{"id":"81fc9049-8445-41db-9a5c-ac1624c410e6","arxiv_id":"2505.23289","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Quantum annealing on D-Wave hardware reproduces mean marker incidences and pairwise correlations of a learned epigenetic Ising model as accurately as classical Boltzmann sampling, while generating TAD-like structural motifs.","lead":"Researchers used a quantum annealer to sample configurations of an epigenetic Ising model of chromatin, and the results matched empirical statistics for epigenetic marker co-occurrence. The work tests whether quantum annealing can be a practical sampler for biological models that classical computers find difficult.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Cartesian exclusion of mixed marker–nucleosome couplings (Eq. 2) is asserted without reported evidence; if those couplings are non-negligible, the learned model and the claimed reproduction of chromatin statistics are misspecified.","rationale":"The reader identified Eq. (2) as the weakest assumption, and I agree that it is the most load-bearing point. The central claim requires that the learned epigenetic Ising model captures the statistics relevant to TAD-like chromatin organization. The paper learns parameters from three categories of statistics—mean incidences, intra-nucleosome correlations, and same-marker inter-nucleosome correlations—then evaluates success on those same categories. That circularity is mitigated if the model is treated as a proof of concept for quantum-annealing sampling, but it means the model's biological validity rests entirely on the assertion that all omitted mixed couplings are negligible. That assertion is not supported by any reported computation: no cross-marker spatial correlation matrix is shown, and the paper simply states that such couplings are not statistically relevant and would complicate embedding. This is a missing-support issue, not a disagreement with consensus, because the omitted terms are measurable from the same binarized dataset already used in the paper. If those correlations are substantial, then the model is structurally incapable of representing patterns such as bivalent domains or cross-talk between adjacent nucleosomes carrying different marks, and the TAD-like motifs in Fig. 14 may be artifacts of the wrong interaction graph. Furthermore, the Cartesian graph product structure is explicitly used to justify the favorable embedding properties in Section III.2 and Appendix A, so the assumption is load-bearing both for biological fidelity and for the feasibility of the quantum-annealing approach. The concrete test above would settle the question directly by measuring the omitted correlations and, if needed, testing a model that includes them. Because this concern is already reflected in the reader's CONDITIONAL verdict, I do not recommend changing the verdict; the paper should either supply the missing evidence or restrict its claims accordingly.","tokens_in":21312,"tokens_out":5411,"duration_ms":63529,"concrete_test":"Compute the full binned cross-correlation tensor from the binarized IMR90 data used in Appendix B: for every marker pair (m, m') with m != m' and every distance l = 1,...,5, evaluate rho_{m,m'}(l) = (1/N) sum_n x_n^m x_{n+l}^{m'} (with appropriate normalization and mean subtraction), and compare with the included same-marker correlations S_l^m and intra-nucleosome correlations R_{mm'}. If the mean absolute value of rho_{m,m'}(l) is comparable to the included terms, Eq. (2) is empirically unsupported; as a follow-up, fit a small model (e.g., [4, 7, 1]) including mixed couplings and check whether quantum-annealing sampling reproduces held-out statistics better than the Cartesian model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is the Cartesian decomposition in Eq. (2): all couplings between different epigenetic markers at different nucleosome positions are set to zero. The paper asserts that these mixed couplings 'do not bear significant statistical relevance' (Section II.2) but never reports the corresponding empirical correlations. The learning objective in Section II.4 only targets mean incidences, intra-nucleosome marker correlations, and same-marker inter-nucleosome correlations, so the evaluation in Fig. 11 cannot detect misspecification in the omitted sector. If cross-marker spatial correlations in the IMR90 data are comparable to the included terms, the learned model is not a faithful maximum-entropy model of the empirical statistics, and the sampled TAD-like motifs—which are themselves only shown for selected examples—could arise from an incomplete interaction structure. The Cartesian structure is also what makes the objective graph sparse and embeddable on Pegasus (Section III.2); restoring mixed couplings would increase graph density and chain lengths, directly threatening the practical quantum-annealing advantage claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a pipeline that binarizes IMR90 ChIP-seq data into 12 epigenetic markers over 200-bp bins, learns an Ising/QUBO model whose parameters are per-marker biases, intra-nucleosome marker couplings, and same-marker inter-nucleosome distance couplings, and embeds this Cartesian model onto the D-Wave Pegasus topology. The parameters are fitted by classical Boltzmann sampling against empirical mean incidences and correlations. The authors then benchmark quantum annealing samples against classical Boltzmann samples and empirical statistics, reporting qualitative agreement in Fig. 11, grid-search results for annealing time, chain strength, boundary conditions, and coupling threshold, plus reverse-annealing and cluster-parallelization experiments. The abstract claims that QA reproduces mean marker incidences and intra- and inter-nucleosome correlations while generating TAD-like structural motifs, rather than reconstructing exact TAD size distributions or insulation scores.","tokens_in":21505,"tokens_out":5619,"duration_ms":65970,"significance":"If the missing quantitative support were supplied, this would be a useful proof-of-concept that a D-Wave quantum annealer can sample an epigenetically learned Ising model as well as a classical Boltzmann sampler for model sizes up to [12, 25, 5]. The paper has clear strengths: an explicit QUBO mapping, a detailed embedding analysis with chain-length statistics across Pegasus and Zephyr topologies, use of real quantum hardware, and transparent hyperparameter sweeps over T_A, J_C, boundary conditions, and coupling threshold. However, the currently reported evidence is essentially qualitative: Fig. 11 has no error bars or reported R2 values, no statistical tests are given, and the omitted mixed-coupling sector is not validated. The paper also conflates model learning with hardware sampling and overstates timing gains by comparing anneal time to full Markov-chain time. These issues are fixable within the scope of the manuscript, but they must be addressed before the central claim can be accepted.","major_comments":[{"comment":"The assertion that mixed marker-nucleosome couplings Q^{nn'}_{mm'} for m != m' and n != n' 'do not bear significant statistical relevance' is not supported by any reported analysis. The learning objective in Section II.4 fits only mean incidences, intra-nucleosome marker correlations, and same-marker inter-nucleosome correlations, and Fig. 11 evaluates exactly those quantities; therefore neither the learning procedure nor the evaluation can detect whether the omitted couplings are negligible. Please add the empirical cross-marker spatial correlation matrix, or a summary of its magnitude relative to the included terms, and if those correlations are non-negligible, either include the terms or demonstrate that their omission does not change the conclusions. This point is load-bearing because the Cartesian structure is what keeps the objective graph sparse enough for the minor-embedding analysis in Section III.2.","section":"Section II.2, Eq. (2)"},{"comment":"The central claim of statistical reproduction rests on a qualitative visual comparison. No error bars, confidence intervals, per-category R2 values, or statistical tests are reported; the R2 metric introduced in the grid-search paragraph is not used in Fig. 11 or in any table, and no statement about run-to-run variability is made. Please report R2 or a comparable error metric with uncertainties from repeated QA and classical runs, for each model size and statistic category, and state explicitly whether the QA-CBS differences are within noise. Because the displayed agreement is obtained after a grid search over T_A, J_C, boundary conditions, and delta, a sensitivity analysis or holdout evaluation is needed to establish that the agreement is not an artifact of parameter selection.","section":"Figure 11 and Section IV.1"},{"comment":"Because the model parameters are learned from the same empirical means and correlations that are used as the evaluation target in Fig. 11, the reproduction of these statistics by the classical Boltzmann sampler is partly by construction. For the QA claim, what is actually demonstrated is that QA sampling from the learned Hamiltonian approximates the classical sampler's target distribution. The manuscript should separate these two statements, for example by reporting QA-vs-exact-model statistics (sampled model means and correlations versus their Boltzmann expectations) in addition to QA-vs-empirical statistics. Without that separation, the abstract's phrasing that QA 'reproduces' empirical statistics conflates model learning with hardware sampling.","section":"Section II.4 and Section IV.1"},{"comment":"The claim that the sampled configurations exhibit 'TAD-like structural motifs' is supported only by three selected sample images. No automated or statistical definition of a TAD-like motif is given, no quantitative measure of motif frequency is reported, and there is no comparison with the frequency of such motifs in random configurations or in classical Boltzmann samples. Please provide a quantitative criterion for TAD-likeness, its distribution over many samples, and a null-model comparison; otherwise the motif claim remains anecdotal.","section":"Section IV.1, Fig. 14"},{"comment":"The timing comparison is not apples-to-apples. The reported '100x' speed-up from cluster-parallelized QA counts only annealing time, while the classical side appears to include full Markov-chain step time; QPU programming, readout, and classical post-processing are not included in the quantum time. A meaningful comparison should report end-to-end wall-clock time for both methods, or explicitly qualify the speed-up as anneal-time-only and acknowledge that this does not support the 'order-of-magnitude reductions in sampling time' statement in the conclusion.","section":"Section IV.3"}],"minor_comments":[{"comment":"The sentence 'QA faithfully reproduces the TAD length distributions and insulation score profiles while delivering an order-of-magnitude speedup' contradicts the abstract and Section II.3, which state that exact TAD size distributions and insulation scores are not reconstructed; please align these statements.","section":"Introduction, first paragraph after Section I"},{"comment":"The figure lacks labeled axes, a legend explaining the colored categories, and a statement of whether the y-axis is on a logarithmic scale; the acronyms CBS and QAS should be expanded in the caption.","section":"Figure 11"},{"comment":"The caption text describing which panel shows chain strength and which shows annealing time appears to be swapped relative to the panel layout; please correct the description and label both axes explicitly.","section":"Figure 12"},{"comment":"There are notational inconsistencies in the QUBO indices, including the conditions m > m' and n > n' in Eq. (1) versus the sums in Eq. (3), and the threshold condition in Appendix E is written as delta > max(|J_i|), which does not match the subsequent description of removing couplings below delta; please standardize the notation.","section":"Equations (2)-(3) and Appendix E"},{"comment":"Reproducibility details are missing: the binarization threshold for ChIP-seq peak strength is not specified, the Boltzmann sampling parameters (nsteps, beta, learning rate) are not given, and no data or code availability statement is provided.","section":"Section II.4 and Appendix B"},{"comment":"There are small typos such as 'Zephr' instead of 'Zephyr' in Fig. 20, and the caption 'max(CL)' in Fig. 24 should be 'max(C_L)' or similar; please proofread the figure text.","section":"Figures 20 and 22"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to contain an internal contradiction between the Introduction's strong claim about reproducing TAD length distributions and insulation scores and the more modest abstract claim; the authors should reconcile these in revision. The Cartesian-truncation issue in Eq. (2) is the most substantive technical concern and should be resolved with empirical evidence rather than assertion. The paper would be strengthened by a clear separation of model-learning performance from quantum-sampling performance, especially by comparing QA outputs to exact Boltzmann statistics of the learned model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a legitimate proof-of-concept that quantum annealing can sample from a learned epigenetic Ising model and reproduce the mean marker incidences and intra-/inter-nucleosome correlations about as well as classical Boltzmann sampling. The abstract is careful—it disclaims exact TAD size distributions and insulation scores—but the introduction overclaims, saying QA \"faithfully reproduces the TAD length distributions and insulation score profiles.\" That sentence does not match anything in the results. The authors need to fix that before this goes anywhere.\n\nWhat is genuinely new: the D-Wave embedding of the Cartesian model, the chain-length scaling analysis across Pegasus and Zephyr topologies, the cluster-parallelized single-readout sampling, and the reverse-annealing experiments. The embedding analysis is the strongest part of the paper—thorough, quantitative, and directly useful to anyone trying to map similar grid-like Ising models onto current hardware. The cluster parallelization trick is neat and the validation that it gives nearly identical statistics is a real point in the paper's favor.\n\nSoft spots, in proportion: First, the Cartesian assumption in Eq. (2) sets all mixed marker–nucleosome couplings to zero, and the paper just asserts they \"do not bear significant statistical relevance\" without reporting the corresponding empirical correlations. The learning objective only targets means, intra-nucleosome correlations, and same-marker inter-nucleosome correlations, so the evaluation in Fig. 11 cannot detect whether the omitted couplings matter. This limits the biological fidelity claim, but it does not undermine the sampling proof-of-concept: the paper is really demonstrating that QA can handle this particular sparse model, not that this model is the right biology. Second, Fig. 11 has no error bars, no final R2 values, and no statistical tests; the grid search over annealing time and chain strength used R2, but the final values are not reported. Combined with the absence of code and detailed parameter values, this makes the central comparison hard to assess. These are addressable issues, not fatal ones. Third, the circularity burden is real but mild: parameters are learned from the same statistics used for evaluation, so the classical sampler's reproduction is partly by construction. What is still meaningful is that the quantum annealer, despite chain breaks and hardware noise, lands close to the classical result.\n\nWho this is for: people working on quantum annealing applications and anyone interested in sampling from learned Ising models of chromatin. The biological conclusions should be read cautiously, but the methods and embedding analysis are sound.\n\nRecommendation: send it to peer review. The core demonstration is honest, the execution time discussion acknowledges that classical methods remain competitive, and the issues are fixable. With the introduction toned down, error bars added, and code/data released, this would be a useful reference point for hybrid quantum–classical workflows in epigenomics.","headline":"A credible proof-of-concept that D-Wave annealing can sample a learned epigenetic Ising model about as well as classical Boltzmann sampling, but the introduction overclaims relative to the results, and the load-bearing Cartesian coupling assumption is never validated.","tokens_in":22049,"tokens_out":2124,"would_cite":false,"duration_ms":24076,"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":"The paper claims that quantum annealing can sample chromatin configurations from an epigenetic Ising model accurately enough to reproduce empirical marker statistics and produce TAD-like structural motifs, matching classical Boltzmann…","keywords":["quantum annealing","chromatin","topologically associating domains","epigenetic Ising model","QUBO","Boltzmann sampling","nucleosome correlations"],"falsifier":"A direct test would compute, from the ChIP-seq data, the joint correlation of two different markers at two different nucleosome positions, terms the model sets to zero, and compare them with the same correlations in samples drawn from the Cartesian model; a systematic, threshold-dependent mismatch would show that the omitted mixed couplings carry TAD-relevant information and break the central claim.","tokens_in":21046,"feed_emoji":"🧬","tokens_out":7270,"duration_ms":74042,"temperature":0.7,"pith_summary":"This paper asks whether a quantum annealer can serve as a practical sampler for models of chromatin folding. It takes an Ising-type model in which each nucleosome carries binary epigenetic markers, with couplings learned from empirical ChIP-seq data, and embeds it onto quantum annealing hardware. The central claim is that the annealer reproduces the statistical features that matter for TAD formation, namely mean marker incidences and intra- and inter-nucleosome correlations, about as well as classical Boltzmann sampling, while also producing configurations with TAD-like structural motifs. If true, it means near-term quantum hardware can generate candidate intermediate chromatin states without reconstructing exact TAD boundaries, opening a sampling-based route into epigenetic modeling.","feed_headline":"Quantum annealing reproduces chromatin domain statistics","feed_subtitle":"An epigenetic Ising model on a quantum annealer matches classical sampling on marker correlations and yields TAD-like structures.","key_machinery":"The load-bearing object is the Cartesian QUBO/Ising objective function, which restricts couplings to two kinds: intra-nucleosome couplings between different markers at the same nucleosome, and inter-nucleosome couplings between the same marker at different nucleosomes, with all mixed marker-and-position couplings set to zero. Because of this Cartesian product structure, the objective graph is sparse and low-dimensional, which makes it embeddable on the annealing hardware with modest qubit chains. The argument runs through four steps: learning the biases and couplings from empirical data by classical Boltzmann sampling and gradient updates, minor-embedding the objective graph onto the processor topology, tuning annealing time, chain strength, boundary conditions, and a coupling threshold, and evaluating the sampled ensembles against empirical statistics using the coefficient of determination $R^2$ on log-transformed values.","core_discovery":"The paper's discovery is that the difficult part of this task, sampling from a frustrated and densely coupled epigenetic energy landscape, can be shifted to the annealer while the biologically meaningful statistics remain intact. Concretely, for model configurations up to [12, 25, 5] (twelve markers, twenty-five nucleosomes, five-nucleosome coupling range), quantum annealing samples match empirical mean incidences, within-nucleosome marker correlations, and same-marker across-nucleosome correlations at a level comparable to a classical Metropolis-based Boltzmann sampler. The sampled configurations form block-like structures resembling TADs, and by adding a bias toward an empirical domain or using reverse annealing, the sampler explores states in the structural vicinity of that domain. The paper is careful to say that it does not recover exact TAD size distributions or insulation scores; the match is at the level of statistical features and structural motifs.","pith_inferences":["A testable extension the paper leaves implicit is to include the strongest mixed couplings (different markers at different nucleosomes) and re-learn the model; if $R^2$ improves on cross-marker spatial correlations, the Cartesian restriction is a genuine loss, and if not, the restriction is validated.","If the reverse-annealing depth parameter controls how far sampled states sit from an empirical TAD, then varying it may produce an ensemble that can be compared directly with Hi-C maps of intermediate folding states, which the paper motivates but does not test.","The reported speed-up is specific to fitting many copies of a small model on one chip; as hardware connectivity grows, the same cluster-replication idea would extend to larger $[M, N, L]$ windows, making genome-wide sampling a question of qubit count rather than algorithm redesign.","The learned parameters come from a classical Boltzmann loop; a fully quantum loop, in which quantum annealing samples supply the gradients for parameter updates, is the natural next step and would remove the classical sampling bottleneck entirely."],"forward_implications":["Quantum annealing can be used as a drop-in replacement for classical Boltzmann sampling in this epigenetic model when the goal is statistical fidelity rather than exact boundary prediction.","Because the embedding depends only on the model size $[M, N, L]$ and not on the parameter values, one computed embedding serves many independent sampling runs, amortizing the embedding cost.","When the hardware can fit replicas of the objective graph, all 100 samples can be drawn in a single anneal, yielding an effective 100-fold speed-up in annealing time.","The optimal annealing time for sampling is finite: too long an anneal drives the system into the ground state and destroys the Boltzmann diversity needed for computing statistics.","Pruning weak couplings below a threshold near $\\delta = 0.25$ simplifies the embedding and can improve performance, while pruning beyond that removes biologically relevant interactions and degrades the match."],"supporting_citations":[{"why":"Supplies the entropy-based chromatin-domain modeling framework and the iterative parameter-learning scheme that this paper adapts to a QUBO/Ising objective.","marker":"[23]"},{"why":"Provides the binarized ChIP-seq epigenetic marker data used as the empirical reference for the statistics.","marker":"[30]"},{"why":"Defines the IMR90 human fetal lung fibroblast cell line whose chromosome 9 data are used in the experiments.","marker":"[31]"},{"why":"Describes the post-processing used to reduce sampling noise in quantum-annealing readouts.","marker":"[28]"},{"why":"Provides the error-mitigation approach applied to the annealer samples before statistical analysis.","marker":"[29]"},{"why":"Documents the reverse-annealing schedule used to sample states in the vicinity of an empirical TAD.","marker":"[47]"},{"why":"Supplies the clique-minor embedding strategy used for the complete-graph intra-nucleosome subgraph.","marker":"[49]"},{"why":"Is the minorminer heuristic used to compute the minor embeddings onto the hardware graph.","marker":"[54]"}],"fun_headline_variants":["Quantum annealing matches classical sampling on chromatin markers","QA reproduces TAD-like structures from epigenetic data","Quantum sampler captures chromatin domain statistics","D-Wave annealer samples epigenetic Ising model","Quantum annealing matches empirical marker correlations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that two different epigenetic markers at two different nucleosome positions never interact directly, so the model sets all such couplings to zero; if these omitted cross terms carry information needed for TAD formation, the sampled configurations will be distorted.","fun_headline_variants_meta":{"raw":{"variants":["Quantum annealing matches classical sampling on chromatin markers","QA reproduces TAD-like structures from epigenetic data","Quantum sampler captures chromatin domain statistics","D-Wave annealer samples epigenetic Ising model","Quantum annealing matches empirical marker correlations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000475,"raw_usage":{"total_tokens":2332,"prompt_tokens":894,"completion_tokens":1438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1372}},"tokens_in":510,"tokens_out":1438,"duration_ms":12646,"temperature":1.0,"reasoning_tokens":1372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:47:59.554802+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would compute, from the ChIP-seq data, the joint correlation of two different markers at two different nucleosome positions, terms the model sets to zero, and compare them with the same correlations in samples drawn from the Cartesian model; a systematic, threshold-dependent mismatch would show that the omitted mixed couplings carry TAD-relevant information and break the central claim.","supporting_citations":[{"cited_title":"Mater- nal–infant interaction quality is associated with child nr3c1 cpg site methylation at 7 years of age,","cited_arxiv_id":null,"evidence_quote":"Supplies the entropy-based chromatin-domain modeling framework and the iterative parameter-learning scheme that this paper adapts to a QUBO/Ising objective."},{"cited_title":"Using quantum annealing to design lattice proteins,","cited_arxiv_id":null,"evidence_quote":"Provides the binarized ChIP-seq epigenetic marker data used as the empirical reference for the statistics."},{"cited_title":"Efficient quantum algorithm for lat- tice protein folding,","cited_arxiv_id":null,"evidence_quote":"Defines the IMR90 human fetal lung fibroblast cell line whose chromosome 9 data are used in the experiments."},{"cited_title":"Comparative study of the performance of quantum annealing and simulated anneal- ing,","cited_arxiv_id":null,"evidence_quote":"Describes the post-processing used to reduce sampling noise in quantum-annealing readouts."},{"cited_title":"Folding lattice proteins with quantum annealing,","cited_arxiv_id":null,"evidence_quote":"Provides the error-mitigation approach applied to the annealer samples before statistical analysis."},{"cited_title":"Exact correspondence be- tween renyi entropy flows and physical flows,","cited_arxiv_id":null,"evidence_quote":"Documents the reverse-annealing schedule used to sample states in the vicinity of an empirical TAD."},{"cited_title":"Error mitigation in brainbox quan- tum autoencoders,","cited_arxiv_id":null,"evidence_quote":"Supplies the clique-minor embedding strategy used for the complete-graph intra-nucleosome subgraph."}],"review_version":1}