{"id":"bf482129-e935-4af2-8be5-2882f7ca0593","arxiv_id":"2505.12358","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AbFlowNet trains a diffusion-based antibody CDR designer with a GFlowNet Trajectory Balance term so that sampled CDRs are rewarded for lower Rosetta-estimated binding energy, improving energy and reconstruction metrics over DiffAb.","lead":"This paper fuses diffusion models with GFlowNet training to design antibody CDR loops that score better on a Rosetta binding-energy estimate, without requiring online reinforcement learning on the test set. The method improves sequence recovery, structural RMSD, and energy-based Top-1 metrics over its diffusion baseline in the reported experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rosetta reward and Rosetta evaluation are the same quantity, so the reported binding-energy gain is not evidence of true affinity improvement.","rationale":"I read AbFlowNet as a methods paper: GFlowNet Trajectory Balance is used to reweight a diffusion model toward reference CDRs with favorable precomputed Rosetta energies, and the experiments compare against DiffAb under a matched gradient budget. The most important empirical claim is the energy improvement; the AAR/RMSD improvements are valuable but secondary. The weakest point is that the reward used in training and the energy metric used in Table 1 come from the same InterfaceAnalyzer estimator. Because the paper itself cites evidence that Rosetta energies correlate only moderately with true binding affinity, an improvement on that estimator is compatible with no real affinity improvement. This is especially concerning because the reward is a deterministic transform of the evaluation score, so the model is being trained on the same estimator used to score the output. The reader's weakest_assumption identified exactly this loop, and I agree. I also considered the reversed conditioning in Eq. 17 (p(St|St-1) is not the defined denoising transition p(St-1|St)) and the N=100-vs-N=2528 comparison to AbDPO in Table 1; both are real issues, but Eq. 17 may be a transcription error that code can resolve, and the AbDPO comparison is not the paper's central claim. The Rosetta loop, by contrast, cannot be resolved without independent measurement. Because the reader already made the verdict conditional on this concern, my read does not change the verdict.","tokens_in":16212,"tokens_out":14525,"duration_ms":156044,"concrete_test":"Using the released code, generate 100 CDR-H3s per RAbD complex from AbFlowNet and from DiffAb with identical sampling and side-chain packing, then re-score the same generated complexes with an independent binding-energy predictor not used in training, e.g., PRODIGY or FoldX (or OpenMM Yank on a 10-complex subset). Compare median DeltaG and IMP under the independent predictor; if AbFlowNet no longer outperforms DiffAb, the energy claim is an artifact of the Rosetta reward/evaluation loop.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step for the central energy claim is the identity between training reward and evaluation metric. Equation 17 defines the terminal reward as R(S0)=exp(-alpha*BindingEnergy(S0)), and Section 5 computes binding energy by Rosetta InterfaceAnalyzer; Table 1's CDR Etotal and CDR-Ag DeltaG are also InterfaceAnalyzer outputs. Section 2 states such estimators have only 'moderate correlation with the real binding energy', and Appendix E calls InterfaceAnalyzer 'an unreliable estimator of binding energy'. Thus the reported 24.8%/38.1% Top-1 improvements show that AbFlowNet scores better under the same surrogate it was trained to maximize, not that it produces tighter-binding antibodies. The AAR and RMSD gains are independent of this issue and support the reconstruction claims, but they do not validate the energy headline. This is a correctness risk for the paper's central contribution, not merely a missing baseline comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AbFlowNet, a CDR design model that adds a GFlowNet Trajectory Balance objective to the DiffAb diffusion backbone, using a precomputed terminal reward exp(-alpha * BindingEnergy(S0)) based on Rosetta InterfaceAnalyzer energies of the training complexes. Experiments on RAbD and a 19-complex DiffAb test set report average improvements of +3.06% amino acid recovery, +20.40% RMSD, and +3.60% IMP over DiffAb, together with Top-1 CDR total-energy and binding-energy reductions of 24.8% and 38.1% at a sampling budget of N=100. The authors release code and model weights and emphasize that AbFlowNet avoids online RL and test-set pseudo-labeling.","tokens_in":16304,"tokens_out":5941,"duration_ms":60307,"significance":"If the structural results are reliable, the article makes a modest but useful contribution: it demonstrates that a trajectory-balance term can be appended to a diffusion antibody model without sacrificing reconstruction quality, while improving a Rosetta-based energy surrogate. The release of code and weights, the matched gradient-update comparison with DiffAb, and the explicit reporting of hyperparameter sweeps are strengths. However, the binding-energy headline is not supported as a claim about true affinity, because the training reward and the evaluation metric are the same Rosetta InterfaceAnalyzer quantity, and the manuscript itself identifies this estimator as unreliable. The AAR and RMSD gains are independent of that issue and support the reconstruction claims, but they do not validate the energy claim.","major_comments":[{"comment":"The terminal reward in Eq. (17) is R(S0) = exp(-alpha * BindingEnergy(S0)) with BindingEnergy computed by Rosetta InterfaceAnalyzer, and the Top-1 CDR Etotal and CDR-Ag Delta G metrics in Table 1 are the same InterfaceAnalyzer outputs. Since Section 2 and Appendix E state that InterfaceAnalyzer is an unreliable estimator of binding energy, the reported 24.8% and 38.1% improvements show that AbFlowNet optimizes the surrogate it was trained on, not that it produces tighter-binding antibodies. I request a concrete guard: evaluate on the subset of complexes with experimental affinity data, or use an independent estimator, or at minimum reframe the headline as 'Rosetta InterfaceAnalyzer energy' and remove the affinity implications from the abstract and conclusion.","section":"Section 5, Section 6.1, Eq. (17)"},{"comment":"The trajectory-balance weight w = 5e-6 in Eq. (18) is selected by evaluating on the RAbD dataset, which is the same test set used for all main results in Table 1 and Table 2. This makes the central comparison vulnerable to test-set peeking during hyperparameter selection. Please choose w on a held-out validation split derived from the SAbDab training set, then report final evaluation on RAbD and the DiffAb test set only for the fixed configuration.","section":"Section A.2, Figure 4"},{"comment":"Equation (17) writes log[ Z_theta * prod_t p(St | St-1; theta) / (R(S0) * prod_t q(St-1 | St)) ]^2, but Eqs. (13)-(16) define q(St | St-1) as the forward noising process and p(St-1 | St) as the reverse denoising process, so the TB-loss arguments appear reversed and p(St | St-1; theta) is not defined by Eq. (14). In addition, Section 7 states that backpropagation is performed for only one random timestep, which deviates from the full Trajectory Balance objective and makes the implemented loss an approximation. Because the GFlowNet objective is the core novelty, please correct the indexing and analyze or quantify the effect of the single-step gradient approximation on reward propagation.","section":"Section 4.2, Eqs. (13)-(17), Section 7"},{"comment":"The claim that AbFlowNet is 'competitive with AbDPO' is based on an apples-to-oranges comparison: AbDPO is evaluated at N=2,528 while AbFlowNet is evaluated at N=100, and Top-1 metrics are explicitly sampling-budget-sensitive, as the paper itself notes for DiffAb. The comparison against DiffAb at N=100 is fair, but the AbDPO competitiveness claim should either be removed, or supported by running AbFlowNet at the same N=2,528 budget on at least a subset of the test complexes.","section":"Table 1, Section 6.1"}],"minor_comments":[{"comment":"The phrase 'online Reinforcement Learning (RL) pipelines rely heavily on unreliable binding energy estimators' is a grammatical fragment; also, 'ABFlowNet' in the abstract should be 'AbFlowNet' for consistency.","section":"Abstract"},{"comment":"The caption says the generated CDRs were 'selected the one with the highest Delta G', but selecting the highest Delta G would pick the least favorable binding-energy value; this should read 'lowest Delta G' if the displayed values are the best energies.","section":"Figure 3 caption"},{"comment":"The column headers 'CDR+Etotal' and 'CDR-Ag+Delta G' are confusing, and the percent reductions 89.5% and 149.7% are unexplained; please clarify the normalization and the direction of improvement.","section":"Table 4"},{"comment":"The sentence 'we compute all forward q(St-1 | St) and backward probabilities p(St | St-1; theta)' uses notation reversed with respect to Eqs. (13)-(16); please make the time indexing consistent throughout Section 4.","section":"Section 4.2"},{"comment":"The text says AbFlowNet 'outperforms DiffAb in all three metrics', but CDR-H3 IMP (14.38 vs 12.65) and CDR-L3 IMP (36.98 vs 34.70) both decrease; the claim should be qualified as 'on average across the six CDR regions'.","section":"Section 6.2, Table 2"},{"comment":"References [2] and [3] are duplicate entries for the same RAbD paper; please deduplicate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the reward-evaluation identity: the paper's headline energy improvements are measured with the same Rosetta estimator used as the training reward, and the authors themselves acknowledge this estimator is unreliable. This is fixable by reframing the claims and adding a guard experiment, so I do not recommend rejection, but the current abstract and conclusion overstate what is shown. The structural improvements are the most credible contribution and should be foregrounded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: legitimate new application, honest reporting, but the energy headline overstates what is shown.\n\nThe new thing here is transferring GFlowNet Trajectory Balance from image diffusion alignment to antibody CDR co-design, using a precomputed Rosetta energy reward. That is a real contribution. The training scheme is simple and the code and weights are public. The authors also tried Detailed Balance, hit a wall, and reported the negative result. That transparency counts.\n\nTable 2 is the strongest part: AbFlowNet beats DiffAb on AAR, RMSD, and IMP under the same training budget, and those metrics don't depend on the energy surrogate. The RMSD gain on L3 is dramatic; I'd want a few seeds before trusting it, and the paper only mentions one reseed in the appendix.\n\nThe main weakness is the energy claim. The training reward and the evaluation metric are both Rosetta InterfaceAnalyzer, which the paper itself calls unreliable. So the 24.8% and 38.1% improvements show the model optimizes that surrogate, not that it makes better binders. The paper is upfront about this in Section 7 and Appendix E, but the abstract still says 'binding energy errors.' That's misleading and should be fixed.\n\nMinor issues: the AbDPO comparison mixes sampling budgets, but they label it as reference only. The one-timestep backprop approximation to TB is disclosed. The TB weight was tuned; the sweep is there but no error bars in the main tables.\n\nThis is a paper for the antibody-design community. It deserves peer review. I'd recommend major/minor revision asking for reframing of the energy metric, seed variance, and ideally a small experimental affinity check.","headline":"Solid new application of GFlowNet TB to diffusion-based CDR design, with honest reporting; the energy improvements are real only for the Rosetta surrogate, not for true binding affinity.","tokens_in":16969,"tokens_out":3782,"would_cite":true,"duration_ms":36057,"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":"AbFlowNet shows that a diffusion model for antibody CDR design can be trained to optimize binding energy by adding a GFlowNet Trajectory Balance objective whose terminal reward is a precomputed Rosetta energy score.","keywords":["antibody design","CDR","diffusion models","GFlowNet","Trajectory Balance","binding energy","Rosetta","protein generation"],"falsifier":"Retrain AbFlowNet under the identical pipeline but with the terminal reward replaced by experimental in vitro binding-affinity measurements (or by an independently validated energy function) for the training complexes, then compare generated CDRs against DiffAb on both Rosetta energy and measured affinity. If the reported gains vanish when the reward is decoupled from the evaluation metric, the binding-energy improvement is an artifact of reward–metric coupling.","tokens_in":15912,"feed_emoji":"🧬","tokens_out":7296,"duration_ms":62582,"temperature":0.7,"pith_summary":"AbFlowNet aims to show that antibody CDR design can be made to optimize binding energy directly, without the expensive online reinforcement-learning stage used by prior methods. It fuses a diffusion model for CDR sequence and structure with a GFlowNet, treating each denoising step as a state and using the Trajectory Balance objective to propagate a terminal reward—the Rosetta-computed binding energy of the reference CDR, precomputed once per training complex—back through the whole trajectory. On the RAbD test set, averaged over all six CDR regions, AbFlowNet improves amino acid recovery by 3.06%, backbone RMSD by 20.40%, and the fraction of CDRs with better-than-reference binding energy by 3.60% relative to the same diffusion baseline at the same training budget. At 100 samples per complex, its Top-1 total energy and binding energy improve by 24.8% and 38.1% over the baseline, without using test-set structures during training. A sympathetic reading is that the paper thereby unifies reconstruction and reward optimization in one training procedure.","feed_headline":"Diffusion plus GFlowNet cuts antibody binding energy 38%","feed_subtitle":"A Trajectory Balance reward built into denoising improves CDR design without online RL or test-set labels.","key_machinery":"The load-bearing mechanism is the GFlowNet Trajectory Balance objective applied to the diffusion trajectory. In its standard form it enforces $Z_\\theta \\prod_t P_F(s_t|s_{t-1}) = R(s_n)\\prod_t P_B(s_{t-1}|s_t)$, where $Z_\\theta$ is a learned estimate of the initial-state flow, $P_F$ is the diffusion model's denoiser, $P_B$ is the fixed forward noising process, and $R(s_n)$ is the exponential Rosetta binding-energy reward. Applied to CDR design, each partially denoised CDR is a state and the per-residue transition probabilities factor across amino-acid type, 3D coordinate, and SO(3) orientation. The objective propagates the terminal energy reward back through all 100 denoising steps while the diffusion reconstruction losses preserve fidelity to the training distribution.","core_discovery":"On its own terms, the paper's central claim is that a diffusion model for antibody CDR design can be trained to optimize binding energy by adding a GFlowNet Trajectory Balance term to the standard denoising losses. The GFlowNet state is a partially denoised CDR; the forward edge flow is the product of the learned denoising probabilities over all CDR residues, the backward edge flow is the product of the fixed noising probabilities, and the terminal state's flow is $\\exp(-\\alpha \\cdot \\text{BindingEnergy}(S_0))$ precomputed from Rosetta InterfaceAnalyzer for each training CDR. Enforcing parity between forward and backward trajectory flows makes the sparse terminal reward influence every denoising step. The paper reports that, with identical hyperparameters and gradient-update counts, this joint objective outperforms the base diffusion model on every averaged reconstruction and binding metric and is competitive with an online DPO-based method while using no test-set sampling and far less compute.","pith_inferences":["Because the training reward and the reported energy metrics are both computed with Rosetta InterfaceAnalyzer, the measured binding-energy improvements may partly reflect overfitting to that estimator rather than true affinity gains; a test with experimental binding data or a different energy function would separate the two.","The same diffusion-as-GFlowNet framing could transfer to other sparse-reward protein generation tasks—side-chain packing, protein–protein docking, or enzyme design—whenever a terminal score can be precomputed once per training example.","The training schedule that activates the Trajectory Balance term only for the final 5,000 steps suggests the reward acts as a refinement signal on an already-trained generative model; ablating the schedule's length and position might reveal whether earlier reward injection helps or destabilizes reconstruction.","The paper's own appendix reports that Detailed Balance was infeasible because side-chain packing makes per-state energy evaluation too slow; a fast neural surrogate for Rosetta's interface energy would make local, per-step reward objectives viable and could be a concrete next step."],"forward_implications":["Binding-energy optimization no longer requires sampling new CDRs and scoring them during training; a single precomputed energy per training complex suffices, cutting compute by orders of magnitude relative to online RL.","Because the reward is computed only on training complexes, the method avoids the test-set pseudo-labeling used by some RL baselines, reducing data-leakage concerns.","Joint optimization preserves reconstruction quality: unlike DPO-based post-training, which lowers amino acid recovery and worsens RMSD, AbFlowNet improves both relative to the base diffusion model.","If more accurate or experimental binding data became available for training complexes, the same pipeline could use them directly, since energies are needed only once per CDR rather than iteratively for generated CDRs.","The Top-1 energy gains at a 100-sample budget suggest that larger sampling budgets would likely produce further improvements, as the reference baseline itself improves from 480 to 211 kcal/mol when its budget grows from 100 to 2,528 samples."],"supporting_citations":[{"why":"Provides the base diffusion model DiffAb, its transformer architecture, conditioning, and the baseline that all comparisons are measured against.","marker":"[30]"},{"why":"AbDPO, the online DPO-based RL method whose test-set sampling and AAR/RMSD degradation motivate AbFlowNet and with which it is compared.","marker":"[53]"},{"why":"Introduces GFlowNets and the flow-matching framework used to define diffusion steps as states and transitions as flows.","marker":"[4]"},{"why":"Introduces the Trajectory Balance objective that constitutes AbFlowNet's reward-propagation mechanism.","marker":"[31]"},{"why":"PyRosetta InterfaceAnalyzer, the software that computes the binding-energy reward for training CDRs and the evaluation metric for the reported energy gains.","marker":"[6]"},{"why":"SAbDab, source of the 9,410 training antibody-antigen complexes.","marker":"[11]"},{"why":"RAbD, the benchmark/test set and metric definitions (AAR, RMSD, IMP) used for evaluation.","marker":"[3]"},{"why":"Provides the denoising diffusion objective used for 3D coordinate generation.","marker":"[16]"},{"why":"Provides the multinomial diffusion process used for amino-acid type generation.","marker":"[17]"}],"fun_headline_variants":["Diffusion-GFlowNet fusion cuts antibody binding error 38%","GFlowNet inside diffusion trains antibodies for lower energy","Unified diffusion and reward: antibody CDR design without RL","Binding energy optimized directly in denoising steps via GFlowNet","AbFlowNet: joint denoising and reward improves antibody design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the precomputed Rosetta binding-energy score used as the training reward is a useful signal for real binding affinity; since the same estimator also defines the reported energy improvements, the training loop and the evaluation metric are the same quantity, so if that estimator is unreliable the binding-energy gains may not reflect true affinity.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-GFlowNet fusion cuts antibody binding error 38%","GFlowNet inside diffusion trains antibodies for lower energy","Unified diffusion and reward: antibody CDR design without RL","Binding energy optimized directly in denoising steps via GFlowNet","AbFlowNet: joint denoising and reward improves antibody design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1741,"prompt_tokens":955,"completion_tokens":786,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":699}},"tokens_in":571,"tokens_out":786,"duration_ms":7673,"temperature":1.0,"reasoning_tokens":699,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:34:24.074625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain AbFlowNet under the identical pipeline but with the terminal reward replaced by experimental in vitro binding-affinity measurements (or by an independently validated energy function) for the training complexes, then compare generated CDRs against DiffAb on both Rosetta energy and measured affinity. If the reported gains vanish when the reward is decoupled from the evaluation metric, the binding-energy improvement is an artifact of reward–metric coupling.","supporting_citations":[{"cited_title":"Flow network based generative models for non-iterative diverse candidate generation.Advances in Neural Information Processing Systems, 34:27381–27394, 2021","cited_arxiv_id":null,"evidence_quote":"Introduces GFlowNets and the flow-matching framework used to define diffusion steps as states and transitions as flows."},{"cited_title":"Trajectory balance: Improved credit assignment in gflownets.Advances in Neural Information Processing Systems, 35:5955–5967, 2022","cited_arxiv_id":null,"evidence_quote":"Introduces the Trajectory Balance objective that constitutes AbFlowNet's reward-propagation mechanism."},{"cited_title":"Rosettaantibodydesign (rabd): A general framework for computational antibody design.PLoS computational biology, 14(4): e1006112, 2018","cited_arxiv_id":null,"evidence_quote":"RAbD, the benchmark/test set and metric definitions (AAR, RMSD, IMP) used for evaluation."}],"review_version":1}