{"id":"6a7dcf2a-f3b0-47fd-970f-c250f10f74b8","arxiv_id":"2605.27413","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ProtLiD² is a ligand-conditioned discrete diffusion model for joint protein sequence-structure generation that reports higher TM-score, pLDDT, and ligand-aware docking pass rates than prior methods on whole-protein and pocket design tasks.","lead":"The paper introduces ProtLiD², a discrete diffusion model that jointly generates protein amino acid sequences and structure tokens while conditioning on ligand chemical and geometric features via cross-attention. A smart generalist might read it because ligand-aware protein co-design could accelerate the creation of proteins that bind specific small molecules for drug development or industrial enzymes.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly flags the standard data-shift risk for any large-scale conditional generative model. Because the supplied text contains no contradictory internal evidence and the claim is scoped to reported metrics on (presumably) held-out data, no additional load-bearing technical objection is warranted beyond the already-noted generalization caveat.","tokens_in":1944,"tokens_out":274,"duration_ms":18199,"concrete_test":"Re-run the pocket co-design benchmark after filtering the test ligands to Tanimoto similarity <0.4 with the training set; if the pass-rate gap versus PocketGen shrinks below 2×, the headline generalization claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract presents a coherent engineering claim: a ligand-conditioned discrete diffusion model trained on >1M complexes yields measurable gains in fold confidence (TM-score 0.672→0.802) and pocket metrics (BB-RMSD 3.4Å→1.97Å, pass-rate 14.86%→59.73%). No internal contradiction, missing identity, or parameter-free derivation failure is visible from the supplied text. The reader's weakest assumption (training-set diversity) is a standard generalization caveat rather than an identifiable flaw in the argument as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces ProtLiD², a ligand-conditioned discrete diffusion model for joint protein sequence and structure co-design. It generates amino-acid sequences and discrete structure tokens while incorporating ligand information via geometry-aware cross-attention, trained on over one million complexes. The work proposes maximum confidence-margin guided ReMask decoding for inference. Reported results include improved whole-protein design metrics over Complexa (TM-score 0.672→0.802, pLDDT 64.55→73.00) and pocket co-design gains over FAIR/PocketGen (active-site BB-RMSD reduced to 1.97Å; ligand-aware pass rates increased to 59.73% and 23.49% under varying docking thresholds).","tokens_in":2063,"tokens_out":451,"duration_ms":19303,"significance":"If the empirical gains are robust, the work demonstrates that discrete diffusion in token space can support effective ligand-aware protein co-design, providing a scalable alternative to continuous diffusion or flow-based methods. The scale of training data and the inference-time correction mechanism represent practical strengths for applications in functional protein engineering and ligand-binding design.","major_comments":[{"comment":"Abstract: the central empirical claims of metric improvements (TM-score 0.672 to 0.802, BB-RMSD to 1.97Å, pass-rate gains from 14.86% to 59.73%) are reported without error bars, statistical significance tests, details on data splits, hyperparameter search procedures, or evaluation threshold selection criteria, preventing full verification of the performance claims.","section":"Abstract"},{"comment":"Results and Methods sections: the manuscript provides no quantitative assessment of training-set diversity, potential memorization, or distribution shift for the >1M complexes, which is load-bearing for the generalization claims to novel ligands and targets.","section":"Results/Methods"}],"minor_comments":[{"comment":"Abstract: the model name expansion (ProtLiD²) and the description of ReMask decoding could be expanded slightly for immediate clarity to readers unfamiliar with discrete diffusion variants.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for these constructive comments on the empirical reporting and generalization analysis. We address each point below and will incorporate revisions to strengthen the manuscript.","responses":[{"response":"We agree that the abstract alone does not provide sufficient context for verification. In the revised manuscript we will (i) report error bars and statistical significance (e.g., paired t-tests across independent runs) for all headline metrics in the Results section, (ii) expand the Methods section with explicit descriptions of data splits, hyperparameter search ranges and selection criteria, and (iii) clarify how docking-score thresholds were chosen. The abstract will be updated to reference these supporting details or to present the improvements more conservatively.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central empirical claims of metric improvements (TM-score 0.672 to 0.802, BB-RMSD to 1.97Å, pass-rate gains from 14.86% to 59.73%) are reported without error bars, statistical significance tests, details on data splits, hyperparameter search procedures, or evaluation threshold selection criteria, preventing full verification of the performance claims."},{"response":"We acknowledge this gap. The current manuscript does not contain quantitative checks for training-set diversity, memorization, or distribution shift. In revision we will add a dedicated subsection reporting (a) sequence-identity and TM-score distributions within the training set, (b) nearest-neighbor similarity analysis between generated outputs and training examples on held-out ligands, and (c) a brief discussion of potential distribution shift between training and evaluation ligands. These additions will directly support the generalization claims.","revision_made":"yes","referee_comment":"[Results/Methods] Results and Methods sections: the manuscript provides no quantitative assessment of training-set diversity, potential memorization, or distribution shift for the >1M complexes, which is load-bearing for the generalization claims to novel ligands and targets."}],"tokens_in":1539,"tokens_out":428,"duration_ms":20312,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this paper puts explicit ligand conditioning into a masked discrete diffusion model for protein sequence and structure co-design. It uses geometry-aware cross-attention to bring in the small molecule's chemical and spatial features, adds a maximum-confidence-margin ReMask step at inference, and trains on over a million complexes. The abstract shows better numbers than Complexa, FAIR, and PocketGen on TM-score, pLDDT, active-site BB-RMSD, and ligand-aware docking pass rates.\n\nWhat is actually new is the specific combination of discrete token diffusion with direct ligand conditioning and the ReMask decoding trick. Earlier discrete diffusion protein models lacked the ligand part, so this is a straightforward engineering extension that keeps everything in token space. The paper does well by demonstrating that the conditioning produces measurable lifts in both global fold quality and pocket accuracy, and the scale of the training data is substantial.\n\nThe soft spots are in the experimental reporting. The abstract gives no error bars, no statistical tests, and no information on data splits, hyperparameter search, or how evaluation thresholds were set. Without those, the size of the reported gains is hard to judge for robustness. The assumption that the million-complex training set is diverse enough to generalize to new ligands is also left as a standard caveat rather than something tested with clear out-of-distribution cases. These issues are real but not load-bearing if the full methods section fills them in.\n\nThis is for people working on conditional generative models for proteins in drug design or enzyme engineering. Readers who follow discrete diffusion or ligand-aware design will find the architecture and decoding details worth seeing. The work shows clear thinking on how to integrate the ligand signal and deserves a serious referee so the methods and numbers can be checked properly.","headline":"ProtLiD² adds ligand geometry-aware cross-attention to masked discrete diffusion for joint seq-structure co-design and reports benchmark gains, but the abstract leaves the empirical claims under-specified.","tokens_in":2604,"tokens_out":437,"would_cite":false,"duration_ms":26558,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Ligand-conditioned discrete diffusion jointly designs protein sequences and structures that bind given small molecules more accurately than prior token models.","keywords":["protein design","discrete diffusion","ligand conditioning","sequence-structure co-design","protein-ligand complexes","pocket design","diffusion models"],"falsifier":"Evaluation on a held-out test set of ligands and protein targets whose binding pockets share no structural or sequence similarity with any training example, showing no improvement over unconditioned baselines on TM-score or docking pass rate.","tokens_in":2836,"feed_emoji":"🧬","tokens_out":683,"duration_ms":17648,"temperature":0.7,"pith_summary":"The paper establishes that conditioning a discrete diffusion process on ligand geometry and chemistry allows simultaneous generation of amino-acid sequences and structure tokens that form compatible, functional binding sites. A reader would care because this token-space approach avoids the need to separately optimize sequence and coordinates while still respecting explicit ligand constraints. The method extends masked diffusion with cross-attention that injects ligand features and adds an inference-time ReMask step that iteratively corrects low-confidence tokens. Results show measurable gains on both global fold metrics for full proteins and local active-site accuracy plus docking success for pockets.","feed_headline":"Ligand diffusion raises protein TM-score from 0.67 to 0.80","feed_subtitle":"Token-space model conditions discrete diffusion on small-molecule geometry, cutting pocket RMSD and lifting docking success rates.","key_machinery":"Geometry-aware cross-attention inside a masked discrete diffusion process over sequence and structure tokens, plus maximum confidence-margin guided ReMask decoding.","core_discovery":"ProtLiD² jointly generates amino-acid sequence and discrete structure tokens while incorporating ligand chemical and geometric information through geometry-aware cross-attention. Trained on over one million ligand-protein complexes, it extends masked discrete diffusion to ligand-aware functional protein design and uses maximum confidence-margin guided ReMask decoding at inference. This yields higher TM-score and pLDDT than Complexa on whole-protein design and lower active-site backbone RMSD plus higher ligand-aware pass rates than FAIR and PocketGen on pocket co-design.","pith_inferences":["The approach could be tested on designing proteins that bind entirely new ligand scaffolds never seen in the training distribution.","ReMask decoding might transfer to other discrete diffusion tasks such as antibody or enzyme active-site redesign.","If the model truly captures ligand geometry in token space, it could reduce reliance on post-hoc docking validation during design campaigns."],"forward_implications":["Whole-protein designs achieve higher global fold confidence (TM-score rising from 0.672 to 0.802).","Pocket designs produce active-site backbones with lower RMSD (1.97 Å versus 3.4 Å).","Ligand-aware docking success rates increase substantially under both standard and stricter thresholds.","The same token-space framework supports both full-length and local pocket co-design without separate coordinate optimization."],"fun_headline_variants":["ProtLiD2 reaches 0.80 TM-score via ligand diffusion","Ligand diffusion cuts pocket RMSD to 1.97A","Discrete diffusion conditions on ligand geometry","ProtLiD2 lifts docking pass rate to 59.73 percent"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The training set of over one million ligand-protein complexes is assumed to be sufficiently diverse and representative that the learned conditional distribution generalizes to novel ligands and protein targets without significant distribution shift or memorization.","fun_headline_variants_meta":{"raw":{"variants":["ProtLiD2 reaches 0.80 TM-score via ligand diffusion","Ligand diffusion cuts pocket RMSD to 1.97A","Discrete diffusion conditions on ligand geometry","ProtLiD2 lifts docking pass rate to 59.73 percent"]},"model":"grok-4.3","cost_usd":0.006091,"raw_usage":{"total_tokens":2953,"prompt_tokens":818,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":60912000,"prompt_tokens_details":{"text_tokens":818,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2067,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":818,"tokens_out":68,"duration_ms":16319,"temperature":1.0,"reasoning_tokens":2067,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T19:45:35.989665+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Evaluation on a held-out test set of ligands and protein targets whose binding pockets share no structural or sequence similarity with any training example, showing no improvement over unconditioned baselines on TM-score or docking pass rate.","supporting_citations":[],"review_version":1}