The cost of quantum algorithms for biochemistry: A case study in metaphosphate hydrolysis

Abstract: We evaluate the quantum resource requirements for ATP/metaphosphate hydrolysis, one of the most important reactions in all of biology with implications for metabolism, cellular signaling, and cancer therapeutics. In particular, we consider three algorithms for solving the ground state energy estimation problem: the variational quantum eigensolver, quantum Krylov, and quantum phase estimation. By utilizing exact classical simulation, numerical estimation, and analytical bounds, we provide a current and future outlook for using quantum computers to solve impactful biochemical and biological problems. Our results show that variational methods, while being the most heuristic, still require substantially fewer overall resources on quantum hardware, and could feasibly address such problems on current or near-future devices. We include our complete dataset of biomolecular Hamiltonians and code as benchmarks to improve upon with future techniques.

• The paper provides detailed resource estimates for three quantum algorithms—ADAPT-VQE, Krylov, and QPE—applied to metaphosphate hydrolysis, showing that ADAPT-VQE can achieve chemical accuracy with only a few thousand two-qubit gates, potentially feasible with current or near-future hardware. The Krylov and QPE methods, with cost estimates to be bound through different algorithms, are beyond current capabilities.
• Analysis of the quantum resource requirements involves rigorous preprocessing steps, including DUCC truncation and Jordan-Wigner mapping, suggesting that quantum algorithms like ADAPT-VQE are advantageous but cracking troughs make Krylov and QPE methods beyond the currently accessible.
• The study demonstrates that despite classical methods' substantial disagreement, quantum algorithms offer a higher-accuracy treatment of electronic structure problems in biochemistry, highlighting the potential of quantum computing in solving complex biochemical reactions.
• The authors mitigates dependence on obtained results via numerical validation and their benchmark tools thus making estimates reliable and generalizable.

Motivation and scope

ATP hydrolysis is among the most consequential reactions in biology, releasing approximately $-7.3$ kcal/mol per phosphate cleavage and underpinning metabolism, cellular signaling, and a broad class of pharmaceutical targets. Despite decades of computational effort, first-principles prediction of its free energy profile remains unresolved: prior QM/MM studies employing DFT have either significantly under- or overshot the known free energy difference and activation energy, sometimes predicting the reaction to be endothermic overall. This paper quantifies, end to end, what it would cost to solve the electronic structure of this reaction class on quantum computers. The authors take magnesium-coupled metaphosphate hydrolysis (PO$_3^-$ lysing H$_2$O) as a tractable proxy for full ATP hydrolysis — 50 electrons in 78 orbitals in a 6-31G(d) basis — and estimate resources for three ground state energy estimation algorithms spanning three hardware eras: ADAPT-VQE for NISQ devices, quantum Krylov for MegaQuop machines, and single-ancilla iterative quantum phase estimation (QPE) for fault-tolerant application-scale systems.

A notable methodological point is that classical methods disagree substantially on this problem. B3LYP predicts an activation barrier of 18.92 kcal/mol versus 22.95 kcal/mol from CCSD(T), and while B3LYP's reaction energy ($-10.57$ kcal/mol) closely matches the CCSD(T) benchmark ($-10.90$ kcal/mol), CAM-B3LYP predicts a significantly more exothermic value of $-14.77$ kcal/mol. This dispersion across electronic structure theories at the transition state is itself evidence that the problem is quantum-mechanically difficult and motivates higher-accuracy treatment.

Problem setup and preprocessing

All three algorithms target chemical accuracy, $\epsilon = 10^{-3}$ Ha, on effective Hamiltonians produced by DUCC downfolding. The downfolding solves the external orbital space classically via a full-system CCSD calculation, folds dynamical correlation into an effective Hamiltonian via the DUCC(3) truncation (correct through third-order perturbation theory), and projects onto a 32-electron, 22-orbital active space — a 44-qubit problem under Jordan–Wigner mapping. The resulting Hamiltonians contain roughly $5.8 \times 10^5$ Pauli terms each, sorted into about 2,700 commuting groups by greedy sorted insertion.

This preprocessing step is essential to the paper's conclusions: without it, none of the three algorithms would be feasible on near-term hardware. It also introduces an assumption worth noting — the accuracy of the effective Hamiltonian depends on truncated BCH expansions and approximate external cluster amplitudes, so the reported resource counts apply to the downfolded model rather than the full electronic problem.

Resource estimates

The headline results are summarized below (all-to-all connectivity assumed).

System	Qubits	Pauli terms	ADAPT-VQE $n_{2Q}$	Krylov $n_{2Q}$*	QPE $_3^-$0
PO$_3^-$1H$_3^-$2O	44	575,751	6,508	$_3^-$3	$_3^-$4
H$_3^-$5PO$_3^-$6	44	579,542	7,200	$_3^-$7	$_3^-$8

*Estimated lower bounds; †estimated upper bounds.

ADAPT-VQE resources were obtained exactly, by explicitly optimizing CEO-ADAPT ansätze against the downfolded Hamiltonians using a sparse wavefunction circuit simulator. Reaching chemical accuracy required only ~6,500–7,200 two-qubit gates with no ancillas — well within the scale of recent hardware demonstrations. The authors state plainly that important biological questions are therefore within reach of current or near-future devices, though they caveat that variational algorithms offload substantial cost to classical processors, whose exponential scaling remains an open concern even if practical workarounds exist.

Quantum Krylov requires one ancilla qubit and Hadamard-test circuits for matrix elements. Because direct simulation of the metaphosphate systems was impossible, subspace dimension $_3^-$9 and Trotter count $_2$0 were estimated by exact simulation of hydrogen chains followed by extrapolation; the authors take these as lower bounds given hydrogen chains are simpler Hamiltonians. With $_2$1 and $_2$2, the largest circuit comprises roughly $_2$3 two-qubit gates — beyond current hardware but plausibly reachable on MegaQuop-scale machines with algorithmic improvements. A significant weakness of the analytical convergence bound is exposed here: applying the Epperly et al. bound to a six-qubit hydrogen chain yields $_2$4, whereas numerically $_2$5 suffices. Numerical estimation rather than analytical bounding is thus indispensable throughout.

QPE costs are bounded above via a perturbative analysis of Trotter error propagated through phase estimation, yielding $_2$6–$_2$7 Trotter steps and hence $_2$8 two-qubit gates. These figures are consistent with recent independent estimates for molecular Hamiltonians, but they make clear that Trotterized time evolution of these dense, high-term-count Hamiltonians is the dominant bottleneck for both Krylov and QPE — even when compiled with the Paulihedral framework, which the authors show produces markedly fewer two-qubit gates than naive or grouped exponentiation strategies.

Measurement and compilation overheads

For energy estimation, shot counts to reach $_2$9 Ha precision were computed using DMRG-generated matrix product states with increasing bond dimensions until convergence. Converged estimates are roughly $-10.57$0 shots for PO$-10.57$1H$-10.57$2O and $-10.57$3 for H$-10.57$4PO$-10.57$5 — three orders of magnitude below the maximally-mixed-state worst case of $-10.57$6. The authors note that finer-grained commutativity partitioning, non-Clifford diagonalization, ghost/fluid fragments, and low-rank factorization methods could reduce these further, and they release code and data explicitly as benchmarks for such improvements.

Hardware-aware compilation adds nontrivial overhead. Mapping Trotter circuits from all-to-all connectivity to IBM's heavy-hex topology roughly doubles two-qubit gate counts at the ~40-qubit scale, implying at least a factor-of-two penalty for the metaphosphate systems on superconducting hardware. For ADAPT circuits, a comparison of Bqskit, Tket, and Qiskit compilation flows shows that the best compiler for all-to-all targets is not necessarily best for heavy-hex targets, and that the optimal strategy varies with ansatz depth — a concrete argument that resource comparisons must be benchmarked per-hardware rather than assumed.

Limitations and open questions

The paper is explicit about several caveats. Gate counts exclude diagonalization circuits for energy estimation (which scale as $-10.57$7 per commuting group) and initial state preparation for Krylov and QPE, both of which can be substantial since those algorithms require strong ground-state overlap to converge. All resource counts are quantum-only; classical optimization costs for VQE, though empirically manageable at hours-to-days for these system sizes, carry exponential scaling in general. The Krylov and QPE figures rest on extrapolation from hydrogen-chain simulations and perturbative bounds respectively, and the appendix shows that thresholding — likely necessary for numerical stability in practice — increases the required subspace dimension relative to the unthresholded values reported. Error correction overheads are not included despite being certainly required at these circuit scales. Finally, the downfolded Hamiltonian itself embeds approximations whose fidelity to the full biochemical problem is not independently validated here.

Conclusion

This work provides the first detailed quantum resource estimates for metaphosphate hydrolysis, comparing three algorithmic regimes on equal footing with a combination of exact simulation, numerical estimation, and analytical bounds. Its central finding is a stark separation of feasibility: ADAPT-VQE requires only thousands of two-qubit gates on 44 qubits and is plausibly executable today, whereas quantum Krylov needs $-10.57$8 gates and Trotterized QPE $-10.57$9, dominated by the cost of time-evolving Hamiltonians with hundreds of thousands of terms. The released dataset of biomolecular Hamiltonians and reproducible code positions subsequent measurement-reduction, compilation, and time-evolution improvements to be evaluated directly against a biochemically meaningful target.