{"id":"57613a7e-ac4e-4ba0-85a1-606a7d82508a","arxiv_id":"2411.15183","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CMOMO is an evolutionary framework that balances multi-property optimization with constraint satisfaction in latent molecular space and outperforms five baselines on four tasks.","lead":"This paper introduces CMOMO, an evolutionary method that optimizes multiple molecular properties at once while enforcing drug-like constraints. It reports better success rates and solution diversity than five existing methods on four optimization tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CMOMO's comparison is confounded by an external Bank-library initialization that baselines do not receive; reported SR/HV margins may reflect information leakage rather than algorithmic superiority.","rationale":"The reader's weakest assumption concerns the smoothness and decoding validity of the pretrained latent space. That is a real concern, but it applies to most baselines as well (QMO, MSO, and MOMO also operate in continuous latent spaces), so it does not specifically undermine the claimed relative advantage of CMOMO. The more load-bearing weakness is the asymmetric access to external high-property molecules through the Bank library. This enters directly into the main empirical claim: if the Bank already contains molecules that satisfy the success thresholds, then the crossover initialization hands CMOMO near-solved starting points, and the reported SR and HV comparisons in Figures 3A-3D are not apples-to-apples. The paper explicitly states that all methods use the same population size and iterations, but it does not state that baselines receive an equivalent initialization or external database screen. The ablation removing the Bank library is insufficient because it is not benchmarked against the five comparison methods. I retain the reader's conditional verdict rather than escalating to rejection because the fairness issue is concrete and testable with the released code: a controlled re-run with matched initialization would settle whether the claimed advantage survives. If that check fails, the central performance claim should be downgraded.","tokens_in":17876,"tokens_out":8034,"duration_ms":94644,"concrete_test":"Re-run Tasks 1 and 2 with CMOMO's Bank-library initialization replaced by P random latent perturbations of each lead (or by the same initialization used by each baseline), keeping the rest of the CMOMO pipeline and all evaluation metrics unchanged; if CMOMO's SR/HV margins over MOMO, GB-GA-P, and Molfinder shrink substantially or disappear, the reported superiority is attributable to the external Bank library rather than to the proposed algorithm.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that CMOMO outperforms five state-of-the-art methods is not supported under a fair comparison. In Section 4.2, CMOMO builds, for each lead molecule, a Bank library of high-property molecules screened from public databases, and initializes its population by linear crossover between the lead vector and each Bank vector (Fig. S6). The five baselines are not given this external high-property knowledge; Section 2.2.2 only states that all methods use the same population size, number of samples, and iterations. If Bank molecules already satisfy the success thresholds (e.g., Task 1: QED≥0.85, PlogP_imp≥3, Similarity≥0.3), the large SR and HV gaps in Fig. 3 could be caused by starting from near-solution molecules rather than by the proposed dynamic cooperative optimization or ranking-aggregation constraint handling. The ablation in Section 2.7 removes the Bank library (CMOMO_nobank) but reports results only against CMOMO on Task 1, not against the baselines, so it does not establish that a Bank-free CMOMO still dominates. Without a matched-initialization comparison, the performance claim is unverified as stated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CMOMO, a constrained multi-objective molecular optimization framework. CMOMO initializes a population by screening a Bank library of high-property molecules from public databases and performing linear crossover in the latent space of a pretrained encoder-decoder; it then evolves molecules using a vector fragmentation-based evolutionary reproduction (VFER) strategy. A two-stage dynamic constraint handling strategy first performs unconstrained NSGA-II-style selection and then switches to a ranking-aggregation selection that interpolates property-preference and constraint-preference ranks via a cosine-decayed weight. The framework is evaluated on four tasks (two benchmark tasks and two docking/inhibition tasks) against QMO, MolFinder, MOMO, MSO, and GB-GA-P, using success rate, number of successfully optimized molecules, hypervolume, and mean property values. The paper reports that CMOMO achieves the highest success rate and hypervolume on all four tasks.","tokens_in":18131,"tokens_out":4739,"duration_ms":50995,"significance":"If the comparison is fair, CMOMO is a useful contribution to constrained multi-objective molecular optimization: it combines latent-space evolutionary search with explicit constraint handling, and the two-stage ranking aggregation is a sensible design. The practical docking tasks and the case studies add applied value, and the authors promise public code. The main performance claim, however, currently rests on a comparison in which CMOMO receives an external Bank-library initialization that the baselines do not receive; until that confound is removed or controlled, the quantitative superiority claim is not established. The paper's positive aspects include the clear problem formulation, the ablation study (though only on Task 1), and the attention to practical drug-like constraints.","major_comments":[{"comment":"The claim that all methods use the same population size, number of samples, and iterations does not make the comparison fair, because CMOMO additionally receives a Bank-library initialization. In Section 4.2, CMOMO builds a Bank library of high-property molecules screened from public databases for each lead molecule and then initializes its population by linear crossover between the lead vector and each Bank vector (Fig. S6). The five baselines are not described as receiving this external high-property knowledge. If the Bank molecules already satisfy the success thresholds, the large SR and HV margins reported in Fig. 3 could be caused by starting from near-solution molecules rather than by the proposed dynamic cooperative optimization or ranking-aggregation constraint handling. The ablation in Section 2.7 removes the Bank library (CMOMO_nobank) but reports results only against CMOMO on Task 1, not against the baselines, so it does not establish that a Bank-free CMOMO still dominates. Please add a matched-initialization comparison in which baselines receive the same Bank-derived initial population, or alternatively compare CMOMO_nobank against all baselines on all tasks, and report how many Bank molecules already satisfy the success thresholds per lead.","section":"Section 4.2 and Section 2.2.2"},{"comment":"All reported performance metrics are point estimates without error bars, multiple runs, or statistical significance testing. The phrase 'significantly better performance' in Section 2.3 is therefore not supported in the statistical sense. Since the compared methods are stochastic evolutionary algorithms, run-to-run variance is expected. Please report mean and standard deviation over repeated runs, and support the SR/HV ordering with paired tests across lead molecules or bootstrapped confidence intervals.","section":"Section 2.3 and Fig. 3"},{"comment":"The success thresholds used to define SR are hand-chosen (e.g., Task 1: QED ≥ 0.85, PlogP_imp ≥ 3, Similarity ≥ 0.3; Task 2: all four scores ≥ 0.5 or 0.3), and the hypervolume reference point is set to the zero vector. Because the SR comparison depends entirely on these thresholds, the observed ordering could be an artifact of threshold placement. Please provide a sensitivity analysis over threshold values, or report a threshold-independent evaluation such as the feasible hypervolume across varying thresholds and the full property distributions of the optimized sets.","section":"Section 2.2.3"},{"comment":"The method relies on the pretrained encoder-decoder being smooth and on decoded offspring being valid and evaluable, but no validity rate is reported for CMOMO or for the baselines that also operate in the same latent space. If decoding frequently fails or produces invalid molecules, the comparison may conflate search quality with decoder behavior. Please report the fraction of generated offspring that decode to valid SMILES and the fraction that are feasible, and state how these rates compare across methods. This is important for interpreting the VFER ablation in Section 2.7, whose benefit could partly reflect differences in offspring validity rather than in search efficiency.","section":"Section 4.3 and Section 4.2"}],"minor_comments":[{"comment":"There are several typos: 'matrial science' in the Introduction, 'as can bee seen' in Section 2.3, and 'access the quality' where 'assess' is meant. These should be corrected in a revision.","section":"Introduction and Section 2.3"},{"comment":"The text says 'A link to access the supplementary materials will be provided in the published article.' For an arXiv submission, the supplementary file should be included or the repository should contain the missing figures and experimental details (e.g., Bank library screening thresholds and the exact baselines' implementations).","section":"Supplementary Materials"},{"comment":"The histograms are labeled 'for each lead molecule,' but the x-axis labels are difficult to parse (e.g., 'Mol 1 10050 7525'). Please provide a clearer axis label, a sorted lead index, or a caption that explains the grouping.","section":"Fig. 3E and 3F"},{"comment":"The cosine decay for alpha is introduced without a sensitivity analysis. Since Eq. (7) controls the entire balance between property and constraint preferences, a brief robustness check over different schedules (e.g., fixed alpha, linear decay) would strengthen the claim that the dynamic schedule is beneficial.","section":"Section 4.4.2, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The Bank-library initialization confound is the decisive issue. I would like to see matched-initialization experiments and repeated-run statistics before publication. If those experiments support the current conclusions, the paper would be a reasonable contribution to the molecular optimization literature. The method itself appears sound; the weakness is in the evaluation design rather than in the proposed algorithm."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a mostly competent extension of the authors' earlier MOMO to constrained multi-property optimization. The new mechanism is two-stage search: first unconstrained NSGA-II in the latent space, then a ranking-aggregation selection that blends property ranks with constraint-violation ranks. They also add a fragmentation mutation operator and a Bank-library initialization that seeds the population with linear crossovers between the lead and high-property molecules screened from public databases. The ablation shows each component helps on Task 1. Code and data are available.\n\nWhat it does well: the problem formulation is sensible, the experimental setup covers four tasks including two with docking and a GSK3β surrogate, and the property evaluators are independent of the optimizer. The case studies and trajectory plots give useful qualitative insight.\n\nNow the soft spots. The main one is the comparison. The Bank library gives CMOMO a head start that the five baselines do not receive. The paper claims fairness because all methods use the same population size, samples, and iterations, but that does not control for initialization. If many Bank molecules already satisfy the success thresholds, the reported SR and HV gaps could reflect the quality of the external database rather than the algorithmic machinery. The ablation does not fix this: CMOMO_nobank is only compared against CMOMO, not against the baselines, so we still do not know whether a Bank-free CMOMO would beat QMO, MOMO, MSO, etc. This is a load-bearing flaw in the central performance claim.\n\nSecondary issues: there are no error bars or significance tests over multiple runs, and the success thresholds look hand-chosen. The latent-space smoothness assumption is inherited from Winter et al. without new validation. The supplementary parameter details are deferred to a later publication, which makes reproduction harder.\n\nEven with those caveats, the framework idea is plausibly useful and the implementation is careful enough to deserve referee time. The paper would need a matched-initialization comparison (e.g., give baselines the same Bank-seeded starting population or run CMOMO without Bank against all baselines) and some statistical grounding before the performance claim is established.\n\nVerdict: worth a serious peer review, but the authors should be asked to address the confound. I would not cite it yet as evidence of superiority; I would cite it as a method description.\n\nBest.","headline":"Useful extension of MOMO to constrained multi-objective optimization, but the headline performance claim is confounded by an unequal Bank-library initialization and lacks statistical grounding.","tokens_in":18637,"tokens_out":2234,"would_cite":false,"duration_ms":23023,"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 two-stage evolutionary framework balances molecular property optimization with drug-like constraints and outperforms five baselines.","keywords":["constrained multi-objective optimization","molecular optimization","latent space evolution","drug-like constraints","evolutionary algorithm","vector fragmentation mutation","ranking aggregation","drug discovery"],"falsifier":"Run CMOMO's VFER strategy on a benchmark set and record, at every generation, the fraction of decoded offspring that are valid SMILES and satisfy the ring-size and toxic-substructure constraints; if offspring validity or feasibility collapses under latent-space mutation, or if CMOMO's advantage over baselines disappears when the pretrained encoder-decoder is replaced by a non-smooth representation, the central claim would be refuted.","tokens_in":17678,"feed_emoji":"🧪","tokens_out":7644,"duration_ms":73970,"temperature":0.7,"pith_summary":"This paper proposes CMOMO, a constrained multi-property molecular optimization framework that treats each drug-like criterion as a constraint rather than folding it into an aggregated objective. The central claim is that by first optimizing multiple molecular properties without constraints, then gradually shifting selection pressure toward constraint satisfaction, CMOMO finds molecules that are both high-quality and feasible more often than existing methods. On two benchmark tasks and two drug-target tasks, CMOMO reports higher success rates, larger hypervolumes, and more successfully optimized molecules than five comparison methods. The practical payoff is a way to generate candidate ligands and inhibitors that respect drug-like constraints, which is what makes an optimized molecule actually usable in drug discovery.","feed_headline":"Two-stage search beats five methods on constrained drug optimization","feed_subtitle":"CMOMO raises success rates to 94.8% on benchmark Task 1 while keeping molecules within drug-like constraints.","key_machinery":"The load-bearing mechanism is the dynamic cooperative optimization loop, which alternates between a continuous implicit space of latent vectors and the discrete chemical space of SMILES strings. Offspring are generated in latent space by VFER (vector-fragmentation-based evolutionary reproduction), which applies blended linear crossover and then mutates one randomly chosen fragment of the latent vector; decoded molecules are evaluated for properties and selection feeds back into latent space. The dynamic constraint-handling strategy splits evolution into an unconstrained stage that uses non-dominated sorting and crowding distance to build convergence and diversity, and a constrained stage that ranks each candidate twice — once by properties and once by constraint dominance — then combines the two ranks with a weight α that decays from 1 to 0 across generations, so the search gradually shifts from pure property optimization to constraint satisfaction.","core_discovery":"On the paper's own terms, the discovery is that constrained multi-property molecular optimization is best modeled as a constrained multi-objective problem, and that a two-scenario evolutionary search can solve it better than aggregating objectives with constraints or discarding infeasible candidates. In the unconstrained scenario, CMOMO evolves latent vectors using non-dominated sorting and crowding distance to build a diverse, well-converged set; in the constrained scenario, it ranks candidates by properties and by constraint dominance and combines the ranks with a decaying weight, finally returning feasible molecules on the constrained Pareto front. Empirically, CMOMO reports the highest success rate on both benchmark tasks — 94.8% on Task 1 and 70.87% on Task 2 — and the largest mean hypervolume among the multi-objective comparison methods. On the practical tasks, it reports 75% success for β2-adrenoceptor ligands versus 59% for the best baseline, and 52.7% for GSK3β inhibitors versus at most 24.6% for baselines, with optimized molecules reaching docking energies below -7 kcal/mol. The ablation study attributes the largest performance drop to removing the vector-fragmentation-based mutation, suggesting that fragment-level mutation in the latent space is the key enabler.","pith_inferences":["Editorial extension: The paper's comparison does not isolate the contribution of the Bank-library initialization from the two-scenario search; a natural follow-up would vary Bank size and similarity thresholds to measure how much of the success-rate gain comes from initialization quality.","Editorial extension: The smooth-latent-space premise inherited from the pretrained encoder-decoder is not stress-tested in the paper; a direct experiment tracking the validity rate and property distribution of decoded offspring over generations would show how much of CMOMO's search power depends on that premise.","Editorial extension: The decaying weight α implements a soft transition from property preference to constraint preference, suggesting the framework could be adapted to time-varying or soft constraints by replacing the fixed schedule with a feedback controller that responds to the current feasible/infeasible ratio of the population."],"forward_implications":["If CMOMO is correct, constrained multi-property molecular optimization can be treated as a two-stage search: first build a diverse set of good-property molecules, then apply constraint pressure gradually, rather than mixing constraints into a single fitness from the start.","The vector-fragmentation-based mutation suggests that mutating a block of latent dimensions at once is more effective for molecule generation than single-gene mutation, reducing the search-dimension problem in long latent vectors.","Because the framework is agnostic to how properties are computed, its two-scenario design should transfer to other objectives, constraint sets, and property predictors without re-architecting the search.","The demonstrations on β2-adrenoceptor and GSK3β targets imply the method can propose candidate ligands and inhibitors that satisfy structural constraints and reach binding-energy thresholds used in drug-likeness screening."],"supporting_citations":[{"why":"Supplies the pretrained encoder-decoder that maps SMILES into the continuous latent space where CMOMO performs crossover and mutation.","marker":"[34]"},{"why":"Provides the blended linear crossover operator used to generate offspring latent vectors.","marker":"[35]"},{"why":"Supplies the non-dominated sorting and crowding-distance selection scheme used in the unconstrained optimization stage.","marker":"[36]"},{"why":"Defines the 163 toxic/uncommon substructures used in constraint C2 and provides a comparison baseline with an aggregation strategy.","marker":"[32]"},{"why":"QMO is a comparison method; demonstrates molecular optimization in the same latent space, against which CMOMO reports success rate.","marker":"[29]"},{"why":"MolFinder is a comparison method; an evolutionary SMILES-based optimizer against which CMOMO reports success rate and hypervolume.","marker":"[30]"},{"why":"MOMO is a comparison method; a multi-objective evolutionary optimizer in implicit chemical space against which CMOMO reports all metrics.","marker":"[31]"},{"why":"GB-GA-P is a comparison method; a graph-based Pareto genetic algorithm that discards infeasible molecules, against which CMOMO reports all metrics.","marker":"[33]"},{"why":"Supplies the QED drug-likeness objective used in Tasks 1, 3, and 4.","marker":"[37]"},{"why":"Supplies the structural scores (dissimilarity, molecular weight, rotatable bonds) used as objectives in Task 2.","marker":"[40]"}],"fun_headline_variants":["CMOMO: two-scenario evolution outperforms five methods on constrained drug optimization","Constrained molecular optimization: CMOMO hits 94.8% success on Task 1","Better drug candidates: CMOMO balances properties and constraints","Two-scenario evolution for constrained drug optimization: CMOMO wins","Constrained multi-property optimization: CMOMO beats five methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole search runs in a latent space produced by a pretrained encoder-decoder, and the method's success depends on that space being smooth enough that linear blends and fragment mutations decode back into valid, property-relevant molecules.","fun_headline_variants_meta":{"raw":{"variants":["CMOMO: two-scenario evolution outperforms five methods on constrained drug optimization","Constrained molecular optimization: CMOMO hits 94.8% success on Task 1","Better drug candidates: CMOMO balances properties and constraints","Two-scenario evolution for constrained drug optimization: CMOMO wins","Constrained multi-property optimization: CMOMO beats five methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001155,"raw_usage":{"total_tokens":4828,"prompt_tokens":1033,"completion_tokens":3795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":3700}},"tokens_in":649,"tokens_out":3795,"duration_ms":28670,"temperature":1.0,"reasoning_tokens":3700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:49:53.081164+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CMOMO's VFER strategy on a benchmark set and record, at every generation, the fraction of decoded offspring that are valid SMILES and satisfy the ring-size and toxic-substructure constraints; if offspring validity or feasibility collapses under latent-space mutation, or if CMOMO's advantage over baselines disappears when the pretrained encoder-decoder is replaced by a non-smooth representation, the central claim would be refuted.","supporting_citations":[{"cited_title":"Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained encoder-decoder that maps SMILES into the continuous latent space where CMOMO performs crossover and mutation."},{"cited_title":"A crossover operator using independent component analysis for real- coded genetic algorithms","cited_arxiv_id":null,"evidence_quote":"Provides the blended linear crossover operator used to generate offspring latent vectors."},{"cited_title":"A fast and elitist multiobjective genetic algorithm: NSGA-II","cited_arxiv_id":null,"evidence_quote":"Supplies the non-dominated sorting and crowding-distance selection scheme used in the unconstrained optimization stage."},{"cited_title":"Efficient multi-objective molecular optimization in a continuous latent space","cited_arxiv_id":null,"evidence_quote":"Defines the 163 toxic/uncommon substructures used in constraint C2 and provides a comparison baseline with an aggregation strategy."},{"cited_title":"Optimizing molecules using efficient queries from property evaluations","cited_arxiv_id":null,"evidence_quote":"QMO is a comparison method; demonstrates molecular optimization in the same latent space, against which CMOMO reports success rate."},{"cited_title":"MolFinder: an evolutionary algorithm for the global optimization of molec- ular properties and the extensive exploration of chemical space using SMILES","cited_arxiv_id":null,"evidence_quote":"MolFinder is a comparison method; an evolutionary SMILES-based optimizer against which CMOMO reports success rate and hypervolume."},{"cited_title":"Graph-based molecular Pareto optimisation","cited_arxiv_id":null,"evidence_quote":"GB-GA-P is a comparison method; a graph-based Pareto genetic algorithm that discards infeasible molecules, against which CMOMO reports all metrics."},{"cited_title":"Quantifying the chemical beauty of drugs","cited_arxiv_id":null,"evidence_quote":"Supplies the QED drug-likeness objective used in Tasks 1, 3, and 4."},{"cited_title":"GuacaMol: benchmarking models for de novo molecular design","cited_arxiv_id":null,"evidence_quote":"Supplies the structural scores (dissimilarity, molecular weight, rotatable bonds) used as objectives in Task 2."}],"review_version":1}