{"id":"ceb7ea9e-eb42-4d91-a995-f7f2b7f7acfc","arxiv_id":"2505.07086","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"MOG-DFM uses rank-directional scoring and an adaptive hypercone filter to guide discrete flow matching toward sequences with balanced multi-objective improvements.","lead":"This paper introduces MOG-DFM, a method that steers pretrained discrete flow matching models to generate biological sequences that balance several desired properties at once. It reports improved peptide binders and enhancer DNA sequences on computational benchmarks, though most property scores are produced by the same predictors used for guidance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Appendix D guarantee is definitional and does not match the argmax sampling rule, so Pareto-front convergence is unsupported.","rationale":"The reader's verdict is conditional and highlights predictor reliability. That concern is real, but I think the more load-bearing issue is that the paper's only formal argument for Pareto-efficient guidance is essentially definitional and does not apply to the sampling rule actually used. The positive-expected-improvement claim follows from admitting only transitions with positive dot product with omega; it is the same property any scalarizing filter would have. The mismatch between the softmax measure in Appendix D and the argmax selection in Section 3.4 means the theorem, even if accepted, is about a different algorithm. The Phi > pi/2 case also invalidates the proof's claim that all cone members have positive directional improvement. These are internal correctness risks, not disagreements with consensus, and they sit upstream of the predictor-bias issue: even with perfect predictors, the paper has not shown convergence to or near the Pareto front. I do not recommend changing the conditional verdict because the empirical results may still support a practical engineering claim, and the reader already required stronger formal support and independent validation. The proposed synthetic check would settle whether the algorithm actually achieves Pareto-front behavior and whether the formal proof's stochastic sampler is essential.","tokens_in":20855,"tokens_out":6055,"duration_ms":68611,"concrete_test":"Run MOG-DFM on a small synthetic discrete instance with known Pareto front, e.g., d=4, K=16, N=3, random additive objective functions so the full 65,536-state space is enumerable. Compute the fraction of MOG-DFM outputs that are nondominated in the true set and the hypervolume of outputs, and compare three controls: (a) greedy scalarized coordinate ascent with random restarts using the same omega sampling, (b) the same hypercone filter but with a uniform token-proposal base model replacing the pretrained DFM, and (c) the stochastic softmax sampler described in Appendix D instead of the argmax rule in Algorithm 1. If MOG-DFM is not closer to the true Pareto front than controls (a) or (b), the claimed Pareto-guidance advantage is not supported. If (c) differs materially from the argmax version, the proof does not cover the deployed algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's only formal support for the central Pareto claim is Appendix D. That argument shows that, if the sampler only accepts transitions whose improvement vector makes an acute angle with omega, the expected improvement in direction omega is positive. This is true by construction for any greedy scalarized coordinate ascent; it does not establish progress toward the Pareto front, let alone Pareto efficiency, because moving along one trade-off direction can degrade other objectives and no monotonic or convergence statement is proved. Two additional internal problems make the proof inapplicable to the submitted algorithm: (i) the proof defines mu_i as a softmax over accepted candidates (exp(Delta S)/Z) and takes expectation under that distribution, whereas Section 3.4 and Algorithm 1 select y_best = argmax over accepted candidates and only use the guided rate to decide whether to move at all (Eq. 23); (ii) the proof assumes Phi in (0,pi), but for Phi > pi/2 the cone can contain transitions with negative dot product, so the key inequality D>0 fails unless Phi is restricted to an acute cone (as in the experiments, Phi_max=75 deg). Thus the central formal claim is not established even before considering whether the property predictors are accurate. The empirical Pareto-efficiency claims are also evaluated with the same score functions used for guidance, so they inherit any predictor bias.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MOG-DFM, a guidance framework for pretrained discrete flow matching models that aims to generate biological sequences satisfying multiple, possibly conflicting objectives. At each sampling step, MOG-DFM scores candidate token replacements with a hybrid rank-directional score, reweights the base model's transition rates, applies an adaptive hypercone filter to restrict moves to those roughly aligned with a user-selected trade-off vector, and then evolves the CTMC by Euler sampling. The authors train two unconditional base models, PepDFM for peptides and EnhancerDFM for enhancer DNA, and evaluate MOG-DFM on peptide binder design with five objectives (hemolysis, non-fouling, solubility, half-life, binding affinity) and on enhancer DNA generation guided by class and shape. They also compare against classical MOO baselines and report ablations. Appendix D claims a formal guarantee that MOG-DFM induces positive expected improvement in the direction of the weight vector.","tokens_in":1761,"tokens_out":1494,"duration_ms":61949,"significance":"If the empirical results were independently confirmed, MOG-DFM would be a useful contribution to controllable discrete sequence generation. The paper includes trained base models, a planned public codebase, ablations isolating the hypercone mechanism, and some external checks with ADMET-AI and AlphaFold3/VINA. However, the formal support is definitional rather than a genuine convergence guarantee, and the main evaluations reuse the same predictors that provide guidance, so the strength of the Pareto-efficiency claims is not currently supported. The framework itself is plausible and the empirical trends are internally consistent, but the manuscript needs substantial revision of both the theoretical claim and the evaluation design before the central claims can be accepted.","major_comments":[{"comment":"The formal guarantee is not established for the implemented algorithm. The proof takes expectation under the softmax distribution mu_i over feasible transitions, but Section 3.4 and Eq. (23) select y_best = argmax over accepted candidates and use the guided rate only to decide whether to move; these are different stochastic processes. In addition, the key inequality D(y_i,x;omega)>0 is claimed to follow from alpha_i <= Phi with Phi in (0,pi), but arccos of a dot product is nonnegative only when Phi <= pi/2; for Phi > pi/2 the cone contains vectors with negative dot product. The fallback rule in Eq. (18) selects among candidates with alpha_i < pi, which can include transitions with D<0, so the proof's assertion that the fallback chooses D>0 'by construction' is false. At best, the argument shows positive expected scalarized improvement for a softmax sampler restricted to an acute cone, which is a direct consequence of the filter rather than an independent convergence proof. It does not establish monotonic progress toward the Pareto front, since moves with positive dot product along one weight vector can degrade individual objectives.","section":"Appendix D and Algorithm 1"},{"comment":"The headline improvements are evaluated with the same predictors used for guidance. The hemolysis, non-fouling, solubility, half-life, and affinity scores in Tables 2, 3, 7, 8, 9, and 10 are produced by the models described in Section B, which are exactly the score functions used in Eq. (14) during sampling. The external ADMET-AI and AlphaFold3/VINA checks cover only solubility, half-life, and binding on selected structures; non-fouling and hemolysis have no independent check, and the binding check is limited to a small number of examples. The reported Pareto gains may therefore reflect predictor artifacts rather than genuine molecular improvements, and the claim that MOG-DFM 'consistently produces the most favorable trade-offs' is not supported by the experiments as designed.","section":"Section 4.3 and Section B"},{"comment":"Most quantitative comparisons lack error bars, standard deviations, or significance tests. In particular, Table 4 evaluates each DNA guidance setting on only five sequences, and Table 10 varies one hyperparameter at a time on different targets, so cross-setting comparisons are confounded. Without variance estimates, it is impossible to judge whether the large half-life differences in Table 3 (for example, 27.227 h vs 4.449 h for target 1B8Q) or the ablation differences in Table 7 are meaningful or within sampling noise. The paper should report distributions or confidence intervals and use a consistent evaluation setup when comparing settings.","section":"Tables 3, 4, 7, 10"}],"minor_comments":[{"comment":"The column layout is confusing: the 'Shape' label appears multiple times and the Task 1 and Task 2 values are presented side by side without a clear separator, making the ablation rows hard to read.","section":"Table 4"},{"comment":"The sentence 'the best EnhancerDFM model is achieved within 20 training epochs, while the best EnhancerDFM is obtained only in around 1400 training epochs' presumably should refer to the best Dirichlet FM model in the second clause; as written it is self-contradictory.","section":"Section 4.1"},{"comment":"The angle definition in Eq. (16) is undefined when the improvement vector Delta s(y_i,x) is zero; the paper should specify the convention for zero improvement vectors.","section":"Eq. (16)"},{"comment":"The abstract and Section 3 refer to steering toward 'Pareto-efficient' solutions, while the definition of the Pareto front in Section 3 is followed by the parenthetical '(not guaranteed to be Pareto optimal)'. These statements are in tension and should be reconciled in the revised text.","section":"Abstract and Section 3"},{"comment":"MOG-DFM is substantially slower than the baselines (43-90 s vs 8-37 s per sequence); since the conclusion emphasizes scalability, the paper should discuss whether this runtime gap is acceptable for practical use.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the circular evaluation and the definitional proof. I would advise the editor that the theoretical claim in Appendix D should be substantially rewritten or downgraded to a statement about scalarized expected improvement, and that at least one independent predictor for the non-fouling and hemolysis objectives, or a clear disclaiming of external validity, is needed before the Pareto-efficiency claims can be taken at face value. The paper's novelty relative to PepTune and Gumbel-Softmax flow matching should also be clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThis one is a genuine new method: as far as I can tell, MOG-DFM is the first multi-objective guidance scheme for discrete flow matching. The rank-directional score plus adaptive hypercone filter is distinct from Nisonoff's single-objective guidance, from ParetoFlow's continuous-space approach, and even from their own PepTune, which is diffusion-based. The algorithm is clearly specified with pseudocode, and they put real engineering effort into the two base models, PepDFM and EnhancerDFM, and into ablations on the hypercone module. The comparisons against NSGA-III, SPEA2, and others show large empirical wins on half-life, solubility, and non-fouling, and the ADMET-AI cross-check for solubility/half-life and the AlphaFold3/docking spot checks for binding are a good-faith attempt at external validation.\n\nNow the soft spots, in order of severity.\n\nThe Appendix D proof does not support the paper's Pareto-front claim, and it doesn't even describe the algorithm. The proof defines mu_i as a softmax over accepted transitions and takes an expectation under that distribution, but Algorithm 1 and Eq. 23 select the argmax and then move to it with probability governed by the guided rate. The proof also asserts that every accepted transition has D>0 because the cone angle is less than pi; that is only true when Phi<pi/2, and the fallback branch can pick a transition with negative D. So the formal claim is essentially true by construction, and even that construction only gives positive expected drift in one weight direction, not convergence to the Pareto front. I wouldn't ask for a full convergence proof in an applied paper, but the current appendix overclaims.\n\nThe evaluation is partly circular. Hemolysis, non-fouling, solubility, half-life, and affinity are all measured with the same predictors that generate the guidance signal. Only solubility and half-life get an independent check (ADMET-AI), and binding on a handful of targets. The DNA experiments use five sequences per condition with no error bars. And despite the contribution list promising flow-based baselines, the comparison set is evolutionary algorithms only; ParetoFlow is dismissed because it needs continuous inputs.\n\nThis is a solid engineering contribution, not a field-reorganizing one. I'd send it to review: a good referee can force the authors to drop or repair the convergence claim, add proper baselines, and either release code or say why they can't. The method itself is worth having in the literature.\n\nRecommendation: send to peer review, expect major revisions.","headline":"MOG-DFM is a real algorithmic contribution for steering discrete flow samplers, but the formal guarantee is essentially definitional and the evaluation is mostly circular; send it to review, but expect major revisions.","tokens_in":21692,"tokens_out":3949,"would_cite":true,"duration_ms":38963,"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":"Sampling-time guidance turns any pretrained discrete flow matching model into a multi-objective sequence designer.","keywords":["discrete flow matching","multi-objective optimization","Pareto front","peptide binder design","enhancer DNA design","guided generation","controllable sequence design","CTMC sampling"],"falsifier":"Run the full pipeline with the property scorers replaced by random noise; if the guided sequences still show large balanced gains on the original predictors, the apparent Pareto improvement is an artifact of the guidance loop rather than real optimization. A direct lab assay of designed binders, measuring hemolysis, solubility, half-life, and affinity, would also settle the question, because predictor-only gains would fail to reproduce in measurement.","tokens_in":1845,"feed_emoji":"🧬","tokens_out":3909,"duration_ms":93527,"temperature":0.7,"pith_summary":"At its core, MOG-DFM is a sampling-time steering layer for discrete flow matching. Instead of retraining a generator, it reweights the model's token-level transition velocities so that, step by step, the partially generated sequence moves in a direction that improves several competing properties at once. The paper shows that, under its scoring and filtering scheme, each accepted token replacement has positive expected improvement along a chosen trade-off weight vector, which in expectation directs generation toward the Pareto front. This is demonstrated on therapeutic peptide design with five simultaneous objectives, and on enhancer DNA design with class and shape objectives, where the method outperforms classical multi-objective optimizers and flow baselines in producing balanced profiles.","feed_headline":"One sampler balances five conflicting drug properties in peptides","feed_subtitle":"A rank-directional score plus adaptive hypercone filter steers flow matching toward Pareto-efficient trade-offs.","key_machinery":"The load-bearing objects are the guided transition score and the adaptive hypercone filter. For each candidate token replacement $y_i$ at position $i$, the score is $\\Delta S(y_i,x,\\omega) = \\mathrm{Norm}\\left(\\frac{1}{N}\\sum_{n=1}^N i_n I_n(y_i,x)\\right) + \\lambda\\, \\mathrm{Norm}\\left(\\Delta s(y_i,x)\\cdot\\omega\\right)$, where $I_n$ is the rank-normalized improvement of objective $n$ among all tokens at that position, $\\Delta s$ is the vector of score changes, and $\\omega$ is the sampled trade-off weight. The guided velocity is $\\beta u_t^i(y_i,x)\\exp(\\Delta S)$, which preserves the non-negativity and zero-sum rate conditions, so the process remains a valid continuous-time Markov chain. The hypercone accepts only candidates with angle $\\alpha_i = \\arccos\\left(\\frac{\\Delta s\\cdot\\omega}{\\|\\Delta s\\|\\|\\omega\\|}\\right) \\le \\Phi_t$, and $\\Phi_t$ adapts via an exponential moving average of the rejection rate, widening when too many candidates are rejected and narrowing when too few are. Together, these mechanisms bias the dynamics toward locally improving, directionally aligned moves and prevent the trajectory from drifting off the chosen trade-off direction.","core_discovery":"The paper claims that multi-objective control of discrete flow matching can be achieved purely at sampling time. The generator is treated as a black box; a precomputed set of weight vectors is sampled, and at each sampling step one position in the sequence is chosen and every alternative token at that position is scored. The score combines a rank-normalized per-objective improvement and a directional term that measures how well the improvement vector aligns with the sampled weight vector. The original transition velocity is multiplied by the exponent of this score, while keeping the non-negativity and zero-sum rate conditions intact. Then a hypercone filter retains only those token replacements whose improvement vector lies within an angular tolerance of the weight vector, and that tolerance adapts based on how many candidates are rejected. The paper proves that, if the candidate set is nonempty, the expected improvement along the weight vector is strictly positive. On peptide binders, the method yields simultaneous gains in hemolysis reduction, non-fouling, solubility, half-life, and affinity across multiple targets; on enhancer DNA, it achieves the requested class and shape values, and ablations show that removing any guidance criterion collapses that property.","pith_inferences":["Editorial inference: because the guidance only touches the velocity field and requires only scalar scorers, the same recipe should transfer to other discrete sequence domains such as RNA design, antibody CDR optimization, or promoter design without modifying the base model, though the paper does not demonstrate this.","Editorial inference: the proof establishes positive expected improvement along a single weight direction, but it does not establish a convergence rate to the full Pareto set; whether many runs with different weight vectors cover the front depends on the predictor landscape, so a formal coverage guarantee would be a natural next step.","Editorial inference: the paper cross-checks solubility and half-life with an independent predictor and binding with structure tools, but hemolysis and non-fouling rest on the original classifiers alone; a wet-lab assay or an additional orthogonal predictor for those two properties would strengthen the claim.","Editorial inference: because rank normalization amplifies small score differences, high-noise predictors such as the hemolysis classifier (F1 $=0.58$) could inject noise into the guidance signal; ensembling predictors or making the rank term uncertainty-aware is a plausible robustness extension the paper does not test."],"forward_implications":["Any pretrained discrete flow matching model over a finite vocabulary can be steered toward multi-objective trade-offs without retraining, by reweighting its token-level velocities at sampling time.","User-supplied weight vectors from a Das-Dennis simplex lattice let separate runs explore different regions of the Pareto front, so a single base generator can cover diverse trade-off preferences.","In peptide binder design, simultaneous guidance across hemolysis, non-fouling, solubility, half-life, and binding affinity moves all five scores in the improving direction over 100 iterations, with gains corroborated by an independent ADMET predictor and by docking and structure-prediction tools.","In enhancer DNA design, the same framework steers sequences toward a specified enhancer class and DNA shape value (HelT or Rise), and ablations show that removing a guidance criterion collapses that property.","The hypercone ablation shows that removing the filter collapses half-life gains from roughly 30-35 hours down to 4-13 hours, indicating that angular filtering is what allows gradual, coordinated objectives to progress."],"supporting_citations":[{"why":"Supplies the continuous-time Markov chain and factorized velocity-field formalism that MOG-DFM reweights for guidance.","marker":"[25]"},{"why":"Supplies the enhancer DNA dataset, the Dirichlet flow matching baseline used for comparison, and the enhancer class predictor used for guidance.","marker":"[27]"},{"why":"Supplies the single-objective rate-based guidance idea for pretrained discrete samplers that MOG-DFM extends to multiple objectives.","marker":"[30]"},{"why":"Supplies the prior multi-objective masked-diffusion framework the paper positions against and which motivates the discrete flow matching direction.","marker":"[23]"},{"why":"Supplies the continuous Pareto flow matching method that would be the natural baseline but cannot be applied because it requires continuous inputs.","marker":"[20]"},{"why":"Provides the independent ADMET-AI predictions used to cross-check the solubility and half-life of the designed binders.","marker":"[36]"},{"why":"Provides structure-prediction-based validation of binding for the designed peptides.","marker":"[34]"},{"why":"Provides docking-based validation of binding for the designed peptides.","marker":"[35]"}],"fun_headline_variants":["Pareto-efficient peptide and DNA design via multi-objective flow matching","Sampling-time guidance steers discrete flow matching to Pareto-optimal sequences","Five drug properties balanced in one peptide sampler","Multi-objective guidance for discrete flow matching in biomolecules","Rank-directional score plus hypercone filter hits multiple biophysical targets"],"cache_read_input_tokens":23808,"weakest_assumption_plain":"The quantitative claims assume that the machine-learning predictors used both to guide and to evaluate the sequences faithfully reflect real biological properties; if those predictors are biased, the reported Pareto improvements could be artifacts of the score functions rather than genuine gains in molecule quality.","fun_headline_variants_meta":{"raw":{"variants":["Pareto-efficient peptide and DNA design via multi-objective flow matching","Sampling-time guidance steers discrete flow matching to Pareto-optimal sequences","Five drug properties balanced in one peptide sampler","Multi-objective guidance for discrete flow matching in biomolecules","Rank-directional score plus hypercone filter hits multiple biophysical targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001109,"raw_usage":{"total_tokens":4644,"prompt_tokens":990,"completion_tokens":3654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3568}},"tokens_in":606,"tokens_out":3654,"duration_ms":23331,"temperature":1.0,"reasoning_tokens":3568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:25:28.737810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full pipeline with the property scorers replaced by random noise; if the guided sequences still show large balanced gains on the original predictors, the apparent Pareto improvement is an artifact of the guidance loop rather than real optimization. A direct lab assay of designed binders, measuring hemolysis, solubility, half-life, and affinity, would also settle the question, because predictor-only gains would fail to reproduce in measurement.","supporting_citations":[{"cited_title":"Discrete flow matching","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous-time Markov chain and factorized velocity-field formalism that MOG-DFM reweights for guidance."},{"cited_title":"Dirichlet flow matching with applications to dna sequence design.Proceedings of the 41st International Conference on Machine Learning (ICML), 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the enhancer DNA dataset, the Dirichlet flow matching baseline used for comparison, and the enhancer class predictor used for guidance."},{"cited_title":"Unlocking guid- ance for discrete state-space diffusion and flow models","cited_arxiv_id":null,"evidence_quote":"Supplies the single-objective rate-based guidance idea for pretrained discrete samplers that MOG-DFM extends to multiple objectives."},{"cited_title":"Peptune: De novo generation of therapeutic peptides with multi-objective-guided discrete diffusion","cited_arxiv_id":null,"evidence_quote":"Supplies the prior multi-objective masked-diffusion framework the paper positions against and which motivates the discrete flow matching direction."},{"cited_title":"ParetoFlow: Guided Flows in Multi-Objective Optimization","cited_arxiv_id":"2412.03718","evidence_quote":"Supplies the continuous Pareto flow matching method that would be the natural baseline but cannot be applied because it requires continuous inputs."},{"cited_title":"Admet-ai: a machine learning admet platform for evaluation of large-scale chemical libraries","cited_arxiv_id":null,"evidence_quote":"Provides the independent ADMET-AI predictions used to cross-check the solubility and half-life of the designed binders."}],"review_version":1}