{"id":"082589e8-6a7f-47aa-83ae-68c4c45dcc3b","arxiv_id":"2506.14488","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"READ couples contrastively aligned latent diffusion with pocket-similarity retrieval to generate 3D ligands, reporting Rank 1 on CBGBench and lower Vina energies than native ligands.","lead":"A drug-design model called READ combines retrieval of similar pocket-ligand templates with an equivariant diffusion generator, and reports top scores on the CBGBench benchmark. The paper argues that aligning a latent chemistry space with retrieved scaffolds yields valid, pocket-fitting ligands, but the benchmark protocol and the claimed 'first' status need scrutiny.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Rank 1 result may rest on retrieval of near-native templates: the retrieval graph is built from the CBGBench training split without any reported train/test pocket-similarity analysis, so the 'surpass native ligands' gains could reflect template recall rather than generative generalization.","rationale":"The reader identified the same load-bearing concern: the evaluation assumes the CBGBench training split used for retrieval does not contain pockets so similar to test targets that near-native templates are supplied. My reading of the paper confirms this is the weakest point of the central claim. The method is otherwise internally coherent, the equations are plausible, and the ablations show that alignment and retrieval both contribute. However, the entire Rank 1 and 'surpass native' argument depends on interpreting the retrieval-augmented generation as a genuinely generative process rather than a template-recall mechanism. The paper provides no analysis of train/test pocket similarity, no control retrieval condition, and no release of code or data to independently verify the split. These omissions make the central claim conditional. I agree with the reader's verdict of CONDITIONAL and do not see a reason to change it; the concrete test proposed would either substantiate or refute the concern. Secondary issues such as the overstated validity-rate claim and lack of error bars further support a conditional verdict but are not the single most load-bearing concern.","tokens_in":13442,"tokens_out":4119,"duration_ms":43256,"concrete_test":"For every test pocket, compute the structural similarity (TM-score from TM-align or Z-score from DaliLite) to its top-40 retrieved training pockets, and correlate this similarity with the per-pocket Dock-mode improvement over the native ligand. Then run a control experiment: rebuild the retrieval graph using only training pockets whose similarity to the query is below a strict threshold (e.g., TM-score < 0.5), and compare READ-2k's mean EVina, MPBG, and Rank 1 status against the result with unrestricted retrieval. If the performance collapses to the no-RAG baseline or loses Rank 1, the central claim is explained by near-native template retrieval rather than generative alignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that READ achieves Rank 1 on CBGBench and surpasses native ligands because retrieval and diffusion are co-optimized. The load-bearing assumption is that the retrieval graph, constructed from the CBGBench training split (Section 3.3(i)), does not supply near-native templates for test pockets. The paper never reports the distribution of TM-align/DaliLite similarities between test pockets and their retrieved training pockets, and it provides no control experiment (e.g., retrieval from dissimilar pockets or shuffled ligand labels). Because sampling retrieves top-4 ligands from the 40 nearest neighbors of each query pocket and injects their embeddings at every reverse step (Section 3.3), if a test pocket is homologous to a training pocket, the retrieved ligands can be structurally similar to the native binder. The reported mean Vina improvements and MPBG in Table 1 and Figure 5 are computed from the best of 100 generated molecules after drug-likeness filtering, which maximizes the chance that a near-native template, if retrieved, is selected. Consequently, the observed 'surpass native' gains could be an artifact of template recall rather than evidence of generative generalization. The ablations in Table 2 do not address this, since they vary alignment and retrieval presence but keep the same retrieval graph and therefore the same potential for template leakage. Secondary overclaims (e.g., 'validity rate nearly twice' while Table 1 shows 0.98 vs 0.96) do not affect the ranking argument but are inaccurate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces READ, a retrieval-enhanced SE(3)-equivariant diffusion model for pocket-conditioned de novo ligand generation. A contrastive encoder pretrained on six million ZINC molecules defines a latent chemical space; at inference, a pocket-ligand graph built from the CBGBench training split is queried with TM-align and DaliLite, and the top-4 retrieved ligand embeddings are injected into the denoiser at every reverse-diffusion step. The model is evaluated on CBGBench against twelve baselines using AutoDock Vina scores, PLIP interaction fidelity, validity, and substructure statistics. The authors claim that READ-2k achieves Rank 1 in the benchmark, surpasses native ligands for most pockets, and that retrieval and diffusion are co-optimized.","tokens_in":13777,"tokens_out":6123,"duration_ms":64979,"significance":"If the reported gains are genuine, the combination of contrastive latent alignment with pocket-matched retrieval is a useful and timely contribution to structure-based drug design, and the paper gives a reasonably detailed account of the method and ablations. The work is one of relatively few attempts to integrate retrieval-augmented generation with 3D diffusion for ligand design, and it explicitly addresses chemical validity through latent alignment rather than handcrafted post hoc filters. However, the central Rank 1 claim is not yet established: the evaluation protocol leaves open a template-leakage pathway from the retrieval graph, the best-of-100 selection is not shown to be applied identically to baselines, and no error bars or significance tests support the ranking. The paper also contains a clear factual overstatement about validity rates. These issues are fixable in principle, but they currently prevent the central claim from being accepted as stated.","major_comments":[{"comment":"The retrieval graph is constructed from the CBGBench training split, and at inference each query pocket retrieves the top-4 ligands from its 40 nearest neighbors. The paper never reports the TM-align/DaliLite similarity between test pockets and the retrieved training pockets, nor does it provide a control (for example, retrieval from structurally distant pockets, or shuffled ligand labels). If test pockets share detectable homology with training pockets, the top-4 templates can be near-native binders, and the best-of-100 selection used in Fig. 5 would make the 'surpass native' gains and Rank 1 in Table 1 an artifact of template recall rather than generative generalization. Please report the similarity distribution between test pockets and retrieved training pockets and run the controls; the ablations in Table 2 do not address this issue because the retrieval graph is fixed across all rows.","section":"Sec. 3.3(i), Fig. 5, Table 1"},{"comment":"All READ docking numbers are reported as the best of 100 generated molecules after filtering, but the manuscript does not state how many samples each baseline generated or whether the same best-of-N selection protocol was applied. Without this information, the comparisons in Table 1 and the improvement rates in Fig. 5 are not apples-to-apples. In addition, no error bars, independent seeds, or significance tests are reported for any metric, so the Rank 1 claim is currently supported only by point estimates and a composite score in which Dock-mode metrics are arbitrarily double-weighted. Please provide per-seed variability, statistical comparisons, and a sensitivity analysis of the ranking weights.","section":"Sec. 4, Table 1"},{"comment":"The Discussion states that READ-2k has 'a validity rate nearly twice that of the next best method.' Table 1 reports validity 0.98 for READ-2k and 0.96 for TargetDiff (with MolCraft at 0.95), so the improvement is a few percent, not nearly twofold. This is a factual overstatement that should be corrected in the Discussion, and it also affects the paper's overall framing of the method's advantage in chemical validity.","section":"Sec. 5, Discussion vs. Table 1"}],"minor_comments":[{"comment":"The text says 'LiGAN retains the best raw Vina scores under Score and Minimize modes,' but LiGAN appears neither in the baseline list nor in Table 1. The table shows VoxBind as the best for Vina Score and Vina Min. Please clarify which model is meant or remove the sentence.","section":"Sec. 4.1, Table 1"},{"comment":"The claim that READ is 'the first retrieval-augmented diffusion framework for de novo ligand design' should be reconciled with the discussion of MolR [40], which is described as projecting retrieved molecules into pre-trained latent spaces to inform diffusion trajectories; please clarify the precise novelty claim and cite any closer prior work.","section":"Sec. 2, Related Work"},{"comment":"Reference [35] is cited for retrieval-augmented generation but points to 'Human RAG mutations: Biochemistry and clinical implications,' which is unrelated; please replace it with an appropriate RAG citation, such as Lewis et al. (2020) or a comparable source.","section":"References"},{"comment":"The contrastive objective is written with g as the original graph embedding and g+ as the augmented positive, but the text says the encoder processes 'view pairs'; please clarify which augmentation is treated as the anchor and which as the positive in Eq. (5).","section":"Sec. 3.2, Eq. (5)"},{"comment":"The sentence 'We release two READ variants' should include the repository or model availability information; currently no code or data link is provided, which limits reproducibility.","section":"Sec. 4, Sampling cost"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about template leakage is real and load-bearing: the retrieval graph is built from the CBGBench training split with no reported pocket-similarity analysis between training and test targets, and the best-of-100 selection makes the reported gains maximal. I would ask the authors for a control experiment and a full description of the baseline sampling protocol before this can be accepted. The validity-rate overstatement and the phantom 'LiGAN' baseline should also be corrected. The novelty overlap with MolR and the irrelevant citation [35] should be checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is the specific combination: contrastive latent alignment, hierarchical pocket retrieval, and SE(3)-equivariant diffusion, tested on CBGBench. The ablations in Table 2 are the strongest part of the paper—they show that alignment alone does little, retrieval alone can hurt, and the two together produce a large jump. That is a genuine, nontrivial interaction, and it deserves credit.\n\nBut the central Rank-1 claim has a soft spot the paper never addresses. The retrieval graph is built from the CBGBench training split (Sec. 3.3(i)), and the method retrieves the top-4 ligands from the 40 nearest pockets to the query. If test pockets are homologous to training pockets, those retrieved ligands can be near-native templates, and the reported 'surpass native ligands' gains—computed as the best of 100 generated molecules after drug-likeness filtering—could reflect template recall rather than generative skill. The paper reports no similarity analysis between test pockets and retrieved training pockets, and no control experiment (e.g., retrieval from dissimilar pockets, or shuffled ligand labels). The stress-test note has this exactly right. It is an omission, not a demonstrated flaw, but it is load-bearing for the strongest claim.\n\nTwo overclaims are easy to confirm and should be fixed. The Discussion says READ has a 'validity rate nearly twice that of the next best method,' but Table 1 shows 0.98 vs. 0.96—not nearly twice. And the abstract's 'first to merge molecular RAG with SE(3)-equivariant diffusion' is contradicted by the paper's own citations: f-RAG and KGDiff are both retrieval-guided diffusion methods. Those need rephrasing.\n\nThere are also standard but real weaknesses: no error bars or significance tests, no code or data released, and the best-of-100 selection protocol is described for READ but not clearly equated for all baselines. The Friedman ranking in Table 1 is a reasonable attempt to aggregate, but the raw numbers show LiGAN or MolCraft beating READ on some Score-mode metrics, so the Rank-1 placement depends on the weighting scheme, which is fine but should be stated as such.\n\nOverall, this is a serious engineering effort with an internally consistent method and a meaningful ablation. It is not ready as-is because the headline result may not survive a template-leakage control. But the idea is worth testing properly, and the community would benefit from a revision that adds the control, fixes the overclaims, and releases the code.\n\nRecommendation: send to peer review. A serious referee should push on the retrieval-graph leakage question first, then the baseline equivalence and the missing error bars.","headline":"A plausible retrieval+diffusion system whose headline Rank-1 result is undercut by an unexamined template-leakage risk and a couple of overclaims.","tokens_in":14314,"tokens_out":1218,"would_cite":false,"duration_ms":18628,"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":"A retrieval-enhanced aligned diffusion model, READ, claims the top rank in pocket-conditioned ligand generation by steering each reverse-diffusion step with pocket-matched scaffolds.","keywords":["retrieval-augmented generation","diffusion model","molecular generation","structure-based drug design","SE(3) equivariance","latent representation alignment","pocket-conditioned ligand design","CBGBench"],"falsifier":"Rebuild the retrieval graph after deleting every training pocket whose structural similarity to a test pocket exceeds a threshold such as a TM-score of 0.7, and re-run the CBGBench evaluation; if READ's docking and binding-gap advantages largely vanish, template leakage rather than generative generalization explains the results.","tokens_in":13256,"feed_emoji":"🧬","tokens_out":9922,"duration_ms":98482,"temperature":0.7,"pith_summary":"The paper tries to establish that the long-standing trade-off in structure-based drug design — geometric fit to a protein pocket versus chemical validity and synthetic accessibility — can be resolved by coupling a diffusion generator with retrieved chemical priors. It introduces READ, a retrieval-enhanced aligned diffusion model: a contrastively pretrained encoder maps molecules into a drug-like latent space, and at inference a pocket-ligand similarity graph retrieves scaffold embeddings that steer every reverse-diffusion step. On the CBGBench benchmark, READ is reported to achieve the strongest mean docking energy and the largest percentage binding gap among twelve baselines, reaching Rank 1 while producing valid, diverse ligands that often dock better than the native crystal binder. If correct, this would mean retrieval and diffusion can be co-optimized so that real-world chemical knowledge is injected exactly where generative models tend to fail, without post-hoc fixes.","feed_headline":"Retrieval-guided diffusion beats native ligands on 97% of pockets","feed_subtitle":"READ fuses contrastive latent pretraining with pocket-matched scaffold retrieval to rank first on CBGBench.","key_machinery":"The load-bearing mechanism is the coupling of three components: a contrastively pretrained SE(3)-equivariant (rotation- and translation-covariant) encoder that maps molecules into a drug-like latent space, trained on six million MMFF-optimized ZINC molecules with coordinate and masking augmentations; a pocket-ligand bipartite graph whose pocket-pocket edges are weighted by the average of TM-align and DaliLite structural similarity and queried by a coarse-to-fine MSA-then-neighbors routine; and a dual-space diffusion process over atomic coordinates and atom types whose hidden states are aligned layer-wise to the retrieved ligand embeddings through an InfoNCE-style objective. The alignment loss is what makes retrieval effective: the ablation shows that adding the same retrieval signal to an unaligned diffusion backbone degrades the raw Vina score, whereas after alignment it expands the binding gap by roughly an order of magnitude.","core_discovery":"READ is, to the authors' knowledge, the first retrieval-augmented diffusion framework for de novo ligand design. Its central claim is that a contrastively pretrained SE(3)-equivariant encoder defines a drug-like latent manifold in which chemical validity is already encoded, and that aligning the diffusion model's layer-wise hidden states to embeddings retrieved from this manifold lets real-world chemical priors guide each reverse-diffusion step. On the CBGBench benchmark the authors report that the 2000-step variant READ-2k achieves the strongest mean dock energy and the largest percentage binding gap, ranks first among twelve baselines, and improves docking over the native ligand on roughly 97% of pockets with mean Vina reductions near 3 kcal/mol.","pith_inferences":["An implication the paper leaves implicit is that the docking improvements should shrink as the 166,000-ligand retrieval library is made smaller; ablating library size would directly test whether retrieval diversity drives the gain.","The same recipe of latent alignment plus pocket-similarity retrieval should transfer to other 3D generation tasks such as linker design or fragment elaboration, since the guidance term only requires a pretrained encoder and a pocket-ligand graph.","The ablation predicts that applying READ's retrieval guidance to an off-the-shelf diffusion model with an unaligned latent space would hurt, not help, binding scores; this is a direct, checkable consequence of the paper's two-component argument.","Because top-1 selection across 100 generated samples drives the per-pocket improvement probabilities, a learned reranker on top of READ's candidates is a promising way to preserve those gains at a smaller sampling budget."],"forward_implications":["Pocket-conditioned generation can reach top benchmark rank without handcrafted validity filters; contrastive pretraining alone embeds chemical validity into the latent manifold.","Retrieval guidance is only beneficial after the diffusion backbone has been aligned to the pretrained latent space; on an unaligned backbone, the same retrieval signal degrades the raw Vina score.","Combining alignment with retrieval expands the mean percentage binding gap by roughly an order of magnitude relative to either component alone.","READ improves the docking energy over the native reference on roughly 97% of CBGBench pockets, with mean reductions around 3 kcal/mol in Score, Minimize, and Dock modes.","The guidance signal compounds with longer denoising schedules, making the 2000-step READ variant the strongest model in the benchmark."],"supporting_citations":[{"why":"Supplies the CBGBench benchmark, its training/test splits, and the evaluation protocol against twelve baselines.","marker":"[9]"},{"why":"Provides the MMFF force field used to optimize the six-million-molecule pretraining corpus.","marker":"[30]"},{"why":"Defines TM-align structural similarity, one of the two averaged scores weighting pocket-pocket edges in the retrieval graph.","marker":"[32]"},{"why":"Provides DaliLite structural alignment, the second score averaged into pocket-pocket similarity.","marker":"[33]"},{"why":"Supplies the bipartite-graph construction used to index pockets and their cognate ligands in the retrieval space.","marker":"[36]"},{"why":"Supplies the EGNN backbone used for local structural reasoning in the denoiser hidden state.","marker":"[44]"},{"why":"Implements the MSA routine that selects the coarse entry pocket before local neighbor refinement.","marker":"[46]"},{"why":"Defines AutoDock Vina Score, Minimize, and Dock modes that produce all reported docking-energy metrics.","marker":"[47]"},{"why":"Implements PLIP interaction profiling used to compute the Jensen-Shannon divergence and error metrics for interaction fidelity.","marker":"[52]"}],"fun_headline_variants":["Retrieval-augmented diffusion beats native ligands on 97% of pockets","READ: first retrieval-augmented diffusion to top CBGBench","Better docking on 97% of pockets with READ's retrieval-aware diffusion","Retrieval-guided reverse diffusion ranks first on CBGBench","CBGBench: retrieval-augmented diffusion surpasses native ligands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes the training pockets used to build the retrieval graph are not already near-duplicates of the test pockets, because if they were, the reported improvements over native ligands could come from retrieving near-native templates rather than from the model learning to generate.","fun_headline_variants_meta":{"raw":{"variants":["Retrieval-augmented diffusion beats native ligands on 97% of pockets","READ: first retrieval-augmented diffusion to top CBGBench","Better docking on 97% of pockets with READ's retrieval-aware diffusion","Retrieval-guided reverse diffusion ranks first on CBGBench","CBGBench: retrieval-augmented diffusion surpasses native ligands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000915,"raw_usage":{"total_tokens":3889,"prompt_tokens":868,"completion_tokens":3021,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2927}},"tokens_in":484,"tokens_out":3021,"duration_ms":22709,"temperature":1.0,"reasoning_tokens":2927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:36.924474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rebuild the retrieval graph after deleting every training pocket whose structural similarity to a test pocket exceeds a threshold such as a TM-score of 0.7, and re-run the CBGBench evaluation; if READ's docking and binding-gap advantages largely vanish, template leakage rather than generative generalization explains the results.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MMFF force field used to optimize the six-million-molecule pretraining corpus."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines TM-align structural similarity, one of the two averaged scores weighting pocket-pocket edges in the retrieval graph."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides DaliLite structural alignment, the second score averaged into pocket-pocket similarity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EGNN backbone used for local structural reasoning in the denoiser hidden state."},{"cited_title":"(2015).msa: An R package for multiple sequence alignment","cited_arxiv_id":null,"evidence_quote":"Implements the MSA routine that selects the coarse entry pocket before local neighbor refinement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines AutoDock Vina Score, Minimize, and Dock modes that produce all reported docking-energy metrics."},{"cited_title":"J., Adasme, M","cited_arxiv_id":null,"evidence_quote":"Implements PLIP interaction profiling used to compute the Jensen-Shannon divergence and error metrics for interaction fidelity."}],"review_version":1}