Entropy quantum computing for fixed-backbone protein design Authors: Babak Emami, Wesley Dyk, David Haycraft, Jenn Robinson, Lac Nguyen, Mohammad-Ali Miri, David J. Huggins Published: 2026 ABSTRACT Computational protein design (CPD) is a central problem in biotechnology, with applications in enzyme engineering and therapeutic design, but its combinatorial complexity poses a significant challenge for classical optimization methods. In this work, we formulate fixed–backbone CPD as a quadratic Hamiltonian over rotamer variables, enabling solution using Quantum Computing Inc.’s photonic entropy computing platform, Dirac-3. We evaluate solution quality by benchmarking against an exact classical cost function network (CFN) solver, which provides provably optimal baselines. On a set of standard benchmark proteins ranging from 493 to 943 variables, Dirac-3 produces best-observed solutions (over 100 samples per instance) within 0.16–2.47% of the optimal energies. These results show that the proposed formulation can identify low-energy configurations on directly solvable CPD instances, while sample-level energy distributions vary by instance. Runtime behavior is reported over the tested regime, where CFN remains faster in absolute terms, while Dirac-3 exhibits moderate growth in runtime with problem size. This study focuses on optimization performance as measured by energy relative to exact baselines under a pairwise fixed–backbone energy model. Exploratory experiments on larger instances using decomposition-based approaches are provided in the Supplementary Material. Overall, the results establish a benchmark for entropy-based optimization on CPD formulations within the directly solvable regime. Introduction Proteins are fundamental to biological systems, governing functions such as catalysis, cellular structure, and molecular recognition. The ability to design novel proteins with specified functions therefore has broad applications in drug discovery, synthetic biology, biomaterials, and enzyme engineering1–5. Fixed-backbone Computational Protein Design (CPD) considers the problem of selecting amino acid identities and side-chain conformations that minimize an energy model for a given backbone structure. However, this task suffers from severe combinatorial complexity: for a protein with N positions and multiple rotamer choices per position, the number of possible sequence–conformation combinations grows exponentially. It has been shown that fixed-backbone CPD is NP-hard6. Classical approaches have made significant progress in rendering CPD tractable for moderate-sized systems. Traoré et al. 7 formulated fixed-backbone CPD as a Cost Function Network (CFN) and solved it exactly using the toulbar2 solver, recovering Global Minimum-Energy Conformations (GMECs) for a range of benchmark proteins. Such exact methods provide rigorous baselines for solution quality and are highly efficient on many instances, although their computational cost can increase substantially with problem size, rotamer library expansion, and interaction complexity. In parallel, quantum and quantum-inspired optimization approaches8 are being explored for biomolecular problems including protein folding, docking, and molecular design9. Prior work includes QUBO formulations for peptide–protein docking10,11, variational quantum algorithms for coarse-grained folding12, and hybrid quantum-classical pipelines for drug discovery13. While these approaches are promising, their performance relative to strong classical baselines on standard CPD formulations remains an open question. In this work, we formulate fixed-backbone CPD as a quadratic Hamiltonian over rotamer variables and solve it using Quantum Computing Inc.’s photonic entropy computing platform, Dirac-3. We evaluate solution quality by benchmarking against an exact CFN baseline on a set of standard protein design instances. For directly solvable problems ranging from 493 to 943 variables, Dirac-3 produces best-observed solutions, over 100 samples per instance, within 0.16–2.47% of the optimal energies. These results demonstrate that the proposed formulation can identify low-energy configurations on established CPD benchmarks. This study focuses on optimization performance as measured by objective energy relative to exact baselines under a pairwise fixed-backbone energy model. Exploratory experiments on larger instances using decomposition-based approaches are provided in the Supplementary Material and are not part of the main benchmark analysis. Accordingly, the results presented here establish an optimization benchmark for entropy-based methods on directly solvable CPD instances rather than a comprehensive evaluation of large-scale protein design performance. Background and related work The CPD problem has been studied for more than three decades, leading to the development of multiple algorithmic frameworks and software platforms14–16. These methods often focus on the fixed-backbone formulation, where the polypeptide backbone coordinates are held constant and the side-chain rotamer variables are optimized to minimize an empirical force field. Despite its combinatorial complexity, fixed-backbone CPD has yielded biologically validated designs ranging from enzymes to binding proteins17,18. In this work, we adopt the standard fixed-backbone CPD formulation based on pairwise decomposable energy functions over rotamer variables. This representation underlies many established design frameworks and enables comparison with exact optimization baselines such as CFN. However, it is an approximation of full protein energetics, as it does not explicitly capture effects such as backbone flexibility, many-body polarization, or broader environment-dependent interactions. Accordingly, the present study evaluates optimization performance within this commonly used modeling framework, rather than attempting to represent all aspects of protein physics. Computational approaches to molecular design often utilize a funnel where designs from a more-approximate model that can search more sequence/conformational space (such as using a fixed-backbone with a rotamer library) can be advanced to more-complex models such as MD and or free-energy calculations in order to make more accurate predictions. Proteins are made from amino acids connected by peptide bonds and can be defined by their amino-acid sequence. The connected amino acids form a chain, termed the protein backbone, and different amino acids form branches with different protein sidechains. The backbone and sidechain of one amino acid is termed a residue (Fig. 1a).Figure 1Schematic overview of the fixed-backbone protein design setup. (a) Illustration of protein structure with backbone atoms boxed in magenta, a residue boxed in green, a sidechain boxed in cyan, and the sequence boxed in orange. (b) Folded protein structure, with the backbone shown as a magenta ribbon and sidechains shown as cyan sticks. Five selected sidechains are shown in space-filling representation to highlight sidechain packing. (c) Schematic illustration of folding-energy components. The unfolded state has no pairwise rotamer–rotamer interaction energies, while the folded state includes single-rotamer energies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{i,r}$$\end{document} and pairwise interaction energies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{i,r,j,s}$$\end{document}. (d) Example energy matrix E for two residues and a solution vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}$$\end{document}. Solving for the GMEC, and for the optimal sequence in CPD, is equivalent to minimizing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x} E \textbf{x}^T$$\end{document} subject to selecting one rotamer per residue. The figure is intended as a schematic illustration and has been checked for readability at journal print size. Schematic overview of the fixed-backbone protein design setup. (a) Illustration of protein structure with backbone atoms boxed in magenta, a residue boxed in green, a sidechain boxed in cyan, and the sequence boxed in orange. (b) Folded protein structure, with the backbone shown as a magenta ribbon and sidechains shown as cyan sticks. Five selected sidechains are shown in space-filling representation to highlight sidechain packing. (c) Schematic illustration of folding-energy components. The unfolded state has no pairwise rotamer–rotamer interaction energies, while the folded state includes single-rotamer energies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{i,r}$$\end{document} and pairwise interaction energies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{i,r,j,s}$$\end{document}. (d) Example energy matrix E for two residues and a solution vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}$$\end{document}. Solving for the GMEC, and for the optimal sequence in CPD, is equivalent to minimizing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x} E \textbf{x}^T$$\end{document} subject to selecting one rotamer per residue. The figure is intended as a schematic illustration and has been checked for readability at journal print size. For many proteins, the backbone folds into a distinct conformation such that the sidechains pack together to form intramolecular interactions (Fig. 1b). Importantly, because proteins can be formed of hundreds or even thousands of residues, the 3D fold-space of a protein can be huge19. Determining the fold of a protein from its sequence is thus a very challenging problem20. Numerous approaches such as molecular dynamics have been applied to this problem, with success when applied to small proteins21. However, more recently, machine-learning approaches such as Alphafold22, Boltz23, and Chai24 have revolutionized the field of protein folding. The fixed-backbone approach using a rotamer library In this work, we solve two problems directly related to protein folding: (1) identifying the global minimum energy conformation of a protein with a known fold and (2) identifying the amino acid sequence that minimizes the folding energy for a protein with a known fold. This approach has been studied many times previously5, typically using a pairwise additive forcefield25. For the purposes of solving this problem, the protein backbone is treated as being fixed and the sidechains for each residue can adopt a number of distinct conformations. These conformations are defined by a rotamer library and in this work we consider the widely used Dunbrack-Karplus rotamer library26. Each of the 20 natural amino acids is represented by a one-letter code or a three-letter code, and each can have multiple rotamers, which affect the pairwise interaction energies used to calculate the GMEC. These one- and three-letter codes and the rotamer numbers are included in the Supplementary Material (see Table S1). In fixed-backbone CPD, the protein backbone and any fixed sidechains have a defined energy (Efixed), each rotamer has a different interaction energy with the fixed protein backbone (Ei), and each pair of rotamers has a different pairwise interaction energy (Eij). The CPD problem is then suitable for optimization in this form, but as above it can lead to very large search spaces. As an example of the complexity: a five-residue protein with the sequence LSHVF from Fig. 1a would have 6561 possible combinations of rotamers using the Dunbrack-Karplus rotamer library. If one expands this to a CPD problem and allows each of the 20 natural amino acids at each position, there are now 27,675,731,591,168 possible combinations of rotamers. We aim to calculate (i) the folding energy of a protein to its GMEC and (ii) for CPD the sequence yielding the most favorable folding energy. The folding energy is equal to the difference between the energy of the unfolded state and the folded state. For the unfolded state, the optimal (lowest energy) rotamer is selected at each position. For the non-fixed portions of the system, the energies for all rotamers at all positions can be formed into an energy matrix E for both the folded and unfolded states (Fig. 1c). Any solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}$$\end{document} can be represented as a binary column vector specifying which rotamer is selected for each residue (Fig. 1c). The GMEC can be found by selecting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}$$\end{document} to minimize \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x} E \textbf{x}^T$$\end{document} (Fig. 1d) while respecting the constraint of selecting only one rotamer per residue. In this form, the problem is equivalent to the multiple-choice quadratic knapsack problem27–29 and has similarities to the Potts model as well as quadratic linearly constrained binary optimization (QLCBO). We aim to solve this problem using Dirac-3. The description above defines a discrete optimization problem in which exactly one rotamer is selected at each residue position. In the following sections, we introduce a continuous relaxation of this problem in which selection variables are allowed to take values in the interval [0, 1], with additional penalty terms used to encourage discrete, one-hot solutions. This relaxation enables a natural mapping to a quadratic Hamiltonian suitable for entropy-based optimization on Dirac-3. Classical and quantum optimization approaches Early CPD efforts relied on heuristic search strategies such as Monte Carlo sampling30–32 and dead-end elimination (DEE)33–37, which systematically prune rotamers incompatible with global optima38,39. These approaches, combined with stochastic minimization, powered widely adopted tools such as Rosetta40. More recently, exact optimization approaches have been developed41. Notably, Traoré et al.7 introduced a Cost Function Network (CFN) encoding of CPD and solved it exactly using the open-source solver toulbar2. Their results demonstrated that for mid-sized proteins, CFN can recover the GMEC efficiently, often within seconds42. Other classical frameworks include integer linear programming formulations43–45, branch-and-bound search, and graph-based decomposition techniques46,47. These methods achieve high accuracy but scale poorly with protein size and rotamer library expansion. Thus, while state-of-the-art classical solvers represent a rigorous baseline, the exponential scaling remains a fundamental limitation. Along different lines, quantum and quantum-inspired methods have recently been proposed as alternatives for tackling CPD and related biomolecular problems. Perdomo-Ortiz et al.48 demonstrated that lattice-based protein folding models can be mapped to quadratic unconstrained binary optimization (QUBO) and solved using quantum annealers. Babbush et al.49 further developed adiabatic quantum simulations for protein folding, highlighting the potential of Hamiltonian-based encodings. More recently, variational quantum algorithms (VQAs) have been proposed for coarse-grained folding and docking tasks, exploiting hybrid classical–quantum pipelines to balance expressivity and near-term device constraints12,13. These studies suggest that quantum optimization devices may provide an advantage in navigating the high-dimensional energy landscapes of biomolecules, though large-scale demonstrations remain elusive. Entropy computing was recently introduced as a quantum optimization paradigm that minimizes Hamiltonians within an open photonic quantum system that utilizes photon number state encoding and Fock basis readouts50. While such a paradigm can enable quantum-parallel computation51, its intermediate hybrid implementation through electronic measurement-and-feedback demonstrated efficient optimization with all-to-all graph connectivity and a high dynamic range50,52. This allows the native encoding of dense, multi-scale protein interaction graphs without the need for rescaling or complex embedding. The Quantum Computing Inc.’s entropy computing device, Dirac-3, is described in greater details in the Supplementary Material. Results Hardware parameters and hyperparameter tuning We first formulated the problem for use on Dirac-3, as described in the Methods section below. We then explored the effect of hardware parameters and hyperparameters. The goal was to identify robust operating regimes that retain energy quality while minimizing runtime. In what follows, the results of a thorough hyperparameter study on the benchmark problem 1MJC is presented. The mean photon number is an important hardware-level parameter on Dirac-3 that controls the optical intensity used to encode Hamiltonian coefficients within the photonic subsystem. Intuitively, it sets the average number of photons present in the system during computation, which in turn determines the signal-to-noise ratio and the precision with which energy differences can be resolved. In general, low mean photon numbers yield high-fidelity energy landscapes but longer runtimes due to weaker optical signals, whereas higher mean photon numbers accelerate convergence by strengthening signal amplitudes but risk saturating the device or introducing noise if pushed too far. Relaxation schedule is another important hardware parameter. In the context of QCi’s Dirac-3 device, relaxation schedule refers to the sequence of entropy-driven updates applied during the minimization process, effectively controlling how gradually the system relaxes from an initial high-entropy state to a low-entropy configuration that encodes a solution. In general, shallow schedules (few steps) yield faster convergence but risk trapping in local minima, while deeper schedules (more steps) provide improved energy accuracy at the cost of longer runtimes. The sample solution energies and their distributions at various values of relaxation schedule and mean photon number are shown in Fig. 2(a) and (b), respectively. The corresponding runtimes and solution energies are shown in Tables S3 and S4 of the Supplementary Material. We observe that, for this benchmark instance, a mean photon number of 0.003 provides a favorable trade-off between solution quality and runtime; using smaller values increases runtime without observable energy improvement. We also observe that increasing the relaxation schedule depth from 1 to 4 increases runtime without substantial energy improvement beyond schedule 2 for this instance.Figure 2Impact of (a) mean photon number and (b) relaxation schedule on Dirac-3 solution energy for the 1MJC protein. Energies are reported in kcal/mol. Unless varied on the corresponding axis, parameters are fixed as listed in Table 1. Impact of (a) mean photon number and (b) relaxation schedule on Dirac-3 solution energy for the 1MJC protein. Energies are reported in kcal/mol. Unless varied on the corresponding axis, parameters are fixed as listed in Table 1. Next, we investigated the impact of the dynamic range on the solution quality. The dynamic range, defined as the ratio between the largest and smallest coefficients in the Hamiltonian (expressed in decibels(dB)), significantly influences numerical stability and optimization difficulty. Notably, Dirac-3 has a known performance limitation at 23 dB, making it particularly relevant to evaluate how it performs when handling Hamiltonians that exceed this range. Table S2 of the Supplementary Material reports problem sizes and coefficient dynamic ranges for all of the proteins studies in this paper. The GMEC is known for each of these test cases, allowing effective benchmarking. It is generally expected that pruning very large and very small coefficients helps reduce dynamic range, but aggressive pruning risks losing important interactions. The results are summarized in Tables S5 and S6 of the Supplementary Material. We observe that values of max energy threshold in the 5–10 range yield better solution energies by altering the Hamiltonian just enough to keep the dynamic range manageable; pushing to 100 degrades solution energy due to larger dynamic range values. In addition, values of min energy threshold around \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-4}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-3}$$\end{document} are conservative and stable; while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-1}$$\end{document} happens to give a slightly lower energy here, that behavior is instance-specific and risks over-pruning. We have also studied the impact of penalty multipliers alpha and beta (α and β in Eq. 1) as shown in Tables S7 and S8 of the Supplementary Material. Similar hyperparameter studies on the other small-sized proteins studied here (1CSK, 1SHF, 1SHG, 1NXB, 1TEN, and 1POH) yielded similar results. Exploratory experiments on larger instances using decomposition-based approaches are provided in the Supplementary Material and are not part of the main benchmark analysis. Whilst there is some system-dependence on the performance based on hyperparameter selection, we selected a consistent set of hyperparameters based on this analysis to generate the results below (Table 1).Table 1Hyperparameters selected to generate final results. Energy thresholds are in kcal/mol.HyperparameterValueMean photon number0.003Relaxation schedule2Max energy threshold10.0Min energy threshold1.0e−3Alpha20.0Beta0 Table 1Hyperparameters selected to generate final results. Energy thresholds are in kcal/mol.HyperparameterValueMean photon number0.003Relaxation schedule2Max energy threshold10.0Min energy threshold1.0e−3Alpha20.0Beta0 Hyperparameters selected to generate final results. Energy thresholds are in kcal/mol. Optimization results We evaluate the proposed Dirac-3 workflow on a canonical fixed-backbone CPD benchmark suite of directly solvable proteins: 1MJC, 1CSK, 1SHF, 1SHG, 1NXB, 1TEN, and 1POH. We compare against the GMEC from the classical CFN baseline implemented in toulbar2, a strong reference for weighted-constraint encodings of CPD7,42,53. Figure 3 shows the solution energies for small protein problems, sorted by problem size, obtained by Dirac-3 and in comparison with the known optimal solutions from CFN 7. For each problem, the results were collected from 100 runs using the hyperparameters in Table 1.Figure 3Energy distribution across 100 Dirac-3 samples for each directly solvable protein benchmark, compared with the optimal CFN energy 7. Dirac-3 samples were generated using the hyperparameters in Table 1. Energies are reported in kcal/mol. Percentage gaps referenced in the text are computed relative to the CFN optimum as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(E_{\textrm{D3}} - E_{\textrm{CFN}})/|E_{\textrm{CFN}}|$$\end{document}, such that positive values indicate higher (worse) energies than the optimum. Energy distribution across 100 Dirac-3 samples for each directly solvable protein benchmark, compared with the optimal CFN energy 7. Dirac-3 samples were generated using the hyperparameters in Table 1. Energies are reported in kcal/mol. Percentage gaps referenced in the text are computed relative to the CFN optimum as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(E_{\textrm{D3}} - E_{\textrm{CFN}})/|E_{\textrm{CFN}}|$$\end{document}, such that positive values indicate higher (worse) energies than the optimum. The best-observed Dirac-3 solutions are within 0.16–2.47% of the GMEC, with a mean absolute best-sample gap of 1.21% and median best-sample gap of 1.08%. The discrepancy between Dirac-3 and CFN energies can be seen explicitly in Table 2.Table 2Energy and runtime comparison against the CFN baseline for directly solvable benchmark proteins. Dirac-3 energy statistics are computed over 100 samples per protein and reported in kcal/mol. The best Dirac-3 energy is the lowest-energy sample observed among the 100 samples.ProblemSizeAvg. runtime (s)CFN runtime (s)Dirac-3 Energy Mean ± SDBest Dirac-3 EnergyCFN energy1MJC4934.230.08-66.97 ± 3.71-70.45-70.711CSK6163.540.15-48.75 ± 5.12-55.41-55.501SHF6384.380.17-54.68 ± 3.46-60.07-61.601SHG7379.380.25-61.07 ± 6.53-65.13-65.561NXB8003.860.24-31.30 ± 6.23-37.35-38.271TEN8086.110.33-81.27 ± 11.21-91.45-92.451POH94317.450.45-87.17 ± 17.90-97.53-98.78Table 3Additional distributional statistics for Dirac-3 energies across 100 samples per directly solvable protein benchmark. IQR denotes the interquartile range of sampled energies. Hit rates report the percentage of samples whose energy gap relative to the CFN optimum is at most 2%, 3%, or 5%, respectively. Energies are reported in kcal/mol.ProblemMedian Dirac-3 EnergyIQRHit łe2%Hit łe3%Hit łe5%1MJC-68.531.6815.0%41.0%78.0%1CSK-49.994.012.0%2.0%12.0%1SHF-55.765.150.0%3.0%9.0%1SHG-63.053.1322.0%47.0%55.0%1NXB-33.677.380.0%2.0%13.0%1TEN-86.157.554.0%5.0%29.0%1POH-94.9012.7845.0%46.0%51.0% Table 2Energy and runtime comparison against the CFN baseline for directly solvable benchmark proteins. Dirac-3 energy statistics are computed over 100 samples per protein and reported in kcal/mol. The best Dirac-3 energy is the lowest-energy sample observed among the 100 samples.ProblemSizeAvg. runtime (s)CFN runtime (s)Dirac-3 Energy Mean ± SDBest Dirac-3 EnergyCFN energy1MJC4934.230.08-66.97 ± 3.71-70.45-70.711CSK6163.540.15-48.75 ± 5.12-55.41-55.501SHF6384.380.17-54.68 ± 3.46-60.07-61.601SHG7379.380.25-61.07 ± 6.53-65.13-65.561NXB8003.860.24-31.30 ± 6.23-37.35-38.271TEN8086.110.33-81.27 ± 11.21-91.45-92.451POH94317.450.45-87.17 ± 17.90-97.53-98.78 Energy and runtime comparison against the CFN baseline for directly solvable benchmark proteins. Dirac-3 energy statistics are computed over 100 samples per protein and reported in kcal/mol. The best Dirac-3 energy is the lowest-energy sample observed among the 100 samples. Table 3Additional distributional statistics for Dirac-3 energies across 100 samples per directly solvable protein benchmark. IQR denotes the interquartile range of sampled energies. Hit rates report the percentage of samples whose energy gap relative to the CFN optimum is at most 2%, 3%, or 5%, respectively. Energies are reported in kcal/mol.ProblemMedian Dirac-3 EnergyIQRHit łe2%Hit łe3%Hit łe5%1MJC-68.531.6815.0%41.0%78.0%1CSK-49.994.012.0%2.0%12.0%1SHF-55.765.150.0%3.0%9.0%1SHG-63.053.1322.0%47.0%55.0%1NXB-33.677.380.0%2.0%13.0%1TEN-86.157.554.0%5.0%29.0%1POH-94.9012.7845.0%46.0%51.0% Additional distributional statistics for Dirac-3 energies across 100 samples per directly solvable protein benchmark. IQR denotes the interquartile range of sampled energies. Hit rates report the percentage of samples whose energy gap relative to the CFN optimum is at most 2%, 3%, or 5%, respectively. Energies are reported in kcal/mol. In addition to the best-observed energies, Tables 2 and 3 report distributional statistics across the 100 Dirac-3 samples for each benchmark instance. The best-observed Dirac-3 solutions are within 0.16–2.47% of the GMEC, but the sample distributions vary substantially by instance. Hit rates within 2%, 3%, and 5% of the CFN optimum are therefore reported to characterize the frequency with which near-optimal samples are obtained. In several instances, the energy distributions exhibit substantial spread, indicating variability in solution quality across samples despite strong best-observed performance. Although the best-observed Dirac-3 energies are close to the CFN energies in absolute energy units for several instances, these numerical energy gaps should not be interpreted as evidence of equivalent biological or structural outcomes. For the larger systems, we apply the graph-partitioning framework described in the Supplementary Methods Section, solving sub-problems on Dirac-3. Results can be found in the Supplementary Material. Whilst these results illustrate the behavior of the decomposition-based workflow, they should be interpreted as exploratory and not as a direct assessment of end-to-end optimization performance. Runtime comparison and scaling behavior The reported Dirac-3 runtimes correspond to the average wall-clock time per sample, including Hamiltonian loading, device execution, and readout. Preprocessing steps such as Hamiltonian construction and coefficient pruning are not included. Similarly, CFN runtimes correspond to solver execution time as reported by toulbar2, excluding preprocessing. Figure 4 compares wall-clock runtimes for the classical exact CFN solver and the Dirac-3 workflow as a function of problem size (number of rotamer variables). The plot includes observed Dirac-3 runtimes over the directly solvable regime (up to 943 variables). Over this range, Dirac-3 exhibits moderate growth in runtime with problem size, consistent with behavior dominated by device-level overheads and fixed schedule costs rather than strongly size-dependent combinatorial search. Within the directly solvable regime, CFN remains faster in absolute runtime than Dirac-3. However, CFN runtimes increase sharply across the tested instances, consistent with the observed increase in computational effort required by exact discrete optimization methods on these instances. The comparison presented here is therefore intended to document empirical runtime trends on the benchmark set rather than to establish a demonstrated crossover point. Further evaluation on larger directly solvable instances would be required to assess relative performance beyond the current regime.Figure 4Comparison of runtime scaling between the classical exact CFN solver and the Dirac-3 workflow on protein design problems. CFN runtimes correspond to exact optimization on the full problem, whereas Dirac-3 runtimes include only directly solved instances within the hardware limit. Results obtained using decomposition-based approaches (reported in the Supplementary Material) are not included here and are not directly comparable to the exact CFN baseline. Comparison of runtime scaling between the classical exact CFN solver and the Dirac-3 workflow on protein design problems. CFN runtimes correspond to exact optimization on the full problem, whereas Dirac-3 runtimes include only directly solved instances within the hardware limit. Results obtained using decomposition-based approaches (reported in the Supplementary Material) are not included here and are not directly comparable to the exact CFN baseline. Conclusions and outlook This work demonstrates that entropy-based optimization on Dirac-3 can produce low-energy solutions for fixed-backbone computational protein design when evaluated on standard benchmark instances. For directly solvable problems with known optima, Dirac-3 achieves best-observed solution energies, over 100 samples per instance, within 0.16–2.47%. These results indicate that the proposed Hamiltonian formulation and continuous-variable optimization approach are capable of identifying low-energy regions of the combinatorial energy landscape of fixed-backbone CPD. Importantly, the energy improvements reported here are evaluated within a pairwise fixed-backbone energy model and are not directly validated against experimental observables such as sequence recovery, folding stability, or functional activity. As such, the results should be interpreted as optimization performance within a surrogate energy model rather than as direct indicators of biological design success. The present study focuses on optimization performance as measured by objective energy relative to exact baselines under a pairwise fixed-backbone energy model. It does not assess downstream biological properties such as sequence recovery, structural fidelity, or experimental validation of designed proteins. In addition, while runtime behavior is reported for the directly solvable regime, no crossover with classical methods is established within the range of tested problem sizes. Exploratory experiments on larger instances using decomposition-based approaches are provided in the Supplementary Material; however, these rely on additional heuristic steps and are not directly comparable to the end-to-end formulation studied here. Future work will focus on extending the range of directly solvable instances, improving robustness with respect to hyperparameter selection, and evaluating the relationship between energy optimization and biologically relevant design outcomes. We suggest that the problem of fixed-backbone protein design using a rotamer library is well-suited to quantum-computational approaches and these or other test cases could be used as a standard benchmark in the future.