A deep learning-based biosynthetic pathway prediction method and system
By employing a deep learning approach based on the Transformer model, the challenge of elucidating the synthetic routes of high-value molecules in organisms has been solved. This approach enables rapid and accurate prediction of biosynthetic pathways, reduces human intervention and errors, and improves the reliability of enzyme reaction prediction.
Patent Information
- Application Number
- CN202310808179.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-03
- Publication Date
- 2026-01-06
- Estimated Expiration
- 2043-07-03
AI Technical Summary
Existing technologies are insufficient to effectively analyze the synthetic routes of high-value molecules in living organisms, resulting in artificial retrosynthetic analysis being resource-intensive and prone to errors. Furthermore, existing deep learning methods are difficult to apply to complex enzyme reactions.
We employ a deep learning approach based on the Transformer model to acquire and preprocess enzyme-catalyzed reaction data, generate reaction fingerprints, train the model using a loss function, predict precursor molecules, and compare reaction fingerprints to generate biosynthetic pathways, thereby reducing human intervention and errors.
It enables rapid and accurate prediction of biosynthetic routes, improves the accuracy and reliability of predictions, saves scientific research time and resources, and provides protease information to guide experimental verification.
Smart Images
Figure CN116825202B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of synthetic biology and artificial intelligence, and more specifically, to a method and system for predicting biosynthetic pathways based on deep learning. Background Technology
[0002] Biosynthesis, achieved through the catalytic action of proteases within living organisms, holds great promise for the preparation of high-value molecules such as drugs and additives needed by humans. However, the synthetic routes of most molecules required by humans within living organisms have not yet been elucidated, which hinders the application of biosynthesized high-value small molecules.
[0003] For the biosynthesis of high-value small molecules, retrosynthesis is typically employed to design synthetic routes. First, the target molecule is analyzed and broken down into multiple smaller fragments. These fragments can combine through certain reactions to form the target molecule, often called a precursor molecule. Then, these precursor molecules are further broken down into even smaller fragments, and this process is repeated until easily obtainable small molecules, such as oxygen or carbon dioxide, are synthesized—the building blocks. By working backward from this process, the desired target molecule can be synthesized using readily available small molecules. However, retrosynthetic analysis typically requires extensive experience from researchers to determine the rationality and feasibility of each step of the breakdown. Furthermore, the methods for breaking down different types of molecules may vary. Analysis relying on artificial retrosynthesis is extremely labor-intensive and requires significant resources for experimental verification.
[0004] Currently, there are proposed retrosynthetic analysis methods based on reaction rules. These methods first analyze and summarize existing chemical reactions to derive universal reaction rules, which are then used to replace manual separation in retrosynthetic analysis. However, these reaction rules still require manual summarization. Enzyme reaction data contains many cofactors and irrelevant products in addition to the main components, making the reaction information more complex, increasing the difficulty of summarizing reaction rules, and drastically increasing the workload. Furthermore, since reaction rules mainly rely on manually finding patterns in existing data, it is difficult to avoid errors caused by subjective factors; that is, molecular separation using reaction rules may still be unreasonable or infeasible. There are also proposed retrosynthetic prediction methods based on deep learning. However, these methods focus on organic synthesis. For biological enzyme reactions, due to the aforementioned problems of more complex reaction information, transferring existing deep learning-based retrosynthetic methods to enzyme reactions is quite difficult. Summary of the Invention
[0005] To overcome at least one of the defects described in the prior art, the present invention provides a method and system for predicting biosynthetic pathways based on deep learning.
[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0007] A deep learning-based method for predicting biosynthetic pathways includes:
[0008] S1. Obtain raw data of organic chemical reactions and enzyme-catalyzed reactions and preprocess them to obtain reaction group data; the reaction group data includes product molecule data and corresponding real precursor molecule data;
[0009] S2. Represent the reaction group data using SMILES sequences and generate corresponding reaction fingerprints to form a fingerprint database;
[0010] S3. After encoding the SMILES sequence of the product molecule data, input it into the Transformer model to obtain the SMILES sequence of the predicted precursor molecule. Then, based on the SMILES sequence encoding of the predicted precursor molecule and the SMILES sequence encoding corresponding to the real precursor molecule data, train the Transformer model based on the loss function.
[0011] S4. Input the product molecule to be predicted into the trained Transformer model to predict the precursor molecule, and obtain m predicted precursor molecules; where m is a positive integer greater than or equal to 1.
[0012] S5. Combine the product molecule to be predicted with any predicted precursor molecule to form a complete reaction and generate a corresponding predicted reaction fingerprint. Match and compare the m predicted reaction fingerprints with the reaction fingerprints in the fingerprint library respectively. Add the predicted precursor molecule and enzyme protein corresponding to the reaction with the highest similarity to the synthesis route.
[0013] The predicted precursor molecules and enzyme proteins in the synthetic route are sorted in descending order according to their similarity.
[0014] S6. Determine whether the predicted precursor molecule output by the current Transformer model is a preset building block or whether the preset number of iterations has been reached: if yes, output the current synthesis route; otherwise, use the current predicted precursor molecule as the product molecule to be predicted and jump to step S4.
[0015] Furthermore, this invention proposes a deep learning-based biosynthetic pathway prediction system, which applies the biosynthetic pathway prediction method proposed in this invention, including:
[0016] The data acquisition module is used to acquire raw data of organic chemical reactions and enzyme-catalyzed reactions and preprocess them to obtain reaction group data;
[0017] The fingerprint database consists of corresponding reaction fingerprints generated from the reaction group data.
[0018] The prediction module is equipped with a trained Transformer model, which is used to predict the biosynthetic pathway of the input product molecule to be predicted and output m predicted precursor molecules.
[0019] In the prediction module, the product molecule to be predicted is combined with any predicted precursor molecule to form a complete reaction and generate a corresponding predicted reaction fingerprint. The m predicted reaction fingerprints are matched and compared with the reaction fingerprints in the fingerprint library. The predicted precursor molecule and enzyme protein corresponding to the reaction with the highest similarity are added to the synthesis route. Then, it is determined whether the predicted precursor molecule output by the current Transformer model is a preset building block or whether the preset number of iterations has been reached. If so, the current synthesis route is output. Otherwise, the current predicted precursor molecule is input into the Transformer model again as the product molecule to be predicted for prediction.
[0020] Compared with existing technologies, the beneficial effects of the present invention are as follows: The present invention applies the Transformer model to the prediction of enzyme reactions, and completes the prediction of biosynthetic routes through the prediction cycle. It can replace artificial retrosynthetic design or reaction rule extraction with data training, and can predict the biosynthetic routes of any high-value small molecule, and provide information on the proteases required in the route to guide experimental verification. The present invention effectively improves the accuracy and reliability of biosynthetic pathway prediction by introducing reaction fingerprint comparison in the process of evaluating the reliability of predicted reactions. Attached Figure Description
[0021] Figure 1 This is a flowchart of the biosynthetic pathway prediction method of the present invention.
[0022] Figure 2 This is a flowchart for preprocessing raw data.
[0023] Figure 3 A schematic diagram of the process for extracting atom-to-atom mappings.
[0024] Figure 4 This is a diagram of the Transformer model framework.
[0025] Figure 5 This is a diagram illustrating the architecture of the biosynthetic pathway prediction system of the present invention. Detailed Implementation
[0026] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent; it is understandable that some well-known descriptions in the drawings may be omitted in order to better illustrate this embodiment.
[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0028] Example 1
[0029] This embodiment proposes a deep learning-based method for predicting biosynthetic pathways, such as... Figure 1 The diagram shown is a flowchart of the deep learning-based biosynthetic pathway prediction method in this embodiment.
[0030] The deep learning-based biosynthetic pathway prediction method proposed in this embodiment includes the following steps:
[0031] S1. Obtain raw data of organic chemical reactions and enzyme-catalyzed reactions and preprocess them to obtain reaction group data; the reaction group data includes product molecule data and corresponding real precursor molecule data.
[0032] S2. Represent the reaction group data using SMILES sequences and generate corresponding reaction fingerprints to form a fingerprint database.
[0033] S3. After encoding the SMILES sequence of the product molecule data, input it into the Transformer model to obtain the SMILES sequence of the predicted precursor molecule. Then, based on the SMILES sequence encoding of the predicted precursor molecule and the SMILES sequence encoding corresponding to the real precursor molecule data, train the Transformer model based on the loss function.
[0034] S4. Input the product molecule to be predicted into the trained Transformer model to predict the precursor molecule, and obtain m predicted precursor molecules.
[0035] Where m is a positive integer greater than or equal to 1.
[0036] S5. Combine the product molecule to be predicted with any predicted precursor molecule to form a complete reaction and generate a corresponding predicted reaction fingerprint. Match and compare the m predicted reaction fingerprints with the reaction fingerprints in the fingerprint library. Add the predicted precursor molecule and enzyme protein corresponding to the reaction with the highest similarity to the synthesis route.
[0037] The predicted precursor molecules and enzyme proteins in the synthetic route are sorted in descending order according to their similarity.
[0038] S6. Determine whether the predicted precursor molecule output by the current Transformer model is a preset building block or whether the preset number of iterations has been reached: if yes, output the current synthesis route; otherwise, use the current predicted precursor molecule as the product molecule to be predicted and jump to step S4.
[0039] In this embodiment, the complex reaction data is first preprocessed by extracting only the main components of the reaction and removing irrelevant factors to obtain high-quality reaction data. Then, the Transformer model is used to learn the reaction rules directly from the reaction data. This not only allows new data to be incorporated into the model for training at any time, but also reduces manual costs and errors, thereby achieving rapid and accurate retrosynthetic analysis. Furthermore, the hidden rules learned by the model can be used to find proteases during the analysis process.
[0040] This embodiment applies the Transformer model to predict enzyme reactions and completes the prediction of biosynthetic routes through a prediction cycle. Data training can replace manual retrosynthetic design or reaction rule extraction, saving time and resources in scientific research. Furthermore, for any high-value small molecule, this embodiment can predict its biosynthetic route and provide information on the proteases required in the route to guide experimental verification.
[0041] Furthermore, this embodiment introduces reaction fingerprint comparison in the process of evaluating the reliability of predicted reactions, effectively improving the accuracy and reliability of biosynthetic pathway prediction. While the Transformer model can learn potential reaction rules from data, its interpretability is poor, and the prediction results lack a rigorous source of evidence, making its reliability assessment difficult. This embodiment uses reaction fingerprint comparison to determine the similarity between the predicted reaction and known reactions, thereby verifying the reliability of the reaction to some extent.
[0042] In one alternative embodiment, existing organic chemical reaction and enzyme-catalyzed reaction data are collected from databases and academic papers, and the collected reaction data includes information about enzyme proteins.
[0043] In an optional embodiment, the raw data is preprocessed, including:
[0044] 1) Remove data from the original data where carbon atoms are not conserved;
[0045] 2) Based on the atom-to-atom mapping method, the atoms before and after the reaction in the original data are mapped, and the atoms of the precursor molecules and product molecules in each reaction are numbered so that they correspond one-to-one;
[0046] 3) For each product molecule in any reaction, determine whether any precursor molecule contains an atom from the corresponding product. If so, retain the corresponding precursor molecule; otherwise, discard the corresponding precursor molecule. That is, retain the precursor molecule containing the atom number of the product.
[0047] Furthermore, in an optional embodiment, common substructure matching is used to extract the common parts in the precursor molecule data and product molecule data, which are used to directly generate atomic maps; for atoms other than the common substructure, bipartite graph matching is used for atomic mapping.
[0048] like Figure 2 The diagram shown is a flowchart of the original data preprocessing in this embodiment.
[0049] This embodiment eliminates carbon-non-conservation reactions by calculating the number of carbon atoms in the precursor and product molecules, and then removes duplicate reactions. An atom-to-atom mapping method is used to number the atoms of the precursor and product molecules in each reaction, ensuring a one-to-one correspondence. Only the precursor molecule containing the product atom number is retained. For example... Figure 3 The diagram shows a flowchart of atom-to-atom mapping extraction. When a reaction has multiple products, for each product, the precursor molecule containing its atom number must be retained. Finally, principal component extraction is performed on all reactions, followed by deduplication to complete data preprocessing.
[0050] Further, optionally, the preprocessed reaction group data can be divided into a training set, a validation set, and a test set for model training and testing.
[0051] In this embodiment, considering that complex cofactors and byproducts in enzyme-catalyzed reactions often have long smiles sequences, directly applying them to Transformer model training would increase complexity and introduce interference. In this embodiment, each preprocessed reaction data point contains only one product, and each carbon atom of the product can be traced back to the precursor molecule. This ensures the integrity of the reaction while simplifying data complexity, thereby improving the speed and accuracy of Transformer model predictions.
[0052] In an optional embodiment, encoding the SMILES sequence of the product molecule data includes:
[0053] The SMILES sequence of the reaction group data is split into individual characters, and all characters are sorted in descending order of their frequency of occurrence to construct a vocabulary of size n; where n is the total number of characters after deduplication.
[0054] The SMILES sequence of the product molecule data is encoded according to the position in the vocabulary; wherein, any character in the SMILES sequence of the product molecule data is represented by an array of length n; in the array, the position corresponding to the current character in the vocabulary is set to 1, and the other positions are set to 0.
[0055] As an example, during the encoding process, if the split character is located at the 5th position of the vocabulary, then only the 5th position of the array encoding that character is 1, and the rest are 0.
[0056] Further, in an optional embodiment, training the Transformer model based on a loss function includes:
[0057] Based on the predicted precursor SMILES sequence encoding and the corresponding SMILES sequence encoding of the real precursor molecule data, the cross-entropy loss is calculated, and the Transformer model is trained with the goal of minimizing the cross-entropy.
[0058] In an optional embodiment, matching the predicted reaction fingerprint with reaction fingerprints in the fingerprint database includes:
[0059] The similarity between the predicted reaction fingerprint and any reaction fingerprint in the fingerprint database is evaluated by calculating the Euclidean distance between them, and the reaction fingerprint with the smallest Euclidean distance is selected from the fingerprint database as the reference reaction.
[0060] Optionally, a reaction fingerprint conversion tool can be used to generate a corresponding predicted reaction fingerprint from the reaction consisting of the product molecule to be predicted and the predicted precursor molecule.
[0061] In this embodiment, Euclidean distance is selected to evaluate the similarity between the predicted reaction fingerprint and the reaction fingerprint in the fingerprint database.
[0062] As an example, the Euclidean distance for a reaction fingerprint consisting of a one-dimensional vector of length 256 is calculated as follows:
[0063]
[0064] In the formula, x i y represents the number of fingerprints of dimension i in a given reaction fingerprint in the fingerprint database. i The number of dimensions i in the predicted reaction fingerprint is represented.
[0065] In this embodiment, several possible precursor molecules are predicted using the Transformer model. Then, the precursor molecules and the input product are combined to form a reaction and generate a reaction fingerprint. This fingerprint is compared with an existing reaction fingerprint database. The reaction with the highest similarity is taken as a reference, and the enzyme protein of that reaction is extracted and output. The reaction is then sorted by similarity to form a synthetic route to improve the reliability of the predicted reaction.
[0066] Example 2
[0067] This embodiment uses the biosynthetic pathway prediction method proposed in Example 1 for further explanation.
[0068] like Figure 4 The diagram shown is a framework diagram of the Transformer model in this embodiment.
[0069] To predict the biosynthetic route of a target molecule such as glutaric acid, first input its SMILES.
[0070] “O=C(O)CCCC(=O)O” or hand-drawn structure:
[0071]
[0072] Convert this to SMILES: "O=C(O)CCCC(=O)O", specifying that the number of predicted precursor molecules is 3 each time, and the building blocks are the default 368 common small molecules. The Transformer model will output the top three predicted precursor molecules SMILES: "O=C(O)CCCC(=O)[CoA]".
[0073] The molecular structures of “O=C(O)CCCC(=O)C(=O)O” and “NC(=O)CCCC(=O)O” are as follows:
[0074]
[0075] The precursor and target molecule are combined separately to form a complete reaction, and represented in the form of SMILES:
[0076] The following reactions were analyzed using a reaction fingerprinting tool: “O=C(O)CCCC(=O)O>>O=C(O)CCCC(=O)[CoA]”, “O=C(O)CCCC(=O)O>>O=C(O)CCCC(=O)C(=O)O”, and “O=C(O)CCCC(=O)O>>NC(=O)CCCC(=O)O”. The molecular fingerprints of these three reactions were calculated, and their maximum similarity to reactions in the fingerprint database was calculated and ranked according to this maximum similarity. The ranking result showed that the first reaction had the highest similarity, meaning the first predicted precursor molecule had the highest probability of synthesizing the target molecule. Furthermore, the enzyme required for the reaction with the highest similarity in the fingerprint database could also be considered a potential enzyme catalyzing that reaction. Therefore, the predicted precursor molecule (“O=C(O)CCCC(=O)[CoA]”) was used as the target molecule, and the above steps were repeated until the predicted precursor molecule output by the model was included in the predefined building blocks or the preset number of iterations was reached. The final synthetic route given by the Transformer model was as follows:
[0077] "O=C(O)CCCC(=O)O>>O=C(O)CCCC(=O)[CoA]", "O=C(O)CCCC(=O)[CoA]>>O=C(O)CCCC(=O)C(=O)O", that is:
[0078]
[0079] Except for the target molecule, which needs to be specified manually, the above process can be run automatically by the device, which will automatically generate a synthetic route diagram in structural form.
[0080] Example 3
[0081] This embodiment proposes a deep learning-based biosynthetic pathway prediction system, applying the biosynthetic pathway prediction method proposed in Embodiment 1. For example... Figure 5 The diagram shown is an architecture diagram of the biosynthetic pathway prediction system in this embodiment.
[0082] The deep learning-based biosynthetic pathway prediction system proposed in this embodiment includes:
[0083] The data acquisition module is used to acquire raw data of organic chemical reactions and enzyme-catalyzed reactions and preprocess them to obtain reaction group data.
[0084] The fingerprint database consists of corresponding reaction fingerprints generated from the reaction group data.
[0085] The prediction module is equipped with a trained Transformer model, which is used to predict the biosynthetic pathway of the input product molecule to be predicted and output m predicted precursor molecules.
[0086] In the prediction module, the product molecule to be predicted is combined with any predicted precursor molecule to form a complete reaction and generate a corresponding predicted reaction fingerprint. The m predicted reaction fingerprints are matched and compared with the reaction fingerprints in the fingerprint library. The predicted precursor molecule and enzyme protein corresponding to the reaction with the highest similarity are added to the synthesis route. Then, it is determined whether the predicted precursor molecule output by the current Transformer model is a preset building block or whether the preset number of iterations has been reached. If so, the current synthesis route is output. Otherwise, the current predicted precursor molecule is input into the Transformer model again as the product molecule to be predicted for prediction.
[0087] In an optional embodiment, the system further includes an input component for specifying product molecules by inputting SMILES sequences or hand-drawn structures, wherein, for hand-drawn structures, the hand-drawn structures are converted into normalized SMILES sequences using a conversion tool; and / or for setting building blocks by inputting SMILES sequences or hand-drawn structures.
[0088] In an optional embodiment, the system further includes a parameter specification component for configuring parameters in the prediction module; the parameters include the number of predicted precursor molecules output by the Transformer model at one time, the number of iterations to complete the prediction of biosynthetic pathways, and / or the number of output routes.
[0089] In an alternative embodiment, the system further includes a display module for displaying the synthetic routes output by the prediction module in the form of an interactive network, and / or for displaying the enzyme protein for each predicted reaction.
[0090] It is understood that the system in this embodiment corresponds to the method in Embodiment 1 above, and the options in Embodiment 1 above are also applicable to this embodiment, so they will not be described again here.
[0091] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A deep learning-based biosynthetic pathway prediction method, characterized in that, The application relates to a method for predicting precursor molecules and enzymes in a synthetic route. The method comprises the following steps: S1, obtaining original data of organic chemical reactions and enzyme-catalyzed reactions and preprocessing the original data to obtain reaction group data; the reaction group data comprises product molecule data and corresponding real precursor molecule data; S2, representing the reaction group data by using SMILES sequences and generating corresponding reaction fingerprints to form a fingerprint library; S3, inputting the SMILES sequence of the product molecule data into a Transformer model after encoding, obtaining a SMILES sequence of a predicted precursor molecule, encoding the SMILES sequence of the predicted precursor molecule according to the SMILES sequence of the real precursor molecule data, and training the Transformer model based on a loss function; S4, inputting a product molecule to be predicted into the trained Transformer model to predict a precursor molecule, and obtaining m predicted precursor molecules; wherein m is a positive integer greater than or equal to 1; S5, forming a complete reaction by using the product molecule to be predicted and any predicted precursor molecule and generating a corresponding predicted reaction fingerprint, respectively matching and comparing m predicted reaction fingerprints with reaction fingerprints in the fingerprint library, and adding the predicted precursor molecule and the enzyme protein corresponding to the reaction with the highest similarity to the synthetic route; The predicted precursor molecules and enzyme proteins in the synthetic route are sorted in descending order according to the similarity; 2. The biosynthetic pathway prediction method according to claim 1, wherein, S6, judging whether the predicted precursor molecule output by the current Transformer model is a preset building block or whether the preset iteration number is reached: if yes, outputting the current synthetic route; otherwise, taking the current predicted precursor molecule as the product molecule to be predicted and jumping to execute the S4 step. In the S1 step, the preprocessing of the original data comprises the following steps: Eliminating data in which the number of carbon atoms is not conserved in the original data; Mapping atoms before and after reactions in the original data based on an atom-atom mapping method; 3. The biosynthetic pathway prediction method according to claim 2, wherein, For each product molecule data in any reaction, judging whether any precursor molecule contains atoms in the corresponding product, and if yes, retaining the corresponding precursor molecule, otherwise eliminating the corresponding precursor molecule. The atom mapping before and after reactions in the original data based on the atom-atom mapping method comprises the following steps: Using common substructure matching to extract parts common to precursor molecule data and product molecule data for directly generating atom mapping; 4. The biosynthetic pathway prediction method according to claim 1, wherein, For atoms outside the common substructure, atom mapping is performed by using bipartite graph matching. In the S3 step, the SMILES sequence of the product molecule data is encoded, comprising the following steps: Splitting the SMILES sequence of the reaction group data into single characters, and constructing a vocabulary table with a size of n by sorting all characters in descending order according to the number of character occurrences; wherein n is the total number of characters after deduplication; According to the position of the vocabulary, the SMILES sequence of the product molecule data is encoded; wherein, for any character in the SMILES sequence of the product molecule data, an array with a length of n is used for representation; in the array, the position corresponding to the current character in the vocabulary is set to 1, and the rest of the positions are set to 0.
5. The biosynthetic pathway prediction method according to claim 1, wherein, In the S3 step, the Transformer model is trained based on a loss function, including: According to the SMILES sequence encoding of the predicted precursor molecule and the SMILES sequence encoding corresponding to the real precursor molecule data, the cross-entropy loss is calculated, and the Transformer model is trained with the goal of minimizing the cross-entropy.
6. The biosynthetic pathway prediction method according to claim 1, wherein, In the S5 step, the predicted reaction fingerprint is matched and compared with the reaction fingerprints in the fingerprint library, including: The similarity between the predicted reaction fingerprint and any reaction fingerprint in the fingerprint library is evaluated by calculating the Euclidean distance, and the reaction fingerprint with the smallest Euclidean distance is selected as the reference reaction.
7. A deep learning-based biosynthetic pathway prediction system applying the biosynthetic pathway prediction method according to any one of claims 1 to 6, characterized by, Including: A data acquisition module for obtaining raw data of organic chemical reactions and enzyme-catalyzed reactions and preprocessing the raw data to obtain reaction group data; A fingerprint library composed of reaction fingerprints generated from the reaction group data; A prediction module having a trained Transformer model configured thereon for predicting the biosynthesis pathway of an input product molecule to be predicted and outputting m predicted precursor molecules; In the prediction module, the product molecule to be predicted and any predicted precursor molecule form a complete reaction to generate a corresponding predicted reaction fingerprint, and the m predicted reaction fingerprints are matched and compared with the reaction fingerprints in the fingerprint library, respectively, and the predicted precursor molecule and enzyme protein corresponding to the reaction with the highest similarity are added to the synthesis route; then it is judged whether the predicted precursor molecule output by the current Transformer model is a preset building block or whether the preset iteration number is reached: if yes, the current synthesis route is output; otherwise, the current predicted precursor molecule is input into the Transformer model again for prediction as the product molecule to be predicted.
8. The biosynthetic pathway prediction system of claim 7, wherein, The system further includes an input component for specifying a product molecule by inputting a SMILES sequence or a hand-drawn structure, wherein for a hand-drawn structure, the hand-drawn structure is converted into a standardized SMILES sequence by a conversion tool; and / or for setting a building block by inputting a SMILES sequence or a hand-drawn structure.
9. The biosynthetic pathway prediction system of claim 7, wherein, The system further includes a parameter specification component for configuring parameters in the prediction module; the parameters include the number of predicted precursor molecules output by the Transformer model at a time, the iteration number for completing the biosynthesis pathway prediction, and / or the number of output routes.
10. The biosynthetic pathway prediction system of claim 7, wherein, The system further includes a display module for displaying the synthesis route output by the prediction module in the form of an interactive network, and / or for displaying the enzyme protein of each predicted reaction.
Citation Information
Patent Citations
Protein ligand affinity prediction method, related device and equipment
CN115116538A
Marine nutritional ingredient biosynthetic pathway mining method, device, equipment and medium
CN116072227A