Hybrid Classical–Quantum Folding

• Hybrid classical–quantum folding approaches are algorithms that minimize energy landscapes by integrating quantum search methods with classical optimization techniques.
• They leverage quantum routines such as variational eigensolvers and quantum annealing to address protein, RNA, circuit, and chemical folding challenges.
• These methods enhance resource efficiency and noise resilience, though scalability is constrained by hardware limitations and increasing computational complexity.

Hybrid classical–quantum folding approaches constitute a rapidly evolving class of algorithms that combine quantum and classical resources in order to address computationally intractable folding problems in physical, chemical, and biological systems. These hybrid schemes leverage the strengths of quantum algorithms—such as superposition-based search or quantum sampling—while delegating tasks like parameter optimization, model construction, and auxiliary computation to classical processors, thereby mitigating the severe qubit and circuit-depth requirements of purely quantum implementations. Applications span protein folding on lattice and off-lattice models, RNA secondary structure prediction (including pseudoknots), quantum circuit partitioning for resource-efficient workload management, and quantum chemistry simulations via Hamiltonian downfolding. The following sections provide an in-depth account of the main methodologies, mathematical foundations, representative workflows, algorithmic variants, performance metrics, and scaling behavior of these hybrid folding paradigms.

1. Mathematical Foundations of Hybrid Folding

Hybrid classical–quantum folding typically involves expressing the folding problem as an energy-minimization over a high-dimensional configuration space, rendered as a suitable cost Hamiltonian $H$ or quadratic unconstrained binary optimization (QUBO) objective whose ground state encodes the folded state.

In lattice protein folding, e.g., (Robert et al., 2019, Chandarana et al., 2022, Irbäck et al., 2022), the chain's conformational degrees of freedom are encoded via binary variables or qubits. For a tetrahedral-lattice protein, the full Hamiltonian decomposes into geometric/turn constraints, chirality, and non-local interaction terms, each mapped to local and few-body terms in the Pauli-$Z$ basis:

$$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$

where $q_{\rm cf}$ encodes the chain's geometry and $q_{\rm in}$ encodes pairwise contacts.

RNA secondary structure prediction leverages a QUBO modeling framework, where each candidate base pair or stem is assigned a binary variable $q_i\in\{0,1\}$, and the global structure cost adopts the canonical form

$$H_{\rm QUBO}(q) = \sum_i h_i q_i + \sum_{i<j} J_{ij} q_i q_j,$$

with $h_i$ representing per-stem stability and $J_{ij}$ penalizing overlaps/pseudoknots (Zaborniak et al., 2022). Parameterization can be tuned to various physical and thermodynamic models.

In quantum chemistry, orbital-space partitioning is performed, followed by coupled-cluster-based downfolding of dynamical correlation into an active subspace, yielding an effective Hamiltonian $H_{\rm eff}$ for further quantum processing (Bauman et al., 1 Jul 2025).

Hybrid circuit folding, in the context of circuit partitioning and workload reduction, operates on the graph-structured representation of quantum circuits, employing module discovery and meta-graph construction to guide “folding” of repeated circuit regions (Kan et al., 2024).

2. Core Hybrid Algorithms and Workflow Structure

The archetypal hybrid folding workflow comprises a repeated, iterative loop in which classical and quantum subcomponents interact in a tightly coupled optimization or sampling scheme.

Variational Quantum Folding (protein/polymer, e.g., (Robert et al., 2019, Chandarana et al., 2022)):

1. Hamiltonian Construction: Classical preprocessing computes lattice/geometry and energetic parameters, assembling the cost Hamiltonian $Z$0.
2. Quantum Phase: A parameterized quantum circuit (e.g., CVaR-VQE, CD-inspired ansatz) prepares variational states; measurements yield cost-function samples for basis states.
3. Classical Optimization: Parameters are updated via differential evolution, gradient descent, or Adam-like optimizers based on estimates of a cost function (e.g., $Z$1).
4. Termination: Loop continues until energy convergence or iteration budget is met.

Hybrid Quantum Annealing (RNA folding, HP folding, (Zaborniak et al., 2022, Irbäck et al., 2022)):

1. QUBO/Ising Mapping: Problem is encoded as a quadratic form in binary variables, suitable for submission to quantum annealers.
2. Hybrid Oracle: A cloud- or middleware-based hybrid solver partitions the problem for both quantum annealing and local-search classical updates.
3. Parameter Training: For RNA, classical SPSA (simultaneous perturbation stochastic approximation) is used to optimize Hamiltonian parameters against annotated structure datasets.
4. Iterative Anneal/Local Search: Candidate solutions are iteratively improved, with quantum subcalls embedded within classical heuristics.

Quantum Walk/Deep Learning Hybrid (off-lattice folding, (Casares et al., 2021)):

1. Classical Initializer: Neural network (AlphaFold-style) produces per-residue angular priors, focusing search on high-probability regions.
2. Quantum Subroutine: A coined Szegedy quantum walk, applying projective Metropolis acceptance in a quantum register, explores the conformational space.
3. Classical Evaluation/Monitoring: Energy landscape and acceptance rates are monitored, with hybrid transitions between neural network, quantum walk, and classical update.

Circuit Folding and Meta-Graph Partitioning (Kan et al., 2024):

1. Module Discovery: Automated identification and folding of repeated low-level subcircuits (subgraphs).
2. Meta-Graph Construction: Each module is treated as a meta-node, with weighted edges representing cut points and resource linkage.
3. Partition Optimization: Greedy and local-refinement heuristics assign subcircuits to available quantum hardware, minimizing

$Z$2

with $Z$3 quantum resource usage and $Z$4 classical communication/overhead.

1. Integration with State-of-the-Art Circuit Knitting: Use of quasi-probability decompositions and wire-cut/ancilla-free protocols in cut evaluation.

3. Representative Folding Models Across Domains

Folding Domain	Hybrid Model/Encoding	Quantum Subroutine
Protein (lattice)	Tetrahedral/square lattice encoding	VQE/CVaR, digitized-CD, QA
RNA (secondary)	QUBO models (stacked stems, polymer terms)	Hybrid quantum annealing
Quantum chemistry	CC downfolded Hamiltonian (active space)	VQE, ADAPT-VQE, GCIM
Circuit partitioning	Module meta-graph (DAG folding)	Circuit knitting, QPD

Protein folding approaches (Robert et al., 2019, Chandarana et al., 2022, Casares et al., 2021) focus on encoding the full suite of lattice/geometric/chemical constraints directly in the quantum Hamiltonian, thus enabling either variational quantum eigensolver (VQE) or counter-diabatic circuit approaches. RNA folding (Zaborniak et al., 2022) parameterizes loop penalties, stem energies, and pseudoknot interactions within the QUBO framework, with hybrid annealing and SPSA training. Chemical systems deploy similarity-transformed coupled cluster frameworks for downfolding, then apply quantum active-space solvers, e.g., ADAPT-VQE (Bauman et al., 1 Jul 2025). Circuit workload folding (Kan et al., 2024) exploits structural repetition at the gate/module level to minimize both hardware and classical resource overhead.

4. Performance, Resource Scaling, and Hardware Results

Hybrid folding algorithms demonstrate substantial resource savings and improved scalability relative to purely classical or quantum-only options but are strongly limited by current quantum hardware capabilities and model mapping complexity.

• For 10-residue Angiotensin peptide, variational approaches required 22 qubits and $Z$5 Pauli terms; convergence to the exact ground state achieved in $Z$6 generations with up to 1024 shots per generation. Population ground-state probabilities reached $Z$7 at the best (Robert et al., 2019).
• QUBO/Ising-based hybrid annealing for lattice HP chains achieved exact ground states (100% hit rate) for $Z$8, and lowest-known energies for $Z$9 using D-Wave Hybrid with tailored penalty strengths and run-time budgets. Pure quantum annealing alone folded up to $$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$0 at exponentially decaying hit rates (Irbäck et al., 2022).
• Circuit-folding resource overheads decreased by up to 799.2% on large benchmark circuits (e.g., 190-qubit adder) under tight qubit constraints by meta-module folding, compared to wire-cutting baselines. Cut-finding times remained under 10 s (Kan et al., 2024).
• RNA hybrid quantum annealing yielded median Matthews correlation coefficients of $$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$1–$$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$2 for 40-nucleotide test sequences (with and without pseudoknots), with no significant model-dependent differences (post-training) (Zaborniak et al., 2022).
• Hybrid downfolding in quantum chemistry enabled active-space quantum solves that captured $$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$3–$$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$4 of correlation energy for molecules like benzene or porphyrin using 12–14 qubits and shallow adaptive circuits. Hardware results on Quantinuum H1-1 reached chemical accuracy (mean error $$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$520 m$$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$6) after error mitigation (Bauman et al., 1 Jul 2025).

5. Algorithmic Advantages, Limitations, and Hybrid Synergy

Hybrid folding frameworks offer several advantages:

• Quantum acceleration (e.g., polynomial speedup in mixing/optimization) is realized by confining quantum effort to the intractable subspace or tail behavior (e.g., CVaR, quantum Metropolis), while exploiting classical preprocessing for initialization, penalty estimation, or model reduction.
• Flexibility to tailor problem size to available quantum devices, for example by downfolding high-dimensional Hamiltonians onto smaller active spaces or partitioning circuits based on detected modularity.
• Improved noise resilience and lower measurement overhead; CVaR targets ground-state energies with reduced statistical cost, while hybrid solvers average out hardware fluctuations.
• Integration of advanced resource-mapping, e.g., ring-topology entanglers, ancilla-free wire cutting, and greedy meta-graph partitioning, to minimize both quantum and classical communication costs.

However, all approaches face distinct limitations:

• Scalability is ultimately limited by the brute-force scaling of the Hamiltonian (e.g., $$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$7 Pauli terms in lattice models), QUBO variable proliferation (RNA, $$H(q) = H_{\rm gc}(q_{\rm cf}) + H_{\rm ch}(q_{\rm cf}) + H_{\rm in}(q_{\rm cf}, q_{\rm in}),$$8), and hardware restrictions on gate fidelity and parallelism.
• Substantial preprocessing (stem enumeration, active-space selection) can pose significant classical costs in large systems.
• Optimization landscapes remain susceptible to local minima and barren-plateaus, especially as system size increases.
• In certain architectures (e.g., low-connectivity superconducting circuits), SWAP overhead and circuit depth can significantly degrade performance compared to all-to-all topologies.
• Current hybrid quantum annealers often return single samples per anneal, and stochasticity limits reproducibility for large, hard instances (Zaborniak et al., 2022, Irbäck et al., 2022).

6. Future Directions and Outlook

Continued progress in hybrid classical–quantum folding research is anticipated along several axes:

• Scaling to larger system sizes via hierarchical problem decomposition (multi-active-space, partitioned QUBO modeling) and improved circuit/graph partitioning heuristics.
• Integration of advanced error-mitigation protocols (zero-noise extrapolation, symmetry exploitation, probabilistic error cancellation) to further suppress hardware noise without deep circuits.
• Meta-learning or reinforcement learning for variational ansatz selection, initialization, and gauge-potential component weighting in digitized counterdiabatic methods.
• Exploration of hybrid folding in domains beyond biochemistry and quantum chemistry, e.g., large-scale quantum circuit synthesis, machine learning model partitioning, or other combinatorial optimization tasks requiring interplay between classical sampling and quantum search.
• Systematic parameterization of resource models, enabling automated tailoring of algorithm depth, qubit count, classical/quantum workload fraction, and expected solution quality given hardware specifications.

Hybrid folding brings together disparate advances across quantum information, algorithmic optimization, structural biology, and quantum chemistry to address persistent computational bottlenecks. As quantum hardware evolves toward higher fidelities and logical qubit counts, these hybrid strategies are positioned as a leading route to early quantum advantage in complex folding and energy minimization problems (Robert et al., 2019, Chandarana et al., 2022, Irbäck et al., 2022, Zaborniak et al., 2022, Kan et al., 2024, Casares et al., 2021, Bauman et al., 1 Jul 2025).