{"id":"b6fb37ac-b261-4d92-9ed1-d752074abcf1","arxiv_id":"2608.11483","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SABLE is a modular, open-source agentic framework that combines LLM orchestration, reaction-constrained analog enumeration, Bayesian optimization, and computational affinity and ADMET scoring to prioritize synthesis-feasible hit-to-lead candidates.","lead":"SABLE is an open-source system that uses a language model to coordinate chemistry tools that propose and score new drug-like molecules. It automatically generates synthesis-feasible analogs of a starting compound and ranks them with Bayesian optimization, testing only a fraction of the possible candidates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central enrichment claim lacks a random/background control; BO-guided gains over the seed may reflect HEALER's analog-library composition rather than SABLE's selection.","rationale":"The reader's CONDITIONAL verdict is reasonable. I agree that the unvalidated Boltz-2 oracle is a serious limitation for any claim of better drug candidates, and the authors are appropriately cautious about it. My stress-test pass, however, locates the most load-bearing gap one step earlier: the stated claim is about computational enrichment, and that claim is not yet distinguished from the trivial effect of sampling a library of close analogs. Without a random baseline, the reported figures show that SABLE can find good candidates, not that its selection strategy enriches over chance. This concern is concrete and testable computationally, and it does not require experimental data. I would keep the verdict CONDITIONAL: the framework is well engineered and openly released, with code, Dockerfiles, prompts, and state dumps supporting reproducibility, but the evaluation should add random and uncertainty-agnostic baselines, and ideally repeated seeds, before the enrichment language is taken at face value.","tokens_in":16092,"tokens_out":9083,"duration_ms":85495,"concrete_test":"Re-run the CAMKK2 single-objective campaign (Section 2.2) and the three-objective campaign (Section 2.4) with the same HEALER library and the same characterization budget, but replace the qLogEI/qLogNEHVI acquisition with uniform random sampling, repeated 5-10 times; compare best-observed Boltz-2 logIC50 values and Pareto hypervolumes. If random sampling attains equal or better values, the enrichment claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that SABLE enriches candidate sets for user-defined computational objectives while evaluating only a subset of the enumerated space (Abstract; Discussion). The only quantitative evidence offered for enrichment is improvement of the best-observed scores relative to the seed compound (Figs. 2A, 3A-C, S3) and, in the QED benchmark, recovery of a known global maximum (Fig. S4). No experiment compares SABLE's selected batches to random selection, to uncertainty-agnostic acquisition, or to the score distribution of the full HEALER library. This matters because HEALER is explicitly designed to generate close, synthetically feasible analogs of an active seed (Section 5.3). In such a library, the unconditional chance that a randomly drawn analog improves a predicted affinity or QED score can be high; observing that a BO-guided subset improves on the seed is therefore not, by itself, evidence of enrichment by the agent. The multi-objective campaigns (Sections 2.3-2.4) report Pareto-frontier coverage but provide no random-sampling hypervolume baseline, so the same ambiguity applies. The oracle-accuracy caveat acknowledged by the authors (Section 2.7, Discussion) is important for external validity, but it is secondary: if the selection mechanism cannot be shown to beat random library sampling, the computational enrichment claim is underdetermined even before asking whether the oracle is trustworthy. The released code and state dumps are a real strength, but they do not substitute for the missing control.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SABLE, an open-source agentic framework for synthetically constrained hit-to-lead optimization. An LLM orchestrates a workflow in which HEALER enumerates reaction-templated analogs from Enamine building blocks, RDKit/STOPLIGHT compute physicochemical and ADMET properties, Boltz-2 predicts binding affinity, and Bayesian optimization (qLogEI or qLogNEHVI) selects batches for evaluation. The authors report single-, dual-, and multi-objective campaigns showing improvements over seed compounds in predicted logIC50, QED, and CNS activity; a QED benchmark on one-million-compound libraries; and retrospective case studies on METTL3 and four Boltz-2 validation targets. The paper claims that SABLE enriches candidate sets for user-defined computational objectives while evaluating only a subset of the enumerated space, and that all proposed molecules are constructible from commercially available building blocks under known reaction templates.","tokens_in":16266,"tokens_out":4976,"duration_ms":45641,"significance":"If the enrichment and efficiency claims survive proper baseline testing, SABLE would be a valuable open-source contribution: it cleanly separates LLM-based intent parsing from quantitative model computation, restricts the search space to synthetically tractable analogs, provides provenance of each numerical output, and ships with Dockerfiles, Apptainer definitions, prompts, and JSON state dumps for reproduction. The modular BYOT architecture is a genuine strength and distinguishes SABLE from monolithic LLM-as-scientist systems. However, the current empirical evidence is too weak to establish the central 'enrichment' claim: there are no random-sampling or alternative-acquisition baselines, no repeated runs or error bars, and the primary affinity objective is the same Boltz-2 oracle being optimized. The paper is transparent about several limitations, but the Results sections present the oracle-score improvements as validation rather than as an internal-consistency check.","major_comments":[{"comment":"The central enrichment claim lacks any random or non-agentic baseline. In every single- and multi-objective campaign, the only quantitative evidence is the best-observed score relative to the seed compound, and in the QED benchmark it is recovery of a known global maximum. Because HEALER is explicitly designed to generate close, synthetically feasible analogs of an active seed (§5.3), the unconditional probability that a randomly drawn analog improves predicted affinity or QED may be high. The paper should compare SABLE's selected batches to random sampling of the same enumerated library, to an uncertainty-agnostic acquisition (e.g., greedy selection by surrogate mean), and, for multi-objective campaigns, to a random-sampling hypervolume baseline. Without these comparisons, the stated conclusions that SABLE 'enriches candidate sets' and 'produces useful enrichment after evaluating only a small fraction of the enumerated space' are underdetermined.","section":"§2.2, §2.5, Figs. 2–3, S3–S4"},{"comment":"Every campaign is a single unrepeated run without error bars or variance estimates. The claim that the cumulative distribution of the 20 best candidates 'plateaus' after approximately five iterations (§2.2) is a qualitative reading of one trajectory. Repeated runs with different random seeds are needed to show that the plateau is systematic and that reported improvements (e.g., the 1.03 log-unit improvement in §2.2) are not within run-to-run noise. Similarly, the Pareto-frontier claims in §2.3 and §2.4 would be substantially strengthened by reporting hypervolume or dominated-volume statistics averaged over repeated runs, with standard deviations.","section":"§2.2–§2.4, Figs. 2–3"},{"comment":"The primary demonstration is self-referential: the acquisition function maximizes the same Boltz-2 score that is later reported as the 'improvement,' so the best-observed oracle value is guaranteed to improve by construction. The paper correctly acknowledges this in the Discussion, but the Results sections present these improvements as evidence of SABLE's value. To make the value claim load-bearing, the paper should include a control in which the same enumerated library is optimized by a cheap baseline (e.g., selecting the top-k candidates by the surrogate mean without acquisition exploration, or random selection) and compare oracle-score improvement per evaluation. Without such a control, the framework's contribution reduces to implementing a standard Bayesian optimizer, which is not the paper's stated novelty.","section":"§2.2, §2.7, Discussion"}],"minor_comments":[{"comment":"The phrase 'Across single, and multi-objective optimization studies' contains a stylistic comma error; it should read 'Across single- and multi-objective optimization studies.'","section":"Abstract and §2.1"},{"comment":"There is a typo in 'SubstanceParamaterobjects'; it should be 'SubstanceParameter objects.' Also, the expansion 'qLogNoisyExpectedHypervolumeImprovement' should be written as 'qLogNEHVI' consistently, including the parenthetical.","section":"§5.4"},{"comment":"The sentence 'The relative savings in oracle or experimentation calls, grow approximately linearly with library size in the regime of〈10 3 to 106)' has broken math formatting and grammar; please rewrite to specify the regime clearly, e.g., 'library sizes from 10^3 to 10^6.'","section":"§2.5"},{"comment":"The text says all experiments were performed 'in a Dockerized environment on a single laptop,' but Boltz-2 inference was carried out on a GPU node with four NVIDIA L40 GPUs; please clarify which components ran on the laptop and which ran on the remote compute node.","section":"§5.10"},{"comment":"The t-SNE visualization lacks axis labels, a stated perplexity value, and a random seed; without these details, the visualization cannot be reproduced or compared across runs.","section":"Fig. 2B"},{"comment":"The paper reports predicted concentrations in µM and nM units (e.g., '58.88 µM' and '370 nM') as if they were quantitative IC50 estimates. Given that the authors acknowledge Boltz-2 is not calibrated in absolute terms, the P00519 row (seed predicted at 97.72 nM versus ChEMBL 2.95 µM) should be discussed explicitly as evidence that these values should not be interpreted as absolute affinity predictions, and the table caption should state this caveat.","section":"§2.7 and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The authors are commendably transparent about limitations, including the oracle-accuracy caveat and the absence of experimental validation. However, the Results section's language ('SABLE produced improvements on every tracked objective') goes beyond what the evidence supports given the lack of baselines and single unrepeated runs. The central fix is to add random-sampling and simple-acquisition baselines to the empirical evaluation; this is within the scope of a revision and would substantially increase the paper's value. I also note that the paper's framing as 'hit-to-lead optimization' may be too strong for a system whose only demonstrated output is enrichment against a computational oracle, but that is a framing issue rather than a technical flaw."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: SABLE is a genuinely useful open-source integration of HEALER, BayBE, and Boltz-2 with LLM orchestration, and the code release is serious. But the paper's central enrichment claim lacks the one control that would make it stick: comparison to random or naive selection. The optimizer may just be re-ranking a library of close analogs that are already likely to improve on the seed.\n\nWhat's actually new: the clean separation between the LLM as a translator of user intent and the quantitative tools as the decision-makers, the BYOT config interface, and the synthesis-constrained search space via HEALER are a real step beyond AutoLead, ChemCrow, and Coscientist. The state dumps, Dockerfiles, and prompts make the work reproducible, which is more than most papers in this area.\n\nThe soft spots are concentrated in the evaluation. Campaigns are single runs without error bars, and no baseline is reported. The one-million-compound QED benchmark showing recovery of the global max in three to seven iterations is suggestive, but even a random sampler will eventually find it; what matters is sample efficiency relative to a random baseline, and we don't get that. So 'enrichment' is underdetermined—it may be a property of HEALER's analog library rather than SABLE's selection. The second issue is the oracle: every affinity result is a Boltz-2 prediction, never experimentally validated. The METTL3 retrospective helps, but the best SABLE analog was not tested. To their credit, the authors explicitly acknowledge both limitations. The paper is honest about what it can and cannot claim.\n\nWho this is for: anyone building agentic workflows for drug discovery or interested in the division of labor between LLMs and classical optimization. It deserves a serious referee. A revision should add a random-sampling baseline, repeated runs, and ideally one closed-loop experimental validation. With those, the enrichment claim would actually be supported. I'd send it out for review, not desk reject.","headline":"Solid framework paper with a real gap: no random-sampling baseline, so enrichment is asserted rather than demonstrated.","tokens_in":16953,"tokens_out":2889,"would_cite":true,"duration_ms":24328,"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":"SABLE shows that an LLM-orchestrated ligand-optimization workflow can improve every tracked objective while keeping every proposed molecule synthesizable from commercial building blocks.","keywords":["hit-to-lead optimization","agentic AI","Bayesian optimization","synthetic accessibility","reaction templates","multi-objective optimization","Boltz-2","drug discovery"],"falsifier":"Synthesize a handful of the top-ranked SABLE analogs from the METTL3 and Boltz-2-validated-target campaigns and measure their IC50s in the corresponding assays; if most fail to beat the seed by a meaningful margin, the enrichment claim would be shown to be an artifact of the Boltz-2 objective rather than genuine lead improvement.","tokens_in":15803,"feed_emoji":"🧪","tokens_out":8098,"duration_ms":71608,"temperature":0.7,"pith_summary":"SABLE is a modular, agentic workflow for hit-to-lead drug optimization that combines natural-language orchestration with specialized computational tools: reaction-templated analog enumeration, property and ADMET prediction, structure-based affinity scoring, and Bayesian optimization. The paper aims to establish that this division of labor, with the language model handling intent and dedicated models handling quantitative decisions, lets a user specify an optimization campaign in plain English and receive a prioritized set of synthesizable analogs that improve on the starting compound across the user's chosen objectives. Across single- and multi-objective campaigns, SABLE improved every tracked objective relative to the seed while evaluating only a fraction of the enumerated analog space, and every proposed molecule is constructible from commercially available building blocks under known reaction templates. A sympathetic reader would care because the framework directly addresses two bottlenecks that keep generative design out of the lab: synthetic infeasibility of proposed molecules and the difficulty of balancing competing drug properties.","feed_headline":"Agentic drug optimizer improves molecules and keeps them synthesizable","feed_subtitle":"SABLE improves every tracked objective while evaluating a fraction of the analog space.","key_machinery":"The load-bearing mechanism is the separation of the language model from quantitative tools. The LLM only parses the user request into an optimization configuration (seed molecule, protein target, objectives, iteration count) and summarizes results; it never proposes analogs, scores affinity, or judges quantitative outcomes. Candidate enumeration is delegated to HEALER, a retrosynthetic reaction-templated enumerator that assembles analogs from commercially available building blocks, making synthetic feasibility a property of the search space rather than a post hoc filter. Selection is delegated to Bayesian optimization with a Gaussian-process surrogate and the qLogEI (single-objective) or qLogNEHVI (multi-objective) acquisition functions, which balances exploration and exploitation and returns batch recommendations that are validated against the search space and the experiment ledger. Characterization is delegated to purpose-built predictors: Boltz-2 for structure-based affinity scores, RDKit for physicochemical descriptors and QED, and STOPLIGHT for ADMET-related liabilities. The modular design, with tools registered through configuration files and state managed by a checkpointed workflow graph, is what makes the pipeline swappable and traceable.","core_discovery":"The paper's central claim is that a language model can orchestrate, but not dominate, scientific decision-making in molecular optimization: SABLE separates intent translation from candidate generation, quantitative characterization, and selection. Starting from a user's seed compound, HEALER builds a search space of analogs that are guaranteed to be synthesizable from purchasable building blocks through known reaction templates; a Bayesian optimization loop then selects small batches of unevaluated molecules to be scored by Boltz-2 for affinity, RDKit for physicochemical descriptors, and STOPLIGHT for ADMET-related properties. Across single-objective (affinity), dual-objective (affinity plus QED), and multi-objective (affinity, QED, and CNS activity) campaigns, the evaluated candidates improved on the seed for every tracked objective, and the best-observed candidates formed a Pareto frontier of nondominated trade-offs. In retrospective case studies on METTL3 and on four Boltz-2-validated targets, SABLE moved from experimentally weak seed compounds to analogs with substantially lower predicted concentrations, in the same range as the best compounds in the experimental series. The authors state the intended scope explicitly: SABLE enriches against the computational objectives, not against unmeasured biochemical truth, and is designed as a prioritization engine to support medicinal chemistry review and eventual experimental closed-loop deployment.","pith_inferences":["Beyond the paper: the same 'LLM as translator, specialists as judges' split could transfer to other closed-loop design tasks, such as reaction-condition optimization or materials formulation, where a language model interfaces with quantitative simulators; the paper's evidence is limited to ligand optimization, so this is an extrapolation.","Beyond the paper: because the search space is determined by HEALER's reaction templates, the ceiling of SABLE's performance is set by the enumerator's coverage; expanding the template set or adding a second enumerator would likely widen the Pareto frontier, and comparing frontier sizes across enumerators would test this.","Beyond the paper: a direct experimental follow-up would be to synthesize the top-ranked METTL3 analog (predicted 102 nM) and measure its IC50; if it lands within an order of magnitude of the prediction, that would support Boltz-2 as a viable ranking oracle for prioritization within HEALER's space."],"forward_implications":["Because every enumerated analog is assembled from purchasable building blocks via reaction templates, the top-ranked candidates can be ordered for synthesis without a separate synthetic-feasibility triage step.","Replacing the affinity oracle with a wet-lab assay or a better predictor requires only a node-level adapter, so the same orchestration logic can drive experimental closed-loop design-make-test-analyze cycles as robotic infrastructure matures.","The Pareto-frontier output gives medicinal chemists a set of trade-off alternatives rather than a single candidate, so campaign-specific priorities can be applied after optimization.","Because only a subset of the enumerated space is evaluated before convergence, the framework's computational cost scales more gently than exhaustive screening, with savings growing roughly linearly with library size in the tested regime."],"supporting_citations":[{"why":"Supplies HEALER, the reaction-templated enumerator that restricts the search space to analogs constructible from commercially available building blocks.","marker":"[11]"},{"why":"Supplies Boltz-2, the structure-based affinity oracle used for every affinity objective, and its per-assay validation data that motivates the four-target case study.","marker":"[15]"},{"why":"Provides the qLogEI acquisition function used for single-objective Bayesian optimization.","marker":"[32]"},{"why":"Provides the qLogNEHVI acquisition function used for multi-objective optimization.","marker":"[33]"},{"why":"Supplies STOPLIGHT, the ADMET and property predictor used for objectives such as CNS activity.","marker":"[14]"},{"why":"Defines the QED drug-likeness objective used in dual- and multi-objective campaigns.","marker":"[31]"},{"why":"Provides the METTL3 retrospective series, including the seed compound and published UZH2, used to check SABLE against a real medicinal-chemistry trajectory.","marker":"[37]"},{"why":"Supplies RDKit, the cheminformatics toolkit computing descriptors and QED values used as objectives and molecular encodings.","marker":"[21]"}],"fun_headline_variants":["Agentic framework SABLE optimizes multi-objective drug leads, synthesizable","LLM orchestrates multi-objective drug optimization with synthesis constraints","SABLE: agentic ligand explorer for synthesizable hit-to-lead optimization","Drug optimizer uses LLM to guide Pareto-front analog exploration","Modular agentic system improves drug analogs, ensures synthesizability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire affinity optimization signal rests on the assumption that Boltz-2's predicted scores correctly rank true binding affinity within each chemical series, and the paper never tests that assumption against experimental measurements, so if that ranking is wrong, the reported affinity improvements are improvements only against a miscalibrated oracle.","fun_headline_variants_meta":{"raw":{"variants":["Agentic framework SABLE optimizes multi-objective drug leads, synthesizable","LLM orchestrates multi-objective drug optimization with synthesis constraints","SABLE: agentic ligand explorer for synthesizable hit-to-lead optimization","Drug optimizer uses LLM to guide Pareto-front analog exploration","Modular agentic system improves drug analogs, ensures synthesizability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000882,"raw_usage":{"total_tokens":3835,"prompt_tokens":991,"completion_tokens":2844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2753}},"tokens_in":607,"tokens_out":2844,"duration_ms":17789,"temperature":1.0,"reasoning_tokens":2753,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:03.422376+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize a handful of the top-ranked SABLE analogs from the METTL3 and Boltz-2-validated-target campaigns and measure their IC50s in the corresponding assays; if most fail to beat the seed by a meaningful margin, the enrichment claim would be shown to be an artifact of the Boltz-2 objective rather than genuine lead improvement.","supporting_citations":[{"cited_title":"Journal of Clinical Epidemiology , volume =","cited_arxiv_id":null,"evidence_quote":"Supplies HEALER, the reaction-templated enumerator that restricts the search space to analogs constructible from commercially available building blocks."},{"cited_title":"and Ma, Jianzhu , year = 2024, month = jun, number =","cited_arxiv_id":null,"evidence_quote":"Provides the qLogNEHVI acquisition function used for multi-objective optimization."},{"cited_title":"doi:10.1093/nar/gkae236 , urldate =","cited_arxiv_id":null,"evidence_quote":"Supplies STOPLIGHT, the ADMET and property predictor used for objectives such as CNS activity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the METTL3 retrospective series, including the seed compound and published UZH2, used to check SABLE against a real medicinal-chemistry trajectory."}],"review_version":1}