KR20220022059A - Apparatus for substructure-based neural machine translation for retrosynthetic prediction and method thereof - Google Patents

Abstract

With the rapid development of machine translation, neural network machine translation has played an important role in the field of retrosynthesis to find a rational synthesis route for a target molecule. Previous studies have shown that using the sequence-to-sequence framework of neural network machine translation is a promising way to solve a problem with retrosynthetic schemes. The present invention reconstructs a problem with retrosynthetic schemes as a problem with language translation by using a sequence-to-sequence model that does not require a template. A model of the embodiment is trained in an end-to-end manner and is a fully data-centered method. Unlike a conventional method for translating simplified molecular input line entry system (SMILES) strings of reactants and products, the present invention provides a method for representing chemical reactions based on molecular fragments. Therefore, more advanced prediction results can be calculated than existing calculation methods. In addition, an embodiment of the present invention can solve problems with existing retrosynthetic methods such as generating many invalid SMILES strings. Specifically, an embodiment of the present invention can predict very similar reactant molecules with an accuracy of 57.7%. In addition, the present invention can provide more robust prediction than existing methods. A retrosynthetic prediction comprises a step of training a neural machine translation (NMT) model by using a final product-reactant pair.

Description

Substructure-based neural network machine translation apparatus for prediction of inverse synthesis and a translation method using the same

Reverse synthesis refers to a process of tracing a target compound having a specific chemical structure from which starting material through which route in a synthesis reaction targeting an organic compound. To this end, the starting material is reached by identifying a precursor that can arrive at the chemical structure of the target compound in the opposite direction from the actual reaction path, and considering other precursors that arrive at the precursor again. In particular, in drug design, after virtual screening of a ligand candidate group capable of binding to a target protein, for example, based on binding affinity, it is an important step in finding a substance that can actually be synthesized or commercially available from among these candidate groups.

Although knowledge of organic chemistry has been accumulated through many years of research, designing efficient synthetic pathways for target molecules still remains an important task in organic synthesis. The retrosynthetic approach provides a logical synthetic route for generating target molecules from a set of available reactants and reagents. Since inverse synthesis transforms require sequential operations, this approach is naturally iterative and recursive. Reverse synthetic transformations are simpler, and occur recursively until commercially available molecules are identified.

Computational retrosynthesis analysis was first formalized as an algorithmic method by Corey and Wipke in 1969 (Corey, EJ; Todd Wipke, W. Computer-assisted design of complex organic syntheses. Science 1969, 166, 178-192). . The algorithm considers all possible fragments into a known chemical reaction, reduces the complexity of the product and proceeds with processing until a chemically rational pathway is specified. These segments are hand-made minimal transformation rules, widely known as reaction templates. Manually encoding these transformation rules requires in-depth chemistry expertise and intuition. Considering the large size (>10000) of transformation rules that must be manually coded, manual management of synthetic knowledge is a very complex task. Furthermore, dependence on response templates potentially limits prediction accuracy. In particular, if the reaction takes place outside the template region. Subsequent work allowed chemists to find better routes faster by enabling automatic extraction of reaction templates. However, this did not address the aforementioned limitations resulting from previous studies. A computer-aided synthesis scheme has recently been documented in a number of studies.

See Kayala, M. A.; Azencott, C.-A.; Chen, J. H.; Baldi, P. Learning to Predict Chemical Reactions. J. Chem. Inf. Model. 2011, 51, 2209-2222 and Kayala, M. A.; Baldi, P. ReactionPredictor: Prediction of Complex Chemical Reactions at the Mechanistic Level Using Machine Learning. J. Chem. Inf. Model. 2012, 52, 2526-2540], the reaction predictor is the first template-free approach. This was a strategy at the mechanistic level that applied the ideas of rule-based modeling and machine learning to the framework of synthetic planning. See Jin, W.; Coley, C. W.; Barzilay, R.; Jaakkola, T. Predicting organic reaction outcomes with weisfeiler-lehman network. Adv. Neur. In. 2017, 2017-Decem, 2608-2617] proposed a new template-free method based on complete data based on the Weisfeiler-Lehman network. Both approaches provide an end-to-end solution for generating candidate products. Cadeddu, A.; Wylie, E. K.; Jurczak, J.; Wampler-Doty, M.; Grzybowski, B. A. Organic chemistry as a language and the implications of chemical linguistics for structural and retrosynthetic analyses. Angew. Chem. Int. Edit. 2014, 53, 8108-8112], other template-free for the task of predicting forward or reverse response using various types of neural machine translation (NMT) architectural approaches. further development of methods. Based on the explicit analogy between sentences in the linguistic corpus and molecules in the chemical corpus, for example, in chemical space, Cadeddu et al. showed that the rank-frequency distribution of a substructure as a block constituting a molecule is similar to that of a word in a language corpus. This validation means that the concept of linguistic analysis can be easily applied to find solutions to the problems of forward and backward prediction. From this point of view, reverse synthesis prediction is suitable for applying the sequence-to-sequence framework of machine translation.

Sequence-to-sequence learning uses a recurrent neural network (RNN) layer to map source sequences of arbitrary length into fixed-dimensional context vectors made up of real numbers. The context vector contains information about the syntax and semantic structure of the source sequence. Connected to this RNN layer, another RNN decodes the context vector into the target sequence. In this respect, the two RNN units act together as a pair of encoder-decoder systems. See Sutskever, I.; Vinyals, O.; Le, Q. V. Sequence to sequence learning with neural networks. Adv. Neur. In. 2014, 4, 3104-3112] showed that long short-term memory (LSTM)-based architectures can solve common sequence-to-sequence problems through their ability to handle long-range relationships of sequences. Liu, B.; Ramsundar, B.; Kawthekar, P.; Shi, J.; Gomes, J.; Luu Nguyen, Q.; Ho, S.; Sloane, J.; Wender, P.; Pande, V. Retrosynthetic Reaction Prediction Using Neural Sequence-to-Sequence Models. ACS Cent. Sci. 2017, 3, 1103-1113] proposed the first multilayer LSTM-based sequence-to-sequence model for reverse synthesis prediction. Its gated recurrent unit (GRU) variant was proposed by Nam and Kim for forward reaction prediction (Nam, J.; Kim, J. Linking the Neural Machine Translation and the Prediction of Organic Chemistry Reactions. 2016, 1- 19]).

Recently, the best neural machine translation (NMT) has enhanced the performance for long sentences by including an attention mechanism as a configuration of the neural network architecture. There is also the inverse synthesis prediction implemented in the transformation architecture based only on the attention mechanism. Encoder-decoder models introduced similar strategies in addressing the nature of the translation task, especially after the attention mechanism was introduced. A simplified molecular input line entry system (SMILES) representation of a molecular structure is a typical input for sequence-to-sequence based models. However, previous studies have not shown models that focus on translation at the substructure, fragment, or level.

Embodiments of the present invention provide a template-free approach for predicting a retrosynthetic reaction by learning chemical changes at the substructure level. Embodiments use a molecular access system (MACS) key to represent a molecule in a sentence based on a set of substructures corresponding to words. In addition, embodiments provide a unique tokenization structure that adequately eliminates problems arising from SMILES-based tokenization. Embodiments consist of bidirectional LSTM cells, are fully data-based without preceding response classification information, and are learned in an end-to-end manner. The examples thoroughly discuss all aspects of the methodology, including the dataset and descriptor curation steps. We present the evaluation results based on three datasets acquired from the USPTO response dataset.

The embodiment also provides a novel method of tokenization prior to curation with analysis of datasets and descriptors. In the embodiment of the present invention, the model architecture and evaluation procedure for accuracy calculation are briefly described, the results of the translation experiment set are described with emphasis on the advantages of MACCS key-based molecular expression, and the advantages and challenges of the present embodiment are described. .

Various embodiments of the present invention are intended to provide a new type of reverse synthesis method. Specifically, it is intended to provide a template-free reverse synthesis method that does not require a template by applying machine translation to the reverse synthesis method.

In addition, the embodiments are intended to provide data preprocessing methods for expressing molecules as text for a reverse synthesis method through neural network machine translation.

In addition, we intend to provide a data preprocessing method for machine translation at the molecular substructure level.

In addition, embodiments provide a unique tokenization structure that adequately eliminates problems arising from SMILES-based tokenization.

An embodiment of the present invention provides a method for predicting reverse synthesis using neural network machine translation that can be implemented in a reverse synthesis system including a communication unit, a control unit, and a memory unit, generating an initial product-reactant pair, the communication unit Receiving a reaction data set for the product-reactant from an external server through the; expressing the received response data set as a descriptor composed of characters; and storing the initial product-reactant pair represented by the descriptor in the memory unit, generating the initial product-reactant pair; filtering by applying one or more filters to the initial product-reactant pair; One-to-one mapping of the filtered product-reactant pair; aligning the product-reactant pair to which the one-to-one mapping is applied to generate a final product-reactant pair; training a neural machine translation (NMT) model using the final product-reactant pair; evaluating the learned NMT model; and when a new product is input to the NMT model, predicting and calculating a candidate reactant corresponding to the new product.

In addition, the step of expressing the received response data set as a descriptor composed of characters may include: preparing a molecular access system (MACS) key representing a molecule; generating a character descriptor for the MACCS key by allocating a predetermined character to each of the MACCS keys; expressing the received response data set with the MACCS key; and converting the response dataset expressed by the MACCS key into the character descriptor.

In addition, the preparing of the MACCS key may include generating a curated MACCS key by removing MACCS keys having a frequency of occurrence less than or equal to a predetermined among the MACCS keys.

In addition, the generating of the character descriptor for the MACCSS key provides a desynthesis prediction method in which the predetermined character is assigned to each of the curated MACCS keys.

In addition, the filtering by applying one or more filters to the initial product-reactant pair includes a filter that removes a pair having three or more reactants, a filter that removes the same product-reactant pair, a filter that removes internal twins, It may include one or more filters among filters that remove pairs having a sequence length of 100 or more.

In addition, in the one-to-one mapping of the filtered product-reactant pair, a one-to-many mapping for the filtered product-reactant pair is checked, and a molecule having the shortest sequence length among the one-to-many mappings A method for predicting inverse synthesis is provided, which reduces to a one-to-one mapping by selecting .

In addition, the step of aligning the product-reactant pair to which the one-to-one mapping is applied to generate the final product-reactant pair includes, when two or more reactants of the product-reactant pair are arranged in descending order according to the length, A method for predicting reverse synthesis is provided, which connects the reactants with symbols.

In addition, the neural network machine translation model provides a sequence-to-sequence model, a reverse synthesis prediction method.

In addition, the neural network machine translation model may include two bidirectional long short-term memories (LSTMs), and one of the two bidirectional LSTMs may function as an encoder and the other as a decoder.

In addition, the encoder and the decoder may be connected through an attention mechanism.

In addition, the bidirectional LSTM provides a method for predicting inverse synthesis, each including a hidden layer, and further including a dropout layer after the hidden layer.

In addition, the neural network machine translation model may further include gradient clipping so that the norm of the gradient does not exceed a threshold value during backpropagation.

In addition, the learning rate of the neural network machine translation model may be reduced by a constant coefficient every three epochs.

In addition, the evaluating the learned NMT model comprises calculating the similarity between the candidate reactant predicted through the NMT model and the actual reactant obtained through the final product-reactant pair to evaluate the prediction accuracy, reverse synthesis prediction method provides

Also, the predicting and calculating of the candidate reactant may further include predicting and calculating a secondary candidate reactant for synthesizing the candidate reactant.

Embodiments of the present invention provide a template-free desynthesis method, which is learned in an end-to-end manner, based on complete data.

In addition, embodiments of the present invention can verify the similarity between machine translation of language and molecular translation based on substructure, and provide a reverse synthesis method using machine translation based on this.

1 is a diagram schematically illustrating an embodiment of a reverse synthesis system.
2 is a diagram illustrating a frequency distribution of MACCS keys according to an embodiment.
3 is a diagram illustrating an example of a character assigned to a MACCS key.
4 is a diagram exemplarily illustrating a distribution of pair lengths.
5 is a diagram illustrating a process for generating a product-reactant pair according to an embodiment.
6 is a diagram illustrating a relationship between a character key and other keys according to an embodiment.
7 is a diagram illustrating a learning progress of an inverse synthesis prediction model according to an embodiment.
8 shows randomly selected predictions according to an embodiment.
9 shows selected predictions according to another embodiment.
10 shows selected predictions according to another embodiment.
11 shows an example of candidates for the first reactant of reaction 4 of FIG. 8 .

The present invention will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings herein. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, and only these embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention belongs It is provided to fully inform the possessor of the scope of the invention, and the present invention is only defined by the scope of the claims. On the other hand, the terms used herein are for the purpose of describing the embodiments and are not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, "comprises" or "comprising" does not exclude the presence or addition of one or more other elements, steps, in addition to the recited elements, steps. Terms such as first, second, etc. may be used to describe various elements, but the elements should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another.

Hereinafter, various embodiments of the present invention will be exemplarily described with reference to the drawings.

Fig. 1 is a block diagram showing an example of the configuration of a reverse synthesis system according to the embodiment, and conceptually shows parts related to the present embodiment. Each configuration may be provided in one device and may be processed independently, but the present invention is not limited thereto, and may also include that each configuration is performed in a separate device by being connected through a network.

The external server 20 may be connected to each other with the reverse synthesis system 10 through a network, and protein structure information, ligand structure information, protein-ligand complex structure information, protein-ligand complex binding site binding affinity information, Information such as genetic information, intermolecular interaction information, and/or protein structure similarity information may be provided. The external server 20 may also provide datasets of molecules denoted by SMILES notation or MACCS key. Alternatively, for example, the external server 20 may be a molecular database provided by the USPTO or a preprocessed dataset. The external server 20 may be, for example, a database for the prediction processing of the inverse synthesis system 10 or a server providing the same.

The inverse synthesis system 10 may include a control unit 11 , a communication unit 12 , an input/output interface unit 13 , and a memory unit 14 .

The control unit 11 is a component that controls the entire inverse synthesis system 10 , and may include, for example, a processing unit such as a CPU or a GPU. The controller 11 may learn models to be described later by using the information stored in the memory unit 14 , and may also calculate a predicted value for a new input through the learned model. Specifically, the controller 11 may control a neural machine translation (NMT) model. To this end, the control unit 11 may include an internal memory for storing a control program such as an operating system (OS), a program defining various processing sequences, and the like, and data. Then, the control unit 11 can perform information processing for executing various processing by these programs or the like.

In addition, the communication unit 12 may include an interface that can be connected to a communication device such as a router connected to a communication line or the like, and can control communication between the reverse synthesis system 10 and the external server 20 . there is.

The input/output interface unit 13 may be an interface connected to the input unit 15 and the display unit 16 . A user may communicate with the prediction system 10 through the input/output interface unit 13 . For example, the display unit 16 may be a display means for displaying a display screen such as an application (eg, a display composed of liquid crystal or organic EL or the like, a monitor, a touch panel, etc.). Also, the input unit 15 may be, for example, a key input unit, a touch panel, a control pad (eg, a touch pad, a game pad, etc.), a mouse, a keyboard, and a microphone.

Also, the memory unit 14 may be a device for storing various databases or tables. For example, the memory unit is a filtered US patent reaction dataset, a dataset of molecules marked with SMILES notation or MACCS key, descriptor curation information representing the molecule, various filter units, protein structure information, ligand structure information, protein-ligand complex It may include information such as structural information, binding affinity information according to the protein-ligand complex binding site, genetic information, intermolecular interaction information, and/or protein structure similarity information.

dataset

In the examples, a filtered US patent response dataset obtained with a text mining approach was used. Schwaller, P.; Gaudin, T.; Lanyi, D.; Bekas, C.; Laino, T. "Found in Translation": predicting outcomes of complex organic chemistry reactions using neural sequence-to-sequence models. Chem. Sci. 2018, 9, 6091-6098] removed duplicate response strings from the dataset without atom-mapping. They also removed 780 responses that were not standardized with RDKit. An essential limitation of the data is that most of the entries are single product reactions. Therefore, the examples were considered only for a single product corresponding to 92% of the dataset.

SMILES line notation represents a molecular structure as a linear sequence of letters, numbers, and symbols. Therefore, from a linguistic point of view, SMILES can be regarded as a language with grammatical specifications. However, in the embodiment, a molecule is expressed as a set of fragments using MACCS keys composed of a predetermined 166 substructures. This binary bit-based molecular descriptor transforms the molecule into a 166-bit vector length, each bit associated with a feature extracted from a predetermined dictionary of SMARTS (SMILES arbitrary target specification) patterns.

Descriptor curation

In an embodiment, a molecule may be represented as a set of fragments using a MACCS key. The example looked at the number of occurrences of each MACCS key in the example's dataset. In addition, the results obtained were compared with one million randomly selected drug-like small molecules, a subset of the Generated Data Base-13 (GDB-13) consisting of 975 million molecules. Figure 2 shows the normalized frequency distribution of MACCS keys in two databases. Direct pair comparison in FIG. 2 may reduce the number of MACCS keys. In an embodiment, 5 keys that never occurred and 9 keys that did not appear frequently in the USPTO database were omitted. According to the comparative analysis, 26 keys that were not identified or rarely identified in the GDB-13 database were additionally excluded.

Since the embodiment is to develop a model for the analysis of drug-like molecules, not all MACCS keys need be used. Removal of unnecessary keys based on occurrence analysis has obvious advantages. This shortens the length of the source and target sentences and further improves the rank distribution of keys used in translation work. In an embodiment, all molecules are represented by 126 MACCS keys, which can adequately represent 98% of the randomly sampled subset of GDB-13. In machine translation tasks handled by chemists, source and target molecules are placeholders that correspond interchangeably to reactants and products. The choice depends on the desired analysis. For the reverse synthesis prediction task, the source and target sentences represent products and reactants, respectively.

reaction preprocessing

The embodiment considered only non-zero indicators of the curated MACCS key. To form a unique artificial “word”, English characters were assigned to non-zero-ranked MACCS keys based on the rank frequency distribution of MACCS keys. In other words, the product and reactant sentences of the examples consist of only 126 words, ie, each sentence is an ordered list of independent pieces. Single-letter words were generated using upper and lower cases of the most frequent letters top-21 of English. The double-letter words are formed by concatenating “x” and “z” for every 42 single letters, ensuring coverage for all 126 MACCS keys. That is, the literal fragment vocabulary of the embodiment has a fixed length of 126. 3 is an exemplary diagram illustrating characters assigned to MACCS keys. For example, it can be seen that 'mz' is assigned to the MACCS key 58 and "G" is assigned to the MACCS key 110. The allocation table of FIG. 3 is an example for explaining an embodiment, and the present invention is not limited thereto, and it is obvious that modifications, additions, deletions, etc. that can be easily derived by those skilled in the art are also included in the technical spirit of the present invention. In addition, although the embodiment of the present invention displays the numerator by allocating characters based on the MACCS key, this is one of the embodiments and it is also possible to express the numerator by allocating characters for other expression methods.

Table 1 below shows an exemplary process for generating product-reactant pairs. The same procedure was followed for all reactions in the dataset. Table 1 shows the data preparation steps to obtain product and reactant sentences for the reverse synthesis prediction task.

Step Example (entry # 20032, Lowe's USPTO data set)

reaction

REACTION SMILES
(standardization)

Reactants > reagents · solvents > products
C=CC.CC(=O)O>CC(=O)OC(C)=O.O=O>CC(=O)OCC(C)OC(C)=O

Reactants > reagents · solvents > products
C=CC.CC(=O)O>CC(=O)OC(C)=OO=O>CC(=O)OCC(C)OC(C)=O Removal of reagents and solvents [['C=CC','CC(=O)O'],'CC(=O)OCC(C)OC(C)=O'] MACCS domain
(Nonzero keys) [[34, 99, 160],[123, 139, 154, 157, 159, 160, 164]],
[72, 108, 109, 115, 116, 123, 126, 132, 136, 140, 141, 146, 149, 152, 153, 154, 155, 157, 159, 160, 164] character assignment
(Frequency basis) [['Uz','dz','s'],['Y','Ex','c','r','h','s','t']],
['iz','Fx','fx','ez','Gx','Y','E','hx','ux','Hx','Ix','O','H','v','u','c','w','r','h','s','t'] product iz Fx fx ez Gx Y E hx ux Hx Ix O H v u c w r h s t reactant(s) Y Ex c r h s t - Uz dz s

reaction data set curation

The product-reactant pair dataset undergoes additional curation before being processed by the translation model of the example. After representing all molecules with 126 edited MACCS keys, several filters are applied to remove identical product-reactant pairs and inner twins. An inner twin refers to a data entry pair in which the product and reactant statements are identical. These occur when the chemical change exceeds the sensitivity of the MACCS key-based representation of the embodiment. Some information is lost because the embodiment has linked molecules to MACCS keys to operate in the subspace of the substructure. The example pretreatment procedure resulted in 5748 internal tweens, which were removed from the dataset. Additionally, reactions with three or more reactants were excluded. The length of the longest pair was set to 100 as shown in FIG. 4 to avoid too long fragment sequences.

The product-reactant pairs are then fed to an injective map generator to ensure a one-to-one correspondence between product and reactant sentences. If the reactant sentence consists of two reactants, the embodiment sorts them in descending order according to their sequence length. Reactants are separated by a “-” sign. The curated dataset contains a total of 352,546 product-reactant pairs, further divided into two separate subsets according to the number of reactant molecules in each pair. That is, it is subdivided into datasets of single reactant and double reactant. Organizing the dataset in this way is important for independently evaluating the performance of the model. Figure 5 illustrates the steps of curation of a dataset along with the size of the dataset.

5 is a diagram schematically illustrating a process of deriving a final product-reactant pair according to the above description. As shown in Figure 5, the dataset curation process for obtaining product-reactant pairs for training and testing is shown.

For example, in the reaction preprocessing step, 1,002,970 molecules input from the USPTO dataset (N=1,002,970) are expressed as 126 MACCS keys. This may create an initial product-reactant pair. In the reaction dataset curation stage, it goes through a number of filters. The example applied, for example, 4 filters. Cases of 3 or more reactants were removed (N = 922,823), overlapping pairs were removed (N = 786,219), inner twins were removed (N = 780,471), and cases with a pair length greater than 100 were removed (N = 429,889). In addition, filters other than the disclosed filters may be applied. The filtering of this embodiment is only some examples of possible filtering, but is not limited thereto. Additional filtering may be further applied, or only a portion of the disclosed filtering may be applied.

The filtered product-reactant pairs are fed to a one-to-one map generator (N=352,546), which is sorted to produce the final product-reactant pairs. The resulting product-reactant pairs are fed into the model to be trained and tested.

model architecture

The sequence-to-sequence neural network of the embodiment includes two bidirectional LSTMs. One for the encoder and the other for the decoder. In addition, a unidirectional LSTM was also used to quantify the degree of improvement in model performance using the bidirectional LSTM. The embodiment connected the encoder and the decoder through Luong's global attention mechanism to ensure non-local relationships between all configurations of source sequences (Luong, MT; Pham, H.; Manning, CD Effective approaches to Attention-based neural machine translation. Conf. Proc. - EMNLP 2015 Conf. Empir. Methods Nat. Lang. Process. 2015, 1412-1421]). The attention mechanism allows the neural network to focus on different parts of the source sentence, and can take into account non-linear relationships between words during the learning process. The global attention mechanism applied in the embodiment is essentially the proposed method for machine translation tasks [Bahdanau, D.; Cho, K. H.; Bengio, Y. Neural machine translation by jointly learning to align and translate. 3rd Int. Conf. Learn. Represent. ICLR 2015 - Conf. Track Proc.2015, 1-15] is similar to the first attention mechanism. The global approach concentrates on “attention” on all words in the source sentence to compute a global context vector for each target word at each time step of the decoder unit. Therefore, the global context vector represents the weighted sum of all source hidden states. This contextual information can lead to improved prediction accuracy.

Learning Details

The curated datasets of the examples are randomly divided in a ratio of 9:1 for training and testing, respectively. The validation set is randomly sampled from the training set (10%). Word embeddings are used to represent literal fragments within a vocabulary. After the embedding layer is created, a learnable tensor containing a 126-dimensional fixed-length dense vector is randomly initialized. The embedding class method then evaluates the embedding of each word through a lookup on the tensor. The embodiment trained all parameters of the encoder-decoder model using a stochastic gradient descent algorithm. The cross entropy function was used as the loss function.

For each dataset, the Example ran a series of tests within the bounds of the hyperparameter space as described in Table 2 to obtain optimal performance. Table 2 shows the hyperparameter space and hyperparameters for the optimal model.

parameter
(Parameter) possible values
(Possible Values) Optimal model parameters
(Best Model Parameters) RNN Cell Type LSTM or Bi-LSTM Bi-LSTM (encoder and decoder) number of layers 2, 4, or 6 2 number of units 500, 1000, 2000 2000 running rate 0.1 - 8 4 Decay factor 0.50 - 0.90 0.85 Dropout 0.1 - 0.5 0.1 Attention type Luong's global attention mechanism

Based on preliminary experiments, the embodiment creates an encoder and a decoder with two bidirectional LSTM layers, each of which has 2000 hidden units. A dropout layer with a dropout rate of 0.1 was included following the hidden layer to avoid being vulnerable to overfitting. To avoid the potential exploding gradient problem, the embodiment introduces gradient clipping to ensure that the norm of the gradient does not exceed a threshold value (0.25) during backpropagation. The initial learning rate was set to 4.0 and decreased by a factor of 0.85 every three epochs.

With these hyper-parameter settings, a batch size of 64 on a single NVIDIA RTX 2080Ti GPU card resulted in an average learning rate of about 3300 words per second. Large batch sizes due to memory limitations were not tested, and the same applies to the size of the hidden layer. The example trained the model for a minimum of 30 epochs, each epoch taking about 2 hours on a curated dataset of 320,000 sentence pairs.

The embodiment was implemented in PyTorch version 1.3.0 and Python version 3.6.8. An open source RDKit module version 2020.03.1 was used to obtain MACCS keys and similarity maps.

evaluation process

In order to evaluate the performance of the inverse synthesis model of the example, the Tanimoto coefficient, which is known as one of the most excellent metrics for calculating structural similarity, was selected as a similarity metric (Bajusz, D.; Racz, A.; Heberger, K. Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations? J. Cheminformatics 2015, 7, 1-13]). Pairwise similarity between the predicted sequence and the measured data of all test molecules was calculated. The Tanimoto coefficient ( T _c ) is measured between two chemical structures with values between 0 and 1. If molecules do not share a common fragment, the coefficient is 0, and identical molecules have a Tanimoto coefficient of 1. These are cases at both ends of the Tanimoto similarity metric, but there is no single criterion that defines similar and dissimilar molecules. The Example defined three thresholds (0.50, 0.70 and 0.85) to evaluate the quality of the translation experiment. The similarity between the predicted sequence and the ground truth is computed at the end of each epoch for all pairs appearing in the validation set using the Tanimoto similarity measure (Equation 1).

[Formula 1]

All reactions are included in the combined dataset, so the example yields predictions with one or two reactants. Therefore, comparing the predicted sequence with the ground truth has multiple possibilities. Potential pairs for evaluation corresponding to the number of reactants are listed in Table 3. Tanimoto similarity between predicted sequences and all possible pairs of ground truth data was calculated. Thereafter, the pair(s) having the highest similarity was selected based on the assumption that the more similar the structures are, the higher the matching probability is. In Table 3, possible pairs between the prediction sequence and the ground truth are presented. The similarity of each pair was calculated by Equation 1.

actual data prediction list of possible pairs P -> R _A + R _B P -> P _A + P _B
P -> P _C [( R _A P _A - R _B P _B ), ( R _A P _B - R _B P _A )]
[ R _A P _C , R _B P _C ] P -> R _C P -> P _A + P _B
P -> P _C [ R _C P _A , R _C P _B ]
[ R _C P _C ]

Results and discussion

Prediction Accuracy

The performance of the examples was evaluated based on three datasets. single reactant, double reactant, and combined test set. Table 4 shows the evaluation results for the test set. The prediction quality of each test dataset is expressed as the pairwise Tanimoto similarity. The Examples introduced three criteria to evaluate the success rate of the translation model: 1) number of exact matches ( T _c = 1.0), 2) bioactive quasi-matches (0.85 < T _c < 1.00), 3) Overall success rate expressed as mean Tanimoto similarity between prediction and ground truth sequences (series of fragments) for all test molecules. Table 4 shows the success rates of single reactant, double reactant, and combined test sets based on the Tanimoto similarity metric.

dataset single double Sum size learning pair 88,151 229,141 317,292 test pair 9,794 25,460 35,254 test molecules 9,794 50,911 55,958 average pair length 74 73 74 success rate Bi-LSTM T _c = 1.0 29.0% 27.9% 25.3% 0.85 < T _c < 1.00 ^a 28.7% 10.5% 12.9%

^b

^b 0.84 0.66 0.68 LSTM T _c = 1.0 22.9% 21.6% 19.4% 0.85 < T _c < 1.00 ^a 29.7% 10.2% 12.5%

^b 0.82 0.62 0.64

^a bioactive similar molecule ^b average similarity

For single reactant reactions, the bidirectional LSTM model of the Examples achieved an accuracy of 57.7% by combining the first two criteria. Accurate prediction and physiologically similar matching were 29.0% and 28.7%, respectively. The mean T _c between the predicted values and the ground truth sequence is 0.84. These results show that the model of the Example predicts with high accuracy the single reactant response.

For double reactant reactions, the success rate of accurate predictions (27.9%) is approximately the same as for single reactant reactions. However, the success rate of high similarity prediction deteriorated from 28.5% to 10.5%. For the combined set, 25.3% of the predictions were correct, and 12.9% of the predictions were high similarity predictions. Similarly, the mean T _c value dropped from 0.84 to 0.66, and the dataset combined with the double and bound reactants showed 0.68.

One of the reasons for the low accuracy in the double and combined sets seems to be that the "-" sign was not predicted properly. Another cause seems to be that the frequency occurrence of small molecules is represented by a small number of MACCS keys in the dataset. Indeed, 477 molecules in 61822 different reactions are represented by less than 7 MACCS keys. More specifically, 3944 responses include reactants represented by one of the seven MACCS keys shown in FIG. 6 . However, the number of unique structures corresponding to the above-described keys is only 29. Because these small and simple structures are densely populated in the dataset, erroneously predicted fragments have a significant impact on the success rate (zero values in the 1-bit case).

The results of the examples also show that the bidirectional LSTM based model outperforms the unidirectional LSTM based model. The success rate of exact matching is consistently lower by about 6% for the entire dataset. This can be seen because the molecular expression based on the MACCS key of the embodiment does not depend on the order of the keys. In other words, most of the information about molecules and chemical reactions is embedded in the co-occurrence of keys.

Since response classification information was not previously provided to the model of Examples, the prediction accuracy of Examples was compared with other retrosynthetic prediction methods that did not consider response class labels. Several recent reports have summarized the prediction accuracy of various models. Lin, K.; Xu, Y.; Pei, J.; Lai, L. Automatic retrosynthetic route planning using template free models. Chem. Sci. 2020, 11, 3355-3364] showed that the top-1 accuracy was 28.3% (Liu, B.; Ramsundar, B.; Kawthekar, P.; Shi, J.; Gomes, J.; Luu Nguyen). , Q.; Ho, S.; Sloane, J.; Wender, P.; Pande, V. Retrosynthetic Reaction Prediction Using Neural Sequence-to-Sequence Models. ACS Cent. Sci. 2017, 3, 1103-1113], USPTO of 54.1% (Lin, K.; Xu, Y.; Pei, J.; Lai, L. Automatic retrosynthetic route planning using template free models. Chem. Sci. 2020, 11, 3355-3364] of the Transformer model for the USPTO MIT dataset). As an alternative approach, see Coley, C. W.; Rogers, L.; Green, W. H.; Jensen, K. F. Computer-Assisted Retrosynthesis Based on Molecular Similarity. ACS Cent. Sci. 2017, 3, 1237-1245] achieved a top-1 accuracy of 37.3% for the USPTO's 50,000 dataset.

To confirm how the example model learns the grammar of chemical reactions, FIG. 7 shows the evolution of prediction accuracy for thresholds according to learning epochs for a single reactant validation set. Specifically, Fig. 7 shows that the network has successfully learned the response rules by capturing molecular modifications at the substructure level. The number of correct predictions ( T _c = 1.0) rose sharply during the first 10 epochs. After 20 epochs, the value has almost tripled. The probability of making better predictions for each piece is higher during learning. This is a clear indicator of successful learning. The improvement in accurate predictions appears to be the result of each decrease in non-exact matches, except for very bad predictions ( T _c < 0.50). The quality of the bad predictions (5% of the validation set) did not improve, probably due to the lack of information, complexity, and noise contained in the data. These results were similarly repeated for all other datasets.

In the examples, it is assumed that a candidate reactant with T _c > 0.85 is sufficiently similar to its actual counterpart. To validate this assumption, the Examples evaluated the quality of the candidate reactants by comparing them to their actual reactants. The Examples investigated whether the following factors are correct: functional group interconversion or bond severing, reactive functional groups, core structure, and substituents. The accuracy of the lateral substituents was considered less important in matching the functionality of the reactants, especially when it was a simple alkyl.

8 shows randomly selected predictions to account for possible prediction cases. A similarity map was provided to visualize the similarity between the candidates and the actual reactant. Specifically, FIG. 8 shows that non-exact candidates differ according to their degree of similarity. 8 shows the similarity score calculation and the similarity map using the Morgan fingerprint and Tanimoto metrics. The color represents the atomic level contribution to the overall similarity. Specifically, green indicates an increase in the similarity score, red indicates a decrease in the similarity score, and colorless indicates no effect.

Reaction 1 derived a reaction product in which a main chain composed of 8 carbons and α, β-unsaturated aldehyde groups were precisely induced at the correct site. Although an ester rather than an aldehyde was expected, aldehyde reduction can also provide the same target alcohol. This indicates that the predictions of the Examples correctly confirmed the functional group interconversion. On the other hand, one olefin was missing and the site and number of methyl groups of two out of four were misinterpreted.

In Reaction 2, except for the position of the ester group, the core heterocyclic ring, pyridine and thiazole, and their linkages were created correctly. In the actual reactant, the methyl ester group is attached to C6 of the pyridine, whereas the ethylester group is attached to C4 of the thiazole ring in the candidate. If the position of the ethyl ester group is correct, a single step of reduction will be required to obtain the alcohol group.

In Reaction 3, the core structure of pyrazole and its methyl ester group were accurately predicted. However, chloride, one of the reactive functional groups, and a substituent on the pyrazole ring are missing, and the structure of the thiol is misinterpreted.

The predicted results of Reaction 4 show that the Example accurately predicted the core structure, bond cleavage, and reactive functional groups. However, the number and sites of halides were wrong.

In reaction 5, one reactant was predicted correctly but the other reactant was partially predicted incorrectly. In the erroneously predicted candidate, the (2-naphthyl)vinyl group was erroneously predicted as a (phenyl)methyl group, but the reactive functional group, acylhydrazine, was correctly predicted.

The results of reaction 6 show an exact match for N-hydroxyphthalimide as a precursor of O-hydroxylamine. However, the structure of the alkyl halide lacks a phenylene group. Estimates of the core structure have largely failed in this reaction. On the other hand, the reactive functional group and bond severing have been precisely proposed.

A quantitative summary of the evaluations described above is presented in Table 5. Three criteria: functional group interconversion or dissociation, core structure, and reactive functional groups were weighted equally. Together with the similarity score, they are utilized to form a chemically reasonable score. With regard to the core structure, the longest carbon chain and/or ring capable of bearing heteroatoms such as O, N, S is contemplated. Since functional group interconversion or dissociation is the most important part of reverse synthesis analysis, the exact position of the reaction site is strictly scored with true/false values corresponding to 1 and 0, respectively. Related functional groups receive the same weight. The examples scored each candidate reactant individually and averaged the results to derive a final score for each criterion.

In Table 5, the FGI or bond dissociation and reactive functional groups columns show accuracy in a true (1)/false (0) manner. The core structure column indicates the average accuracy of the core structure of candidate molecules by securing the accuracy of the core structure itself as well as the type and position of the side substituents. The source of error is given in parentheses. C1 points to candidate 1, and C2 points to candidate 2. For example, “C2=0.33, 2/3 fragments” means that candidate reactant 2 has an accuracy of 0.33 because 2 out of 3 fragments were predicted incorrectly. The mean of the three criteria is shown in separate columns. The Tc column represents the average Tc value of the candidate reactants.

Reaction No. FGI or
bond dissociation core structure reactive functional groups Average T _c One 1.00 0.33 (2/3 pieces) 1.00 0.78 0.64 2 1.00 0.67 (1/3 piece area) 1.00 0.89 0.85 3 1.00 0.69 (C1=0.88, lateral substructure of piece; C2=0.5, 1/2 piece) 0.50 (1 for thiol, 0 for chloride) 0.73 0.57 4 1.00 0.96 (C1=0.92, region of the lateral substructure; C2=1.0, excluding Cl) 1.00 0.99 0.86 5 1.00 0.83 (C1=0.67, 1/3 piece; C2 is correct) 1.00 0.94 0.84 6 1.00 0.67 (C1 is correct; C2=0.33, 2/3 pieces) 1.00 0.89 0.73

It is noteworthy that the examples accurately predicted functional group interconversion or dissociation for all six reactions. Except for reaction 3, the reactive functional groups were accurately reflected. It can be seen that the prediction error affecting the score is mainly related to the core structure. The Examples applied this structural data-based scoring strategy to a more specific set comprising 10 randomly selected responses in which the candidate reactants, on average, were in similar bioactive domains (T _c = 0.87). 9, 10, and Table 6 specifically disclose this. The results clearly show that the Example predicts not only the reactive functional group but also the functional group interconversion or dissociation very accurately for the reactant candidates that are physiologically and chemically similar.

Chemical examination of the reactions indicated a close correlation with the mean similarity score and the score generated based on the structural data. The scoring approach of the Examples provides a clear idea that the quality of candidate reactants and similarity scores is consistent with those obtained passively. The similarity measure scores lower than the structural data-based score may be due to the inclusion of side chains and geometric elements. Although the interpretation of the similarity score is somewhat difficult for objective evaluation, it can be used to evaluate the predictive quality of retrosynthesis. A high similarity score indicates that the desired molecule is more synthetically accessible according to the rules of organic chemistry.

In Table 6, the FGI or bond dissociation and reactive functional groups columns show accuracy in a true (1)/false (0) manner. The core structure column indicates the average accuracy of the core structure of candidate molecules by securing the accuracy of the core structure itself as well as the type and position of the side substituents. The source of error is given in parentheses. C1 points to candidate 1, and C2 points to candidate 2. For example, “C2=0.33, 2/3 fragments” means that candidate reactant 2 has an accuracy of 0.33 because 2 out of 3 fragments were predicted incorrectly. The mean of the three criteria is shown in separate columns. The Tc column represents the average Tc value of the candidate reactants.

Reaction No. FGI or
bond dissociation core structure reactive functional groups Average T _c One 1.00 0.98 (C1=1.00; C2=0.95, alkyl #C 6/5) 1.00 0.99 0.87 2 1.00 0.83 (C1=1.00; C2=0.67, 1/3 piece) 1.00 0.94 0.81 3 1.00 1.00 1.00 1.00 0.84 4 1.00 0.75 (C1=1.00; C2=0.50, 1/2 piece) 1.00 0.92 0.87 5 1.00 0.79 (C1=0.75, fragment area 1/2; C2=0.83, fragment area 1/3) 1.00 0.93 0.86 6 1.00 0.88 (C1=0.75, slice area 1/2, C2=1.00) 1.00 0.96 0.94 7 1.00 0.96 (C1=0.97, ring #C 5/6; C2=0.94, alkyl #C 5/4) 1.00 0.99 0.91 8 1.00 1.00 1.00 1.00 0.87 9 1.00 1.00 1.00 1.00 0.83 10 1.00 0.97 (C1=1.00; C2=0.94, region of the lateral substructure) 1.00 0.99 0.85

Advantages of embodiments

The main advantage of the word-based MACCS key method of the embodiment over the character-based SMILES method is that it requires the network to learn relatively simple grammatical rules (ascending and co-occurrence of keys) to produce meaningful results. In the SMILES-based method, the network must understand not only the complex grammar of SMILES but also the standardized representation to predict synthetically correct sequences. Liu, B.; Ramsundar, B.; Kawthekar, P.; Shi, J.; Gomes, J.; Luu Nguyen, Q.; Ho, S.; Sloane, J.; Wender, P.; Pande, V. Retrosynthetic Reaction Prediction Using Neural Sequence-to-Sequence Models. ACS Cent. Sci. 2017, 3, 1103-1113], the difficulty of learning the syntactic structure of SMILES notation may lead to problematic results such as inappropriate SMILES. In general, existing text-based models have the problem of generating literal invalid, literally valid but chemically irrational, or literal and chemically valid but impractical candidates. The embodiment solves these problems by projecting the SMILES representation of the molecular structure into the substructural domains.

In general, the probability of making a correct retrosynthetic prediction is relatively low. In fact, the accuracy of the reverse synthesis planning task is about twice as low as the level of accuracy achieved in the forward prediction task. This is especially true given that several possible synthetic routes are available for forward reactions. It is worth noting that the contents of the dataset used in the reverse mapping may also be responsible for the behavior of the network. Given the level of abstraction for describing molecules in the dataset of examples, mapping reactants from reactant domains to product domains and then vice versa does not necessarily give the original reactants. The existence of a one-to-many mapping from product to reactant domains can cause confusion during the learning process. Based on these observations, a simple idea was adopted to ensure a stronger pairwise functional relationship between domains. To this end, the embodiment checks all one-to-many mappings, selects the molecule with the shortest sequence length (presumed to be the molecule with the lowest level of structural complexity) and maps it to one-to-one mapping. reduced to

In particular, the model of the embodiment provides consistent predictions. In each independent run on the same input molecule, the model of the Example consistently gave the same output. The robustness of the example model can be attributed to its low complexity and good interpretability of molecular descriptors. In general, retrosynthetic models use top-N accuracy scores to represent overall model performance. However, recently, Schwaller, P.; Petraglia, R.; Zullo, V.; Nair, V. H.; Haeuselmann, R. A.; Pisoni, R.; Bekas, C.; Iuliano, A.; Laino, T. Predicting retrosynthetic pathways using transformer based models and a hyper-graph exploration strategy. Chem. Sci. 2020, 11, 3316-3325], the top-N accuracy score is because for each proposal the model tends to produce predicted answers from the dataset rather than chemically more meaningful predictions. It may not be a suitable metric to evaluate the retrosynthesis model. Although MACCS keys have been criticized for poor performance in similarity benchmarks, the advantage of this descriptor is that there is a one-to-one correspondence between bits and substructures compared to the fingerprint obtained by an exhaustive generation algorithm and subsequent hashing procedure. Therefore, the MACCS key was a natural choice for testing the proof-of-concept level of the translation methodology.

Using substructure-based representations makes it difficult to convert predicted sequences into molecules. The SMILES representation of the molecular structure is not unique, but it is reversible. On the other hand, MACCS key representations exhibit unique, irreversible properties. For this reason, the embodiment must perform a lookup to access the molecular structure in the database of pre-computed MACCS keys. The example constructs a lookup table from the USPTO reaction dataset and associates it with the predicted MACCS key for the reactant molecule. Candidate reactants may also be obtained from commercially available chemical databases or reaction databases based on similarity searches. Some such as PubChem (Nakata, M.; Shimazaki, T. PubChemQC Project: A Large-Scale First-Principles Electronic Structure Database for Data-Driven Chemistry. Journal of Chemical Information and Modeling 2017, 57, 1300-1308) Online chemistry databases may also be used for the same purpose by allowing MACCS key based queries.

11 shows reaction 4 of FIG. 8 in more detail. The similarity score was expressed using a circular fingerprint (Morgan) with a radius of 2 and the Tanimoto metric. Unique pieces marked with SMARTS patterns. FIG. 11 also shows five candidates obtained from the USPTO reaction dataset for the first reactant of reaction 4 of FIG. 8 . All five reactants are associated with other reactions in the database. The MACCS key representation of the obtained molecules is the same. This means that sometimes it is likely to find one or more matches corresponding to the prediction sequence. These closely related similarities can be aligned by computing Tanimoto coefficients using path-based or circular fingerprints. Because they will be different for the same set. To this end, the embodiment uses a circular fingerprint of radius 2 as a bit vector. In the example, the molecule having the highest similarity value among the candidates was selected as the final result.

conclusion

The embodiment provides a sequence-to-sequence NMT model for automatically extracting reaction rules of chemical reactions by learning relationships at the sub-structure level. By constructing an abstract language with a short fixed-length vocabulary with non-zero constructs of MACCS keys, the three conceptual problems are addressed and solved together. (i) irregular prediction: SMILES-based representation easily exposes model results to error, (ii) synthetic availability: predicted molecule is not synthetically accessible, (iii) top-N accuracy metric: suggestion by model may change depending on the execution of the model. Comparisons and quality checks show that the Examples successfully provide candidate reactants in the range of 0.85 < T _c ≤ 1.00, and achieve a high overall level of accuracy, particularly for functional group interconversion or dissociation and reactive functional groups. The approach according to the embodiment has high potential to be applied to various fields of organic chemistry.

The embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the computer software field. Examples of the computer-readable recording medium include a hard disk, a magnetic medium such as a floppy disk and a magnetic tape, an optical recording medium such as a CD-ROM and DVD, and a magneto-optical medium such as a floppy disk. media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules for carrying out the processing according to the present invention, and vice versa.

In addition, the embodiments according to the present invention described above may be a set of program instructions that can be executed through various computer components and a user application itself for executing them. Specifically, it may be a program itself that can be downloaded through a server or a storage medium and installed on a client computer.

In the above, the present invention has been described with specific matters such as specific components and limited embodiments and drawings, but these are provided to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments. , various modifications and variations can be devised from these descriptions by those of ordinary skill in the art to which the present invention pertains.

Therefore, the spirit of the present invention should not be limited to the above-described embodiments, and not only the claims described below, but also all modifications equivalently or equivalently to the claims described below belong to the scope of the spirit of the present invention. will do it

10: reverse synthesis system 11: control unit
12: communication unit 13: input/output interface unit
14: memory unit 15: input unit
16: display unit 20: external server

Claims (15)

A method for predicting reverse synthesis using neural network machine translation that can be implemented in a reverse synthesis system including a communication unit, a control unit, and a memory unit, the method comprising:
A step of generating an initial product-reactant pair, the step of receiving a reaction data set for the product-reactant from an external server through the communication unit; expressing the received response data set as a descriptor composed of characters; and storing the initial product-reactant pair represented by the descriptor in the memory unit, generating the initial product-reactant pair;
filtering by applying one or more filters to the initial product-reactant pair;
one-to-one mapping of the filtered product-reactant pairs;
generating a final product-reactant pair by aligning the product-reactant pair to which the one-to-one mapping is applied;
training a neural machine translation (NMT) model using the final product-reactant pair;
evaluating the learned NMT model; and
When a new product is input to the NMT model, predicting and calculating a candidate reactant corresponding to the new product. The method of claim 1,
The step of expressing the received response data set as a descriptor composed of characters,
Preparing a MACCS (molecular access system) key representing a molecule;
generating a character descriptor for the MACCS key by allocating a predetermined character to each of the MACCS keys;
expressing the received response data set with the MACCS key; and
and converting the response dataset expressed by the MACCS key into the character descriptor. 3. The method of claim 2,
The preparing of the MACCS key includes generating a curated MACCS key by removing MACCS keys having a frequency of occurrence less than or equal to a predetermined among the MACCS keys. 4. The method of claim 3,
In the generating a character descriptor for the MACCS key, the predetermined character is assigned to each of the curated MACCS keys. The method of claim 1,
Filtering by applying one or more filters to the initial product-reactant pair,
A reverse synthesis comprising at least one of a filter that removes pairs having 3 or more reactants, a filter that removes identical product-reactant pairs, a filter that removes inner twins, and a filter that removes pairs with a sequence length of 100 or more. Prediction method. The method of claim 1,
The one-to-one mapping of the filtered product-reactant pair comprises:
Reverse synthesis prediction, confirming the one-to-many mapping for the filtered product-reactant pair, and reducing the one-to-one mapping by selecting the molecule having the shortest sequence length among the one-to-many mappings method. The method of claim 1,
The step of aligning the product-reactant pair to which the one-to-one mapping is applied to generate a final product-reactant pair comprises:
When there are two or more reactants of the product-reactant pair, they are arranged in descending order according to length, and symbols are connected between the reactants. The method of claim 1,
The neural network machine translation model is a sequence-to-sequence model, a reverse synthesis prediction method. 9. The method of claim 8,
The neural network machine translation model includes two bidirectional long short-term memories (LSTMs),
and one of the two bidirectional LSTMs functions as an encoder and the other functions as a decoder. 10. The method of claim 9,
wherein the encoder and the decoder are connected through an attention mechanism. 10. The method of claim 9,
Each of the bidirectional LSTMs includes a hidden layer,
Inverse synthesis prediction method, further comprising a dropout layer after the hidden layer. 12. The method according to claim 10 or 11,
The neural network machine translation model further comprises gradient clipping so that a norm of a gradient does not exceed a threshold value during backpropagation. 13. The method of claim 12,
The learning rate of the neural network machine translation model is reduced by a constant coefficient every three epochs, inverse synthesis prediction method. The method of claim 1,
Evaluating the learned NMT model comprises:
A reverse synthesis prediction method for evaluating prediction accuracy by calculating a similarity between a candidate reactant predicted through the NMT model and an actual reactant obtained through the final product-reactant pair. The method of claim 1,
The step of predicting and calculating the candidate reactant,
The method further comprising predicting and calculating a secondary candidate reactant for synthesizing the candidate reactant.

Images