A method for joint optimization of synthetic and conversion rate prediction of drug chemical reactions
By jointly training a drug chemical reaction synthesis and conversion rate prediction task, and utilizing hierarchical sequence encoding and uncertainty estimation, the problems of information correlation neglect and insufficient interpretability in existing methods are solved, thus achieving more efficient compound development.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH BEIJING
- Filing Date
- 2022-12-05
- Publication Date
- 2026-06-05
AI Technical Summary
Existing methods for predicting the synthesis and conversion rates of medicinal chemical reactions ignore the correlation between the two and lack interpretability and the ability to handle uncertainty.
We employ hierarchical sequence coding technology combined with ON-LSTM and uncertainty estimation. By jointly training the task of drug chemical reaction synthesis and conversion rate prediction, we use ON-LSTM for hierarchical sequence coding feature extraction and normal inverse gamma distribution for uncertainty estimation. We construct a multi-task loss function to optimize the model.
It improves the interpretability of the model and its robustness to uncertain data, reduces the cost of compound research and development, and improves the efficiency of compound research and development.
Smart Images
Figure CN115831246B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of pharmaceutical chemical reaction synthesis and conversion rate prediction technology, and in particular to a joint optimization method for pharmaceutical chemical reaction synthesis and conversion rate prediction. Background Technology
[0002] With the advent of the big data era, people have high hopes for the ability of deep learning to analyze large amounts of existing chemical data and derive models to predict various aspects of chemical reactions. Among these, chemical reaction synthesis prediction and conversion rate prediction are crucial and challenging problems in chemical reaction problems. Chemical reaction synthesis is the process of constructing a target product from a set of existing reactants and reagents. Since chemical reactions require a significant amount of time and money, it is particularly important to use machines to predict the products of chemical reactions. For the general process of chemical reaction synthesis reasoning, the model constructs the product by analyzing the chemical properties of the reactants and combining them with reaction conditions. In order to extend its application to new chemical reactants, the model needs to have a stronger adaptability to uncommon chemical reaction types or even new categories. The chemical reaction synthesis prediction task is to predict the product in a given pharmaceutical chemical reaction expression, given the reactants. For this task, recurrent neural networks and attention-based Transformer models are generally used. Recurrent neural networks emphasize the long-term temporal context dependence of chemical reaction expressions, while Transformer models mainly focus on global structural information.
[0003] Chemical reaction conversion rate prediction is the problem of inferring the conversion rate of reactants from a chemical reaction equation. Conversion rate is the ratio of the amount of reactants converted to the total amount of reactants. Considering economic and time factors, chemists tend to find reaction pathways with higher conversion rates. However, due to the complexity of chemical reactions, there are many pathways from reactants to products. Therefore, designing chemical synthesis pathways with high conversion rates usually requires a significant amount of time for exploration. Thus, using deep learning to predict the conversion rate of chemical reactions in a short time is particularly important. Similar to chemical reaction synthesis prediction, chemical reaction conversion rate prediction is generally based on recurrent neural networks and Transformers.
[0004] Current research on these two tasks using deep learning employs relatively basic network architectures and trains models with large amounts of data to obtain satisfactory results. However, in real-world scenarios, considering the safety and reliability requirements of medicinal chemical production, there is an urgent need to develop interpretable model structures and learning algorithm designs. Furthermore, existing methods lack corresponding mechanisms to address the high uncertainty in the synthesis results and conversion rates of chemical products caused by measurement errors and changes in the experimental environment in real-world data.
[0005] Once chemical formulas are converted to the SMILES format, existing sequence model algorithms can be readily adapted for tasks involving chemical reaction synthesis prediction and conversion rate prediction. Currently, commonly used models for chemical reaction synthesis prediction and conversion rate prediction include those utilizing recurrent neural networks and Transformer models based on attention mechanisms.
[0006] Recurrent Neural Networks (RNNs) are neural networks with short-term memory capabilities. Neurons can receive information from other neurons and from themselves, forming a network structure with loops. This model is naturally suited for processing sequential information. It recursively calculates the output of the next time step based on the output of the previous time step, ultimately outputting a sequence of arbitrary length. The network is trained using linear sequences of chemical formulas under the SMILES rule, guiding the model to find the products of chemical reactions. However, commonly used RNNs can only model temporal processes and struggle to extract hierarchical structural information of chemical reactions. Information from different levels, such as functional groups and chemical bonds, is crucial for the representation of chemical formulas. The Transformer is a model based on a multi-head attention mechanism. Its structure abandons traditional CNNs and RNNs; the entire network structure is based entirely on attention mechanisms. In the problem of predicting chemical reaction synthesis, the Transformer-based method does not model through hidden information transmission but directly calculates the relationships between each sequence feature, enabling simultaneous information transmission between chemical molecule sequences. It focuses on the global feature representation of the sequence, similarly ignoring local hierarchical information representation.
[0007] In summary, existing techniques for predicting the synthesis and conversion rates of chemical reactions neglect the interconnected information that these two tasks can provide. Both tasks are highly correlated with bond breaking and functional group recombination, while jointly learning these two tasks is beneficial for guiding model parameters towards the desired, interpretable objective. Furthermore, existing techniques have not yet considered how to handle uncertain data. Summary of the Invention
[0008] This invention provides a joint optimization method for predicting the synthesis and conversion rate of pharmaceutical chemical reactions, which addresses the technical problems of existing methods for predicting pharmaceutical chemical reactions neglecting the interconnected information that can be provided by the two tasks of predicting chemical reaction synthesis and conversion rate, and the lack of interpretability and consideration of uncertainties in chemical reactions.
[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0010] On one hand, the present invention provides a joint optimization method for drug chemical reaction synthesis and conversion rate prediction, the joint optimization method for drug chemical reaction synthesis and conversion rate prediction includes:
[0011] The SMILES expressions of the reactants in the drug chemical reaction are obtained, and the SMILES expressions of the reactants are segmented into words. The word segments are then embedded into the expression to obtain the word segmentation features of the SMILES expressions of the reactants.
[0012] Hierarchical sequence encoding is performed on the word segmentation features of the reactants' SMILES expressions to obtain the hierarchical sequence encoding features of the reactants, thereby increasing the weight of atoms surrounding the chemical bonds that undergo changes in the chemical formula;
[0013] The synthesis prediction and conversion rate prediction tasks of medicinal chemical reactions are combined and trained simultaneously to achieve joint optimization of the synthesis and conversion rate prediction tasks. The synthesis prediction task predicts the product based on the hierarchical sequence coding features of the reactants, while the conversion rate prediction task predicts the conversion rate based on the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the product output by the synthesis prediction task.
[0014] Furthermore, hierarchical sequence encoding is performed on the word segmentation features of the reactant's SMILES expression, including:
[0015] The word segmentation features of the SMILES expression of the reactants are used for hierarchical sequence encoding using ON-LSTM.
[0016] Furthermore, hierarchical sequence encoding of the word segmentation features of the reactant's SMILES expression also includes:
[0017] A three-layer ON-LSTM was used to hierarchically encode the SMILES expression of the reactants, and the hierarchical encoding features were learned by the intermediate layer. The hierarchical encoding features were obtained by weighted summation of the latent variable features of the intermediate layer.
[0018] Furthermore, when obtaining hierarchical coding features by weighted summation of the latent variable features of the intermediate layer, the magnitude of the weights quantifies the probability of chemical bond breakage in drug chemical reactions.
[0019] Furthermore, the method of combining the tasks of drug chemical reaction synthesis prediction and conversion rate prediction, and simultaneously training both tasks to achieve joint optimization of drug chemical reaction synthesis and conversion rate prediction, includes:
[0020] This paper combines the tasks of predicting the synthesis and conversion rate of medicinal chemical reactions, training both tasks simultaneously to construct a joint prediction model for the synthesis and conversion rate of medicinal chemical reactions; it includes:
[0021] ON-LSTM is used as the generator network to predict the products based on the hierarchical sequence encoding features of the reactants. During the training of the generator network, the input of the previous time step is directly used as the true value to prevent error accumulation. During inference, the product sequence generated is generated by bundle search regression, thereby completing the retrosynthesis of the compound.
[0022] After obtaining the predicted results of the products, the hierarchical sequence coding features of the products are calculated; and after obtaining the hierarchical sequence coding features of the products, the hierarchical sequence coding features of the reactants are spliced and fused with the hierarchical sequence coding features of the products to obtain the global features of the chemical reaction expression.
[0023] Based on the global features of the obtained chemical reaction expression, a conversion rate prediction network is used to predict the conversion rate of drug chemical reactions. During the training process, evidence-based deep learning is used to sample the predicted conversion rate values from the normal inverse gamma distribution, and a multi-task loss function is used to train the model, so that the model can accurately predict while having the ability to estimate uncertainty.
[0024] Furthermore, the generating network is a three-layer ON-LSTM.
[0025] Further, the hierarchical sequence coding features of the product output by the generating network are calculated, including:
[0026] Extract the latent variable features of the intermediate layers of the generated network;
[0027] The hierarchical coding features of the product are obtained by weighted summation of the latent variable features of the intermediate layers of the generator network.
[0028] Furthermore, the conversion rate prediction value p ~ NIG(α, β, γ, ν) is sampled from the normal inverse gamma distribution; where α, β, γ, and ν are parameters in the distribution, and 0 < γ < 1, β > 0, ν > 0, α > 1. γ and α, β, and ν are calculated by two independent multilayer perceptrons. According to the constraints, the outputs of the two networks are activated by the Sigmoid and Softplus functions, respectively. The optimization objective for the overall uncertainty estimation of the model is:
[0029] L = L BM +kL UE
[0030] Among them, L BM The cross-entropy loss is the point estimate γ and the true value, where k is a hyperparameter;
[0031] L UE =L NLL +L LMSE +C
[0032] Among them, L NLL L is the negative logarithm of the marginal likelihood.LMSE Let C be the Lipschitz-corrected mean squared error (MSE) loss function, and C be the evidence regularization term.
[0033] In another aspect, the present invention also provides an electronic device comprising a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described method.
[0034] In another aspect, the present invention also provides a computer-readable storage medium storing at least one instruction that is loaded and executed by a processor to implement the above-described method.
[0035] The beneficial effects of the technical solution provided by this invention include at least the following:
[0036] The advantages of this invention are that it combines the tasks of chemical reaction synthesis prediction and conversion rate prediction, trains both tasks simultaneously, and allows them to mutually guide and update the network, resulting in strong interpretability and robustness to uncertain data. The joint optimization method for drug chemical reaction synthesis and conversion rate prediction of this invention has significant application value in scenarios that aim to reduce compound development costs and improve compound development efficiency, laying a model foundation for the widespread application of comprehensive and accurate compound reaction synthesis and conversion rate prediction. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is a schematic diagram illustrating the implementation principle of the combined optimization method for drug chemical reaction synthesis and conversion rate prediction provided in this embodiment of the invention;
[0039] Figure 2 This is a schematic diagram of the execution flow of the joint optimization method for drug chemical reaction synthesis and conversion rate prediction provided in the embodiments of the present invention;
[0040] Figure 3 This is the molecular diagram of ciprofloxacin. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0042] First Embodiment
[0043] To address the shortcomings of existing methods for predicting pharmaceutical chemical reactions, such as limited interpretability and failure to consider uncertainties in the reaction process, this embodiment provides a joint optimization method for predicting the synthesis and conversion rates of pharmaceutical chemical reactions based on uncertainty estimation. Figure 1 As shown, the method consists of three parts: chemical formula SMILES format conversion, hierarchical sequence modeling, and uncertainty-based joint optimization. The method jointly learns two interrelated tasks: drug chemical reaction synthesis prediction and conversion rate prediction. During the learning process, it provides guidance information on the interpretability of the model based on the tasks and uses uncertainty estimation methods to train the model.
[0044] Specifically, the execution flow of this method is as follows: Figure 2 As shown, it includes the following steps:
[0045] S1, obtain the SMILES expression of the reactants in the drug chemical reaction, segment the SMILES expression of the reactants into words, and embed the segmented words into the expression to obtain the segmentation features of the SMILES expression of the reactants;
[0046] It should be noted that the method in this embodiment first requires converting the reactants and products of the chemical reaction from molecular diagrams into SMILES molecular format, for example, for Figure 3 Ciprofloxacin, in particular, has the following SMILES format: N1CCN(CC1)C(C(F)=C2)=CC(=C2C4=O)N(C3CC3)C=C4C(=O)O. The reactant's SMILES expression is segmented, and the segmented words are embedded to obtain the segmentation features of the reactant's SMILES expression. The embedding matrix is then connected to the model in subsequent steps for end-to-end learning.
[0047] S2, hierarchical sequence encoding is performed on the word segmentation features of the SMILES expression of the reactants to obtain the hierarchical sequence encoding features of the reactants, so as to increase the weight of atoms around the chemical bonds that change in the chemical formula;
[0048] It should be noted that the method in this embodiment requires sequence encoding of the word segmentation features of the chemical formula SMILES expression. Since chemical reactions generally do not completely disrupt the order of atoms, but rather break some chemical bonds and recombine functional groups, it is necessary to adaptively learn the information of functional groups where chemical bonds are broken, thereby improving the model's ability to extract information and increasing computational efficiency. To this end, this embodiment utilizes ON-LSTM (ordered neurons LSTM) for multi-level encoding, increasing the weights of atoms surrounding the chemical bonds that undergo changes in the chemical formula, and representing the chemical expression according to a certain semantic hierarchy. ON-LSTM is a language model that can adaptively encode multiple hierarchical structures, such as complete sentence information, phrase-level information, and word segmentation-level information. In fact, these hierarchical levels are implicit and have no explicit meaning. After training the model based on different tasks, its hierarchical semantics will exhibit task-related characteristics. ON-LSTM is a variant of the Long Short-Term Memory network LSTM, and its input and output are the same as LSTM.
[0049] Specifically, in this embodiment, a three-layer ON-LSTM is used to hierarchically encode the SMILES expressions corresponding to the chemical formulas of the reactants. Since the first and third layers are affected by input and output and cannot learn good hierarchical features, this embodiment uses an intermediate layer to learn the hierarchical encoding features. Let the latent variable feature of the intermediate layer be h. t Hierarchical coding features are obtained by weighted summation of latent variable features:
[0050]
[0051] Among them, a t As the weights, based on the master forget gate f in ON-LSTM t The calculation yielded:
[0052]
[0053] Where D represents the feature dimension of the forget gate. Weight a t The larger the value, the more likely it is to be a high-level semantic segmentation point. In this embodiment, since the model is guided by chemical reaction synthesis prediction and conversion rate prediction tasks, the magnitude of this weight quantifies the probability of chemical bond breaking in pharmaceutical chemical reactions.
[0054] S3 combines the tasks of drug chemical reaction synthesis prediction and conversion rate prediction, and trains both tasks simultaneously to achieve joint optimization of drug chemical reaction synthesis and conversion rate prediction. The synthesis prediction task predicts the product based on the hierarchical sequence coding features of the reactants; the conversion rate prediction task predicts the conversion rate based on the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the product output by the synthesis prediction task.
[0055] It should be noted that, since both the chemical synthesis process and the conversion rate of a chemical reaction depend on the breaking of these chemical bonds, jointly learning these two interrelated tasks—synthesis and conversion rate prediction—can improve the model's expressive power and effectiveness. This embodiment uses the functional group hierarchy in hierarchical chemical formula sequence modeling to unsupervisedly locate the positions of chemical bond breaks, thereby finding key information affecting chemical reaction synthesis and conversion rates, making the model interpretable.
[0056] Based on the above, in this embodiment, the implementation process of S3 is as follows:
[0057] This paper combines the tasks of predicting the synthesis and conversion rate of medicinal chemical reactions, training both tasks simultaneously to construct a joint prediction model for the synthesis and conversion rate of medicinal chemical reactions; it includes:
[0058] 1. Prediction of drug chemical reaction synthesis
[0059] After hierarchical sequence encoding of the reactants, ON-LSTM is used as the sequence generation network model to predict the products based on the hierarchical sequence encoding features of the reactants. During training, the input from the previous time step is directly taken as the true value to prevent error accumulation. During inference, the product sequence generated by bundle search regression is used to complete the retrosynthesis of the compound. After obtaining the prediction results of the product, its hierarchical sequence encoding features are calculated.
[0060] Specifically, in this embodiment, a three-layer ON-LSTM is used as the decoder, i.e., the generative network model, to generate the product. During training, the input to the decoder is the hierarchical encoded features x of the reactants. enc The decoder's input at each time step is the true value of the output from the previous step. That is, given the SMILES expression features at the word segmentation level of the previous time step, the decoder outputs the features of the current time step. For the next step of reaction conversion rate prediction, the latent variable features of the intermediate layer of the decoder are extracted, and the hierarchical sequence code y of the product is calculated in the same way as the reaction hierarchical coding features calculated in S2. enc The product sequences were generated using bundle search regression during testing.
[0061] 2. Prediction of drug chemical reaction conversion rate
[0062] After obtaining the hierarchical sequence coding features of the products, the hierarchical sequence coding features of the reactants and the products are concatenated and fused to obtain the global feature of the chemical reaction expression g = [x enc y enc After obtaining the global features of the expression, a conversion rate prediction network model is used to predict the conversion rate of drug chemical reactions based on uncertainty. During the training process, evidence-based deep learning is used to sample the predicted conversion rate values from the normal inverse gamma distribution, and the model is trained using a multi-task loss function, so that the model can accurately predict while having the ability to estimate uncertainty.
[0063] Specifically, in this embodiment, in order to estimate the uncertainties in the reaction process and the real environment, the proposed model is trained using evidence-based deep learning by sampling the predicted conversion rate p from a normal inverse gamma distribution:
[0064] p~NIG(α,β,γ,ν),
[0065] Where (α, β, γ, δ) are parameters in the distribution, and 0 < γ < 1, ν, β > 0, α > 1. γ and α, β, ν are calculated by two independent multilayer perceptrons. According to the constraints, the outputs of the two networks are activated by the Sigmoid and Softplus functions, respectively. The optimization objective for the overall uncertainty estimation is:
[0066] L = L BM +kL UE ,
[0067] Among them, L BM The cross-entropy loss is the point estimate γ and the true value, where k is a hyperparameter;
[0068] L UE =L NLL +L LMSE +C
[0069] Among them, L NLL L is the negative logarithm of the marginal likelihood. LMSE Let C be the Lipschitz-corrected mean squared error (MSE) loss function, and C be the evidence regularization term.
[0070] In summary, the method in this embodiment combines two tasks: chemical reaction synthesis prediction and conversion rate prediction. It introduces hierarchical sequence modeling technology, allowing the two tasks to mutually guide the optimization of model interpretability parameters, thereby improving model training efficiency and performance. Furthermore, the method in this embodiment incorporates uncertainty estimation to address the interference caused by uncertain data in real-world scenarios. Therefore, the method in this embodiment has significant application value in scenarios that aim to reduce compound development costs and improve compound development efficiency, laying a model foundation for the widespread application of comprehensive and accurate compound reaction synthesis and conversion rate prediction.
[0071] Second Embodiment
[0072] This embodiment provides an electronic device, which includes a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment.
[0073] The electronic device can vary considerably depending on its configuration or performance, and may include one or more processors (central processing units, CPUs) and one or more memories, wherein the memories store at least one instruction that is loaded by the processor and executed in accordance with the above method.
[0074] Third Embodiment
[0075] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc. The instruction stored therein can be loaded and executed by a processor in a terminal.
[0076] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.
[0077] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0079] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.
[0080] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.
Claims
1. A method for joint optimization of pharmaceutical chemical reaction synthesis and conversion rate prediction, characterized in that, include: The SMILES expressions of the reactants in the drug chemical reaction are obtained, and the SMILES expressions of the reactants are segmented into words. The word segments are then embedded into the expression to obtain the word segmentation features of the SMILES expressions of the reactants. Hierarchical sequence encoding is performed on the word segmentation features of the reactants' SMILES expressions to obtain the hierarchical sequence encoding features of the reactants, thereby increasing the weight of atoms surrounding the chemical bonds that undergo changes in the chemical formula; The synthesis prediction and conversion rate prediction tasks of medicinal chemical reactions are combined and trained simultaneously to achieve joint optimization of the synthesis and conversion rate prediction tasks. The synthesis prediction task predicts the product based on the hierarchical sequence coding features of the reactants, while the conversion rate prediction task predicts the conversion rate based on the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the product output by the synthesis prediction task. The method of combining the tasks of drug chemical reaction synthesis prediction and conversion rate prediction, and training both tasks simultaneously to achieve joint optimization of drug chemical reaction synthesis and conversion rate prediction, includes: This paper combines the tasks of predicting the synthesis and conversion rate of medicinal chemical reactions, training both tasks simultaneously to construct a joint prediction model for the synthesis and conversion rate of medicinal chemical reactions; it includes: ON-LSTM is used as the generator network to predict the products based on the hierarchical sequence encoding features of the reactants. During the training of the generator network, the input of the previous time step is directly used as the true value to prevent error accumulation. During inference, the product sequence generated is generated by bundle search regression, thereby completing the retrosynthesis of the compound. After obtaining the predicted results of the products, the hierarchical sequence coding features of the products are calculated; and after obtaining the hierarchical sequence coding features of the products, the hierarchical sequence coding features of the reactants are spliced and fused with the hierarchical sequence coding features of the products to obtain the global features of the chemical reaction expression. Based on the global features of the obtained chemical reaction expression, a conversion rate prediction network is used to predict the conversion rate of drug chemical reactions. During the training process, evidence-based deep learning is used to sample the predicted conversion rate values from the normal inverse gamma distribution, and a multi-task loss function is used to train the model, so that the model can accurately predict while having the ability to estimate uncertainty. The predicted conversion rate p ~ NIG(α,β,γ,v) is sampled from a normal inverse gamma distribution; where α, β, γ, and v are parameters in the distribution, and 0 < γ < 1, β > 0, v > 0, and α > 1. γ and α, β, and v are calculated by two independent multilayer perceptrons. According to the constraints, the outputs of the two networks are activated by the Sigmoid and Softplus functions, respectively. The optimization objective for the overall uncertainty estimation of the model is: L=L BM +kL UE Among them, L BM The cross-entropy loss is the point estimate γ and the true value, where k is a hyperparameter; L UE L NLL +L LMSE +C Among them, L NLL L is the negative logarithm of the marginal likelihood. LMSE Let C be the Lipschitz-corrected mean squared error (MSE) loss function, and C be the evidence regularization term.
2. The method for joint optimization of drug chemical reaction synthesis and conversion rate prediction as described in claim 1, characterized in that, Hierarchical sequence encoding is performed on the word segmentation features of the reactant's SMILES expression, including: The word segmentation features of the SMILES expression of the reactants are used for hierarchical sequence encoding using ON-LSTM.
3. The method for joint optimization of drug chemical reaction synthesis and conversion rate prediction as described in claim 1, characterized in that, Hierarchical sequence encoding is performed on the word segmentation features of the reactant's SMILES expression, including: A three-layer ON-LSTM is used to hierarchically encode the SMILES expression of the reactants, and the hierarchical encoding features are learned by the intermediate layer. The hierarchical encoding features are obtained by weighted summation of the latent variable features of the intermediate layer.
4. The method for joint optimization of drug chemical reaction synthesis and conversion rate prediction as described in claim 3, characterized in that, When obtaining hierarchical coding features by weighted summation of the latent variable features of the intermediate layer, the magnitude of the weights quantifies the probability of chemical bond breakage in drug chemical reactions.
5. The method for joint optimization of drug chemical reaction synthesis and conversion rate prediction as described in claim 1, characterized in that, The generating network is a three-layer ON-LSTM.
6. The method for joint optimization of drug chemical reaction synthesis and conversion rate prediction as described in claim 5, characterized in that, Calculating the hierarchical sequence coding features of the product output by the generating network includes: Extract the latent variable features of the intermediate layers of the generated network; The hierarchical coding features of the product are obtained by weighted summation of the latent variable features of the intermediate layers of the generator network.