{"id":"0afef6bb-7c28-48f3-95c3-74623c0a9991","arxiv_id":"2506.09496","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Markov-bridge inverse folding model fine-tuned with energy-based preference pairs and an explicit ΔΔG loss designs lower-energy protein complex sequences while keeping sequence recovery close to state-of-the-art.","lead":"EnerBridge-DPO is a new AI method for designing protein sequences that fold into a given 3D shape while also having low energy, which usually means the protein is more stable. It combines Markov bridge sequence generation with preference optimization, training the model to prefer low-energy sequences and to predict energy changes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Bridge-DPO objective (Eq 7) is an unvalidated L2 proxy for the log-likelihood ratio and samples z_t without conditioning on Y_w/Y_l, so the central claim of energy-guided preference learning is not yet established; the reported energy gains could stem from the supervised energy loss.","rationale":"The reader's weakest assumption identifies the same load-bearing point: the Bridge-DPO loss is an unvalidated approximation. My stress-test sharpens it in two ways: (1) the L2 prediction error is not the negative log-likelihood used in the pretraining objective, so the DPO likelihood-ratio derivation does not carry over; (2) Eq 7 and Algorithm 1 use different sampling distributions for z_t, so even the intended forward-process approximation is internally inconsistent. Both issues touch the core method rather than peripheral evaluation details. The paper's own Appendix B.4 acknowledges the simplification, which strengthens the concern rather than resolving it. I do not see the approximation as a demonstrated fatal flaw: it could be that L2 error is a reasonable proxy in practice, but the paper provides no evidence for that. The absence of code, error bars on the 26-complex energy comparison, and a clearly defined test-set selection process further limit verification, but these are secondary to the methodological gap. Since the reader already returned CONDITIONAL for closely related reasons, my read does not change the verdict; it reinforces the need for a concrete derivation or controlled comparison before the central claim can be accepted.","tokens_in":16446,"tokens_out":3573,"duration_ms":43476,"concrete_test":"Re-derive Eq 7 from the path-level DPO loss in Eq 15 using the actual pretraining objective (cross-entropy, Eq 6) and the conditional forward bridge q(z_t|X,S,Y). If the resulting objective differs materially from Eq 7, implement the corrected DPO loss and retrain on a single complex or a small subset. Then re-evaluate the 26-complex energy table: if the energy advantage over Bridge-IF disappears or shrinks substantially, the L2 proxy is responsible for the reported gains; if it persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty of the paper is that energy-based preferences are injected into a Markov bridge model via Bridge-DPO. The validity of that claim rests on Eq 7, which is not shown to be a faithful DPO objective. The pretraining loss in Eq 6 is a cross-entropy negative log-likelihood, but Eq 7 replaces the log-likelihood ratio with differences of L2 prediction errors ||Y - phi(z_t, S, t)||^2. For a probability-vector output, L2 error is not proportional to negative log-likelihood, so the substitution does not preserve the DPO gradient structure. Appendix B.4 explicitly calls this a 'simplification' and gives no quantitative justification. Separately, Eq 7 samples z_t^w ~ q(z_t^w|X,S) and z_t^l ~ q(z_t^l|X,S), without conditioning on the target sequences Y_w and Y_l, while Algorithm 1 samples q(z_t|X,S,Y_w) and q(z_t|X,S,Y_l). For a discrete Markov bridge, these are different distributions: dropping the endpoint changes the noisy state at each timestep, so the winner/loser errors are evaluated at states that do not correspond to the intended bridge paths. Because the paper's headline energy improvement is attributed to this preference fine-tuning, the unsupported proxy is load-bearing: the measured energy reductions on the 26 complexes could be driven by the explicit energy loss L_energy or by a surrogate objective unrelated to preference optimization. The paper does not provide an ablation separating these effects on the energy metric used in Table 2.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents EnerBridge-DPO, an inverse folding model that fine-tunes a Markov-bridge pretrained model (Bridge-IF style) with a Direct Preference Optimization (DPO) objective using energy-based preference pairs, together with an explicit energy-constraint loss. The authors report that EnerBridge-DPO maintains perplexity and recovery rates comparable to Bridge-IF, produces lower predicted energies (FoldX, Rosetta, BA-Cycle) on 26 selected protein complexes, and achieves Delta-Delta-G prediction performance comparable to BA-DDG. The central methodological novelty is the Bridge-DPO loss in Eq. (7), whose derivation in Appendix B substitutes L2 prediction error for the DPO log-likelihood ratio.","tokens_in":16783,"tokens_out":8039,"duration_ms":83058,"significance":"If the central derivation were valid and the ablations clean, the paper would be a useful step toward energy-aware inverse folding with generative models, with practical value for protein engineering. The paper includes comparisons against many baselines, evaluations with three independent energy functions, and ablation studies. The pretraining stage itself is a reasonable use of Markov bridges. However, the validity of the Bridge-DPO objective and the attribution of the energy improvements to the DPO mechanism are not established by the present evidence; these points are load-bearing for the paper's claimed contribution.","major_comments":[{"comment":"The Bridge-DPO loss in Eq. (7) replaces the DPO log-likelihood ratio with differences of L2 prediction errors ||Y - phi_theta(z_t,t)||^2, while the pretraining objective in Eq. (6) is a cross-entropy negative log-likelihood. For a probability-vector output, L2 error is not proportional to negative log-likelihood, so the substitution does not preserve the DPO gradient structure. Appendix B.4 explicitly calls this a 'simplification' but provides no quantitative or empirical justification. Because the paper's headline energy improvement is attributed to this preference fine-tuning, this unsupported proxy is load-bearing for the paper's central claim.","section":"Section 3.3.2, Eq. (7), Appendix B.4"},{"comment":"Equation (7) samples z_t^w ~ q(z_t^w|X,S) and z_t^l ~ q(z_t^l|X,S), i.e., without conditioning on the target sequences Y_w and Y_l. Algorithm 1 lines 7-8, however, sample z_t^w ~ q(z_t|X,S,Y_w) and z_t^l ~ q(z_t|X,S,Y_l). For a Markov bridge, these are different distributions: dropping the endpoint changes the noisy state distribution at each timestep. Appendix B.4 acknowledges the simplification but does not argue that the resulting loss preserves the intended preference ordering. The measured energy reductions in Table 2 could therefore be an artifact of evaluating winner/loser errors at mismatched bridge states.","section":"Section 3.3.2, Eq. (7), Algorithm 1"},{"comment":"The Delta-Delta-G prediction claim in Section 4.3 is not evidence of emergent physical understanding. Equation (8) uses a learnable k_BT scaling factor and Equation (9) trains the model with an explicit MAE loss against experimental Delta-Delta-G labels. Table 3 therefore largely measures supervised fitting ability. Moreover, the construction of the experimental label for a pair (Y_w,Y_l) in Eq. (9) is not specified: SKEMPI provides per-mutation Delta-Delta-G values, and the paper does not state whether the label is the difference between the two mutants' values or some other quantity. Both the interpretation and the reproducibility of Table 3 need to be addressed.","section":"Section 3.4, Section 4.3, Table 3"},{"comment":"The ablation study in Section 4.4 does not report FoldX, Rosetta, or BA-Cycle energies for the 26 complexes used in Table 2. The w/o DPO and w/o Energy variants are evaluated only on perplexity/recovery and Delta-Delta-G metrics. As a result, the central claim that the DPO preference mechanism, rather than the explicit energy loss L_energy, is responsible for the lower predicted energies in Table 2 is not supported by the presented experiments. An ablation reporting the Table 2 energy metrics for all three variants is needed.","section":"Section 4.4, Tables 4 and 5, Table 2"}],"minor_comments":[{"comment":"Equation (7) writes phi_theta(z_t,t) without the structure argument S, whereas Algorithm 1 and Eq. (6) use phi_theta(z_t,S,t); please make the notation consistent.","section":"Section 3.3.2, Eq. (7)"},{"comment":"The text reports '47,297 million preference pairs' for BindingGym, which appears to be a typo; the intended scale should be clarified.","section":"Appendix D.2"},{"comment":"The ZScore formula is garbled and should be rewritten clearly so that the normalization across models and methods is unambiguous.","section":"Section 4.1.4"},{"comment":"The claim that EnerBridge-DPO produces 'significantly lower' energies is not backed by statistical tests; please report paired significance tests or confidence intervals given the high standard deviations.","section":"Table 2"},{"comment":"The criteria for selecting the 26 protein complexes are not described; please specify the selection protocol or report results on the full test set to rule out cherry-picking.","section":"Section 4.2"},{"comment":"Figure 2 shows only four complexes; it would be helpful to report aggregate structural validity metrics (e.g., TM-score/RMSD distributions) for the designed sequences across all 26 complexes.","section":"Figure 2"},{"comment":"The checkmark notation in the ablation tables is ambiguous; please spell out the exact model configuration for each row.","section":"Tables 4 and 5"},{"comment":"The paper does not include a code or data availability statement; adding one would aid reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central contribution is the Bridge-DPO objective, but the derivation in Appendix B contains an admitted simplification that is not validated. I recommend asking the authors to either provide conditions under which the L2 proxy preserves the DPO gradient structure or to supply an ablation isolating the DPO contribution on the Table 2 energy metrics. The Delta-Delta-G claim should be reframed as supervised regression rather than emergent biophysical understanding. The paper's scope and citation practice are appropriate for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper stacks three known pieces — Bridge-IF, Diffusion-DPO, and BA-DDG — into one inverse folding model that reports lower predicted energies while keeping recovery near Bridge-IF. That specific combination is new relative to the cited literature, and the writing is refreshingly direct. The core idea, starting from a structure-derived prior sequence and fine-tuning with energy-based preferences plus an explicit energy loss, is reasonable. Perplexity and recovery on the MPNN benchmark stay close to Bridge-IF, which is the right sanity check. The ΔΔG numbers on SKEMPI are competitive with BA-DDG, though that is a supervised fit with a learnable kBT scaling factor.\n\nThe soft spots are concentrated in the Bridge-DPO derivation. Appendix B replaces the log-likelihood ratio with L2 prediction error, and that substitution is not justified. The pretraining loss in Eq 6 is cross-entropy, not L2, so the two are not interchangeable for a probability-vector output. That makes Eq 7 a heuristic rather than a faithful DPO objective. There is also a mismatch: Eq 7 samples z_t from q(z_t|X,S), while Algorithm 1 conditions on Y_w/Y_l. Those are different distributions for a Markov bridge, so the winner/loser errors are evaluated at states that do not correspond to the intended bridge paths. This is load-bearing because the energy improvement in Table 2 is attributed to preference fine-tuning, yet the ablations in Tables 4 and 5 do not report energy on the 26-complex test set. The w/o DPO variant would be the control that separates the effect of L_energy from the DPO term. Without that, the energy drop could come from the supervised energy loss alone.\n\nOther issues are real but less severe: the 26 complexes are 'specially selected' with no objective selection rule, Table 2 lacks error bars, and no code or data is provided. These are fixable. The central gap is the unvalidated DPO proxy, which is addressable with a derivation or a careful ablation.\n\nThis is a serious attempt, not a sloppy one. The paper deserves referee time, but the referee should press hard on the DPO derivation and ask for an energy ablation that isolates the preference loss from the energy constraint.","headline":"A sensible integration of bridge inverse folding, DPO, and energy loss, but the central DPO derivation has an unproven L2 proxy and the energy gains are not cleanly attributed to preference learning.","tokens_in":17316,"tokens_out":1602,"would_cite":false,"duration_ms":19017,"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 energy-guided preference fine-tuning of a Markov bridge sequence generator yields low-energy protein designs without sacrificing recovery fidelity.","keywords":["protein inverse folding","Markov bridge","direct preference optimization","energy-guided sequence design","binding free energy prediction","protein stability","protein complex design","generative protein design"],"falsifier":"Compute the exact log-likelihood ratio of the bridge process for a sample of preference pairs and compare it with the squared-error proxy used in the Bridge-DPO loss; if the two disagree in sign on a non-negligible fraction of pairs, the objective is not actually direct preference optimization. A simpler check is to train the model with only the energy-constraint loss and see whether it reproduces the reported energy improvement on the 26-complex benchmark.","tokens_in":16206,"feed_emoji":"🧬","tokens_out":7310,"duration_ms":75475,"temperature":0.7,"pith_summary":"Protein inverse folding asks which amino-acid sequences will fold into a given backbone; most learned solvers optimize sequence recovery and ignore the energetic stability of what they generate. EnerBridge-DPO is proposed as a way to make energy itself part of the objective: a Markov bridge generator is first pretrained to recover native-like sequences from a structure-derived prior, then fine-tuned with a preference loss in which lower-energy sequences are treated as winners and a supervised energy-constraint loss that teaches the model to predict binding free-energy changes. The paper reports that on protein complexes the designed sequences have lower predicted energy than those of all compared models, while recovery and perplexity stay comparable to the strongest baseline, and that the same model predicts $\\Delta\\Delta G$ at a level comparable to a dedicated predictor. A sympathetic reader would take away that inverse folding can be steered toward thermodynamic stability directly, without a separate post-hoc filtering step.","feed_headline":"Energy-guided protein design cuts energy, keeps recovery","feed_subtitle":"Markov bridge plus preference optimization steers sequences toward stable complexes and predicts binding changes.","key_machinery":"The central object is the Bridge-DPO loss, an adaptation of direct preference optimization to the Markov bridge process. The Markov bridge is a generative process that starts from a structure-conditioned prior sequence and iteratively refines it toward a target distribution; the DPO step uses energy-based winner/loser pairs and a frozen reference model, and replaces the intractable path-likelihood ratio with the difference in squared prediction errors between the fine-tuned and reference models. Alongside it, an energy-constraint loss derived from a Boltzmann-aligned inverse folding predictor regresses predicted onto experimental $\\Delta\\Delta G$, forcing the model to learn quantitative energy values. A pretrained structure encoder supplies the informative prior sequence that anchors both stages.","core_discovery":"On its own terms, the paper's central claim is that EnerBridge-DPO simultaneously optimizes sequence-structure compatibility, low predicted energy, and quantitative binding-energy prediction in one inverse folding model. The discovery is that an energy oracle can replace human feedback in direct preference optimization for a Markov bridge generative model: preference pairs are built by ranking candidate sequences by measured or predicted energy, and the DPO objective is adapted so that the bridge model is pushed to assign lower reconstruction error to the low-energy winner than to the higher-energy loser, relative to the frozen pretrained reference. The explicit energy-constraint loss adds regression to experimental $\\Delta\\Delta G$ values, which the paper argues makes the model capture quantitative energy features rather than only relative rankings. The evidence offered is lower predicted energy on 26 protein complexes according to three computational energy predictors, recovery and perplexity on the MPNN benchmark comparable to the leading bridge baseline, and three-fold cross-validated $\\Delta\\Delta G$ predictions comparable to a dedicated predictor.","pith_inferences":["Editorial inference: the reported energy gains are measured with computational energy predictors, not experimental stability; wet-lab validation on a handful of designed complexes would be the decisive next test.","Editorial inference: because the bridge-DPO loss substitutes squared prediction error for the true bridge log-likelihood ratio, it is plausible that similar energy reductions could be achieved by the supervised energy loss alone; an ablation that isolates the preference term on the energy benchmark would clarify the mechanism.","Editorial inference: the energy-as-preference recipe could generalize to other structure-conditioned generative models, since the preference signal is carried by the energy oracle rather than by any bridge-specific property."],"forward_implications":["If EnerBridge-DPO works as reported, protein designers can fold energy optimization into the generative model itself instead of screening or rescoring designs after generation.","The same checkpoint can serve two purposes: produce low-energy sequences and predict $\\Delta\\Delta G$ for mutations, potentially removing the need for a separate predictor.","Because recovery and perplexity stay at the level of the base bridge model, energy alignment does not have to trade away the model's ability to find sequences compatible with the backbone.","The preference-pair construction from mutational datasets suggests that any data source with energy or fitness rankings could be used to steer the generator toward more stable designs.","The approach is in principle agnostic to the energy oracle, so swapping in a more accurate or faster scorer would directly retarget the model."],"supporting_citations":[{"why":"Supplies the Markov bridge framework and simplified pretraining objective that EnerBridge-DPO fine-tunes.","marker":"[11]"},{"why":"Introduces direct preference optimization, whose Bradley-Terry objective is adapted into the Bridge-DPO loss.","marker":"[17]"},{"why":"Shows how DPO is adapted to diffusion models by replacing path likelihoods with prediction errors; the paper transfers this move to Markov bridges.","marker":"[36]"},{"why":"Provides the Boltzmann-aligned inverse folding predictor underlying the energy-constraint loss and the energy evaluation cycle.","marker":"[39]"},{"why":"Supplies the large-scale mutational dataset used to construct winner/loser preference pairs.","marker":"[40]"},{"why":"Supplies the SKEMPI benchmark on which $\\Delta\\Delta G$ prediction is evaluated.","marker":"[41]"},{"why":"Provides one of the computational energy functions used to compare designed sequences.","marker":"[54]"},{"why":"Provides another computational energy function used to compare designed sequences.","marker":"[55]"}],"fun_headline_variants":["Energy preferences steer protein design to lower-energy sequences","Markov bridge plus energy-guided DPO yields stable protein sequences","Inverse folding with energy constraints cuts predicted energy, keeps recovery","Energy-aware preference optimization improves protein stability prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a model's squared prediction error correctly represents how likely it is to generate a sequence, so that ranking sequences by these errors is the same as ranking them by preference; if that substitution fails, the claimed preference optimization may reduce to the supervised energy loss.","fun_headline_variants_meta":{"raw":{"variants":["Energy preferences steer protein design to lower-energy sequences","Markov bridge plus energy-guided DPO yields stable protein sequences","Inverse folding with energy constraints cuts predicted energy, keeps recovery","Energy-aware preference optimization improves protein stability prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1328,"prompt_tokens":956,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":308}},"tokens_in":572,"tokens_out":372,"duration_ms":4823,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:46:57.304847+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact log-likelihood ratio of the bridge process for a sample of preference pairs and compare it with the squared-error proxy used in the Bridge-DPO loss; if the two disagree in sign on a non-negligible fraction of pairs, the objective is not actually direct preference optimization. A simpler check is to train the model with only the energy-constraint loss and see whether it reproduces the reported energy improvement on the 26-complex benchmark.","supporting_citations":[{"cited_title":"Bridge-IF: Learning Inverse Protein Folding with Markov Bridges","cited_arxiv_id":"2411.02120","evidence_quote":"Supplies the Markov bridge framework and simplified pretraining objective that EnerBridge-DPO fine-tunes."},{"cited_title":"Direct preference opti- mization: Your language model is secretly a reward model,","cited_arxiv_id":null,"evidence_quote":"Introduces direct preference optimization, whose Bradley-Terry objective is adapted into the Bridge-DPO loss."},{"cited_title":"Diffusion model alignment using direct preference optimization,","cited_arxiv_id":null,"evidence_quote":"Shows how DPO is adapted to diffusion models by replacing path likelihoods with prediction errors; the paper transfers this move to Markov bridges."},{"cited_title":"Bindinggym: A large-scale mutational dataset toward deciphering protein-protein interactions,","cited_arxiv_id":null,"evidence_quote":"Supplies the large-scale mutational dataset used to construct winner/loser preference pairs."},{"cited_title":"Skempi 2.0: an updated benchmark of changes in protein–protein binding energy, kinetics and thermodynamics upon mutation,","cited_arxiv_id":null,"evidence_quote":"Supplies the SKEMPI benchmark on which $\\Delta\\Delta G$ prediction is evaluated."},{"cited_title":"Foldx 5.0: working with rna, small molecules and a new graphical interface,","cited_arxiv_id":null,"evidence_quote":"Provides one of the computational energy functions used to compare designed sequences."},{"cited_title":"The rosetta all-atom energy function for macromolecular modeling and design,","cited_arxiv_id":null,"evidence_quote":"Provides another computational energy function used to compare designed sequences."}],"review_version":1}