{"id":"4caf1c94-ed7c-430d-b733-09c60d6c7b40","arxiv_id":"2412.17780","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PepTune introduces Monte Carlo Tree Guidance for masked discrete diffusion, generating peptide SMILES simultaneously optimized for binding, permeability, solubility, hemolysis, and non-fouling.","lead":"PepTune is a new AI system that designs therapeutic peptide molecules by combining masked diffusion with a search algorithm that optimizes several drug properties at once. It generates chemically modified and cyclic peptides and claims binders with better in silico docking scores than existing GLP-1R drugs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed multi-objective optimization is evaluated with the same classifiers used as rewards; the only external check, docking, is applied to a hand-picked subset with no random or matched-length baseline, so the central claim rests on unvalidated proxy monotonicity.","rationale":"The central claim is that PepTune simultaneously optimizes several therapeutic properties. This requires the reward classifiers to be informative proxies. The paper gives no independent assay or external benchmark for permeability, solubility, hemolysis, or non-fouling; all reported gains come from the same functions used in Equation 14. For binding, Vina docking is independent, but the selection procedure (top predicted scores, non-dominated sequences) and the absence of any random or length-matched control make the reported docking superiority uninterpretable. I therefore align with the reader's weakest assumption, with the added observation that the docking check itself is not a clean external test. I found no more load-bearing internal inconsistency; the invalid-loss softmax notation issue is real but secondary. The CONDITIONAL verdict should stand: accept only after a blinded external validation of at least one property, or a docking comparison on the full generated set with baselines.","tokens_in":45167,"tokens_out":7596,"duration_ms":80218,"concrete_test":"Run a blinded docking validation for the GLP-1R objective: from a single PepTune run, dock all 100 generated peptides with AutoDock Vina under the exact protocol of Appendix F.3; in parallel dock 100 PepMDLM unconditional outputs, 100 random training-set peptide SMILES matched by length, and semaglutide/liraglutide. Compare the full docking-score distributions and compute the Spearman correlation between the binding-classifier score and Vina score within the PepTune set. If the PepTune distribution is not significantly better than the unconditional/random baselines, or if the classifier score is not positively correlated with Vina score, the external-validation claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the reward signal in Equation 14, computed from the trained property classifiers, is monotonically related to the true biological properties. All multi-objective gains are reported with those same classifiers (Section 5, Appendix A), so the improvement is partly self-fulfilling: maximizing s_k will raise s_k even if s_k is not a faithful proxy. The binding regressor, the only objective with an external check, has held-out Spearman 0.633 on 1,806 PepLand sequences (Section 4). The docking validation does not break the circularity: Appendix F.3 docks only non-dominated valid sequences, and Section 5 selects the peptides with the highest predicted binding-affinity scores before docking. No random sample of PepMDLM outputs or of training-set peptides is docked under the same protocol, and the clinical comparators are not length-matched. Therefore the GLP-1R comparison (-7.4/-7.0 vs -5.7/-5.1 kcal/mol) and the TfR/GLAST tables are uninterpretable as evidence that the optimized peptides bind better than existing drugs or than unconditionally generated peptides.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PepTune, a discrete masked diffusion model for de novo peptide SMILES generation, together with Monte Carlo Tree Guidance (MCTG), an inference-time algorithm for multi-objective optimization. The unconditional generator, PepMDLM, uses a bond-dependent masking schedule and an invalid-peptide loss to improve structural validity. The MCTG procedure builds a search tree over partially unmasked sequences, rolls out candidates, scores them with trained property classifiers for binding affinity, permeability, solubility, hemolysis, and non-fouling, and returns a Pareto front of non-dominated sequences. The authors report that PepTune generates chemically modified and cyclic peptides, reaches 100% validity after 20 MCTS iterations, and produces in silico docking scores for GLP-1R, TfR, GLAST, GFAP, NCAM1, AMHR2, and RBX1, including GLP-1R scores claimed to be better than those of semaglutide and liraglutide.","tokens_in":45365,"tokens_out":5515,"duration_ms":55684,"significance":"If the empirical claims hold, MCTG would be a useful and modular contribution to guided discrete diffusion, particularly because it avoids gradient estimation in discrete spaces and can combine arbitrary reward signals. The manuscript ships several strengths: the theoretical derivations of the bond-dependent reverse posterior and NELBO in Appendix G are presented in detail, the validity-filter tool is released, and the code is publicly available. The main limitation is that the multi-objective improvements are measured with the same classifiers that define the MCTG reward, and the only external check, docking, is applied to a hand-picked subset without matched-length or random-sequence baselines. The central claim of simultaneous optimization of true therapeutic properties is therefore not yet established, although the methodological framework itself is plausible and worth further validation.","major_comments":[{"comment":"The central claim that PepTune improves multi-objective therapeutic properties is evaluated with the same trained classifiers that define the MCTG reward. Since Eq. (14) computes rewards from these classifiers and the Pareto front is updated from the same scores, increases in solubility, hemolysis, non-fouling, and predicted binding affinity over iterations are partly by construction. The only external validation, Autodock Vina docking, is applied in Section 5 and Appendix F.3 only to non-dominated sequences with the highest predicted binding-affinity scores; no random sample of PepMDLM outputs or training-set peptides is docked under the same protocol, and the clinical comparators are not length-matched. The GLP-1R comparison (-7.4 and -7.0 kcal/mol for PepTune versus -5.7 and -5.1 kcal/mol for semaglutide and liraglutide) is therefore uninterpretable as evidence that optimized peptides bind better than unconditionally generated or known peptides. This should be fixed by docking random PepMDLM samples and training peptides of matched length under the identical protocol and reporting the full distribution of scores.","section":"Section 5, Eq. (14), Appendix F.3"},{"comment":"There is a load-bearing notational inconsistency in the invalid-loss derivation. Section 2.1 defines x_theta(zt,t) as a vector of token probabilities in the simplex Delta^|V|, but Eq. (6) applies a softmax to x_theta, and Eq. (7) is derived as the derivative of softmax(x_theta,k) with respect to x_theta,k. If x_theta is already a probability vector, applying softmax is not the correct transformation and Eq. (7) is not the gradient of Eq. (6). If x_theta instead denotes logits, then the simplex notation and the use of x_theta as probabilities in Eq. (4) are inconsistent. The authors should clarify this distinction and restate the invalid loss and its gradient accordingly, since the claimed novel gradient mechanism depends on this step.","section":"Section 2.3, Eq. (6)-(7), Proposition 2.3"},{"comment":"The docking validation is not controlled for ligand length or sequence background. Vina scores are not size-normalized, and the GLP-1R comparators semaglutide and liraglutide are over 30 residues while the PepTune binders are shorter; the TfR/GLAST table reports only selected binders without docking scores for PepMDLM baselines or for random peptides from the training set. Without matched-length baseline docking, the absolute score comparisons in Table 3 and Figure 3 do not support the conclusion that the generated peptides bind better than existing drugs or than unconditionally generated peptides. A matched-length random baseline is needed before these numbers can be cited as evidence of improved binding.","section":"Appendix F.3, Table 3, Figure 3"}],"minor_comments":[{"comment":"The appendix overview states that the proofs for Proposition 2.2 appear in G.2 and for Proposition 2.1 in G.3, but the actual sections are reversed relative to the main-text numbering; please correct the mapping.","section":"Appendix overview, G.2-G.3"},{"comment":"The abstract says PepTune 'ensures valid peptide structures', but Table 1 reports PepMDLM validity of 0.450 and PepTune validity of 1.000 only after MCTS; the unconditional model does not by itself ensure validity, so the wording should be qualified.","section":"Abstract, Table 1"},{"comment":"For non-fouling, the reported accuracy of PeptideBERT (0.873) is higher than the reported accuracy of the proposed model (0.766); the text should acknowledge that the improvement is on F1, not on accuracy.","section":"Table 2"},{"comment":"Equation (7) contains a typographical period in the subscript SM(x^(ell)_theta.k); it should be a comma: SM(x^(ell)_theta,k).","section":"Eq. (7)"},{"comment":"In Algorithm 6, the dominance flag d is defined using the weak inequality s_k(x_s) >= s*_k, but the definition of dominance in Eq. (10) requires strict improvement in at least one objective; the flag should use strict inequality to match the formal definition.","section":"Algorithm 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to interest the machine-learning-for-biology community, and the MCTG framework is a reasonable contribution if the evaluation is strengthened. My main concern is not the method's internal consistency but the gap between the claims and the evidence: the property-classifier circularity and the uncontrolled docking comparisons are fixable with additional experiments, but they are central to the paper's stated conclusions. I would be willing to reconsider after the authors add matched-length random baselines, dock a random sample of unconditional outputs, and clarify the invalid-loss notation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, the method is genuinely novel: a bond-dependent masking schedule for peptide SMILES, a global invalid-loss that pushes the generator away from non-peptide strings, and Monte Carlo Tree Guidance for inference-time multi-objective discrete diffusion. The math in Appendix G for the bond-dependent NELBO and reverse posterior checks out; this is not hand-waving. Second, the biological headline—GLP-1R binders that out-dock semaglutide and liraglutide—is weaker than it looks, because the same classifiers that provide the MCTG reward also select which few peptides get docked. That is a circularity problem, not a fabricated result, and it is fixable.\n\nWhat it does well: the ablation (Table 13) shows both the masking schedule and invalid loss matter (validity 0.40 vs 0.16/0.21 without them). MCTG is a sensible answer to multi-objective guidance in discrete spaces, avoiding the gradient-estimation issues of prior predictor-guidance methods. The training-data curation—11M modified and cyclic peptide SMILES—is a contribution in itself, and code is released.\n\nWhere it is soft, in proportion: the classifier-reward circularity is the main issue. All property-score improvements over iterations (Figures 5-10) are measured with the same XGBoost/Transformer models used as rewards in Equation (14), so those curves are partly self-fulfilling. The docking validation breaks the loop only partially: it is an external tool, but peptides are selected for high predicted affinity, and no random sample of PepMDLM outputs or training-set peptides is docked under the same protocol. The GLP-1R comparison is also not length-matched (semaglutide/liraglutide are ~30 residues; the generated peptides appear shorter), which biases Vina scores. Minor but real: Proposition 2.3 differentiates softmax as if x_theta were logits, while Equation (6) calls it a probability vector; softmax of a probability is not standard. That needs a corrected derivation, though it does not undermine the main NELBO result.\n\nWho profits: anyone working on discrete diffusion for molecular or peptide design; the method section is worth the read. The docking claims should be treated as in silico hypotheses, not validated binders. I would send it to review; a competent referee will ask for random-baseline docking, matched-length comparators, and a cleaned-up invalid-loss derivation. Conditional accept territory.","headline":"Novel MCTS-guided discrete diffusion for peptide SMILES, with a real circularity problem in the biological validation; worth reviewing, not ready as claimed.","tokens_in":45949,"tokens_out":3423,"would_cite":true,"duration_ms":33743,"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":"A discrete diffusion generator with Monte Carlo tree guidance designs chemically modified peptides while optimizing several therapeutic properties at once.","keywords":["peptide design","discrete diffusion","Monte Carlo tree search","multi-objective optimization","peptide SMILES","non-natural amino acids","binding affinity prediction","therapeutic peptides"],"falsifier":"Generate a separate set of peptides with the binding-affinity regressor's reward replaced by a random or reversed scorer and compare docking distributions: if the optimized set still docks as well or better than the unconditional set, the apparent affinity gains come from the generator's prior rather than from the guidance. A cheaper check is to compute the rank correlation between the regressor's predictions and AutoDock Vina scores on the paper's own generated binders; a near-zero or negative correlation would mean the two validation layers disagree.","tokens_in":44909,"feed_emoji":"🧬","tokens_out":9546,"duration_ms":85560,"temperature":0.7,"pith_summary":"PepTune aims to show that therapeutic peptides containing non-natural amino acids and cyclic modifications can be generated from scratch while simultaneously optimizing several drug-relevant properties, not just binding to one target. The paper's central claim is that a masked discrete diffusion model trained on peptide SMILES, aided by a bond-dependent masking schedule and an invalid-sequence loss, can generate chemically modified peptides, and that an inference-time Monte Carlo Tree Guidance (MCTG) algorithm can steer generation toward Pareto-optimal trade-offs across predicted binding affinity, membrane permeability, solubility, hemolysis, and non-fouling. If correct, this would allow researchers to design short, modified peptide leads for targets with no known binders, including disordered or intracellular proteins, without requiring a 3D target structure or retraining the generator for each new objective set. The evidence is computational: two GLP-1R candidates receive docking scores of -7.4 and -7.0 kcal/mol, compared with -5.7 for semaglutide and -5.1 for liraglutide, and seven dual TfR/GLAST binders all dock at or below -7.5 kcal/mol.","feed_headline":"One diffusion model designs peptides for five drug properties at once","feed_subtitle":"Tree-guided diffusion picks the trade-off set, with no retraining for new targets or objectives.","key_machinery":"PepTune's load-bearing mechanism is Monte Carlo Tree Guidance (MCTG), an inference-time algorithm that reformulates classifier-guided discrete diffusion as a search over unmasking steps. Starting from a fully masked sequence, each MCTS iteration selects a child node via a Pareto selection score that balances normalized reward vectors with the unconditional reverse-posterior probability and an exploration term, expands the node into $M = 50$ distinct Gumbel-perturbed partial sequences, rolls each out greedily to a clean peptide SMILES, scores it with peptide-SMILES property predictors for binding affinity, permeability, solubility, hemolysis, and non-fouling, updates the Pareto-optimal set, and back-propagates reward vectors to ancestors. Two supporting components make the search space tractable: a bond-dependent masking schedule in which peptide-bond tokens survive longer under $\\alpha_t = 1 - t^w$ with $w = 3$, so the model learns backbone structure first, and an invalid peptide loss that propagates penalties from argmax-invalid SMILES back through softmax probabilities. The property predictors themselves are part of the contribution, including a cross-attention binding-affinity regressor and XGBoost classifiers for the other properties, trained specifically on peptide SMILES data.","core_discovery":"On the paper's own terms, the discovery is that multi-objective guidance for discrete diffusion can be carried out entirely in the discrete token space by treating partial unmaskings as nodes of a search tree. MCTG rolls out each candidate unmasking path to a complete peptide SMILES, scores it with trained classifiers and regressors for the desired properties, retains only Pareto-non-dominated sequences, and back-propagates reward vectors up the tree so subsequent iterations favor unmasking steps that lead to high-scoring, valid peptides. Paired with a bond-dependent masking schedule that masks peptide-bond tokens more slowly, using $\\alpha_t = 1 - t^w$ with $w = 3$, and a softmax-scaled invalid loss, the generator reaches 100% valid peptide SMILES after about 20 MCTS iterations while maintaining diversity comparable to the unconditional model. The paper reports two PepTune GLP-1R binders with docking scores of -7.4 and -7.0 kcal/mol, compared with -5.7 for semaglutide and -5.1 for liraglutide, and seven dual TfR/GLAST binders with docking scores at or below -7.5 kcal/mol, the best reaching -10.5 kcal/mol for TfR and -9.2 kcal/mol for GLAST.","pith_inferences":["An implication left implicit by the paper is that MCTG is a general recipe for any discrete sequence domain, including DNA, RNA, or small-molecule SMILES, wherever a scorer exists and data is too sparse for classifier-free training; the machinery itself does not depend on peptide chemistry.","The strongest caveat is that every optimized property is measured by a learned proxy, and the binding-affinity regressor was trained on only 1,806 PepLand sequences with a held-out Spearman correlation of 0.633; if those proxies are biased, the Pareto front may be optimal for the predictors rather than for biology.","A natural testable extension would be to synthesize a small set of top-scoring binders and measure affinity, permeability, and hemolysis in vitro, then compare whether the ordering predicted by the classifiers matches the experimental ordering.","Because MCTG returns the full Pareto front in one run, downstream users could apply a secondary filter, such as synthesis cost, route availability, or a patent landscape, to choose from the non-dominated set without rerunning generation."],"forward_implications":["If PepTune works as claimed, researchers can generate chemically modified, cyclic peptide candidates for targets with no existing binders or no stable 3D structure, since guidance needs only the target sequence and trained property predictors.","A single trained unconditional diffusion model can be steered toward new objectives or new target sequences at inference time, because MCTG requires no retraining of the generator and no gradient estimates in the discrete space.","Optimizing several therapeutic properties together should reduce the binding-versus-developability trade-off; the paper's TfR/GLAST and GFAP/RBX1 cases show predicted property scores improving together over MCTS iterations.","Shorter GLP-1R peptide leads with in silico docking scores better than approved agonists could point toward cheaper, lower-immunogenicity alternatives, if the docking and predicted-affinity signals hold up in experiments.","The time-dependent guidance variant described in the appendix suggests that practitioners can prioritize one property first and layer in additional objectives later, which may help when some properties are much harder to optimize than others."],"supporting_citations":[{"why":"Supplies the Masked Diffusion Language Model training and sampling framework that PepMDLM extends.","marker":"Sahoo et al., 2024"},{"why":"Provides the state-dependent masking generalization from which the bond-dependent schedule is derived.","marker":"Shi et al., 2024"},{"why":"Defines Monte-Carlo Tree Search, whose selection-expansion-rollout-backpropagation loop MCTG adapts to unmasking steps.","marker":"Coulom, 2007"},{"why":"Supplies the PepLand binding-affinity dataset of 1,806 sequences used to train the cross-attention affinity regressor.","marker":"Zhang et al., 2023"},{"why":"Provides PeptideCLM embeddings and the SMILES pair-encoding vocabulary used to tokenize peptide SMILES and feed the property predictors.","marker":"Feller & Wilke, 2024"},{"why":"CycloPs generates the roughly 10-million-sequence chemically modified and cyclic peptide training set.","marker":"Duffy et al., 2011"},{"why":"AutoDock Vina computes the in silico docking scores used to validate the GLP-1R, TfR, and GLAST binders.","marker":"Eberhardt et al., 2021"},{"why":"PeptideBERT is the baseline whose solubility, hemolysis, and non-fouling prediction performance PepTune's classifiers are benchmarked against.","marker":"Guntuboina et al., 2023"}],"fun_headline_variants":["Tree-guided diffusion designs peptides for five drug traits at once","Discrete diffusion tunes peptides for multiple drug properties in one go","Search-tree steering guides peptide diffusion to Pareto-optimal sets","Tree-guided diffusion hits five drug properties without retraining","One diffusion model, many drug traits: tree-guided trade-offs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the trained property predictors, especially the binding-affinity regressor with a held-out Spearman correlation of only 0.633 on 1,806 training sequences, rank true therapeutic quality well enough that reward-guided unmasking toward higher predicted scores also moves real binding, permeability, solubility, hemolysis, and fouling behavior in the right direction.","fun_headline_variants_meta":{"raw":{"variants":["Tree-guided diffusion designs peptides for five drug traits at once","Discrete diffusion tunes peptides for multiple drug properties in one go","Search-tree steering guides peptide diffusion to Pareto-optimal sets","Tree-guided diffusion hits five drug properties without retraining","One diffusion model, many drug traits: tree-guided trade-offs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2628,"prompt_tokens":978,"completion_tokens":1650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":1568}},"tokens_in":594,"tokens_out":1650,"duration_ms":11993,"temperature":1.0,"reasoning_tokens":1568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:08.261003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a separate set of peptides with the binding-affinity regressor's reward replaced by a random or reversed scorer and compare docking distributions: if the optimized set still docks as well or better than the unconditional set, the apparent affinity gains come from the generator's prior rather than from the guidance. A cheaper check is to compute the rank correlation between the regressor's predictions and AutoDock Vina scores on the paper's own generated binders; a near-zero or negative correlation would mean the two validation layers disagree.","supporting_citations":[],"review_version":1}