{"id":"705e81d0-c49f-419c-a829-caab4c031a36","arxiv_id":"2607.25156","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Vilya-2 predicts bound structures of chemically diverse peptides and small molecules at state-of-the-art accuracy using an all-atom diffusion transformer, recovering 59.1% of peptide interfaces to sub-2 Å backbone RMSD.","lead":"Vilya-2 is an AI model that predicts the 3D shapes of peptides and small molecules bound to proteins, handling unusual chemical modifications that standard models miss. It outperforms existing co-folding models on a new benchmark of 88 peptide complexes and generalizes to molecules much larger than anything in its training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training-set leakage through CPSea augmentation may inflate the 59.1% Riptides result; the 2021-09-30 cutoff is only stated for PDB entries, not for CPSea cyclic-peptide complexes.","rationale":"Read in good faith: the paper proposes a genuinely interesting all-atom diffusion transformer, and the Riptides benchmark release is a useful community resource. The central quantitative claim is the 59.1% sub-2Å peptide interface recovery and the improvement over Boltz-2. That claim requires that the Riptides benchmark be held out from training. The manuscript says training was restricted to PDB entries released on or before 2021-09-30, but only after mentioning that protein-peptide complexes from CPSea were added. The sentence \"training was restricted to PDB entries released on or before 2021–09–30\" is grammatically ambiguous: it could mean all PDB-sourced training data (including CPSea) were filtered, or only the PDB-derived complexes. CPSea is a cyclic peptide-protein complex dataset, and Riptides is dominated by non-canonical cyclic peptides. If CPSea contains complexes released after the cutoff, the held-out status of Riptides is compromised. This is the load-bearing weakness because it directly affects the main numerical result; the reader's concern about manual curation is related but secondary. I recommend keeping the conditional verdict, with the explicit condition that the authors provide a CPSea release-date audit and a similarity screen against Riptides. The paper should also clarify whether the 'Oracle' miniprotein results are intended as evidence of practical prediction or merely as sampling capability.","tokens_in":16410,"tokens_out":12237,"duration_ms":114009,"concrete_test":"Audit the CPSea training split: obtain the complete list of PDB IDs and release dates in CPSea, and confirm the 2021-09-30 filter was applied to all CPSea-derived complexes. Then run a similarity screen between each of the 88 Riptides entries and every CPSea/PDB training complex (e.g., sequence identity on the peptide, ligand maximum common substructure, and TM-score on the receptor pocket). If any Riptides entry has a close analog (e.g., identical peptide sequence or receptor pocket TM-score >0.8), re-run the Riptides comparison after excluding those entries. If the 59.1% success rate or the margin over Boltz-2 drops materially, the central claim is an artifact of leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 states the training data was augmented with protein-peptide complexes from CPSea [28], then says \"training was restricted to PDB entries released on or before 2021–09–30.\" It is ambiguous whether this cutoff applied to CPSea-derived complexes or only to the PDB-derived portion. Because Riptides is released after 2023-06-01 and 51/88 entries are non-canonical cyclic peptides—exactly the modality CPSea covers—if CPSea was not filtered by release date, Riptides may contain near-duplicates of training complexes. That would inflate the headline 59.1% success rate and the gap over Boltz-2, making the central generalization claim untrustworthy. This is a concrete, checkable missing detail, not an ad hominem accusation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Vilya-2, a diffusion transformer that represents proteins, peptides, and small molecules uniformly as atomic graphs rather than residue-level tokens. It is trained for two tasks: single-molecule conformer generation and target-conditioned interface prediction, with a shared architecture later fine-tuned for confidence estimation, activity prediction, and property prediction. The central empirical claims are: (i) on a new, author-curated Riptides benchmark of 88 protein–peptide complexes, Vilya-2 recovers 54.1% of bound peptide backbones below 2 Å RMSD at 100 samples and 59.1% at 1000 samples, outperforming Boltz-2 even when Boltz-2 is given the receptor crystal structure as a template; (ii) Vilya-2 is state-of-the-art at small-molecule docking on PoseBusters, Runs N' Poses, and PoseX, including cross-docking; (iii) the conformer generator generalizes to macrocycles and disulfide-stapled miniproteins larger than training examples; and (iv) fine-tuned heads improve enrichment in an internal hit-to-lead campaign. The paper also releases the Riptides benchmark and benchmarking code.","tokens_in":16646,"tokens_out":5837,"duration_ms":58537,"significance":"If the results hold, Vilya-2 would be a meaningful advance: it demonstrates that a single all-atom representation can support peptide, small-molecule, macrocycle, and miniprotein modeling without residue-level or MSA-derived features, and that inference-time sampling with calibrated confidence scoring is effective for interface prediction. The paper includes useful controls: template-conditioned Boltz-2 baselines, Schrödinger docking comparisons, PoseBusters physical-validity filtering, explicit calibration plots, and sampling-scaling analysis. The release of the Riptides benchmark is a community asset. However, the headline generalization claims rest on a self-curated benchmark with an ambiguous training-data cutoff for the CPSea augmentation, and the reported confidence intervals are computed by resampling poses rather than benchmark targets. These issues are checkable and fixable, but they are load-bearing for the paper's central quantitative claims.","major_comments":[{"comment":"The training-data cutoff is ambiguous. The text says the dataset was expanded with protein–peptide complexes from CPSea and then states: 'training was restricted to PDB entries released on or before 2021–09–30.' It is not clear whether this restriction applies to the CPSea-derived complexes as well. This matters directly because Riptides — the benchmark used for the headline 59.1% result — is constructed from PDB entries released after 2023-06-01 and contains 51/88 non-canonical cyclic peptides, exactly the modality covered by CPSea. If CPSea complexes were not filtered by the 2021–09–30 cutoff, near-duplicates of training complexes can appear in the test set. Please (a) state explicitly whether the September 2021 cutoff was applied to CPSea-derived complexes; (b) if it was not, rerun the Riptides evaluation with any CPSea-derived training complexes released after 2021-09-30 excluded, or","section":"§2.2"},{"comment":"The reported '95% confidence intervals' are obtained by resampling 100 poses from a fixed pool of 1000 poses per target, then recomputing the success rate over the 88 targets. This procedure captures only pose-selection stochasticity; it does not reflect uncertainty about the benchmark success rate or about the difference between Vilya-2 and Boltz-2 across targets. For example, the Riptides 2 Å estimate of 54.1 ± 4.6% would have a substantially wider target-level binomial interval given n = 88. Please either relabel these intervals as within-system sampling variability, or add target-level bootstrap/binomial confidence intervals for the headline numbers and for the Vilya-2 vs. Boltz-2 comparison.","section":"§2.4"},{"comment":"The text states: 'in all of these cases, none of the experimental structures we compare Vilya-2’s predictions against were present in the training data on which the model was trained.' At least one of the comparison structures, PDB 7S5G (the progenitor of Lipfendra), may have been released before the 2021–09–30 training cutoff; the manuscript does not provide release dates for the comparison structures. Please verify the release date of 7S5G and of the internal structures, and provide a table of release dates. If any comparison structure is in the training set, qualify the claim accordingly or exclude that comparison.","section":"§3.1, Figure 1A"},{"comment":"The Riptides construction includes a manual curation step: entries with unresolved residues or 'inadequate contact' with the receptor are dropped. The criteria for 'inadequate contact' are not quantified, and the manual step is performed by the authors who are also proposing the benchmark and evaluating their own model. Because this benchmark is the primary evidence for the central claim of peptide-modeling superiority, please (a) release the full list of candidate entries with the reasons for exclusion, (b) define the contact criterion in terms of a reproducible threshold (e.g., number of atom pairs, buried surface area), and (c) report whether the Vilya-2 vs. Boltz-2 comparison is stable under alternative inclusion thresholds or when the manually excluded entries are included.","section":"§2.3"}],"minor_comments":[{"comment":"There is a typo in the final sentence: 'the design and evaluation of de novo of peptide therapeutics' should read 'de novo peptide therapeutics'.","section":"Abstract"},{"comment":"The text uses 'released on or before 2021–09–30' for the PDB restriction but does not define whether 'released' means PDB release date or deposition date. Please state which date is used and apply the same convention consistently.","section":"§2.2"},{"comment":"The columns 'PB-Valid = True' and 'PB-Valid prefilter' are clear from the text but would benefit from a one-line definition in the table caption: the former excludes poses that fail PoseBusters checks, while the latter discards invalid poses before confidence ranking so that the output always passes all checks.","section":"Table 1"},{"comment":"The Riptides selection criteria refer to 'at least 20 unique atom pairs with a distance < 5 Å between the ligand and its closest protein receptor chain.' It would be useful to state whether this includes only heavy atoms and how symmetry-related copies of the ligand were handled.","section":"§2.3"}],"recommendation":"major_revision","confidential_remarks":"The CPSea training-cutoff ambiguity is the most important issue: if CPSea data were not date-filtered, the Riptides results could be substantially inflated, and the manuscript would need re-evaluation. I do not see evidence of deliberate circularity, but the ambiguity is serious enough to require clarification and possible recomputation. The paper is otherwise strong and well within scope for the journal; I recommend major revision rather than rejection because the concern is concrete, checkable, and fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"John—quick take on Vilya-2. The paper is worth engaging: it extends a unified all-atom diffusion transformer from single-molecule conformer generation to protein-ligand interface prediction, and it does a lot of the evaluation right. Bootstrapped confidence intervals, calibration plots, sampling curves, head-to-head against Boltz-2 (with templates) and Schrödinger's dockers, and a released benchmark (Riptides) all are above the usual bar for this literature. The architectural choices—removing triangle attention, more 1D updates, cuEquivariance kernels—are concrete and get real speedups. The generalization claims on macrocycles and miniproteins are supported by external structures.\n\nBut there is one spot that could be load-bearing: the training-data cutoff. In Section 2.2 the paper says training was augmented from the PDB and from CPSea, then says \"training was restricted to PDB entries released on or before 2021-09-30.\" That phrasing suggests the date filter was applied only to PDB entries, not to CPSea complexes. Riptides is built from PDB entries released after 2023-06-01, and 51 of its 88 entries are non-canonical cyclic peptides—exactly what CPSea covers. If CPSea was added without the same cutoff, then Riptides could contain near-duplicates of training complexes, which would inflate the 59.1% success rate and the gap over Boltz-2. This is not an accusation; it is a missing detail that is easy to fix and should be fixed before anyone trusts the headline.\n\nThe reader's take is otherwise reasonable. The self-curated Riptides benchmark with manual exclusions is a moderate concern, but the paper at least documents the filters. The absence of model weights or inference code is a real reproducibility gap for a \"foundation model\" paper. And the abstract overclaims (\"unprecedented,\" \"paradigm shift\")—ignore that. The activity-prediction section uses proprietary data, but that is a common arrangement and not a fatal flaw.\n\nSo: the central idea is clean, the evaluation is mostly careful, and the one serious technical question is empirical. If the CPSea cutoffs are clarified and the numbers are re-run on a leak-free split, this could be a strong paper. If the leakage is real, the 59.1% and the comparative claims need to be recomputed. Either way it deserves a serious referee, but a referee who will force the date-filter issue to be addressed.\n\nI would bring it to our reading group. I'd cite it if I were working on peptide structure prediction—if only for the Riptides benchmark—but with a note about the unresolved leakage question.\n\nRecommendation: send it out; don't desk reject.","headline":"Vilya-2 is a real advance, but the CPSea training-date ambiguity could poison the headline Riptides number.","tokens_in":17141,"tokens_out":3960,"would_cite":true,"duration_ms":36391,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Vilya-2, a diffusion transformer that represents all molecules as atom-and-bond graphs, predicts the bound structures of chemically diverse peptides and small molecules with accuracy that far exceeds co-folding baselines, even when those ba","keywords":["all-atom representation","diffusion transformer","protein–peptide interface prediction","conformer generation","macrocycles","confidence calibration","small-molecule docking","foundation model"],"falsifier":"Run Vilya-2 on all post-2023-06-01 PDB peptide–protein complexes that pass only automated filters (resolution < 3 Å, at least 20 atom contacts within 5 Å, receptor chain ≥ 50 residues) without the authors' manual exclusion of unresolved residues or weak-contact entries, comparing success rates to Boltz-2 under identical template conditioning; if the accuracy gap narrows to near parity, the benchmark curation rather than the model would be the source of the reported advantage.","tokens_in":16335,"feed_emoji":"💊","tokens_out":5152,"duration_ms":42786,"temperature":0.7,"pith_summary":"The paper introduces Vilya-2, a diffusion transformer that represents peptides, small molecules, and protein targets identically as atom-and-bond graphs, with no residue tokens or molecule-type annotations. The authors claim this unified all-atom representation, combined with diverse pose sampling and a calibrated confidence score, lets the model predict bound peptide conformations far more accurately than co-folding models such as Boltz-2, even when those models are given the receptor crystal structure as a template. On the new Riptides benchmark, Vilya-2 recovers the peptide backbone within 2 Å RMSD in 59.1% of cases with 1000 samples, versus 40.9% for Boltz-2. The same architecture generalizes to small-molecule docking, macrocycles, and miniproteins larger than any training example, and can be fine-tuned to rank active compounds in drug-discovery campaigns.","feed_headline":"Vilya-2 recovers 59% of peptide interfaces under 2 Å","feed_subtitle":"Treating every molecule as one atom graph beats co-folding baselines and generalizes to unseen chemistries.","key_machinery":"The key machinery is the unified all-atom graph representation inherited from Vilya-1: the input is a chemical graph whose nodes are heavy atoms and whose edges are covalent bonds, with no residue tokens, molecule-type annotations, or multiple-sequence-alignment information. A diffusion transformer runs the diffusion process through the whole architecture and can generate diverse structural ensembles; a separately trained confidence head predicts per-atom lDDT from the chemical graph and predicted coordinates, and this plddt-ligand score ranks the sampled poses. For interface prediction, the model is additionally conditioned on a sparse distance matrix derived from receptor Cα and nucleic-ac","core_discovery":"The paper's central claim is that a single all-atom, residue-free chemical representation—where a peptide, a macrocycle, a small molecule, and a protein are all atom-and-bond graphs—combined with ensemble sampling and a calibrated confidence score, is enough to accurately model protein–ligand interfaces across diverse chemistries. On the authors' new Riptides benchmark, Vilya-2 recovers the bound peptide backbone within 2 Å RMSD for 59.1% of cases when 1000 poses are sampled and ranked, compared with 40.9% for the co-folding baseline Boltz-2 even when Boltz-2 is given the receptor crystal structure as a template. The paper further claims that this advantage comes from both diversity in sampl","pith_inferences":["If the all-atom representation delivers on its promise, the traditional separation between protein structure prediction, docking, and conformer generation may collapse into one modeling task, simplifying the software stack of structure-based drug design.","The receptor backbone conditioning means Vilya-2 assumes a known or reliably modeled target structure; a natural stress test is fully unbound receptor prediction with no structural prior, which is not covered by the reported benchmarks.","The PoseBusters prefilter—discarding physically invalid poses before confidence ranking—is a simple, general recipe that other generative structure models could adopt, and it suggests that physical-validity filters complement learned scoring rather than compete with it.","A testable extension would be to apply Vilya-2 to peptides with backbone N-methylation or ester linkages and to covalent inhibitors, to check whether the atom-graph representation handles bond-order and chirality variations beyond those in Riptides."],"forward_implications":["Peptide therapeutics with non-canonical residues, macrocyclic topologies, and unusual covalent linkages can be structurally modeled without retraining or residue-level featurization, making them addressable by de novo design pipelines.","Inference-time scaling through diffusion sampling is a practical lever for interface accuracy: Vilya-2 keeps improving at 1000 samples on peptide targets, whereas co-folding baselines do not benefit from extra sampling.","A well-calibrated confidence score provides an absolute stop-criterion—designers can trust a pose with high plddt-ligand and know when more sampling is needed, rather than relying on relative ranking alone.","The same pretrained model transfers across small-molecule docking, conformer generation of large organics, and macrocycles, suggesting a single foundation model can serve multiple modalities in drug discovery.","Fine-tuning on fewer than a thousand experimental potency measurements yields roughly 3× enrichment for active compounds, indicating that the learned structural representations carry signal relevant to binding affinity."],"fun_headline_variants":["Vilya-2: One atom graph to model all drug chemistries","All-atom model beats co-folding on peptides and docking","Vilya-2 generalizes to macrocycles, miniproteins, small molecules","Beyond proteins: Vilya-2 models diverse interfaces accurately","Vilya-2: 59% peptide interfaces under 2 Å, beats Boltz-2"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the Riptides benchmark, manually curated by the authors from PDB entries after mid-2023 with unresolved residues and weak-contact entries removed, fairly represents the distribution of therapeutically relevant peptide interfaces and does not systematically favor Vilya-2 over co-folding baselines; if the curation biases the comparison, the reported accuracy gap could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Vilya-2: One atom graph to model all drug chemistries","All-atom model beats co-folding on peptides and docking","Vilya-2 generalizes to macrocycles, miniproteins, small molecules","Beyond proteins: Vilya-2 models diverse interfaces accurately","Vilya-2: 59% peptide interfaces under 2 Å, beats Boltz-2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1521,"prompt_tokens":858,"completion_tokens":663,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":602,"tokens_out":663,"duration_ms":5929,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T03:15:47.762729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Vilya-2 on all post-2023-06-01 PDB peptide–protein complexes that pass only automated filters (resolution < 3 Å, at least 20 atom contacts within 5 Å, receptor chain ≥ 50 residues) without the authors' manual exclusion of unresolved residues or weak-contact entries, comparing success rates to Boltz-2 under identical template conditioning; if the accuracy gap narrows to near parity, the benchmark curation rather than the model would be the source of the reported advantage.","supporting_citations":[],"review_version":1}