Inverse synthesis prediction method based on similarity feature constraint, medium and device
By incorporating the graph structure features of product molecules into retrosynthesis prediction and using feature similarity constraints of graph neural networks and Transformer models for training, the problem of insufficient generalization of existing methods is solved, and the prediction accuracy and efficiency are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2022-11-23
- Publication Date
- 2026-06-02
AI Technical Summary
Existing template-based retrosynthesis prediction methods lack generalization ability, while deep learning-based retrosynthesis prediction methods fail to fully consider molecular structure information, and the accuracy of end-to-end methods is limited.
By incorporating the graph structure features of the product molecules into the prediction model, the graph structure features of the product SMILES sequence are extracted using a graph neural network and fused with word vectors. The Transformer model is then used for feature similarity constraint training to form a loss function to improve prediction accuracy.
It improves the accuracy of retrosynthetic prediction, captures common features between product and reactant sets, saves training time, and improves the model's prediction accuracy.
Smart Images
Figure CN115758758B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of inverse synthesis prediction technology, and more specifically, to an inverse synthesis prediction method, medium, and device based on similarity feature constraints. Background Technology
[0002] Organic synthesis has become one of the most important disciplines in chemistry, with its research spanning materials science, energy, life sciences, and other fields, playing a vital role in social development and daily life. Organic synthesis refers to the process of using chemical methods to transform elements, simple inorganic substances, or simple organic compounds into more complex organic compounds. In recent years, computer-aided synthetic planning (CASP) technology has developed rapidly, especially retrosynthetic design, which has greatly facilitated chemists in drug synthesis. Retrosynthetic design aims to find a series of commercially available reactants for a given product molecule; retrosynthetic prediction, on the other hand, predicts the reaction process of a given target organic compound.
[0003] Traditional retrosynthetic prediction methods are mostly template-based. These methods extract a retrosynthetic reaction center template from existing chemical reactions, then match the products with the extracted template, and predict the set of reactants based on the matched template. However, because templates extracted by template-based methods are based on known chemical reactions, some reactions will inevitably fail to match the extracted templates. For these reactions, the retrosynthetic prediction results obtained by the model are usually unsatisfactory. Therefore, template-based methods suffer from insufficient generalization.
[0004] In recent years, with the rise and development of artificial intelligence, new attempts to apply deep learning have gradually come into view, and the task of inverse synthesis prediction has also been regarded as a deep learning task for research. Methods using deep learning for inverse synthesis prediction are mainly divided into two categories: one is based on inverse synthesis analysis, and the other is end-to-end methods. Methods based on inverse synthesis analysis divide the inverse synthesis prediction task into two steps: the first step uses a model to identify the reaction centers of the products, and then disconnects the reaction centers to obtain multiple synthons; the second step uses another model to convert the multiple synthons into the corresponding set of reactants. Methods based on inverse synthesis analysis have a certain degree of interpretability, but the number of reaction centers can only be limited to a preset value, and it is overly dependent on the accuracy of the first step. End-to-end methods aim to achieve the inverse synthesis prediction task in one step; current end-to-end methods are basically based on Transformer models (sequence-to-sequence models).
[0005] The Transformer model, proposed by Google in 2017, is a classic model in the field of natural language processing. The Transformer model consists of N encoders and N decoders. The source language sentence, after word embedding and positional encoding, is first input into N cascaded encoders, and then the encoded results are input into N cascaded decoders for decoding. Finally, the target language sentence is obtained based on the output probabilities. Methods based on the Transformer model treat inverse synthesis prediction as a translation task. Specifically, both products and reactants can be represented as a defined sequence called the Simplified Molecular Input Line Entry Specification (SMILES). Therefore, inverse synthesis prediction can be viewed as a conversion process from product SMILES sequences to reactant SMILES sequences. Sequence-to-sequence model-based methods can achieve inverse synthesis prediction in one step, but using only SMILES sequences cannot fully consider the structural information of the molecules.
[0006] In summary, there is still room for improvement in current deep learning-based inverse synthesis prediction methods. Summary of the Invention
[0007] To overcome the shortcomings and deficiencies of existing technologies, the present invention aims to provide a retrosynthetic prediction method, medium, and device based on similarity feature constraints. This method incorporates the graph structure features of product molecules into the prediction model, thereby improving the prediction accuracy. By learning the commonalities between reactant and product sets through feature constraints, the common features are used to better assist the Transformer model in retrosynthetic prediction, further improving the model's prediction accuracy.
[0008] To achieve the above objectives, the present invention provides the following technical solution: an inverse synthesis prediction method based on similarity feature constraints, comprising the following steps:
[0009] S1. Convert the product to be predicted into the corresponding product SMILES sequence; use a graph neural network to extract the graph structure features h corresponding to all characters in the product SMILES sequence. i ′;The graph structure feature h i The word vector representation of the product SMILES sequence is fused with the word vector representation of the product SMILES sequence to obtain the fused vector t for each character in the product SMILES sequence. i ′, and then obtain the fusion vector matrix corresponding to the product SMILES sequence;
[0010] S2. Input the fusion vector matrix corresponding to the product SMILES sequence into the Transformer encoder; then input the encoding result of the Transformer encoder into the Transformer decoder for decoding to obtain the reactant set SMILES sequence; the Transformer encoder and Transformer decoder refer to the trained Transformer encoder and Transformer decoder; in the training of the Transformer encoder and Transformer decoder, the cosine function is used to measure the similarity between the fusion vector of the product SMILES sequence and the word vector representation of the reactant set SMILES sequence, and the similarity loss is used as the feature similarity constraint of the original loss to form the loss function L for training;
[0011] S3, the sequence of the reactant set SMILES is transformed to obtain the corresponding reactants, thereby obtaining the retrosynthetic prediction results.
[0012] Preferably, S1 refers to:
[0013] Convert the product to be predicted into the corresponding product SMILES sequence;
[0014] Obtain the atomic information, edge information, and adjacency information between atoms corresponding to each character in the product SMILES sequence; utilize the information h of atom i corresponding to the i-th character. i Information h of adjacent atom j j And edge information e ij Calculate the weight β ij After calculating the weights β of all adjacent atoms corresponding to the i-th character... ij Then, the information h of all adjacent atoms is... j The fusion is performed by weighted summation, and the graph structure feature h corresponding to the i-th character of the product SMILES sequence is calculated using the fused information. i ′;
[0015] The word vector t corresponding to the i-th character of the product SMILES sequence is represented by... i The graph structure feature h corresponding to the i-th character of the product SMILES sequence i The sequences are concatenated, and feature fusion is performed using matrix mapping to obtain the fusion vector t for each character in the resulting SMILES sequence. i ′, and then obtain the fusion vector matrix corresponding to the product SMILES sequence.
[0016] Preferably, the weight β ij Graph structural features h i ′ are respectively:
[0017]
[0018]
[0019] Where, β ij The symbol represents the weight of atom j relative to atom i; || represents the vector concatenation operation; h i / h j / h k Represents information about atoms i / j / k; e ij / e ik Represents the edge information between atom i and atom j / k; A i / A j Let A represent the set of all adjacent atoms of atom i / j, where atom j belongs to A. i ;r T and W g σ represents the learnable parameters; σ represents the activation function.
[0020] Preferably, the fusion vector t i 'for:
[0021] t i ′=W c [t i ||h i ′]
[0022] Among them, W c These are learnable parameters.
[0023] Preferably, the method for forming the loss function L is as follows:
[0024] The dataset contains m pairs of product-reactant sets of SMILES sequences; the fusion vector t of the product SMILES sequences in each product-reactant set is obtained respectively. i The word vector representation of the sequence of reactant sets SMILES. i ;
[0025] The fusion vector t corresponding to all characters in the SMILES sequence of the product in the k-th product-reactant pair is used. i The summation and averaging are performed to obtain the characteristic representation v of the entire product SMILES sequence. p k ;
[0026] Let t be the word vector representation of all characters in the SMILES sequence of the reactant set in the k-th product-reactant pair. i Summing and averaging yields the characteristic representation v of the reactant set SMILES sequence. r k ;
[0027] Calculate the cosine similarity s of the SMILES sequences for the k-th product-reactant pair. k And similarity loss l k ;
[0028] Calculate the original loss L t And similarity loss L s ;
[0029] The similarity loss L s As the original loss L t The feature similarity constraint is used to automatically balance the loss to form the loss function L.
[0030] Preferably, the characteristic representation v of the entire product SMILES sequence p k for:
[0031]
[0032] The characteristic representation of the reactant set SMILES sequence v r k for:
[0033]
[0034] Where, N k and M k represents the number of characters contained in the product SMILES sequence and the reactant SMILES sequence in the k-th product-reactant pair, respectively.
[0035] The cosine similarity s corresponding to the SMILES sequence of the k-th product-reactant set k And similarity loss l k They are respectively:
[0036]
[0037] l k =1-s k
[0038] The original loss L t And similarity loss L s They are respectively:
[0039]
[0040]
[0041] Where n is the total number of characters contained in the entire reactant set SMILES sequence, and y′ iThis represents the predicted probability corresponding to the true label of the i-th character out of n characters.
[0042] Preferably, the loss function L is:
[0043]
[0044] Here, α1 and α2 are learnable weight coefficients.
[0045] A storage medium storing a computer program that, when executed by a processor, causes the processor to perform the aforementioned inverse synthesis prediction method based on similarity feature constraints.
[0046] A computing device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described inverse synthetic prediction method based on similarity feature constraints.
[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0048] 1. The retrosynthesis prediction method of this invention incorporates the graph structure features of product molecules into the prediction features, thereby improving the prediction accuracy. There are many common structures between reactants and product sets. Theoretically, there should be a certain similarity in the features between reactants and product sets. This invention learns the commonalities between reactants and product sets through feature constraints, and uses the common features to better help the Transformer model to perform retrosynthesis prediction, thereby further improving the prediction accuracy of the model.
[0049] 2. The retrosynthetic prediction method of this invention not only transfers the structural features of the correlation graph of the products to the feature representation of the reactant set through feature similarity constraints, capturing the common features between the products and the reactant set; it also balances the original loss L through weighting coefficients α1 and α2. t And similarity loss L s The weight coefficients α1 and α2 are automatically adjusted to match the original loss L at different stages of the training process. t And similarity loss L s The importance of the changes is reflected in the weight coefficients, which are more reasonable and save training time. Attached Figure Description
[0050] Figure 1 This is a schematic diagram of the model architecture of the inverse synthesis prediction method based on similarity feature constraints of the present invention;
[0051] Figure 2 This is a flowchart illustrating the inverse synthesis prediction method based on similarity feature constraints of the present invention.
[0052] Figure 3This is a schematic diagram of graph structure feature extraction for the inverse synthesis prediction method based on similarity feature constraints of the present invention. Detailed Implementation
[0053] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0054] Example 1
[0055] This embodiment presents an inverse synthesis prediction method based on similarity feature constraints, the model architecture of which is as follows: Figure 1 As shown, the method flow is as follows: Figure 2 As shown, it includes the following steps:
[0056] S1. Convert the product to be predicted into the corresponding product SMILES sequence; use a graph neural network to extract the graph structure features h corresponding to all characters in the product SMILES sequence. i ′;The graph structure feature h i The word vector representation of the product SMILES sequence is fused with the word vector representation of the product SMILES sequence to obtain the fused vector t for each character in the product SMILES sequence. i ′, and then obtain the fusion vector matrix corresponding to the product SMILES sequence.
[0057] Specifically, S1 refers to converting the product to be predicted into the corresponding product SMILES sequence;
[0058] Obtain the atomic information, edge information, and adjacency information between atoms corresponding to each character in the product SMILES sequence; utilize the information h of atom i corresponding to the i-th character. i Information h of adjacent atom j j And edge information e ij Calculate the weight β ij After calculating the weights β of all adjacent atoms corresponding to the i-th character... ij Then, the information h of all adjacent atoms is... j The fusion is performed by weighted summation, and the graph structure feature h corresponding to the i-th character of the product SMILES sequence is calculated using the fused information. i ′:
[0059]
[0060]
[0061] Where, β ij The symbol represents the weight of atom j relative to atom i; || represents the vector concatenation operation; h i / h j / h k Represents information about atoms i / j / k; e ij / e ik Represents the edge information between atom i and atom j / k; A i / A j Let A represent the set of all adjacent atoms of atom i / j, where atom j belongs to A. i ;r T and W g σ represents the learnable parameters; σ represents the activation function.
[0062] like Figure 3 As shown, the molecule corresponding to the SMILES sequence "S=C(Cl)Cl" contains two Cl atoms, one S atom, one C atom, two single bond edges, and one double bond edge. This atomic information, edge information, and adjacency information are input into a graph neural network, and the weight β is used... ij Graph structural features h i The formula for ' can be used to obtain the graph structure features of each character (atom);
[0063] The fusion of graph structure features aims to integrate graph structure features into a Transformer-based inverse synthesis prediction model. Assuming a SMILES sequence consists of N characters, these characters can be divided into two main categories: atomic characters and non-atomic characters. To fuse graph structure features with the word vector representation of the SMILES sequence, each character should have a corresponding graph structure feature. The graph structure features corresponding to atomic characters have already been obtained in the above process. For non-atomic characters, since they are not part of the molecular graph structure, their corresponding graph structure feature vectors are set to all zeros.
[0064] Next, feature fusion is performed: the word vector t corresponding to the i-th character of the product SMILES sequence is represented as... i The graph structure feature h corresponding to the i-th character of the product SMILES sequence i The sequences are concatenated, and feature fusion is performed using matrix mapping to obtain the fusion vector t for each character in the resulting SMILES sequence. i ′, and thus obtain the fusion vector matrix corresponding to the product SMILES sequence:
[0065] t i ′=W c [t i ||h i ′]
[0066] Among them, W c These are learnable parameters.
[0067] S2. Input the fusion vector matrix corresponding to the product SMILES sequence into the Transformer encoder; then input the encoding result of the Transformer encoder into the Transformer decoder for decoding to obtain the reactant set SMILES sequence.
[0068] S3, the sequence of the reactant set SMILES is transformed to obtain the corresponding reactants, thereby obtaining the retrosynthetic prediction results.
[0069] The Transformer is a classic model in the field of natural language processing, proposed by Google in 2017. The Transformer consists of N encoders and N decoders. After word embedding and positional encoding, the source language sentence is first input into the N concatenated encoders, and then the encoded results are input into the N concatenated decoders for decoding. Finally, the target language sentence is obtained based on the output probability.
[0070] The Transformer encoder and Transformer decoder refer to the trained Transformer encoder and Transformer decoder. In the training of the Transformer encoder and Transformer decoder, the cosine function is used to measure the similarity between the fusion vector of the product SMILES sequence and the word vector representation of the reactant set SMILES sequence, and the similarity loss is used as the feature similarity constraint of the original loss to form the loss function L for training.
[0071] The loss function L is formed as follows:
[0072] The dataset contains m pairs of product-reactant sets of SMILES sequences; the fusion vector t of the product SMILES sequences in each product-reactant set is obtained respectively. i The word vector representation of the sequence of reactant sets SMILES. i ;
[0073] The fusion vector t corresponding to all characters in the SMILES sequence of the product in the k-th product-reactant pair is used. i The summation and averaging are performed to obtain the characteristic representation v of the entire product SMILES sequence. p k :
[0074]
[0075] Let t be the word vector representation of all characters in the SMILES sequence of the reactant set in the k-th product-reactant pair. i Summing and averaging yields the characteristic representation v of the reactant set SMILES sequence.r k :
[0076]
[0077] Where, N k and M k represents the number of characters contained in the product SMILES sequence and the reactant SMILES sequence in the k-th product-reactant pair, respectively.
[0078] Calculate the cosine similarity s of the SMILES sequences for the k-th product-reactant pair. k And similarity loss l k :
[0079]
[0080] l k =1-s k
[0081] Calculate the original loss L t And similarity loss L s :
[0082]
[0083]
[0084] Where n is the total number of characters contained in the entire reactant set SMILES sequence, and y′ i This represents the predicted probability corresponding to the true label of the i-th character out of n characters;
[0085] If we consider the similarity loss L s As the original loss L t Given the feature similarity constraint, the loss function L is:
[0086] L=α1L t +α2L s
[0087] Here, α1 and α2 are weighting coefficients. By using feature similarity constraints, the structural features of the correlation graph of the products can be transferred to the feature representation of the reactant set, capturing the common features between the products and the reactant set.
[0088] From the similarity loss calculation formula, this loss tends to make the product SMILES sequences and the reactant set SMILES sequences completely identical in characteristics. However, the product and reactant sets should have both commonalities and unique characteristics. Therefore, the original loss L t And similarity loss L sThere exists an adversarial balancing process. This invention balances the two losses using weight coefficients α1 and α2. However, this method has two problems: First, α1 and α2 need to be manually set, and the parameters lack theoretical basis, requiring multiple experiments to find the optimal combination of α1 and α2, resulting in high time costs. Second, while the weight coefficients are fixed throughout the training process, the importance of the loss may change at different stages, making it more reasonable to use different weight coefficients at different stages. To address these problems, this invention, based on existing automated loss balancing technology, automatically adjusts the values of α1 and α2 during training, resulting in a new loss function L:
[0089]
[0090] Here, α1 and α2 are learnable weight coefficients. During the optimization of L, α1 and α2 are dynamically changing as the gradient is backpropagated. t and L s Its importance is also in a dynamic process of change.
[0091] This loss function not only transfers the correlation graph structure features of the products to the feature representation of the reactant set through feature similarity constraints, capturing the common features between the products and the reactant set; it also balances the original loss L using weight coefficients α1 and α2. t And similarity loss L s The weight coefficients α1 and α2 are automatically adjusted to match the original loss L at different stages of the training process. t And similarity loss L s The importance of the changes is reflected in the weight coefficients, which are more reasonable and save training time.
[0092] The advantage of this invention lies in addressing the fact that current retrosynthetic prediction methods do not consider the relationship between product and reactant sets. Building upon sequence-to-sequence models, it proposes a retrosynthetic prediction method based on feature similarity constraints. By learning the commonalities between products and reactants through similarity constraints, this common feature is used to better assist the model in retrosynthetic prediction, thereby improving the accuracy of the prediction results. Using this invention for reasonable retrosynthetic prediction helps chemists design synthetic routes for targeted molecules, greatly assisting them in designing pathways for synthesizing new molecules.
[0093] Example 2
[0094] This embodiment provides a storage medium, characterized in that the storage medium stores a computer program, which, when executed by a processor, causes the processor to execute the inverse synthesis prediction method based on similarity feature constraints as described in Embodiment 1.
[0095] Example 3
[0096] This embodiment provides a computing device, including a processor and a memory for storing processor-executable programs. The processor, when executing the program stored in the memory, implements the inverse synthesis prediction method based on similarity feature constraints as described in Embodiment 1.
[0097] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A method for inverse synthesis prediction based on similarity feature constraints, characterized in that: Includes the following steps: S1. Convert the product to be predicted into the corresponding product SMILES sequence; use a graph neural network to extract the graph structure features corresponding to all characters in the product SMILES sequence. ;Graph structural features By fusing the word vector representation of the product SMILES sequence with the word vector representation of the product SMILES sequence, a fused vector for each character in the product SMILES sequence is obtained. This leads to the fusion vector matrix corresponding to the product SMILES sequence; S2. Input the fusion vector matrix corresponding to the product SMILES sequence into the Transformer encoder; then input the encoding result of the Transformer encoder into the Transformer decoder for decoding to obtain the reactant set SMILES sequence; the Transformer encoder and Transformer decoder refer to the trained Transformer encoder and Transformer decoder; in the training of the Transformer encoder and Transformer decoder, the cosine function is used to measure the similarity between the fusion vector of the product SMILES sequence and the word vector representation of the reactant set SMILES sequence, and the similarity loss is used as the feature similarity constraint of the original loss to form the loss function. To conduct training; S3, the reaction set SMILES sequence is transformed to obtain the corresponding reactants, thereby obtaining the retrosynthetic prediction results; The loss function The formation method is: The dataset contains For the SMILES sequences of the product-reactant pairs, obtain the fusion vector of the product SMILES sequences for each pair of product-reactant pairs. Word vector representation of the reactant set SMILES sequence ; The fusion vector corresponding to all characters in the SMILES sequence of the product in the k-th product-reactant pair is... Summation and averaging are performed to obtain the characteristic representation of the entire product SMILES sequence. ; Represent the word vectors corresponding to all characters in the SMILES sequence of the reactant set in the k-th product-reactant pair. Summation and averaging yield the characteristic representation of the reactant set SMILES sequence. ; Calculate the first Cosine similarity of the SMILES sequences corresponding to the product-reactant sets and similarity loss ; Calculate the original loss and similarity loss ; Similarity loss As the original loss The feature similarity constraint is used to automatically balance the loss to form the loss function. ; The characteristic representation of the entire product SMILES sequence for: ； The characteristic representation of the reactant set SMILES sequence for: ； in, and They represent the first The number of characters contained in the product SMILES sequence and the reactant SMILES sequence in the product-reactant set; The first Cosine similarity of the SMILES sequences corresponding to the product-reactant sets and similarity loss They are respectively: ； ； The original loss and similarity loss They are respectively: ； ； in, The total number of characters contained in the SMILES sequence of all reactants. express The first character in the nth character The predicted probability corresponding to the true label of each character; The loss function for: ； in, and These are the learnable weight coefficients.
2. The inverse synthesis prediction method based on similarity feature constraints according to claim 1, characterized in that: S1 refers to: Convert the product to be predicted into the corresponding product SMILES sequence; Obtain the atomic information, edge information, and adjacency information between atoms corresponding to each character in the product SMILES sequence; use the first Atoms corresponding to each character Information Adjacent atoms Information and edge information Calculate the weights ; After calculating the first The weights of all adjacent atoms corresponding to each character Then, information about all adjacent atoms. The fusion was performed using weighted summation, and the SMILES sequence of the product was calculated using the fused information. Graph structure features corresponding to each character ; The product SMILES sequence number Word vector representation of each character and the product SMILES sequence number Graph structure features corresponding to each character The sequences are concatenated, and feature fusion is performed using matrix mapping to obtain the fused vector for each character in the resulting SMILES sequence. This leads to the fusion vector matrix corresponding to the product SMILES sequence.
3. The inverse synthesis prediction method based on similarity feature constraints according to claim 2, characterized in that: The weight Sum graph structural features They are respectively: ； ； in, Indicates the use of atoms to measure Relative to atoms The weights; This represents a vector concatenation operation; / / Represents atoms / / Information; / Represents atoms and atoms / Edge information between them; / Represents atoms / The set of all adjacent atoms, atoms belong ; and These are learnable parameters; This represents the activation function.
4. The inverse synthesis prediction method based on similarity feature constraints according to claim 2, characterized in that: The fusion vector for: ； in, These are learnable parameters.
5. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, causes the processor to perform the inverse synthesis prediction method based on similarity feature constraints as described in any one of claims 1-4.
6. A computing device, comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the inverse synthesis prediction method based on similarity feature constraints as described in any one of claims 1-4.