{"id":"4ca37d79-5ff8-40e0-951e-1bc77c6483b2","arxiv_id":"2411.09216","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Trajectory reweighting and implicit differentiation are combined to fit coarse-grained DNA, RNA, and DNA-protein models to structural, mechanical, and thermodynamic experimental targets in a single gradient-based framework.","lead":"This paper demonstrates an automatic-differentiation framework that fits coarse-grained DNA and RNA simulation models to experimental measurements such as helix pitch, stiffness, and melting temperature. It matters because molecular force fields are usually tuned by hand, and this method makes the fitting process systematic, extensible, and potentially reproducible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Thermodynamic gradient bias from the omitted boundary term at the base-pair threshold is unquantified; a finite-difference check is needed before the melting-temperature demonstrations can be taken as validating the framework.","rationale":"The reader's weakest_assumption correctly identifies the unquantified approximation in the melting-temperature gradient as the most load-bearing concern. The central claim is that the framework can fit structural, mechanical, and thermodynamic properties in a systematic and reproducible way; if the thermodynamic gradient is biased by an omitted boundary term, then a core branch of the demonstration is not established. The paper itself flags the approximation, but provides no numerical evidence of its size, so the concern is internal to the argument rather than a disagreement with external consensus. I considered alternative concerns: the absence of released code and data weighs against the 'frictionless reproducibility' wording, and the 'micro- and millisecond timescales' claim in the abstract is not obviously supported by the reported simulation lengths. However, those issues are secondary to the correctness of the thermodynamic gradient because even perfect code would not resolve a biased estimator. The paper has real independent support: DiffTRE is a published estimator, the JAX implementation is benchmarked against standalone oxDNA, implicit differentiation is standard, and the mechanical-structure optimizations do not rely on the disputed threshold approximation. Those strengths do not remove the need to test the thermodynamic approximation. The reader's conditional verdict already captures this uncertainty, so no change in verdict is warranted; the recommendation is to require the finite-difference check before accepting the thermodynamic results as evidence for the framework.","tokens_in":32915,"tokens_out":6796,"duration_ms":86264,"concrete_test":"Recompute the melting-temperature gradient for a representative hydrogen-bonding parameter using the full Eq. 90 expression with the B-gradient term omitted, and compare it to a central finite-difference estimate of dTm/dθ or dLoss/dθ obtained by rerunning the melting calculation at θ±ε with the same reference states and resampling protocol. If the relative error exceeds about 10-20% or the sign disagrees, the omitted boundary term is not negligible and the thermodynamic optimization branch is biased; if the error is small across several parameters and target conditions, the approximation is justified. An alternative analytical check is to replace the hard threshold in B(i) with a smooth sigmoid of width ε in energy, compute the gradient in the ε→0 limit, and compare with the paper's gradient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is in the thermodynamic gradient computation described in Supporting Information Section F.1, around Eq. 90. The paper computes melting-temperature gradients by differentiating the DiffTRE reweighting weights and the temperature-extrapolation exponential, but explicitly sets the gradient of the thresholded base-pair indicator B(i) to zero, calling this 'a reasonable approximation in practice.' This is not merely a benign discretization choice: the base-pair count N(R) enters the order parameter through delta functions in Eq. 78 and Eq. 90, so the observable is a discontinuous function of the force-field parameters. For such an observable, the true gradient of its expectation contains a boundary term concentrated on configurations where an infinitesimal parameter change flips B(i) between 0 and 1. The reweighted expectation computed by DiffTRE captures only the smooth part of the gradient; the boundary term is dropped. Whether this omission is acceptable depends on the density of sampled configurations near the bonding threshold and on how strongly the threshold-crossing energies depend on the optimized hydrogen-bonding parameters. The paper provides no finite-difference or exact-gradient comparison to quantify the bias. Because the melting-temperature optimization is one of the three core target classes and is also used in the joint optimization demonstrations, a large unquantified bias in this gradient would directly weaken the central claim that the framework can systematically fit coarse-grained potentials to thermodynamic data. The convergence plots show loss decreasing, but biased gradients can still decrease a proxy loss while moving parameters in directions that do not actually minimize the true melting-temperature error. This concern is specific, testable, and currently unresolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a systematic framework for fitting coarse-grained molecular force fields to macroscopic experimental data using differentiable trajectory reweighting (DiffTRE), implicit differentiation of summary-statistic fitting procedures, and conflict-free gradient updates from multi-task learning. The authors implement oxDNA and oxRNA energy functions in JAX, benchmark JAX-MD against the standalone codes, and demonstrate parameter optimization for structural properties (pitch, propeller twist, RMSD to a nanopore structure), mechanical properties (persistence length, torsional modulus, effective stretch modulus, twist-stretch coupling), thermodynamic properties (duplex and hairpin melting temperatures and melting width), and joint multi-property optimizations, including extensions to RNA and DNA-protein complexes with a custom sequence-specific Morse potential.","tokens_in":33216,"tokens_out":4268,"duration_ms":52161,"significance":"If the demonstrated gradients are unbiased and the convergence results are robust, the framework is a valuable contribution to coarse-grained model parameterization: it avoids differentiating through long trajectories, handles enhanced sampling and external force/torque ensembles, makes line-fitting procedures differentiable, and supplies a concrete recipe for multi-objective fitting. The sensitivity analyses based on gradient magnitudes are a useful methodological feature, and the use of existing simulation packages for reference-state sampling lowers the barrier to adoption. The paper also reports concrete performance gains over naive gradient estimators and demonstrates transfer to RNA and DNA-protein systems. However, the validation is weakened by an unquantified approximation in the thermodynamic gradient, by a partially circular evaluation against targets included in the loss, by omission of outlier iterations and error estimates in key convergence plots, and by the absence of a code/data availability statement despite the paper's central reproducibility claim.","major_comments":[{"comment":"The text after Eq. (90) states that the derivative of the thresholded base-pair indicator B(i) with respect to the force-field parameters is neglected because B is not continuous, calling this 'a reasonable approximation in practice.' This is an unquantified approximation for a discontinuous observable: the exact gradient of the reweighted expectation of delta(N = i) contains boundary terms from configurations where an infinitesimal parameter change flips B(i) between 0 and 1, and DiffTRE as written captures only the smooth part. Because the melting-temperature optimization is one of the three core target classes and is also used in the joint optimization in Figure 6B, the paper should provide a finite-difference or exact-gradient check for a small duplex or hairpin system to bound the resulting bias. At minimum, the magnitude of the omitted term relative to the retained gradient should be reported.","section":"Supporting Information F.1, Eq. (90)"},{"comment":"The agreement of C, Seff, and g with experimental values is obtained by optimizing the loss function toward exactly those experimental values, so the reported negative g (overwinding) is a fitted outcome rather than a prediction. The statements that the method 'successfully models the overwinding of DNA under low force' and yields parameters that 'agree with experiments' should be reframed as consistency checks of the fitting procedure, not as validation of the model. To support a predictive claim, the paper should include an out-of-sample evaluation, for example fitting to a subset of mechanical targets and reporting held-out properties, or testing the optimized parameters on sequences and salt conditions not used in the loss.","section":"Section 2.3, Figure 3C; Section 2.5, Figure 6B"},{"comment":"The captions disclose that outlier iterations are omitted and replaced with dashed lines (twist-stretch coupling at iterations 52-53 in Figure 3C, torsional modulus at iterations 53-55 in Figure S10). Omitting these points and not reporting their magnitudes prevents the reader from assessing the stability and variance of the optimization, which is load-bearing for the claims that conflict-free updates are more stable and that values converge 'within experimental error' or 'with low variance.' The outlier values and iteration-level variability should be reported, together with error bars on the converged property estimates.","section":"Section 2.3, Figure 3C caption; Section S10 caption"},{"comment":"The paper's central motivation and title claim is 'frictionless reproducibility' of coarse-grained model fitting, yet the manuscript contains no data availability statement, code repository link, or description of where the JAX implementations and optimization scripts can be obtained. Without these artifacts, the reproducibility claim cannot be evaluated and the framework cannot be built upon by the community. A public code repository with the implementation of the energy functions, DiffTRE gradient calculations, and scripts for at least the main optimization demonstrations should be provided.","section":"Abstract and Section 1"}],"minor_comments":[{"comment":"The first sentence contains a typo: 'in silica biophysical modeling' should be 'in silico biophysical modeling.'","section":"Section 1"},{"comment":"The caption states 'We use the ConFIG method of Zhang et al. [59]' while Section 2.5 and Section 4.3 attribute ConFIG to Liu et al. ([59] and [56], respectively). These attributions should be made consistent.","section":"Figure 6 caption"},{"comment":"Equation (15) has notation issues: the right-hand side uses 'gi' rather than 'g_i' and the matrix M is introduced without definition; the sentence beginning 'where wi > 0 for all i and M−⊤ is the pseudoinverse of the transposed matrix M⊤' should be completed and the indices corrected.","section":"Section 4.3, Eq. (15)"},{"comment":"The denominator is written as 'A4A1 − A2^3' but A2 was explicitly omitted for notational consistency and A3 is the torque-twist slope; the expressions for C and g appear to require 'A4 A1 − A3^2'. Please correct this typo or clarify the notation.","section":"Supporting Information E.2, Eqs. (S71)-(S72)"},{"comment":"The description of the WLC optimization states that the persistence length is constrained via the passive calculation, but the corresponding text in the main text does not specify that this constraint is a hard equality or how the uncertainty in the passive Lps value affects the reported stretch modulus; a sentence clarifying this would improve reproducibility.","section":"Section 2.3, Figure S6"}],"recommendation":"major_revision","confidential_remarks":"The core methodology is sound and potentially impactful, but the paper currently overstates the validation: several headline agreements are fitted outcomes rather than predictions, and the thermodynamic branch rests on an unverified gradient approximation. The absence of a code/data availability statement is particularly conspicuous given the 'frictionless reproducibility' framing; I would regard its addition as a necessary condition for acceptance. The paper fits the journal's scope, but the authors should be asked to either add out-of-sample validation or substantially soften the predictive claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ryan – send this one to referees. The paper combines DiffTRE with implicit differentiation to fit oxDNA-class potentials to structural, mechanical, and thermodynamic target data, and it works: the optimizations converge, the sensitivity analyses are sensible, and the extension to summary statistics (persistence length, stretch/torsional moduli, melting temperatures) via implicit differentiation is a real step beyond the piecemeal fitting that dominates this field. The JAX implementation of oxDNA is itself a useful resource. Credit where due: the authors are transparent that DiffTRE comes from Zhang et al. and Thaler/Zavadlav, and they show clearly why pathwise and score-function estimators fail here. The conflict-free gradient adaptation for multi-objective fitting is a practical addition.\n\nThe soft spots, in order. First, the thermodynamic gradient: in SI F.1 they explicitly set ∇θB(i)=0 for the thresholded base-pair count and call it 'reasonable.' That is an unverified approximation on a load-bearing branch of the method. States within an energy tolerance of the bond threshold contribute boundary terms, and the paper gives no finite-difference or exact-gradient check to show those terms are small. This is easily fixable and should be fixed before publication; as it stands, the Tm optimization results are not fully supported.\n\nSecond, the validation is partly circular. The paper touts agreement with experimental C, Seff, g, and overwinding, but those are exactly the values the optimizer is told to hit. That demonstrates the fitting machinery, not a predictive model. The same goes for the zinc-finger RMSD optimization, where the target structures are also the training data. Not a flaw in the method, but the framing 'successfully models' overstates.\n\nThird, reproducibility is under-delivered. No code or data is shipped, despite the 'frictionless reproducibility' slogan. Outlier iterations are dropped from plots without comment (twist-stretch coupling, C in Figure S10), and there are no error bars on the fitted properties. For a methods paper, that matters.\n\nOverall, the core estimator is sound, the combination is new, and the paper deserves a serious referee. Whoever handles it should ask for the finite-difference check, the code/data, and a rewritten validation section that treats these as fitting demonstrations, not predictions. This is for people who build or fit coarse-grained nucleic-acid models, and for method developers in differentiable simulation.","headline":"A genuinely useful framework for gradient-based CG force-field fitting, with one unquantified thermodynamic approximation and a validation framing that overstates what is predicted.","tokens_in":33731,"tokens_out":2602,"would_cite":true,"duration_ms":28553,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Gradient descent can replace hand-tuning in coarse-grained force-field fitting.","keywords":["coarse-grained force fields","automatic differentiation","trajectory reweighting","oxDNA model","parameter fitting","molecular dynamics","multi-task optimization","melting temperature"],"falsifier":"Run a duplex melting-temperature optimization and compare the DiffTRE gradient of $T_m$ with respect to the hydrogen-bonding parameters against a central finite-difference estimate of the same gradient; if the difference exceeds the statistical error of the reweighted estimate, the neglected boundary term in $\\nabla_\\theta B(i)$ is not negligible and the thermodynamic optimization is biased.","tokens_in":32743,"feed_emoji":"🧬","tokens_out":6772,"duration_ms":65476,"temperature":0.7,"pith_summary":"The paper establishes a systematic pipeline for fitting coarse-grained molecular force fields to macroscopic experimental data by gradient descent, rather than by hand-tuning. Using the oxDNA DNA model as a testbed, it shows that structural (pitch, propeller twist, structure RMSD), mechanical (persistence length, stretch and torsional moduli, twist-stretch coupling), and thermodynamic (melting temperature and curve width) targets can all be optimized with the same gradient machinery. The method combines differentiable trajectory reweighting, implicit differentiation of line-fitting summary statistics, and conflict-free multi-task gradients so that several properties can be fit simultaneously without one objective degrading another. The stated payoff is frictionless reproducibility: force fields that can be transparently updated as new data arrive, demonstrated also on RNA and on DNA-protein hybrid models.","feed_headline":"Gradient descent now fits force fields to many experiments at once","feed_subtitle":"Trajectory reweighting and automatic differentiation tune structural, mechanical, and thermodynamic targets together.","key_machinery":"The load-bearing object is differentiable trajectory reweighting (DiffTRE), a stochastic gradient estimator that rewrites an expected observable as a weighted sum over reference states sampled under a previous parameter set, with weights equal to ratios of unnormalized Boltzmann probabilities. Because the weights depend smoothly on the parameters while the sampled states do not, gradients flow through energy evaluations only, not through unrolled trajectories, avoiding the memory blow-up and exploding gradients of pathwise differentiable molecular dynamics. Around this core the paper wraps two further mechanisms: implicit differentiation for summary statistics obtained by line fitting (persistence length, moduli, melting temperature), and a conflict-free gradient projector (ConFIG) that maps per-objective gradients to a single update decreasing all objectives simultaneously.","core_discovery":"The central claim is that the obstacle to systematic force-field fitting is not the physics but the gradient: if one can compute low-variance derivatives of any experimentally accessible equilibrium average with respect to model parameters, then structural, mechanical, and thermodynamic data can be folded into a single differentiable objective and minimized with standard optimizers. The paper demonstrates this for a 103-parameter coarse-grained DNA model, recovering or exceeding the original hand-fit accuracy and correcting known failures, such as predicting overwinding under low force where the original model predicted underwinding. The same machinery extends to RNA and to a DNA-protein model with a newly introduced sequence-specific Morse interaction, illustrating that the approach does not depend on the specific energy function.","pith_inferences":["Because DiffTRE needs only unnormalized state probabilities, the same pipeline should extend to atomistic and polarizable force fields, where reference states can be harvested from existing production simulations and reweighted onto a parametric correction.","The recurring cross-stacking sensitivity, if confirmed by exact-gradient checks, implies that the original oxDNA parameterization may be underdetermined along that direction; refitting with the full joint objective could yield a model with a different balance between stacking, hydrogen bonding, and cross-stacking.","A natural testable extension is to replace the thresholded base-pair count in melting calculations with a continuously differentiable soft order parameter, which would remove the unverified gradient approximation and let melting-curve fits be checked against finite differences.","If the workflow matures into a shared library of experimental targets and simulation protocols, the fragmented landscape of bespoke coarse-grained models could give way to a single continuously updated model class, though that would require community standardization that this paper does not itself supply."],"forward_implications":["Any equilibrium observable whose unnormalized probability is known can in principle become a training target, including observables from umbrella sampling, constant-force ensembles, and temperature extrapolation.","Mechanical quantities that require microsecond simulations and downstream fits, including persistence length, torsional modulus, stretch modulus, and twist-stretch coupling, become differentiable end-to-end without differentiating the simulator.","Joint optimization with conflict-free gradients keeps every target monotonically decreasing, removing the trade-off that plagues summed objectives and roughly halving the number of gradient updates in the paper's examples.","Sensitivity analysis of per-parameter gradients becomes a routine diagnostic, and on this model it repeatedly points to the cross-stacking interaction as the most influential term across structural, mechanical, and thermodynamic properties.","The workflow ports to new models: the paper fits RNA to four simultaneous targets and fits DNA-protein complexes to three zinc-finger structures, suggesting the protocol is energy-function-agnostic."],"supporting_citations":[{"why":"Introduced differentiable trajectory reweighting (DiffTRE) for molecular dynamics, the gradient estimator the whole pipeline is built on.","marker":"[29]"},{"why":"Independently derived the same reweighted-gradient estimator for differentiable Monte Carlo at the level of unnormalized distributions.","marker":"[30]"},{"why":"Provides the automatic implicit differentiation used to differentiate line-fitting summary statistics like persistence length and melting temperature.","marker":"[68]"},{"why":"Supplies the ConFIG conflict-free gradient update the paper uses to impose hard constraints in joint optimization.","marker":"[59]"},{"why":"Provides the general conflict-free inverse-gradients formulation from multi-task learning that the paper's joint-update scheme builds on.","marker":"[56]"},{"why":"Defines the oxDNA2 sequence-averaged DNA force field whose 103 parameters serve as the exemplar fitting target.","marker":"[26]"},{"why":"Supplies the stretch-torsion simulation protocol and analytical formulas used to compute the effective stretch modulus, torsional modulus, and twist-stretch coupling.","marker":"[20]"},{"why":"Defines the oxRNA model and the structural and mechanical observable protocols used in the RNA extension.","marker":"[24]"},{"why":"Provides the coarse-grained DNA-protein model that the paper extends with sequence-specific Morse interactions.","marker":"[25]"},{"why":"Defines the original oxDNA potential and the finite-size melting correction on which the thermodynamic calculations rest.","marker":"[69]"}],"fun_headline_variants":["Autodiff unifies force field fitting across many experimental targets","Gradient-based fitting corrects DNA force field and extends to RNA","One differentiable loss fits many experimental observations for molecular models","Low-variance gradients from autodiff enable multi-objective force field fitting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"In the melting-temperature derivation (Supporting Information F.1, after Eq. 90), the paper sets the gradient of the thresholded base-pair count to zero and calls this a reasonable approximation in practice, so the thermodynamic optimization is only valid insofar as states near the bonded/unbonded energy threshold contribute negligibly to melting-temperature gradients.","fun_headline_variants_meta":{"raw":{"variants":["Autodiff unifies force field fitting across many experimental targets","Gradient-based fitting corrects DNA force field and extends to RNA","One differentiable loss fits many experimental observations for molecular models","Low-variance gradients from autodiff enable multi-objective force field fitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":3012,"prompt_tokens":854,"completion_tokens":2158,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":2085}},"tokens_in":470,"tokens_out":2158,"duration_ms":20472,"temperature":1.0,"reasoning_tokens":2085,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:54:40.720151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a duplex melting-temperature optimization and compare the DiffTRE gradient of $T_m$ with respect to the hydrogen-bonding parameters against a central finite-difference estimate of the same gradient; if the difference exceeds the statistical error of the reweighted estimate, the neglected boundary term in $\\nabla_\\theta B(i)$ is not negligible and the thermodynamic optimization is biased.","supporting_citations":[{"cited_title":"Efficient and modular implicit differentiation.Advances in neural information processing systems, 35:5230–5242, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the automatic implicit differentiation used to differentiate line-fitting summary statistics like persistence length and melting temperature."},{"cited_title":"Towards impartial multi-task learning","cited_arxiv_id":null,"evidence_quote":"Provides the general conflict-free inverse-gradients formulation from multi-task learning that the paper's joint-update scheme builds on."},{"cited_title":"Coarse-grained modelling of DNA and DNA self-assembly","cited_arxiv_id":null,"evidence_quote":"Defines the original oxDNA potential and the finite-size melting correction on which the thermodynamic calculations rest."}],"review_version":1}