Methods, media, and products for retrosynthetic route planning
By constructing a candidate pool and an independent validation process, combined with a generative model and route search state update, the problem of generating unexecutable routes in existing technologies is solved, achieving more efficient retrosynthetic route planning that is adaptable to a variety of molecular inventories.
Patent Information
- Application Number
- CN202611552802.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-09
- Publication Date
- 2026-11-06
AI Technical Summary
Existing retrosynthetic planning techniques fail to effectively utilize the inventory evidence associated with the current molecule when generating molecular inventory information. This results in generated routes that may be unexecutable or unable to connect to the molecular inventory, and the model needs to be retrained to adapt to different molecular inventories, reducing the applicability of the system.
By acquiring target molecules, molecular inventory, and route search status, a candidate pool is constructed based on molecular structure association. Generative models are used to generate retrosynthetic proposals, and an independent validation process is combined to update the route search structure, ensuring that the generated routes comply with chemical rules and that materials can be obtained from the inventory.
It increases the probability of generating closable routes, reduces invalid expansion, adapts to various molecular inventories without retraining the model, and ensures that the generated routes have verifiable computational basis.
Smart Images
Figure CN123053280A_ABST
Abstract
Description
Technical Field
[0001] This invention generally relates to the field of computer-aided synthesis planning based on artificial intelligence technology, and specifically to methods, media, and products for inverse synthesis route planning. Background Technology
[0002] Retrosynthetic analysis starts with the target molecule and progressively breaks it down into simpler precursor molecules until the raw material that can be obtained directly is obtained. Computer-aided retrosynthetic analysis typically includes steps such as reaction step generation, reaction validity assessment, route search, and raw material availability assessment.
[0003] Some approaches employ reaction templates, single-step prediction models, or heuristic value functions to expand candidate reactions in search trees or hypergraphs. Another approach utilizes generative models, such as large language models, to directly generate one or more retrosynthetic steps or route fragments. To avoid directly treating correctly formatted but chemically non-executable text as routes, some systems can further configure parsers, reaction template validators, and AND / OR route searchers to perform post-filtering of the generated results and query commercially available molecule inventories at route leaf nodes.
[0004] However, information from the existing molecular inventory is typically only used to determine whether leaf nodes are purchasable, score candidate routes, or decide whether to terminate the search after the retrosynthetic step or route fragment generation is complete. Generative models do not obtain actual molecular evidence from the molecular inventory associated with the current molecule to be expanded when selecting the current bond-breaking and precursor molecules. Therefore, even if precise library closure rules are eventually set, the limited search budget may still be consumed on branches that are difficult to connect to the molecular inventory, outputs that cannot be resolved, or reaction steps that cannot be chemically verified.
[0005] Some schemes allow users to specify a starting material (raw material) or screen commercially available molecules from a pool of generated candidates, but this requires retraining the generative model when the molecule inventory comes from different suppliers, lab inventory, or user-defined sets, thus reducing the applicability of the system.
[0006] Therefore, there is a need in this field for improved inverse synthesis planning techniques. Summary of the Invention
[0007] This invention is provided to offer an improved inverse synthetic programming technique.
[0008] One aspect of the present invention provides a method for retrosynthetic route planning, comprising the following steps: S1: obtaining a target molecule, a molecular inventory, and a route search state associated with a route search structure for the current search round, the route search structure including a retrosynthetic route from the target molecule to the original material, the route search state including a verified partial route history and an unresolved molecule set; S2: identifying the target molecule or a molecule to be expanded based on the unresolved molecule set as a query molecule; S3: obtaining a candidate pool from the molecular inventory based on at least one molecular structure association between the query molecule and molecules in the molecular inventory, the candidate pool including multiple inventory candidate records; S4: constructing the multiple inventory candidate records into multiple candidate molecule records, each candidate molecule record... The molecule selection record includes the identifier of the queried molecule, the canonical structural representation of the molecules in the molecule inventory, the candidate type, the ranking score, and the availability label; S5: Determine inventory guidance information based on the multiple candidate molecule records; S6: Construct route generation input based on the verified partial route history, the unresolved molecule set, the queried molecule, and the inventory guidance information; S7: Determine inverse synthesis proposals based on the route generation input using a generative model; S8: Perform an independent verification process on the inverse synthesis proposals; S9: Determine verified search actions based on the inverse synthesis proposals that have passed the independent verification process; and S10: Update the route search structure based on the verified search actions, wherein the inverse synthesis route to be output is generated at least in part based on the updated route search structure.
[0009] As described above, step S3 includes: S31: obtaining a candidate pool from the molecular inventory based on at least one of the following molecular structure associations between the queried molecule and the molecules in the molecular inventory: molecular fingerprint neighborhood, molecular skeleton association, substructure association, maximum common subgraph association, and molecular fragment association.
[0010] As described in any of the above methods, the candidate pool is also obtained based on a quantity limit, which indicates the quantity of inventory candidates to be obtained.
[0011] As described in any of the preceding methods, S4 includes: S41: performing direct projection on the candidate molecule record to obtain a compact record line, the compact record line including a canonical structural representation of a molecule in the molecule inventory, the candidate type, the availability tag, and a weak guidance indicator, the weak guidance indicator being used to indicate that the corresponding molecule in the molecule inventory is used only if there is a chemical basis for bond breaking, wherein the inventory guidance information is determined based on the compact record line.
[0012] As described in any of the preceding methods, S4 includes: S42: performing a summary process on the same candidate molecule records in the candidate molecule records to determine a summary, the summary including useful inventory directions, key constraints and route generation instructions, wherein the inventory guidance information is determined based on the summary.
[0013] As described in any of the above methods, the retrosynthetic proposal includes: current product molecules, reactant molecules, current molecule set, updated molecule set, reaction representation, and step description.
[0014] As described in any of the above methods, the independent verification process includes: format and field parsing checks; consistency checks between the current product and the molecule to be expanded; consistency checks between the molecular set and route state updates before and after the reaction; molecular structure legality, atom mapping, valence state, or stereochemistry checks; reaction template suitability checks; consistency checks between the generated reactants and the template execution results; consistency checks between inventory identifiers and precise membership relationships; and checks for repeated steps, loop routes, or other routes that do not meet the preset route rules.
[0015] As described in any of the above methods, S10 includes: S101: inserting the verified search action into the route search structure; S102: removing successfully expanded molecules from the route search structure; and S103: adding reactants that do not meet the conditions for terminating the search to the set of unresolved molecules.
[0016] The method described in any of the above methods further includes: S11: determining whether to terminate the search.
[0017] As described in any of the above methods, S11 includes: S111: determining whether the search budget has been exhausted; S112: terminating the search in response to determining that the search budget has been exhausted.
[0018] As described in any of the above methods, S11 further includes: S113: determining whether the search route is closed in response to determining that the search budget has not been exhausted; S114: terminating the search in response to determining that the search route is closed; S115: reselecting the molecule to be expanded in response to determining that the search route is not closed.
[0019] The method as described in any of the above further includes: S12: in response to determining that the search will not be terminated, performing S1 to S10 with the route search state associated with the updated route search structure.
[0020] The method described in any of the above further includes: S13: in response to determining to terminate the search, determining whether the inverse synthesis route in the updated route search structure meets the conditions; and S14: based on the determination result of whether the inverse synthesis route in the updated route search structure meets the conditions, selectively outputting the inverse synthesis route or a portion of the inverse synthesis route.
[0021] As described in any of the above methods, S13 includes: S131: determining whether each terminal molecule in the updated route search structure exists in the molecule inventory; and S132: determining whether the reaction steps in the retrosynthetic route in the updated route search structure have all passed the independent verification process.
[0022] As described in any of the above methods, S14 includes: S141: in response to the determination result indicating that not every terminal molecule in the updated route search structure exists in the molecular inventory, outputting the unclosed partial routes; S142: in response to the determination result indicating that not all reaction steps in the retrosynthetic route have passed the independent verification process, outputting the verified partial routes; S143: in response to the determination result indicating that every terminal molecule in the updated route search structure exists in the molecular inventory and all reaction steps in the retrosynthetic route have passed the independent verification process, outputting the retrosynthetic routes in the route search structure of the current search round.
[0023] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the preceding claims.
[0024] Another aspect of the present invention provides a computer program product, including a computer program that, when executed, implements the steps of the method as described in any of the preceding claims.
[0025] The method for retrosynthetic route planning according to some embodiments avoids model outputs that are formatted correctly but chemically invalid, increases the probability of early discovery of closable routes, and can be adapted to a variety of molecular stockpiles without retraining generative models. Attached Figure Description
[0026] Various embodiments of the present invention are described in conjunction with the accompanying drawings.
[0027] Figure 1 This is a schematic diagram of a process for inverse synthesis route planning according to some embodiments.
[0028] Figure 2 This is a schematic diagram illustrating the process of obtaining a candidate pool from a molecular inventory based on the molecular structure association between a queried molecule and molecules in the molecular inventory, according to some embodiments.
[0029] Figure 3 This is a schematic diagram of the composition of candidate molecule record 300 according to some embodiments.
[0030] Figure 4 This is a schematic diagram illustrating the process of determining inventory guidance information using direct projection according to some embodiments.
[0031] Figure 5 This is a schematic diagram illustrating the process of determining inventory guidance information using summary projection according to some embodiments.
[0032] Figure 6 This is a schematic diagram of the composition of a reverse synthesis proposal according to some embodiments.
[0033] Figure 7 This is a schematic diagram of the composition of an independent verification process according to some embodiments.
[0034] Figure 8 This is a schematic diagram illustrating the process of updating the route search structure based on verified search actions, according to some embodiments.
[0035] Figure 9 This is a flowchart illustrating the process of determining whether to terminate the search, based on some embodiments.
[0036] Figure 10 This is a flowchart of the output process of inverse synthesis route planning according to some embodiments.
[0037] Figure 11 This is a flowchart of a method for inverse synthesis route planning according to some embodiments.
[0038] Figure 12 This is a flowchart of a first process associated with a method for inverse synthetic route planning according to some embodiments.
[0039] Figure 13 This is a flowchart of a second process associated with a method for inverse synthetic route planning, according to some embodiments.
[0040] Figure 14 This is a flowchart of a third process associated with a method for inverse synthetic route planning, according to some embodiments.
[0041] Figure 15 This is a flowchart of a fourth process associated with a method for inverse synthetic route planning, according to some embodiments.
[0042] Figure 16 This is a flowchart of a fifth process associated with a method for inverse synthetic route planning, according to some embodiments.
[0043] Figure 17This is a flowchart of a sixth process associated with a method for inverse synthetic route planning, according to some embodiments.
[0044] Figure 18 This is a block diagram of a computer-readable storage medium according to some embodiments.
[0045] Figure 19 It is a block diagram of a computer program product according to some embodiments. Detailed Implementation
[0046] In this application, the term "target molecule" refers to the molecule for which a synthetic route is to be planned. As an example, the target molecule may be represented by a normalized Simplified Molecular Input Line Entry System (SMILES) string, an IUPAC International Chemical Identifier (InChI), a molecular diagram, or other form that can uniquely or normally represent the molecular structure.
[0047] In this application, the term "molecular inventory" refers to a collection of molecular records of raw materials that can serve as the endpoint of a route. As an example, each inventory record in a molecular inventory may include a molecular structure identifier. Furthermore, a molecular inventory may also include inventory source, availability tags, supplier identifiers, internal location identifiers, etc.
[0048] In this application, the term "route search state" refers to a data object used to describe the current search round.
[0049] In this application, the term "query molecule" refers to a molecule used to retrieve records in a molecule inventory.
[0050] In this application, the term "candidate molecule record" refers to a record selected from a molecular inventory.
[0051] In this application, the term "retrosynthetic proposal" refers to one or more reverse reaction steps or pathway segments output by a generative model for the current query molecule.
[0052] In this application, the term "inventory guidance information" refers to information based on the construction of candidate molecule records and added as route generation input prior to retrosynthesis proposal.
[0053] In this application, the term "independent verification process" refers to the analysis and verification process outside the generative model, which is not based on the confidence level or explanatory text provided by the generative model itself.
[0054] Figure 1 This is a schematic diagram of a process for inverse synthesis route planning according to some embodiments.
[0055] This process may involve iteratively updating the route search structure 110 through route searching. The route search state 120 may be associated with the route search structure 110 and may be updated as the route search structure 110 is updated.
[0056] In some embodiments, route search state 120 may include a verified partial route history 122, which may include results of routes verified in previous search rounds. In some embodiments, route search state 120 may include an unresolved molecule set 124, which may include molecules to be searched in the current search round. As an example, in Figure 1 The route search structure 110 is shown as a search tree, where solid circles and solid arrows indicate routes that have been validated in previous search rounds, and hollow circles and dashed arrows indicate a set of unresolved molecules to be searched in the current search round. As an example, other forms of route search structures may also be used, such as AND-OR trees, hypergraphs, or other data structures capable of expressing the dependencies between a product molecule and a set of reactants.
[0057] In some embodiments, target molecule 134, molecular inventory 140, and route search state 120 associated with route search structure 110 for the current search round may be obtained. Route search structure 110 may include a retrosynthetic route from target molecule 134 to the original material.
[0058] In some embodiments, the molecule to be expanded 132 may be determined based on the unresolved molecule set 124. The molecule to be expanded 132 is the molecule to be expanded in the current search round.
[0059] In some embodiments, a target-level implementation may be adopted, in which the target molecule 134 is always identified as the query molecule 130, thereby providing an inventory neighborhood related to the overall target structure before the generation of the retrosynthetic proposal 185 begins.
[0060] In some embodiments, a dynamic implementation may be adopted, in which the molecule to be expanded 132 is identified as the query molecule 130, so that whenever the search exposes a new unresolved intermediate molecule, the search can be re-executed for that intermediate molecule, so that the evidence of the molecule inventory 140 changes with the update of the route search state 120.
[0061] In some embodiments, the target-level implementation and the dynamic implementation described above can be used separately in the same inverse synthesis route planning process. In some embodiments, the target-level implementation and the dynamic implementation can be combined in the same inverse synthesis route planning process.
[0062] In some embodiments, a candidate pool 150 can be obtained from the molecular inventory 140 based on at least one molecular structural association between the query molecule 130 and molecules in the molecular inventory 140. The candidate pool 150 may include multiple inventory candidate records (e.g., inventory candidate record 1, ..., inventory candidate record k). As an example, k inventory candidate records can be selected from the molecular inventory 140, and the inventory candidate records can be sorted according to scalar score or lexicographical order score, retaining a preset number (e.g., k) of inventory candidate records. The following will combine... Figure 2 Describe further implementation details of obtaining the candidate pool from the molecular inventory.
[0063] In some embodiments, multiple inventory candidate records (e.g., inventory candidate record 1, ..., inventory candidate record k) can be constructed as multiple candidate molecule records (e.g., candidate molecule record 1, ..., candidate molecule record m). The following will combine... Figure 3 Describe further implementation details of candidate molecule recording.
[0064] In some embodiments, inventory guidance information 160 can be determined based on multiple candidate molecule records (e.g., candidate molecule record 1, ..., candidate molecule record m). The following will combine... Figure 4 , Figure 5 Further implementation details are described for determining inventory guidance information based on multiple candidate molecule records.
[0065] In some embodiments, route generation input 170 may be constructed based on empirical partial route history 122 of route search state 120, unresolved molecule set 124, query molecule 130 and inventory guidance information 160.
[0066] In some embodiments, a generative model 180 may be used to determine a retrosynthetic proposal 185 based on route generation input 170. As an example, the generative model 180 may include a large language model, a language model fine-tuned from chemical route data, a sequence generation model, or other generative models capable of outputting reaction steps based on conditional information. The following will combine... Figure 6 Further implementation details of inverse synthesis proposal 185 are described.
[0067] In some embodiments, an independent verification process 190 may be performed on the inverse synthesis proposal 185.
[0068] In some embodiments, a verified search action 195 may be determined based on an inverse synthesis proposal 185 obtained through an independent verification process 190.
[0069] In some embodiments, the route search structure 110 may be updated based on the verified search action 195. The inverse synthesis route to be output is generated at least in part based on the updated route search structure. This will be discussed in conjunction with... Figure 8 This describes further implementation details of updating the route search structure 110 based on the verified search action 195.
[0070] Some embodiments move the availability of molecules in the molecular stock from the termination condition after the retrosynthetic route to the route generation input for the generative model to determine the retrosynthetic proposal, thus avoiding the generative model mechanically copying candidate molecule records from the stock as chemically unreasonable precursors.
[0071] Some implementations use a dynamic inventory guidance mechanism that redetermines query molecules and inventory candidates based on new unresolved molecules after the route search status is updated. This reduces ineffective expansion under a fixed search budget and increases the probability of discovering inventory closable routes earlier.
[0072] Some implementations provide molecular structural evidence related to the actual molecular stock before the generative model selects the current break, giving the generative model a better chance to propose precursors that can be linked to the stock.
[0073] Some implementations combine an inventory-driven mechanism with an independent validation process. The generative model only proposes definitive retrosynthetic solutions, while the results of the independent validation process determine whether to use reaction steps to update the route search structure and associated route search states. This decouples the molecular inventory from the training parameters of the generative model. In other words, commercial inventory, lab inventory, or user-specified inventory can be replaced during the route planning phase without retraining the generative model for each inventory change.
[0074] In some embodiments, if the reverse synthesis proposal 185 fails the independent verification process 190, the reverse synthesis proposal 185 may be marked as “rejected” instead of being written into the verified partial route history 122.
[0075] In some embodiments, after updating the route search structure 110 (and consequently updating the route search state 120), it can be determined whether to terminate the search. If not, the next search round continues as described above with the updated route search state. If yes, it can be determined whether the result meets the requirements, and the result can be selectively output based on the determination result. The following will combine... Figure 9 Describe further implementation details of how to determine whether to terminate the search, and combine them with... Figure 10 Describe further implementation details regarding whether the judgment result meets the requirements and selectively output the result.
[0076] Some implementations combine the open output of generative models with deterministic analysis, reaction verification, search state updates, and precise inventory closure to form a complete computational closed loop of "retrieval-generation-verification-update-closure", making the output inverse synthesis route have a verifiable computational basis.
[0077] In some embodiments, the route generation input 170 may also be constructed based on an optional reference route 162. In some embodiments, the route generation input 170 may also be constructed based on an optional output constraint 164.
[0078] Figure 2 This is a schematic diagram illustrating the process of obtaining a candidate pool from a molecular inventory based on the molecular structure association between a queried molecule and molecules in the molecular inventory, according to some embodiments.
[0079] In some embodiments, the query can be based on molecule 130 and molecule inventory (e.g., Figure 1 At least one molecular structural association between molecules in the molecular inventory 140 is used to obtain the candidate pool 150.
[0080] In some embodiments, at least one molecular structure association may include a molecular fingerprint neighborhood 210. As an example, for binary fingerprints a and b, the Tanimoto similarity T(a,b) can be calculated by the following equation (1): T(a,b)=|a∧b| / (|a|+|b|-|a∧b|) (1) Here, "∧" represents a bitwise AND operation.
[0081] As an example, the extended connection fingerprints of query molecule 130 and molecules in the molecule inventory can be calculated, the Tanimoto similarity of the extended connection fingerprints of the two can be calculated, and the molecules in the molecule inventory can be sorted using the calculated Tanimoto similarity.
[0082] In some embodiments, at least one molecular structure association may include a molecular skeleton association 220. As an example, the Bemis–Murcko molecular skeleton (Bemis–Murckoscaffold) of the query molecule 130 and molecules in the molecular inventory may be extracted. As an example, molecules in the molecular inventory with exact skeleton matches may be selected first, followed by molecules in the molecular inventory with substructure associations between their skeletons. As an example, sorting may also be performed based on the fingerprint similarity of the two.
[0083] In some embodiments, at least one molecular structure association may include a substructure association 230. As an example, molecules in the molecular inventory that satisfy either "the query molecule is a substructure of a molecule in the molecular inventory" or "molecules in the molecular inventory are substructures of the query molecule" may be selected and further sorted according to fingerprint similarity.
[0084] In some embodiments, at least one molecular structure association may include a maximum common subgraph association 240. As an example, the maximum common subgraph of the subgraph of the query molecule 130 and the subgraphs of molecules in the molecular inventory may be calculated, and sorted using the node coverage, edge coverage, and fingerprint similarity of the maximum common subgraph as dictionary sorting keys.
[0085] In some embodiments, at least one molecular structure association may include molecular fragment association 250. As an example, the Breaking of Retrosynthetically Interesting Chemical Substructures (BRICS) method can be used to decompose the query molecule and molecules in the molecular inventory into fragments, with the number of shared fragments, the atomic coverage of the shared fragments to the query molecule, and fingerprint similarity forming the sorting criteria.
[0086] In some embodiments, the at least one molecular structure association described above may be used individually. In some embodiments, the at least one molecular structure association described above may be calculated in parallel, and the results may be combined. In some embodiments, the at least one molecular structure association described above may be selected for use alternately. In some embodiments, the at least one molecular structure association described above may be selected for use based on the structural features of the query molecule 130.
[0087] In some embodiments, the candidate pool 150 is also obtained based on a quantity limit 260. This quantity limit 260 indicates the number of inventory candidates to be obtained. In some embodiments, the quantity limit 260 may be determined based on a search budget.
[0088] Figure 3 This is a schematic diagram of the composition of candidate molecule record 300 according to some embodiments.
[0089] In some embodiments, the candidate molecule record 300 may include an identifier 310 for the queried molecule, a canonical structural representation 320 for molecules in the molecule inventory, a candidate type 330, a ranking score 340, and an availability tag 350.
[0090] Figure 4 This is a schematic diagram illustrating the process of determining inventory guidance information using direct projection according to some embodiments.
[0091] In some embodiments, direct projection 410 may be performed on candidate molecule record 300 to obtain compact record line 420. In some embodiments, compact record line 420 may include canonical structural representations 320 of molecules in a molecule inventory, candidate types 330, availability tags 350, and weak guidance indicators 422. Weak guidance indicators 422 are used in subsequent generative models (e.g., Figure 1 Generative model 180 in the middle generates inverse synthesis proposals (e.g., Figure 1 In retrosynthesis proposal 185, the generative model is instructed to use only these molecules as weak guides, i.e., the corresponding molecules in the molecular stock are used only when bond breaking has a chemical basis. In the direct projection implementation, the stock guide information 160 can be determined based on compact record line 420.
[0092] Figure 5 This is a schematic diagram illustrating the process of determining inventory guidance information using summary projection according to some embodiments.
[0093] In some embodiments, summarization processing 510 may be performed on the same candidate molecule records (e.g., candidate molecule record 1, candidate molecule record 2) to determine a summary 520. In some embodiments, the summary 520 may include useful inventory directions 522, key constraints 524, and route generation instructions 526. In a summary projection implementation, inventory guidance information 160 may be determined based on the summary 520.
[0094] In some embodiments, Figure 4 The implementation of direct projection described in [the document] and Figure 5 The implementation of the summary projection described herein can be used individually (e.g., one of them can be used in a single search round). In some embodiments, Figure 4 The implementation of direct projection described in [the document] and Figure 5 The implementations of the summary projection described herein can be used in combination (e.g., using both in a single search round).
[0095] Figure 6 This is a schematic diagram of the composition of a reverse synthesis proposal according to some embodiments.
[0096] In some embodiments, retrosynthesis proposal 185 may include current product molecule 610, reactant molecule 620, current molecule set 630, updated molecule set 640, reaction representation 650, and step description 660.
[0097] In some embodiments, the inverse synthesis proposal 185 may take the form of a list of key-value pairs, a graph structure, or any other structured form that can be deterministically parsed by a parser.
[0098] Figure 7This is a schematic diagram of the composition of an independent verification process 190 according to some embodiments.
[0099] In some embodiments, the standalone validation process 190 may include format and field parsing checks 710.
[0100] In some embodiments, the independent verification process 190 may include a consistency check 720 between the current product and the molecule to be expanded.
[0101] In some embodiments, the independent verification process 190 may include a consistency check 730 of the update of molecular assemblies and route states before and after the reaction.
[0102] In some embodiments, the independent verification process 190 may include molecular structure legitimacy, atom mapping, valence state or stereochemistry checks 740.
[0103] In some embodiments, the independent verification process 190 may include a reaction template suitability check 750.
[0104] In some embodiments, the independent verification process 190 may include a consistency check 760 between the generated reactants and the template execution results.
[0105] In some embodiments, the independent verification process 190 may include a consistency check 770 of inventory identification and precise membership.
[0106] In some embodiments, the independent verification process 190 may include repeated steps, loop routes, or other checks 780 that do not meet preset route rules.
[0107] In some embodiments, the template router may use pattern fingerprint (PatternFP) and asymmetric Tversky similarity index to screen and sort product-side and reactant-side templates, and then the applicability is checked by the symbolic reaction execution tool.
[0108] In some embodiments, if any of the above checks fail, the inverse synthesis proposal can be marked as "rejected" and the inverse synthesis proposal will not be written to the route search state (e.g., Figure 1 Empirical evidence of partial route history (e.g., route search status 120) in the route search status 120 Figure 1 The verified portion of the route history (122) is shown. If all the above checks pass, a verified search action that can be executed in the route search state of the current search round can be formed (e.g., Figure 1 195 verified search actions in the data.
[0109] Figure 8This is a schematic diagram illustrating the process of updating the route search structure based on verified search actions, according to some embodiments.
[0110] In some embodiments, a verified search action 195 may be inserted into a route search structure 110.
[0111] In some embodiments, successfully expanded molecules 132 may be removed from the route search structure 110.
[0112] In some embodiments, reactants 810 that do not meet the termination search criteria may be added to the unresolved molecule set 124 of the route search state 120. As an example, this can be achieved by adding reactants that do not meet the termination search criteria to the route search structure 110, and updating the unresolved molecule set 124 of the route search state 120 as the route search structure 110 is updated.
[0113] Figure 9 This is a flowchart illustrating the process of determining whether to terminate the search, based on some embodiments.
[0114] In some embodiments, it can be determined whether the search budget has been exhausted.
[0115] In some embodiments, the search can be terminated in response to determining that the search budget has been exhausted.
[0116] In some embodiments, in response to determining that the search budget has not been exhausted, it can be determined whether the search route is closed.
[0117] In some embodiments, the search can be terminated in response to determining that the search route is closed.
[0118] In some embodiments, in response to determining that the search path is not closed, the molecule to be expanded can be reselected to perform the next search round.
[0119] Figure 10 This is a flowchart of the output process of inverse synthesis route planning according to some embodiments.
[0120] In some embodiments, it can be determined whether each terminal molecule in the updated route search structure exists in the molecule inventory.
[0121] In some embodiments, in response to the result of the determination indicating that not every terminal molecule exists in the molecule inventory in the updated route search structure, a partial route in an unclosed state may be output.
[0122] In some embodiments, in response to the determination result indicating that each terminal molecule in the updated route search structure exists in the molecular stock, it can be determined whether the reaction steps in the retrosynthetic route in the updated route search structure have all passed the independent verification process.
[0123] In some embodiments, in response to the result of the determination indicating that the reaction steps in the inverse synthesis route in the updated route search structure have not all passed the independent verification process, the verified partial routes may be output.
[0124] In some embodiments, in response to the result of the determination indicating that the reaction steps in the inverse synthesis route in the updated route search structure have all passed the independent verification process, the inverse synthesis route (i.e., the complete route) can be output.
[0125] According to one aspect of the present invention, a method for inverse synthesis route planning is provided.
[0126] Figure 11 This is a flowchart of a method for inverse synthesis route planning according to some embodiments. This method can be combined with the above. Figure 1 The process described corresponds to inverse synthesis route planning, but the scope of the invention is not limited thereto.
[0127] In some embodiments, the method may include step S1: obtaining the target molecule, a molecular inventory, and the route search state associated with the route search structure for the current search round. The route search structure may include a retrosynthetic route from the target molecule to the original material. The route search state may include a validated partial route history and an unresolved set of molecules.
[0128] In some embodiments, the method may include step S2: identifying the target molecule or the molecule to be expanded based on the unresolved molecule set as the query molecule.
[0129] In some embodiments, the method may include step S3: obtaining a candidate pool from the molecular inventory based on at least one molecular structural association between the query molecule and molecules in the molecular inventory. The candidate pool may include multiple inventory candidate records.
[0130] In some embodiments, the method may include step S4: constructing multiple inventory candidate records into multiple candidate molecule records. Each candidate molecule record may include an identifier of the queried molecule, a canonical structural representation of the molecule in the molecule inventory, a candidate type, a ranking score, and an availability tag.
[0131] In some embodiments, the method may include step S5: determining inventory guidance information based on multiple candidate molecule records.
[0132] In some embodiments, the method may include step S6: constructing route generation input based on verified partial route history, unresolved molecule set, query molecules, and inventory guidance information.
[0133] In some embodiments, the method may include step S7: using a generative model to determine an inverse synthesis proposal based on route generation input.
[0134] In some embodiments, the method may include step S8: performing an independent verification process on the inverse synthesis proposal.
[0135] In some embodiments, the method may include step S9: determining a verified search action based on an inverse synthesis proposal obtained through an independent verification process.
[0136] In some embodiments, the method may include step S10: updating the route search structure based on verified search actions. The inverse synthetic route to be output is generated at least in part based on the updated route search structure.
[0137] In some embodiments, step S3 may include step S31: obtaining a candidate pool from the molecular inventory based on at least one of the following molecular structure associations between the queried molecule and the molecules in the molecular inventory: molecular fingerprint neighborhood, molecular skeleton association, substructure association, maximum common subgraph association, and molecular fragment association.
[0138] In some embodiments, a retrosynthetic proposal may include: current product molecule, reactant molecule, current set of molecules, updated set of molecules, reaction representation, and step description.
[0139] In some embodiments, the independent verification process may include: format and field parsing checks; consistency checks between the current product and the molecule to be expanded; consistency checks between the molecular set and route state updates before and after the reaction; checks on the legality of molecular structure, atomic mapping, valence state, or stereochemistry; checks on the suitability of the reaction template; consistency checks between the generated reactants and the template execution results; consistency checks on the inventory identifiers and precise membership relationships; and checks on repeated steps, loop routes, or other routes that do not meet the preset route rules.
[0140] Figure 12 This is a flowchart of a first process associated with a method for inverse synthetic route planning according to some embodiments. This first process may be... Figure 11 The specific implementation of step S4 in the method, and can be combined with the above. Figure 4 , Figure 5 The process described corresponds to this, but the scope of the present invention is not limited thereto.
[0141] In some embodiments, the first process may include step S41: performing direct projection on the candidate molecule record to obtain a compact record line. The compact record line may include the canonical structural representation of the molecules in the molecule inventory, candidate type, availability tag, and weak guidance indicator. The weak guidance indicator is used to indicate that the corresponding molecule in the molecule inventory should only be used if bond breaking has a chemical basis. Inventory guidance information is determined based on the compact record line.
[0142] In some embodiments, the first process may include step S42: performing summarization processing on the same candidate molecule records in the candidate molecule records to determine a summary. The summary may include useful inventory directions, key constraints, and route generation instructions. Inventory guidance information is determined based on the summary.
[0143] Figure 13 This is a flowchart of a second process associated with a method for inverse synthetic route planning, according to some embodiments. This second process may be... Figure 11 The specific implementation method of step S10 in the method, and can be combined with the above. Figure 8 The process described corresponds to this, but the scope of the present invention is not limited thereto.
[0144] In some embodiments, the second process may include step S101: inserting the verified search action into the route search structure.
[0145] In some embodiments, the second process may include step S102: removing the successfully expanded molecule from the route search structure.
[0146] In some embodiments, the second process may include step S103: adding reactants that do not meet the conditions for terminating the search to the unresolved molecule set.
[0147] Figure 14 This is a flowchart of a third process associated with a method for inverse synthetic route planning, according to some embodiments. This third process may be... Figure 11 Additional steps in the method described.
[0148] In some embodiments, the third process may include step S11: determining whether to terminate the search.
[0149] In some embodiments, the third process may include step S12: in response to determining that the search is not terminated, performing the next search round with the route search state associated with the updated route search structure (e.g., performing...). Figure 11 Steps S1-S10 of the method in the text.
[0150] In some embodiments, the third process may include step S13: in response to determining to terminate the search, determining whether the inverse synthesis route in the updated route search structure satisfies the conditions.
[0151] In some embodiments, the third process may include step S14: selectively outputting an inverse synthesis route or a portion thereof based on the determination result of whether the inverse synthesis route in the updated route search structure meets the conditions.
[0152] Figure 15 This is a flowchart of a fourth process associated with a method for inverse synthetic route planning, according to some embodiments. This fourth process may be... Figure 14 The specific implementation of step S11 in the third process is described, but the scope of the present invention is not limited thereto.
[0153] In some embodiments, the fourth process may include step S111: determining whether the search budget has been exhausted.
[0154] In some embodiments, the fourth process may include step S112: terminating the search in response to determining that the search budget has been exhausted.
[0155] In some embodiments, the fourth process may include step S113: determining whether the search route is closed in response to determining that the search budget has not been exhausted.
[0156] In some embodiments, the fourth process may include step S114: terminating the search in response to determining that the search route is closed.
[0157] In some embodiments, the fourth process may include step S115: in response to determining that the search route is not closed, reselecting the molecule to be expanded.
[0158] Figure 16 This is a flowchart of a fifth process associated with a method for inverse synthetic route planning, according to some embodiments. This fifth process may be... Figure 14 The specific implementation of step S13 in the third process is described, but the scope of the present invention is not limited thereto.
[0159] In some embodiments, the fifth process may include step S131: determining whether each terminal molecule in the updated route search structure exists in the molecule inventory.
[0160] In some embodiments, the fifth process may include step S132: determining whether all reaction steps in the retrosynthetic route in the updated route search structure have passed the independent verification process.
[0161] Figure 17 This is a flowchart of a sixth process associated with a method for inverse synthetic route planning, according to some embodiments. This sixth process may be... Figure 14 The specific implementation of step S14 in the third process is described, but the scope of the present invention is not limited thereto.
[0162] In some embodiments, the sixth process may include step S141: in response to the result of the determination indicating that not every terminal molecule exists in the molecule inventory in the updated route search structure, outputting the partially unclosed route.
[0163] In some embodiments, the sixth process may include step S142: in response to the result of the determination indicating that the reaction steps in the retrosynthetic route have not all passed the independent verification process, outputting the verified partial route.
[0164] In some embodiments, the sixth process may include step S143: in response to the result of the determination indicating that each terminal molecule in the updated route search structure exists in the molecular inventory and that the reaction steps in the retrosynthetic route have passed the independent verification process, outputting the retrosynthetic route in the route search structure for the current search round.
[0165] In some embodiments, Figure 16 The steps of the fifth process and Figure 17 The steps of the sixth process can be combined with the above text. Figure 10 The described logic is executed in combination.
[0166] According to another aspect of the present invention, a computer-readable storage medium is provided.
[0167] Figure 18 This is a block diagram of a computer-readable storage medium 1800 according to some embodiments.
[0168] In some embodiments, the computer-readable storage medium 1800 may include a computer program 1850 stored thereon. When executed by a processor, the computer program 1850 implements the above-described combination. Figures 1-17 The methods or processes described.
[0169] According to another aspect of the present invention, a computer program product is provided.
[0170] Figure 19 This is a block diagram of a computer program product 1900 according to some embodiments.
[0171] In some embodiments, computer program product 1900 may include computer program 1850. When executed by a processor, computer program 1850 implements the above-described combination. Figures 1-17 The methods or processes described.
[0172] The technical effects of the embodiments of the present invention were verified through experiments.
[0173] First, the experimental results of different inventory usage methods were validated. These methods were distinguished by the different entry positions of inventory guidance information during the inverse synthesis route planning process. Three control groups were included in the experiment: the basic scheme, reference route selection based on Tanimoto similarity, and reference route rearrangement based on terminal inventory hits. Validation parameters included the number of solved objectives per 100 objectives (“solved objectives / 100”), the proportion of inventory at leaf nodes, the proportion of inventory-closed routes, and the proportion of effective steps. The experiment used a generative model fine-tuned with approximately 100,000 route samples, and was compared on the inverse synthesis test set with a maximum of 20 iterations. Table 1 below shows the experimental results for different entry positions of inventory guidance information.
[0174]
[0175] The experimental results above show that, under the same post-validation and precise inventory closure rules, all indicators are better than those of the case where inventory guidance information is provided directly in the proposal generation stage, which only uses inventory to select reference routes or performs post-rearrangement of reference routes.
[0176] Secondly, to further verify the above experimental conclusions, a control inventory record experiment was conducted. In the control experiment, which maintained consistency with the generative model and other conditions, the random inventory records retained the same recording format but removed the structural relevance to the target molecule, while the non-inventory similar records retained structural proximity but removed inventory availability. Under the budgets of 10, 20, 50, and 100 iterations on the two test sets, the overall change in the solved objective relative to the base scheme for both control conditions was close to zero or negative, failing to reproduce the significant gain brought by target-related inventory candidates.
[0177] The results of this control experiment demonstrate that the aforementioned effect does not solely stem from adding several molecular representations to the model input, but is related to the combination of "selecting relevant records from the molecular inventory for the queried molecule and providing them before generation."
[0178] Embodiments of the invention have been described with reference to the accompanying drawings. These embodiments are illustrative and not restrictive.
Claims
1. A method for inverse synthesis route planning, characterized in that, Includes the following steps: S1: Obtain the target molecule, molecule inventory, and route search status associated with the route search structure of the current search round. The route search structure includes retrosynthetic routes from the target molecule to the original material. The route search status includes a verified partial route history and an unresolved molecule set. S2: The target molecule or the molecule to be expanded based on the unresolved molecule set is determined as the query molecule; S3: Based on at least one molecular structure association between the queried molecule and the molecules in the molecular inventory, a candidate pool is obtained from the molecular inventory, the candidate pool including multiple inventory candidate records; S4: Construct the plurality of inventory candidate records into a plurality of candidate molecule records, each candidate molecule record including the identifier of the queried molecule, the canonical structural representation of the molecules in the molecule inventory, the candidate type, the ranking score and the availability tag; S5: Determine inventory guidance information based on the multiple candidate molecule records; S6: Construct route generation input based on the verified partial route history, the unresolved molecule set, the queried molecule, and the inventory guidance information; S7: Utilize a generative model to generate inputs based on the route to determine inverse synthesis proposals; S8: Perform an independent verification process on the inverse synthesis proposal; S9: Determine the verified search action based on the inverse synthesis proposal obtained through the independent verification process; as well as S10: Update the route search structure based on the verified search action, wherein the inverse synthesis route to be output is generated at least in part based on the updated route search structure.
2. The method as described in claim 1, characterized in that, S3 includes: S31: Based on at least one of the following molecular structure associations between the queried molecule and the molecules in the molecular inventory, a candidate pool is obtained from the molecular inventory: molecular fingerprint neighborhood, molecular skeleton association, substructure association, maximum common subgraph association, and molecular fragment association.
3. The method as described in claim 2, characterized in that, The candidate pool is also obtained based on a quantity limit, which indicates the number of inventory candidates to be obtained.
4. The method as described in claim 1, characterized in that, S4 includes: S41: Perform direct projection on the candidate molecule record to obtain a compact record line, the compact record line including the canonical structural representation of the molecules in the molecule inventory, the candidate type, the availability tag, and a weak guidance indicator, the weak guidance indicator being used to indicate that the corresponding molecule in the molecule inventory is used only when there is a chemical basis for bond breaking, wherein the inventory guidance information is determined based on the compact record line.
5. The method as described in claim 1, characterized in that, S4 includes: S42: Perform summary processing on the same candidate molecule records in the candidate molecule records to determine a summary, the summary including useful inventory directions, key constraints and route generation instructions, wherein the inventory guidance information is determined based on the summary.
6. The method as described in claim 1, characterized in that, The retrosynthesis proposal includes: current product molecule, reactant molecule, current set of molecules, updated set of molecules, reaction representation, and step description.
7. The method as described in claim 1, characterized in that, The independent verification process includes: format and field parsing checks; consistency checks between the current product and the molecule to be expanded; consistency checks between the molecular set and route state updates before and after the reaction; checks on the legality of molecular structure, atomic mapping, valence state, or stereochemistry; checks on the applicability of the reaction template; consistency checks between the generated reactants and the template execution results; consistency checks on the inventory identifiers and precise membership relationships; and checks on repeated steps, loop routes, or other routes that do not meet the preset route rules.
8. The method as described in claim 1, characterized in that, S10 includes: S101: Insert the verified search action into the route search structure; S102: Remove the successfully expanded molecule from the route search structure; and S103: Add reactants that do not meet the conditions for terminating the search to the set of unresolved molecules.
9. The method as described in claim 1, characterized in that, Also includes: S11: Determine whether to terminate the search.
10. The method as described in claim 9, characterized in that, S11 includes: S111: Determine if the search budget has been exhausted; S112: Terminate the search in response to determining that the search budget has been exhausted.
11. The method as described in claim 10, characterized in that, S11 further includes: S113: In response to determining that the search budget has not been exhausted, determine whether the search route is closed; S114: In response to determining that the search route is closed, terminate the search; S115: In response to the determination that the search route is not closed, a new molecule to be expanded is selected.
12. The method as described in claim 9, characterized in that, Also includes: S12: In response to determining that the search will not be terminated, execute S1 to S10 with the route search state associated with the updated route search structure.
13. The method as described in claim 9, characterized in that, Also includes: S13: In response to determining the termination of the search, determine whether the inverse synthesis route in the updated route search structure satisfies the condition; as well as S14: Based on the determination result of whether the inverse synthesis route in the updated route search structure meets the conditions, selectively output the inverse synthesis route or a part of the inverse synthesis route.
14. The method as described in claim 13, characterized in that, S13 includes: S131: Determine whether each terminal molecule in the updated route search structure exists in the molecule inventory; and S132: Determine whether all reaction steps in the retrosynthetic route in the updated route search structure have passed the independent verification process.
15. The method as described in claim 14, characterized in that, S14 includes: S141: In response to the result of the determination indicating that not every terminal molecule exists in the molecule inventory in the updated route search structure, output the partial route in the unclosed state; S142: In response to the result of the determination indicating that not all reaction steps in the retrosynthetic route have passed the independent verification process, output the verified partial route; S143: In response to the result of the determination indicating that each terminal molecule in the updated route search structure exists in the molecule inventory and that the reaction steps in the retrosynthetic route have all passed the independent verification process, the retrosynthetic route in the route search structure for the current search round is output.
16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-15.
17. A computer program product, comprising a computer program, characterized in that, When the computer program is executed, it performs the steps of the method according to any one of claims 1-15.