Can PQCF help genome research?
Measured, and mostly no.
A research branch beside the core platform: every promising photonic route for genomics measured on real data against the best classical method, with a proven optimum as the reference.
Can PQCF help genome research?
Measured on a virus, a bacterium and a human chromosome.
We started a research branch, developed in parallel with the core platform, to answer one question: can a photonic quantum simulator like PQCF be used for genome research? We used several research papers on quantum computing for genomics, then measured every promising route on real genomic data, against the best classical methods, with every answer checked against a proven optimum.
The papers:
- The critical review: Maurizio and Mazzola, arXiv:2507.04111. Grover-style search for read alignment loses its speedup once the cost of loading the data is counted. Realistic candidates are hard optimization problems of modest size, and any claimed advantage must be measured against the best classical solver.
- Genome assembly on a quantum annealer and on a simulated optical Ising machine: Boev et al., arXiv:2004.06719.
- Quantum sequence alignment: Creevey, Jing, Hollenberg, arXiv:2506.22775, after Hollenberg's original proposal, arXiv:quant-ph/0002076.
- Encoding genomes as quantum states: Creevey et al., arXiv:2508.06184.
- DNA sequences as quantum images: Kösoglu-Kind et al., arXiv:2303.13608.
Where photonics could fit. PQCF is not a qubit machine, so it does not run Grover-style search. Photonics has two tools for the problems the review names:
- Gaussian boson sampling (GBS), which samples dense subgraphs and cliques of a graph (arXiv:1803.10730).
- Coherent Ising machines, optical solvers for Ising problems such as haplotype phasing (simulated as SimCIM, arXiv:1901.08927, and with amplitude-heterogeneity correction, arXiv:1810.12565).
What we measured, in five steps:
- The classical baselines. Assembly, haplotype phasing and dense modules of a gene network, built from public data, with their exact optima.
- Translated into the binary form annealers need, assembly becomes far harder: a problem solved exactly in a millisecond defeats every solver at 24 reads.
- Gaussian boson sampling, simulated exactly by PQCF. For dense modules of a gene network, it needs 2 to 21 times fewer samples than the best classical sampler (Oh et al., arXiv:2302.00536), a lead growing with size.
- But simulating it costs 800 to 3,000 times more time.
- An advantage needs a programmable device delivering 17,000 to 450,000 usable samples per second, and none exists.
- Coherent Ising machines. No advantage on any instance class.
- Short-read phasing is solved exactly in linear time by dynamic programming.
- Long-read phasing is won by classical parallel tempering, about 4 times faster than the best CIM-derived solver.
- The GBS graph kernel (arXiv:1905.12646), to tell a virus, a bacterium and a human chromosome apart from short windows of DNA.
- At these sizes its features are computed exactly without a quantum device.
- They reach 47–60% accuracy, against 84–96% for plain counts of three-letter words.
- Genomes as quantum states. We reproduce the published φX174 result (bond dimension 98).
- φX174, SARS-CoV-2, E. coli and human chromosome 21 need exactly what a random sequence of the same length needs: they do not compress at all.
- Loading a genome into a quantum computer costs as much as the genome is long.
A worked example on public genomes, run end to end:
| task | data | result |
|---|---|---|
| assembly | SARS-CoV-2, whole genome: 599 reads | exact, 0.05 s, equal to the reference. As an annealer problem: 358,801 variables. |
| assembly | E. coli K-12, the lac operon region (50,000 bases): 1,000 reads | exact, 0.12 s, equal to the reference |
| phasing | a simulated individual on 1 million bases of human chromosome 21 (648 variants) | short reads: 110 separate pieces. Long reads: one piece, exact in 0.17 s, no errors. |
| protein modules | E. coli's DNA replication and repair network (45 proteins, STRING) | found with PQCF's simulated GBS: the replication machinery (DnaA, DnaC, DnaG, DnaE, DnaN, DnaX) and the RecFOR–RuvAB repair pathway |
| quantum encoding | SARS-CoV-2, E. coli, human chromosome 21 | no compression: as costly as a random sequence |
The answer, plainly:
- Classical exact methods solve these genomic tasks in under a second.
- The photonic routes either need hardware that does not exist or lose to classical methods.
- Under every quantum route sits the cost of loading the genome, which we measured to be as large as the genome itself.
PQCF's part is an exact, checked simulation of the photonic devices and a measurement discipline that can be rerun when hardware appears: a precise testbed for photonic ideas, not a faster genome tool.
Mistakes we made, also written down.
- Our first measurement made haplotype phasing look hard. The exact solver had been run on a single core; on eight, and by dynamic programming, the problem is easy.
- One run measured its solvers against a target that was not the true optimum. Our checks caught it, and the run was repeated.
The research, every measurement and the code: the summary, the five steps and the worked example (its results), on the research/genome branch.
How to reproduce the results: a step-by-step page: the genomes from NCBI checked by SHA-256, then the worked example (about 6 minutes on 8 cores) and each study G0–G4; every result must equal the published one.