{"id":"399ce846-2ce0-4157-8dbc-97e5934a7523","arxiv_id":"2411.13280","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MolJO applies joint gradient guidance to Bayesian Flow Networks, achieving state-of-the-art success rate (51.3%) in structure-based molecule optimization on CrossDocked2020.","lead":"A new machine learning method called MolJO guides a 3D molecule generator with gradients from desired drug properties, improving both atom types and 3D coordinates at once. It reports state-of-the-art results on the CrossDocked2020 benchmark and can be extended to lead-optimization tasks like R-group design and scaffold hopping.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim requires the property regressor to be trained without test-set labels, but Appendix D.1 omits the train/test split and the code link points to MolCRAFT, not MolJO; the reported 51.3% Success Rate is unverified until the split is confirmed.","rationale":"The central claim is an empirical SOTA on CrossDocked2020. The strongest threat to such a claim is not the smoothness of the surrogate in OOD regions (the reader's weakest assumption) but whether the evaluation itself is uncontaminated. The paper's own description of regressor training (Appendix D.1, Eq. 20) is silent on the train/test split, and the code link is to MolCRAFT rather than a MolJO-specific repository. These two omissions together mean the paper does not establish that the guidance used no test-set labels. This is a load-bearing condition: if the regressor saw the 100 test proteins' ligands, the reported Success Rate would be an artifact of test-set memorization rather than a valid optimization result. I emphasize that this is a good-faith ambiguity, not an allegation; the authors most likely used the training split, but the manuscript should state it explicitly, and the code should make it checkable. I therefore keep the reader's CONDITIONAL verdict: the claim is plausible but not fully established until the split is confirmed. The reader's concern about surrogate gradients in OOD regions is real but secondary: even if the regressor is imperfect, the empirical benchmark result would still stand as long as the evaluation is clean; conversely, a test-set leak would invalidate the benchmark regardless of surrogate quality. For that reason I see my concern as more load-bearing and mark agreement as partial.","tokens_in":29199,"tokens_out":12566,"duration_ms":126061,"concrete_test":"Inspect the released code (or ask the authors) for the property-regressor training script and verify that the regressor training set is exactly the 100,000 training poses (proteins clustered at 30% sequence identity, RMSD < 1A filter) and excludes the 100 test proteins and their ligands. If the script uses the full CrossDocked set, a random split, or the test split, the benchmark claim is invalid. If code is unavailable, rerun regressor training on the train split only with the paper's hyperparameters (Adam, lr=0.0005, batch size 8, 5 epochs, TargetDiff-style architecture) and re-evaluate Table 1 row 14 on the 100 test proteins; a Success Rate drop below the MolCRAFT baseline (26.8%) or a material change in ranking would confirm the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Success Rate 51.3%, Vina Dock -9.05, SA 0.78, Table 1) is only valid if the differentiable property regressor E(theta, p, t) used for guidance is trained on the CrossDocked training split and has no access to the 100 test proteins or their reference ligands. Appendix D.1, 'Training Property Regressors,' states only that the regressor is trained 'based on the molecules and their properties (Vina Score, QED, SA) in CrossDocked dataset' (Eq. 20) and does not specify whether p_data is the 100,000 training poses or the full dataset including the test proteins. Section 5.1 defines the train/test split for the generative model, but it is not stated that the same split is used for the regressor. If the regressor was fit on the full dataset, then during inference the gradient guidance could steer toward test-reference-like solutions, and the 4x-vs-TAGMol and 2x-Me-Better claims would reflect test-set memorization rather than generalized optimization. The provided code link (https://github.com/AlgoMole/MolCRAFT) points to the MolCRAFT repository and does not clearly contain MolJO artifacts, so this condition cannot be verified from the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MolJO, a gradient-based structure-based molecule optimization framework built on Bayesian Flow Networks (BFNs). The method derives a joint guidance signal over continuous atomic coordinates and discrete atom types by applying product-of-experts sampling with a first-order Taylor expansion of a learned property energy function (Prop. 4.1), and introduces a 'backward correction' sampling strategy that updates the current belief from a sliding window of past history rather than only the previous step. The authors claim state-of-the-art results on CrossDocked2020 (Success Rate 51.3%, Vina Dock -9.05, SA 0.78), a 4x improvement over the gradient-based TAGMol baseline, a 2x 'Me-Better' ratio relative to 3D baselines, and additional results for multi-objective optimization, R-group optimization, and scaffold hopping.","tokens_in":29462,"tokens_out":11027,"duration_ms":103688,"significance":"If the empirical claims hold, MolJO would be a strong contribution: it extends classifier-guidance ideas to the continuous-discrete belief space of BFNs, provides an equivariance proof, and demonstrates broad applicability beyond unconstrained optimization. The paper includes useful empirical safeguards: 95% confidence intervals for the main metrics (Table 6), paired t-tests for the backward-correction ablation (Table 12), an inference-time comparison (Table 16), and a thoughtful analysis of molecular-size confounds (Appendix E). The backward-correction ablation is informative and shows a clear performance gain. However, the central SOTA and comparison claims rest on several details that are either underspecified or framed in a way that overstates the contribution, notably the training split of the property regressor, the backbone mismatch in the 4x claim, and the lack of algorithmic detail for constrained optimization. These issues are fixable but are load-bearing for the paper's headline conclusions.","major_comments":[{"comment":"The training split for the property regressor E(theta,p,t) is not specified: Eq. (20) trains on 'the molecules and their properties ... in CrossDocked dataset' without stating that p_data is restricted to the 100,000 training poses and excludes the 100 test proteins and their reference ligands. If the regressor saw test-pocket ligand complexes, the guidance gradients could steer toward test-reference-like solutions, and the reported Success Rate 51.3% and Vina Dock -9.05 would not demonstrate generalization. Please state the split explicitly and release MolJO-specific code (the current link points to the MolCRAFT repository) so this can be verified.","section":"Appendix D.1 / Eq. (20)"},{"comment":"The headline 'more than 4x improvement in Success Rate compared to the gradient-based counterpart' compares MolJO (51.3%) with TAGMol (11.1%), but the two methods use different generative backbones (MolCRAFT vs TargetDiff). The within-backbone gain is 26.8% (MolCRAFT) to 51.3% (MolJO), i.e., 1.9x, so the 4x framing overstates the contribution of the guidance method itself. Please either restrict the claim to a same-backbone comparison or clearly state that the comparison is across different generative backbones.","section":"Abstract, Sec. 5.2, Table 1"},{"comment":"There is a mismatch between the derived guidance and the implemented guidance for the discrete modality. Prop. 4.1 defines g_yv = g_theta_v * d(theta_v)/d(y_v), but Sec. 4.2 states that 'We further bypass the derivative d(theta_v)/d(y_v) = theta_v(1-theta_v) to stabilize the gradient flow.' Algorithm 1 line 9 then samples y_v using Eq. 5 without the Jacobian term. Please clarify whether Eq. 5 is intended to include the chain-rule gradient or the bypassed gradient, and adjust the proposition or the algorithm accordingly.","section":"Sec. 4.1-4.2, Prop. 4.1, Algorithm 1"},{"comment":"The claim that MolJO is '2x as effective' in proposing 'me-better' candidates (39.8% vs below 17% for other baselines) is not backed by a table. Please define the exact criterion for 'all-better' (which properties must improve, and compared against which reference), report the ratio for every baseline, and provide confidence intervals.","section":"Sec. 5.2, Fig. 1B"},{"comment":"The constrained optimization experiments are said to be 'achieved by infilling (details in Appendix D.2)', but Appendix D.2 only describes R-group and scaffold hopping task definitions and RDKit fragmentation; it does not explain how the masked or infilled regions are integrated into the guided BFN sampling of Algorithm 1. Please provide the algorithmic details (masking, conditional update of theta, and any changes to backward correction) needed to reproduce these results.","section":"Sec. 5.3, Appendix D.2"},{"comment":"No validation accuracy or gradient sanity check is reported for the learned property regressor E(theta,p,t). Since guidance scale 50 can drive samples outside the training distribution of theta, the first-order Taylor expansion in Prop. 4.1 may be inaccurate in those regions; please report held-out regression error (e.g., R^2 or MAE on a validation split of noised BFN beliefs) and, ideally, an analysis of how the surrogate gradient behaves under large guidance.","section":"Appendix D.1"}],"minor_comments":[{"comment":"The instruction 'Sample theta_x according to Eq. 4 and 13' is ambiguous; Eq. 4 is the guided one-step transition while Eq. 13 is the backward-corrected update. Please clarify how the two equations combine in the sampling loop.","section":"Algorithm 1, line 8"},{"comment":"The notation sigma_x is used inconsistently: in Eq. 4 and Prop. 4.1 it appears as a variance parameter, while Eq. 13 gives the variance as Delta_beta / rho_n^2. Please define sigma_x and sigma_v unambiguously in terms of the BFN schedules.","section":"Eq. 4, Prop. 4.1, Eq. 13"},{"comment":"There are several typos and infelicities, including 'optimziation' in the Impact Statement, 'synthesizeability' in Sec. 1, 'utlizing' and 'unsatifactory' in Appendix G, and 'Cummulative' in the Fig. 9 caption. A careful proofread is recommended.","section":"Throughout"},{"comment":"The FLAG row reports Vina Score 45.85 in Table 1 but 45.978 in Table 6; please reconcile these numbers or explain the difference in evaluation.","section":"Table 1 vs Table 6"},{"comment":"Please specify whether the property regressor uses the same atom featurization and kNN graph construction as TargetDiff, and how the noised BFN belief theta is featurized before being passed to that network.","section":"Appendix D.1"}],"recommendation":"major_revision","confidential_remarks":"The key gatekeeping issue is the property-regressor data split: if the regressor was trained on the full CrossDocked set including the 100 test proteins, the main empirical claim would be invalid. Please verify this with the authors before further consideration. Also, the code link points to MolCRAFT, so the MolJO implementation must be released to make the claims checkable. The 4x-vs-TAGMol framing should be softened to avoid overstating the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine methods contribution, and the central guidance idea is real. The reported 51.3% Success Rate is the one number to check, and the paper currently leaves an ambiguity that makes it impossible to verify from the text alone.\n\nWhat's new: MolJO is the first to push gradient guidance through the BFN belief parameters for both continuous coordinates and discrete atom types jointly, instead of guiding only coordinates (TAGMol) or assuming a Gaussian classifier (DiGress). The first-order Taylor product-of-experts derivation in Prop. 4.1 is clean. The backward-correction sampling with a sliding window is a sensible extension that unifies the k=1 and k=n samplers and empirically stabilizes guidance; the ablation is convincing, and they provide 95% confidence intervals plus paired t-tests. The SE(3)-equivariance argument is standard but correctly transferred. Credit is also due for extending to R-group optimization and scaffold hopping.\n\nSoft spots, in order of size. First, the property regressor E(θ,p,t) is the engine of the method, but Appendix D.1 does not state whether it was trained on the 100k training split or on the full CrossDocked set including the 100 test proteins. If it saw test proteins, the gradient guidance could exploit test-reference-like solutions, and the '4x' claim would reflect memorization rather than optimization. The stress-test note lands here. I do not see positive evidence of leakage, but the omission is exactly what a referee should force into the next version. Second, the headline comparison to TAGMol is not apples-to-apples: different backbones (MolCRAFT vs TargetDiff). The 4x Success Rate is plausible but not a clean attribution of the guidance mechanism; the joint-modality ablation helps but does not fix the cross-backbone confound. Third, the code link points to the MolCRAFT repository; the MolJO-specific artifacts are not clearly there, so the main result is not independently reproducible from the link. Fourth, the backward correction is a heuristic without a rigorous bias or convergence argument, and the guidance scale/window are tuned on the benchmark. Minor: the size restriction in Table 1 is defensible but complicates direct comparison with methods that explore larger molecules.\n\nWho this is for: anyone working on gradient guidance for structured generative models, or on structure-based lead optimization. The method deserves a serious referee; the main empirical claim needs the split confirmed and ideally code cleanup, but the work is worth engaging seriously.\n\nRecommendation: send to peer review with requests to clarify the regressor train/test split, release MolJO code with the trained regressor, and provide a same-backbone comparison with TAGMol if possible.","headline":"Solid new guidance mechanism for BFN-based molecule optimization; the SOTA claim hinges on an unstated regressor train/test split that must be fixed before believing it.","tokens_in":30010,"tokens_out":2477,"would_cite":true,"duration_ms":26516,"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":"MolJO guides molecule optimization on the Bayesian belief space of a flow network, jointly steering continuous coordinates and discrete atom types, and reports state-of-the-art success on CrossDocked2020.","keywords":["structure-based molecule optimization","Bayesian flow networks","gradient guidance","joint continuous-discrete guidance","SE(3)-equivariant sampling","backward correction","CrossDocked2020","scaffold hopping"],"falsifier":"Replace the learned regressor with the exact oracle (Vina docking score, QED, and SA computed on decoded molecules) at each guidance step, or measure the gradient-field smoothness of the learned regressor on beliefs sampled from the guided trajectory at scales 50–100; if the success rate collapses or the gradient field is demonstrably non-smooth on those out-of-distribution beliefs, the guidance mechanism is the bottleneck.","tokens_in":29004,"feed_emoji":"🧬","tokens_out":9972,"duration_ms":81608,"temperature":0.7,"pith_summary":"This paper aims to make gradient-guided optimization work for structure-based molecule design, where a molecule's atoms have both continuous 3D coordinates and discrete atom types. It claims that Bayesian Flow Networks offer a continuous, differentiable 'belief' space, the posterior parameters θ, on which one can compute joint gradients over both modalities at once, without relaxing the discrete atom types into Gaussians or guiding only coordinates. The resulting method, MolJO, reaches state-of-the-art results on the CrossDocked2020 benchmark: 51.3% success rate, Vina Dock -9.05, SA 0.78, a fourfold improvement in success rate over the coordinate-only gradient baseline TAGMol, and a 2x higher fraction of 'me-better' molecules that improve affinity, drug-likeness, and synthesizability together. The paper also introduces a backward-correction sampling strategy that re-optimizes a sliding window of past belief states, which it argues aligns guidance gradients across steps and balances exploration with exploitation. If these claims hold, MolJO would be the first gradient-based SBMO method that jointly and equivariantly guides both modalities, and it extends naturally to constrained tasks such as R-group optimization and scaffold hopping.","feed_headline":"Gradient-guided molecule optimizer hits 51.3% success on docking benchmark","feed_subtitle":"MolJO steers both atom coordinates and types with one gradient signal, setting a new bar on CrossDocked2020.","key_machinery":"The central object is the Bayesian belief θ, the receiver's posterior parameters over continuous coordinates θ^x and discrete atom-type logits θ^v, updated through the Bayesian update function h(θ_{i-1}, y, α_i). Guidance is exerted on θ via the score ∇_θ log p_E(θ) with p_E ∝ exp[-sE(θ,p,t)], and Proposition 4.1 reduces the guided transition kernel to shifting the Gaussian of θ^x by σ^x g_{θx} and the Gaussian of y^v by σ^v g_{yv}, where the gradients come from a learned SE(3)-equivariant energy regressor. The second mechanism is backward correction: rather than sampling θ_i only from θ_{i-1}, the update aggregates accuracy over a window of k past steps, p_U(θ_n | θ_{n-k}, x̂_n; Σ_{i=n-k+1}^n α_i), so the optimized reconstruction x̂_n replaces earlier, suboptimal estimates. Together, these two mechanisms, belief-space gradient guidance and windowed backward correction, constitute MolJO.","core_discovery":"The central claim is that the correct object to guide in a Bayesian Flow Network is not the noisy latent y nor the clean molecule, but the aggregated Bayesian posterior belief θ, which lives in continuous space for coordinates and on the probability simplex for atom types. Proposition 4.1 derives, via a first-order Taylor expansion of an energy function E(θ,p,t), a guided transition kernel in which the Gaussian mean of the coordinate belief shifts by the coordinate gradient and the Gaussian latent for discrete counts shifts by the type gradient, so both modalities are steered simultaneously while keeping the discrete belief on the simplex. Proposition 4.4 shows that this guided sampling is SE(3)-equivariant whenever the backbone and energy function are equivariant and the protein is centered. The paper further claims that a backward-correction update, which aggregates the accuracy over a sliding window of past steps and conditions the current belief on the optimized reconstruction, unifies previous BFN sampling strategies and empirically improves both unguided sampling and the effectiveness of guidance. On CrossDocked2020, MolJO reports the best success rate (51.3%), Vina Dock (-9.05), and SA (0.78) among compared optimization methods, with 39.8% of generated molecules improving affinity, QED, and SA simultaneously.","pith_inferences":["The paper leaves implicit that the same belief-space guidance recipe should transfer to other BFN-based generative models beyond molecules, since the derivation only uses the Gaussian or simplex structure of θ and an energy function; materials, graphs, or unified 3D models trained as BFNs could inherit MolJO-style joint guidance by swapping in a suitable regressor.","A testable extension is to penalize the learned regressor for non-smooth gradients on out-of-distribution beliefs reached at guidance scales 50–100; if the first-order Taylor approximation is the bottleneck, such a penalty should push the reported success rate higher.","The backward-correction update can be read as a sliding-window proximal optimizer over the belief chain, suggesting that the window size k tunes a bias-variance trade-off between error accumulation (small k) and stale, over-smoothed updates (large k), which would predict an optimal window near the value 130 used here.","If joint θ-guidance is the causal driver of the 2x 'me-better' ratio, then coordinate-only guidance applied to the same backbone should reproduce the failure mode of good Vina but poor SA/QED, while type-only guidance should produce the opposite, a contrast the ablations hint at but do not fully disentangle."],"forward_implications":["If the reported numbers hold, gradient guidance becomes a viable plug-and-play optimizer for structure-based drug design without per-target oracle simulations or RL fine-tuning of the backbone.","Joint guidance over coordinates and atom types resolves the modality imbalance that limited coordinate-only guidance, so optimized molecules can improve drug-likeness and synthesizability rather than affinity alone.","The backward-correction window size k gives a single knob that interpolates between exploration and exploitation, letting practitioners trade sample diversity against optimization intensity.","The framework extends to constrained settings such as R-group optimization, fragment growing, and scaffold hopping through infilling, meaning the same gradient-guided sampler handles lead-optimization tasks beyond de novo design.","Equipping MolJO with top-of-N oracle selection raises success rate to 70.3% on the same benchmark, indicating that the gradient signal and a small screening budget are complementary."],"supporting_citations":[{"why":"Defines Bayesian Flow Networks, including the sender-receiver setup, Bayesian update function h, and the accuracy-additivity property used to derive backward-correction updates.","marker":"(Graves et al., 2023)"},{"why":"MolCRAFT, the pretrained BFN backbone for structure-based drug design that MolJO guides without finetuning; its sampling strategy corresponds to the k=n extreme of backward correction.","marker":"(Qu et al., 2024)"},{"why":"Classifier guidance for diffusion models, the template for gradient guidance that MolJO adapts to the BFN belief space.","marker":"(Dhariwal & Nichol, 2021)"},{"why":"TAGMol, the coordinate-only gradient-guided baseline that MolJO improves upon by 4x in success rate; the contrast shows the cost of ignoring discrete types.","marker":"(Dorna et al., 2024)"},{"why":"Supplies the equivariance argument about equivariant gradient flows that Proposition 4.4 relies on.","marker":"(Bao et al., 2022)"},{"why":"SDE-based BFN sampling, used as an ablation baseline and as the alternative to backward correction.","marker":"(Xue et al., 2024)"},{"why":"TargetDiff, a continuous-discrete diffusion baseline and the origin of the benchmark metrics, including Success Rate thresholds and the Vina Dock protocol.","marker":"(Guan et al., 2022)"},{"why":"CrossDocked2020, the dataset used to train the backbone and property regressors and to run the reported evaluations.","marker":"(Francoeur et al., 2020)"}],"fun_headline_variants":["Guiding Bayesian flow beliefs optimizes molecules with gradients","MolJO: gradient-guided molecule optimization hits 51.3% success","Gradient guidance on Bayesian posterior boosts docking success","MolJO: SE(3)-equivariant gradient optimization for drug design","Backward-correction gradient guidance achieves 51.3% docking success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned property regressor E(θ,p,t), trained on noised BFN beliefs from CrossDocked, gives accurate and smooth gradients across the whole guided trajectory, including out-of-distribution beliefs reached at guidance scale 50; if the regressor is inaccurate or non-smooth there, the first-order Taylor approximation in Proposition 4.1 will steer sampling incorrectly.","fun_headline_variants_meta":{"raw":{"variants":["Guiding Bayesian flow beliefs optimizes molecules with gradients","MolJO: gradient-guided molecule optimization hits 51.3% success","Gradient guidance on Bayesian posterior boosts docking success","MolJO: SE(3)-equivariant gradient optimization for drug design","Backward-correction gradient guidance achieves 51.3% docking success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1404,"prompt_tokens":1044,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":273}},"tokens_in":660,"tokens_out":360,"duration_ms":3938,"temperature":1.0,"reasoning_tokens":273,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:37:46.225158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the learned regressor with the exact oracle (Vina docking score, QED, and SA computed on decoded molecules) at each guidance step, or measure the gradient-field smoothness of the learned regressor on beliefs sampled from the guided trajectory at scales 50–100; if the success rate collapses or the gradient field is demonstrably non-smooth on those out-of-distribution beliefs, the guidance mechanism is the bottleneck.","supporting_citations":[],"review_version":1}