A constraint preserving mixing strategy called an XY mixer — which enforces the one guide per gene rule structurally in a quantum circuit, bypassing tunable penalty terms — beats three penalty tuned variants in simulation and holds up better on
A new paper pits two ways of teaching a quantum computer to pick CRISPR guide RNAs for a small multiplex-editing problem, and shows what happens to each one when the simulator hands off to the chip.
The work, COMET, runs on a 12-qubit circuit built to pick guide RNAs for three immune-checkpoint genes at once: PD-1, LAG-3, and TIM-3 (the gene symbols PDCD1, LAG3, and HAVCR2). Multiplex editing is the workhorse of functional-genomics screens, and the question of which guides to use is combinatorial. Quantum optimizers are a natural fit in principle.
Both approaches encode the choice as a QUBO (quadratic unconstrained binary optimization), the standard formulation a quantum annealer or gate-based optimizer can chew on. The difference is how they enforce the constraint that each gene gets exactly one guide. The penalty approach adds a soft term to the cost function that grows with violations. The XY-mixer bakes the constraint into the circuit's mixing operation, so feasibility holds by construction without any tuning.
In noiseless simulation, the choice is decisive. The XY-mixer crosses 95% probability of the optimum at circuit depth p=3. All three penalty variants, tested across an order of magnitude of penalty strengths, stay below 6% at every depth the authors measured (paper).
Then the authors put the same circuit on IBM's Heron r2 processor, ibm_kingston (processor guide). The structural guarantee of the XY-mixer largely holds: the simulator-versus-hardware energy gap stays inside roughly 0.8 across depths. The worst-tuned penalty variant drifts to about +53.9 on hardware, a divergence that turns a near-optimal simulation pick into noise on the chip.
Two caveats the authors flag themselves. The instance, three genes and twelve qubits, is "classically trivial." They are not claiming quantum advantage over a laptop. The structural guarantee of the XY-mixer "partially breaks under gate-level noise," which is the practical limit of "by construction." Both admissions are part of why the methodological comparison holds up.
As quantum optimizers start to land on real biological problems, the comparison framework matters more than any absolute number. The COMET setup, same circuit family with the same problem instance and multiple constraint strategies and paired simulator and hardware numbers, is a template for benchmarking how an algorithm respects biological constraints when it leaves the simulator.
What it does not claim is that quantum computers are now useful for picking guides at scale, that PD-1 / LAG-3 / TIM-3 therapy has a new tool, or that one of the strategies is retired. The penalty approach is heuristic by construction and noisy on near-term devices. The structural approach buys you feasibility in simulation and most of the way on hardware. The next step is the one the authors do not run here: scaling to the gene counts a real functional-genomics screen actually uses, where the constraint-preserving advantage has to be re-measured rather than assumed.