Processing method and device for carrying out synthesizable optimization on material molecules

Abstract

The embodiment of the invention relates to a processing method and device for carrying out synthesizable optimization on material molecules. The method comprises the following steps: setting a synthesizable chemical information base; constructing a synthesizable optimization model in combination with a synthesizable chemical information base; constructing a model training data set to train the synthesizable optimization model; and after the model training is finished, carrying out synthesizable optimization processing on the molecular graph of any material molecule input by the user based on the synthesizable optimization model to obtain a corresponding optimized material molecular graph, and feeding back the optimized material molecular graph to the current user. According to the invention, the synthesizability of material molecules can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for processing material molecules for synthetic optimization. Background Art

[0002] Material molecular design plays an important role in the discovery and development of new materials, especially in the fields of energy, catalysts and nanomaterials. Generative models have become an effective tool for material molecular design in recent years, but due to the limited rules of material chemical reactions, the material molecules obtained by generative models often have poor synthesizability. Summary of the invention

[0003] The purpose of the present invention is to provide a processing method, device, electronic device and computer-readable storage medium for optimizing the synthesizability of material molecules in view of the defects of the prior art. The present invention pre-sets two information libraries for material molecules: a building block structure library and a reaction rule library, and merges the two information libraries to obtain a synthesizable chemical information library; and constructs a processing model for improving the synthesizability of material molecules in combination with the synthesizable chemical information library, which is recorded as a synthesizable optimization model; and constructs a model training data set to train the synthesizable optimization model; and after the model training is completed, the molecular graph of any material molecule input by the user is synthesized and optimized based on the synthesizable optimization model to obtain the corresponding optimized material molecule graph and feedback it to the current user. The synthesizable optimization model provided by the present invention can optimize the structure of material molecules on the premise of meeting the synthesizability requirements. After a material molecule is obtained through the generation model, the synthesizable optimization model of the present invention can improve the synthesizability of the currently generated molecule.

[0004] To achieve the above-mentioned purpose, a first aspect of an embodiment of the present invention provides a method for synthesizing and optimizing material molecules, the method comprising:

[0005] Setting a building block structure library and a reaction rule library corresponding to the material molecule; and forming a corresponding synthesizable chemical information library from the building block structure library and the reaction rule library; and constructing a processing model for improving the synthesizability of the material molecule in combination with the synthesizable chemical information library, which is recorded as a corresponding synthesizable optimization model;

[0006] Constructing a model training data set as a corresponding first data set; and training the synthesizable optimization model based on the first data set;

[0007] After the model training is completed, the molecular graph of any material molecule input by the user is subjected to synthesizable optimization processing based on the synthesizable optimization model to obtain the corresponding optimized material molecular graph and feed it back to the current user.

[0008] Preferably, the building block structure library includes a plurality of first building block records; each of the first building block records corresponds to a synthesizable material molecular building block; the first building block record includes a first block identifier, a first block structure and a first molecular fingerprint; the first block identifier is a unique identifier of the current building block; the first block structure is a molecular graph of the current building block; the first molecular fingerprint is a molecular fingerprint of the current building block;

[0009] The reaction rule library includes multiple first reaction rule records; each of the first reaction rule records corresponds to a synthesis reaction type of a class of material molecules; the first reaction rule record includes a first rule identifier, a first reaction rule type, a first reactant sequence and a second reactant sequence; the first and second reactant sequences each contain multiple corresponding first and second reactant identifiers; the first and second reactant identifiers are each a first block identifier.

[0010] Preferably, the synthesizable optimization model is used to perform synthesizable optimization processing on the original material molecular graph input into the model and output the corresponding optimized material molecular graph;

[0011] The model input end of the synthesizable optimization model is used to receive the original material molecular graph input by the model, and the model output end is used to output the corresponding optimized material molecular graph;

[0012] The synthesizable optimization model includes an encoder, a decoder, a decoding output module, a molecular fingerprint prediction module, a building block processing module, a reaction rule prediction module, a reaction rule processing module, a synthesizable chemical information library and a molecular graph stack;

[0013] The input end of the encoder is connected to the input end of the model, and the output end is connected to the first input end of the decoder; the second and third input ends of the decoder are respectively connected to the output end of the building block processing module and the output end of the reaction rule processing module, and the first and second output ends of the decoder are connected to the first and second input ends of the decoding output module; the first, second and third output ends of the decoding output module are respectively connected to the input end of the molecular fingerprint prediction module, the input end of the reaction rule prediction module and the output end of the model; the output end of the molecular fingerprint prediction module is connected to the input end of the building block processing module; the output end of the reaction rule prediction module is connected to the input end of the reaction rule processing module; the building block processing module and the reaction rule processing module are also respectively connected to the synthesizable chemical information library; the decoding output module, the building block processing module and the reaction rule processing module are also respectively connected to the molecular graph stack;

[0014] The encoder is implemented based on the Uni-Mol model; the encoder is used to perform atomic-level embedded coding processing on the original material molecular graph to obtain a corresponding coding tensor X and send it to the decoder;

[0015] The decoder is implemented based on the Decoder model of the Transformer model; the decoder is formed by sequentially connecting an embedded coding module, N decoding modules, a linear layer and a Softmax layer, and the preset total number of decoding layers N is a positive integer; the first and second input ends of the embedded coding module are respectively connected to the second and third input ends of the decoder; the first input end of the first decoding module is connected to the output end of the embedded coding module; except for the first decoding module, the first input end of each of the remaining decoding modules is connected to the output end of the previous decoding module; the second input end of each decoding module is connected to the first input end of the decoder; the output end of the last decoding module is respectively connected to the input end of the linear layer and the first output end of the decoder; the output end of the linear layer is connected to the input end of the Softmax layer; the output end of the Softmax layer is connected to the second output end of the decoder;

[0016] The embedded coding module is used to set the corresponding marker type s0 based on the preset start marker before the first step prediction; and to set the corresponding position encoding vector PE based on the preset position embedded coding rule during the i-th step prediction. i , the single-step index i is an integer greater than or equal to 1; and the marker type s corresponding to the i-1th step i-1 Identify; if the tag type s i-1 is the start marker, then the encoding vector e is embedded based on the preset start type start Set the corresponding marker encoding vector e i ; If the marker type s i-1 To construct a block marker, a preset first MLP network is used to construct the marker type s i-1 The corresponding building block molecular fingerprint p i-1 Perform embedded coding to obtain the corresponding marker encoding vector e i ; If the marker type s i-1 is a reaction marker, then the marker type s is embedded in the encoding rule based on the preset reaction rule i-1 The corresponding reaction rule type r i-1 Perform embedded coding to obtain the corresponding marker encoding vector e i ; and the position encoding vector PE obtained this time i and the marker encoding vector e i The starting decoded tensor H that constitutes the i-th stepi,0 Sending to the first decoding module;

[0017] H i,0 =PE i +e i ,

[0018]

[0019] Among them, f MLP1 () is the model inference function corresponding to the first MLP network, θ MLP1 is the model parameter of the first MLP network; Embed() is the encoding rule for embedding the reaction rule;

[0020] The first decoding module is used to decode the encoding tensor X and the starting decoding tensor H according to the encoding tensor X and the starting decoding tensor H i,0 Decode to get the corresponding decoding tensor H i,1 Send to the second decoding module; the second decoding module is used to generate a decoded tensor based on the encoding tensor X and the decoding tensor H i,1 Decode to get the corresponding decoding tensor H i,2 Send to the third decoding module; and so on, until the Nth decoding module is used to decode the tensor according to the encoding tensor X and the decoding tensor H i,N-1 Decode to get the corresponding decoding tensor H i,N Sending to the linear layer and the decoding output module;

[0021] The linear layer is used to decode the tensor H i,N Perform full connection calculation to obtain the corresponding full connection vector L i Send to the Softmax layer;

[0022] The Softmax layer is used to calculate the fully connected vector L i Predict the marker type of step i to get the corresponding marker type s i Send to the decoding output module; the marker type s i Includes building block tags, response tags, and end tags;

[0023] The decoding output module is used to decode the marker type s predicted in step i i Identify; if the tag type s i is a building block marker, then the marker type s i and the corresponding decoded tensor H i,N Send to the molecular fingerprint prediction module; if the marker type s i If it is a reaction marker, the marker type s iand the corresponding decoded tensor H i,N Send to the reaction rule prediction module; if the marker type s i is an end marker, a pop operation is performed on the molecular graph stack to obtain a popped molecular graph, and the current popped molecular graph is sent to the model output end as the corresponding optimized material molecular graph;

[0024] The molecular fingerprint prediction module is used to use a preset second MLP network according to the decoded tensor H i,N Perform molecular fingerprint prediction processing to obtain the corresponding predicted fingerprint p pre,i ; and set the tag type s i and the predicted fingerprint p pre,i Sending to the building block processing module;

[0025] p pre,i =sigmoid(f MLP2 (H i,N θ MLP2 ));

[0026] Among them, the sigmoid() function is the activation function of the second MLP network, f MLP2 () is the model inference function corresponding to the second MLP network, θ MLP2 are the model parameters of the second MLP network;

[0027] The building block processing module is used to compare each of the first molecular fingerprints and the predicted fingerprint p in the building block structure library of the synthesizable chemical information library. pre,i The similarity is calculated to obtain the corresponding first similarity; the first block structure and the first molecular fingerprint of the first building block record corresponding to the first similarity with the largest similarity are extracted as the corresponding current molecular graph and current molecular fingerprint; and the current molecular graph is pushed into the molecular graph stack by performing a push operation on the molecular graph stack; and the current molecular fingerprint is used as the corresponding building block molecular fingerprint p i ; and set the tag type s i and the corresponding building block molecular fingerprint p i Sending to the embedded coding module;

[0028] The reaction rule prediction module is used to use a preset third MLP network according to the decoding tensor H i,N Perform reaction rule type prediction processing to obtain the corresponding prediction type r pre,i ; and set the tag type s i and the prediction type r pre,i Sending to the reaction rule processing module;

[0029] r pre,i =softmax(f MLP3 (H i,N θ MLP3 ));

[0030] Among them, the softmax() function is the activation function of the third MLP network, f MLP3 () is the model inference function corresponding to the third MLP network, θ MLP3 are model parameters of the third MLP network;

[0031] The reaction rule processing module is used to perform two pop operations on the molecular graph stack to obtain two popped molecular graphs as corresponding A and B reactant molecular graphs; and identify whether the A and B reactant molecular graphs are both empty; if at least one of the A and B reactant molecular graphs is empty, the prediction type r is invalid based on the preset invalid reaction rule type. pre,i reset; if both the A and B reactant molecule graphs are not empty, then the first reaction rule type in the reaction rule library of the synthesizable chemical information library and the predicted type r pre,i The first reaction rule record matched is recorded as the corresponding current reaction rule record, and whether the current reaction rule record is empty is identified; if the current reaction rule record is empty, the prediction type r is changed based on the invalid reaction rule type. pre,i reset; if the current reaction rule record is not empty, the first block identifier corresponding to the A and B reactant molecule graphs is recorded as the corresponding A and B block identifiers, and the A and B block identifiers and the first and second reactant sequence information of the current reaction rule record are identified; if the A block identifier satisfies the first reactant sequence but the B block identifier does not satisfy the second reactant sequence, or the A block identifier satisfies the second reactant sequence but the B block identifier does not satisfy the first reactant sequence, or the B block identifier satisfies the first reactant sequence but the A block identifier does not satisfy the second reactant sequence, or the B block identifier satisfies the second reactant sequence but the A block identifier does not satisfy the first reactant sequence, or the A and B block identifiers do not satisfy the first and second reactant sequences, then based on the invalid reaction rule type, the prediction type r pre,i Reset; and in the prediction type r pre,i If it is not the invalid reaction rule type, the prediction type r is used based on the RDKit tool. pre,iThe current reaction rule is used to perform a synthetic compound prediction calculation with the A and B reactant molecular graphs as the current reactant combination, and the synthetic molecular graph obtained by this prediction calculation is used as the corresponding current molecular graph, and the current molecular graph is pushed into the molecular graph stack by performing a stacking operation on the molecular graph stack; and the prediction type r pre,i As the corresponding reaction rule type r i ; and set the tag type s i and the corresponding reaction rule type r i Sending to the embedded coding module;

[0032] The molecular graph stack is used to store the molecular graph pushed into the stack and put it on the top of the stack when processing the push operation, and when the total number of other molecular graphs in the stack is not zero, the other molecular graphs are shifted one step toward the bottom of the stack in sequence;

[0033] The molecular graph stack is also used to output the molecular graph at the top of the stack as the corresponding popped molecular graph when processing a pop operation and simultaneously remove the currently output molecular graph from the stack, and when the total number of remaining molecular graphs in the stack is not zero, shift the remaining molecular graphs one step toward the top of the stack in sequence.

[0034] Preferably, the first data set includes a plurality of first data records; the first data record includes a first training molecular graph and a first label sequence; the first label sequence consists of a plurality of first labels; the first label includes a label index j, a label marker type and label data The tag index j is a positive integer; the tag marker type Including building block markers, reaction markers, and end markers; the tag marker type When building a block marker, the label data is a corresponding molecular fingerprint tag The tag marker type When it is a reaction tag, the tag data A corresponding reaction rule type label The tag marker type When it is the end tag, the tag data Is empty.

[0035] Preferably, the model building training data set is recorded as the corresponding first data set, which specifically includes:

[0036] Step 51, obtaining a plurality of first collected data groups by collecting material molecule graphs and corresponding synthesis path information from a public material molecule synthesis database;

[0037] Wherein, the first acquisition data group includes a first material molecular graph and a first synthesis path; the first synthesis path is formed by sequentially sorting a plurality of first path point information; the first path point information includes a first reactant molecular graph, a second reactant molecular graph, a first reactant molecular fingerprint, a second reactant molecular fingerprint, a synthetic product molecular graph, a synthetic product molecular fingerprint and a synthetic reaction rule type; the first reactant molecular fingerprint is a molecular fingerprint of the first reactant molecular graph; the second reactant molecular fingerprint is a molecular fingerprint of the second reactant molecular graph; the synthetic product molecular fingerprint is a molecular fingerprint of the synthetic product molecular graph; the synthetic reaction rule type satisfies the reaction rule type range of the synthesizable chemical information library; one and only one of the first and second reactant molecular graphs in the latter first path point information is the synthetic product molecular graph in the previous first path point information; the synthetic product molecular graph in the last first path point information matches the first material molecular graph;

[0038] Step 52, taking the first of the first collected data groups as the corresponding current data group;

[0039] Step 53, taking the first material molecule graph and the first synthesis path of the current data group as the corresponding current material molecule graph and current synthesis path; taking the first first path point information of the current synthesis path as the corresponding current path point information; initializing the first index to 1; and initializing an empty sequence as the corresponding first label sequence;

[0040] Step 54: Set the tag index j as the first index and set a corresponding tag marker type. To build a block marker, and set a corresponding tag data is the first reactant molecular fingerprint of the current path point information; and the label index j and the label marker type obtained this time and the label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index;

[0041] Step 55: Set the tag index j as the first index; and set a corresponding tag marker type To build a block marker, and set a corresponding tag data is the second reactant molecular fingerprint of the current path point information; and the label index j and the label marker type obtained this time and the label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index;

[0042] Step 56: Set the tag index j as the first index; and set a corresponding tag marker type For the reaction tag, set a corresponding tag data is the synthesis reaction rule type of the current path point information; and the label index j and the label marker type obtained this time and the label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index;

[0043] Step 57, identifying whether the current path point information is the last of the first path point information of the current synthetic path; if so, setting a label index j as the first index and setting a corresponding label marker type As the end marker, set a corresponding tag data is empty, and the label index j obtained this time, the label marker type and the label data The first label is added into the first label sequence, and the current material molecular graph is used as a corresponding first training molecular graph, and the first training molecular graph corresponding to the current synthesis path and the first label sequence are used to form a corresponding first data record, and the process goes to step 58; if not, the next first path point information of the current synthesis path is extracted as the new current path point information, and the process goes back to step 54;

[0044] Step 58, identifying whether the current data group is the last one of the first collected data groups; if not, taking the next first collected data group as the new current data group and returning to step 53; if so, forming the corresponding first data set from all the first data records obtained.

[0045] Preferably, the training of the synthesizable optimization model based on the first data set specifically includes:

[0046] Step 61, based on a preset first segmentation ratio, the first data set is divided into two sub-data sets which are recorded as a corresponding first training set and a first evaluation set;

[0047] Wherein, both the first training set and the first evaluation set are composed of a plurality of the first data records; the ratio of the total number of records of the first training set to the first evaluation set satisfies the first segmentation ratio;

[0048] Step 62, taking the first first data record of the first training set as the corresponding current training record;

[0049] Step 63, input the first training molecular graph of the current training record as the corresponding raw material molecular graph into the synthesizable optimization model for synthesizable optimization processing; and during the model processing, the marker type s generated by the model prediction at each step i Data collection is performed, and for each marker type s specifically a building block marker i The corresponding predicted fingerprint p pre,i Data collection is performed, and for each marker type s specifically a reaction marker i The corresponding prediction type r pre,i Perform data collection; and based on all collected data and the maximum label index j of the first label sequence of the current training record max Performing prediction sequence assembly to obtain a corresponding first prediction sequence;

[0050] Among them, j max = max(j); the first prediction sequence is composed of a plurality of first prediction arrays; the first prediction array includes a prediction marker type s j and predicted data d j ; The predicted marker type s j Including building block markers, reaction markers, and end markers; the prediction marker type s j When the block marker is constructed, the prediction data d j is a corresponding predicted fingerprint p pre,i ; The predicted marker type s j When it is a reaction marker, the predicted data d j is a corresponding prediction type r pre,i ; The predicted marker type s j When it is the end marker, the predicted data d j Empty; if the tag type s collected i The total number of tags is less than the maximum tag index j max , then the first prediction sequence is predicted based on the preset invalid prediction array until the first prediction sequence is consistent with the sequence length of the current first label sequence; if the collected marker type s i The total number is greater than the maximum label index jmax , then based on the invalid prediction array, the current first label sequence is label-filled until the current first label sequence is consistent with the first prediction sequence and the sequence length is consistent with the first prediction sequence, and the maximum label index j is filled based on the sequence length of the first prediction sequence. max Perform a reset;

[0051] Step 64: Substitute the first prediction sequence and the current first label sequence into a preset first model loss function L M1 ; and based on the preset first model optimizer, move towards making the first model loss function L M1 Performing a round of modulation on the model parameters of the encoder, the decoder, the molecular fingerprint prediction module and the reaction rule prediction module of the synthesizable optimization model in the direction of reaching the minimum value;

[0052] Among them, the first model loss function L M1 By the first type loss function L S , the first fingerprint loss function L P And the first rule loss function L R composition;

[0053] L M1 =L S +L P +L R ,

[0054]

[0055] Indicator function

[0056] Indicator function

[0057] L CE () is the cross entropy loss function, L BCE () is a binary cross entropy loss function; n is the type of the label marker in the current first label sequence, specifically the building block marker The total number of, m is the tag marker type of the current first tag sequence specifically the reaction marker Total number of;

[0058] The first model optimizer includes at least an SGD optimizer and an Adam optimizer;

[0059] Step 65, identifying whether the current training record is the last first data record of the first training set; if so, going to step 66; if not, extracting the next first data record of the first training set as the new current training record, and returning to step 63;

[0060] Step 66, perform a round of traversal on all the first data records of the first evaluation set; and in this round of traversal, use the first data record currently traversed as the corresponding current evaluation record; and input the first training molecular graph of the current evaluation record as the corresponding raw material molecular graph into the synthesizable optimization model for synthesizable optimization processing; and in the process of model processing, the marker type s generated by the model prediction at each step i Data collection is performed, and for each marker type s specifically a building block marker i The corresponding predicted fingerprint p pre,i Data collection is performed, and for each marker type s specifically a reaction marker i The corresponding prediction type r pre,i Perform data collection; and based on all collected data and the maximum tag index j of the first tag sequence of the current evaluation record max Perform prediction sequence assembly to obtain a corresponding second prediction sequence; and form a corresponding prediction-label sequence pair from the second prediction sequence and the current first label sequence; and at the end of this round of traversal, estimate the accuracy, precision, recall and F1 score based on all the obtained prediction-label sequence pairs to obtain the corresponding first accuracy, first precision, first recall and first F1 score;

[0061] Step 67, identifying the first accuracy, the first precision, the first recall and the first F1 score; if the first accuracy does not meet the preset first accuracy range, or the first accuracy does not meet the preset first accuracy range, or the first recall does not meet the preset first recall range, or the first F1 score does not meet the preset first F1 score range, return to step 62 to continue training; if the first accuracy meets the first accuracy range, and the first accuracy meets the first accuracy range, and the first recall meets the first recall range, and the first F1 score meets the first F1 score range, stop training and confirm that the model training is completed.

[0062] Preferably, the synthesizable optimization model is used to perform synthesizable optimization processing on the molecular graph of any material molecule input by the user to obtain the corresponding optimized material molecular graph and feed it back to the current user, specifically including:

[0063] The molecular graph of any material molecule input by the user is input into the synthesizable optimization model as the corresponding original material molecular graph to perform synthesizable optimization processing to obtain the corresponding optimized material molecular graph; and the optimized material molecular graph is fed back to the current user as the corresponding optimized material molecular graph.

[0064] The second aspect of the embodiment of the present invention provides a device for implementing the method for synthesizing and optimizing material molecules as described in the first aspect, the device comprising: a chemical information library and model building module, a model training module and a model application module;

[0065] The chemical information library and model building module are used to set the building block structure library and reaction rule library corresponding to the material molecule; and the building block structure library and the reaction rule library form a corresponding synthesizable chemical information library; and a processing model for improving the synthesizability of the material molecule is constructed in combination with the synthesizable chemical information library, which is recorded as a corresponding synthesizable optimization model;

[0066] The model training module is used to construct a model training data set recorded as a corresponding first data set; and train the synthesizable optimization model based on the first data set;

[0067] The model application module is used to perform synthesizable optimization processing on the molecular graph of any material molecule input by the user based on the synthesizable optimization model after the model training is completed, obtain the corresponding optimized material molecular graph and feed it back to the current user.

[0068] A third aspect of an embodiment of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;

[0069] The processor is used to couple with the memory, read and execute instructions in the memory, so as to implement the method steps described in the first aspect above;

[0070] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0071] A fourth aspect of an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions. When the computer instructions are executed by a computer, the computer executes the instructions of the method described in the first aspect above.

[0072] The embodiment of the present invention provides a processing method, device, electronic device and computer-readable storage medium for optimizing the synthesizability of material molecules. As can be seen from the above content, the embodiment of the present invention pre-sets two information libraries for material molecules: a building block structure library and a reaction rule library, and merges these two information libraries to obtain a synthesizable chemical information library; and combines the synthesizable chemical information library to construct a processing model for improving the synthesizability of material molecules, which is recorded as a synthesizable optimization model; and constructs a model training data set to train the synthesizable optimization model; and after the model training is completed, the molecular graph of any material molecule input by the user is synthesized and optimized based on the synthesizable optimization model to obtain the corresponding optimized material molecule graph and feedback it to the current user. The synthesizable optimization model provided by the embodiment of the present invention can optimize the structure of material molecules on the premise of meeting the synthesizability requirements. After obtaining a material molecule through the generation model, the synthesizable optimization model of the embodiment of the present invention can improve the synthesizability of the currently generated molecule. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A schematic diagram of a method for synthesizing and optimizing material molecules provided in Example 1 of the present invention;

[0074] Figure 2 A module structure diagram of a synthesizable optimization model provided in the first embodiment of the present invention;

[0075] Figure 3 A module structure diagram of a processing device for synthesizing and optimizing material molecules provided in the second embodiment of the present invention;

[0076] Figure 4 A schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] Embodiment 1 of the present invention provides a method for synthesizing and optimizing material molecules, such as Figure 1 As shown in the schematic diagram of a method for synthesizing and optimizing material molecules provided in the first embodiment of the present invention, the method mainly comprises the following steps:

[0079] Step 1, setting a building block structure library and a reaction rule library corresponding to the material molecule; and forming a corresponding synthesizable chemical information library from the building block structure library and the reaction rule library; and constructing a processing model for improving the synthesizability of the material molecule in combination with the synthesizable chemical information library, which is recorded as a corresponding synthesizable optimization model;

[0080] Specifically comprising: step 11, setting a building block structure library and a reaction rule library corresponding to the material molecule; and forming a corresponding synthesizable chemical information library from the building block structure library and the reaction rule library;

[0081] Here, the building block structure library of the embodiment of the present invention includes a plurality of first building block records; each first building block record corresponds to a synthesizable material molecular building block; the first building block record includes a first block identifier, a first block structure and a first molecular fingerprint; the first block identifier is a unique identifier of the current building block; the first block structure is a molecular graph of the current building block; the first molecular fingerprint is a molecular fingerprint of the current building block;

[0082] The reaction rule library of the embodiment of the present invention includes a plurality of first reaction rule records; each first reaction rule record corresponds to a type of synthesis reaction of a class of material molecules; the first reaction rule record includes a first rule identifier, a first reaction rule type, a first reactant sequence and a second reactant sequence; the first and second reactant sequences each include a plurality of corresponding first and second reactant identifiers; the first and second reactant identifiers are each a first block identifier;

[0083] Step 12, and construct a processing model for improving the synthesizability of material molecules in combination with the synthesizable chemical information library, which is recorded as the corresponding synthesizable optimization model.

[0084] Here, the synthesizable optimization model of the embodiment of the present invention is used to perform synthesizable optimization processing on the original material molecular graph input to the model and output the corresponding optimized material molecular graph; the model input end of the synthesizable optimization model is used to receive the original material molecular graph input to the model, and the model output end is used to output the corresponding optimized material molecular graph; Figure 2 The module structure diagram of the synthesizable optimization model provided in the first embodiment of the present invention is shown.

[0085] like Figure 2 As shown, the model components of the synthesizable optimization model include an encoder, a decoder, a decoding output module, a molecular fingerprint prediction module, a building block processing module, a reaction rule prediction module, a reaction rule processing module, a synthesizable chemical information library, and a molecular graph stack.

[0086] The connection relationship of the model components is as follows: the input end of the encoder is connected to the input end of the model, and the output end is connected to the first input end of the decoder; the second and third input ends of the decoder are respectively connected to the output end of the building block processing module and the output end of the reaction rule processing module, and the first and second output ends of the decoder are connected to the first and second input ends of the decoding output module; the first, second and third output ends of the decoding output module are respectively connected to the input end of the molecular fingerprint prediction module, the input end of the reaction rule prediction module and the output end of the model; the output end of the molecular fingerprint prediction module is connected to the input end of the building block processing module; the output end of the reaction rule prediction module is connected to the input end of the reaction rule processing module; the building block processing module and the reaction rule processing module are also respectively connected to the synthesizable chemical information library; the decoding output module, the building block processing module and the reaction rule processing module are also respectively connected to the molecular graph stack.

[0087] The component functions of the model components are described below.

[0088] 1) Encoder:

[0089] The encoder of the embodiment of the present invention is implemented based on the Uni-Mol model. The encoder is used to perform atomic level embedding coding on the original material molecular graph to obtain the corresponding coding tensor X and send it to the decoder.

[0090] Here, the Uni-Mol model is a three-dimensional molecular graph encoding model implemented by an Encoder model based on the Transformer model. The Uni-Mol model is an encoding model that can perform atomic-level embedding encoding on molecular graphs. For details, please refer to the technical document "Uni-Mol: A Universal 3D Molecular Representation Learning Framework". I will not elaborate on it again.

[0091] 2) Decoder:

[0092] The decoder of the embodiment of the present invention is implemented based on the Decoder model of the Transformer model; Figure 2As shown, similar to the standard structure of the Decoder model based on the Transformer model, the decoder of the embodiment of the present invention is also composed of an embedded coding module, N decoding modules, a linear layer and a Softmax layer connected in sequence, and the preset total number of decoding layers N is a positive integer; the first and second input ends of the embedded coding module are respectively connected to the second and third input ends of the decoder; the first input end of the first decoding module is connected to the output end of the embedded coding module; except for the first decoding module, the first input end of each of the remaining decoding modules is connected to the output end of the previous decoding module; the second input end of each decoding module is connected to the first input end of the decoder; the output end of the last decoding module is respectively connected to the input end of the linear layer and the first output end of the decoder; the output end of the linear layer is connected to the input end of the Softmax layer; the output end of the Softmax layer is connected to the second output end of the decoder.

[0093] a. The embedded coding module is used to set the corresponding marker type s0 based on the preset start marker before the first step prediction; and to set the corresponding position encoding vector PE based on the preset position embedding coding rule during the i-th step prediction. i , the single-step index i is an integer greater than or equal to 1; and the marker type s corresponding to the i-1th step i-1 Identify; if the marker type s i-1 is the start marker, then the encoding vector e is embedded based on the preset start type start Set the corresponding marker encoding vector e i ; If the marker type s i-1 To construct a block marker, a preset first MLP network is used for the marker type s i-1 The corresponding building block molecular fingerprint p i-1 Perform embedded coding to obtain the corresponding marker encoding vector e i ; If the marker type s i-1 If it is a reaction marker, the marker type s is embedded in the encoding rule based on the preset reaction rule. i-1 The corresponding reaction rule type r i-1 Perform embedded coding to obtain the corresponding marker encoding vector e i ; And the position encoding vector PE obtained this time i and the marker encoding vector e i The starting decoded tensor H that constitutes the i-th step i,0 Send to the first decoding module.

[0094] Here, the starting decoded tensor H i,0 The expression is as follows:

[0095] H i,0 =PEi +e i ,

[0096]

[0097] Among them, f MLP1 () is the model inference function corresponding to the first MLP network, θ MLP1 are the model parameters of the first MLP network; Embed() is the reaction rule embedding encoding rule.

[0098] b. The first decoding module is used to decode the encoding tensor X and the starting decoding tensor H i,0 Decode to get the corresponding decoding tensor H i,1 Send to the second decoding module; the second decoding module is used to decode the tensor according to the encoding tensor X and the decoding tensor H i,1 Decode to get the corresponding decoding tensor H i,2 Send to the third decoding module; and so on, until the Nth decoding module is used to decode the tensor according to the encoding tensor X and the decoding tensor H i,N-1 Decode to get the corresponding decoding tensor H i,N Send to the linear layer and the decoding output module.

[0099] c. The linear layer is used to decode the tensor H i,N Perform full connection calculation to obtain the corresponding full connection vector L i Send to the Softmax layer.

[0100] d. Softmax layer is used to calculate the fully connected vector L i Predict the marker type of step i to get the corresponding marker type s i Sent to the decoding output module; where the marker type s i Includes building block tags, response tags, and end tags.

[0101] 3) Decoding output module:

[0102] The decoding output module of the embodiment of the present invention is used to predict the marker type s in the i-th step. i Identify; if the marker type s i For building block markers, set the marker type s i and the corresponding decoded tensor H i,N Send to the molecular fingerprint prediction module; if the marker type s i For a reaction marker, set the marker type to s i and the corresponding decoded tensor H i,N Send to the reaction rule prediction module; if the marker type s iIf is the end marker, a pop operation is performed on the molecular graph stack to obtain a popped molecular graph, and the current popped molecular graph is sent to the model output end as the corresponding optimized material molecular graph.

[0103] 4) Molecular fingerprint prediction module:

[0104] The molecular fingerprint prediction module of the embodiment of the present invention is used to use a preset second MLP network to decode the tensor H i,N Perform molecular fingerprint prediction processing to obtain the corresponding predicted fingerprint p pre,i ; and set the tag type s i and predicted fingerprint p pre,i Send to the building block processing module.

[0105] Here, the predicted fingerprint p pre,i The prediction process is:

[0106] p pre,i =sigmoid(f MLP2 (H i,N θ MLP2 ));

[0107] Among them, the sigmoid() function is the activation function of the second MLP network, f MLP2 () is the model inference function corresponding to the second MLP network, θ MLP2 are the model parameters of the second MLP network.

[0108] 5) Building block processing module:

[0109] The building block processing module of the embodiment of the present invention is used to process each first molecular fingerprint and the predicted fingerprint p in the building block structure library of the synthesizable chemical information library. pre,i The similarity is calculated to obtain the corresponding first similarity; the first structure and the first molecular fingerprint of the first building block record corresponding to the first similarity with the largest similarity are extracted as the corresponding current molecular graph and current molecular fingerprint; and the current molecular graph is pushed into the molecular graph stack by performing a push operation on the molecular graph stack; and the current molecular fingerprint is used as the corresponding building block molecular fingerprint p i ; and set the tag type s i and the corresponding building block molecular fingerprint p i Send to the embedded coding module.

[0110] 6) Reaction rule prediction module:

[0111] The reaction rule prediction module of the embodiment of the present invention is used to use a preset third MLP network according to the decoding tensor H i,N Perform reaction rule type prediction processing to obtain the corresponding prediction type r pre,i; and set the tag type s i and prediction type r pre,i Send to the reaction rule processing module.

[0112] Here, the prediction type r pre,i The prediction process is:

[0113] r pre,i =softmax(f MLP3 (H i,N θ MLP3 ));

[0114] Among them, the softmax() function is the activation function of the third MLP network, f MLP3 () is the model inference function corresponding to the third MLP network, θ MLP3 are the model parameters of the third MLP network.

[0115] 7) Reaction rule processing module:

[0116] The reaction rule processing module of the embodiment of the present invention is used to perform two pop operations on the molecular graph stack to obtain two popped molecular graphs as corresponding A and B reactant molecular graphs; and identify whether both A and B reactant molecular graphs are not empty; if at least one of the A and B reactant molecular graphs is empty, the prediction type r is set based on the preset invalid reaction rule type. pre,i Reset; if both A and B reactant molecule graphs are not empty, the first reaction rule type in the reaction rule library of the synthesizable chemical information library and the predicted type r pre,i The first matched reaction rule record is recorded as the corresponding current reaction rule record, and whether the current reaction rule record is empty is identified; if the current reaction rule record is empty, the prediction type r is set based on the invalid reaction rule type. pre,i Reset; if the current reaction rule record is not empty, the first block identifier corresponding to the A and B reactant molecule graphs is recorded as the corresponding A and B block identifiers, and the A and B block identifiers and the first and second reactant sequence information of the current reaction rule record are identified; if the A block identifier satisfies the first reactant sequence but the B block identifier does not satisfy the second reactant sequence, or the A block identifier satisfies the second reactant sequence but the B block identifier does not satisfy the first reactant sequence, or the B block identifier satisfies the first reactant sequence but the A block identifier does not satisfy the second reactant sequence, or the B block identifier satisfies the second reactant sequence but the A block identifier does not satisfy the first reactant sequence, or the A and B block identifiers do not satisfy the first and second reactant sequences, then the prediction type r is based on the invalid reaction rule type. pre,i reset; and in the prediction type r pre,i If it is not an invalid reaction rule type, the RDKit tool is used to predict the type r pre,iFor the current reaction rule, a synthetic compound prediction calculation is performed with the A and B reactant molecular graphs as the current reactant combination, and the synthetic molecular graph obtained by this prediction calculation is used as the corresponding current molecular graph, and the current molecular graph is pushed into the molecular graph stack by performing a stacking operation on the molecular graph stack; and the prediction type r pre,i As the corresponding reaction rule type r i ; and set the tag type s i and the corresponding reaction rule type r i Send to the embedded coding module.

[0117] 8) Molecular graph stack:

[0118] The molecular graph stack of the embodiment of the present invention is used to store the molecular graph pushed into the stack and place it on the top of the stack when processing the push operation, and to shift the other molecular graphs one step toward the bottom of the stack in sequence when the total number of other molecular graphs in the stack is not zero.

[0119] The molecular graph stack is also used to output the molecular graph at the top of the stack as the corresponding popped molecular graph when processing a pop operation and remove the currently output molecular graph from the stack at the same time, and when the total number of remaining molecular graphs in the stack is not zero, the remaining molecular graphs are shifted one step toward the top of the stack in sequence.

[0120] Step 2, constructing a model training data set and recording it as the corresponding first data set; and training the synthesizable optimization model based on the first data set;

[0121] Specifically comprising: step 21, constructing a model training data set and recording it as the corresponding first data set;

[0122] The first data set includes a plurality of first data records; the first data record includes a first training molecular graph and a first label sequence; the first label sequence consists of a plurality of first labels; the first label includes a label index j, a label marker type and label data The label index j is a positive integer; the label marker type Includes building block tags, response tags, end tags; tag tag types When building a block marker, the label data is a corresponding molecular fingerprint tag Tag marker type When it is a response tag, the tag data A corresponding reaction rule type label Tag marker type When it is the end tag, the tag data is empty;

[0123] Specifically, it includes: step 211, obtaining a plurality of first collected data groups by collecting material molecule graphs and corresponding synthesis path information from a public material molecule synthesis database;

[0124] Among them, the first collected data group includes the first material molecular graph and the first synthesis path; the first synthesis path is formed by sequentially sorting a plurality of first path point information; the first path point information includes the first reactant molecular graph, the second reactant molecular graph, the first reactant molecular fingerprint, the second reactant molecular fingerprint, the synthetic product molecular graph, the synthetic product molecular fingerprint and the synthetic reaction rule type; the first reactant molecular fingerprint is the molecular fingerprint of the first reactant molecular graph; the second reactant molecular fingerprint is the molecular fingerprint of the second reactant molecular graph; the synthetic product molecular fingerprint is the molecular fingerprint of the synthetic product molecular graph; the synthetic reaction rule type satisfies the reaction rule type range of the synthesizable chemical information library; in the first and second reactant molecular graphs in the latter first path point information, there is and only one reactant molecular graph that is the synthetic product molecular graph in the previous first path point information; the synthetic product molecular graph in the last first path point information matches the first material molecular graph;

[0125] Step 212, taking the first first collected data group as the corresponding current data group;

[0126] Step 213, taking the first material molecular graph and the first synthesis path of the current data group as the corresponding current material molecular graph and the current synthesis path; taking the first first path point information of the current synthesis path as the corresponding current path point information; initializing the first index to 1; and initializing an empty sequence as the corresponding first label sequence;

[0127] Step 214, set a tag index j as the first index and set a corresponding tag marker type To build a block marker, set a corresponding label data is the first reactant molecular fingerprint of the current path point information; and the label index j and label marker type obtained this time and label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index;

[0128] Step 215, set a tag index j as the first index; and set a corresponding tag marker type To build a block marker, set a corresponding label data is the second reactant molecular fingerprint of the current path point information; and the label index j and label marker type obtained this time and label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index;

[0129] Step 216, set a tag index j as the first index; and set a corresponding tag marker type Respond to the marker and set a corresponding tag data is the synthetic reaction rule type of the current path point information; and the label index j and label marker type obtained this time and label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index;

[0130] Step 217, identify whether the current path point information is the last first path point information of the current synthetic path; if so, set a label index j as the first index and set a corresponding label marker type For the end marker, set a corresponding tag data is empty, and the label index j and label marker type obtained this time and label data A corresponding first tag is formed and added to the first tag sequence, and the current material molecular graph is used as a corresponding first training molecular graph, and the first training molecular graph corresponding to the current synthesis path and the first tag sequence are formed into a corresponding first data record, and the process goes to step 218; if not, the next first path point information of the current synthesis path is extracted as the new current path point information, and the process goes back to step 214;

[0131] Step 218, identifying whether the current data group is the last first collected data group; if not, taking the next first collected data group as the new current data group and returning to step 213; if so, forming a corresponding first data set from all the obtained first data records;

[0132] Step 22, training the synthesizable optimization model based on the first data set;

[0133] Specifically, it includes: step 221, dividing the first data set into two sub-data sets based on a preset first division ratio and recording them as a corresponding first training set and a first evaluation set;

[0134] The first segmentation ratio is a preset ratio parameter, such as 9:1; the first training set and the first evaluation set are both composed of a plurality of first data records; the ratio of the total number of records of the first training set to the first evaluation set satisfies the first segmentation ratio;

[0135] Step 222, taking the first first data record of the first training set as the corresponding current training record;

[0136] Step 223, input the first training molecular graph of the current training record as the corresponding original material molecular graph into the synthesizable optimization model for synthesizable optimization processing; and during the model processing, the marker type s generated by each step of the model prediction is i Perform data collection and for each specific building block marker type s i The corresponding predicted fingerprint p pre,i Perform data collection and for each specific marker type s for the reaction marker i Corresponding prediction type r pre,i Perform data collection; and based on all collected data and the maximum label index j of the first label sequence of the current training record max Performing prediction sequence assembly to obtain a corresponding first prediction sequence;

[0137] Among them, j max = max(j); the first prediction sequence is composed of a plurality of first prediction arrays; the first prediction array includes a prediction marker type s j and predicted data d j ; Prediction token type s j Includes building block markers, response markers, end markers; prediction marker types j When building a block marker, predict the data d j is a corresponding predicted fingerprint p pre,i ; Prediction token type s j When it is a reaction marker, the predicted data d j is a corresponding prediction type r pre,i ; Prediction token type s j When it is the end marker, the predicted data d j Empty; if the collected marker type is s i The total number of is less than the maximum label index j max , then the first prediction sequence is supplemented with a prediction array based on the preset invalid prediction array until the first prediction sequence is consistent with the sequence length of the current first label sequence; if the collected marker type s i The total number is greater than the maximum label index j max , then the current first label sequence is label-filled based on the invalid prediction array until the current first label sequence is consistent with the first prediction sequence and the maximum label index j is calculated based on the sequence length of the first prediction sequence. max Perform a reset;

[0138] Step 224: Substitute the first prediction sequence and the current first label sequence into the preset first model loss function L M1 ; and based on the preset first model optimizer, move towards making the first model loss function L M1A round of modulation is performed on the model parameters of the encoder, decoder, molecular fingerprint prediction module and reaction rule prediction module of the synthesizable optimization model in the direction of reaching the minimum value;

[0139] Here, the first model loss function L of the embodiment of the present invention is M1 By the first type loss function L S , the first fingerprint loss function L P And the first rule loss function L R The composition is specifically:

[0140] L M1 =L S +L P +L R ,

[0141]

[0142] Indicator function

[0143] Indicator function

[0144] Among them, L CE () is the cross entropy loss function, L BCE () is the binary cross entropy loss function; n is the label marker type of the current first label sequence, specifically the building block marker The total number of, m is the tag marker type of the current first tag sequence specifically the reaction marker Total number of;

[0145] The first model optimizer of the embodiment of the present invention includes at least an SGD optimizer and an Adam optimizer;

[0146] Step 225, identifying whether the current training record is the last first data record of the first training set; if so, going to step 226; if not, extracting the next first data record of the first training set as the new current training record, and returning to step 223;

[0147] Step 226, perform a round of traversal on all first data records of the first evaluation set; and in this round of traversal, use the first data record currently traversed as the corresponding current evaluation record; and input the first training molecular graph of the current evaluation record as the corresponding original material molecular graph into the synthesizable optimization model for synthesizable optimization processing; and in the process of model processing, the marker type s generated by the model prediction at each step i Perform data collection and for each specific building block marker type s i The corresponding predicted fingerprint p pre,iPerform data collection and for each specific marker type s for the reaction marker i Corresponding prediction type r pre,i Perform data collection; and based on all collected data and the maximum label index j of the first label sequence of the current evaluation record max The prediction sequence is assembled to obtain the corresponding second prediction sequence; and the second prediction sequence and the current first label sequence form a corresponding prediction-label sequence pair; and at the end of this round of traversal, the accuracy, precision, recall and F1 score are estimated based on all the obtained prediction-label sequence pairs to obtain the corresponding first accuracy, first precision, first recall and first F1 score;

[0148] Step 227, identify the first accuracy, the first precision, the first recall rate and the first F1 score; if the first accuracy does not meet the preset first accuracy range, or the first accuracy does not meet the preset first accuracy range, or the first recall rate does not meet the preset first recall rate range, or the first F1 score does not meet the preset first F1 score range, return to step 222 to continue training; if the first accuracy meets the first accuracy range, and the first accuracy meets the first accuracy range, and the first recall rate meets the first recall rate range, and the first F1 score meets the first F1 score range, stop training and confirm that the model training is completed.

[0149] Here, the first accuracy range, the first precision range, the first recall range and the first F1 score range of the embodiment of the present invention are four preset value ranges.

[0150] Step 3: After the model training is completed, the molecular graph of any material molecule input by the user is subjected to synthesizable optimization processing based on the synthesizable optimization model to obtain the corresponding optimized material molecular graph and feedback it to the current user;

[0151] Specifically, it includes: inputting the molecular graph of any material molecule input by the user as the corresponding original material molecular graph into the synthesizable optimization model for synthesizable optimization processing to obtain the corresponding optimized material molecular graph; and feeding back the optimized material molecular graph to the current user as the corresponding optimized material molecular graph.

[0152] Figure 3 The module structure diagram of a processing device for synthesizing and optimizing material molecules provided in the second embodiment of the present invention is a terminal device or server that implements the aforementioned method embodiment, and can also be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiment. For example, the device can be a device or chip system of the aforementioned terminal device or server. Figure 3 As shown, the device includes: a chemical information library and model building module 201, a model training module 202 and a model application module 203.

[0153] The chemical information library and model construction module 201 is used to set the building block structure library and reaction rule library corresponding to the material molecules; and the building block structure library and the reaction rule library form a corresponding synthesizable chemical information library; and a processing model for improving the synthesizability of material molecules is constructed in combination with the synthesizable chemical information library, which is recorded as the corresponding synthesizable optimization model.

[0154] The model training module 202 is used to construct a model training data set recorded as the corresponding first data set; and train the synthesizable optimization model based on the first data set.

[0155] The model application module 203 is used to perform synthesizable optimization processing on the molecular graph of any material molecule input by the user based on the synthesizable optimization model after the model training is completed, and obtain the corresponding optimized material molecular graph to feed back to the current user.

[0156] An embodiment of the present invention provides a processing device for synthesizing and optimizing material molecules, which can execute the method steps in the above method embodiment. Its implementation principle and technical effects are similar and will not be repeated here.

[0157] It should be noted that it should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the chemical information library and model building module can be a separate processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a processing element of the above device. The function of the above-mentioned module is determined. The implementation of other modules is similar. In addition, these modules can be integrated together in whole or in part, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in the processor element or instructions in software form.

[0158] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASIC), or one or more digital signal processors (DSP), or one or more field programmable gate arrays (FPGA). For another example, when a module above is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0159] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the above method embodiments are generated. The above computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The above-mentioned computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the above-mentioned computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, Bluetooth, microwave, etc.) methods. The above-mentioned computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The above-mentioned available medium can be a magnetic medium (such as a floppy disk, a hard disk, a tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0160] Figure 4 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device may be a terminal device or a server for implementing the method of the aforementioned embodiment, or may be a terminal device or a server for implementing the method of the aforementioned embodiment connected to the aforementioned terminal device or server. Figure 4As shown, the electronic device may include: a processor 301 (such as a CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transceiver 303. Various instructions may be stored in the memory 302 to complete various processing functions and implement the processing steps described in the aforementioned embodiment method. Preferably, the electronic device involved in the embodiment of the present invention also includes: a power supply 304, a system bus 305 and a communication port 306. The system bus 305 is used to realize the communication connection between components. The above-mentioned communication port 306 is used for connecting and communicating between the electronic device and other peripherals.

[0161] exist Figure 4 The system bus 305 mentioned in the figure can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The communication interface is used to realize the communication between the database access device and other devices (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM), and may also include non-volatile memory (Non-Volatile Memory), such as at least one disk storage.

[0162] The above-mentioned processor can be a general-purpose processor, including a central processing unit CPU, a network processor (Network Processor, NP), a graphics processing unit (Graphics Processing Unit, GPU), etc.; it can also be a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0163] It should be noted that an embodiment of the present invention further provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the method and processing process provided in the above embodiments.

[0164] The embodiment of the present invention provides a processing method, device, electronic device and computer-readable storage medium for optimizing the synthesizability of material molecules. As can be seen from the above content, the embodiment of the present invention pre-sets two information libraries for material molecules: a building block structure library and a reaction rule library, and merges these two information libraries to obtain a synthesizable chemical information library; and combines the synthesizable chemical information library to construct a processing model for improving the synthesizability of material molecules, which is recorded as a synthesizable optimization model; and constructs a model training data set to train the synthesizable optimization model; and after the model training is completed, the molecular graph of any material molecule input by the user is synthesized and optimized based on the synthesizable optimization model to obtain the corresponding optimized material molecule graph and feedback it to the current user. The synthesizable optimization model provided by the embodiment of the present invention can optimize the structure of material molecules on the premise of meeting the synthesizability requirements. After obtaining a material molecule through the generation model, the synthesizable optimization model of the embodiment of the present invention can improve the synthesizability of the currently generated molecule.

[0165] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0166] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for synthesizing and optimizing material molecules, characterized in that: The method comprises: Setting a building block structure library and a reaction rule library corresponding to the material molecule; and forming a corresponding synthesizable chemical information library from the building block structure library and the reaction rule library; and constructing a processing model for improving the synthesizability of the material molecule in combination with the synthesizable chemical information library, which is recorded as a corresponding synthesizable optimization model; Constructing a model training data set as a corresponding first data set; and training the synthesizable optimization model based on the first data set; After the model training is completed, the molecular graph of any material molecule input by the user is subjected to synthesizable optimization processing based on the synthesizable optimization model to obtain the corresponding optimized material molecular graph and feed it back to the current user.

2. The method for synthesizing and optimizing material molecules according to claim 1, characterized in that: The building block structure library includes a plurality of first building block records; each of the first building block records corresponds to a synthesizable material molecule building block; the first building block record includes a first block identifier, a first block structure and a first molecular fingerprint; The first block identifier is a unique identifier of the current building block; the first block structure is a molecular graph of the current building block; the first molecular fingerprint is a molecular fingerprint of the current building block; The reaction rule library includes multiple first reaction rule records; each of the first reaction rule records corresponds to a synthesis reaction type of a class of material molecules; the first reaction rule record includes a first rule identifier, a first reaction rule type, a first reactant sequence and a second reactant sequence; the first and second reactant sequences each contain multiple corresponding first and second reactant identifiers; the first and second reactant identifiers are each a first block identifier.

3. The method for synthesizing and optimizing material molecules according to claim 2, characterized in that: The synthesizable optimization model is used to perform synthesizable optimization processing on the original material molecular graph input into the model and output the corresponding optimized material molecular graph; The model input end of the synthesizable optimization model is used to receive the original material molecular graph input by the model, and the model output end is used to output the corresponding optimized material molecular graph; The synthesizable optimization model includes an encoder, a decoder, a decoding output module, a molecular fingerprint prediction module, a building block processing module, a reaction rule prediction module, a reaction rule processing module, a synthesizable chemical information library and a molecular graph stack; The input end of the encoder is connected to the input end of the model, and the output end is connected to the first input end of the decoder; the second and third input ends of the decoder are respectively connected to the output end of the building block processing module and the output end of the reaction rule processing module, and the first and second output ends of the decoder are connected to the first and second input ends of the decoding output module; the first, second and third output ends of the decoding output module are respectively connected to the input end of the molecular fingerprint prediction module, the input end of the reaction rule prediction module and the output end of the model; the output end of the molecular fingerprint prediction module is connected to the input end of the building block processing module; the output end of the reaction rule prediction module is connected to the input end of the reaction rule processing module; the building block processing module and the reaction rule processing module are also respectively connected to the synthesizable chemical information library; the decoding output module, the building block processing module and the reaction rule processing module are also respectively connected to the molecular graph stack; The encoder is implemented based on the Uni-Mol model; the encoder is used to perform atomic-level embedded coding processing on the original material molecular graph to obtain a corresponding coding tensor X and send it to the decoder; The decoder is implemented based on the Decoder model of the Transformer model; the decoder is formed by sequentially connecting an embedded coding module, N decoding modules, a linear layer and a Softmax layer, and the preset total number of decoding layers N is a positive integer; the first and second input ends of the embedded coding module are respectively connected to the second and third input ends of the decoder; the first input end of the first decoding module is connected to the output end of the embedded coding module; except for the first decoding module, the first input end of each of the remaining decoding modules is connected to the output end of the previous decoding module; the second input end of each decoding module is connected to the first input end of the decoder; the output end of the last decoding module is respectively connected to the input end of the linear layer and the first output end of the decoder; the output end of the linear layer is connected to the input end of the Softmax layer; the output end of the Softmax layer is connected to the second output end of the decoder; The embedded coding module is used to set the corresponding marker type s0 based on the preset start marker before the first step prediction; and to set the corresponding position encoding vector PE based on the preset position embedded coding rule during the i-th step prediction. i , the single-step index i is an integer greater than or equal to 1; and the marker type s corresponding to the i-1th step i-1 Identify; if the tag type s i-1 is the start marker, then the encoding vector e is embedded based on the preset start type start Set the corresponding marker encoding vector e i ; If the marker type s i-1 To construct a block marker, a preset first MLP network is used for the marker type s i-1 The corresponding building block molecular fingerprint p i-1 Perform embedded coding to obtain the corresponding marker encoding vector e i ; If the marker type s i-1 is a reaction marker, then the marker type s is embedded in the encoding rule based on the preset reaction rule i-1 The corresponding reaction rule type r i-1 Perform embedded coding to obtain the corresponding marker encoding vector e i ; and the position encoding vector PE obtained this time i and the marker encoding vector e i The starting decoded tensor H that constitutes the i-th step i,0 Sending to the first decoding module; A i,0 =PE i +e i , Among them, f MLP1 () is the model inference function corresponding to the first MLP network, θ MLP1 is the model parameter of the first MLP network; Embed() is the encoding rule for embedding the reaction rule; The first decoding module is used to decode the encoding tensor X and the starting decoding tensor H according to the encoding tensor X and the starting decoding tensor H i,0 Decode to get the corresponding decoding tensor H i,1 Send to the second decoding module; the second decoding module is used to generate a decoded tensor based on the encoding tensor X and the decoding tensor H i,1 Decode to get the corresponding decoding tensor H i,2 Send to the third decoding module; and so on, until the Nth decoding module is used to decode the tensor according to the encoding tensor X and the decoding tensor H i,N-1 Decode to get the corresponding decoding tensor H i,N Sending to the linear layer and the decoding output module; The linear layer is used to decode the tensor H i,N Perform full connection calculation to obtain the corresponding full connection vector L i Send to the Softmax layer; The Softmax layer is used to calculate the fully connected vector L i Predict the marker type of step i to get the corresponding marker type s i Send to the decoding output module; the marker type s i Includes building block tags, response tags, and end tags; The decoding output module is used to decode the marker type s predicted in step i i Identify; if the tag type s i is a building block marker, then the marker type s i and the corresponding decoded tensor H i,N Send to the molecular fingerprint prediction module; if the marker type s i If it is a reaction marker, the marker type s i and the corresponding decoded tensor H i,N Send to the reaction rule prediction module; if the marker type s i is an end marker, a pop operation is performed on the molecular graph stack to obtain a popped molecular graph, and the current popped molecular graph is sent to the model output end as the corresponding optimized material molecular graph; The molecular fingerprint prediction module is used to use a preset second MLP network according to the decoded tensor H i,N Perform molecular fingerprint prediction processing to obtain the corresponding predicted fingerprint p pre,i ; and set the tag type s i and the predicted fingerprint p pre,i Sending to the building block processing module; p pre,i =sigmoid(f MLP2 (H i,N ;θ MLP2 )); Among them, the sigmoid() function is the activation function of the second MLP network, f MLP2 () is the model inference function corresponding to the second MLP network, θ MLP2 are the model parameters of the second MLP network; The building block processing module is used to compare each of the first molecular fingerprints and the predicted fingerprint p in the building block structure library of the synthesizable chemical information library. pre,i The similarity is calculated to obtain the corresponding first similarity; the first block structure and the first molecular fingerprint of the first building block record corresponding to the first similarity with the largest similarity are extracted as the corresponding current molecular graph and current molecular fingerprint; and the current molecular graph is pushed into the molecular graph stack by performing a push operation on the molecular graph stack; and the current molecular fingerprint is used as the corresponding building block molecular fingerprint p i ; and set the tag type s i and the corresponding building block molecular fingerprint p i Sending to the embedded coding module; The reaction rule prediction module is used to use a preset third MLP network according to the decoding tensor H i,N Perform reaction rule type prediction processing to obtain the corresponding prediction type r pre,i ; and set the tag type s i and the prediction type r pre,i Sending to the reaction rule processing module; r pre,i =softmax(f MLP3 (H i,N ;θ MLP3 )); Among them, the softmax() function is the activation function of the third MLP network, f MLP3 () is the model inference function corresponding to the third MLP network, θ MLP3 are model parameters of the third MLP network; The reaction rule processing module is used to perform two pop operations on the molecular graph stack to obtain two popped molecular graphs as corresponding A and B reactant molecular graphs; and identify whether the A and B reactant molecular graphs are both empty; if at least one of the A and B reactant molecular graphs is empty, the prediction type r is invalid based on the preset invalid reaction rule type. pre,i reset; if both the A and B reactant molecule graphs are not empty, then the first reaction rule type in the reaction rule library of the synthesizable chemical information library and the predicted type r pre,i The first reaction rule record matched is recorded as the corresponding current reaction rule record, and whether the current reaction rule record is empty is identified; if the current reaction rule record is empty, the prediction type r is changed based on the invalid reaction rule type. pre,i reset; if the current reaction rule record is not empty, the first block identifier corresponding to the A and B reactant molecule graphs is recorded as the corresponding A and B block identifiers, and the A and B block identifiers and the first and second reactant sequence information of the current reaction rule record are identified; if the A block identifier satisfies the first reactant sequence but the B block identifier does not satisfy the second reactant sequence, or the A block identifier satisfies the second reactant sequence but the B block identifier does not satisfy the first reactant sequence, or the B block identifier satisfies the first reactant sequence but the A block identifier does not satisfy the second reactant sequence, or the B block identifier satisfies the second reactant sequence but the A block identifier does not satisfy the first reactant sequence, or the A and B block identifiers do not satisfy the first and second reactant sequences, then based on the invalid reaction rule type, the prediction type r pre,i Reset; and in the prediction type r pre,i If it is not the invalid reaction rule type, the prediction type r is used based on the RDKit tool. pre,i The current reaction rule is used to perform a synthetic compound prediction calculation with the A and B reactant molecular graphs as the current reactant combination, and the synthetic molecular graph obtained by this prediction calculation is used as the corresponding current molecular graph, and the current molecular graph is pushed into the molecular graph stack by performing a stacking operation on the molecular graph stack; and the prediction type r pre,i As the corresponding reaction rule type r i ; and set the tag type s i and the corresponding reaction rule type r i Sending to the embedded coding module; The molecular graph stack is used to store the molecular graph pushed into the stack and put it on the top of the stack when processing the push operation, and when the total number of other molecular graphs in the stack is not zero, the other molecular graphs are shifted one step toward the bottom of the stack in sequence; The molecular graph stack is also used to output the molecular graph at the top of the stack as the corresponding popped molecular graph when processing a pop operation and simultaneously remove the currently output molecular graph from the stack, and when the total number of remaining molecular graphs in the stack is not zero, shift the remaining molecular graphs one step toward the top of the stack in sequence.

4. The method for synthesizing and optimizing material molecules according to claim 3, characterized in that: The first data set includes a plurality of first data records; the first data record includes a first training molecular graph and a first label sequence; the first label sequence consists of a plurality of first labels; the first label includes a label index j, a label marker type and label data The tag index j is a positive integer; the tag marker type Including building block markers, reaction markers, and end markers; the tag marker type When building a block marker, the label data is a corresponding molecular fingerprint tag The tag marker type When it is a reaction tag, the tag data A corresponding reaction rule type label The tag marker type When it is the end tag, the tag data Is empty.

5. The method for synthesizing and optimizing material molecules according to claim 4, characterized in that: The model building training data set is recorded as the corresponding first data set, which specifically includes: Step 51, obtaining a plurality of first collected data groups by collecting material molecule graphs and corresponding synthesis path information from a public material molecule synthesis database; Wherein, the first acquisition data group includes a first material molecular graph and a first synthesis path; the first synthesis path is formed by sequentially sorting a plurality of first path point information; the first path point information includes a first reactant molecular graph, a second reactant molecular graph, a first reactant molecular fingerprint, a second reactant molecular fingerprint, a synthetic product molecular graph, a synthetic product molecular fingerprint and a synthetic reaction rule type; the first reactant molecular fingerprint is a molecular fingerprint of the first reactant molecular graph; the second reactant molecular fingerprint is a molecular fingerprint of the second reactant molecular graph; the synthetic product molecular fingerprint is a molecular fingerprint of the synthetic product molecular graph; the synthetic reaction rule type satisfies the reaction rule type range of the synthesizable chemical information library; one and only one of the first and second reactant molecular graphs in the latter first path point information is the synthetic product molecular graph in the previous first path point information; the synthetic product molecular graph in the last first path point information matches the first material molecular graph; Step 52, taking the first of the first collected data groups as the corresponding current data group; Step 53, taking the first material molecule graph and the first synthesis path of the current data group as the corresponding current material molecule graph and current synthesis path; taking the first first path point information of the current synthesis path as the corresponding current path point information; initializing the first index to 1; and initializing an empty sequence as the corresponding first label sequence; Step 54: Set the tag index j as the first index and set a corresponding tag marker type. To build a block marker, and set a corresponding tag data is the first reactant molecular fingerprint of the current path point information; and the label index j and the label marker type obtained this time and the label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index; Step 55: Set the tag index j as the first index; and set a corresponding tag marker type To build a block marker, and set a corresponding tag data is the second reactant molecular fingerprint of the current path point information; and the label index j and the label marker type obtained this time and the label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index; Step 56: Set the tag index j as the first index; and set a corresponding tag marker type For the reaction tag, set a corresponding tag data is the synthesis reaction rule type of the current path point information; and the label index j and the label marker type obtained this time and the label data Form a corresponding first tag and add it to the first tag sequence; and add 1 to the first index; Step 57, identifying whether the current path point information is the last of the first path point information of the current synthetic path; if so, setting a label index j as the first index and setting a corresponding label marker type As the end marker, set a corresponding tag data is empty, and the tag index j obtained this time, the tag marker type and the label data The first label is added into the first label sequence, and the current material molecular graph is used as a corresponding first training molecular graph, and the first training molecular graph corresponding to the current synthesis path and the first label sequence are used to form a corresponding first data record, and the process goes to step 58; if not, the next first path point information of the current synthesis path is extracted as the new current path point information, and the process goes back to step 54; Step 58, identifying whether the current data group is the last one of the first collected data groups; if not, taking the next first collected data group as the new current data group and returning to step 53; if so, forming the corresponding first data set from all the first data records obtained.

6. The method for synthesizing and optimizing material molecules according to claim 4, characterized in that: The training of the synthesizable optimization model based on the first data set specifically includes: Step 61, based on a preset first segmentation ratio, the first data set is divided into two sub-data sets which are recorded as a corresponding first training set and a first evaluation set; Wherein, both the first training set and the first evaluation set are composed of a plurality of the first data records; the ratio of the total number of records of the first training set to the first evaluation set satisfies the first segmentation ratio; Step 62, taking the first first data record of the first training set as the corresponding current training record; Step 63, inputting the first training molecular graph of the current training record as the corresponding raw material molecular graph into the synthesizable optimization model for synthesizable optimization processing; and during the model processing, the marker type s generated by the model prediction at each step i Data collection is performed, and for each marker type s specifically a building block marker i The corresponding predicted fingerprint p pre,i Data collection is performed, and for each marker type s specifically a reaction marker i The corresponding prediction type r pre,i Perform data collection; and based on all collected data and the maximum label index j of the first label sequence of the current training record max Performing prediction sequence assembly to obtain a corresponding first prediction sequence; Among them, j max = max(j); the first prediction sequence is composed of a plurality of first prediction arrays; the first prediction array includes a prediction marker type s j and predicted data d j ; The predicted marker type s j Including building block markers, reaction markers, and end markers; the prediction marker type s j When the block marker is constructed, the prediction data d j is a corresponding predicted fingerprint p pre,i ; The predicted marker type s j When it is a reaction marker, the predicted data d j is a corresponding prediction type r pre,i ; The predicted marker type s j When it is the end marker, the predicted data d j Empty; if the tag type s collected i The total number of tags is less than the maximum tag index j max , then the first prediction sequence is supplemented with a prediction array based on a preset invalid prediction array until the first prediction sequence is consistent with the sequence length of the current first label sequence; if the collected marker type s i The total number is greater than the maximum label index j max , then based on the invalid prediction array, the current first label sequence is label-filled until the current first label sequence is consistent with the first prediction sequence and the sequence length is consistent with the first prediction sequence, and the maximum label index j is filled based on the sequence length of the first prediction sequence. max Perform a reset; Step 64: Substitute the first prediction sequence and the current first label sequence into a preset first model loss function L M1 ; and based on the preset first model optimizer, move towards making the first model loss function L M1 Performing a round of modulation on the model parameters of the encoder, the decoder, the molecular fingerprint prediction module and the reaction rule prediction module of the synthesizable optimization model in the direction of reaching the minimum value; Among them, the first model loss function L M1 By the first type loss function L S , the first fingerprint loss function L P And the first rule loss function L R composition; L M1 =L S +L P +L R , Indicator function Indicator function L CE () is the cross entropy loss function, L BCE () is a binary cross entropy loss function; n is the type of the label marker in the current first label sequence, specifically the building block marker The total number of, m is the tag marker type of the current first tag sequence specifically the reaction marker Total number of; The first model optimizer includes at least an SGD optimizer and an Adam optimizer; Step 65, identifying whether the current training record is the last first data record of the first training set; if so, going to step 66; if not, extracting the next first data record of the first training set as the new current training record, and returning to step 63; Step 66, perform a round of traversal on all the first data records of the first evaluation set; and in this round of traversal, use the first data record currently traversed as the corresponding current evaluation record; and input the first training molecular graph of the current evaluation record as the corresponding raw material molecular graph into the synthesizable optimization model for synthesizable optimization processing; and in the process of model processing, the marker type s generated by the model prediction at each step i Data collection is performed, and for each marker type s specifically a building block marker i The corresponding predicted fingerprint p pre,i Data collection is performed, and for each marker type s specifically a reaction marker i The corresponding prediction type r pre,i Perform data collection; and based on all collected data and the maximum tag index j of the first tag sequence of the current evaluation record max Perform prediction sequence assembly to obtain a corresponding second prediction sequence; and form a corresponding prediction-label sequence pair from the second prediction sequence and the current first label sequence; and at the end of this round of traversal, estimate the accuracy, precision, recall and F1 score based on all the obtained prediction-label sequence pairs to obtain the corresponding first accuracy, first precision, first recall and first F1 score; Step 67, identifying the first accuracy, the first precision, the first recall and the first F1 score; if the first accuracy does not meet the preset first accuracy range, or the first accuracy does not meet the preset first accuracy range, or the first recall does not meet the preset first recall range, or the first F1 score does not meet the preset first F1 score range, return to step 62 to continue training; if the first accuracy meets the first accuracy range, and the first accuracy meets the first accuracy range, and the first recall meets the first recall range, and the first F1 score meets the first F1 score range, stop training and confirm that the model training is completed.

7. The method for synthesizing and optimizing material molecules according to claim 3, characterized in that: The synthesizable optimization model is used to perform synthesizable optimization processing on the molecular graph of any material molecule input by the user to obtain the corresponding optimized material molecular graph and feed it back to the current user, specifically including: The molecular graph of any material molecule input by the user is input into the synthesizable optimization model as the corresponding original material molecular graph to perform synthesizable optimization processing to obtain the corresponding optimized material molecular graph; and the optimized material molecular graph is fed back to the current user as the corresponding optimized material molecular graph.

8. A device for executing the method for processing material molecules for synthesis optimization as claimed in any one of claims 1 to 7, characterized in that: The device comprises: a chemical information library and model building module, a model training module and a model application module; The chemical information library and model building module are used to set the building block structure library and reaction rule library corresponding to the material molecule; and the building block structure library and the reaction rule library form a corresponding synthesizable chemical information library; and a processing model for improving the synthesizability of the material molecule is constructed in combination with the synthesizable chemical information library, which is recorded as a corresponding synthesizable optimization model; The model training module is used to construct a model training data set recorded as a corresponding first data set; and train the synthesizable optimization model based on the first data set; The model application module is used to perform synthesizable optimization processing on the molecular graph of any material molecule input by the user based on the synthesizable optimization model after the model training is completed, obtain the corresponding optimized material molecular graph and feed it back to the current user.

9. An electronic device, characterized in that: include: memory, processors, and transceivers; The processor is used to couple with the memory, read and execute instructions in the memory, so as to implement the method according to any one of claims 1 to 7; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a computer, the computer is enabled to execute the method according to any one of claims 1 to 7.

Images