{"id":"400fc595-217c-4ac4-85fe-9fdc99f39b93","arxiv_id":"2602.04119","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"S3-GFN generates synthesizable SMILES molecules with over 95% positivity and competitive rewards by adding a contrastive replay-buffer loss to GFlowNet post-training of a pretrained chemical language model.","lead":"This paper introduces S3-GFN, a method that trains a SMILES-based generative model with a soft penalty so most molecules it invents are judged synthesizable by retrosynthesis tools. The interest is that it claims to match or beat reaction-based generators while keeping the flexibility of simple string generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthesizability claim is anchored to a single retrosynthesis oracle used for both training labels and the primary metric; for SBDD the independent AiZynthFinder check is reported only on a filtered Top-100 subset, so the ≥95% raw-output claim is not yet established.","rationale":"The reader's weakest assumption identifies the retrosynthesis oracle as the linchpin of the synthesizability claim, and I agree. My pass sharpens this concern: the paper's strongest independent evidence (AiZynthFinder) is not reported on the raw SBDD generation distribution, only on a filtered positive subset, so the most general version of the claim — 'S3-GFN learns to generate synthesizable molecules ≥95%' — is not yet fully supported on the SBDD benchmark. The sEH results are more convincing because AiZynthFinder is evaluated on random samples and reaches 0.99, but the Positive Ratio there is 0.945, slightly below the abstract's ≥95%. I do not see an internal inconsistency in the GFlowNet training: the RTB objective on positive samples plus the contrastive auxiliary loss is a reasonable soft-constraint mechanism, and the off-policy replay design is coherent. The main risk is empirical: label noise and selective evaluation can inflate synthesizability. The paper honestly discloses that validation is in silico, and the code is provided; still, a raw-distribution synthesizability measurement on SBDD is a concrete, low-cost check that would settle whether the headline claim extends beyond the filtered subset. Given this gap, the reader's CONDITIONAL verdict remains appropriate: the method is promising and mostly well-supported, but the synthesizability claim should be verified on unfiltered outputs before full acceptance.","tokens_in":21649,"tokens_out":8480,"duration_ms":99795,"concrete_test":"For each of the five LIT-PCBA targets, compute and report the raw Positive Ratio (training oracle) and AiZynthFinder success rate on all 64,000 generated molecules before Top-100, QED, and Tanimoto filtering, with means and standard deviations over the three seeds. If raw AiZynthFinder success is ≥95% on all five targets, the synthesizability claim for SBDD is supported; if it is materially lower, the headline should be restricted to the selected subset or revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that S3-GFN generates ≥95% synthesizable molecules depends on how synthesizability is defined and measured. The training labels in Algorithm 1 come from a heuristic retrosynthesis search over 105 reaction templates and the Enamine Stock library (maximum 3 steps), and the primary 'Positive Ratio' metric uses exactly the same oracle. Thus the 0.945 Positive Ratio in Table 1 is partly enforced by construction, not independently verified. The paper does report AiZynthFinder as an external check, and on sEH the raw random-sample AiZynthFinder success is 0.990 (Table 6), which substantially mitigates the circularity for that task. However, for the SBDD benchmark (Table 2), AiZynthFinder success is reported only for the post-hoc selected Top-100 positive candidates, after negative molecules have been explicitly excluded (Appendix B.3). The raw synthesizability of the 64,000-molecule generation distribution is never reported for any LIT-PCBA target. The abstract's '≥95% synthesizable' is therefore only directly supported for one task, and even there the measured Positive Ratio is 0.945—below the stated threshold. If the retrosynthesis oracle is permissive or incomplete relative to true synthetic accessibility, both the training signal and the headline metric inherit the same systematic error; the high AiZynthFinder agreement on sEH reduces but does not eliminate this risk, since AiZynthFinder is itself a computational retrosynthesis proxy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes S3-GFN, a sequence-based SMILES GFlowNet that induces synthesizability as a soft distributional constraint. Starting from the pretrained GP-MolFormer prior, the method applies relative trajectory balance (RTB) only to molecules labeled as synthesizable by a retrosynthesis heuristic, and adds a contrastive auxiliary loss over replay buffers of positive and negative samples to suppress unsynthesizable regions. The authors report that S3-GFN achieves over 95% synthesizability, outperforms reaction-based GFlowNets on sEH proxy optimization, structure-based docking benchmarks on five LIT-PCBA targets, and sample-limited PMO tasks, and enables fast adaptation to changed synthesis constraints. The paper includes a 2D grid-world proof of concept, extensive ablations, and released code.","tokens_in":21970,"tokens_out":5348,"duration_ms":52391,"significance":"If the central claim holds, S3-GFN offers a flexible and scalable alternative to reaction-based GFlowNet MDPs by leveraging pretrained SMILES language models, with the ability to adapt to evolving synthesizability constraints without redesigning the action space. The paper's strengths include the broad experimental coverage, the inclusion of an external AiZynthFinder check on the sEH task, the controlled reward-shaping baseline, the off-policy genetic-exploration extension, and the public code release. The main weakness is that the headline synthesizability claim rests on a metric that is partly defined by the same oracle that provides training labels, and raw-output synthesizability is not reported for several benchmark settings. These issues are addressable with additional evaluation and clarification.","major_comments":[{"comment":"The abstract's '>=95% synthesizable' claim is not established for the structure-based drug design task. AiZynthFinder success in Table 2 is reported only for the diverse Top-100 candidates after 'negative samples' have been explicitly excluded (Appendix B.3). The synthesizability of the full 64,000-molecule generation distribution is never reported for any LIT-PCBA target. Because the contrastive training explicitly penalizes molecules that the oracle labels negative, the high Top-100 AiZynthFinder rates do not imply that the raw sampler output is >=95% synthesizable. Please report Positive Ratio and/or AiZynthFinder success on the full generated set (before filtering) for all five targets.","section":"§5.1.2, Table 2, Appendix B.3"},{"comment":"There is an inconsistency in the metric supporting the '>=95%' claim. On sEH, the template-based Positive Ratio is 0.945±0.009 (Table 1), below 95%, while the external AiZynthFinder success on random samples is 0.990±0.008 (Table 6). If AiZynthFinder is the intended operational definition of synthesizability, this should be stated explicitly and reported consistently across all tasks and tables. If Positive Ratio is the intended definition, the headline claim should be revised or the threshold clarified.","section":"§5.1.1, Table 1 vs. Abstract"},{"comment":"The training signal and the primary 'Positive Ratio' metric both come from the same heuristic retrosynthesis oracle (105 templates + Enamine Stock, max 3 steps). The model is trained to imitate that oracle's labels, so the Positive Ratio is partly enforced by construction and does not independently validate synthesizability. The independent AiZynthFinder result on sEH (0.990 raw) substantially mitigates this concern for that task; extending the same raw-output external evaluation to the SBDD and PMO benchmarks would address the concern for the paper's general claim.","section":"§3, Problem definition and Algorithm 1"},{"comment":"The baseline numbers for SynFlowNet, RGFN, and RxnFlow are taken from Seo et al. (2025), while S3-GFN is evaluated with the protocol described in Appendix B.3, including Uni-Dock and exclusion of negative samples before Top-100 selection. Please confirm that the docking software, receptor preparation, selection criteria, and exclusion policy are identical for all methods, or explicitly state any differences. If the baseline numbers were produced under a different evaluation pipeline, the performance comparison (especially AiZynthFinder success rates) may be confounded.","section":"Table 2"}],"minor_comments":[{"comment":"Several reported means have zero standard deviation across three runs (e.g., S3-GFN Pos. Top100 sEH 1.009±0.000, Diversity 0.764±0.000). This is surprising for independent runs and should be explained. If it is due to reward saturation and deterministic Top-100 selection, a sentence clarifying this would help.","section":"Table 1 / Table 6"},{"comment":"If the negative replay buffer B− is empty in early training, the contrastive loss Laux in Eq. (4) is undefined. Please state the initialization or fallback behavior (e.g., skip auxiliary loss until B− has samples).","section":"Eq. (4) and Algorithm 1"},{"comment":"The sentence 'We omit VE' appears incomplete. It likely refers to omitting a specific oracle from the PMO benchmark; please complete or rephrase.","section":"Appendix B.5"},{"comment":"There is a typo: 'Lipinski et al. (2012))' has a double closing parenthesis. Also, the curated reaction list would benefit from a table or SMARTS list in the appendix.","section":"Appendix B.4"},{"comment":"The caption states 'SFN guarantees 100% validity' but the figure compares SynFlowNet; please use the full method name consistently.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The method is plausible and the experiments are broad, but the headline synthesizability claim needs stronger raw-output validation and the baseline comparison needs explicit protocol harmonization. I do not see grounds for rejection; the requested changes are within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading and worth refereeing. The core idea — keep a sequence-based SMILES GFlowNet, warm-start from a pretrained SMILES language model, and enforce synthesizability through contrastive replay rather than a reaction-based MDP — is genuinely useful. It is not a radical departure: it builds on RTB post-training (Venkatraman et al.; Pandey et al.) and off-policy replay (Kim et al.), but the combination evaluated against SynFlowNet and RxnFlow fills a real gap. It shows that you do not need hard reaction constraints to get high synthesizability and better task rewards.\n\nThe paper does several things properly. It includes a sanity-check grid world, ablates a reward-shaping baseline with the same prior and replay setup, reports diversity, SA score, QED, and molecular weight, and explicitly discloses that all validation is in silico. The fast-realignment experiment under changed reaction sets and drug-likeness constraints is a nice argument for soft constraints. Code is promised, and the description of the method is precise enough to reimplement.\n\nThe stress-test concern is mostly on target, though not fatal. The positive/negative labels come from the same retrosynthesis heuristic (105 templates, Enamine Stock, max 3 steps) that defines the primary Positive Ratio metric. If that oracle is permissive or incomplete, both training and the headline metric share the same error. What partly rescues it: on sEH, AiZynthFinder success on random samples is 0.990 (Table 6), an independent proxy, so the method really is biasing generation toward retrosynthetic accessibility on that task. But for the SBDD benchmark, AiZynthFinder is only reported on the post-hoc selected Top-100 positives after negatives are excluded; the raw 64,000-sample synthesizability is never reported. So the abstract's \"≥95% synthesizable\" is directly supported only by sEH, and even there Table 1 says 0.945. That is a verbal overclaim, easily fixed with more careful wording and by reporting raw AiZynthFinder for every task. Also, some baseline numbers in Table 2 are lifted from Seo et al. with different generation budgets; rerunning or clearly stating the protocol would strengthen the comparison.\n\nWho is this for: anyone working on synthesizable molecular generation or GFlowNet post-training; the contrastive replay separation is a reusable mechanism beyond chemistry. It deserves a serious referee. My recommendation: engage with it, do not desk-reject, and ask for distribution-level synthesizability metrics on all SBDD targets and aligned baselines.","headline":"Soft-constraint GFlowNet post-training with contrastive replay is a credible, reusable alternative to reaction-based synthesizable generation, but the synthesizability claims need a careful referee to check oracle-dependence and evaluation filtering.","tokens_in":22561,"tokens_out":2029,"would_cite":true,"duration_ms":21571,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"S3-GFN shows that synthesizability in molecular generation can be learned as a soft, distributional constraint—via contrastive replay of positive and negative SMILES trajectories—rather than hard-coded into reaction-based state spaces, achi","keywords":["GFlowNets","molecular generation","SMILES","synthesizability","contrastive learning","replay buffers","relative trajectory balance","drug discovery"],"falsifier":"Take the molecules the model labels as positive and pass them through an independent retrosynthesis planner built on a different reaction set, or attempt actual laboratory synthesis of a sample; if the positive ratio or external synthesis-success rate falls well below the claimed 95%, the central claim that the soft constraint induces synthesizability is falsified.","tokens_in":21483,"feed_emoji":"🧪","tokens_out":5060,"duration_ms":47721,"temperature":0.7,"pith_summary":"The paper claims that synthesizability does not have to be hard-wired into a reaction-by-reaction construction process. It trains a sequence-based GFlowNet that generates SMILES strings (a text notation for molecular structure), starting from a pretrained chemical language model, and uses two complementary signals: a trajectory-balance objective applied only to molecules flagged as synthesizable by a retrosynthesis oracle, and a contrastive auxiliary loss that pushes down the model's likelihood on unsynthesizable trajectories sampled from replay buffers. The result, in the paper's experiments, is a sampler whose output is at least 95% synthesizable and that achieves higher rewards on the tested molecular design tasks than reaction-based GFlowNets. A sympathetic reader would take this as evidence that rich pretrained chemical priors plus soft distributional regularization can replace hard reaction constraints for practical de novo design.","feed_headline":"Soft constraint pushes GFlowNet molecules past 95% synthesizable","feed_subtitle":"A sequence-based GFlowNet with contrastive replay beats reaction-based generation on reward and synthesizability.","key_machinery":"The load-bearing pieces are the relative trajectory balance (RTB) objective, which reweights a pretrained prior toward the reward-scaled posterior, and the contrastive replay loss, which separates positive (synthesizable) and negative (unsynthesizable) trajectories by comparing their sequence-level log-likelihoods. Two replay buffers supply the contrast: D+ stores synthesizable trajectories with reward-prioritized sampling, D− stores unsynthesizable ones uniformly, and local mutations of positives generate additional informative negatives. The contrastive term is applied only during replay, so the reward objective and constraint objective are optimized separately rather than entangled as in","core_discovery":"S3-GFN establishes that synthesizability can be induced at the level of the sampling distribution rather than the state space. Starting from a pretrained SMILES prior, the model is post-trained so that posterior probability is proportional to reward times prior inside the synthesizable set; on-policy updates use relative trajectory balance on positive samples only, while replay updates combine RTB on positives with a contrastive loss that suppresses negative trajectories. The auxiliary loss is what prevents probability mass from leaking into unsynthesizable regions, and the off-policy replay buffers let the model incorporate external search operators such as mutations and genetic search. Acr","pith_inferences":["The same replay-and-contrast recipe could be applied to other soft constraints (toxicity, ADMET, novelty filters) by changing only the oracle that labels the buffers, not the MDP or the loss.","If the pretrained prior already carries a strong implicit bias toward synthesizable chemistry, then the contrastive term may be acting mainly at the boundary; a testable prediction is that the required auxiliary coefficient shrinks as the prior improves.","The reported synthesizability numbers inherit the oracle's definition; a stricter or independent route-finder could lower them, so the headline '≥95%' should be read as '≥95% according to the retrosynthesis heuristic used for labeling'.","Because negative samples are generated by mutating positives, the model may learn to avoid local neighborhoods of synthesizable molecules; whether this shrinks accessible chemical space relative to reaction-based methods is worth measuring explicitly."],"forward_implications":["A sequence-based GFlowNet can reach or exceed the synthesizability and reward of reaction-based GFlowNets, without needing a combinatorial reaction action space.","The contrastive replay loss suppresses unsynthesizable probability mass while preserving shared substructures, avoiding the over-suppression seen with reward shaping.","When synthesizability constraints change (e.g., a curated reaction set plus drug-likeness filters), the model can realign in about 100 replay-only steps, keeping most of its reward and diversity.","Off-policy replay makes it straightforward to inject externally discovered high-reward positives (via genetic search) and locally mutated negatives, improving sample-limited optimization.","Generated molecules receive high success rates under an independent retrosynthesis planning metric—higher than reaction-based baselines—indicating the soft constraint transfers beyond the training reaction set."],"fun_headline_variants":["Soft-constrained GFlowNet attains 95% synthesizable molecules","GFlowNet soft constraint yields 95% synthesizable compounds","S3-GFN: rich priors plus soft constraint hit 95% synthesizability","Contrastive replay buffers push GFlowNet to 95% synthesizable","Sequence-based GFlowNet with rich priors achieves 95% synthesis"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire training signal—which molecules count as synthesizable—comes from an automated retrosynthesis search procedure; if that oracle is permissive or incomplete, the claimed synthesizability of the model's output inherits its errors.","fun_headline_variants_meta":{"raw":{"variants":["Soft-constrained GFlowNet attains 95% synthesizable molecules","GFlowNet soft constraint yields 95% synthesizable compounds","S3-GFN: rich priors plus soft constraint hit 95% synthesizability","Contrastive replay buffers push GFlowNet to 95% synthesizable","Sequence-based GFlowNet with rich priors achieves 95% synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000921,"raw_usage":{"total_tokens":3754,"prompt_tokens":681,"completion_tokens":3073,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":2985}},"tokens_in":425,"tokens_out":3073,"duration_ms":20537,"temperature":1.0,"reasoning_tokens":2985,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:42:22.801969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the molecules the model labels as positive and pass them through an independent retrosynthesis planner built on a different reaction set, or attempt actual laboratory synthesis of a sample; if the positive ratio or external synthesis-success rate falls well below the claimed 95%, the central claim that the soft constraint induces synthesizability is falsified.","supporting_citations":[],"review_version":1}