{"id":"dbb7a823-2b62-4fda-812c-8eb24fb995ad","arxiv_id":"2508.17815","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DrugFlow, a flow-matching plus Markov-bridge generative model, reports state-of-the-art distributional fidelity for structure-based drug design and adds uncertainty, size adaptation, side-chain flexibility, and preference alignment.","lead":"A new generative model called DrugFlow combines continuous flow matching with discrete Markov bridges to generate drug-like molecules inside protein pockets, and it reports closer matches to the training data distribution than three existing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparability is unverified: filtered training set and validity filtering may inflate baseline distances, undermining the SOTA distribution-learning claim.","rationale":"The reader's weakest_assumption exactly identifies the baseline-comparability problem: DrugFlow is trained and evaluated on a PoseBusters-filtered training set, while the three baselines are not shown to have been retrained or filtered in the same way. This is the most load-bearing assumption because the paper's headline conclusion, 'consistently achieves state-of-the-art distribution learning performance,' depends entirely on the distance comparisons in Tables 1–3. My stress-test pass confirms this is a genuine and unresolved issue: the methods section explicitly states the filtering for DrugFlow's training set and validity filtering for DrugFlow's samples, but is silent on the same steps for baselines. This is not a mere theoretical quibble — if the baselines were trained on the unfiltered CrossDocked split, then the reference distribution itself differs, and every reported baseline distance is systematically biased. The Appendix's bootstrap and t-test analyses only measure sampling variability of the generated sets, not this systematic protocol mismatch. The concern is addressable, not fatal: adding explicit statements about baseline training data and applying identical validity filtering would resolve it, or retraining baselines on the filtered set would provide a definitive apples-to-apples comparison. Since the reader already judged the paper CONDITIONAL with moderate confidence, and my analysis supports that judgment without identifying a different or stronger concern, the verdict should remain unchanged.","tokens_in":28734,"tokens_out":3097,"duration_ms":34026,"concrete_test":"Obtain the official released checkpoints for Pocket2Mol, DiffSBDD, and TargetDiff, generate 100 molecules per test target, apply the same RDKit validity filter used for DrugFlow, and recompute the Wasserstein and JSD distances in Tables 1–3 against the PoseBusters-filtered training set. If DrugFlow still wins on the large majority of metrics, the SOTA distribution-learning claim is supported; if the baseline distances drop substantially (e.g., JSD_all changes by more than 0.05) or the ranking changes, the reported comparison is confounded by the filtering protocol and must be re-run with baselines retrained on the filtered set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DRUGFLOW is a state-of-the-art distribution learner on CrossDocked. Section 3.1 states that 'Ligands that do not pass all PoseBusters filters were removed from the training set,' and that DrugFlow samples were filtered by RDKit validity. The paper never states that Pocket2Mol, DiffSBDD, or TargetDiff were retrained on this PoseBusters-filtered set, nor that their generated molecules were subjected to the same validity filter. Because every Wasserstein and JSD distance in Tables 1–3 is computed against the filtered training set, this is a systematic protocol difference. If the baselines were trained on the original unfiltered CrossDocked distribution, their samples are matched to a different target distribution; comparing them to the filtered set inflates their distances. Similarly, if baseline samples include RDKit-invalid molecules while DrugFlow samples are filtered, the baseline distributions are contaminated. The bootstrap significance tests in Appendix B.10 resample from the same generated sets and cannot correct this systematic bias. The SOTA claim therefore rests on an unverified apples-to-apples assumption between training data, validity filtering, and reference distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DrugFlow, a generative model for structure-based drug design that combines continuous flow matching for atom coordinates (and, in the FlexFlow variant, for protein side-chain torsion angles on the torus) with Markov bridge models for discrete atom and bond types. It adds three methodological contributions: an end-to-end trained per-atom uncertainty estimate intended to flag out-of-distribution samples, an adaptive size selection mechanism based on virtual 'no atom' nodes, and a multi-domain preference alignment (MDPA) loss adapted from DPO for both continuous and discrete domains. The evaluation is centered on distribution-matching metrics (Wasserstein distances and Jensen-Shannon divergences) against Pocket2Mol, DiffSBDD, and TargetDiff on CrossDocked, and the paper claims consistent state-of-the-art distribution learning performance.","tokens_in":28991,"tokens_out":7733,"duration_ms":75674,"significance":"If the claims hold, the paper makes a valuable contribution by shifting evaluation of structure-based drug design models toward distribution learning, which better reflects the training objective of likelihood-based generative models. The paper ships a reproducible codebase, transparent derivations in the appendices, and a set of falsifiable distribution-matching benchmarks. The uncertainty estimate, if validated quantitatively, could be practically useful for ranking generated molecules. The preference alignment scheme is a sensible extension of DPO to multi-domain generative models. The main risk is that the headline SOTA distribution-learning claim rests on a comparison protocol that may not be apples-to-apples.","major_comments":[{"comment":"The paper states that ligands failing PoseBusters filters were removed from the training set and that DrugFlow samples were filtered by RDKit validity, but it does not state whether the baselines (Pocket2Mol, DiffSBDD, TargetDiff) were retrained on this filtered training set or whether their generated molecules were subjected to the same validity filter. Because every Wasserstein and Jensen-Shannon distance in Tables 1-3 is computed against the filtered training set, any mismatch in training distribution or in post-generation filtering would systematically inflate baseline distances. The bootstrap t-tests in Appendix B.10 resample from the same generated samples and cannot correct this systematic protocol difference. The authors should provide explicit details or re-run the baselines under identical conditions; without this, the central 'state-of-the-art distribution learner' claim is not fully supported.","section":"Section 3.1, Tables 1-3"},{"comment":"The paper claims that DrugFlow 'consistently achieves state-of-the-art distribution learning performance across various orthogonal metrics,' but the comparison includes only three baselines, all from 2022-2023. More recent structure-based drug design models are not compared, and the paper itself notes a few metrics (e.g., logP and hydrophobic interactions in Tables 1 and 2) where another method ranks first. The authors should either broaden the baseline set or temper the 'state-of-the-art' wording to accurately reflect the scope of the comparison.","section":"Section 5, Conclusion"},{"comment":"The out-of-distribution detection claim is supported only by qualitative histograms and correlation plots; no quantitative metric such as AUROC or precision-recall for OOD detection is reported. Given the abstract states that the uncertainty estimate is 'able to detect out-of-distribution samples,' the authors should provide a quantitative evaluation, especially because the uncertainty scores have a narrow range (0.85-0.92) as acknowledged in Section 3.2.","section":"Section 3.2"}],"minor_comments":[{"comment":"The empirical size prior p(N|M) is sampled from the training set; this is a practical heuristic rather than an end-to-end learned size selection. The phrase 'end-to-end size estimation' used in Section 2.2 and the Introduction should be qualified, as the model still requires a pre-specified maximum Nmax.","section":"Appendix A.5, Section 2.2"},{"comment":"The authors acknowledge that with Nmax=10 the model 'fails to remove more' than about 10 atoms and that a fully adaptive regime would require scaling up Nmax. This limitation should be stated clearly in the main text rather than only in the discussion of Figure 3.","section":"Section 3.3"},{"comment":"Tables 14-16 contain typographical issues where standard deviations are concatenated with means without proper spacing (e.g., '0.082 (0.004)0.045(0.006)'); these should be fixed for readability.","section":"Appendix B.10, Tables 14-16"},{"comment":"The MDPA loss in Eq. 5 is presented with weights lambda_c, lambda_w, lambda_l, but the relationship between Eq. 5 and the derivation in Appendix A.4 (Eqs. 47-49) is not fully spelled out in the main text, making the loss harder to follow.","section":"Section 2.4, Appendix A.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well-written and the appendices provide transparent derivations with code available. The main concern is the evaluation protocol for baselines. I recommend requesting a detailed description of the baseline setup (training data filtering, validity filtering of generated samples, sampling hyperparameters) as a condition of acceptance. If the authors cannot confirm that baselines were treated identically, the distribution-learning results should be presented with the qualification that the comparison is not fully controlled, and the SOTA claim should be softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jan,\n\nHere's my take on the DrugFlow paper. It is a genuinely strong method paper, and the new pieces are real: the per-atom uncertainty head with a clean MAP derivation, the virtual 'no atom' node type for adaptive molecule size, the side-chain angle flow on the torus with a polynomial scheduler, and the multi-domain DPO applied across flow and Markov bridge losses. The architecture is a sensible heterogeneous GNN with GVP layers. Appendix A's math is coherent, the empirical evaluation is broad with bootstrap and t-tests, and code is promised. This deserves a serious referee.\n\nThe soft spot is the baseline-comparability assumption. The paper removes training molecules that fail PoseBusters filters and filters DrugFlow samples by RDKit validity, but never says whether Pocket2Mol, DiffSBDD and TargetDiff were retrained on the filtered set or validity-filtered during evaluation. If they were not, every Wasserstein and JSD distance to the filtered training set is inflated for the baselines. That data-protocol gap is load-bearing for the SOTA distribution-learning claim; the bootstraps can't fix systematic bias. This is fixable, and I'd want it addressed before I believed the headline number. The OOD detection is suggestive but not a true OOD test—it's mostly tail-detection on the training distribution. The size adaptation experiment is honest about its 10-atom 'bandwidth.' The preference alignment results are interesting but come with a 10-20% validity drop.\n\nAlso, the 'Practical relevance' paragraph in the introduction is duplicated word-for-word, which is sloppy but not substantive.\n\nBottom line: the method is good, the math holds, and the components are reusable. The headline claim is conditional on an unverified baseline protocol. I'd accept it for peer review and push for a revised evaluation. I'd cite it for the components, but I'd be careful about citing the SOTA claim until the comparability question is resolved.\n\nBest.","headline":"A genuine method paper with reusable components, but the SOTA distribution-learning claim is conditional on an unverified baseline-comparability assumption.","tokens_in":29536,"tokens_out":2025,"would_cite":true,"duration_ms":21526,"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 introduces DrugFlow, which combines flow matching with Markov bridges, and claims it reproduces the chemical, geometric, and physical distribution of protein–ligand complexes better than three established baselines.","keywords":["de novo drug design","structure-based drug design","flow matching","Markov bridge","distribution learning","uncertainty estimation","preference alignment","protein flexibility"],"falsifier":"Retrain or fine-tune Pocket2Mol, DiffSBDD, and TargetDiff on the same filtered CrossDocked training set that DrugFlow used, then recompute the Wasserstein, Jensen-Shannon, and FCD distances to that filtered set; if any baseline then matches or beats DrugFlow's distances, the paper's central distribution-learning claim is falsified.","tokens_in":28511,"feed_emoji":"💊","tokens_out":6779,"duration_ms":64377,"temperature":0.7,"pith_summary":"The paper argues that generative models for structure-based drug design should be judged by how faithfully they reproduce the training distribution of protein–ligand complexes, not by absolute property scores. It introduces DrugFlow, a model that samples atom coordinates with conditional flow matching while generating atom and bond types with Markov bridges, and whose experiments show it tracks the filtered CrossDocked training set across geometry, drug-likeness, binding-efficiency, and interaction metrics more closely than Pocket2Mol, DiffSBDD, and TargetDiff. DrugFlow also adds an uncertainty head that flags out-of-distribution samples, a virtual atom type that lets molecule size be decided during sampling, and a flexible variant, FlexFlow, that simultaneously samples side-chain conformations. The paper further proposes a preference-alignment scheme, applied jointly to the continuous and discrete generative processes, that shifts sampling toward user-desired property values. If these results hold, the practical payoff is a generative model that can be retrained or fine-tuned on curated data to steer drug design while preserving physical plausibility.","feed_headline":"DrugFlow reproduces drug–protein data closer than three baselines","feed_subtitle":"A flow-matching plus Markov-bridge model tops chemistry, geometry, and binding-score distribution matches.","key_machinery":"The load-bearing machinery is a heterogeneous graph neural network that, at each denoising step, operates on a 3D graph with ligand atoms, protein residues represented as C-alpha nodes carrying full residue vector features, and three edge types, using geometric vector perceptrons to keep coordinate predictions equivariant. It drives two coupled generative processes: Euclidean conditional flow matching for atom coordinates and side-chain torsion angles on a torus, and Markov bridge models for atom and bond types. A virtual node type acts as a remove-this-atom label, so molecule size can emerge during sampling; an uncertainty head outputs per-atom sigma, and the loss is modified to make sigma large where the predicted vector field is poor; preference alignment uses a DPO-style contrast between winning and losing samples, computed as loss differences in the coordinate flow and the discrete Markov bridges. A NERF step converts sampled side-chain angles back to atomic coordinates at every training and sampling step.","core_discovery":"DrugFlow is a multi-domain generative model for structure-based drug design that learns the joint distribution of ligand atom coordinates, atom types, and bond types conditioned on a protein pocket. The central discovery claimed is that this combination—continuous flow matching for Euclidean coordinates and discrete Markov bridges for graph types—yields a distribution learner whose samples lie closer to the training distribution than Pocket2Mol, DiffSBDD, and TargetDiff, across bond lengths and angles, atom and bond type frequencies, ring systems, QED, SA, logP, rotatable bonds, Vina and Gnina efficiency, and normalized protein-ligand interaction counts. The paper reports the largest gap in Fréchet ChemNet Distance (4.278 versus at least 11.637 for baselines) and the lowest Jensen-Shannon divergence for the joint distribution of QED, SA, logP, and Vina efficiency. It also claims the model learns the conditional distribution of molecule sizes, detects out-of-distribution samples through per-atom uncertainty, and, in its FlexFlow extension, recovers bound side-chain rotamer distributions; a preference-alignment update improves target properties beyond fine-tuning, at a moderate cost in molecular validity.","pith_inferences":["An implicit testable extension is to scale the virtual-node bandwidth: the paper shows the model removes about Nmax/2 atoms on average and fails beyond that bandwidth, so training with a larger Nmax or a heavy-tailed virtual-node distribution should enable fully adaptive sizing on large, crowded pockets.","Because the uncertainty score is trained end-to-end with the flow-matching loss, it could double as a training signal: weighting the flow-matching loss by the inverse predicted variance is a natural next step the paper does not explore.","The same multi-domain alignment scheme should transfer to other mixed continuous-discrete generative tasks, such as linker design or co-design of ligands with protein backbones, whenever a DPO-style win/lose comparison can be defined per modality.","The comparison also leaves a protocol question open: the reported margin assumes the baselines were trained on the same filtered distribution as DrugFlow, so retraining the baselines on that filtered set would be the cleanest way to confirm the margin is a property of the model and not of the evaluation setup."],"forward_implications":["Distribution-centered evaluation becomes the primary benchmark for generative drug design: a model that wins on these distances can be retrained on curated data and will inherit the curated distribution's properties.","Molecule size no longer needs to be pre-specified: the virtual-node mechanism lets the model delete excess atoms during sampling, and clashes stay low within the trained deletion bandwidth.","Uncertainty scoring gives a usable out-of-distribution flag and correlates with ligand efficiency and steric clashes, so generated molecules can be ranked before expensive docking.","FlexFlow extends the same framework to side-chain conformational sampling, making target-aware generation possible for pockets that may be in unbound or flexible states.","The preference-alignment loss lets users push QED, SA, or Vina efficiency in a chosen direction while keeping the continuous and discrete domains coupled.",""],"supporting_citations":[{"why":"Supplies the conditional flow matching objective used for ligand coordinates.","marker":"Lipman et al., 2022"},{"why":"Supplies the Markov bridge model used for atom and bond types.","marker":"Igashov et al., 2023"},{"why":"Pocket2Mol, the autoregressive baseline in the distribution comparisons.","marker":"Peng et al., 2022"},{"why":"DiffSBDD, the diffusion baseline in the distribution comparisons.","marker":"Schneuing et al., 2022"},{"why":"TargetDiff, the diffusion baseline in the distribution comparisons.","marker":"Guan et al., 2023a"},{"why":"Provides the Riemannian flow matching framework for side-chain torsion angles.","marker":"Chen & Lipman, 2023"},{"why":"Provides geometric vector perceptrons, the equivariant building blocks of the neural network.","marker":"Jing et al., 2020"},{"why":"Introduces direct preference optimization, the foundation of the alignment loss.","marker":"Rafailov et al., 2023"},{"why":"Adapts DPO to diffusion models, which the paper extends to Markov bridges.","marker":"Wallace et al., 2024"},{"why":"Supplies the PoseBusters filters used to curate the training set and score validity.","marker":"Buttenschoen et al., 2024"}],"fun_headline_variants":["Flow matching plus Markov bridges for distribution-aware drug design","DrugFlow learns joint drug-protein distribution closer than baselines","Uncertainty-aware drug design detects out-of-distribution samples","Preference alignment improves drug properties without fine-tuning","FlexFlow samples drug molecules and protein side chains jointly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparison assumes the three baseline models were trained and evaluated on the same filtered training set as DrugFlow; if the published baselines were trained on the original, unfiltered data, their distances to the filtered set would be inflated by data curation rather than by model deficiency, and the claimed margin would shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["Flow matching plus Markov bridges for distribution-aware drug design","DrugFlow learns joint drug-protein distribution closer than baselines","Uncertainty-aware drug design detects out-of-distribution samples","Preference alignment improves drug properties without fine-tuning","FlexFlow samples drug molecules and protein side chains jointly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000389,"raw_usage":{"total_tokens":2012,"prompt_tokens":871,"completion_tokens":1141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":1063}},"tokens_in":487,"tokens_out":1141,"duration_ms":11547,"temperature":1.0,"reasoning_tokens":1063,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:00:22.355711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain or fine-tune Pocket2Mol, DiffSBDD, and TargetDiff on the same filtered CrossDocked training set that DrugFlow used, then recompute the Wasserstein, Jensen-Shannon, and FCD distances to that filtered set; if any baseline then matches or beats DrugFlow's distances, the paper's central distribution-learning claim is falsified.","supporting_citations":[],"review_version":2}