one-row-per-person tables do not have. A single donor contributes many cells, so a guarantee about one row covers only a small part of what that donor contributed. Privacy has to be measured per donor.

Germline data is next. With Sage Bionetworks we are measuring the privacy risk in synthetic germline data for rare diseases (RECOMB-Privacy 2026).

Fig. 2

How private, how useful

Drag the dial
Fidelity100%
Traceable100%
A toy generator makes one synthetic point per donor by adding noise to it. A synthetic point is marked traceable (red) when its nearest real donor is the donor it came from. Turn the dial up and the dense clusters stop being traceable first, while the isolated donor at the top right stays traceable long after the shape of the data has blurred. Real generators do not work point by point, which is why their leaks have to be measured with attacks.