{"id":"f1c8d076-014e-4e77-ad1a-3e8af7085168","arxiv_id":"2508.03619","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FlowBack-Adjoint fine-tunes a flow-matching backmapping model with molecular mechanics energy gradients, reducing clashes and bond errors and producing lower-energy all-atom protein reconstructions.","lead":"A new machine learning tool rebuilds all-atom protein structures from coarse-grained C-alpha traces while using physics-based energy and force feedback to produce more realistic, lower-energy configurations. It could make molecular dynamics simulations of proteins more practical by generating starting structures that need less energy relaxation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adjoint-matching target p_phi = p_theta e^{-lambda U} is never verified, and the actual Euler-discretized, auxiliary-field-augmented implementation does not provably sample it; the advertised Boltzmann-tilt interpretation is therefore unsupported.","rationale":"The reader identified the same load-bearing weakest assumption: the theoretical guarantee of the adjoint correction depends on modeling assumptions that the implemented FlowBack-Adjoint pipeline does not satisfy, and the paper never verifies the implied target distribution. I agree, and I add two concrete aggravations: (i) training uses the stochastic Eq. 1 while inference uses the deterministic Euler map of Algorithm 2, so even a correct control for the SDE is not shown to produce p_phi under the inference procedure; and (ii) the energy-based evaluation is circular in that CHARMM27 is both the reward used to train the energy bias and the metric used to score the result. The paper is otherwise careful: it includes a held-out DESRES pair, a 23-protein OOD suite with sequence-identity screens, an ablation against FlowBack+LJ/BONDS, and explicit discussion of its known limitations, including the chirality-correction pathology and gas-phase energy evaluation. These strengths support a conditional rather than a reject verdict: the method plausibly works as an empirical energy-lowering fine-tune, but the specific claim of Boltzmann-like reweighting by adjoint matching needs an independent distributional test before it can be accepted. Since the reader already returned CONDITIONAL on essentially this basis, my stress-test does not change the verdict.","tokens_in":25996,"tokens_out":6944,"duration_ms":95598,"concrete_test":"On the WW domain at sigma_p = 0.003 nm, draw M = 10^4 configurations from the base flow v_theta (FlowBack+LJ/BONDS) and compute log p_theta(x) for each via the flow's instantaneous change-of-variables along the same 100 Euler steps, as permitted by FlowBack's exact-likelihood claim. Form self-normalized importance weights w(x) proportional to exp(-lambda CHARMM27_U(x)), using the same lambda = 0.01 and energy function as in training. Compare the energy, bond-score, and clash-score distributions of this importance-reweighted base ensemble with direct FlowBack-Adjoint samples using bootstrap uncertainties. If the two disagree beyond sampling error, the actual Euler plus auxiliary-field implementation does not realize the claimed tilt p_phi = p_theta e^{-lambda U}; if they agree, the unbiasedness claim survives this check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that FlowBack-Adjoint samples the reweighted density p_phi(x) = p_theta(x)e^{-lambda U(x)} (Section 2.3). In the adjoint-matching framework, this unbiasedness guarantee is conditional on the memoryless SDE of Eq. 1, a zero-mean Gaussian base, and a base drift v_theta satisfying the associated continuity equation. In practice, v_theta is not the clean learned FlowBack drift: Section 2.2 adds hand-designed chirality, Lennard-Jones, and bond-length velocity fields with time gates, and training minimizes the Euler-discretized residual of Eq. 3 over 100 steps. More importantly, training (Algorithm 1) simulates the stochastic SDE of Eq. 1, while inference (Algorithm 2) uses a purely deterministic Euler map, x_t <- x_t + v_phi Delta_t. A control learned for the stochastic process is not shown to steer the deterministic ODE to p_phi, and no independent check of the output distribution against p_theta e^{-lambda U} is reported. The KL-divergence comparison to DESRES data is not such a check, because the reference uses CHARMM22* while the reward and evaluation both use CHARMM27, and the energy metric is the same objective being optimized. The paper's own ablation (Figure A.8) further shows that the residual energy gap is dominated by the ad hoc chirality correction, weakening the attribution of the observed improvements to adjoint matching itself. The empirical gains are real, but the 'physics-aware energy reweighting' mechanism is currently an interpretation rather than a verified property.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FlowBack-Adjoint augments the pretrained FlowBack flow-matching backmapping model with three auxiliary physics-based velocity fields (chirality, Lennard-Jones, bond-length) and a CHARMM27 energy-gradient adjoint-matching fine-tuning step. The authors evaluate the model on two held-out DESRES mini-protein trajectories and a 23-protein BIOEMU out-of-distribution set, reporting reduced bond-length errors, reduced clashes, lower CHARMM27 single-point energies, improved KL divergence to the MD energy distribution, and improved success rates of MD simulations launched without energy minimization.","tokens_in":26449,"tokens_out":5040,"duration_ms":59107,"significance":"The empirical gains are substantial and the paper includes several commendable practices: explicit out-of-distribution testing with homology-filtered sequences, a modular comparison within the BIOEMU pipeline, an honest ablation of the chirality correction, and an open statement of limitations (gas-phase energies, hydrogen placement). If the reweighting claim were verified, the method would be a lightweight and practical way to inject physical knowledge into pretrained generative models. However, the central theoretical claim about sampling a Boltzmann-tilted distribution is not verified for the actual implementation, and the MD-stability headline is weaker than stated. These gaps are fixable and should be addressed before publication.","major_comments":[{"comment":"The paper claims that adjoint matching steers the generator to sample p_phi(x) = p_theta(x) exp(-lambda U(x)). The unbiasedness guarantee of the lean adjoint ODE requires the memoryless SDE (Eq. 1) with a drift v_theta satisfying the associated continuity equation, and the same stochastic dynamics at inference. In practice, v_theta is the sum of the learned FlowBack drift and three hand-designed auxiliary fields (Section 2.2), training is performed on the stochastic SDE (Algorithm 1), and inference uses a deterministic Euler map (Algorithm 2: x <- x + v_phi Delta t). No test is reported that the output distribution equals p_theta exp(-lambda U), e.g., via reweighting or a separate sampler. Table 1's KL divergence to the DESRES energy distribution is not such a test because the reference frames are evaluated with the same CHARMM27 force field used as the reward. The 'energy-guided reweighting' is therefore currently an interpretation rather than an established property. Please verify the target distribution or revise the theoretical claims accordingly.","section":"Section 2.3, Eq. (2)-(3); Algorithms 1-2"},{"comment":"The abstract's claim that FlowBack-Adjoint produces configurations capable of initializing stable MD 'without requiring energy relaxation' is overstated: the reported stable-run fractions are 82.6% (WW domain) and 65.4% (Protein B) for 20 ps gas-phase runs, and values near 100% are obtained only after removing hydrogen/termini forces or using a united-atom treatment (Section 3.1 and Figure A.5). Please qualify the abstract and conclusions with the actual success rates.","section":"Section 3.1 / Test 5 and Abstract"},{"comment":"The energy-divergence metric is partially circular: the reward R = -lambda U uses CHARMM27, and both the generated and reference ensembles are evaluated with CHARMM27. The KL improvement therefore reflects, in part, the model optimizing the same energy function used for evaluation. The bond, clash, and diversity metrics plus the MD-stability tests provide independent evidence, and the manuscript should explicitly acknowledge this circularity and, ideally, include an energy comparison with an independent force field.","section":"Section 3.1 / Test 4 and Section 2.4.2"}],"minor_comments":[{"comment":"The phrase 'Training of the learned vector field vector field v_gamma' contains a duplicated word; please remove the repetition.","section":"Section 2.1"},{"comment":"The sentence 'we inserted our integrator into the BIOEMU pipeline' is misleading because FlowBack-Adjoint is a generative model, not an integrator; consider rephrasing to 'we inserted our backmapping model into the BIOEMU pipeline'.","section":"Section 3.2"},{"comment":"The description of the periodic box ('expanded by 2 nm in each direction') would benefit from stating the initial box dimensions or the protocol used to define them, to make the energy calculations reproducible.","section":"Section 2.4.2"},{"comment":"The scalar metrics in Table 1 and the force statistics in Figure 2 are reported without uncertainties or repeat counts; providing standard errors or confidence intervals would strengthen the quantitative comparisons.","section":"Table 1 and Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The main barrier to acceptance is the gap between the theoretical reweighting claim and the implementation: the unbiasedness guarantee is not verified for the Euler-discretized, auxiliary-field-augmented flow used at inference. Adding a direct distributional check (or softening the claims) is essential. The other issues, including the abstract's MD-stability overstatement and the circularity of the energy metric, are addressable with revision. The paper is otherwise well within the scope of physics.chem-ph and makes a useful practical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: FlowBack-Adjoint delivers real, measurable improvements in backmapped structures, and the paper is unusually honest about where it falls short. But the central theoretical claim, that adjoint matching reweights the generator to p_phi = p_theta e^{-lambda U}, is not verified, and the actual implementation (auxiliary velocity fields, 100 Euler steps, deterministic inference) does not provably sample that target. Read it as a solid empirical fine-tuning paper whose mechanism is partly interpretation.\n\nWhat is actually new: applying adjoint matching with a molecular mechanics reward to a conditional flow-matching backmapper, plus the time-gated LJ, bond, and chirality velocity fields. The evaluation is thorough: bond scores, clash counts, diversity, energy distributions, MD stability, and a 23-protein out-of-distribution test. The headline numbers are large but credible, with bond errors down 92% and clashes to zero on the DESRES holdouts at sigma_p = 0.003, and the KL divergence down from 0.1397 to 0.0018 on the WW domain. The Figure A.8 ablation is a model of self-criticism: it shows that the chirality correction, not the adjoint matching, is responsible for most of the residual energy gap. That weakens the attribution but also earns trust.\n\nThe soft spots, in order of severity. First, the Boltzmann-tilt claim lacks a direct check. Training uses a stochastic SDE, inference uses a deterministic Euler map, and v_theta includes hand-designed physics terms; the paper never compares samples from the implied target against p_theta e^{-lambda U} from an independent sampler. The KL comparison to DESRES is not that check, since the reference uses CHARMM22* while the reward is CHARMM27. Second, the main energy metric and the reward are both CHARMM27, so part of the improvement is self-consistency; the structural and MD-stability results are independent, so this is a moderate concern, not fatal. Third, the abstract's 'stable MD without relaxation' overstates Test 5: only 82.6% and 65.4% of 20 ps runs were stable for WW domain and Protein B. The paper reports these numbers honestly, but the abstract is too strong. Fourth, no code and no error bars make robustness hard to judge.\n\nWho it is for: anyone building generative backmapping or using flows for molecule generation. It deserves a serious referee. The authors should be asked to re-frame the mechanism as an empirical energy-guided fine-tune, verify or temper the reweighting claim, and soften the abstract. I would send it out.\n\nRecommendation: accept for peer review, with revision.","headline":"Real, well-measured improvements in backmapping, but the energy-reweighting story is not verified as claimed.","tokens_in":26896,"tokens_out":4116,"would_cite":true,"duration_ms":46996,"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":"A post-training physics pass makes FlowBack generate low-energy, clash-free protein structures without retraining from scratch, producing configurations that can directly initialize stable molecular dynamics simulations.","keywords":["adjoint matching","flow matching","protein backmapping","all-atom structure generation","molecular dynamics initialization","coarse-grained models"],"falsifier":"Run the trained FlowBack-Adjoint model on a set of C-alpha traces drawn from a well-characterized molecular dynamics ensemble and compare the energies of the generated structures against an independent Markov-chain Monte Carlo or Hamiltonian Monte Carlo sampler using the same force field; if the generated energy histogram systematically differs from the reference Boltzmann distribution by more than the reported KL divergences, the claimed unbiased reweighting fails.","tokens_in":25802,"feed_emoji":"🔬","tokens_out":2987,"duration_ms":30242,"temperature":0.7,"pith_summary":"This paper claims that a lightweight, post-training adjoint-matching correction can teach a purely structure-trained generative backmapping model to produce physically plausible protein conformations. The central claim is that FlowBack-Adjoint tilts the generated all-atom ensemble toward low-energy configurations, cutting single-point energies by roughly 78 kcal/mol.residue, reducing bond-length errors by over 92%, and eliminating more than 98% of steric clashes, all while preserving the diversity of the original model and producing structures that can launch stable MD runs without energy relaxation. The paper positions this as a general recipe: keep the pretrained flow, add energy and force information only during fine-tuning, and leave inference cost essentially unchanged.","feed_headline":"A light post-training pass steers protein generation toward low-energy, clash-free…","feed_subtitle":"FlowBack-Adjoint cuts energy by ~78 kcal/mol.residue and lets generated proteins launch stable MD runs without relaxation.","key_machinery":"The central machinery is adjoint matching: an adjoint state $a_t$ tracks how a perturbation of the generated configuration would change the terminal energy reward, and a control force $u_t(x) = -\\frac{1}{2}\\sigma_t(t)^2 a_t$ is added to the flow velocity to bias the generated distribution toward $e^{-\\lambda U}$. The method requires a memoryless diffusion noise schedule $\\sigma_t^2(t) = 2(1-t)/t$ with a zero-mean Gaussian base distribution, and assumes the base flow satisfies the continuity equation, so that the derived lean adjoint ODE gives an unbiased tilt toward the reward-weighted distribution. Three auxiliary velocity fields (chirality correction around $t \\geq 0.25$, Lennard-Jones repulsion for $t \\geq 0.85$, and harmonic bond-length restraints for $t \\geq 0.95$) act as physics-based inductive biases that reduce the burden on the energy gradients.","core_discovery":"The paper's central discovery is that a pretrained conditional flow-matching model can be improved by adjoint matching to sample from a reward-weighted distribution $p_\\phi(x) \\propto p_\\theta(x)e^{-\\lambda U(x)}$, where $U$ is a molecular mechanics force-field energy and $\\lambda$ a tuning parameter. Gradients of the energy are backpropagated through the flow trajectory using a lean adjoint ODE, and the resulting control force is added to the learned velocity, producing a fine-tuned model that generates lower-energy configurations while preserving the structural accuracy and diversity learned from data. The paper reports median energy reductions of $\\sim$78 kcal/mol.residue against FlowBack, >92% reduction in bond-length errors, >98% elimination of clashes, strong retention of configurational diversity, and stability of the generated structures when directly used to launch all-atom MD simulations.","pith_inferences":["The paper implicitly suggests that any differentiable energy function can serve as the reward; replacing the gas-phase CHARMM27 potential with an implicit-solvent or machine-learned potential could be a direct test of the method's modularity.","The strong residual energy tail reported in the results could be further reduced by replacing the chirality-correction velocity with a chirality-aware architecture, since the paper's ablation attributes most of the remaining mismatch to that correction.","If adjoint matching reliably tilts toward $e^{-\\lambda U}$, the hyperparameter $\\lambda$ could be tuned to target a specific temperature, moving the generator from \"low-energy structures\" toward an explicit Boltzmann ensemble.","The method's dependence on a memoryless noise schedule implies it may not transfer unchanged to diffusion models with different noise schedules without re-deriving the adjoint dynamics."],"forward_implications":["Generated all-atom ensembles can be closer to a Boltzmann-like distribution, enabling more faithful downstream calculations of binding energies and other thermodynamic quantities.","Backmapped structures can directly initialize stable MD simulations without prior energy minimization, removing a costly and error-prone preprocessing step in coarse-grained workflows.","The post-training energy-guided tilt preserves configurational diversity even at larger prior noise values, breaking the usual accuracy-diversity trade-off.","The adjoint-matching recipe is applicable to other pretrained flow or diffusion generators, since it only requires a differentiable reward and its gradient.","Swapping the CHARMM27 energy for a general machine-learned potential could extend the method beyond proteins to nucleic acids, lipids, and synthetic polymers."],"supporting_citations":[{"why":"Introduces adjoint matching with memoryless stochastic optimal control, providing the theoretical guarantee that the tilt converges to the reward-weighted distribution.","marker":"[Domingo-Enrich et al., 2025]"},{"why":"Defines FlowBack, the pretrained conditional flow-matching backmapping model whose vector field serves as the base distribution $p_\\theta$.","marker":"[Jones et al., 2025]"},{"why":"Supplies CHARMM27, the force field whose energy and gradients define the reward and the Lennard-Jones and bond-length corrections.","marker":"[MacKerell Jr. et al., 2000]"},{"why":"Provides the DESRES fast-folding mini-protein MD trajectories used for adjoint-matching training and the reference ensembles for energy comparison.","marker":"[Lindorff-Larsen et al., 2011]"},{"why":"Supplies the BIOEMU workflow and OOD60 benchmark used for out-of-distribution generalization tests.","marker":"[Lewis et al., 2024]"}],"fun_headline_variants":["Physics-aware adjoint tune cuts protein energy and clashes","Adjoint matching steers flow-matching to low-energy, clash-free proteins","Energy-guided tune slashes protein clashes and preps for stable MD","FlowBack-Adjoint: energy-guided post-training cuts clashes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's guarantee that the corrected flow samples the energy-reweighted distribution depends on the base flow being exactly a memoryless diffusion with a zero-mean Gaussian prior, and on the learned drift satisfying the continuity equation; in practice the flow is approximated with hand-added corrections and 100 explicit Euler steps, so the unbiasedness may not hold exactly.","fun_headline_variants_meta":{"raw":{"variants":["Physics-aware adjoint tune cuts protein energy and clashes","Adjoint matching steers flow-matching to low-energy, clash-free proteins","Energy-guided tune slashes protein clashes and preps for stable MD","FlowBack-Adjoint: energy-guided post-training cuts clashes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001032,"raw_usage":{"total_tokens":4393,"prompt_tokens":1035,"completion_tokens":3358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":3286}},"tokens_in":651,"tokens_out":3358,"duration_ms":28146,"temperature":1.0,"reasoning_tokens":3286,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:19:22.630913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained FlowBack-Adjoint model on a set of C-alpha traces drawn from a well-characterized molecular dynamics ensemble and compare the energies of the generated structures against an independent Markov-chain Monte Carlo or Hamiltonian Monte Carlo sampler using the same force field; if the generated energy histogram systematically differs from the reference Boltzmann distribution by more than the reported KL divergences, the claimed unbiased reweighting fails.","supporting_citations":[],"review_version":1}