{"id":"ef3432d1-e6cc-46e5-bf43-bd9ad90cf9a6","arxiv_id":"2507.18817","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hybrid quantum-classical pipeline co-optimizes codon usage and mRNA secondary structure and runs end-to-end on IBM quantum hardware for short test sequences.","lead":"This paper tests a hybrid quantum-classical workflow that selects mRNA codons for high translation efficiency while using a quantum computer to predict RNA folding. It shows the full pipeline running on IBM's 127-qubit Eagle processor for short sequences up to 30 nucleotides.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reward r and penalty p in the secondary-structure objective Eq. (7) are never specified, leaving the quantum solver's objective undefined; consequently the claimed co-optimization of CAI and MFE and the hardware/simulation agreement cannot be independently verified.","rationale":"The reader's weakest assumption identifies Eq. (7) and the unspecified r and p parameters as the critical point, and this stress-test pass agrees. The missing values are not merely a reproducibility artifact: without them, the objective actually minimized by the quantum subroutine is undefined, so the outer optimization loop cannot be said to minimize Eq. (1), the stated dual objective. The paper's only demonstration of agreement between the quantum-selected fold and the classical MFE fold is the single TLPKAD example; the 100 or so intermediate structures explored by Nelder-Mead are not checked, and Table I's 'success rate' is not defined in terms of MFE alignment. Alternative concerns such as missing code, absent classical baselines, and small qubit counts are real but secondary: they affect reproducibility and significance, whereas unspecified r and p affect the identity of the problem being solved. Because the reader's conditional verdict already requires supplying missing details, and because the concern reinforces rather than redirects that condition, the verdict should remain CONDITIONAL (no adjustment needed).","tokens_in":8022,"tokens_out":6545,"duration_ms":76161,"concrete_test":"Obtain from the authors the exact numerical values of r and p (and any scaling or weight conventions) used in every experiment, together with the generated QUBO for at least one reported case. Then, for each final sequence in Table I, solve Eq. (7) exactly using a classical QUBO/ILP solver (the reported qubit counts are at most 112, so this is computationally trivial), record the optimal dot-bracket, and evaluate it with ViennaRNA's eval_structure. Compare this energy to ViennaRNA's minimum free energy for the same sequence. If any nonzero gap appears, the surrogate is not aligned with the stated objective and the outer-loop selection is distorted. Separately, compare the hardware sample distribution in Fig. 2 to the exact Eq. (7) ground-state distribution to verify the hardware reproduction claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript defines the quantum secondary-structure prediction problem as the QUBO in Eq. (7), which contains two free scalar parameters: a reward r for adjacent stacked quartets and a penalty p for terminal (U,A) pairs. No numerical values, ranges, or scaling conventions for r or p appear anywhere in the text, figures, tables, or references. Equation (7) is therefore not a well-defined optimization problem: a reader cannot reconstruct the objective that the CVaR variational algorithm is supposed to minimize, nor can the claim that hardware sampling reproduced simulation be checked.\n\nThe problem is load-bearing rather than cosmetic. In Function 1, the composite objective is computed as fMFE = ViennaRNA.eval_structure(sequence, dotb), where dotb is the structure returned by the quantum solver. Thus the outer Nelder-Mead loop selects codon sequences based on the energy (under ViennaRNA) of the Eq. (7) optimum, not based on the true minimum free energy of the sequence. This is a valid co-optimization only if, for every sequence visited during the 99-106 function evaluations, the Eq. (7) optimum coincides with the ViennaRNA MFE structure. The paper only demonstrates this coincidence for the final fold in the TLPKAD example. For longer sequences, Table I reports success rates of 70-100%, but no comparison between Eq. (7) optima and ViennaRNA MFE optima is provided, and the definition of 'success' is not given.\n\nConsequently, the central claim that the hardware workflow 'returns the same optimal mRNA constructs as simulation' is internally consistent, but the word 'optimal' refers to an unstated surrogate. The paper does not establish that the returned sequences are optimal for the stated dual objective Eq. (1), which is a necessary part of the feasibility claim for mRNA co-optimization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a hybrid quantum-classical workflow for co-optimizing mRNA codon usage and secondary structure. A classical Nelder-Mead loop updates a parameter vector θ = (θc, θp, θr) that weights GC content, rarity, and repeat penalties in a codon-selection ILP (Eq. 6). The ILP solution yields a nucleotide sequence; the sequence's CAI is computed directly, and its secondary structure is predicted by a CVaR-based variational quantum solver solving the QUBO in Eq. (7). The resulting dot-bracket structure is scored with ViennaRNA's eval_structure, and the combined objective is f = α·CAI + MFE (Eq. 1). The authors report simulation results on seven amino-acid sequences up to 16 residues, with success rates of 70–100%, and hardware runs on IBM's 127-qubit Eagle processors for the TLPKAD example and for a 13-residue sequence sampled on ibm_fez. They conclude that end-to-end execution of the hybrid algorithm on real quantum hardware is feasible and that the hardware results reproduce simulation.","tokens_in":8314,"tokens_out":8603,"duration_ms":97463,"significance":"If fully substantiated, this is a useful systems-integration milestone: it demonstrates that a variational quantum subroutine can be embedded in a closed-loop sequence-design workflow and executed on a 127-qubit processor, with results matching simulation for short sequences. The decomposition is sensible: an exact classical solver handles the easy codon-selection ILP, while the quantum solver targets the more complex secondary-structure QUBO, and the external ViennaRNA evaluator provides a classical check on the predicted fold. The paper also includes a multi-sequence simulation table and a hardware demonstration, which is more than a purely theoretical proposal. However, the current evidence is not yet sufficient to establish the central feasibility claim in a reproducible form: key QUBO parameters are missing, the success metric is undefined, and the simulator-versus-hardware comparison lacks quantitative detail. The results are therefore of interest to the quantum-computing-for-biology community, but the manuscript needs substantial revision before the claims can be verified.","major_comments":[{"comment":"The objective function in Eq. (7) contains two undetermined scalar parameters: the stacking reward r and the terminal-pair penalty p. No numerical values, ranges, units, or scaling conventions are provided anywhere in the text, tables, figures, or references, and the index sets QS(qi) and QUA are not specified precisely enough to reconstruct the second and third sums. Consequently Eq. (7) does not define a concrete optimization problem, and the claim that hardware sampling reproduced simulation cannot be independently verified. This is load-bearing because the reported MFE values and final sequences depend on the structure returned by the solver that minimizes Eq. (7). Please specify r and p explicitly (for example, values from the Turner/NNDB parameters or from Ref. [19]) and clarify the semantics of the QS(qi) and QUA sets.","section":"Section II, Eq. (7)"},{"comment":"The 'success rate' reported in Table I is not defined, and no test is reported comparing the Eq. (7) optimum with the true ViennaRNA MFE structure for the sequences generated during optimization. Since Function 1 scores the quantum-solver structure with eval_structure, the outer loop's sequence choices are meaningful only if the Eq. (7) optimum coincides with the MFE optimum (or if the objective-value discrepancy is negligible). The text states this coincidence only for the final TLPKAD fold. Please define what 'success' counts across the ten runs, report per-sequence agreement between the surrogate optimum and the ViennaRNA MFE optimum, and quantify any impact of disagreement on the composite objective f.","section":"Section III-A, Table I"},{"comment":"The hardware claim rests on one TLPKAD trajectory and one sampling run on ibm_fez, but the paper provides no quantitative comparison between simulator and hardware output distributions, no fidelity metric, and no error bars. Figure 2 does not include a simulator baseline, axis labels, or a description of what the 'objective value' bins represent. Without these, the statement that 'the optimal structure was sampled with the highest frequency' on hardware, and the broader assertion that hardware reproduced simulation, are not supported by the displayed data. Please add side-by-side simulator and hardware distributions, a quantitative agreement measure (e.g., variational distance or Hellinger distance), and details on shot counts and any post-selection.","section":"Section III-B"},{"comment":"The reported final results for the TLPKAD example are internally inconsistent: the text and Fig. 1 give the final nucleotide sequence as ACUCUGCCGAAGGCAGAC with 27 iterations and 106 function evaluations, while Table I lists ACUCUGCCUAAGGCGGAC with 24 iterations and 99 function evaluations. Since this example is the cornerstone of the hardware demonstration, these discrepancies must be reconciled and the correct numbers reported consistently in the text, figure, and table.","section":"Section III-A and Table I"}],"minor_comments":[{"comment":"The symbol n is used for both the number of amino acids and the nucleotide sequence length in the secondary-structure model; using a distinct symbol (e.g., L) for the nucleotide length would avoid ambiguity.","section":"Section II"},{"comment":"The phrase 'Nshots = 213 bitstrings' presumably means 2^13 = 8192 bitstrings; please clarify the shot count and how the raw samples were converted to objective values.","section":"Section III-B"},{"comment":"The p-term writes qi(1−qj) with qj ∈ QUA, which is confusing because QUA is described as a set of stacked quartets; please state directly which variables incur the terminal (U,A) penalty rather than using an implicit complement.","section":"Section II, Eq. (7)"},{"comment":"Please add axis labels, units, and a legend; the claim that the optimal structure is the most frequent cannot be checked from the current figure.","section":"Figure 2"},{"comment":"Add a caption or text defining 'success rate' and reporting the number of runs; the 'ten independent runs' is mentioned only in the text and not in the table caption.","section":"Table I"},{"comment":"Eq. (6) describes θ as the 'initial vector of variational parameters', but θ is the quantity optimized by Nelder-Mead; use 'optimization variable' or 'variational parameter' consistently.","section":"Section II, Eq. (6)"},{"comment":"The relationship between Eq. (6)'s GC/rarity/repeat objective and the CAI term in Eq. (1) is not discussed; a sentence explaining why minimizing the rarity penalty with weights θp is aligned with maximizing CAI would help the reader understand the design.","section":"Section II"}],"recommendation":"major_revision","confidential_remarks":"The main barriers to acceptance are the unspecified r and p parameters, the undefined success metric, the lack of quantitative simulator-hardware comparison, and the TLPKAD data discrepancy. All of these are fixable in revision, so I recommend major revision rather than rejection. The editor may wish to ask for a reproducibility appendix containing the exact r, p values, a precise definition of success rate, per-sequence comparisons between Eq. (7) optima and ViennaRNA MFE optima, and simulator-versus-hardware distributions with statistical measures. The discrepancy between Table I and the text for TLPKAD suggests the table may not have been updated after a code change, and should be checked carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first end-to-end demonstration of an mRNA codon/structure co-optimization loop where the secondary-structure solver runs on real IBM hardware—an 18-nucleotide transcript on Eagle, with longer sequences in simulation and hardware sampling. Second, the paper as written cannot be independently reproduced: Eq. (7) defines the quantum solver's QUBO with two free scalar parameters r and p that are never given values, so the objective being minimized is not well-defined.\n\nWhat is genuinely new: earlier quantum codon-optimization papers were simulation-only, and Ref. [19] is the authors' own structure predictor. Here they integrate that predictor with a classical codon ILP in a variational loop and execute the whole thing on quantum hardware. The final fold for TLPKAD is checked against ViennaRNA's classically computed MFE, and the hardware run reproduces the simulated trajectory. That is an honest engineering milestone, for very short sequences.\n\nThe stress-test concern is on target and load-bearing. Without r and p, no one can reconstruct the Eq. (7) objective. The outer loop selects codons based on whether the Eq. (7) optimum matches the ViennaRNA MFE structure, but the paper only checks this for the final fold of one example. Table I reports success rates of 70–100% without defining what \"success\" means or comparing quantum-predicted structures to ViennaRNA MFE structures for the longer sequences. There is also an objective mismatch the reader caught: the stated dual objective Eq. (1) includes CAI, but the codon ILP Eq. (6) optimizes GC content, rarity, and repeat penalties—CAI is computed after the fact, so the composite f(θ) does not actually contain the CAI of the chosen sequence during the search. Missing code/data, no classical baseline (e.g., the same loop with a classical MFE predictor instead of the quantum subroutine), and the abstract says 30 nucleotides while Table I goes to 48. These are all addressable.\n\nWho is this for? People working at the quantum-for-biology interface and mRNA design toolbuilders curious whether current hardware can be inserted into a real optimization loop. It deserves a serious referee, not a desk reject, but the revision must specify r and p, define success, add a classical baseline, and release code or data.","headline":"Real-hardware hybrid mRNA co-optimization loop is a genuine first, but the quantum solver's QUBO parameters are left unspecified, making the 'optimal' claim unverifiable as written.","tokens_in":8971,"tokens_out":1612,"would_cite":false,"duration_ms":19292,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid quantum-classical algorithm co-optimizes mRNA codon choice and folding, and runs end-to-end on a 127-qubit Eagle processor.","keywords":["mRNA design","codon optimization","secondary structure prediction","hybrid quantum-classical algorithm","CVaR variational optimization","quadratic binary optimization","superconducting quantum processor","codon adaptation index"],"falsifier":"Enumerate all synonymous codon assignments for a short amino-acid sequence, compute the Eq. (7) score and the ViennaRNA MFE for each, and find one sequence where the two disagree about which fold is best; if the outer optimizer then selects a codon choice by the Eq. (7) ranking, the surrogate is steering the design away from the thermodynamically optimal structure.","tokens_in":7790,"feed_emoji":"🧬","tokens_out":9106,"duration_ms":89759,"temperature":0.7,"pith_summary":"An mRNA sequence has to be easy for the ribosome to translate and able to fold into a stable structure, and the two goals fight because both come from the same nucleotide string. This paper tries to settle that fight with one optimization loop: a classical solver picks a codon for each amino acid, a variational quantum subroutine predicts how the resulting RNA folds, and a composite score combining codon adaptation index (CAI) and minimum free energy (MFE) tells both pieces how to improve. The central claim is that this loop is not only simulable but executable on today's hardware: the full workflow, including the quantum fold-prediction step, ran on a 127-qubit Eagle processor and produced the same optimal mRNA constructs for the tested short sequences as the simulator did. That matters because co-optimized mRNAs are the kind of constructs used in therapeutics, where stability in the vial and translation in the cell both depend on getting the sequence right.","feed_headline":"Quantum hardware runs full mRNA co-optimization loop","feed_subtitle":"Classical codon selection plus a quantum folding step reproduces simulator results on a 127-qubit Eagle processor.","key_machinery":"The mechanism that carries the argument is the quadratic binary program of Eq. (7), in which every decision variable is a 'quartet'—two consecutive stacked base pairs—and the objective sums nearest-neighbor free energies from the Turner database, rewards adjacent stacks, and penalizes folds that end in a UA pair. This program is solved by a CVaR variational quantum algorithm that concentrates sampling on low-energy bitstrings, while the outer optimization is a Nelder-Mead loop over the weights θc, θp, θr of the codon-selection linear program, which CPLEX solves exactly and which encodes GC content, rarity, and repeat penalties. The two halves are coupled by the composite objective f = α·CAI + MFE with α = −0.5, so each candidate codon choice is judged by how well the sequence it produces folds.","core_discovery":"The core discovery is that codon usage and secondary structure can be co-optimized in a closed variational loop that survives contact with real quantum hardware. For the six-amino-acid sequence TLPKAD, the loop converged to the 18-nucleotide sequence ACUCUGCCGAAGGCAGAC with CAI 0.718 and MFE −8.100, matching the classically predicted dot-bracket fold, and the hardware run reproduced the simulated optimization trajectory; a 13-amino-acid sequence sampled on a second Eagle processor returned the optimal fold with the highest frequency. Across ten independent runs on sequences of 18 to 48 nucleotides, the method found the classically consistent structure 70–100% of the time. The paper reads this as evidence that the hybrid workflow, with its CVaR variational quantum solver for secondary structure, is a feasible design loop on current superconducting processors.","pith_inferences":["If Eq. (7) ever ranks folds differently from ViennaRNA's MFE on a sequence the optimizer visits, the outer loop would be optimizing toward a distorted target; a direct check would compare the two rankings over all synonymous codon assignments for one amino-acid sequence.","Because the reward r and penalty p in Eq. (7) are never given numerical values, the reported hardware agreement is not independently reproducible without those constants; publishing them would let others rerun the exact experiment.","The same loop could be pointed at other RNA-design objectives—avoiding G-quadruplex motifs, cell-type-specific codon tables, or repeat penalties—by swapping the codon metric while keeping the quantum fold subroutine.","The biological payoff is asserted but not measured: synthesizing the hardware-selected mRNAs and measuring expression and half-life would test whether the co-optimized sequences actually behave better than codon-only designs."],"forward_implications":["The end-to-end loop runs on current 127-qubit hardware, so the quantum secondary-structure subroutine is demonstrated outside pure simulation for short mRNAs.","For the tested sequences, the optimizer's chosen folds agree with classical predictions in 70–100% of independent runs, so the hybrid loop is consistent, not just convergent.","Caching previously evaluated sequences cuts redundant quantum calls, letting the Nelder-Mead loop finish in 12–27 iterations on the tested cases.","The tunable parameter α lets a designer shift the balance between codon optimality and structural stability within the same framework.","Scaling to therapeutically sized mRNAs will face circuit-depth and optimizer-efficiency limits, pointing toward modular decomposition and surrogate models as planned next steps."],"supporting_citations":[{"why":"Supplies the CVaR variational quantum algorithm for secondary-structure prediction that the co-optimization loop embeds.","marker":"[19]"},{"why":"Provides the GC-content, rarity, and repeat-scoring formulation the codon-selection linear program adapts.","marker":"[16]"},{"why":"Provides the nearest-neighbor stacking free energies assigned to quartet variables in Eq. (7).","marker":"[22]"},{"why":"ViennaRNA's eval_structure computes the MFE that scores each candidate fold in the composite objective.","marker":"[25]"},{"why":"Defines the codon adaptation index used as the translation-efficiency half of the objective.","marker":"[14]"},{"why":"CPLEX solves the codon-selection binary linear program exactly at every outer iteration.","marker":"[21]"},{"why":"Nelder-Mead is the classical optimizer that tunes the codon-weight parameters across iterations.","marker":"[23]"},{"why":"Provides the CVaR objective that stabilizes the variational solver on noisy quantum hardware.","marker":"[24]"}],"fun_headline_variants":["Quantum loop co-optimizes codon and mRNA fold","Hybrid quantum workflow tunes mRNA stability","Eagle processor runs mRNA co-optimization loop","Codon and structure co-optimized on quantum chip"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme depends on the simplified fold-scoring formula of Eq. (7), with its two unstated reward and penalty constants, ranking candidate mRNA structures in the same order as true minimum free energy for every sequence the optimizer generates.","fun_headline_variants_meta":{"raw":{"variants":["Quantum loop co-optimizes codon and mRNA fold","Hybrid quantum workflow tunes mRNA stability","Eagle processor runs mRNA co-optimization loop","Codon and structure co-optimized on quantum chip"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000326,"raw_usage":{"total_tokens":1812,"prompt_tokens":920,"completion_tokens":892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":831}},"tokens_in":536,"tokens_out":892,"duration_ms":8959,"temperature":1.0,"reasoning_tokens":831,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:31:14.202790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate all synonymous codon assignments for a short amino-acid sequence, compute the Eq. (7) score and the ViennaRNA MFE for each, and find one sequence where the two disagree about which fold is best; if the outer optimizer then selects a codon choice by the Eq. (7) ranking, the surrogate is steering the design away from the thermodynamically optimal structure.","supporting_citations":[{"cited_title":"mrna secondary structure prediction using utility-scale quantum computers,","cited_arxiv_id":null,"evidence_quote":"Supplies the CVaR variational quantum algorithm for secondary-structure prediction that the co-optimization loop embeds."},{"cited_title":"NNDB: the nearest neighbor parameter database for predicting stability of nucleic acid secondary structure,","cited_arxiv_id":null,"evidence_quote":"Provides the nearest-neighbor stacking free energies assigned to quartet variables in Eq. (7)."},{"cited_title":"The codon adaptation index-a measure of directional synonymous codon usage bias, and its potential applications,","cited_arxiv_id":null,"evidence_quote":"Defines the codon adaptation index used as the translation-efficiency half of the objective."},{"cited_title":"License information","cited_arxiv_id":null,"evidence_quote":"CPLEX solves the codon-selection binary linear program exactly at every outer iteration."},{"cited_title":"A simplex method for function minimiza- tion,","cited_arxiv_id":null,"evidence_quote":"Nelder-Mead is the classical optimizer that tunes the codon-weight parameters across iterations."},{"cited_title":"Improving variational quantum optimization using cvar,","cited_arxiv_id":null,"evidence_quote":"Provides the CVaR objective that stabilizes the variational solver on noisy quantum hardware."}],"review_version":1}