Most cancer-causing mutations that remain undiscovered are rare: each appears in under 2% of patients, so counting mutations cannot separate them from harmless "passenger" mutations. The strongest current methods run a random walk over a gene network, or over a hypergraph in which each patient is one hyperedge joining all of that patient's mutated genes, and rank genes by how often the walker visits them. A recent study found that replacing the classical walker with a quantum walk roughly doubles early retrieval of disease genes on protein networks. In a quantum walk, the walker is a wave that spreads along many paths at once and interferes with itself. The reported benefit is attributed to that interference. Real quantum hardware destroys interference through dephasing, the random scrambling of phases that turns a quantum walk back into ordinary diffusion. The usual framing treats this as a reason quantum walks will fail on hardware. We propose to treat dephasing strength as a tunable parameter, with ordinary diffusion at one end and a pure quantum walk at the other. The question is where along that range rare cancer drivers are ranked best.
The claim to be tested is simple. On a patient–gene mutation hypergraph, a partially dephased quantum walk ranks rare drivers better than both the pure quantum walk and the classical walk. The biology can be settled entirely with exact classical simulation at genome scale, because at this problem size no quantum computer is needed. Hardware enters only as a small, secondary calibration: we measure how much dephasing a real device imposes, using a qubit-remapping trick that changes noise without changing the circuit. If the best ranking occurs at zero dephasing, the result is a clean negative for "noise as regularizer." If it occurs at infinite dephasing, it is a negative for the quantum-walk advantage in cancer. Either answer is informative.
Core Claim
On Quantum Random Walks in Biomolecular Networks motivates this work. It evolves a continuous-time quantum walk under the network's adjacency structure as the Hamiltonian, meaning the matrix that governs how the wave moves. It reports a mean average precision in the top 20 (AP@20) of about 0.30 against 0.14 for classical random walk with restart on autism seed genes. It studies only germline disease seeds on pairwise networks, and only noiseless evolution. DriverRWH and PDRWH supply the representation. They show that treating each tumor as a hyperedge preserves which genes co-occur in the same patient, and that this helps rank rare drivers. Both still use classical Markov walks. Nobody has run a quantum walk on this hypergraph. Nobody has measured how the reported interference benefit degrades as phases are scrambled.
We model the walker's state as a density matrix $\rho$, a description that covers both fully quantum and fully classical mixtures. It evolves as
$$\frac{d\rho}{dt} = -i[H,\rho] + \gamma\left(\sum_j P_j \rho P_j - \rho\right),$$
where $H$ is the Hamiltonian built from the hypergraph, $P_j$ projects onto node $j$, and $\gamma \ge 0$ is the dephasing rate. At $\gamma = 0$ this is the pure quantum walk of Dubovitskii et al. As $\gamma$ grows large, the phase information between nodes is erased. The dynamics then reduce to a classical continuous-time walk with hopping rates proportional to $|H_{ij}|^2/\gamma$, which is the standard strong-dephasing limit. This single dial therefore places quantum and classical walks on one axis. Our hypothesis is that an intermediate $\gamma$ helps. At that setting, interference suppresses the pull of hub genes, while enough phase-scrambling remains to stop the walker oscillating in or getting stuck in localized states on the sparse patient branches where rare drivers sit.
Research Program
We build the Hamiltonian on the star expansion of the mutation hypergraph. This is a two-layer graph with gene nodes and patient nodes, in which each patient node links to that patient's mutated genes. We take this directly from how DriverRWH and PDRWH use hyperedges. Gene–gene edges come from STRING or HumanNet, as in DriverRWH. Hyperedge weights follow PDRWH's expression-similarity weighting where expression data are available. Patient nodes keep co-occurrence context that a clique expansion would blur. We will also test the clique expansion as an ablation.
The walker starts from a known-driver seed set, and genes are scored by time-averaged occupation over a window $[0, T]$. The quantities $\gamma$, $T$, and the balance between patient and protein-network edges are hyperparameters, tuned on held-out cancer types. The whole thing runs as sparse matrix exponentials: Krylov methods for $\gamma = 0$, and an unraveled stochastic simulation or vectorized master-equation solver on reduced subnetworks for $\gamma > 0$. This is feasible for networks of about 20,000 genes on a single GPU node.
The hardware component is deliberately small. We take patient-level subnetworks of 4 to 8 genes, encode the walk in a few qubits with Trotterized evolution, and run it on superconducting hardware of the kind used in Quantum Ensembling Methods for Healthcare and Life Science (ibm_kyiv, with Pauli twirling and dynamical decoupling). Hardware-inspired zero-noise extrapolation permutes logical-to-physical qubit assignments to vary a calibration-derived "circuit error sum" at constant depth, then extrapolates to zero error. We reuse the permutation but not the extrapolation. The permutations serve as a noise dial: we fit an effective $\gamma$ to each run's measured occupation profile and ask whether $\gamma_{\text{eff}}$ is roughly linear in the circuit error sum. If it is, hardware noise can be steered to a chosen point on the dephasing axis rather than only removed. That paper also warns that routing overhead (extra SWAP gates) can cancel gains. We will therefore restrict permutations to connectivity-preserving ones. This component tests whether "noise as a regularizer," which was conjectured in How quantum computing can enhance biomarker discovery and proved for adversarial robustness in Quantum noise protects quantum classifiers against adversaries, corresponds to a controllable physical quantity. It makes no claim of computational speedup.
Baselines
The naive reference point ranks genes by mutation frequency within each cohort. By construction it should do poorly on genes below 2% recurrence, and it sets the floor that every propagation method must clear.
The strongest cohort-level competitor is DriverRWH: classical random walk with restart on the same hypergraph, with the same STRING or HumanNet degrees as node weights. It reports recovering about 50% of known drivers in the top 30 ranked genes for more than half of 31 TCGA cancer types (9,183 tumors), and it remains stable under 50% data perturbation. The strongest personalized competitor is PDRWH, which reports top-ranked-gene precision of 88.4% in breast cancer and 76.0% in kidney cancer across 16 TCGA types. We will rerun both from their published descriptions on identical mutation calls, networks, and gold-standard driver lists. Each gets the same hyperparameter search budget (the same number of configurations over restart probability or its analogues, tuned on the same held-out cancer types) as our $\gamma$ and $T$.
The pure quantum walk of Dubovitskii et al., ported to our hypergraph, is the $\gamma = 0$ endpoint and the direct test of whether dephasing helps. Two ablations isolate what is new. The first is a phase-randomized walk: identical to our method except that $H$ is replaced by $H$ with random signs on edges, averaged over draws. It keeps the magnitudes of the dynamics but breaks coherent interference. The second is a degree-normalized classical diffusion kernel, $e^{-Lt}$ with $L$ a symmetrically normalized Laplacian, tuned over $t$. This tests whether any quantum-walk gain is simply an anti-hub reweighting.
Beating the baselines means the following. On genes below 2% recurrence, the method must show a statistically significant gain in known-driver recall in the top 30 and in AP@20 over DriverRWH and the diffusion kernel. Significance is judged by a paired test across cancer types. The gain must also hold on types not used for tuning.
Why not just…?
Why not just use a classical random walk with a tuned restart probability, as DriverRWH does? This is the default, and it is strong. The only evidence that quantum walks beat it comes from germline disease seeds on pairwise graphs, where AP@20 rose from 0.14 to 0.30 in Dubovitskii et al. That evidence may not transfer to somatic hypergraphs. Our sweep includes the classical limit as $\gamma \to \infty$, and DriverRWH is run at equal tuning budget. If the best $\gamma$ is effectively infinite, we will report that the classical walk suffices.
Why not just use a degree-normalized heat kernel or other spectral filter? This is the most serious threat. Much of what makes a quantum walk avoid hubs could come from its effective reweighting of the graph's spectrum rather than from interference. A well-chosen classical spectral filter might then match it. The phase-randomized ablation and the normalized diffusion kernel are designed to settle this. If either matches the dephased walk on rare-driver recall, the interference story is wrong for this problem and we will say so.
Why not just run the quantum walk on hardware and apply zero-noise extrapolation? At genome scale, the walk is cheap to simulate classically, so hardware adds cost without any benefit to ranking. Extrapolating to zero noise also assumes the noiseless walk is the target, which is exactly what the $\gamma$ sweep tests. Hardware is confined to small subnetworks, where its role is to test whether device noise maps onto $\gamma$.
Why not just train a deep mutation-rate model such as Dig? Dig models local background mutation rates. That is complementary information: a gene can be mutated above background and still sit poorly in the network. We will report whether combining Dig-style background scores with the walk score adds recall. We do not treat Dig as a direct substitute for network propagation.
Proposed Experiments
The primary experiment sweeps $\gamma$ over about six orders of magnitude. It uses the 31 TCGA cancer types and 9,183 samples from DriverRWH and the 16 types with paired expression data from PDRWH. The gold standard is the known-driver lists used by those papers. Evaluation is stratified by recurrence: under 2%, 2–5%, and above 5%. Metrics are recall in the top 30, AP@20, and top-1 precision per patient for the personalized variant. Cancer types are split into tuning and test sets, with $\gamma$ selected on tuning types only.
The decisive check is the shape of the curve of rare-driver AP@20 against $\gamma$ on test types. An interior maximum that beats both endpoints, DriverRWH, and the phase-randomized ablation supports the claim. A curve that rises monotonically toward either endpoint refutes it.
A secondary check reruns everything on DriverRWH-style perturbed data, with 50% of mutations or edges removed. It asks whether the optimal $\gamma$ shifts upward as data get noisier, which is the regularization prediction.
The hardware experiment runs 50 qubit permutations per subnetwork, the pool size Uvarov et al. found sufficient for their fits. For each permutation we fit $\gamma_{\text{eff}}$ to measured occupations and report its correlation with the circuit error sum. For benchmarks such as the germline seed sets or the brown-adipose cell-interaction graph, the file gives no reported dephasing-versus-accuracy results; there is no reported result found on Lacuna for noisy or dephased quantum walks on any biological benchmark, so these measurements would be the first.
References
- On Quantum Random Walks in Biomolecular Networks
- DriverRWH: discovering cancer driver genes by random walk on a gene mutation hypergraph
- A novel hypergraph model for identifying and prioritizing personalized drivers in cancer
- Quantum Ensembling Methods for Healthcare and Life Science
- Mitigating quantum gate errors for variational eigensolvers using hardware-inspired zero-noise extrapolation
- How quantum computing can enhance biomarker discovery
- Quantum noise protects quantum classifiers against adversaries
- Genome-wide mapping of somatic mutation rates uncovers drivers of cancer