A first paper
Earlier this month, my first research paper went live on arXiv:
bioETH-PRS: Confidential Polygenic Risk Scoring without a Trusted Evaluator via Fully Homomorphic Encryption on a Programmable Blockchain
This was also my first collaboration with the University of Texas, working alongside Kimonas Provatas and Ilias Georgakopoulos-Soares. The experience was meaningful on multiple levels, and I’m grateful for the work we did together.
I won’t attempt a full academic walkthrough here. Instead, I want to share what the paper is about, what made it interesting to work on, and what I took away from the process.
The problem
Polygenic Risk Scores (PRSs) are a powerful tool in genomics. They aggregate thousands of genetic effect estimates, typically derived from Genome-Wide Association Studies (GWAS), into a single number that predicts an individual’s susceptibility to a particular disease.
The computation itself is straightforward: a dot product between a person’s genotype vector and a vector of GWAS-derived weights. But the data involved is about as sensitive as it gets. Your DNA.
In most existing approaches, computing a PRS requires exposing raw genotype data to third-party infrastructure. Some solutions use homomorphic encryption to protect the data during computation, but they still rely on a trusted evaluator, a designated party that performs the computation and is assumed not to collude or leak information.
The question we asked was: can we remove that trust assumption entirely?
The approach
bioETH-PRS replaces the trusted evaluator with immutable smart contracts on a blockchain that natively supports Fully Homomorphic Encryption (fhEVM). The core idea:
- The genotype vector is encrypted client-side and submitted as FHE ciphertexts
- The GWAS weights can be published publicly or stored encrypted
- The dot product is computed entirely in the encrypted domain; validators execute the arithmetic without ever seeing plaintext
- The result is released through a noisy oracle that emits only a categorical risk level (Low / Medium / High), never the raw score
No party (not the validators, not the contract deployer, not the model publisher) ever observes the raw genomic data or the exact score.
The architecture
We designed a four-contract system, each handling a distinct concern:
| Contract | Role |
|---|---|
| GenomicRegistry | Stores references to encrypted SNP data with per-address access control |
| ModelMarketplace | Lists GWAS weight vectors, public or private |
| PRSComputeEngine | Executes the chunked FHE dot product |
| ResultOracle | Adds on-chain noise and emits a categorical classification |
This separation ensures that data custody, model publication, computation, and output release are all isolated, both logically and in terms of access control.
We also introduced two computation paths: a classic chunked path (upload all SNPs, then compute) and a streaming path that interleaves upload and computation per chunk. The streaming path reduced gas by roughly 37% in our mock measurements.
Quantization
One of the more interesting technical challenges was representing signed floating-point GWAS weights as unsigned 64-bit integers, the only numeric type available in the TFHE scheme we used.
We developed a three-step fixed-point quantization scheme:
- Scale the floating-point weights to an integer range
- Shift to eliminate negative values
- Encode as
uint64
The scheme achieves machine-epsilon reconstruction accuracy on validated fixtures, and we built an on-chain-compatible decoding path so the final score can be interpreted correctly after decryption.
What I learned
Working on this was different from anything I’d done before. My background is in software engineering: building systems, writing infrastructure code, shipping products. Research has a different rhythm.
A few things stood out:
- Precision of language matters differently. In engineering, clarity serves maintainability. In a paper, every sentence carries weight because reviewers will challenge it. Hedging correctly (“may be cost-competitive” vs. “is cost-competitive”) is a skill.
- The gap between prototype and claim is vast. We had working contracts, gas profiles, and fixture-validated quantization. But turning that into defensible claims about a system’s properties required a level of rigor I hadn’t exercised in the same way before.
- Interdisciplinary work is humbling. Genomics, cryptography, blockchain, privacy. No one person masters all of these. The collaboration forced me to learn deeply in areas outside my comfort zone.
Looking forward
The paper suggests that confidential PRS computation on-chain is feasible today, particularly in low-gas deployment environments. There’s more to explore: formal differential privacy guarantees, multi-party GWAS weight contribution, scaling to larger SNP panels. I’m looking forward to continuing in this space.
If you’re interested, the paper is available on arXiv.
Leave a comment