{"id":"f0ea34ed-71da-416a-8ec0-43de24bb51e9","arxiv_id":"2607.03881","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Smooth %MinMax replaces discrete codon usage and hard min/max branching with probability-weighted averages and a sigmoid gate, enabling gradient-based neural codon harmonization that approximates the hard metric.","lead":"The authors define Smooth %MinMax, a differentiable version of the classic %MinMax codon-usage profile so neural models can optimize codon probabilities by gradient descent. It lets profile-based codon harmonization plug into modern sequence-design networks instead of only discrete search.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-noted discretization gap.","rationale":"The reader's weakest-assumption diagnosis is precise and correctly scoped: β=0.1 is fixed by one-hot fidelity trade-off (§3.4), the loss is pure MAE in probability space (§4), and no discrete sequences or wet-lab outcomes are produced. That is the principal limitation for any stronger claim of 'useful designs,' but the paper's actual strongest claim is only that the smooth metric approximates the hard one and supports gradient-based matching. The mathematics (probability-weighted usage + sigmoid gate) is transparent, the one-hot approximation experiments are thorough, and the neural proof-of-concept shows loss decrease with nonzero gradients. No additional load-bearing flaw (e.g., incorrect normalization, non-differentiable path, or contradictory numerical result) appears. Therefore the CONDITIONAL verdict with high confidence already captures the right posture; no adjustment is warranted. The concrete test above simply operationalizes the missing discretization check the reader already requires.","tokens_in":13568,"tokens_out":622,"duration_ms":5527,"concrete_test":"Take the trained model on a held-out test chunk, sample or argmax a discrete codon sequence from the final synonymous probabilities, recompute conventional hard %MinMax_[h] with the same window=10 and E. coli usage table, and report MAE/Pearson against the original human target profile. If discrete hard-profile MAE remains near the smooth training MAE (~7) rather than collapsing toward random (~30+), the bridge claim is strengthened; a large jump would quantify the discretization gap the reader already notes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is modest and well-supported: under one-hot synonymous distributions, %MinMax_[s] with β=0.1 recovers %MinMax_[h] to high numerical fidelity (MAE ~0.01–0.1 range, Pearson ~1; §3.4, Fig. 1), and the same smooth objective yields nonzero gradients that reduce profile MAE from ~37 to ~7 on the BiLSTM model (§4, Fig. 2). The sigmoid-gated denominator and probability-weighted usage are standard, differentiable relaxations that preserve the signed interpretation of the hard metric. The only material soft spot is exactly the one the reader already flags: the work never converts optimized probabilities into discrete codon sequences, never recomputes hard %MinMax on those sequences, and never compares against CHARMING/MOSAIC or any expression assay. That gap limits claims of practical utility for design, but it does not undermine the stated technical claim that the relaxation approximates and supports gradient-based profile matching in probability space. No hidden mathematical inconsistency or unstated assumption that would invalidate the approximation or the gradient flow was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces Smooth %MinMax (%MinMax_[s]), a differentiable relaxation of the classical hard %MinMax (%MinMax_[h]) profile used in codon harmonization. Discrete codon-usage values are replaced by probability-weighted averages over synonymous codons, and the hard branch between positive and negative normalization denominators is replaced by a sigmoid-gated interpolation controlled by a sharpness parameter β. The authors show that, under one-hot synonymous distributions, %MinMax_[s] recovers %MinMax_[h] to high numerical fidelity (MAE, MSE, Pearson correlation versus β; Section 3.4, Figure 1) and that the smooth objective yields nonzero gradients that reduce profile-matching MAE from ~37 to ~7 in a simple BiLSTM model for human-to-E. coli harmonization (Section 4, Figure 2). Code and data are released.","tokens_in":13812,"tokens_out":934,"duration_ms":7634,"significance":"If the technical claim holds, the work supplies a missing differentiable objective that lets classical profile-based codon harmonization be used inside gradient-based neural codon design. The formulation is transparent, the one-hot approximation study is quantitative, and the training curves confirm both loss decrease and gradient flow. Public code further strengthens reproducibility. The contribution is modest but useful: it does not claim new biological designs or expression gains, only a practical bridge between %MinMax profiles and probabilistic sequence models. That bridge is of clear interest to the synthetic-biology and computational-protein-design communities.","major_comments":[{"comment":"Section 4 and Conclusions: the central practical claim that %MinMax_[s] is a bridge to neural synonymous-sequence design is only partially supported. Optimization is performed entirely in probability space; the manuscript never samples discrete codon sequences from the optimized distributions, never recomputes hard %MinMax_[h] on those sequences, and never compares the resulting profiles (or any other metric) against CHARMING, MOSAIC, or a simple CAI baseline. Without this discretization step the utility for actual gene design remains untested, even though the mathematical differentiability claim itself is sound.","section":null},{"comment":"Section 3.4: β = 0.1 is selected by a visual/MAE trade-off on one-hot human sequences and then fixed for all subsequent experiments. No sensitivity analysis is reported for the neural optimization itself (e.g., whether β = 0.01 or 1.0 changes final profile MAE or gradient norms). Because the gate sharpness directly affects both approximation fidelity and gradient magnitude near Δ̃_t = 0, a short ablation on the training objective would strengthen that the chosen value is not an accidental local optimum.","section":null}],"minor_comments":[{"comment":"Throughout: the notation %MinMax_[s] / %MinMax_[h] is clear, but the manuscript also uses several near-identical glyph variants (e.g., %MinMax, %MinMax[s]) that render inconsistently; a single consistent macro would improve readability.","section":null},{"comment":"Figure 2a: the architecture diagram is helpful, yet the precise definition of the profile-context radius r (and why r = 5 was chosen after a pre-screen) is only stated in prose; adding the radius to the figure legend would make the input construction self-contained.","section":null},{"comment":"Section 3.2: the usage frequency u(c) is defined as occurrences per 1 000 codons; a brief note on which codon-usage table (and organism version) was used for both human and E. coli would aid exact reproduction.","section":null},{"comment":"Supporting Figure S1: six random test profiles are shown; reporting the mean test-set MAE (or Pearson) between smooth and hard profiles after training would quantify generalization more cleanly than visual inspection alone.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a clean methods contribution whose main limitation (no discrete-sequence evaluation) is already acknowledged by the authors in the Conclusions. I would not require wet-lab expression data for acceptance; a short in-silico discretization experiment would be sufficient. Scope fits q-bio.QM / computational synthetic biology well."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this paper solves a real, narrow technical mismatch: classical %MinMax is discrete and non-differentiable, while neural codon design needs gradients. They give a transparent fix—probability-weighted synonymous usage plus a sigmoid-gated blend of the positive and negative denominators—and show it works.\n\nWhat is new is the formulation itself. Under one-hot encodings of real human sequences the smooth metric recovers the hard profile to high fidelity (MAE drops to ~0.01 range, Pearson essentially 1 at β=0.1; Fig. 1). The BiLSTM experiment then shows the loss is minimizable (MAE from ~37 to ~7) with nonzero gradients flowing through the smooth objective (Fig. 2). Math is elementary and correct; the approximation study is quantitative; code and data are released. Citations cover the right prior art (Rodriguez %MinMax, CHARMING, MOSAIC, recent neural codon optimizers). No circularity: the metric is defined independently of the optimization results.\n\nThe soft spot is exactly the one already flagged and is proportionate: they never turn the optimized probabilities into discrete codon sequences, never recompute hard %MinMax on those sequences, and never compare against search-based harmonizers or any expression/folding assay. β=0.1 is a fixed hyperparameter chosen by visual/MAE trade-off; that is fine for a methods paper but leaves open how well the continuous optimum survives discretization. The neural architecture is deliberately minimal, so the experiment only demonstrates that the smooth loss can be optimized, not that the resulting designs are useful.\n\nThis is for people building multi-objective neural codon optimizers who want a drop-in, signed, profile-based harmonization term. It is not a biology paper and does not claim to be. The central technical claim is modest and well-supported; the work is a legitimate methods contribution. I would send it to peer review.","headline":"Clean, usable differentiable relaxation of %MinMax that does exactly what it claims in probability space; the only real gap is the missing discretization step.","tokens_in":14402,"tokens_out":474,"would_cite":true,"duration_ms":4600,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A smooth, differentiable form of the %MinMax codon-usage metric makes neural profile matching possible for codon harmonization.","keywords":["codon harmonization","Smooth %MinMax","differentiable relaxation","synonymous codon design","neural sequence design","profile matching","heterologous expression"],"falsifier":"Take the optimized synonymous-codon probability distributions, sample or greedily decode discrete coding sequences, recompute classical hard %MinMax profiles on those sequences, and check whether they still match the human target profiles under the E. coli codon table; large systematic divergence would refute the claim that the smooth objective is a practical bridge.","tokens_in":14433,"feed_emoji":"🧬","tokens_out":626,"duration_ms":5744,"temperature":0.7,"pith_summary":"Codon harmonization tries to redesign a gene so a foreign host will translate it with the same local mix of common and rare codons that the original gene had, because those patterns can matter for how the protein folds while it is still being made. The usual %MinMax score that measures those local patterns only works on finished, discrete codon sequences, so it cannot be used directly with modern neural designers that update codon probabilities by gradient descent. This paper constructs Smooth %MinMax by replacing each hard codon choice with a probability-weighted average of synonymous-codon usages and replacing the hard positive/negative branch with a soft sigmoid gate. The new score stays close to the classical profile, keeps the same signed meaning (positive = enrichment of frequent codons, negative = enrichment of rare ones), and supplies usable gradients. In human-to-E. coli experiments the smooth profiles track the hard ones and the matching loss falls under gradient updates, showing that classical harmonization profiles can now sit inside neural sequence-design loops.","feed_headline":"Smooth %MinMax lets neural models match codon-usage profiles","feed_subtitle":"A soft version of the classical metric supplies gradients for human-to-E. coli codon harmonization","key_machinery":"Smooth %MinMax (%MinMax_[s]): the residue-level codon-usage value is replaced by its expectation under the synonymous-codon softmax, the window deviation is formed as usual, and the hard max/min denominator switch is replaced by a sigmoid-gated convex combination controlled by a sharpness hyperparameter β.","core_discovery":"Smooth %MinMax, obtained by probability-weighted synonymous-codon usage and a sigmoid-gated interpolation between the positive and negative normalization branches, closely approximates conventional hard %MinMax while remaining fully differentiable with respect to codon probabilities and model parameters, thereby enabling gradient-based profile matching for neural codon harmonization.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Smooth %MinMax enables gradient-based neural codon profile matching","Differentiable soft %MinMax approximates hard metric for codon design","Probability-weighted %MinMax supports neural synonymous sequence optimization","Sigmoid-gated Smooth %MinMax bridges classical harmonization to gradients","Soft %MinMax preserves signed interpretation while staying fully differentiable"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That matching profiles in soft probability space with a fixed smoothing strength will still produce useful discrete codon sequences once the probabilities are later turned into single codon choices.","fun_headline_variants_meta":{"raw":{"variants":["Smooth %MinMax enables gradient-based neural codon profile matching","Differentiable soft %MinMax approximates hard metric for codon design","Probability-weighted %MinMax supports neural synonymous sequence optimization","Sigmoid-gated Smooth %MinMax bridges classical harmonization to gradients","Soft %MinMax preserves signed interpretation while staying fully differentiable"]},"model":"grok-4.5","effort":"low","cost_usd":0.007014,"raw_usage":{"total_tokens":1753,"prompt_tokens":780,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":70140000,"prompt_tokens_details":{"text_tokens":780,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":884,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":780,"tokens_out":89,"duration_ms":6518,"temperature":1.0,"reasoning_tokens":884,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T23:16:21.028659+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Take the optimized synonymous-codon probability distributions, sample or greedily decode discrete coding sequences, recompute classical hard %MinMax profiles on those sequences, and check whether they still match the human target profiles under the E. coli codon table; large systematic divergence would refute the claim that the smooth objective is a practical bridge.","supporting_citations":[],"review_version":1}