{"id":"b16a7e8e-b99b-463b-949f-080d9d09d4d1","arxiv_id":"2411.16623","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A MIP-based molecular generator that automatically converts supported SMARTS inclusion and exclusion patterns into linear constraints and outputs feasible molecules.","lead":"Limeade is an open-source tool that lets users request or forbid chemical substructures, then generates molecules that obey those requests. It is designed to make optimization-based molecular design usable by chemists who do not write optimization models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Consecutive-window inclusion (Eq. 6) is incomplete for multiple disjoint disconnected SMARTS patterns, causing false infeasibility despite valid molecules.","rationale":"The reader correctly identified disconnected SMARTS as the weak spot, but the specific mechanism in their weakest_assumption is too strong. The paper's C3 forces a bond between the first two indices only when a window starts at index 0; a single disconnected pattern can be placed in a later window. However, with two disjoint disconnected patterns, one block must contain indices 0 and 1, and the forced bond makes that block unable to host a no-bond pattern. The path P6 example is a concrete counterexample to completeness. This is a correctness risk for the central inclusion claim, not a mere absence of disclosure, because it turns a feasible request into an infeasible model. It is addressable by either restricting/disclosing the supported class or extending the encoding, so conditional acceptance remains appropriate. The case study and other claims are not affected by this specific issue.","tokens_in":15143,"tokens_out":30132,"duration_ms":291073,"concrete_test":"Enumerate all 6! permutations of the path P6 (atoms C1-C2-C3-N4-N5-N6) and check three conditions: (a) C3 holds for every prefix, (b) C1 and C3 occupy consecutive positions, (c) N4 and N6 occupy consecutive positions. No permutation satisfies all three; this proves the path is infeasible under (6). Then run Limeade with N_atoms=6, atoms=['C','N'], bounds_atoms([3,3],[3,3]), bounds_double_bonds(0,0), bounds_triple_bonds(0,0), bounds_rings(0,0), include_substructures(['C.C','N.N']); if the solver reports infeasible (or the pool omits the path despite other solutions existing), the encoding is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is that constraint (6) is not a complete encoding of SMARTS inclusion when disconnected patterns must be placed disjointly. A single disconnected pattern such as 'C.C' can be placed in a window starting at i>0, so the specific example in the Reader's Report is not fatal. The real failure is with two disjoint disconnected patterns. Consider the six-atom path C1-C2-C3-N4-N5-N6. It contains the non-bonded carbon pair {C1,C3} and the non-bonded nitrogen pair {N4,N6}, so it satisfies the two inclusions 'C.C' and 'N.N'. In any index ordering satisfying C3, {C1,C3} being consecutive forces C2 to be the first index: if C2 comes after C1/C3, then C1 or C3 has no earlier neighbor; if C2 comes before both and is not first, C2 itself has no earlier neighbor. The same argument forces N5 to be first for {N4,N6}. Both cannot be first, so no C3-valid ordering puts both pairs on consecutive indices. Hence the MIP (6) reports infeasible for a molecule that genuinely contains both requested patterns. The paper does not restrict supported SMARTS to connected patterns or disclose that multiple disconnected inclusions are incomplete, so the central claim of general SMARTS support is overstated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Limeade is a mixed-integer programming (MIP) framework for de novo molecular generation. It encodes molecular graphs over a fixed number of heavy atoms with binary variables for atom types, neighbor counts, implicit hydrogens, and bond types, imposes structural constraints largely inherited from prior work, and adds user-facing constraints for atom/bond/ring composition. The claimed new contribution is an automatic translation of SMARTS substructure inclusion and exclusion into linear constraints, together with batch solution pooling and IIS-based infeasibility diagnostics. The paper demonstrates the workflow on an aspirin-like example and on an iterative case study that excludes substructures associated with rare Morgan fingerprints.","tokens_in":15421,"tokens_out":14080,"duration_ms":135099,"significance":"If the SMARTS inclusion encoding were complete, Limeade would be a useful low-barrier MIP-based generation tool: it ships open-source code, provides a Pyomo implementation, reports IIS for infeasible models, and the case study shows a large improvement in the fraction of molecules avoiding rare fingerprints after iterative exclusions. However, the central technical claim of general SMARTS inclusion support is compromised by the incompleteness of the consecutive-index formulation, so the tool's practical scope is narrower than stated.","major_comments":[{"comment":"The consecutive-index inclusion formulation is not a complete encoding of SMARTS inclusion for multiple disjoint disconnected patterns, and it can falsely declare infeasibility. Consider the six-atom path C1-C2-C3-N4-N5-N6. This molecule contains the non-bonded carbon pair {C1,C3} and the non-bonded nitrogen pair {N4,N6}, so it satisfies the two inclusion requirements 'C.C' and 'N.N'. In any ordering satisfying (C3), the pair {C1,C3} can occupy consecutive indices only if C2 is the very first atom of the molecule: if C2 is after either of C1,C3, the earlier of C1,C3 has no earlier neighbor; if C2 is before both but not first, C2 has no earlier neighbor. The same argument forces N5 to be the first atom for {N4,N6}. Since C2 and N5 are distinct, no C3-valid ordering places both pairs on consecutive windows, and the MIP (6) reports infeasible even though the molecule genuinely contains both patterns. This directly undermines the abstract's and Section 3.2's claim of general SMARTS inclusion support.","section":"§2.3, Eq. (6)"},{"comment":"The paper does not restrict the inclusion feature to pattern classes for which Eq. (6) is exact, nor does it disclose the consecutive-index assumption. Section 3.2 presents the supported SMARTS components without any connectivity requirement, and the API example accepts arbitrary SMARTS strings; the abstract states that Limeade 'supports inclusion and exclusion of SMARTS patterns' without qualification. The manuscript should either implement a complete inclusion encoding (e.g., with explicit subgraph-isomorphism variables that allow non-consecutive placements) or explicitly document that inclusion is an approximation that may reject feasible molecules, and characterize the exact cases (for instance, a single connected pattern placed at the beginning of the index ordering). The validation step in Section 3.3 cannot mitigate false infeasibility, because when the MIP is infeasible there are no generated molecules to validate.","section":"§2.3 and §3.2"}],"minor_comments":[{"comment":"The text reads 'Paper structureSection 2 introduces'; it should read 'Paper structure. Section 2 introduces'.","section":"Page 5"},{"comment":"The phrase 'For a molecular with N heavy atoms' should be 'For a molecule with N heavy atoms'.","section":"Page 7"},{"comment":"In the hydrogen-count sum, the term X_{i,I_h_i} should be X_{v,I_h_i}; the atom index v is missing.","section":"Appendix A.4, Eq. (C19)"},{"comment":"The run-together text 'attributeAtomType' should be 'attribute AtomType', referring to the RDKit atom attribute.","section":"§3.2"},{"comment":"The phrase 'no more than 10 5 constraints' should be written as '10^5 constraints' for readability.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"This is a tool/software paper; the open-source release and the IIS diagnostics are useful. The main technical weakness is the unqualified SMARTS-inclusion claim, which I believe can be addressed by a careful statement of the supported pattern classes and by adding a warning when the consecutive-window approximation may cause false infeasibility. The case-study feasibility criterion (rare Morgan fingerprints) is an illustrative operationalization and should be described as such."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Limeade is a genuine tool contribution: it packages the known MIP encoding for molecular graphs and adds a linear encoding of SMARTS-based substructure inclusion/exclusion, with open-source code and a practical case study. The paper does not pretend the structural constraints are new—it credits Zhang et al.—and the genuinely new piece, the component-based linear constraints in Section 2.3, is the right thing to evaluate. That part is mostly sound for connected patterns, and the aspirin walk-through and infeasibility diagnostics show the tool working as advertised.\n\nSoft spots: the consecutive-index inclusion constraint (6) is not a complete encoding for disconnected SMARTS patterns. The reader's 'C.C' example is not itself fatal—that pattern can often be placed in a window starting after the first index. But the stress-test example is correct: a path C-C-C-N-N-N containing the non-bonded pair {C1,C3} and {N4,N6} cannot be indexed so both pairs lie in consecutive windows, because each middle atom would need to be first. So the MIP reports infeasible for a molecule that genuinely satisfies both requested patterns. The paper neither restricts supported SMARTS to connected patterns nor discloses this limitation, so the abstract's claim of general SMARTS support is overstated.\n\nThe case study is weaker than the headline suggests. The iterative exclusion story is nice, but there are no baselines—random generation, a simpler constraint set, or another generative tool—and no repeated runs, so the 2/957 vs. 55,498/63,973 numbers are an existence proof, not a measured improvement. The free parameters (rarity threshold, composition bounds, which patterns to exclude) are chosen post hoc. Minor issue: they drop symmetry-breaking when inclusion constraints are active, so the validation step has to filter duplicates; that is disclosed and acceptable.\n\nBottom line: the core encoding is not load-bearingly wrong for connected patterns, the tool is released, and the limitation is fixable with a restricted-pattern disclosure or a better encoding. This deserves a serious referee. I would send it, with a request to fix or disclose the disconnected-pattern gap and to add a baseline to the case study.","headline":"Useful, mostly sound MIP-generation tool paper; the inclusion encoding is incomplete for disconnected SMARTS patterns, but it deserves a serious referee with targeted revision requests.","tokens_in":15936,"tokens_out":5900,"would_cite":true,"duration_ms":55973,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","92E10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Limeade is a mixed-integer programming tool that converts SMARTS substructure requests into linear constraints and generates a pool of molecules satisfying them.","keywords":["mixed-integer programming","computer-aided molecular design","SMARTS","molecular generation","substructure inclusion","substructure exclusion","constraint encoding","Morgan fingerprints"],"falsifier":"Run Limeade with a disconnected SMARTS pattern, for example include 'C.C' (two separate carbons with no bond) in a two-carbon molecule; the paper's own formulation should declare this infeasible even though such a molecule clearly exists, demonstrating that the consecutive-index assumption is a genuine restriction on the class of supported patterns.","tokens_in":14950,"feed_emoji":"🧪","tokens_out":6827,"duration_ms":57012,"temperature":0.7,"pith_summary":"This paper presents Limeade, a mixed-integer programming (MIP) tool for generating molecules from user-supplied chemical requirements. The central claim is that substructure requests written in SMARTS notation—both 'must include' and 'must exclude' patterns—can be translated automatically into linear constraints, so that solving the MIP produces a pool of molecules that satisfy the requests. The paper demonstrates the encoding for atom-level, bond-level, and graph-level patterns, and shows an iterative workflow in which excluding unwanted substructures raises the share of generated molecules that avoid rare Morgan fingerprints from 2 out of 957 to about 86 percent of 63,973 unique molecules. If correct, the tool lets chemists specify design needs directly and obtain candidate molecules without formulating an optimization model by hand.","feed_headline":"MIP tool turns SMARTS patterns into molecule pools","feed_subtitle":"Users give must-have and forbidden substructures; Limeade returns matching candidates and flags contradictions.","key_machinery":"The central mechanism is the component-matching constraint system built on an integer molecular encoding. A molecule is represented by binary feature variables for atoms (type, neighbor count, hydrogen count, double/triple bond flags) and binary adjacency matrices for bonds. A substructure is reduced to a list of M 'components'—properties such as 'there is a bond', 'there is no bond', 'atom is in set T', 'atom has d neighbors'—each expressible as a linear term in the variables. For a candidate window of n consecutive atom indices, the sum S(V_n) of matched components is compared against M: exclusion requires S < M for every window, while inclusion introduces a binary variable σ with S ≥ M·σ and requires at least one σ = 1. This reduces substructure requirements to O(N − n) linear constraints and binary variables, at the cost of assuming the substructure can be placed on consecutive indices.","core_discovery":"Limeade claims that the molecular design problem is better framed as feasible-region sampling than as single-molecule optimization: the user supplies structural requirements, and the solver returns a diverse pool of molecules satisfying all of them. The enabling mechanism is a set of linear constraints that encode substructure matching. For inclusion, binary variables indicate whether a window of n consecutive atom indices fully matches the given substructure, and the model requires at least one such window; for exclusion, the model requires that no window reaches a full match. Because a perfect match is defined by the sum of linear 'component' terms (bond presence, atom type, hydrogen count, neighbor count, and so on), SMARTS patterns are converted into constraints rather than compared by graph isomorphism. The paper's case study shows that iteratively excluding rare Morgan fingerprints improves feasibility from 2/957 to roughly 86% of 63,973 generated unique molecules.","pith_inferences":["The consecutive-index relaxation means Limeade is best suited to connected, localized substructure patterns; applying it to dispersed or disconnected fragments requires either manual rephrasing of the pattern or relying on the post-hoc validation step rather than the constraints.","Scaling the inclusion constraints to multiple simultaneous substructures without the consecutive-index shortcut would require additional binary variables and reintroduce the symmetry-breaking complications the authors deliberately avoided, so a practical extension would likely need a different formulation.","Pairing Limeade's feasible-region sampling with a learned score (such as a graph neural network) in a rank-then-select loop could sidestep the hardness of optimizing directly over the molecular space, since the tool already provides a diverse set of candidates to score.","The validation-step fallback for large exclusion patterns means the guarantee of exclusion is softened for big substructures; one could measure the empirical frequency of excluded patterns after validation to see where the line between constraint-based and validation-based exclusion should be drawn."],"forward_implications":["A user who can write a SMARTS string can request and forbid substructures without any knowledge of mixed-integer programming.","Iterative exclusion of undesired motifs can rapidly reshape a generated library, as shown by the Morgan fingerprint case study where feasible molecules rose from 2/957 to about 86%.","When user requirements are contradictory, Limeade reports an irreducible inconsistent subsystem that identifies which constraints clash.","Because generation samples the feasible region rather than optimizing a single score, the solution pool naturally offers diverse candidate structures, and batch generation with varied random seeds increases that diversity."],"supporting_citations":[{"why":"Supplies the molecular representation, structural constraints, and integer encoding that Limeade inherits and extends.","marker":"[43]"},{"why":"Provides symmetry-breaking constraints and the numerical estimate of the six-atom molecule count used to justify deferring large exclusion patterns to validation.","marker":"[42]"},{"why":"Defines the SMARTS pattern language that users employ to specify substructure inclusion and exclusion.","marker":"[68]"},{"why":"Used to extract the linear component counts (atom types, bond types, hydrogen counts, neighbor counts) from SMARTS strings.","marker":"[69]"},{"why":"The MIP solver that builds the model, populates the solution pool, and computes an irreducible inconsistent subsystem for infeasible models.","marker":"[70]"},{"why":"The public bioactivity database whose fingerprint frequencies define the feasibility criterion in the case study.","marker":"[72]"}],"fun_headline_variants":["SMARTS patterns become molecule constraints","From SMARTS to a pool of valid molecules","Limeade: encode requirements, generate candidates","MIP turns design rules into molecule sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inclusion mechanism assumes that any molecule containing the requested substructure can be renumbered so that the substructure's atoms sit on consecutive positions; this holds for connected patterns, but fails for disconnected patterns like 'C.C', where the connectivity rules force consecutive atoms to be bonded.","fun_headline_variants_meta":{"raw":{"variants":["SMARTS patterns become molecule constraints","From SMARTS to a pool of valid molecules","Limeade: encode requirements, generate candidates","MIP turns design rules into molecule sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000846,"raw_usage":{"total_tokens":3642,"prompt_tokens":863,"completion_tokens":2779,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":2722}},"tokens_in":479,"tokens_out":2779,"duration_ms":21406,"temperature":1.0,"reasoning_tokens":2722,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:56:03.834881+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Limeade with a disconnected SMARTS pattern, for example include 'C.C' (two separate carbons with no bond) in a two-carbon molecule; the paper's own formulation should declare this infeasible even though such a molecule clearly exists, demonstrating that the consecutive-index assumption is a genuine restriction on the class of supported patterns.","supporting_citations":[{"cited_title":"S.; Feldmann, C.; Sandfort, F.; Mathea, M.; Misener, R","cited_arxiv_id":null,"evidence_quote":"Supplies the molecular representation, structural constraints, and integer encoding that Limeade inherits and extends."},{"cited_title":"S.; Feldmann, C.; Walz, D.; Sandfort, F.; Mathea, M.; Tsay, C.; Misener, R","cited_arxiv_id":null,"evidence_quote":"Provides symmetry-breaking constraints and the numerical estimate of the six-atom molecule count used to justify deferring large exclusion patterns to validation."},{"cited_title":"SMARTS - A Language for Describing Molecular Patterns","cited_arxiv_id":null,"evidence_quote":"Defines the SMARTS pattern language that users employ to specify substructure inclusion and exclusion."},{"cited_title":"https://www.rdkit.org","cited_arxiv_id":null,"evidence_quote":"Used to extract the linear component counts (atom types, bond types, hydrogen counts, neighbor counts) from SMARTS strings."},{"cited_title":"2023; https://www.gurobi.com","cited_arxiv_id":null,"evidence_quote":"The MIP solver that builds the model, populates the solution pool, and computes an irreducible inconsistent subsystem for infeasible models."},{"cited_title":"C\", \"N\",","cited_arxiv_id":null,"evidence_quote":"The public bioactivity database whose fingerprint frequencies define the feasibility criterion in the case study."}],"review_version":1}