{"id":"943b6122-2b1c-407e-92df-4c8786d4830b","arxiv_id":"2412.01174","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A rectified-flow generative model for structure-based drug design reports SOTA Vina Dock (-8.50) but its bond loss equation is constant in the model parameters and hyperparameters are tuned on the test set.","lead":"FlowSBDD applies rectified flow to generate 3D drug molecules for protein pockets, reporting an average AutoDock Vina Dock score of -8.50 and 75% diversity on CrossDocked2020. The method is fast (144 seconds for 100 molecules) and allows swapping the starting noise distribution, but the paper's core bond loss is mis-specified and the best results are selected on the test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) as printed is constant in θ, so the bond loss cannot produce the reported 29.4% Vina Dock / 17.6% QED improvements; the central ablation and best-setting claim rest on an unstated or mis-specified loss.","rationale":"The paper's central claim is a new SOTA Vina Dock score enabled by a flexible rectified-flow framework with an additional bond loss. The bond loss is the specific 'adaptive enhancement' highlighted in Section 3.3, and the reported gains in Table 2 are large. As written, Eq. (4) has no dependence on θ, so it cannot contribute a gradient or any learning signal. This is not a matter of external consensus; it is an internal inconsistency between the method description and the ablation. Unless the authors provide code or a corrected equation making the loss depend on vθ, the empirical improvements and the best-setting result are not reproducible from the text. The independent concern about selecting the noise scale on the test set (Table 3) is also valid, but the constant-loss issue is more fundamental because it affects the validity of the method itself. A single controlled ablation comparing training with and without the printed bond loss would settle whether the concern lands: if the models are identical, the loss is vacuous; if they differ, the implemented loss differs from Eq. (4). The reader's weakest_assumption identified exactly this issue, and the resulting REJECT verdict remains appropriate.","tokens_in":6774,"tokens_out":6194,"duration_ms":54226,"concrete_test":"Run an ablation on CrossDocked2020 with identical hyperparameters and seeds: one model trained with Eq. (1) alone and one with Eq. (4) added exactly as printed. Because Eq. (4) has zero gradient w.r.t. θ, the two models should be identical (same checkpoints and metrics). If Table 2's improvements are reproduced, the authors' implementation uses a different loss and Eq. (4) is a mis-statement; if they are not, the reported gains from the bond loss are invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is the internal inconsistency of the proposed bond loss. Section 2.2 first defines a loss involving the network output, Eq. (3): min_θ E ||ϕ(M1) − ϕ(M0 + v(Mt,t))||², then 'more specifically' writes Eq. (4) as min_θ E_{M0,M1∼D} Σ_i ||b_i(M0) − b_i(M1)||. In Eq. (4) the expectations are over the data distribution M1 and the initial noise M0, and neither term depends on θ; the objective is constant, so ∇θ of Eq. (4) is zero. Table 2 attributes a 29.4% Avg. Vina Dock improvement and a 17.6% Med. QED improvement to adding exactly this bond pairwise loss, and the best setting in Table 1 includes it. If the implemented loss is instead of the form in Eq. (3) using predicted coordinates, the paper must state that and correct Eq. (4); as printed, the central 'adaptive enhancement' is vacuous. The separate test-set selection of the noise scale (Table 3 ranking N(0,0.001) as best) compounds the problem, but even without it the bond loss claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FlowSBDD applies rectified flow to structure-based drug design, learning an ODE that transports noisy atom coordinates and types to ligand coordinates conditioned on a protein pocket. The paper proposes adding a bond-distance loss and allows flexible prior distributions, including non-Gaussian noise and molecules generated by TargetDiff. On CrossDocked2020 the authors report an average Vina Dock score of -8.50 with 75.0% diversity, claim state-of-the-art performance over DecompDiff and TargetDiff, and report faster sampling (144 seconds per 100 molecules). The central contribution is the proposed bond-distance loss and the flexible-prior experiments, but the main claims rest on an ablation whose loss function, as written, is constant with respect to the model parameters.","tokens_in":7063,"tokens_out":8501,"duration_ms":74813,"significance":"If the method were correctly specified and the results reproducible, a rectified-flow formulation for SBDD with flexible priors and auxiliary losses would be a useful complement to diffusion-based methods, especially given the reported sampling speed advantage. The paper also draws on the marginal-preserving properties of rectified flow, which is a principled basis for generative modeling. However, as submitted, the key empirical claims are not supported: the bond loss of Eq. (4) has no dependence on the network parameters, the bond-loss ablation is confounded with a change in noise scale, and the headline setting is selected on the test set. The state-of-the-art claim therefore rests on an experimental setup that is internally inconsistent, and the contribution reduces to a plausible but unverified application of rectified flow.","major_comments":[{"comment":"The loss in Eq. (4), min_θ E_{M0,M1∼D} Σ_i ||b_i(M0) − b_i(M1)||, does not contain the network parameters θ or the velocity field v_θ; its gradient with respect to θ is identically zero. Table 2 therefore cannot support the claim that this bond loss improves Avg. Vina Dock by 29.4% and Med. QED by 17.6%. If the implemented loss is instead the one suggested by Eq. (3), evaluated using predicted coordinates, then Eq. (4) must be corrected and the ablation rerun. As printed, the central 'adaptive enhancement' is vacuous.","section":"Section 2.2, Eq. (4)"},{"comment":"The bond-loss ablation is confounded with a change in the noise scale. The caption of Table 2 states that 'Best w/ B' uses a Gaussian scaled by 0.01, but Table 3 reports identical numbers for N(0, 0.001). Either the scale is misreported, or the comparison changes both the noise scale and the loss at the same time; in neither case can the reported Vina Dock improvement be attributed specifically to the bond loss.","section":"Section 3.3, Tables 2 and 3"},{"comment":"The prior noise scale is selected using the test set. The N(0, 0.001) row is chosen because it 'achieved the best QED and Vina Dock results' on the same 100 test complexes used for the headline benchmark numbers, and no validation split or model-selection procedure is described. The reported -8.50 Avg. Vina Dock is therefore an optimistic selection, and the comparison with baselines is biased in favor of FlowSBDD.","section":"Section 3.3, Table 3"},{"comment":"All benchmark numbers are point estimates with no error bars, standard deviations, or significance tests. The claimed improvement over DecompDiff in Avg. Vina Dock (-8.50 vs -8.39) is small, and without any measure of variability the state-of-the-art claim is not established. This concern is amplified by the test-set noise-scale selection in Table 3.","section":"Section 3.1, Table 1"}],"minor_comments":[{"comment":"Eq. (3) writes v(M_t, t) even though the text states that the velocity field is parameterized by θ; use v_θ(M_t, t) consistently.","section":"Section 2.2, Eq. (3)"},{"comment":"The informal theorem uses dZ_t without defining Z_t; replace with dM_t or define the new process explicitly.","section":"Section 2.2, Theorem 1"},{"comment":"The abstract claims 'state-of-the-art performance' but the conclusion says 'comparable performance to SOTA'; these statements should be reconciled.","section":"Abstract and Conclusion"},{"comment":"The statement that diversity 'exceeded Targetdiff by 4.2%' is ambiguous; report the absolute difference (0.03) as well as the relative percentage.","section":"Section 3.2, Diversity comparison"},{"comment":"The caption says 'Best w/ B' uses a Gaussian scaled by 0.01, but Table 3 shows the same numbers for N(0, 0.001); please correct the inconsistency and ensure all reported settings match their descriptions.","section":"Section 3.3, Table 2 caption"},{"comment":"No code, training hyperparameters, or details of the bond extraction function b_i are provided, which would be needed to reproduce the experiments.","section":"General reproducibility"}],"recommendation":"reject","confidential_remarks":"The paper is an interesting application of rectified flow, but the central methodological inconsistency in Eq. (4) makes the main empirical result unreliable as written. Correcting the loss and re-running the ablations would amount to a substantially new experimental study, so I do not see a routine revision that can repair the paper within the scope of this submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nFlowSBDD takes a sensible idea—rectified flow for structure-based drug design—and shows it can generate molecules quickly (144s per 100 molecules) and with competitive Vina Dock numbers. The flexible-prior experiment, where the model starts from TargetDiff outputs and refines them, is genuinely interesting and points to a useful post-processing strategy. If the headline -8.50 holds up, this is a meaningful practical addition to the SBDD toolkit.\n\nBut the central \"adaptive enhancement\" claim does not hold up as written. Eq. (4), the bond loss credited with large gains in Table 2, is independent of the model parameters: both terms are expectations over data and noise, with no vθ inside. Its gradient is zero. So the 29.4% Vina Dock and 17.6% QED improvements cannot be caused by that loss as printed. Either the equation is mis-stated or the implemented loss is different, and the paper needs to say which. This is load-bearing because Table 1's best setting includes this loss.\n\nThe noise scale is also selected on the test set: Table 3 ranks N(0,0.001) by looking at Vina Dock on the same 100 test pockets used in Table 1. That makes the headline number a fitted value, which is not a fair comparison to baselines that receive no such tuning. A separate validation split or cross-validation is required.\n\nThere are smaller issues: the paper underperforms on Vina Score and on QED/SA relative to several baselines, so the \"maintaining proper molecular properties\" claim is overstated. It also omits citations to prior flow-based 3D molecule generators, which tempers the novelty. These are minor next to the Eq. (4) problem.\n\nThe limitations paragraph is honest about the Vina Score weakness, which I credit. But the two main problems—a vacuous loss in the central ablation and test-set tuning—mean the paper's current evidence does not support its claims. I would send it to reviewers because the framework and the flexible-prior idea deserve a fair hearing; a corrected version with a properly specified loss and a held-out tuning split could become a solid contribution.","headline":"Rectified-flow SBDD with fast sampling and an interesting refine-from-TargetDiff idea, but the bond-loss ablation rests on an equation constant in θ and the noise scale is tuned on the test set.","tokens_in":7565,"tokens_out":3440,"would_cite":false,"duration_ms":28968,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A rectified-flow generative model for structure-based drug design, FlowSBDD, reports state-of-the-art binding affinity (average AutoDock Vina dock score -8.50) and 75.0% diversity on CrossDocked2020 while sampling 100 molecules per pocket…","keywords":["rectified flow","structure-based drug design","ligand generation","3D molecule generation","bond distance loss","CrossDocked2020","AutoDock Vina"],"falsifier":"Re-run the 'Best w/ B' setting with the printed bond loss exactly as written; since the loss is constant with respect to $\\theta$, training with and without it must give identical results. If the reported gains vanish, the stated mechanism is falsified. Alternatively, inspect the training code to see whether the bond loss is applied to the network's output (e.g., $\\phi(M_0 + v_\\theta(M_t,t))$) rather than to $M_0$, in which case the printed equation would be a typo.","tokens_in":6566,"feed_emoji":"💊","tokens_out":5373,"duration_ms":41613,"temperature":0.7,"pith_summary":"This paper tries to establish that rectified flow, a transport-based generative framework, is a strong alternative to diffusion models for structure-based drug design. It proposes FlowSBDD, which learns a velocity field that maps noisy initial atom states to ligand coordinates and types conditioned on a protein pocket, and it adds a bond-distance loss and flexible choice of initial distribution. On the CrossDocked2020 benchmark the model reports an average AutoDock Vina dock score of -8.50 and 75.0% diversity, surpassing the diffusion baselines DecompDiff (-8.39) and TargetDiff (-7.80) while sampling 100 molecules per pocket in 144 seconds. The importance is practical: if true, it means a simpler ODE-based sampler with a flexible objective can deliver comparable or better binding affinity and diversity at a fraction of the sampling cost of diffusion models.","feed_headline":"Rectified flow tops drug docking with -8.50 Vina score","feed_subtitle":"FlowSBDD samples 100 ligands per pocket in 144 seconds, beating DecompDiff and TargetDiff on CrossDocked2020.","key_machinery":"The central object is the rectified-flow ODE $dM_t/dt = v(M_t, t)$ with interpolated state $M_t = tM_1 + (1-t)M_0$, trained by matching the velocity field to $M_1 - M_0$. The paper wraps this in a SE(3)-equivariant graph neural network that updates atom embeddings and coordinates alternately, and it adds two extensions: a 'bond loss' that compares Euclidean distances between bonded atom pairs at the data endpoint and at the transported state, and a flexible initial distribution that replaces the standard Gaussian with other random distributions or with molecules generated by another model such as TargetDiff. The bond loss is the mechanism the paper credits for the largest performance jump in the ablation, and the flexible prior is the mechanism for refining existing samples.","core_discovery":"The paper's central claim is that replacing the diffusion learning objective with a rectified-flow objective, where a network $v_\\theta$ is trained to follow the straight-line interpolation between a noise distribution $M_0$ and the data $M_1$, yields a generative model of ligand molecules that binds to protein pockets at state-of-the-art affinity. The authors further claim that an auxiliary loss on bond Euclidean distances and the ability to start from a non-Gaussian or even a model-generated prior distribution improve the results, with the bond loss raising median QED by 17.6% and average Vina dock by 29.4% in their ablation. FlowSBDD is presented as a more flexible and faster alternative to diffusion-based SBDD, achieving -8.50 average Vina dock and 75.0% diversity on CrossDocked2020 while sampling 100 molecules per pocket in 144 seconds.","pith_inferences":["The stated bond loss in Eq. (4) is independent of the network parameters, so the ablation gains attributed to it cannot come from that printed objective; a corrected loss that depends on $v_\\theta$ would be needed to support the paper's interpretation of the ablation.","The speed of FlowSBDD could make it practical as the generative component in large-scale virtual screening pipelines, where diffusion samplers are often the bottleneck.","The flexible-prior idea points toward a general refinement paradigm: any existing docking or generative sampler's output can be fed as the initial distribution, turning the flow model into a post-processing step that potentially improves affinity without retraining the original model.","The strong dependence on noise scale ($N(0,0.001)$ far outperforming $N(0,1)$) suggests that the practical gains may come more from prior scale tuning than from the rectified-flow formulation itself; a fair comparison would require matching this tuning across baselines."],"forward_implications":["If FlowSBDD's numbers hold, rectified flow offers a faster sampling route than diffusion for SBDD: 144 seconds for 100 molecules per pocket versus 3428 seconds for TargetDiff and 6189 for DecompDiff.","The reported -8.50 average Vina dock and 75.0% diversity would put FlowSBDD ahead of current diffusion baselines on the binding-affinity metric while matching or exceeding their diversity.","The flexible-prior result suggests the same trained velocity field can serve as a refiner of molecules produced by other samplers, improving their dock scores (e.g., 100tar-Flow improves 100-step TargetDiff from -4.04 to -6.69 average Vina dock).","The bond Euclidean distance loss, if correctly implemented, provides a template for injecting task-specific geometric constraints into flow training beyond simple coordinate/type reconstruction."],"supporting_citations":[{"why":"Supplies the rectified-flow framework and the optimal-transport theorem used to justify adding a cost function.","marker":"[13]"},{"why":"Introduces the rectified-flow training objective that the model's velocity field is trained with.","marker":"[14]"},{"why":"TargetDiff: the diffusion baseline FlowSBDD compares against and whose architecture inspires the equivariant GNN.","marker":"[6]"},{"why":"DecompDiff: the strongest diffusion baseline, compared on Vina dock, QED, and sampling time.","marker":"[7]"},{"why":"CrossDocked2020: the benchmark dataset and the filtering/split protocol the experiments follow.","marker":"[4]"},{"why":"AutoDock Vina 1.2.0: the scoring function used to measure binding affinity (Vina Score, Vina Dock).","marker":"[3]"},{"why":"Pocket2Mol: autoregressive baseline and reference for molecular-property comparison.","marker":"[17]"},{"why":"E(n) equivariant graph neural network: the architecture basis for modeling ligand-protein interactions.","marker":"[20]"}],"fun_headline_variants":["FlowSBDD: rectified flow hits -8.50 Vina dock","Rectified flow for drug design: 144s per pocket","No diffusion: rectified flow beats drug docking SOTA","FlowSBDD: better affinity, 75% diversity in 144s"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The bond loss in Eq. (4) has no dependence on the velocity network's parameters, so the paper's central ablation claim that this loss boosts Vina dock by 29.4% and QED by 17.6% rests on the assumption that the implemented loss is different from the printed one and actually contributes a gradient.","fun_headline_variants_meta":{"raw":{"variants":["FlowSBDD: rectified flow hits -8.50 Vina dock","Rectified flow for drug design: 144s per pocket","No diffusion: rectified flow beats drug docking SOTA","FlowSBDD: better affinity, 75% diversity in 144s"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1212,"prompt_tokens":876,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":258}},"tokens_in":492,"tokens_out":336,"duration_ms":3368,"temperature":1.0,"reasoning_tokens":258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:36:22.121405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 'Best w/ B' setting with the printed bond loss exactly as written; since the loss is constant with respect to $\\theta$, training with and without it must give identical results. If the reported gains vanish, the stated mechanism is falsified. Alternatively, inspect the training code to see whether the bond loss is applied to the network's output (e.g., $\\phi(M_0 + v_\\theta(M_t,t))$) rather than to $M_0$, in which case the printed equation would be a typo.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DecompDiff: the strongest diffusion baseline, compared on Vina dock, QED, and sampling time."},{"cited_title":"Journal of chemical information and modeling 60(9), 4200–4215 (2020) 1, 3","cited_arxiv_id":null,"evidence_quote":"CrossDocked2020: the benchmark dataset and the filtering/split protocol the experiments follow."},{"cited_title":"0: New docking methods, expanded force field, and python bindings","cited_arxiv_id":null,"evidence_quote":"AutoDock Vina 1.2.0: the scoring function used to measure binding affinity (Vina Score, Vina Dock)."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Pocket2Mol: autoregressive baseline and reference for molecular-property comparison."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"E(n) equivariant graph neural network: the architecture basis for modeling ligand-protein interactions."}],"review_version":1}