DETAILED ACTION
The applicant’s response, from 07 July 2026, has been fully considered. Amendments to the claims, from 07 July 2026, were received and entered. The following rejections and/or objections are either reiterated or newly applied. They constitute the complete set presently being applied to the instant application.
This action is NON-FINAL.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Status
Claims 4 and 16 are cancelled.
Claims 1-3, 5-15, and 17-20 are currently pending and under examination herein.
Claims 1-3, 5-15, and 17-20 are rejected.
Priority
The instant application claims priority as a continuation of PCT/CN2022/073158 filed 21 January 2022 and foreign priority to CN202110112207.8 filed 24 January 2021. A certified copy of CN202110112207.8 was not found within the filed documents. In this action, claims 1-20 are examined as though they had an effective filing date of 24 January 2021. In future actions, the effective filing date of one or more claims may change, due to amendments to the claims, or further analysis of the disclosure(s) of the priority application(s).
Information Disclosure Statement
The information disclosure statements (IDS) submitted on 10/05/2022 and 04/18/2024 are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Drawings
The drawings filed on 10/05/2022 are accepted.
Claim Rejections - 35 USC § 112
The previously issued 112(b) rejection is withdrawn due to the amendments to claims 3 and 15. Particularly, the amended claims removed the antecedent basis issues.
Claim Rejections - 35 USC § 101
Arguments associated with the previously issued 35 USC 101 rejection are considered unpersuasive (see response to arguments below the rejection). The following rejection is reiterated and modified as has been necessitated by amendment.
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-3, 5-15, and 17-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. In accordance with MPEP 2106, claims found to recite statutory subject matter (Step 1: YES) are then analyzed to determine if the claims recite any concepts that equate to an abstract idea or natural law (Step 2A, Prong 1). Claims 1-3 and 5-12 are directed to a method and Claims 13-15 and 17-20 are directed to systems. In the instant application, the claims recite the following limitations that equate to an abstract idea:
Clain 1 recites the limitation - obtaining a target molecule and determining the target molecule as a root node in a tree structure, the root node being associated with a first leaf node in the tree structure, and the tree structure comprising a retrosynthetic path of the target molecule; recursively processing the predicted molecule set corresponding to the second leaf nodes and determining a terminal node that satisfies a preset condition; traversing path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule, obtaining a first training molecule and a first training synthon corresponding to the first training molecule; determining a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon. Based on the broadest reasonable interpretation, generating molecules, predicting retrosynthesis using a tree structure, and comparing information to a condition could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites expanding the first leaf node through a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position; determining a first loss function based on the node feature and the edge feature; determining a bond breaking probability of the chemical bond between the first training molecule and the first training synthon; determining a second loss function based on the bond breaking probability; and updating a model parameter of the graph neural network according to the first loss function and the second loss function. Based on the broadest reasonable interpretation, using neural and non-neural networks to predict atoms and bonds, and determining functions and probabilities encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 2 recites the limitation - determining a first string of the compound molecule corresponding to the first leaf node; filtering the at least one synthon based on a preset rule to determine the predicted molecule set. Based on the broadest reasonable interpretation, determining a string and filtering information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites converting the first string into a first molecular map; determining the bond breaking position of the compound molecule corresponding to the first leaf node according to the first molecular map through the graph neural network; and determining at least one synthon according to the bond breaking position through the reactant generation network. Based on the broadest reasonable interpretation, converting strings, determining positions with a graph neural network, and determining compounds with a reactant generation network encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 3 recites the limitation - determining bond breaking information according to the first molecular map and a limit on a number of broken bonds through the graph neural network. Based on the broadest reasonable interpretation, determining information with the neural network encompasses equations. This draws the limitation to a mathematical concept, which classifies the limitation as an abstract idea. The claim also recites determining a target key feature according to the target molecule, the target key feature comprising an atom key feature and a bond key feature, the atom key feature comprising at least one of an atom type, a number of bonds, a formal charge, chirality, a number of hydrogen atoms, an atomic hybridization state, aromaticity, an atomic weight, a high frequency reaction center feature, and a reaction type, and the bond key feature comprising at least one of a bond type, a conjugate, a ring bond, and a molecular stereo chemical feature; and extracting key breaking information based on the target key feature to determine the bond breaking position. Based on the broadest reasonable interpretation, determining the information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Claim 5 recites the limitation - splitting the target molecule based on the bond breaking position to obtain at least one synthon molecular map; converting the synthon molecular map into a second string; and determining the at least one synthon according to the third string through the reactant generation network. Based on the broadest reasonable interpretation, the splitting, converting, and determining information with the reactant generation network encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea. The claims also recites updating the second string based on a preset reaction type to obtain a third string. Based on the broadest reasonable interpretation, updating the string could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Claim 6 recites the limitation - obtaining a second training molecule, a second training synthon, and a training reactant; determining a first training string based on a string corresponding to the second training molecule, a string corresponding to the second training synthon, and the preset reaction type; determining a second training string corresponding to the training reactant; and associating the first training string and the second training string to determine a first training sample pair. Based on the broadest reasonable interpretation, determining and associating information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites training the reactant generation network based on the first training sample pair. Based on the broadest reasonable interpretation, training a generic machine learning model encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 7 recites the limitation - obtaining a candidate string predicted by the graph neural network. Based on the broadest reasonable interpretation, predicting information with a neural network encompasses equations. This draws the limitation to a mathematical concept, which classifies the limitation as an abstract idea. The claim also recites adding the candidate string into the string corresponding to the second training synthon to update the first training string to a third training string; training the reactant generation network based on the second training sample pair. Based on the broadest reasonable interpretation, adding information and training a generic machine learning model encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea. The claim also recites associating the third training string and the second training string to determine a second training sample pair. Based on the broadest reasonable interpretation, associating information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Claim 8 recites the limitation - determining a target character format; and updating the first training string based on the target character format. Based on the broadest reasonable interpretation, determining information and updating a string could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Claim 9 recites the limitation - determining a first candidate molecule to which the predicted molecule set corresponding to the second leaf node corresponds under a preset reaction type; and determining that a leaf node corresponding to the second candidate molecule is the terminal node in response to the second candidate molecule satisfying the preset condition. Based on the broadest reasonable interpretation, determining information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites expanding the first candidate molecule through the target retrosynthesis model to obtain third leaf nodes; and recursively processing the predicted molecule set corresponding to the third leaf nodes to determine a second candidate molecule. Based on the broadest reasonable interpretation, using the models to predict retrosynthesis encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 10 recites the limitation - traversing based on the second candidate molecule to obtain a plurality of pathway molecules; and determining that the second candidate molecule satisfies the preset condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node, in response to the pathway molecule being a molecule in a basic molecule set. Based on the broadest reasonable interpretation, obtaining and comparing information could practically be done by the human mind. This draws the limitation a mental process, which classifies the limitation as an abstract idea.
Claim 11 recites the limitation - determining a number of expansions corresponding to the second candidate molecule in the tree structure; and determining that the second candidate molecule satisfies the preset condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node, in response to the number of expansions reaching a preset value. Based on the broadest reasonable interpretation, determining and comparing values encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 12 recites the limitation - wherein the target molecule is a drug molecule, the tree structure is a Monte Carlo tree, and a compound molecule corresponding to the second leaf node is obtained by screening based on at least one of functional group interference, use of a protective group, a reaction feature of a functional group, and a name reaction rule. This limitation specifies the information determined in the judicial exception of claim 1. The refined limitation indicated still represents a judicial expectation.
Claim 13 recites the limitation - obtaining code, configured to cause the at least one processor to obtain a target molecule and determine the target molecule as a root node in a tree structure, the root node being associated with a first leaf node in the tree structure, and the tree structure comprising a retrosynthetic path of the target molecule; recursively process the predicted molecule set corresponding to the second leaf nodes and determine a terminal node that satisfies a preset condition; traverse path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule; obtain a first training molecule and a first training synthon corresponding to the first training molecule; and determine a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon. Based on the broadest reasonable interpretation, generating information on molecules, predicting retrosynthesis using a tree structure, and comparing information to a condition could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites expand the first leaf node through a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position; determine a first loss function based on the node feature and the edge feature; determine a bond breaking probability of the chemical bond between the first training molecule and the first training synthon; determine a second loss function based on the bond breaking probability; and update a model parameter of the graph neural network according to the first loss function and the second loss function. Based on the broadest reasonable interpretation, using neural and non-neural networks to predict atoms and bonds, and determining functions and probabilities encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 14 recites the limitation - determine a first string of the compound molecule corresponding to the first leaf node; and filter the at least one synthon based on a preset rule to determine the predicted molecule set. Based on the broadest reasonable interpretation, determining a string and filtering information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites convert the first string into a first molecular map; determine the bond breaking position of the compound molecule corresponding to the first leaf node according to the first molecular map through the graph neural network; and determine at least one synthon according to the bond breaking position through the reactant generation network. Based on the broadest reasonable interpretation, converting strings, determining positions with a graph neural network, and determining compounds with a reactant generation network encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 15 recites the limitation - determine bond breaking information according to the first molecular map and a limit on a number of broken bonds through the graph neural network. Based on the broadest reasonable interpretation, determining information with the neural network encompasses equations. This draws the limitation to a mathematical concept, which classifies the limitation as an abstract idea. The claim also recites determine a target key feature according to the target molecule, the target key feature comprising an atom key feature and a bond key feature, the atom key feature comprising at least one of an atom type, a number of bonds, a formal charge, chirality, a number of hydrogen atoms, an atomic hybridization state, aromaticity, an atomic weight, a high frequency reaction center feature, and a reaction type, and the bond key feature comprising at least one of a bond type, a conjugate, a ring bond, and a molecular stereo chemical feature; and extract key breaking information based on the target key feature to determine the bond breaking position. Based on the broadest reasonable interpretation, determining the information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Claim 17 recites the limitation - split the target molecule based on the bond breaking position to obtain at least one synthon molecular map; convert the synthon molecular map into a second string; and determine the at least one synthon according to the third string through the reactant generation network. Based on the broadest reasonable interpretation, the splitting, converting, and determining information with the reactant generation network encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea. The claim also recites update the second string based on a preset reaction type to obtain a third string. Based on the broadest reasonable interpretation, updating the string could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Claim 18 recites the limitation - obtain a second training molecule, a second training synthon, and a training reactant; determine a first training string based on a string corresponding to the second training molecule, a string corresponding to the second training synthon, and the preset reaction type; determine a second training string corresponding to the training reactant; and associate the first training string and the second training string to determine a first training sample pair. Based on the broadest reasonable interpretation, determining and associating information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites train the reactant generation network based on the first training sample pair. Based on the broadest reasonable interpretation, training a generic machine learning model encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
Claim 19 recites the limitation - obtain a candidate string predicted by the graph neural network. Based on the broadest reasonable interpretation, predicting information with a neural network encompasses equations. This draws the limitation to a mathematical concept, which classifies the limitation as an abstract idea. The claim also recites add the candidate string into the string corresponding to the second training synthon to update the first training string to a third training string; and train the reactant generation network based on the second training sample pair. Based on the broadest reasonable interpretation, adding information and training a generic machine learning model encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea. The claim also recites associate the third training string and the second training string to determine a second training sample pair. Based on the broadest reasonable interpretation, associating information could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea.
Clain 20 recites the limitation - obtain a target molecule and determining the target molecule as a root node in a tree structure, the root node being associated with a first leaf node in the tree structure, and the tree structure comprising a retrosynthetic path of the target molecule; recursively process the predicted molecule set corresponding to the second leaf nodes and determining a terminal node that satisfies a preset condition; traverse path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule, obtain a first training molecule and a first training synthon corresponding to the first training molecule; determine a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon. Based on the broadest reasonable interpretation, generating molecules, predicting retrosynthesis using a tree structure, and comparing information to a condition could practically be done by the human mind. This draws the limitation to a mental process, which classifies the limitation as an abstract idea. The claim also recites expand the first leaf node through a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position; determine a first loss function based on the node feature and the edge feature; determine a bond breaking probability of the chemical bond between the first training molecule and the first training synthon; determining a second loss function based on the bond breaking probability; and update a model parameter of the graph neural network according to the first loss function and the second loss function. Based on the broadest reasonable interpretation, using neural and non-neural networks to predict atoms and bonds, and determining functions and probabilities encompasses equations and could practically be done by the human mind. This draws the limitation to a mathematical concept and a mental process, which classifies the limitation as an abstract idea.
These limitations recite concepts of determining, predicting, comparing, and identifying information and applying or training machine learning algorithms that are so generically recited that they can be practically performed in the human mind as claimed, which falls under the “Mental processes” and “Mathematical concepts” grouping of abstract ideas. A mathematical concept need not be expressed in mathematical symbols, because words used in a claim operating on data to solve a problem can serve the same purpose as a formula (MPEP 2106.04(a)(2)). Additionally, both product claims and process claims may recite mental processes, which can include a claim that requires a computer (MPEP 2106.04(a)(2)). Therefore, these limitations fall under the “Mental process” and “Mathematical concepts” groupings of abstract ideas. As such, claims 1-3, 5-15, and 17-20 recite an abstract idea (Step 2A, Prong 1: YES).
Claims found to recite a judicial exception under Step 2A, Prong 1 are then further analyzed to determine if the claims as a whole integrate the recited judicial exception into a practical application or not (Step 2A, Prong 2). These judicial exceptions are not integrated into a practical application because the claims do not recite an additional element that reflects an improvement to technology (MPEP § 2106.04(d)(1)). Rather, the claims provide insignificant extra-solution activity (MPEP § 2106.05(g)) and provide mere instructions to apply a judicial exception (MPEP § 2106.05(f)). Specifically, the claims recite the following additional elements:
Claim 1 recites a computer device.
Claim 13 recites memory, a processor, and code.
Claim 20 recites a non-transitory computer-readable storage medium, computer program, and at least one processor.
There are no limitations that indicate that the claimed determining, predicting, comparing, and identifying information and applying or training machine learning algorithms require anything other than generic computing systems. As such, these limitations equate to mere instructions to implement the abstract idea on a generic computer that the courts have stated does not render an abstract idea eligible in Alice Corp., 573 U.S. at 223, 110 USPQ2d at 1983. See also 573 U.S. at 224, 110 USPQ2d at 1984. There is no indication that these steps are affected by the judicial exception in any way and thus do not integrate the recited judicial exception into a practical application. As such, claims 1-3, 5-15, and 17-20 are directed to an abstract idea (Step 2A, Prong 2: NO).
Claims found to be directed to a judicial exception are then further evaluated to determine if the claims recite an inventive concept that provides significantly more than the judicial exception itself (Step 2B). The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception because the claims recite conventional additional elements that equate to mere instructions to apply the recited exception in a generic way or in a generic computing environment. The claims also recite conventional additional elements that represent insignificant extra-solution activities.
As discussed above, there are no additional limitations to indicate that the claimed determining, predicting, comparing, and identifying information and applying or training machine learning algorithms require anything other than generic computer components in order to carry out the recited abstract idea in the claims. Claims that amount to nothing more than an instruction to apply the abstract idea using a generic computer do not render an abstract idea or natural law eligible. Alice Corp., 573 U.S. at 223, 110 USPQ2d at 1983. See also 573 U.S. at 224, 110 USPQ2d at 1984. MPEP 2106.05(f) discloses that mere instructions to apply the judicial exception cannot provide an inventive concept to the claims. As specified in MPEP 2106.05(g), extra-solution activities can be understood as incidental to the primary process or product that are merely a nominal or tangential addition to the claim. Insignificant extra-solution activities include mere data gathering, selecting a particular data source or type of data to be manipulated, and displaying information. Additionally, Feng et al. (2018, Frontiers in Chemistry, Vol. 6: 1-10, cited by previous action) teaches that the use of neural networks, including graph based neural networks, have become standard in the field of synthesis and retrosynthesis of complex molecules (Page 7, Column 1, Paragraph 2: In recent years, deep neural networks have been applied to this field. One characteristic feature is that computers do not need to follow human-defined reaction rules, and instead, they can recomprehend chemical reactions with only millions of reaction examples).
The additional elements do not comprise an inventive concept when considered individually or as an ordered combination that transforms the claimed judicial exception into a patent-eligible application of the judicial exception. Therefore, the claims do not amount to significantly more than the judicial exception itself (Step 2B: No). As such, Claims 1-3, 5-15, and 17-20 are not patent eligible.
Response to arguments
Applicant asserts the independent claims do not recite judicial exceptions (Page 14, Paragraph 4 of remarks). As specified in the rejection above, the independent claims recite at least obtaining data within a tree structure (can be done mentally with pen and paper), predicting retrosynthesis paths (can be done mentally with pen and paper), using generic machine leaning models (requires equations), and determining mathematical functions and probabilities (requires equations).
The instant claims to not recite steps that require analysis of network information as indicated by the cited court case (Page 14, Paragraph 4 of remarks: See SRI Int'l, Inc. v. Cisco Systems, Inc., 930 F.3d 1295, 1304 (Fed. Cir. 2019) (declining to identify the claimed collection and analysis of network data as abstract because "the human mind is not equipped to detect suspicious activity by using network monitors and analyzing network packets as recited by the claims")).
Applicant further asserts the claims cannot be done mentally (Page 14, Paragraph 4 of remarks: the processes draw on reactions that are infinitely variable. See Published Specification, [0003].) However, the scope of this instant claims do not establish any limits on the type of molecules considered. The BRI is no way limited to the types of molecules that can be analyzed by these methods. This allows small/simple molecules that could easily be predicted by hand (few synthetic steps needed).
Applicant asserts the claims do not recite math (Page 14, Paragraph 4 of remarks: it is well settled, that a claim does not recite a mathematical concept (i.e., the claim limitations do not fall within the mathematical concept grouping), if it is only based on or involves a mathematical concept. See, e.g., Thales Visionix, Inc. v. United States, 850 F.3d 1343, 1348-49, 121 USPQ2d 1898, 1902-03 (Fed. Cir. 2017) (determining that the claims to a particular configuration of inertial sensors and a particular method of using the raw data from the sensors in order to more accurately calculate the position and orientation of an object on a moving platform did not merely recite "the abstract idea of using 'mathematical equations for determining the relative position of a moving object to a moving reference frame'.")). Unlike the cited court decision, the instant claims recite determining multiple loss functions and probabilities. These limitations rely on and cannot be formed without mathematical. A mathematical concept need not be expressed in mathematical symbols, because words used in a claim operating on data to solve a problem can serve the same purpose as a formula (MPEP 2106.04(a)(2)). Additionally, machine learning networks, such a neural networks are a collections of data processing steps, which require mathematical equations to process data. Utilizing these machine learning models therefore constitute relying upon mathematical equations.
Applicant asserts example 39 indicates that training a neural network is an additional element (Page 16, last paragraph of the remarks). Example 39 indicates generically training a neural network on image data is not a judicial exception. The instant claims recites calculating loss functions and probabilities as part of the training. Therefore, the instant claims appear more similar to the updated subject matter eligibility example 47 (claim 2) which is ineligible because it recites training a neural network with specified equations (back propagation) to perform a mental process. The instant claims recites training a neural network by calculating loss functions and probability to predict retrosynthesis, which is a mental process explained above. The claims do not use the neural network to perform a task that cannot be a mental process (as in detect anomalies in network traffic (claim 3)). Therefore, the subject matter eligibility example most related indicate the claims are ineligible.
Applicant asserts the claims represent an improvement to technology. Applicant cites the Ex Parte Desjardins decision which addresses the issue of “catastrophic forgetting” encountered in continual learning systems (Page 17, Paragraph 1 of remarks). Loss functions are a common elements of machine learning. Furthermore, they are open ended (i.e., not inherently limited to a certain structure). Therefore, the issue solved by two loss functions in not a problem inherent to loss functions or machine learning as a field because one function can be composed of any number of terms (as indicated by the cited art) to address limitations of the loss function. Therefore, the issue of “escape current local optima” is not analogous to “catastrophic forgetting” or a pervasive issue within the field of machine learning. This indicates the instant claims do not represent an improvement to a technology or field of technology, at least in terms of utilizing two loss functions. The arguments are therefore deemed unpersuasive.
Claim Rejections - 35 USC § 103
Some arguments associated with the previously issued 35 USC 103 rejection are considered persuasive (see response to arguments below the rejection). The following rejection is reiterated and modified as has been necessitated by amendment and previous deficiencies. This action is NON-FINAL.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
The following rejection presents a new ground of rejection.
Claims 1-3, 9-15, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Segler et al. (2018, Nature, Vol. 555: 604-619, cited by previous action), in view of Jin et al. (2017, Advances in Neural Information Processing Systems 30: 1-10, cited by previous action), Kadurin et al. (2017, Molecular Pharmaceutics, Vol. 14: 3098−3104), and Sun et al. (2020, arXiv: 1-13). Italicized text from reference art. Underlined text correspond to amendment.
Applicable claims include:
Claim 1. A method for predicting retrosynthesis of a compound molecule, performed by a computer device, the method comprising: (Claim 1.i) obtaining a target molecule and determining the target molecule as a root node in a tree structure, the root node being associated with a first leaf node in the tree structure, and the tree structure comprising a retrosynthetic path of the target molecule; (Claim 1.ii) expanding the first leaf node through a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position; (Claim 1.iii) recursively processing the predicted molecule set corresponding to the second leaf nodes and determining a terminal node that satisfies a preset condition; and (Claim 1.iv) traversing path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule, wherein the graph neural network is trained by: (Claim 1.v) obtaining a first training molecule and a first training synthon corresponding to the first training molecule; (Claim 1.vi) determining a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon; (Claim 1.vii) determining a first loss function based on the node feature and the edge feature; (Claim 1.viii) determining a bond breaking probability of the chemical bond between the first training molecule and the first training synthon;(Claim 1.ix) determining a second loss function based on the bond breaking probability; and (Claim 1.x) updating a model parameter of the graph neural network according to the first loss function and the second loss function.
Claim 2. (Original): The method according to claim 1, wherein the expanding the first leaf node through a target retrosynthesis model comprises: (Claim 2.i) determining a first string of the compound molecule corresponding to the first leaf node;(Claim 2.ii) converting the first string into a first molecular map; (Claim 2.iii) determining the bond breaking position of the compound molecule corresponding to the first leaf node according to the first molecular map through the graph neural network; (Claim 2.vi) determining at least one synthon according to the bond breaking position through the reactant generation network; and (Claim 2.v) filtering the at least one synthon based on a preset rule to determine the predicted molecule set.
Claim 3. The method according to claim 2, wherein the determining the bond breaking position of the compound molecule comprises: (Claim 3.i) determining bond breaking information according to the first molecular map and a limit on a number of broken bonds through the graph neural network, (Claim 3.ii) determining a target key feature according to the target molecule, the target key feature comprising an atom key feature and a bond key feature, the atom key feature comprising at least one of an atom type, a number of bonds, a formal charge, chirality, a number of hydrogen atoms, an atomic hybridization state, aromaticity, an atomic weight, a high frequency reaction center feature, and a reaction type, and the bond key feature comprising at least one of a bond type, a conjugate, a ring bond, and a molecular stereo chemical feature; and (Claim 3.iii) extracting key breaking information based on the target key feature to determine the bond breaking position.
Claim 9. The method according to claim 1, wherein the recursively processing the predicted molecule set comprises: (Claim 9.i) determining a first candidate molecule to which the predicted molecule set corresponding to the second leaf node corresponds under a preset reaction type; (Claim 9.ii) expanding the first candidate molecule through the target retrosynthesis model to obtain third leaf nodes; (Claim 9.iii) recursively processing the predicted molecule set corresponding to the third leaf nodes to determine a second candidate molecule; and (Claim 9.iv) determining that a leaf node corresponding to the second candidate molecule is the terminal node in response to the second candidate molecule satisfying the preset condition.
Claim 10. The method according to claim 9, wherein the determining that a leaf node corresponding to the second candidate molecule is the terminal node comprises: (Claim 10.i) traversing based on the second candidate molecule to obtain a plurality of pathway molecules; and (Claim 10.ii) determining that the second candidate molecule satisfies the preset condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node, in response to the pathway molecule being a molecule in a basic molecule set.
Claim 11. The method according to claim 9, wherein the determining that a leaf node corresponding to the second candidate molecule is the terminal node comprises: (Claim 11.i) determining a number of expansions corresponding to the second candidate molecule in the tree structure; and (Claim 11.i) determining that the second candidate molecule satisfies the preset condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node, in response to the number of expansions reaching a preset value.
Claim 12. The method according to claim 1, wherein (Claim 12.i) the target molecule is a drug molecule, (Claim 12.ii) the tree structure is a Monte Carlo tree, and (Claim 12.iii) a compound molecule corresponding to the second leaf node is obtained by screening based on at least one of functional group interference, use of a protective group, a reaction feature of a functional group, and a name reaction rule.
Claim 13. An apparatus for predicting retrosynthesis of a compound molecule, comprising: at least one memory configured to store program code; and at least one processor configured to read the program code and operate as instructed by the program code, the program code comprising: (Claim 13.i) obtaining code, configured to cause the at least one processor to obtain a target molecule and determine the target molecule as a root node in a tree structure, the root node being associated with a first leaf node in the tree structure, and the tree structure comprising a retrosynthetic path of the target molecule; (Claim 13.ii) expansion code, configured to cause the at least one processor to expand the first leaf node through a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position;(Claim 13.iii) processing code, configured to cause the at least one processor to recursively process the predicted molecule set corresponding to the second leaf nodes and determine a terminal node that satisfies a preset condition; and (Claim 13.iv) prediction code, configured to cause the at least one processor to traverse path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule wherein the graph neural network is trained by training code configured to cause the at least one processor to:(Claim 13.v) obtain a first training molecule and a first training synthon corresponding to the first training molecule;(Claim 13.vi) determine a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon; (Claim 13.vii) determine a first loss function based on the node feature and the edge feature; (Claim 13.viii) determine a bond breaking probability of the chemical bond between the first training molecule and the first training synthon; (Claim 13.ix) determine a second loss function based on the bond breaking probability; and (Claim 13.x) update a model parameter of the graph neural network according to the first loss function and the second loss function.
Claim 14. The apparatus according to claim 13, wherein the expansion code is configured to cause the at least one processor to:(Claim 14.i) determine a first string of the compound molecule corresponding to the first leaf node; (Claim 14.ii) convert the first string into a first molecular map; (Claim 14.iii) determine the bond breaking position of the compound molecule corresponding to the first leaf node according to the first molecular map through the graph neural network; (Claim 14.iv) determine at least one synthon according to the bond breaking position through the reactant generation network; and (Claim 14.v) filter the at least one synthon based on a preset rule to determine the predicted molecule set.
Claim 15. The apparatus according to claim 14, wherein the program code further comprises: (Claim 15.i) determine bond breaking code configured to cause the at least one processor to determine bond breaking information according to the first molecular map and a limit on a number of broken bonds through the graph neural network, (Claim 15.ii) determine a target key feature according to the target molecule, the target key feature comprising an atom key feature and a bond key feature, the atom key feature comprising at least one of an atom type, a number of bonds, a formal charge, chirality, a number of hydrogen atoms, an atomic hybridization state, aromaticity, an atomic weight, a high frequency reaction center feature, and a reaction type, and the bond key feature comprising at least one of a bond type, a conjugate, a ring bond, and a molecular stereo chemical feature; (Claim 15.iii) and key breaking extraction code configured to cause the at least one processor to extract key breaking information based on the target key feature to determine the bond breaking position.
Claim 20. A non-transitory computer-readable storage medium, storing a computer program that when executed by at least one processor causes the at least one processor to: (Claim 20.i) obtain a target molecule and determining the target molecule as a root node in a tree structure, the root node being associated with a first leaf node in the tree structure, and the tree structure comprising a retrosynthetic path of the target molecule; (Claim 20.ii) expand the first leaf node through a target retrosynthesis model to obtain a plurality of second leaf nodes, the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position; (Claim 20.iii) recursively process the predicted molecule set corresponding to the second leaf nodes and determining a terminal node that satisfies a preset condition; and (Claim 20.iv) traverse path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule, wherein the graph neural network is trained by the computer program executed by the at least one processor to cause the at least one processor to: (Claim 20.v) obtain a first training molecule and a first training synthon corresponding to the first training molecule; (Claim 20.vi) determine a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon; (Claim 20.vii) determine a first loss function based on the node feature and the edge feature; (Claim 20.viii) determine a bond breaking probability of the chemical bond between the first training molecule and the first training synthon; (Claim 20.ix) determine a second loss function based on the bond breaking probability; and (Claim 20.x) update a model parameter of the graph neural network according to the first loss function and the second loss function.
Regarding Claim 1, 13, and 20, Segler et al. teach (Claim 1.i) obtaining a target molecule as a root node in a tree structure, associated with a first leaf node in the tree structure, comprising a retrosynthetic path of the target molecule (Page 611, Column 1, Paragraph 1: Molecules are stored in the search tree as canonical SMILES strings; Monte Carlo tree search has several desirable features, which makes it particularly well suited for retrosynthesis; Page 606, Figure 2: Monte Carlo tree search searches by iterating over four phases (Figure 2 shows structure of the Monte Carl Tree search for retrosynthesis including root and leaf nodes)). Segler et al. also teaches (Claim 1.ii) expanding the first leaf node through a target retrosynthesis model to obtain second leaf nodes (Page 611, Column 2, Paragraph 3: During expansion, the state is processed once via the expansion procedure, and the reduced top 50 successor states are directly added to the tree). Segler et al. also teaches (Claim 1.iii) recursively processing the predicted molecule set corresponding to the second leaf nodes and determining a terminal node that satisfies a preset condition (Page 611, Column 2, Paragraph 4: Before starting the rollout, the state is first checked for being terminal. A state can be terminal if it is solved). Segler et al. also teaches (Claim 1.iv) traversing path information corresponding to the terminal node to determine a retrosynthetic path of the target molecule (Page 611, Column 2, Paragraph 4: Terminal states are directly evaluated with the reward function). Segler et al. teach (Claim 1.vi) determining a node feature and an edge feature of the first training molecule and the first training synthon, the node feature being used for indicating a relationship between atoms of the first training molecule and the first training synthon, and the edge feature being used for indicating a relationship of a chemical bond between the first training molecule and the first training synthon (Page 611, Column 1, Paragraph 1: For processing, molecules are translated from SMILES into molecular graphs, which are vertex-labelled and edge-labelled graphs, with atoms as vertices and bonds as edges. Retrosynthetic transformation rules are productions on graphs). Segler et al. teach (Claim 1.viii) determining a bond breaking probability of the chemical bond between the first training molecule and the first training synthon (Page 611, Column 1, Paragraph 2: A Markov decision processes is a tuple, with states (positions), actions (transformations), a transition model determining the probability of reaching state when taking action in state, and a reward function). The modeling is based on predicting bonds formed/broken. Segler et al. also teaches the method was conducted on a computer (Page 612, Column 1, Paragraph 6: Training was carried out using stochastic gradient descent within 1–2 days on a single NVIDIA K80 graphics processing unit). The use of a computer would indicate the presence of memory, at least one processor, and non-transitory computer readable media to carry out the method. Claims 13 and 20 recites the limitations of Claim 1.
Regarding Claim 2 and 14, Segler et al. teach (Claim 2.i) determining a first string of the compound molecule corresponding to the first leaf node (Page 611, Column 1, Paragraph 1: Molecules are stored in the search tree as canonical SMILES strings). Segler et al. also teach (Claim 2.ii) converting the first string into a first molecular map (Page 611, Column 1, Paragraph 1: For processing, molecules are translated from SMILES into molecular graphs). Segler et al. also teach (Claim 2.iv) determining at least one synthon according to the bond breaking position through the reactant generation network (Page 605, Column 2, Paragraph 2: the transformations with the highest scores are applied to the molecule, which yields the possible precursors (precursor in this context is equivalent to synthon)). Segler et al. also teach (Claim 2.v) filtering the at least one synthon based on a preset rule to determine the predicted molecule set (Page 605, Column 2, Paragraph 1: Prediction with the in-scope filter network; After the search space has been narrowed down by the expansion policy to the most promising transformations, we need to predict whether the corresponding reactions will actually work for a particular molecule). Claim 14 recites the limitations of Claim 2.
Regarding Claim 9, Segler et al. teach (Claim 9.i) determining a first candidate molecule to which the predicted molecule set corresponding to the second leaf node corresponds under a preset reaction type (Page 607, Column 1, Paragraph 1: Figure 3 shows an exemplary six-step route for an intermediate of a drug candidate synthesis reported in 2015, which was found by our algorithm in 5.4 s). Segler et al. also teach (Claim 9.ii) expanding the first candidate molecule through the target retrosynthesis model to obtain third leaf nodes (Page 606, Figure 2a shows expanding the first candidate molecule through the target retrosynthesis model to obtain third leaf nodes). Segler et al. also teach (Claim 9.iii) recursively processing the predicted molecule set corresponding to the third leaf nodes to determine a second candidate molecule (Page 606, Figure 2b shows expansion procedure to determine best precursor (i.e., candidate molecule)). Segler et al. also teach (Claim 9.iv) determining that a leaf node corresponding to the second candidate molecule is the terminal node in response to the second candidate molecule satisfying the preset condition (Page 611, Column 2, Paragraph 3: Before starting the rollout, the state is first checked for being terminal).
Regarding Claim 10, Segler et al. teach (Claim 10.i) traversing based on the second candidate molecule to obtain a plurality of pathway molecules (Page 611, Column 2, Paragraph 3: During expansion, the state is processed once via the expansion procedure and the reduced top 50 successor states are directly added to the tree). Segler et al. also teach (Claim 10.ii) determining that the second candidate molecule satisfies the preset condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node, in response to the pathway molecule being a molecule in a basic molecule set (Page 611, Column 2, Paragraph 3: Before starting the rollout, the state is first checked for being terminal. A state can be terminal if it is solved).
Regarding Claim 11, Segler et al. teach (Claim 11.i) determining a number of expansions corresponding to the second candidate molecule in the tree structure (Page 604, Column 2, Paragraph 3: The first neural network (the expansion policy) guides the search in promising directions by proposing a restricted number of automatically extracted transformations). Segler et al. also teach (Claim 11.ii) determining that the second candidate molecule satisfies the preset condition and determining that the leaf node corresponding to the second candidate molecule is the terminal node, in response to the number of expansions reaching a preset value (Page 611, Column 2, Paragraph 6: In the update phase, the action values and visit counts of the edges traversed in the branch from to the root node are updated).
Regarding Claim 12, Segler et al. teach (Claim 12.i) the target molecule is a drug molecule (Page 607, Column 1, Paragraph 1: Figure 3 shows an exemplary six-step route for an intermediate of a drug candidate synthesis reported in 2015, which was found by our algorithm in 5.4 s). Segler et al. also teach (Claim 12.ii) the tree structure is a Monte Carlo tree (Page 611, Column 2, Paragraph 9: In the first Monte Carlo Tree Search phase, starting at the root node, the tree policy is used to select actions). Segler et al. also teach (Claim 12.iii) a compound molecule corresponding to the second leaf node is obtained by screening based on at least one of functional group interference, use of a protective group, a reaction feature of a functional group, and a name reaction rule (Page 605, Column 1, Paragraph 2: Importantly, such neural networks also learn about the context in which the reactions can occur (functional group tolerance)).
Segler et al. do not teach the target model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node, and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position (Claims 1.ii, 13.ii, 20.ii). Segler et al. also do not teach obtaining a first training molecule and a first training synthon corresponding to the first training molecule (Claims 1.v, 13.v, 20.v). Segler et al. also do not teach determining a first loss function based on the node feature and the edge feature (Claims 1.vii, 13.vii, 20.vii). Segler et al. also do not teach determining a second loss function based on the bond breaking probability (Claims 1.ix, 13.ix, 20.ix). Segler et al. also do not teach updating a model parameter of the graph neural network according to the first loss function and the second loss function (Claims 1.x, 13.x, and 20.x). Segler et al. also do not teach determining the bond breaking position of the compound molecule corresponding to the first leaf node according to the first molecular map through the graph neural network (Claims 2.iii, 14.iii). Segler et al. also do not teach determining bond breaking information according to the first molecular map and a limit on a number of broken bonds through the graph neural network; determining a target key feature according to the target molecule, the target key feature comprising an atom key feature and a bond key feature, the atom key feature comprising at least one of an atom type, a number of bonds, a formal charge, chirality, a number of hydrogen atoms, an atomic hybridization state, aromaticity, an atomic weight, a high frequency reaction center feature, and a reaction type and the bond key feature comprising at least one of a bond type, a conjugate, a ring bond, and a molecular stereo chemical feature; and extracting the key breaking information based on the target key feature to determine the bond breaking position (Claims 3 and 15). Related to Segler et al., Jin et al. disclose additional methods for using neural networks, including graph based, to model and predict molecular synthesis.
Regarding Claim 1, 13, and 20, Jin et al. teach (Claim 1.ii) the target retrosynthesis model comprising a graph neural network and a reactant generation network, the graph neural network being configured for predicting a bond breaking position of a compound molecule corresponding to the first leaf node (Page 3 Figure 2: Overview of our approach. We train a model to identify pairwise atom interactions in the reaction center. Each bond configuration generates a candidate outcome of the reaction; Page 3 Paragraph 1: Our approach builds from graph neural networks to encode graph structures), and the reactant generation network being configured for determining a predicted molecule set based on the bond breaking position (Page 5, Paragraph 5: The set of candidate products are then obtained by enumerating all possible bond configuration changes within the set). Jin et al. teach (Claim 1.v) obtaining a first training molecule and a first training synthon corresponding to the first training molecule (Page 3, Paragraph 3: Specifically, we train a neural network that operates on the reactant graph to predict a reactivity score for every pair of atoms). Jin et al. teach (Claim 1.vii) determining a first loss function based on the node feature and the edge feature (Page 5, Paragraph 4: Both models are trained to minimize the following loss function. For a given reaction with N atoms, we need to predict the reactivity score of O(N2) pairs). Reactivity of atoms (i.e. nodes) related to bonds (i.e. edges) formed/broken. Jin et al. teach (Claim 1.viii) determining a bond breaking probability of the chemical bond between the first training molecule and the first training synthon (Page 5, Paragraph 4: For a given reaction with N atoms, we need to predict the reactivity score of O(N2) pairs). Reactivity of atoms are related to bonds formed/broken. Jin et al. teach (Claim 1.x) updating a model parameter of the graph neural network according to the first loss function and the second loss function (Page 5, Paragraph 4: Both models are trained to minimize the following loss function). This describes model training/developing, updating model parameters are inherently part of the training/validating process and the loss function are in place to optimize the prediction the latent data representation. Claims 13 and 20 have the same limitations as claim 1.
Regarding Claim 2 and 14, Jin et al. teach (Claim 2.iii) determining the bond breaking position of the compound molecule corresponding to the first leaf node according to the first molecular map through the graph neural network (Page 3, Paragraph 3: After identifying the reaction center, we generate possible product candidates by enumerating possible bond configurations between atoms in the reaction center; Page 6, Paragraph 3: A difference graph is defined as a molecular graph which has the same atoms and bonds as (the variable) pi, with atom feature vector). Claim 14 recites the same limitations as claim 2.
Regarding Claims 3 and 15, Jin et al. teach (Claim 3.i) determining bond breaking information according to the first molecular map and a limit on a number of broken bonds through the graph neural network (Page 3, Paragraph 3: we train a neural network that operates on the reactant graph to predict a reactivity score for every pair of atoms. After identifying the reaction center, we generate possible product candidates by enumerating possible bond configurations between atoms in the reaction center subject to chemical constraints). Jin et al. also teach (Claim 3.ii) determining a target key feature according to the target molecule, the target key feature comprising an atom key feature and a bond key feature, the atom key feature comprising at least one of an atom type, a number of bonds, a formal charge, chirality, a number of hydrogen atoms, an atomic hybridization state, aromaticity, an atomic weight, a high frequency reaction center feature, and a reaction type (Page 6, Paragraph 8: Atom-level features include its elemental identity, degree of connectivity, number of attached hydrogen atoms, implicit valence, and aromaticity), and the bond key feature comprising at least one of a bond type, a conjugate, a ring bond, and a molecular stereo chemical feature (Page 6, Paragraph 8: Bond-level features include bond type (single, double, triple, or aromatic), whether it is conjugated, and whether the bond is part of a ring). Jin et al. also teach (Claim 3.iii) extracting the key breaking information based on the target key feature to determine the bond breaking position (Page 3, Paragraph 3: After identifying the reaction center, we generate possible product candidates by enumerating possible bond configurations between atoms in the reaction center subject to chemical constraints; Page 7, Paragraph 5: the global one accurately predicts the reaction center). Claim 15 recites the same limitations as claim 3.
Jin et al. does not explicitly teach determining a second loss function based on the bond breaking probability (Claim 1.ix). Related to the other cited art, Kadurin et al. disclose additional methods for using neural networks to model and predict molecules.
Regarding Claims 1, 13, and 20, Kadurin et al. teach (Claim 1.vii) determining a first loss function based on the node feature and the edge feature (Page 3099, Figure 1: VAE and AAE model architectures. Each with at least two loss functions; Page 3099, Column 2, Paragraph 2: Figure 1 shows the final structure of VAE and AAE, where shaded rectangles with dashed borders indicate components of the cost function minimized during training). The modeling and prediction is based on molecule selection and considered properties. Kadurin et al. teach (Claim 1.ix) determining a second loss function based on the bond breaking probability (Page 3099, Figure 1: VAE and AAE model architectures. Each with at least two loss functions; Page 3099, Column 2, Paragraph 2: Note that VAE directly trains the latent layer to match the standard normal distribution while AAE trains a separate (smaller) network, the discriminator, to distinguish between latent variables and the standard Gaussian). Different loss functions account to different factors and molecular properties within the modeling processes (Page 3099, Column 1, Paragraph 2: the AAE learned to sample from a data set of molecular fingerprints augmented with their concentrations). Kadurin et al. teach (Claim 1.x) updating a model parameter of the graph neural network according to the first loss function and the second loss function (Page 3100, Column 2, Paragraph 1: introduced a new hyperparameter that captured the trade-off between coverage of the original data set with generated samples and reconstruction quality).
Related to the other cited art, Sun et al. disclose additional methods for using neural networks, including graph based, to model and predict molecular synthesis and retrosynthesis. Importantly, Sun et al. also disclose the relatedness and complementarity of synthesis vs retrosynthesis and template free vs template approaches to molecular modeling (also see reason to combine and response to arguments below).
Regarding Claims 1, 13, and 20, Sun et al. suggest determining a first loss function based on the node feature and the edge feature; determining a bond breaking probability of the chemical bond between the first training molecule and the first training synthon; determining a second loss function based on the bond breaking probability; updating a model parameter of the graph neural network according to the first loss function and the second loss function ((Page 5, Paragraph 4: our consistent training is achieved by minimizing the dual loss; Page 5, Paragraph 5: the graph-based method presents chemical molecules, with vertices as atoms and edges as chemical bonds; Page 10, Paragraph 3: The model parameters are updated)).
It would have been obvious to one of ordinary skill in the art at the time of the effective filing date to combine Segler et al. with Jin et al., Kadurin et al., and Sun et al. Jin et al. teach the utilization of a graph neural network which represents an improvement to predicting molecular synthesis compared to prior methods relating to interpreting structural data, suggested by Segler et al. (Page 609, Column 1, Paragraph 4: our system currently does not take reaction mechanisms, equilibria between different forms, such as tautomers, or three-dimensional structures into account). The graph neural network taught by Jin et al. is able to improve on these deficiencies (Page 3, Paragraph 2: The question of molecular graph representation is a key issue in reaction modeling. This recurrent model is derived from the Weisfeiler-Lehman kernel that produces isomorphism-invariant representations of molecular graphs). Kadurin et al. teach their modeling allows for enhanced capabilities and performance, especially when large molecular datasets are considered for drug development (Page 3102, Column 1, Paragraph 2: These experimental validations have suggested that AAE outperforms VAE and significantly enhanced the capacity and efficiency of deep learning models with regard to applications to oncological molecular development). Kadurin et al. also suggest applying their modeling to graph based approaches (Page 3102, Column 2, Paragraph 1: Molecular graphs and other more chemically and biologically relevant representations of the molecular structures may serve as better types of training and discrimination data for the drug discovery pipelines incorporating the various flavors of generative adversarial models). Sun et al. teach methods that combine or suggest to combine various approaches to molecular modeling using neural network related to the other cited art including template vs template free modeling, synthesis vs retrosynthesis, and graph vs non-graph based approaches (Page 10, Paragraph 1: we proposed an unified EBM based framework that integrates multiple sequence- and graph-based variants for retrosynthesis. Assisted by a comprehensive assessment, we provide a critical understanding of different designs. Based on this, we proposed a novel variant—Dual model, which outperforms state-of-the-art in both template-based and template-free setting. Future works include designing new energy-based models that combine both sequence and graph representations, as well as extending it to reaction prediction and other chemistry applications). This teaches how these different approaches can be adapted to function together to model and predict retrosynthesis. Furthermore, one of ordinary skill in the art would predict that the methods could be readily combined with a reasonable expectation of success because all are utilizing related methods within one technical field – modeling molecules and their chemical properties to make predictions using neural networks.
The following rejection presents a new ground of rejection.
Claims 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Segler et al., in view of Jin et al., Kadurin et al., and Sun et al., as applied to Claims 1-3, 9-15, and 20 above, and in further view of Zheng et al. (2020, Journal of Chemical Information and Modeling, Vol. 60: 47-55, cited by previous action). Italicized text from reference art. Underlined text correspond to amendment.
Applicable claims include:
Claims 1-3, 9-15, and 20 are indicated above
Claim 5. The method according to claim 2, wherein the determining at least one synthon comprises: (Claim 5.i) splitting the target molecule based on the bond breaking position to obtain at least one synthon molecular map; (Claim 5.ii) converting the synthon molecular map into a second string; (Claim 5.iii) updating the second string based on a preset reaction type to obtain a third string; and (Claim 5.iv) determining the at least one synthon according to the third string through the reactant generation network.
Claim 6. The method according to claim 5, wherein the reactant generation network is trained by: (Claim 6.i) obtaining a second training molecule, a second training synthon, and a training reactant; (Claim 6.ii) determining a first training string based on a string corresponding to the second training molecule, a string corresponding to the second training synthon, and the preset reaction type; (Claim 6.iii) determining a second training string corresponding to the training reactant; (Claim 6.iv) associating the first training string and the second training string to determine a first training sample pair; and (Claim 6.v) training the reactant generation network based on the first training sample pair.
Claim 7. The method according to claim 6, further comprising: (Claim 7.i) obtaining a candidate string predicted by the graph neural network; (Claim 7.ii) adding the candidate string into the string corresponding to the second training synthon to update the first training string to a third training string; (Claim 7.iii) associating the third training string and the second training string to determine a second training sample pair; and (Claim 7.iv) training the reactant generation network based on the second training sample pair.
Claim 8. The method according to claim 6, further comprising: determining a target character format; and updating the first training string based on the target character format.
Claim 17. The apparatus according to claim 14, wherein the expansion code is configured to: (Claim 17.i) split the target molecule based on the bond breaking position to obtain at least one synthon molecular map; (Claim 17.ii) convert the synthon molecular map into a second string; (Claim 17.iii) update the second string based on a preset reaction type to obtain a third string; and (Claim 17.iv) determine the at least one synthon according to the third string through the reactant generation network.
Claim 18. The apparatus according to claim 17, wherein the reactant generation network is trained by training code configured to: (Claim 18.i) obtain a second training molecule, a second training synthon, and a training reactant; (Claim 18.ii) determine a first training string based on a string corresponding to the second training molecule, a string corresponding to the second training synthon, and the preset reaction type; (Claim 18.iii) determine a second training string corresponding to the training reactant; (Claim 18.iv) associate the first training string and the second training string to determine a first training sample pair; and (Claim 18.v) train the reactant generation network based on the first training sample pair.
Claim 19. The apparatus according to claim 18, wherein the program code is further configured to: (Claim 19.i) obtain a candidate string predicted by the graph neural network; (Claim 19.ii) add the candidate string into the string corresponding to the second training synthon to update the first training string to a third training string; (Claim 19.iii) associate the third training string and the second training string to determine a second training sample pair; and (Claim 19.iv) train the reactant generation network based on the second training sample pair.
Regarding Claims 1-3, 9-15, and 20, these are taught by Segler et al., Jin et al., Kadurin et al. and Sun et al. as described above.
Related to the other cited art, Zheng et al. disclose additional methods for using neural networks, including graph based, to model and predict molecular synthesis and retrosynthesis.
Regarding Claim 5 and 17, Zheng et al. suggest (Claim 5.ii) converting the synthon molecular map into a second string (Page 4, Column 1, Paragraph 2: we also considered a much bigger data set. It contains 480 thousand atom-mapped reactions without predefined reaction classes; Page 50, Column 1, Paragraph 2: In this study, a reaction was described by a variable-length string containing one pair of SMILES notations representing the reactants and target compound; Page 51, Column 1, Paragraph 2: RDKit was used for data preprocessing and molecule canonicalization (RDKit is capable of this function)). Zheng et al. teach (Claim 5.iii) updating the second string based on a preset reaction type to obtain a third string (Page 49, Column 1, Paragraph 2: The source sequence is the product of the reaction with a reaction type token prepended to the sequence). Zheng et al. also teach (Claim 5.iv) determining the at least one synthon according to the third string through the reactant generation network (Page 50, Column 2, Paragraph 3: we further built a transformer-based syntax corrector to automatically correct the syntax of unreasonable SMILES strings for improving the performances). Claim 17 recites the limitations of claim 5.
Regarding Claim 6 and 18, Zheng et al. teach (Claim 6.i) obtaining a second training molecule, a second training synthon, and a training reactant (Page 49, Column 1, Paragraph 2: each reaction was split into a source sequence and target sequence for model training). Zheng et al. also teach (Claim 6.ii) determining a first training string based on a string corresponding to the second training molecule, a string corresponding to the second training synthon, and the preset reaction type (Page 49, Column 1, Paragraph 2: a reaction was described by a variable-length string containing one pair of SMILES notations representing the reactants and target compound). Zheng et al. also teach (Claim 6.iii) determining a second training string corresponding to the training reactant (Page 49, Column 1, Paragraph 2: The source sequence is the product of the reaction with a reaction type token prepended to the sequence, and the target sequence is the reactant set). Zheng et al. also teach (Claim 6.iv) associating the first training string and the second training string to determine a first training sample pair (Page 49, Column 1, Paragraph 3: These sequences are then encoded into one-hot matrices with a token vocabulary). Zheng et al. also teach (Claim 6.v) training the reactant generation network based on the first training sample pair (Page 50, Column 2, Paragraph 5: The system does this by taking ground-truth reactants and invalid reactants generated from the retrosynthesis predictor to produce input−output pairs, which are then used to train a sequence-to sequence transformer model). Claim 18 recites the limitations of claim 6.
Regarding Claim 7 and 19, Zheng et al. teach (Claim 7.ii) adding the candidate string into the string corresponding to the second training synthon to update the first training string to a third training string (Page 49, Column 1, Paragraph 2: In this study, a reaction was described by a variable-length string containing one pair of SMILES notations representing the reactants and target compound). In the study by Zheng et al. stings are of variable length and contain multiple discrete pieces of information, so it would be obvious to generate new strings each time the information was updated. Zheng et al. also teach (Claim 7.iii) associating the third training string and the second training string to determine a second training sample pair (Page 49, Column 1, Paragraph 3: These sequences are then encoded into one-hot matrices with a token vocabulary). Claim 19 recites the limitations of claim 7.
Regarding Claim 8, Zheng et al. also teach determining a target character format; and updating the first training string based on the target character format (Page 48, Column 1, Paragraph 2: The key idea is to use a text representation for the reactants and products (e.g., SMILES)). SMILES is the format selected and strings are updated to this format.
Zheng et al. do not teach splitting the target molecule based on the bond breaking position to obtain at least one synthon molecular map (Claims 5.i, 17.i). Zheng et al. also do not teach obtaining a candidate string predicted by the graph neural network (Claim 7.i, 19.i). Zheng et al. also do not teach training the reactant generation network based on the second training sample pair (Claim 7.iv, 19.iv).
Regarding Claim 5 and 17, Segler et al. teach (Claim 5.i) splitting the target molecule based on the bond breaking position to obtain at least one synthon molecular map (Page 611, Column 1, Paragraph 1: For processing, molecules are translated from SMILES into molecular graphs; Page 606, Column 2, Paragraph 3: This function assigns the lowest cost to steps that split up the molecule into equally sized parts). Claim 17 recites the limitations of claim 5.
Segler et al. do not teach obtaining a candidate string predicted by the graph neural network (Claims 7.i, 19.i). Segler et al. also do not teach training the reactant generation network based on the second training sample pair (Claims 7.iv, 19.i).
Regarding Claim 7 and 19, Jin et al. teach (Claim 7.i) obtaining a candidate string predicted by the graph neural network (Page 5, Paragraph 5: The set of candidate products are then obtained by enumerating all possible bond configuration changes within the set). Jin et al. also teach (Claim 7.iv) training the reactant generation network based on the second training sample pair (Page 5, Paragraph 6: The training set for candidate ranking consists of lists of the reactants, known product, and other enumerated candidate products) This is a separate training set with the other enumerated candidate products being the output of the candidate generation model. Claim 19 recites the limitations of claim 7.
It would have been obvious to one of ordinary skill in the art at the time of the effective filing date to combine Zheng et al. with Segler et al., Jin et al., Kadurin et al. and Sun et al. Zheng et al. teaches detailed methods related to manipulating string data between multiple machine learning models to make predictions on retrosynthesis. Both Segler et al. and Jin et al. convert between strings representing molecules and mathematical representations. Zheng et al. describes methods for working with molecular strings that are applicable to the methods of Jin et al. and Segler et al. Additionally, Zheng et al. describe how their methods are an improvement (Page 48, Column 2, Paragraph 1: our method achieved an accuracy of 47.6%, which was 1.7 times higher than that of other state of-the-art methods), which would have motivated one of ordinary skill in the art to adopt. Zheng et al also suggest applying their methods to recursive retrosynthesis predictions (Page 52, Column 2, Paragraph 4: A possible option would be to recursively decompose the target). Furthermore, one of ordinary skill in the art would predict that the methods could be readily combined with a reasonable expectation of success because all are utilizing related methods within the same technical field – modeling molecules and their chemical properties to make predictions using neural networks.
Response to Arguments
Applicant asserts Jin et al. is silent on any second loss function (Page 21, Paragraph 4 of remarks). The examiner agrees that none the cited references make a dual loss function approach obvious. Therefore, the 103 rejection has been altered to address this limitation by the addition of Kadurin et al. and Sun et al. who teach and suggest using multiple loss functions to enhance data selection and prediction within neural networks. Additionally, Sun et al. teach combining various approaches to molecular modeling using neural network related to the other cited art including template vs template free, synthesis vs retrosynthesis, and graph vs non-graph based approaches. This teaches that different approaches used by the different arts can be adapted to function together to enhance retrosynthesis prediction, as in the instant claims.
Double Patenting
No double patenting issues were identified.
Conclusion
No claims are allowed.
This rejection is non-final (see 35 USC 103 Rejection).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BLAKE H ELKINS whose telephone number is (571)272-2649. The examiner can normally be reached Monday-Friday 8-5PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Karlheinz Skowronek can be reached at (571) 272-9047. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/B.H.E./Examiner, Art Unit 1687
/Karlheinz R. Skowronek/Supervisory Patent Examiner, Art Unit 1687
