{"id":"a5cb5b7d-10a4-4cdc-bb6a-7c0a107f7e78","arxiv_id":"2509.05302","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Sesame uses flow matching to turn apo protein backbones into holo-like backbones, outperforming a Schrödinger-bridge baseline on geometric benchmarks but with limited docking validation.","lead":"Sesame is a machine-learning model that reshapes protein structures without a bound drug (apo) into the shape they adopt when a drug is bound (holo). If it works reliably, it could make drug-screening simulations much cheaper by replacing expensive molecular dynamics runs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pocket-mask ambiguity may violate ligand-agnostic generation and unfair baseline comparison.","rationale":"The reader's weakest assumption—that the pocket mask may be a conditioning input at inference—is exactly the most load-bearing concern. If the mask is used as a model input, the model is not ligand-agnostic and the comparison to SBAlign is unfair. This would undermine the headline empirical claims in Tables 1–3 and the subsequent docking and cryptic-pocket applications. The paper is genuinely ambiguous on this point: Appendix A.3 describes adding and transferring the mask, but Section 3 does not mention it as an architectural input. No code is provided to resolve the ambiguity. Other concerns, such as SBAlign hyperparameter tuning on test sets, are secondary because even a tuned SBAlign might not close the large performance gap; the mask issue, however, goes to the fundamental validity of the comparison. Since the ambiguity could be resolved by an ablation or code inspection, but is not resolved in the paper, the appropriate verdict remains conditional—hence no change from the reader's verdict.","tokens_in":13535,"tokens_out":4149,"duration_ms":44166,"concrete_test":"Obtain the model code or retrain Sesame with the mask removed as an input feature (using only the apo frames, possibly cropping by a fixed length or by sequence order). Then re-evaluate on the D3PM-Large and PDBBind-MD test sets. If the median ΔRMSD drops by more than 0.5 Å or the %RMSD<2.0 falls by more than 10 percentage points relative to Tables 1–2, this confirms the mask is a critical conditioning input and the reported ligand-agnostic performance is not achieved. If no drop occurs, the mask was likely only used for cropping and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Sesame is ligand-agnostic and outperforms baselines depends on how the pocket mask defined in Appendix A.3 is used. The text states that a mask for holo pocket residues (within 8 Å of the ligand) is 'added' and later 'transferred to the apo', with cropping to 512 residues, but it never states whether this mask is fed to the network as a conditioning feature or used only for data preprocessing. If the mask is an input feature at inference, then Sesame is not ligand-agnostic: it receives information about which residues form the binding pocket, information that is not available for a real apo target and is not given to the SBAlign baseline. This could trivially allow the model to focus its generative effort on the pocket, inflating the reported ΔRMSD and RMSD improvements over baselines. Section 3 describes the model architecture without any mention of a mask, making the omission conspicuous. The results in Tables 1–3 would then reflect information leakage rather than genuine apo→holo generation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Sesame, a flow-matching generative model that operates on SE(3)-equivariant backbone frames and is trained to map apo protein conformations to holo-like conformations. The method is evaluated on three datasets: D3PM-Large (large backbone motions), PDBBind-MD (a synthetic dataset of MD-collapsed pockets), and D3PM-Pocket (small pocket motions). Sesame is compared to SBAlign and EGNN, with metrics including Cα RMSD to the holo structure, ΔRMSD, and threshold success rates. Additional experiments assess Sesame-generated structures with PocketMiner for cryptic-pocket detection and with Glide docking on a small set of complexes. The paper claims that Sesame outperforms existing baselines on large and small conformational changes and that its generated conformations are more holo-like and more useful for docking than the original apo structures.","tokens_in":13737,"tokens_out":3981,"duration_ms":45720,"significance":"If the claims are correct, Sesame is a practically valuable contribution: it offers a fast, ligand-agnostic generative model for apo-to-holo backbone adaptation, with potential downstream benefits for virtual screening. The paper includes several strengths: evaluation on an external D3PM-Large test set, cross-inference experiments showing generalization limits, a downstream cryptic-pocket detection analysis, and a docking case study. The core methodology (flow matching on SE(3) frames with FAPE and auxiliary losses) is well grounded in prior work. However, the central claim of ligand-agnostic generation depends on an unresolved methodological ambiguity about how the pocket mask is used, and the baseline comparison involves test-set hyperparameter selection. These issues need to be resolved before the performance claims can be fully accepted.","major_comments":[{"comment":"The pocket-mask handling is load-bearing and ambiguous. Appendix A.3 states: 'we additionally add a mask in the holo structure for pocket residues' and later 'we transfer the holo-defined pocket mask to the apo', but Section 3, which describes the model architecture and losses, never states whether this mask is an input feature to the network or is used only for preprocessing/cropping (e.g., for the 512-residue limit). If the mask is a conditioning input at inference, then Sesame is not ligand-agnostic: it receives holo-derived information about which residues form the binding pocket, information that is unavailable for a real apo target and is not provided to SBAlign. This could directly inflate the reported RMSD and ΔRMSD improvements. Please explicitly specify how the mask is used in the model. If it is only a preprocessing/cropping device, say so unambiguously; if it is a model input","section":"Appendix A.3 and Section 3"},{"comment":"The SBAlign comparison uses test-set hyperparameter tuning. Section 4.2 states that 'inference was performed for the best model hyperparameters', and Appendix B reports results for several values of the diffusion coefficient g on the test sets, with the best-performing value used in Tables 1-3. Selecting hyperparameters based on the test set is a methodological flaw: it reports the upper envelope of SBAlign over the tested g values. This does not explain Sesame's advantage (indeed, it favors SBAlign), but it violates standard evaluation practice and weakens the claim of a fair baseline comparison. Please select g on a validation split or report results across g values without choosing the best based on the test set.","section":"Appendix B, Tables 9-10"},{"comment":"The PDBBind-MD test set is generated by the same synthetic MD collapse protocol used to create the training set, including the same minimum RMSD filter (0.5 Å) and the MD heating/equilibration parameters. Table 2 therefore measures performance on apo states that come from the same simulator as the training distribution, which limits the evidence for generalization to real apo structures. The D3PM-Pocket results (Table 3) partially mitigate this concern because they are an external real apo-holo benchmark, but the paper should explicitly acknowledge this limitation and, ideally, include additional real apo-holo test sets beyond D3PM-Pocket. As written, the abstract and conclusion generalize beyond the evidence in Table 2.","section":"Section 4.2 and Appendix A.2"}],"minor_comments":[{"comment":"Typo: 'substracting' should be 'subtracting'.","section":"Appendix A.3"},{"comment":"Each table heading lists two PDB codes (e.g., 'PDBs: 4ZZI, 4XKQ' and 'PDBs: 4LVT, 1GJH'), but the text discusses only one complex per table. Please clarify the role of the second PDB or remove it from the heading.","section":"Tables 4 and 12"},{"comment":"The indicator function is written as '1{t<0.25}' in Equation (15) but as 't <0.25' in the text of Appendix F. Use consistent notation, and make clear whether the auxiliary loss is applied only in the last 25% of the time steps.","section":"Equation (15)"},{"comment":"The paper notes that ApolloDiff is omitted due to lack of weights/code. This is understandable, but the claim of 'outperforms existing baselines' should be scoped precisely to the methods evaluated (SBAlign and EGNN); as written, the reader might infer a comparison with ApolloDiff.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The mask ambiguity is the most serious issue: if the pocket mask is a model input, the ligand-agnostic claim collapses and the comparison to SBAlign is unfair. The authors may well intend the mask only for cropping, but the text must say so explicitly. The test-set tuning of SBAlign's g is also a clear protocol violation that should be corrected. The paper is otherwise technically plausible and the external D3PM-Large and D3PM-Pocket evaluations provide some support for the claims; with these issues fixed, it could be a solid workshop contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on Sesame. The genuinely new bits are the application of FoldFlow-style SE(3) flow matching to the paired apo→holo task and the new PDBBind-MD dataset of synthetically collapsed pockets. On D3PM-Large, Sesame's numbers look strong—38% of predictions under 2 Å versus 0% for the retrained SBAlign—and the cross-inference analysis is refreshingly honest about the model's limits. Credit where due: this is a sensible baseline for future work in this space.\n\nThe soft spots, in proportion. First, the pocket-mask ambiguity is real. Appendix A.3 says the holo-derived 8 Å pocket mask is transferred to the apo and used to crop to 512 residues, but the methods section never says whether the mask is a conditioning input or just a cropping device. If it's only crop, fine, but the paper needs to say that—right now the 'ligand-agnostic' claim is muddy. If it were a conditioning feature, the ΔRMSD improvements could be inflated. The stress-test note raises exactly this, and I think it's more likely a documentation gap than a fatal flaw, but it's load-bearing enough that the authors must clarify.\n\nSecond, no ApolloDiff comparison. The stated reason—no weights/code—is legitimate, but it means 'outperforms existing baselines' actually means 'outperforms SBAlign and EGNN.' That's a weaker claim. Third, the SBAlign baseline had its diffusion coefficient tuned on the test set (Appendix B), so the comparison isn't clean. Fourth, the PDBBind-MD test set comes from the same MD collapse protocol used for training; the D3PM-Pocket results are the better external check, and there Sesame's advantage is much smaller (median ΔRMSD 0.18 vs 0.001 for SBAlign, but RMSD values are close and the 0.5 Å threshold shows only a modest gap). Finally, the docking evidence is two complexes with model RMSDs of 3.7 and 6.1 Å—not successes by usual standards—and the PocketMiner figures show qualitative improvement but no quantitative curve statistics.\n\nWho gets value? People building generative models for protein conformation change, and anyone needing paired apo–holo data for training. The central RMSD improvement on D3PM-Large likely holds, but the drug-discovery claims need more rigorous validation. It deserves a serious referee, not a desk reject—it's a useful contribution for a workshop, and the mask/cropping question is fixable. I'd want the revised version to state clearly how the mask is used, add ApolloDiff if possible, and soften the conclusions to match the evidence.\n\nRecommendation: send to peer review, expect major revision.","headline":"A promising workshop-scale paper that shows flow matching can turn apo backbones into more holo-like geometries, but the evaluation has enough loose ends—mask handling, missing ApolloDiff comparison, synthetic test data—that the headline claims outrun the evidence.","tokens_in":14286,"tokens_out":3021,"would_cite":false,"duration_ms":36064,"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":"Sesame is a generative model that converts unbound protein backbones into ligand-ready, holo-like shapes without taking the ligand as input.","keywords":["apo-to-holo conformational change","flow matching","SE(3) equivariant frames","generative protein model","molecular docking","virtual screening","cryptic pocket detection","protein backbone generation"],"falsifier":"Retrain or rerun Sesame with the pocket mask removed or replaced by a random residue subset at inference; if the RMSD/ΔRMSD values in Tables 1–3 drop to the apo baseline, the holo-derived mask is carrying the signal. A complementary test runs the pipeline on apo structures whose bound-state pocket is withheld and checks only whether residues near the unseen pocket match the holo geometry.","tokens_in":13382,"feed_emoji":"🧬","tokens_out":11905,"duration_ms":114754,"temperature":0.7,"pith_summary":"Sesame is a generative model that takes an unbound (apo) protein backbone and outputs a bound-like (holo) backbone, designed to be ligand-agnostic. The paper's goal is to show that this learned apo-to-holo transport is accurate enough to replace expensive molecular dynamics simulations for preparing docking targets. On the D3PM-Large set of large motions, Sesame reaches a median Cα-RMSD of 2.87 Å to the experimental holo structure (38% under 2 Å), versus 3.67 Å for the previous baseline; on the small-motion PDBBind-MD and D3PM-Pocket sets, median RMSD drops to 0.18 Å and 0.49 Å, respectively. Downstream, Sesame-generated structures improve cryptic-pocket detection and yield a docking pose closer to the crystal-bound pose than the apo structure in the reported complex. If these results hold, Sesame offers a cheap, scalable way to convert abundant unbound structures into holo-like inputs for virtual screening.","feed_headline":"A generative model turns unbound proteins into ligand-ready shapes","feed_subtitle":"Generated backbones get closer to bound structures, aiding docking and cryptic-pocket discovery from unbound inputs.","key_machinery":"The central object is the SE(3)-equivariant backbone frame: each residue is encoded as a rigid rotation and translation aligning an ideal backbone geometry to the observed atoms, so the whole protein becomes a set of frames in SE(3). A conditional flow-matching network learns a vector field that transports the apo frame distribution to the holo frame distribution, with geodesic interpolation on SO(3) and linear interpolation on R3. The model is trained with an SE(3)-flow-matching loss, a Frame Aligned Point Error term, and auxiliary backbone and pairwise-distance losses, and generates new conformations by integrating the learned ordinary differential equation. This decomposition is what lets","core_discovery":"Sesame frames the apo-to-holo transition as a generative transport problem. Residues are represented as SE(3)-equivariant frames—a rigid rotation and translation per backbone residue—and a conditional flow-matching network interpolates from the apo frame set to the holo frame set, using geodesic paths on SO(3) and straight lines in R3. On the D3PM-Large benchmark, the model reports a median Cα-RMSD to the holo structure of 2.87 Å and a median ΔRMSD of 2.15 Å, with 38% of generated structures under 2 Å, outperforming the SBAlign baseline. On the newly built PDBBind-MD set for small pocket motions, median RMSD is 0.18 Å (82% under 0.2 Å), and on the D3PM-Pocket set it is 0.49 Å, again ahead of","pith_inferences":["The same flow-matching transport could be trained on other paired conformational transitions, such as allosteric state changes or protein-protein binding, where one state is easy to obtain and the other is scarce.","A single model trained on both large and small motions might close the cross-inference gap the paper observes; the authors' planned expansion of MD-generated data is the natural test.","Because only backbones are generated, the docking evaluation depends on a single side-chain reconstruction; evaluating over multiple reconstruction seeds would show how much of the remaining RMSD is side-chain error rather than backbone error."],"forward_implications":["Apo-only targets can be flipped into holo-like starting structures for docking, so virtual screening no longer needs to wait for an experimentally solved ligand-bound complex.","Sesame's runtime is a fraction of an MD protocol, making it practical to pre-process large libraries of protein targets before screening.","The reported gains on D3PM-Large, PDBBind-MD, and D3PM-Pocket imply the method works across motion scales, not just on large hinge-like movements.","Generated backbones improve cryptic-pocket detection, which could make allosteric and transient sites targetable from ordinary unbound structures.","Because only backbones are generated, downstream side-chain reconstruction remains a separate step; the paper's results suggest that better side-chain modeling would close most of the remaining gap to holo."],"supporting_citations":[{"why":"Defines conditional flow matching, the objective Sesame optimizes.","marker":"Lipman et al., 2023"},{"why":"Extends flow matching to Riemannian manifolds, enabling geodesic interpolation in SO(3).","marker":"Chen & Lipman, 2024"},{"why":"Supplies the SE(3)-frame backbone representation and the Frame Aligned Point Error loss.","marker":"Jumper et al., 2021"},{"why":"Provides the SE(3)-flow-matching training framework and loss terms Sesame adapts.","marker":"Bose et al., 2024"},{"why":"Supplies the auxiliary backbone and pairwise-distance losses that refine local geometry.","marker":"Yim et al., 2023"},{"why":"Supplies the D3PM database with the apo-holo pairs used in D3PM-Large and D3PM-Pocket.","marker":"Peng et al., 2022"},{"why":"Defines the SBAlign baseline and the D3PM-Large protocol against which Sesame's large-motion results are compared.","marker":"Somnath et al., 2023"},{"why":"Defines the cryptic-pocket test set and the prediction model used to show generated structures are more holo-like.","marker":"Meller et al., 2023"},{"why":"Establishes that apo structures dock worse and supplies the PDBBind temporal split used to construct the MD-generated training set.","marker":"Corso et al., 2023"}],"fun_headline_variants":["Generative model shapes unbound proteins for docking","Apo to holo in one step: Sesame speeds docking prep","Sesame predicts ligand-ready pockets from apo structures","Flow-matching model opens protein pockets for screening","Sesame: Fast track from unbound to ligand-ready backbones"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper's ligand-agnostic claim rests on the assumption that the holo-derived pocket mask—which residues lie within 8 Å of the ligand in the bound structure—is not an input at inference; if it is, the reported gains could reflect knowledge of the binding site rather than a pure unbound-to-bound transformation.","fun_headline_variants_meta":{"raw":{"variants":["Generative model shapes unbound proteins for docking","Apo to holo in one step: Sesame speeds docking prep","Sesame predicts ligand-ready pockets from apo structures","Flow-matching model opens protein pockets for screening","Sesame: Fast track from unbound to ligand-ready backbones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1151,"prompt_tokens":679,"completion_tokens":472,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":404}},"tokens_in":423,"tokens_out":472,"duration_ms":4657,"temperature":1.0,"reasoning_tokens":404,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:52:06.230188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain or rerun Sesame with the pocket mask removed or replaced by a random residue subset at inference; if the RMSD/ΔRMSD values in Tables 1–3 drop to the apo baseline, the holo-derived mask is carrying the signal. A complementary test runs the pipeline on apo structures whose bound-state pocket is withheld and checks only whether residues near the unseen pocket match the holo geometry.","supporting_citations":[{"cited_title":"SE(3) -stochastic flow matching for protein backbone generation","cited_arxiv_id":null,"evidence_quote":"Extends flow matching to Riemannian manifolds, enabling geodesic interpolation in SO(3)."},{"cited_title":"SE(3) -stochastic flow matching for protein backbone generation","cited_arxiv_id":null,"evidence_quote":"Provides the SE(3)-flow-matching training framework and loss terms Sesame adapts."},{"cited_title":"D3PM : a comprehensive database for protein motions ranging from residue to domain","cited_arxiv_id":null,"evidence_quote":"Supplies the D3PM database with the apo-holo pairs used in D3PM-Large and D3PM-Pocket."},{"cited_title":"Aligned diffusion schr\\\"odinger bridges","cited_arxiv_id":null,"evidence_quote":"Defines the SBAlign baseline and the D3PM-Large protocol against which Sesame's large-motion results are compared."},{"cited_title":"Lotthammer, Felipe Oviedo, Juan Lavista Ferres, and Gregory R","cited_arxiv_id":null,"evidence_quote":"Defines the cryptic-pocket test set and the prediction model used to show generated structures are more holo-like."},{"cited_title":"Jaakkola","cited_arxiv_id":null,"evidence_quote":"Establishes that apo structures dock worse and supplies the PDBBind temporal split used to construct the MD-generated training set."}],"review_version":1}