{"id":"6369629f-ee5b-492d-b421-7afbcc3c709b","arxiv_id":"2412.03424","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Tango* uses a computed molecular similarity reward (TANGO) inside Retro* to solve starting material-constrained retrosynthesis with higher success and fewer expansions than neural value function baselines.","lead":"A computer-aided synthesis planning method, Tango*, adds a chemical similarity score to guide retrosynthetic search toward specified starting materials, and reports higher solve rates and lower computation time than existing neural-guided methods. A generalist might care because it offers a simpler route to repurposing waste or renewable feedstocks into useful drugs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-paper baseline numbers make the claimed efficiency and solve-rate advantages uncontrolled; rerunning baselines under identical conditions is required.","rationale":"The paper is best read as a claim that a cheap cheminformatic cost can replace a learned distance network for constrained retrosynthesis. For that claim to hold, Tango* must beat the baselines under equivalent conditions. The reuse of Yu et al.'s baseline numbers is the weakest link: the paper does use the same single-step model and value network for Tango*, which is a strength, but it does not establish that the baseline runs themselves were performed under the same expansion counting, search implementation, or hardware. The wall-clock time comparison in Table 2 is especially fragile because timings for baselines are presumably from Yu et al. and may not be hardware-comparable. I considered other concerns, including the notation mismatch in Algorithm 1 (c is described as the FMS weight but the formula uses c as the Tanimoto weight), the route-length comparison restricted to routes solved by all methods, and the GPT-4o parsing of compound names in Appendix A.5, but these are secondary and do not by themselves invalidate the central result. The concern is addressable by rerunning baselines, so the reader's CONDITIONAL verdict is appropriate and no verdict change is needed.","tokens_in":13438,"tokens_out":5794,"duration_ms":56173,"concrete_test":"Re-run Retro*, Retro*+D, DESP-F2E, and DESP-F2F in the released Tango* codebase, using the identical single-step and value-network checkpoints, building-block set, expansion-limit definitions, and workstation as Tango*, and recompute Tables 1 and 2. If the baseline numbers do not reproduce, the margins in the reported outperformances must be re-quantified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Tables 1 and 2 is not fully controlled: for all non-Tango methods (Retro*, GRASP, Retro*+D, DESP-F2E/F2F) the paper reuses numbers 'taken from Yu et al.' (Section 4.1), while Tango* runs are performed in the authors' own DESP-based implementation on their workstation (Appendix A.1). The claims that Tango* 'consistently outperforms' Retro*+D and achieves 'lower wall clock times' depend on (i) identical expansion-limit semantics (model calls vs node expansions), (ii) identical single-step model and value-network checkpoints, (iii) the same building-block set, and (iv) comparable hardware for wall-clock timings. None of these are verified in the paper. If Yu et al. count an expansion differently, or if their wall-clock times were measured on slower hardware, the reported margins (e.g., 5.8 s vs 8.3 s on Pistachio Reachable in Table 2) could shrink or invert. This is the load-bearing point because every headline result is a comparison to these external numbers; the paper's own ablations only compare Tango* variants against each other.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Tango*, a node-cost function for starting-material-constrained retrosynthetic planning. TANGO combines Tanimoto similarity and Fuzzy Matching Substructure between a molecule and the specified starting material; the cost is k*(1 - max_sm [c*Tanimoto + (1-c)*FMS]) added to the Retro* cost. The authors integrate this cost into Retro* and into the bidirectional DESP-F2E/F2F algorithms, and evaluate on USPTO-190, Pistachio Reachable, and Pistachio Hard. They report that Tango(1,0)* outperforms the neural-network-guided Retro*+D and matches or exceeds DESP baselines in solve rate, with fewer expansions and lower wall-clock time. They also analyze value-function monotonicity along ground-truth routes and present case studies, including a route to chlorambucil from renewable feedstocks.","tokens_in":13625,"tokens_out":6351,"duration_ms":53728,"significance":"If the reported results are reproducible under controlled comparison, the contribution is notable: a simple, non-learned similarity heuristic can replace a learned pairwise distance network for constrained synthesis planning, improving both efficiency and solve rate. The paper is transparent about using the same single-step and value-network checkpoints as Yu et al. and discloses code. The hyperparameter tuning on one dataset and transfer to others is a strength, as are the literature-validated case-study routes. However, the headline comparisons against Retro*, GRASP, Retro*+D, and DESP rely on numbers imported from a different paper, which currently prevents the efficiency and solve-rate claims from being considered fully controlled.","major_comments":[{"comment":"All baseline numbers for Retro*, GRASP, Retro*+D, and DESP-F2E/F2F are taken from Yu et al. rather than re-run under the same conditions. The paper's central claims of 'consistently outperforms' and 'lower wall clock times' depend on equivalence of expansion-budget semantics (model calls vs node expansions), single-step model and value-network checkpoints, the building-block set, and hardware. The authors use Yu et al.'s checkpoints and an implementation based on the DESP codebase, which makes equivalence plausible, but it is not demonstrated. This is load-bearing because every headline result is a comparison to these external numbers; the authors' own ablations only compare Tango* variants to each other. I ask that the baselines be re-run under the authors' exact evaluation harness, or, failing that, that the paper provide a detailed point-by-point argument for equivalence of all four factors listed above.","section":"Section 4.1, Tables 1 and 2"},{"comment":"The definition of the FMS/Tanimoto weighting parameter c is internally inconsistent. Algorithm 1 computes reward_sm = TanSim·c + FMS·(1−c), which makes c the Tanimoto weight. The text, however, states that 'c defines the FMS weight' and then refers to 'Tango with c = 0.0 as Tango(1, 0)'. Under the formula in Algorithm 1, c=0 gives reward_sm = FMS, i.e., pure FMS guidance, not pure Tanimoto as the name Tango(1,0) implies. This makes the exact cost function ambiguous and harms reproducibility. Please correct either the algorithm or the naming convention.","section":"Section 4.1 and Algorithm 1"},{"comment":"The abstract claims that 'by optimising a single hyperparameter, Tango* outperforms existing methods', but the method in fact tunes two hyperparameters, k and c, and the paper's main results use Tango(1,0) with c=0.0, which is not the value found optimal by the hyperparameter screen (c=0.3, i.e., Tango(0.7,0.3), improves Pistachio Reachable but not the harder datasets). The authors should clarify whether the proposed method is the tuned configuration or the manually selected c=0.0 configuration, and should state the tuning procedure accurately. This matters because the abstract's 'single hyperparameter' claim and the choice of the headline configuration are currently presented inconsistently.","section":"Abstract and Section 4.1"}],"minor_comments":[{"comment":"The text uses 'UPSTO-190' where 'USPTO-190' is meant.","section":"Section 4.2"},{"comment":"The word 'constrainted' should be 'constrained'.","section":"Figure 1 caption"},{"comment":"The CPU model is listed as 'AMD Rysen 9 7900X'; the correct spelling is 'Ryzen'.","section":"Appendix A.1"},{"comment":"The paragraph ending with 'perform substantially better at estimating synthetic distance. unprivileged setting.' contains an incomplete final sentence fragment; please rewrite.","section":"Section 4.4"},{"comment":"The caption states 'Route length comparisons are made on the routes solved by all methods', but it is not specified whether the wall-clock times are also computed only on the common subset or over the entire benchmark; please clarify.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The cross-paper baseline comparison is the single largest risk to the paper's central claim. Even though the authors reuse Yu et al.'s checkpoints, subtle differences in expansion counting, building-block preprocessing, and hardware can change the reported margins. I would encourage the editor to require a re-run of the baselines in the authors' own environment, or, if that is infeasible, a very careful equivalence statement with concrete evidence (e.g., logs, counts of expansions per solved instance). The code release is a positive signal and should help. If the authors can provide controlled baseline numbers, the paper could become a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely attractive: replace the learned distance network in constrained retrosynthesis with a fixed TANGO similarity cost, and show it works in Retro* and even improves the bidirectional DESP methods. I think the idea is real and worth engaging with, but the empirical case is shakier than the abstract suggests, because most baseline numbers are taken from Yu et al. rather than re-run under identical conditions.\n\nWhat's new: TANGO was introduced by the same group for generative design; using it as a node cost in search is new, as is the demonstration that it can replace the trained D network in DESP. The paper also includes a nice analysis of why learned D fails—it overestimates distance on challenging routes, while TANGO shows a more monotonic decrease. That analysis is suggestive, not definitive, but it's a genuine attempt to explain the mechanism.\n\nWhat's done well: they use the same single-step model and Retro* value network as Yu et al., tune the two hyperparameters only on Pistachio Reachable, and report results on three benchmarks. The code link is given. The Chlorambucil case study is a nice practical touch.\n\nSoft spots: the stress-test note is right that the central comparison is against external numbers. The paper never verifies that expansion budgets, model checkpoints, building-block sets, or hardware are identical. The solve-rate differences are often large enough that I'd bet the qualitative conclusion survives, but the wall-clock and expansion-count claims could invert under different conditions. Also, no variance or multiple runs are reported, and a few small differences (e.g., 84.5 vs 84.0) are not meaningful. The abstract says 'outperforms' where the body says 'outperforms or matches.'\n\nNone of this is a dealbreaker. A referee should ask the authors to re-run the baselines or at least document why the comparison is fair. The core idea is solid and the paper deserves review.\n\nFor a reader: it's useful for anyone working on synthesis planning or benchmarking search algorithms. My recommendation: send to peer review, with a request for re-running baselines.","headline":"A simple similarity heuristic beats a learned distance network for constrained retrosynthesis—plausible and worth a referee, but the numbers are partly borrowed from another paper.","tokens_in":14191,"tokens_out":4432,"would_cite":true,"duration_ms":42228,"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":"A computed chemical similarity function, TANGO, added to Retro* solves more starting-material-constrained synthesis planning problems than neural-guided baselines, using fewer expansions and less wall-clock time.","keywords":["retrosynthesis","computer-aided synthesis planning","starting material constrained synthesis","TANGO","Tanimoto similarity","Fuzzy Matching Substructure","Retro*","DESP"],"falsifier":"Re-run the three benchmark suites with identical expansion budgets, model checkpoints, and building-block sets, re-running the baselines in the same codebase; if Retro*+D or DESP-F2F matches or beats Tango* under those controlled conditions, the claimed advantage is an artifact of the comparison. Alternatively, retrain the DESP distance network with balanced negative sampling; if its estimates become monotonic and Tango*'s advantage disappears, the paper's mechanism explanation is wrong.","tokens_in":13203,"feed_emoji":"🧪","tokens_out":6779,"duration_ms":56555,"temperature":0.7,"pith_summary":"The paper claims that starting-material-constrained retrosynthesis—planning a route that must end at a specified building block—does not need a specialized bidirectional algorithm or a learned distance network. The authors add a computed node cost, TANGO (a weighted mix of Tanimoto similarity and fuzzy substructure matching to the target starting material), to the existing Retro* search, and show it solves more benchmark (target, starting material) pairs than the neural-network-guided Retro*+D and the bidirectional DESP baselines, with fewer expansions and lower wall-clock time. The same cost function also improves DESP when swapped in for its learned pairwise-distance network, with the best variant reaching a 99.3% solve rate on one benchmark. The paper's diagnostic plots offer a reason: the TANGO cost decreases monotonically and granularly along ground-truth routes, while the learned distance estimates are noisy and often plateau. If true, constrained planning becomes a cheap add-on to general-purpose planners rather than a separate algorithmic family.","feed_headline":"Similarity score beats neural guides in constrained synthesis planning","feed_subtitle":"Replacing a learned distance network with a chemical similarity score solves more constrained routes, faster.","key_machinery":"TANGO (TANimoto Group Overlap) is the load-bearing object: a node cost function that measures how structurally similar a molecule node is to any enforced starting material, combining Tanimoto similarity and Fuzzy Matching Substructure (FMS) with a weight $c$, scaled by $k$, and added to the Retro* cost. It is computed directly from molecular structure rather than learned, so it needs no training and no special architecture. Plugged into Retro*'s best-first expansion, it biases the search toward nodes that resemble the target starting material; plugged into DESP in place of its learned pairwise distance network, it yields Tango-F2E and Tango-F2F. The argument that it works rests on the empirical monotonicity of TANGO along ground-truth routes compared with the noisy, plateauing estimates of the learned distance network.","core_discovery":"The paper's central claim is that the TANGO node cost function—$k\\cdot(1-\\max_{sm}[c\\cdot \\mathrm{FMS}(node,sm)+(1-c)\\cdot \\mathrm{Tanimoto}(node,sm)])$ plus the Retro* cost—guides constrained retrosynthesis better than learned neural guidance. Empirically, Tango(1,0)* outperforms Retro*+D on all three benchmarks (USPTO-190, Pistachio Reachable, Pistachio Hard) at every expansion limit, with solve rates up to 42.6% on USPTO-190 at 500 expansions and 97.3% on Pistachio Reachable, while using fewer expansions and less wall-clock time. Replacing the DESP learned distance network with TANGO yields Tango-F2F, which reaches 99.3% on Pistachio Reachable and roughly 25% higher solve rates than the next-best DESP on the harder datasets. The authors argue the advantage comes from TANGO being a computed, chemically informed value: it is monotonic and granular along true synthetic routes, whereas the neural distance estimates are noisy, plateau, and consistently overestimate synthetic distance.","pith_inferences":["If TANGO's advantage comes from its monotonicity rather than chemical specificity, then other cheap, well-calibrated structural heuristics (e.g., learned embeddings trained with a ranking objective) might reproduce the gain; the paper does not test this.","The diagnostic failure of the learned distance network suggests its negative-sample training (pairs assigned a fixed distance of 10) may be the culprit; rebalancing those samples could close the gap between learned and computed guidance.","Because TANGO is computed from structure, it should transfer across reaction datasets without re-tuning; the paper only shows transfer of hyperparameters from one benchmark to harder sets, not across different single-step models.","The current constraint is a single molecule at the goal; extending the max over starting materials already works, but enforcing multiple structural constraints along one route would require a different composition rule."],"forward_implications":["Constrained synthesis planning can be achieved by adding a computed similarity term to an existing uni-directional planner like Retro*, with no model retraining.","TANGO can replace learned pairwise-distance networks in bidirectional planners; Tango-F2F becomes the strongest solver in the comparison.","Because searches guided by TANGO expand fewer nodes, wall-clock time stays low even though the similarity computation adds per-node overhead.","The same cost idea can be pointed at other structural goals—key intermediates or substructures—not just full starting materials.","Route lengths are comparable to or shorter than neural-guided baselines, so the efficiency gain does not come at the cost of longer routes."],"supporting_citations":[{"why":"Supplies the Retro* algorithm and its neural value network, which Tango* extends with the TANGO node cost.","marker":"[8]"},{"why":"Provides the DESP baselines, the benchmark datasets, the model checkpoints, and the reference implementation used in the experiments.","marker":"[22]"},{"why":"Introduces the TANGO reward that the paper adapts into a node cost function for search.","marker":"[38]"},{"why":"GRASP is the reinforcement-learning baseline for starting-material-constrained planning that Tango* is compared against.","marker":"[21]"},{"why":"Supplies the renewable/waste feedstock set used in the case study of synthesizing useful compounds from waste feedstocks.","marker":"[17]"},{"why":"Provides evidence that the Retro* value function has limited effect on search performance, motivating the paper's focus on guidance functions.","marker":"[33]"}],"fun_headline_variants":["TANGO's computed value beats neural guidance in synthesis","Chemically informed scores solve more constrained routes","TANGO: faster retrosynthesis with a chemistry-based heuristic","One hyperparameter switch improves constrained synthesis planning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the published baseline numbers for Retro*, GRASP, Retro*+D, and DESP were produced under conditions equal to this paper's runs—same expansion budgets, same single-step model and value network checkpoints, same building-block set, and comparable hardware—so the reported solve-rate and efficiency margins could shrink if those conditions differ.","fun_headline_variants_meta":{"raw":{"variants":["TANGO's computed value beats neural guidance in synthesis","Chemically informed scores solve more constrained routes","TANGO: faster retrosynthesis with a chemistry-based heuristic","One hyperparameter switch improves constrained synthesis planning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1858,"prompt_tokens":952,"completion_tokens":906,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":843}},"tokens_in":568,"tokens_out":906,"duration_ms":8229,"temperature":1.0,"reasoning_tokens":843,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:24:29.606526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the three benchmark suites with identical expansion budgets, model checkpoints, and building-block sets, re-running the baselines in the same codebase; if Retro*+D or DESP-F2F matches or beats Tango* under those controlled conditions, the claimed advantage is an artifact of the comparison. Alternatively, retrain the DESP distance network with balanced negative sampling; if its estimates become monotonic and Tango*'s advantage disappears, the paper's mechanism explanation is wrong.","supporting_citations":[{"cited_title":"Retro*: Learning retrosynthetic planning with neural guided A* search","cited_arxiv_id":null,"evidence_quote":"Supplies the Retro* algorithm and its neural value network, which Tango* extends with the TANGO node cost."},{"cited_title":"It Takes Two to Tango: Directly Optimizing for Constrained Synthesizability in Generative Molecular Design","cited_arxiv_id":"2410.11527","evidence_quote":"Introduces the TANGO reward that the paper adapts into a node cost function for search."},{"cited_title":"Grasp: Navigating retrosynthetic planning with goal-driven policy","cited_arxiv_id":null,"evidence_quote":"GRASP is the reinforcement-learning baseline for starting-material-constrained planning that Tango* is compared against."},{"cited_title":"Herrera, Josef M","cited_arxiv_id":null,"evidence_quote":"Supplies the renewable/waste feedstock set used in the case study of synthesizing useful compounds from waste feedstocks."}],"review_version":1}