{"id":"16f655c1-0c98-4ed8-a3d8-3e0311fa3239","arxiv_id":"2411.14726","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"GraphTRL feeds multiscale weighted colored graphs and persistent images into a dueling deep Q-network and reports small gains on two drug-likeness metrics, not on the claimed binding affinity.","lead":"GraphTRL adds shape and topology descriptors to a reinforcement learning agent that edits drug-like molecules, reporting slightly higher penalized logP and QED scores than MolDQN variants. The paper claims better binding-affinity prediction, but no binding-affinity experiment appears in the evaluation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's binding-affinity claim is untested: Table 1 reports penalized logP and QED only, so the central claim is unsupported regardless of baseline comparability.","rationale":"The reader's verdict is REJECT with 'baseline comparability' as the weakest assumption. I agree with the rejection, but the more load-bearing concern is more fundamental: the paper's stated central claim about binding-affinity prediction is not tested at all. The full text contains no affinity-related experiment, no affinity metric, and no affinity benchmark. This is not a matter of baseline numbers being non-comparable; the entire evaluation is about penalized logP and QED, which are drug-likeness optimization metrics, not binding-affinity prediction. Even if Table 1 were perfectly reproducible and all baselines were re-run under identical settings, the abstract's claim would remain unsupported. The reader's criticism of baseline comparability is a valid secondary issue, but fixing it would not change the outcome. Therefore the verdict remains REJECT, with the primary reason being the missing affinity evaluation rather than baseline comparability.","tokens_in":4731,"tokens_out":2658,"duration_ms":28232,"concrete_test":"Conduct a binding-affinity evaluation using the MWCG/persistent-image features described in Section 2: train a supervised predictor on a standard benchmark (e.g., PDBbind v2019 refined set with the same train/test splits used by EISA-Score or PDFL-ML) and report RMSE, Pearson correlation, and comparison to those baselines under identical protocols. If the intended claim is about the generated molecules' predicted affinity instead, compute predicted affinity scores for GraphTRL-generated and baseline-generated molecules and compare their distributions with error bars. If neither experiment is run, the abstract's binding-affinity claim should be withdrawn.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim, stated in the abstract, is that GraphTRL 'outperforms existing methods in binding affinity prediction.' For this claim to hold, the paper must provide an evaluation of binding-affinity prediction, or at least a comparison of affinity-related scores for generated molecules, against existing methods. No such evaluation appears anywhere in the manuscript. Section 3 describes an experiment on 800 ZINC molecules and reports 'the average improvement in Penalized logP and QED value' in Table 1. Table 1's metrics are penalized logP and QED only; there is no binding-affinity metric, no PDBbind or docking benchmark, no IC50/Kd prediction, and no comparison to the affinity-prediction methods cited in Refs. [6,7] (EISA-Score, PDFL-ML). The MWCG and persistent-image features in Section 2 are used to construct the RL state, but no supervised affinity-prediction model is trained or tested with these features, and neither reward function R1 nor R2 includes affinity. Thus the abstract's central claim is not merely weakly supported; it is completely unaddressed by the experimental section. The reader's concern about baseline comparability is valid for the optimization claim, but even a perfectly controlled re-run of Table 1 would not substantiate a binding-affinity advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GraphTRL, a reinforcement learning framework for molecular optimization that augments SMILES-based states with multiscale weighted colored graphs (MWCG) and persistent images derived from persistent homology. The agent uses a dueling DQN with actions for atom addition, bond addition, and bond removal, and reward functions based on penalized logP, synthetic accessibility, Betti numbers, and Tanimoto similarity. Section 3 reports an experiment on 800 ZINC molecules, comparing top-scoring molecules against JT-VAE, GCPN, ORGAN, and MolDQN variants in terms of penalized logP and QED. The abstract claims that GraphTRL 'outperforms existing methods in binding affinity prediction,' but no binding-affinity metric or benchmark appears in the paper.","tokens_in":1465,"tokens_out":1552,"duration_ms":41860,"significance":"The combination of topological/graph features with RL for molecular generation is plausible and builds on established feature-engineering work in binding-affinity prediction. However, the paper does not substantiate the central claim, and the optimization gains are marginal (0.05 in penalized logP and 0.003 in QED over MolDQN-bootstrap) with no statistical analysis. The manuscript is too incomplete to establish a nontrivial advance, even if the underlying idea has merit.","major_comments":[{"comment":"The abstract states that GraphTRL 'outperforms existing methods in binding affinity prediction,' yet the only experimental evaluation, Table 1, reports penalized logP and QED. No binding-affinity experiment appears anywhere: there is no docking score, no PDBbind or other affinity benchmark, no IC50/Kd prediction, and no comparison to EISA-Score [6] or PDFL-ML [7], which are cited as affinity-prediction methods. Moreover, neither reward function in Section 2.5 includes an affinity term. Thus the paper's central claim is completely unaddressed by its experiments; even a perfectly controlled replication of Table 1 would not support the abstract's binding-affinity assertion.","section":"Abstract; Section 3, Table 1"},{"comment":"The comparison to baselines is not established as a controlled experiment. The text says the setup is 'inspired by the experimental setup of [13] and [5]' but does not state that JT-VAE, GCPN, ORGAN, or the MolDQN variants were re-run under the same action space, reward coefficients (lambda, delta, epsilon), molecule subsets, or random seeds. The reported advantages of GraphTRL (11.89 vs. 11.84 in penalized logP; 0.951 vs. 0.948 in QED) are within typical run-to-run variability for such methods, and no standard deviations, confidence intervals, or number of seeds are provided. Without this information, the table cannot be interpreted as evidence of superiority.","section":"Section 3, Table 1"},{"comment":"The manuscript omits many implementation details needed to reproduce or assess the method. The MWCG parameters (kernel width eta_kk', distance cutoff c, exponent kappa), the persistent-image grid and resolution, the dueling DQN architecture, learning rate, exploration schedule, and the reward coefficients lambda, delta, epsilon, and w are never specified. Since the central claim is empirical, the absence of these details makes the reported results non-reproducible and leaves open the possibility that the performance depends sensitively on choices not described in the paper.","section":"Section 2, Eqs. (4), (5), (7); Section 3"}],"minor_comments":[{"comment":"The text uses 'SMILE' instead of 'SMILES' in Sections 2.3 and 3; please correct throughout.","section":"Throughout"},{"comment":"The exponent is written with both kappa and k: the definition 'kappa > 0' is inconsistent with the formula using k. Please clarify the notation and define kappa (or k) explicitly.","section":"Eq. (4)"},{"comment":"The term B(m, m0) is used to represent topological complexity or a Betti-number difference, but its precise definition is not given. Please define how the Betti number of a molecule is computed and how the difference between m and m0 is measured.","section":"Section 2.5.1, Eq. (7)"},{"comment":"The acronym QED is not defined in the text; please provide the full term (quantitative estimate of drug-likeness) and a citation on first use.","section":"Table 1"},{"comment":"For MolDQN-twosteps, the penalized logP entries are listed as '-'. Please state explicitly why these values are missing (e.g., not reported in the original reference) rather than leaving the reader to infer it.","section":"Table 1"},{"comment":"The phrase 'duel deep neural network' should be 'dueling deep neural network'.","section":"Section 2.4"},{"comment":"The caption says the figure shows a flowchart of state construction, but the figure is not described in the text. Please add a short explanation of the components and how they are concatenated.","section":"Figure 4"},{"comment":"The phrase 'the model was run on each molecule for a single episode' is ambiguous. Please clarify whether one episode means one optimization trajectory of a fixed maximum number of steps and, if so, what that maximum is.","section":"Section 3"}],"recommendation":"reject","confidential_remarks":"The manuscript is a very short workshop-style paper that, in its current form, does not meet the standard for a journal publication. The most serious issue is the mismatch between the abstract's binding-affinity claim and the actual content of the experiments; this is not a matter of degree of support but a complete absence of the claimed evaluation. The baseline comparability and missing-details issues are also severe, though they are secondary to the central unsupported claim. I would encourage the authors to either substantially revise the claims to match the presented results or add a dedicated binding-affinity evaluation, and to provide full implementation details and statistical rigor before resubmitting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on GraphTRL.\n\nThe genuinely useful part is small but real: the author uses SYBYL atom types inside the MWCG state so that the RL agent can literally see when an action changes an atom's hybridization state. That's a sensible way to incorporate structural awareness into a MolDQN-style optimizer, and I haven't seen that exact combination before. The state-construction pipeline is described clearly, and the choice of dueling DQN is reasonable.\n\nWhat's new is an extension, not a new paradigm. The MDP, actions, and reward structure are borrowed from MolDQN and JT-VAE, and the descriptors (MWCG, persistent images, fingerprints) all come from prior work. That's fine, but it means the contribution should be framed as 'does this state representation help?' rather than 'a new framework.'\n\nThe problems start in the abstract. It claims GraphTRL 'outperforms existing methods in binding affinity prediction,' but there is no binding-affinity experiment anywhere in the paper. No docking scores, no PDBbind, no IC50, no comparison to the affinity predictors cited in [6,7]. The experiment section reports only penalized logP and QED. That gap isn't a missing robustness check; it's the central claim left completely untested. Even if Table 1 were perfectly controlled, it would say nothing about binding affinity.\n\nThe Table 1 comparison is also shaky on its own terms. The baselines are reported as fixed numbers, and the paper doesn't state whether they were re-run under identical action spaces, reward coefficients, similarity thresholds, or seeds. With no standard deviations, a 0.05 gain in penalized logP and 0.003 in QED can't be distinguished from noise. And there's no code or hyperparameter list, so the empirical claim isn't reproducible.\n\nOn the plus side, there's no circularity problem: the benchmarks are external, and the results are not derived from fitted parameters. The method section is readable and the idea of tracking atom-type changes is actually clever.\n\nWho should read it? Someone working on descriptor engineering for RL-based molecular design might find the state representation worth a look. But as submitted, the paper does not support its headline claim. I would desk reject it in its current form, and tell the authors to either add a real binding-affinity evaluation or rewrite the abstract to match the optimization results. If they do that, it could become a modest workshop paper.\n\nRecommendation: don't send to peer review as is. Ask for major revision first.","headline":"The paper's one good idea—tracking atom-type changes via SYBYL types—is buried under an untested binding-affinity claim and thin baselines that don't justify the conclusion.","tokens_in":5527,"tokens_out":3982,"would_cite":false,"duration_ms":37432,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GraphTRL claims that a graph-and-topology state representation improves reinforcement learning for drug-like molecule design.","keywords":["reinforcement learning","molecular optimization","persistent homology","multiscale weighted colored graphs","drug discovery","QED","penalized logP"],"falsifier":"Re-running MolDQN-bootstrap and GraphTRL from the same 800 ZINC molecules with identical reward coefficients $\\lambda,\\delta,\\varepsilon$, identical action sets, and identical seeds would settle the performance claim: if the gap between 11.89 and 11.84 penalized logP (or 0.951 and 0.948 QED) is within run-to-run standard deviation, the claimed improvement is falsified. For the abstract's binding-affinity claim, docking or measuring the affinity of GraphTRL-generated versus baseline-generated molecules against a fixed protein target would provide a direct test.","tokens_in":4499,"feed_emoji":"🧪","tokens_out":8801,"duration_ms":116230,"temperature":0.7,"pith_summary":"This paper proposes GraphTRL, a reinforcement-learning method that gives the agent a structural view of a molecule by concatenating three state components: multiscale weighted colored graphs (MWCG), which encode atom-type-specific interactions by distance; persistent images, which summarize the molecule's shape across scales; and molecular fingerprints. The claim is that this graph-and-topology state makes a dueling deep Q-network better at finding drug-like molecules than earlier methods—MolDQN, JT-VAE, GCPN, and ORGAN—while keeping every generated molecule chemically valid. On the reported benchmark, which starts from 800 ZINC molecules, GraphTRL reaches penalized logP 11.89 and QED 0.951, slightly above the best prior scores of 11.84 and 0.948. The abstract goes further, claiming that the structural state also improves binding-affinity prediction; the experiments in the paper, however, measure drug-likeness rather than affinity directly.","feed_headline":"GraphTRL nudges past MolDQN on drug-molecule scores","feed_subtitle":"Adding multiscale graph and persistent-homology features to the RL state lifts penalized logP to 11.89 and QED to 0.951.","key_machinery":"The load-bearing machinery is the state representation, not the policy. MWCG builds element-specific subgraphs whose edges are weighted by a generalized exponential radial basis function $\\Phi_E(\\|r_i-r_j\\|;\\eta_{kk'}) = e^{-(\\|r_i-r_j\\|/\\eta_{kk'})^\\kappa}$, with SYBYL atom types as vertex colors; persistent homology converts the molecule's multi-scale connective structure into persistent images, a vector-valued density representation. These are concatenated with molecular fingerprints to form the MDP state, and a dueling deep Q-network estimates $Q(s_i,a_i)=V(s_i)+A(s_i,a_i)$, separating state value from action advantage. The reward functions are what make the method usable: a constrained reward $R_1$ optimizes penalized logP while penalizing violation of a Tanimoto similarity floor $\\delta$ and a Betti-number floor $\\varepsilon$, and an alternative reward $R_2$ balances a target Betti number against similarity to the starting molecule.","core_discovery":"The central discovery, stated on the paper's own terms, is that persistent homology and multiscale weighted colored graphs can be folded directly into a reinforcement-learning state so that the agent's decisions reflect both local atom environments and global molecular topology. GraphTRL represents each molecule as the concatenation of MWCG features built from SYBYL atom types and a generalized-exponential distance kernel, a persistent-image vector from the molecule's filtration, and molecular fingerprints. The MDP acts by atom addition, bond addition, and bond removal, and the reward function combines penalized logP with penalties for violating a Tanimoto-similarity threshold $\\delta$ and a Betti-number threshold $\\varepsilon$, so the agent is pushed to optimize logP while preserving the scaffold. The experiments report top scores of 11.89 penalized logP and 0.951 QED with 100% validity, the best in the comparison table. The paper's abstract additionally claims better binding-affinity prediction, an outcome not directly measured in the presented experiments.","pith_inferences":["The abstract's binding-affinity claim is not tested in the experiments, which optimize penalized logP and QED; a direct follow-up would be to train a supervised affinity predictor on the same MWCG and persistent-image features and then score GraphTRL-generated molecules against a protein–ligand benchmark.","Table 1 reports top-three scores rather than distributions over runs, so a meaningful next step is to report means and standard deviations across seeds; if the 0.05 penalized logP and 0.003 QED gaps over MolDQN-bootstrap fall inside run-to-run noise, the practical advantage would shrink.","The MWCG kernel width $\\eta_{kk'}$, exponent $\\kappa$, and distance cutoff $c$ are tunable per target; this suggests a natural extension in which the state representation is specialized to a binding site's geometry, which could make the structural reward more informative than generic drug-likeness."],"forward_implications":["Because the constrained reward $R_1$ can be swapped for the target reward $R_2$, the same trained pipeline addresses both lead optimization (high penalized logP, preserved scaffold) and targeted generation (prescribed Betti number or molecular weight).","SYBYL atom-type coloring makes the state sensitive to bond-order changes, so a valid action such as converting a $C.1$ to a $C.2$ atom changes the features; the agent can learn the chemical consequences of its own edits.","The 100% chemical validity reported for GraphTRL means the atom/bond action space plus structural state avoids the invalid-molecule failures of sequence-based generators, lowering the cost of exploration.","On the reported numbers, the added topological features do not sacrifice drug-likeness: GraphTRL's QED and penalized logP exceed those of all compared baselines, including the previous best MolDQN-bootstrap."],"supporting_citations":[{"why":"Defines the chemically valid Markov decision process for atom/bond additions and removals that GraphTRL builds on, and supplies the MolDQN baseline it must beat.","marker":"[5]"},{"why":"Supplies the extended atom-type features and geometric graph learning ideas behind the MWCG edge weighting by nearest-neighbor distance.","marker":"[9]"},{"why":"Provides SYBYL atom types used to color MWCG vertices and to detect bond-order changes from valid actions.","marker":"[10]"},{"why":"Motivates the generalized exponential radial basis function used for multiscale weighted colored-graph edges.","marker":"[11]"},{"why":"Supplies the dueling deep Q-network architecture that separates state value from action advantage in GraphTRL's policy.","marker":"[12]"},{"why":"Defines penalized logP as the optimization target and provides the JT-VAE baseline and experimental setup that the comparisons follow.","marker":"[13]"},{"why":"Supplies the 800 ZINC molecules used as starting points for optimization in the experiments.","marker":"[15]"}],"fun_headline_variants":["GraphTRL: topology-infused RL tops drug benchmarks","Persistent homology folds into RL state for better molecules","Topology-aware RL achieves 11.89 logP, 0.951 QED","GraphTRL beats baseline with multiscale graph and homology","RL gets topological edge: GraphTRL scores 11.89 logP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance comparison assumes the Table 1 baseline scores were produced under the same protocol as GraphTRL—same action space, reward coefficients, starting molecules, and random seeds—but the paper reports them as fixed numbers and does not state that all methods were re-run under identical conditions.","fun_headline_variants_meta":{"raw":{"variants":["GraphTRL: topology-infused RL tops drug benchmarks","Persistent homology folds into RL state for better molecules","Topology-aware RL achieves 11.89 logP, 0.951 QED","GraphTRL beats baseline with multiscale graph and homology","RL gets topological edge: GraphTRL scores 11.89 logP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3355,"prompt_tokens":857,"completion_tokens":2498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2406}},"tokens_in":473,"tokens_out":2498,"duration_ms":18027,"temperature":1.0,"reasoning_tokens":2406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:58:03.843586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running MolDQN-bootstrap and GraphTRL from the same 800 ZINC molecules with identical reward coefficients $\\lambda,\\delta,\\varepsilon$, identical action sets, and identical seeds would settle the performance claim: if the gap between 11.89 and 11.84 penalized logP (or 0.951 and 0.948 QED) is within run-to-run standard deviation, the claimed improvement is falsified. For the abstract's binding-affinity claim, docking or measuring the affinity of GraphTRL-generated versus baseline-generated molecules against a fixed protein target would provide a direct test.","supporting_citations":[{"cited_title":"Optimization of molecules via deep reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Defines the chemically valid Markov decision process for atom/bond additions and removals that GraphTRL builds on, and supplies the MolDQN baseline it must beat."},{"cited_title":"Geometric graph learning with extended atom-types features for protein-ligand binding affinity prediction","cited_arxiv_id":null,"evidence_quote":"Supplies the extended atom-type features and geometric graph learning ideas behind the MWCG edge weighting by nearest-neighbor distance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides SYBYL atom types used to color MWCG vertices and to detect bond-order changes from valid actions."},{"cited_title":"Fast and anisotropic flexibility-rigidity index for protein flexibility and fluctuation analysis","cited_arxiv_id":null,"evidence_quote":"Motivates the generalized exponential radial basis function used for multiscale weighted colored-graph edges."},{"cited_title":"A value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis","cited_arxiv_id":null,"evidence_quote":"Supplies the dueling deep Q-network architecture that separates state value from action advantage in GraphTRL's policy."},{"cited_title":"Junction tree variational autoencoder for molecular graph generation","cited_arxiv_id":null,"evidence_quote":"Defines penalized logP as the optimization target and provides the JT-VAE baseline and experimental setup that the comparisons follow."},{"cited_title":"Zinc: a free tool to discover chemistry for biology","cited_arxiv_id":null,"evidence_quote":"Supplies the 800 ZINC molecules used as starting points for optimization in the experiments."}],"review_version":1}