A method and device for evaluating the synthesizability of biomolecules based on reinforcement learning

Abstract

This invention relates to the fields of biomolecule synthesis technology and artificial intelligence technology, specifically disclosing a method and apparatus for assessing the synthesizability of biomolecules based on reinforcement learning. The method includes: acquiring a preset dataset; merging the preset dataset to generate a merged dataset; performing structural verification and repair operations on the merged dataset to generate a corrected dataset; extracting bioaccessibility features from the corrected dataset to generate a bioaccessibility database; scoring the synthesis paths of the molecules to be analyzed based on a multimodal fusion scoring model and the bioaccessibility database, and determining the optimal path; and performing similarity clustering and replacement on the unsynthesized molecules in the optimal path to optimize the synthesis path. Through systematic data processing, multimodal feature fusion, reinforcement learning path search, and similarity replacement, efficient and accurate synthesis path assessment is achieved.

Description

Technical Field

[0001] This invention relates to the fields of biomolecule synthesis technology and artificial intelligence technology, specifically to a method and apparatus for assessing the syntheticability of biomolecules based on reinforcement learning. Background Technology

[0002] In the field of biomolecule synthesis, the assessment of whether a molecule can be synthesized through a pure metabolic pathway (biosynthesis) or requires chemical finishing steps has long relied on manual judgment and lacks a systematic method, resulting in many significant shortcomings.

[0003] First, the data processing is disorganized. Molecular data is scattered across multiple databases such as MetaNetX and PubChem, with inconsistent data formats. This leads to the same molecule being recorded repeatedly with different identifiers in different databases. Furthermore, irrelevant information such as salt ions is often mixed in, increasing the difficulty of data integration and rendering the data unusable for subsequent analysis. Second, the reliability of molecular structure data is insufficient. Molecular structures in databases often have issues such as unreasonable atomic valence states, incorrect aromatic ring labeling, and unoptimized 3D structures, resulting in low accuracy of structure-based feature extraction and affecting the judgment of molecular synthetic properties. Third, feature extraction is limited in scope. Traditional methods often rely solely on physicochemical properties, neglecting key biological characteristics such as metabolic pathway involvement and enzyme correlations, leading to a one-sided judgment of molecular biosynthesizability and a high risk of misjudgment. Additionally, path evaluation is inefficient. Due to the diversity of molecular synthetic pathways, traditional methods rely on exhaustive search for multiple pathways, consuming significant computational resources and failing to quickly determine the optimal path with the lowest energy. Finally, the lack of processing for unsynthesized molecules in the pathway makes it impossible to efficiently determine their syntheticity. This means that even if a theoretically optimal path is found, it may be impractical due to the presence of unsynthesized molecules. These shortcomings collectively result in low efficiency and poor accuracy in assessing the syntheticity of biomolecules, making it difficult to meet the need for rapid decision-making on synthetic pathways at the start of a project. Summary of the Invention

[0004] To overcome the aforementioned technical problems in the prior art, this invention provides a method and apparatus for assessing the biomolecule synthesizability based on reinforcement learning. Through systematic data processing, feature library construction, path scoring, and optimized replacement, it achieves efficient and accurate assessment of the biomolecule synthesizability.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for assessing the syntheticability of biomolecules based on reinforcement learning. The method includes: acquiring a preset dataset; merging the preset dataset to generate a merged dataset; performing structural verification and repair operations on the merged dataset to generate a corrected dataset; extracting bioaccessibility features from the corrected dataset to generate a bioaccessibility database; scoring the synthesis paths of the molecules to be analyzed based on a multimodal fusion scoring model and the bioaccessibility database and determining the optimal path; and performing similarity clustering replacement on the unsynthesized molecules in the optimal path to optimize the synthesis path.

[0006] Preferably, obtaining the preset dataset includes: selecting a preset standard database, which includes MetaNetX, PubChem, and other optional databases; obtaining data of the molecular set to be analyzed in the preset standard database; and clustering the obtained data to form a preset dataset.

[0007] Preferably, the step of merging the preset dataset to generate a merged dataset includes: acquiring chemical and structural data of the set of molecules to be analyzed in the preset dataset; performing standardization processing on the chemical and structural data to obtain a standardized dataset; merging identical molecules based on the standardized dataset to form a first dataset; and performing ion processing on the first dataset to form the merged dataset.

[0008] Preferably, the step of performing structural verification and repair operations based on the merged dataset and generating a corrected dataset includes: performing chemical rationality verification based on the merged dataset and obtaining a verification dataset; performing aromatic ring correction based on the verification dataset and obtaining a corrected dataset; performing molecular 3D structure analysis based on the corrected dataset and obtaining an initial energy dataset; and performing energy minimization analysis based on the initial energy dataset and obtaining a corrected dataset.

[0009] Preferably, the step of extracting biological accessibility features based on the modified dataset and generating a biological accessibility database includes: acquiring the physicochemical properties and structural features of the modified dataset and generating a first feature set; extracting metabolic and reaction features and enzyme-related features of the modified dataset and generating a second feature set; performing bioavailability prediction features on the modified dataset based on a pre-trained model and generating a third feature set; and constructing the biological accessibility database based on the first feature set, the second feature set, and the third feature set.

[0010] Preferably, the step of scoring the molecular assembly pathways to be analyzed and determining the optimal path based on the multimodal fusion scoring model and the biological accessibility database includes: using a cross-attention mechanism to dynamically weight multimodal feature fusion of graph neural networks and biological pathway embeddings through a fusion layer based on the reaction template encoder to form a multimodal molecular encoder; performing reaction process optimization training based on the multimodal molecular encoder at both the discrete action level and the continuous action level, and determining the reaction process with the highest biological accessibility; performing enhanced search on the reaction process with the highest biological accessibility based on MCTS, determining the corresponding score of the molecular assembly pathway, and determining the optimal path for molecular assembly based on the score.

[0011] Preferably, the graph neural network is used to convert the input molecular graph into a geometry-aware embedding vector; the reaction template encoder is used to encode enzyme reaction rules into context vectors; the biological pathway embedding is used to characterize the location of the metabolic network involved by the molecules; and the dynamic weighted modal feature fusion formula is: .

[0012] Preferably, the enhanced search specifically includes: determining the initial molecular reaction process and its corresponding initial score based on the MCTS algorithm; selecting an initial reaction path with a high UCB score based on the initial score; determining in real time whether a new reaction path exists, and if so, giving it a new score and determining whether the new score is higher than the high UCB score, and if so, replacing the initial molecular reaction process with the new reaction path; training and analyzing the new reaction path based on a neural network model to predict the optimal path and replacing the initial molecular reaction process; filtering invalid actions in real time based on reaction feasibility, optimizing the scoring process of the molecular reaction process, and replacing the initial molecular reaction process.

[0013] Preferably, the step of performing similarity clustering replacement on the unsynthesized molecules in the optimal path to optimize the synthesis path includes: obtaining unsynthesized molecules; performing molecular similarity analysis on the unsynthesized molecules and existing synthesizable molecules; if the similarity value is ≥ a first threshold, then it is determined to be a first synthesizable molecule; scoring the unsynthesized molecules with similarity values ​​< the threshold using a machine learning model; if the score is ≥ a second threshold, then it is determined to be a second synthesizable molecule, otherwise it is determined to be unsynthesized; replacing the unsynthesized molecules in the optimal path based on the first synthesizable molecules and the second synthesizable molecules, and obtaining the optimized optimal synthesis path.

[0014] Accordingly, the present invention also provides a biomolecule synthesizability assessment device based on reinforcement learning. The device includes: an acquisition unit for acquiring a preset dataset; a merging unit for merging the preset dataset and generating a merged dataset; a correction unit for performing structural verification and repair operations on the merged dataset and generating a corrected dataset; a feature library construction unit for extracting bioaccessibility features from the corrected dataset and generating a bioaccessibility database; a path scoring unit for scoring the synthesis paths of the molecules to be analyzed based on a multimodal fusion scoring model and the bioaccessibility database and determining the optimal path; and a path optimization unit for performing similarity clustering and replacement of unsynthesized molecules in the optimal path to optimize the synthesis path.

[0015] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided in the embodiments of the present invention.

[0016] The present invention has at least the following technical effects through the technical solution provided by the present invention:

[0017] By automating the entire process of "data input → feature library construction → path scoring → path optimization", the evaluation efficiency is greatly improved, and errors caused by human intervention are avoided, thus greatly improving the evaluation accuracy and enabling efficient and accurate evaluation of the syntheticity of biomolecules.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is a schematic flowchart of a biomolecule syntheticity assessment method based on reinforcement learning provided in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of a module of a biomolecule syntheticity assessment device based on reinforcement learning provided in an embodiment of the present invention. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0023] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0024] Existing technologies lack a systematic process for assessing the syntheticity of biomolecules, and relying on manual judgment is inefficient and inaccurate.

[0025] Please see Figure 1 This invention provides a method for assessing the syntheticability of biomolecules based on reinforcement learning. The method includes: acquiring a preset dataset; merging the preset dataset to generate a merged dataset; performing structural verification and repair operations on the merged dataset to generate a corrected dataset; extracting bioaccessibility features from the corrected dataset to generate a bioaccessibility database; scoring the synthesis paths of the molecules to be analyzed based on a multimodal fusion scoring model and the bioaccessibility database and determining the optimal path; and performing similarity clustering and replacement on the unsynthesized molecules in the optimal path to optimize the synthesis path.

[0026] In one possible embodiment, firstly, molecules and reaction pathways from a metabolic model are downloaded from MetaNetX, and corresponding molecular smiles, 3D structures, and other data are downloaded from PubChem, along with data collected from other databases with richer datasets, forming a pre-defined dataset. Secondly, since the data in the pre-defined dataset comes from different databases, the data within the pre-defined dataset is standardized, duplicate molecules are merged, salt ions are removed, and a merged dataset is generated. Next, the molecular structure rationality correction (correction of atomic valence states) and 3D structure optimization of molecules are performed on the merged dataset, generating a revised dataset. Then, multi-dimensional features (biological accessibility features) such as physicochemical and metabolic reactions are extracted from the revised dataset, generating a biological accessibility database. Then, based on a multimodal fusion scoring model and the biological accessibility database, the molecular assemblies to be analyzed are scored to quickly determine the optimal (lowest energy) pathway. Finally, the syntheticity of unsynthesized molecules in the optimal pathway is judged through similarity analysis and machine learning scoring, and they are replaced with synthetic molecules, thereby optimizing the pathway feasibility.

[0027] By automating the entire process of "data input → feature library construction → path scoring → path optimization", the evaluation efficiency is greatly improved, and errors caused by human intervention are avoided, thus greatly improving the accuracy of the evaluation.

[0028] The existing databases are scattered and there is no standard process for data collection, which leads to problems such as incomplete data coverage and disordered format.

[0029] In this embodiment of the invention, obtaining the preset dataset includes: selecting a preset standard database, which includes MetaNetX, PubChem, and other optional databases; obtaining data of the molecular set to be analyzed in the preset standard database; and clustering the obtained data to form a preset dataset.

[0030] In one possible embodiment, MetaNetX, PubChem, and other optional databases are selected as the preset standard databases. Then, compound IDs, reaction equations, EC numbers, and other information are downloaded in batches from MetaNetX, and SMILES, InChIKey, 2D / 3D structures, and other data are downloaded from PubChem via CID. Finally, clustering is performed according to "metabolic pathways" to form a preset dataset, avoiding the problem of data dispersion for the same pathway.

[0031] By standardizing the data acquisition process, we ensured coverage of natural metabolites (MetaNetX) and chemical property data (PubChem), thus improving data integrity. Secondly, by clustering and merging the acquired data, we reduced redundant data and improved storage efficiency.

[0032] In existing technologies, problems such as inconsistent data formats, numerous duplicate records, and salt ion interference lead to significant errors in subsequent analyses.

[0033] In this embodiment of the invention, the step of merging the preset dataset and generating a merged dataset includes: acquiring chemical and structural data of the set of molecules to be analyzed in the preset dataset; performing standardization processing on the chemical and structural data to obtain a standardized dataset; merging identical molecules based on the standardized dataset to form a first dataset; and performing ion processing on the first dataset to form the merged dataset.

[0034] In one possible embodiment, chemical data (compound ID, reaction pathway) and structural data (SMILES, 3D structure) are first extracted from a preset dataset. Then, the acquired chemical and structural data are recorded in a standard manner, such as organizing them into a table of "compound ID-reaction pathway-SMILES-3D structure". The SMILES are then uniformly converted to InChIKey to form a standardized dataset. Next, identical molecules in the standardized dataset are merged through InChIKey mapping to form a first dataset. Finally, the StripSalts function of RDKit is used to remove salt ions from the molecular structures of the first dataset to form a merged dataset.

[0035] By standardizing the data, the data format is made the same, which allows data from different databases to be directly compared, thus improving the overall integration efficiency. Then, by removing duplicate molecules and salt ions, the accuracy of feature extraction is improved.

[0036] In existing technologies, errors in molecular structure (such as abnormal valence state or unoptimized 3D structure) can lead to inaccurate feature extraction and reaction prediction.

[0037] In this embodiment of the invention, the step of performing structural verification and repair operations based on the merged dataset and generating a corrected dataset includes: performing chemical rationality verification based on the merged dataset and obtaining a verification dataset; performing aromatic ring correction based on the verification dataset and obtaining a corrected dataset; performing molecular 3D structure analysis based on the corrected dataset and obtaining an initial energy dataset; and performing energy minimization analysis based on the initial energy dataset and obtaining a corrected dataset.

[0038] In one possible implementation, the number of atomic bonds (e.g., carbon ≤ 4, oxygen ≤ 2) is first checked using the MolSanitize function of RDKit, and errors are automatically corrected to obtain a validation dataset. Then, the aromatic rings in the validation dataset are corrected using the SanitizeMol function (e.g., the non-aromatic ring structure of cyclohexene is corrected to a conjugated aromatic structure of benzene rings), and a corrected dataset is obtained. Next, the initial 3D structure of the molecule is generated using RDKit, and an initial energy dataset is obtained. Finally, the initial 3D structure is subjected to energy minimization analysis based on the MMFF94 / UFF force field to obtain the optimized 3D structure, and a corrected dataset is generated.

[0039] By performing structural verification on the merged dataset, the error rate of atomic valence states was reduced, and by correcting the accuracy of aromatic rings, the accuracy of feature extraction and reaction prediction was improved. Secondly, by optimizing the initial 3D structure, its matching degree with molecular conformations in the actual physiological environment was improved, and the accuracy of reaction site prediction was also improved.

[0040] Existing technologies have relatively limited dimensions for feature extraction, which can easily overlook key biological features such as metabolic pathways and enzyme associations, leading to misjudgments of syntheticity.

[0041] In this embodiment of the invention, the step of extracting biological accessibility features based on the modified dataset and generating a biological accessibility database includes: acquiring the physicochemical properties and structural features of the modified dataset and generating a first feature set; extracting metabolic and reaction features and enzyme correlation features of the modified dataset and generating a second feature set; performing bioavailability prediction features on the modified dataset based on a pre-trained model and generating a third feature set; and constructing the biological accessibility database based on the first feature set, the second feature set, and the third feature set.

[0042] In one possible embodiment, the molecular weight, logP, polar surface area, number of rotatable bonds, and other physicochemical properties, as well as functional group frequencies (number of hydroxyl and carboxyl groups), molecular fingerprints (Morgan fingerprint, MACCS bond, etc.), and other structural features of the corrected dataset are first calculated using rdkit.Chem.Descriptors, generating a first feature set. Then, the number of KEGG pathways involved in the molecule is counted, and reaction center atomic labels are marked using RDTools (metabolic and reaction characteristics). Simultaneously, EC numbers associated with the molecule are extracted, and enzyme family distributions are calculated, generating a second feature set. Next, Caco-2 permeability and hepatotoxicity are predicted using ADMETlab, thereby predicting ADME parameters. Simultaneously, drug-likeness features are predicted using drug-likeness rules, generating a third feature set. These drug-likeness rules include, but are not limited to, the Lipinski five rules (rdkit.Chem.Lipinski) and the Veber rule. Finally, the first, second, and third feature sets are stored in a MySQL database, forming a bioaccessibility database.

[0043] By covering information such as physicochemical properties, metabolism, and enzyme associations with multi-dimensional features, the accuracy of syntheticity assessment is improved; the structured storage of the biological accessibility database supports fast querying and feature reuse, thereby improving the efficiency of subsequent path evaluation.

[0044] Existing technologies rely on exhaustive search for path search, resulting in low path search efficiency; furthermore, insufficient feature fusion fails to capture the synergistic relationship between molecules, enzymes, and pathways, leading to errors in determining the optimal path.

[0045] In this embodiment of the invention, the step of scoring the molecular assembly path to be analyzed and determining the optimal path based on the multimodal fusion scoring model and the biological accessibility database includes: a fusion layer based on a reaction template encoder dynamically weights and fuses multimodal features of a graph neural network and biological pathway embeddings through a cross-attention mechanism to form a multimodal molecular encoder; optimizing and training the reaction process based on the multimodal molecular encoder at both the discrete action level and the continuous action level, and determining the reaction process with the highest biological accessibility; performing enhanced search on the reaction process with the highest biological accessibility based on MCTS, determining the corresponding score of the molecular assembly path, and determining the optimal path for molecular assembly based on the score.

[0046] In this embodiment of the invention, the graph neural network is used to transform the input molecular graph into a geometry-aware embedding vector; the reaction template encoder is used to encode enzyme reaction rules into context vectors; the biological pathway embedding is used to characterize the location of the metabolic network involved by the molecules; and the dynamic weighted modal feature fusion formula is: .

[0047] In one possible embodiment, a molecular graph (with atoms as nodes and molecular bonds as edges) is first input into a graph neural network to learn the shape and structure of the molecule, thereby outputting a geometry-aware embedding vector (including 3D information such as bond angles and bond distances). Next, enzyme reaction rules (such as EC numbers and SMARTS reaction centers) from MetaNetX are encoded into context vectors in a reaction template encoder. Then, a pathway embedding pre-trained using the KEGG pathway database (similar to word2vec) is used in the biological pathway embedding to characterize the location of the molecule-involved metabolic network. Finally, multimodal features are dynamically weighted through a cross-attention mechanism in the fusion layer of the template encoder, where the dynamic weighted modal feature fusion formula is: Then, by selecting reaction templates (such as MetaNetX) or functional group transformations to optimize the reaction process at the discrete action level, and by adaptively adjusting reaction regulation (such as temperature, pH, etc.) or conformational parameters (such as dihedral angles) to optimize the reaction process at the continuous action level, the molecular reaction process with the highest biological accessibility is determined through the above two training optimization processes. Finally, the reaction process with the highest biological accessibility is enhanced searched based on MCTS, and its corresponding score is determined. Finally, the optimal path for assembling molecules is determined based on the score.

[0048] By employing a multimodal fusion model, the feature representation capability is enhanced, thereby improving the accuracy of path scoring. After multimodal fusion, the feature's ability to capture molecular-enzyme-pathway relationships is greatly improved, and the matching degree between path scoring and actual synthesis feasibility is also greatly enhanced. Furthermore, a dynamic weighting mechanism is adopted to give more weight to important features in the evaluation, avoiding interference from irrelevant features. Secondly, the path evaluation time is greatly shortened by reinforcement learning + MCTS search, and high-scoring paths can be dynamically focused.

[0049] Existing technologies use traditional MCTS for path search. MCTS search is highly unpredictable, has a high proportion of invalid path calculations, and cannot utilize historical data to accelerate convergence.

[0050] In this embodiment of the invention, the enhanced search specifically includes: determining the initial molecular reaction process and its corresponding initial score based on the MCTS algorithm; selecting an initial reaction path with a high UCB score based on the initial score; determining in real time whether a new reaction path exists, and if so, giving it a new score and determining whether the new score is higher than the high UCB score, and if so, replacing the new reaction path with the initial molecular reaction process; training and analyzing the new reaction path based on a neural network model to predict the optimal path and replacing the initial molecular reaction process; filtering invalid actions in real time based on reaction feasibility, optimizing the scoring process of the molecular reaction process, and replacing the initial molecular reaction process.

[0051] In one possible embodiment, the initial molecular reaction process and its corresponding initial score are first determined based on the existing MCTS algorithm, and the initial reaction path of the molecule with high UCB is selected. Then, it is determined in real time whether there is a new reaction path. If so, a new score is given, and the relationship between the new score and the initial score is determined. If the new score is higher than the initial score, the new reaction path is replaced with the initial molecular reaction process. Then, the new reaction path is trained and analyzed through a neural network model, and the optimal path is predicted. The optimal path is then replaced with the initial molecular reaction process. At the same time, invalid actions are filtered in real time based on reaction feasibility (such as atom mapping verification), the scoring process of the molecular reaction process is optimized, and the initial molecular reaction process is replaced.

[0052] The above content is illustrated with an example as follows: First, the initial reaction process (such as the 3-step enzymatic reaction of curcumin synthesis) is determined using MCTS, with an initial score of 75 out of 100. Then, a new path (achieving a two-step reaction through directed evolution enzymes) is found in the new path detection, with a score of 88, which is higher than the initial score. The new path is then replaced with the current optimal path. Then, a neural network is trained, and each step of the new path is scored (e.g., the first step scores 95, and the second step scores 85). The high-scoring steps are remembered (the high-scoring threshold is 90), and these steps are prioritized in subsequent searches to reduce redundant calculations. Finally, reactions with mismatched atom mappings are eliminated (e.g., the substrate is "C6H6O" but an enzyme catalyzing "C7H8O" is called), avoiding invalid path evaluations and improving the efficiency of each round of search.

[0053] By filtering out invalid actions, the proportion of invalid calculations is significantly reduced, and the consumption of computing resources is also reduced. Secondly, the accuracy of the neural network in predicting the optimal path is improved, and its search convergence speed is greatly enhanced, thereby meeting the needs of enterprises for rapid evaluation.

[0054] In existing technologies, the determination of the syntheticity of unsynthesized molecules mainly relies on manual methods. However, manual methods are inefficient and inaccurate, which leads to the problem that the optimal path may not be feasible in practice.

[0055] In this embodiment of the invention, the step of performing similarity clustering replacement on the unsynthesized molecules in the optimal synthesis path to optimize the synthesis path includes: obtaining unsynthesized molecules; performing molecular similarity analysis between the unsynthesized molecules and existing synthesizable molecules; if the similarity value is ≥ a first threshold, then it is determined to be a first synthesizable molecule; scoring the unsynthesized molecules with similarity values ​​< the threshold using a machine learning model; if the score is ≥ a second threshold, then it is determined to be a second synthesizable molecule, otherwise it is determined to be unsynthesized; replacing the unsynthesized molecules in the optimal synthesis path based on the first synthesizable molecules and the second synthesizable molecules, and obtaining the optimized optimal synthesis path.

[0056] In one possible embodiment, since mixed enzymes can evolve to perform new reactions (e.g., if an enzyme naturally decarboxylates a substrate, it may also evolve to decarboxylate similar substrates), non-synthetic molecules can be replaced with synthetic molecules. First, the types of non-synthetic molecules are determined by the formula "total molecules - existing synthetic molecules = non-synthetic molecules". Second, the fingerprint similarity of non-synthetic molecules is calculated using the Dice coefficient. Then, the molecular fingerprints of each non-synthetic molecule are obtained, and molecules similar to them (specifically, those with a similarity value ≥ 0.7) are searched in a biological accessibility database as the first synthetic molecules. The remaining dissimilar non-synthetic molecules are scored using a machine learning model, which can be a random forest model (with training data of at least 1000 known synthetic molecule features). Non-synthetic molecules with a score ≥ 0.5 are identified as the second synthetic molecules. Finally, the non-synthetic molecules in the optimal path are replaced with either the first or second synthetic molecules to ensure that all molecules in the path are synthesizable and to obtain the optimized optimal synthesis path.

[0057] By combining similarity analysis and machine learning scoring, the accuracy of determining the syntheticity of unsynthesized molecules is significantly improved. Secondly, by replacing the unsynthesized molecules in the optimal path with the first or second synthetic molecules, the actual feasibility of the optimal path is ensured, reducing the risk of rework after the project starts.

[0058] The following description, in conjunction with the accompanying drawings, illustrates a reinforcement learning-based biomolecule syntheticity assessment device provided by an embodiment of the present invention.

[0059] Please see Figure 2 Based on the same inventive concept, this invention provides a biomolecule synthesizability assessment device based on reinforcement learning. The device includes: an acquisition unit for acquiring a preset dataset; a merging unit for merging the preset dataset and generating a merged dataset; a correction unit for performing structural verification and repair operations on the merged dataset and generating a corrected dataset; a feature library construction unit for extracting bioaccessibility features from the corrected dataset and generating a bioaccessibility database; a path scoring unit for scoring the synthesis paths of the molecules to be analyzed based on a multimodal fusion scoring model and the bioaccessibility database and determining the optimal path; and a path optimization unit for performing similarity clustering and replacement of unsynthesized molecules in the optimal path to optimize the synthesis path.

[0060] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described in the embodiments of the present invention.

[0061] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0062] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0063] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0064] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for assessing the syntheticity of biomolecules based on reinforcement learning, characterized in that, The method includes: Get the preset dataset; The preset dataset is used to perform merging processing and generate a merged dataset; Based on the merged dataset, structural verification and repair operations are performed to generate a corrected dataset; Based on the modified dataset, biological accessibility features are extracted and a biological accessibility database is generated; The optimal path is determined by scoring the molecular aggregates to be analyzed based on the multimodal fusion scoring model and the biological accessibility database. The step of scoring the molecular aggregation pathways to be analyzed and determining the optimal pathway based on the multimodal fusion scoring model and the biological accessibility database includes: The fusion layer based on the reaction template encoder performs dynamic weighted multimodal feature fusion of graph neural network and biological pathway embedding through cross attention mechanism, and forms a multimodal molecular encoder. The graph neural network is used to transform the input molecular graph into a geometry-aware embedding vector; the reaction template encoder is used to encode enzyme reaction rules into context vectors; the biological pathway embedding is used to characterize the location of the metabolic network involved by the molecules; the dynamic weighted multimodal feature fusion formula is: ； in, For geometry-aware embedding vectors For enzyme reaction rule context vector, For pre-trained path embedding vectors; Based on the multimodal molecular encoder, reaction process optimization training is performed at both the discrete action level and the continuous action level, and the reaction process with the highest biological accessibility is determined. Based on MCTS, an enhanced search is performed on the reaction process with the highest biological accessibility, and the corresponding score of the molecular assembly path is determined, and the optimal path for molecular assembly is determined based on the score. The synthesis route is optimized by performing similarity clustering and replacement on molecules that were not synthesized in the optimal route; The step of performing similarity clustering and replacement of unsynthesized molecules in the optimal synthesis path to optimize the synthesis path includes: Obtaining unsynthesized molecules; Molecular similarity analysis is performed between the unsynthesized molecule and the existing synthesizable molecule. If the similarity value is greater than or equal to the first threshold, it is determined to be the first synthesizable molecule. For unsynthesized molecules with a similarity value less than the first threshold, a machine learning model is used to score them. If the score is greater than or equal to the second threshold, they are judged as second synthesizable molecules; otherwise, they are judged as unsynthesizable molecules. Based on the first and second synthesizable molecules, the unsynthesized molecules in the optimal synthesis path are replaced to obtain the optimized optimal synthesis path; The enhanced search specifically includes: The initial molecular reaction process and corresponding initial score are determined based on the MCTS algorithm; Based on the initial score, select the initial response path with the highest UCB score; In real time, it is determined whether a new reaction pathway exists. If so, a new score is given to it, and it is determined whether the new score is higher than the high UCB score. If so, the new reaction pathway is replaced with the initial molecular reaction process. The new reaction pathway is trained and analyzed based on a neural network model to predict the optimal path and replace the initial molecular reaction process. Based on the real-time filtering of invalid actions according to reaction feasibility, the scoring process of molecular reaction process is optimized, and the initial molecular reaction process is replaced.

2. The method according to claim 1, characterized in that, The process of obtaining the preset dataset includes: Select a preset standard database; Obtain the data of the molecular set to be analyzed from the preset standard database; The acquired data is clustered and aggregated to form a preset dataset.

3. The method according to claim 1, characterized in that, The process of merging the dataset based on the preset dataset and generating the merged dataset includes: Obtain the chemical and structural data of the molecular set to be analyzed in the preset dataset; The chemical data and structural data are standardized to obtain a standardized dataset. Based on the standardized dataset, identical molecules are merged to form the first dataset; Ion processing is performed on the first dataset to form the merged dataset.

4. The method according to claim 1, characterized in that, The process of performing structural verification and repair operations based on the merged dataset and generating a corrected dataset includes: Based on the merged dataset, chemical rationality verification was performed, and a verification dataset was obtained; Aromatic ring correction is performed based on the validation dataset to obtain the corrected dataset; Molecular 3D structure analysis was performed based on the calibration dataset, and an initial energy dataset was obtained; Energy minimization analysis is performed based on the initial energy dataset to obtain a corrected dataset.

5. The method according to claim 1, characterized in that, The step of extracting biological accessibility features based on the modified dataset and generating a biological accessibility database includes: Obtain the physicochemical properties and structural features of the corrected dataset, and generate a first feature set; Extract metabolic and reaction features and enzyme correlation features from the corrected dataset, and generate a second feature set; Based on the pre-trained model, bioavailability prediction features are generated from the corrected dataset, and a third feature set is produced. The biological accessibility database is constructed based on the first feature set, the second feature set, and the third feature set.

6. A biomolecule synthesizability assessment device based on reinforcement learning, characterized in that, The device includes: The acquisition unit is used to acquire a preset dataset; A merging unit is used to perform merging processing based on the preset dataset and generate a merged dataset; The correction unit is used to perform structural verification and repair operations based on the merged dataset and generate a corrected dataset. The feature library construction unit is used to extract biological accessibility features based on the modified dataset and generate a biological accessibility database; The path scoring unit is used to score the molecular aggregate paths to be analyzed based on the multimodal fusion scoring model and the biological accessibility database, and to determine the optimal path. Specifically, the path scoring unit is used for: The fusion layer based on the reaction template encoder performs dynamic weighted multimodal feature fusion of graph neural network and biological pathway embedding through cross attention mechanism, and forms a multimodal molecular encoder. The graph neural network is used to transform the input molecular graph into a geometry-aware embedding vector; the reaction template encoder is used to encode enzyme reaction rules into context vectors; the biological pathway embedding is used to characterize the location of the metabolic network involved by the molecules; the dynamic weighted multimodal feature fusion formula is: ； in, For geometry-aware embedding vectors For enzyme reaction rule context vector, For pre-trained path embedding vectors; Based on the multimodal molecular encoder, reaction process optimization training is performed at both the discrete action level and the continuous action level, and the reaction process with the highest biological accessibility is determined. Based on MCTS, an enhanced search is performed on the reaction process with the highest biological accessibility, and the corresponding score of the molecular assembly path is determined, and the optimal path for molecular assembly is determined based on the score. The path optimization unit is used to perform similarity clustering and replacement of unsynthesized molecules in the optimal path to optimize the synthesis path; Specifically, the path optimization unit is used for: Obtaining unsynthesized molecules; Molecular similarity analysis is performed between the unsynthesized molecule and the existing synthesizable molecule. If the similarity value is greater than or equal to the first threshold, it is determined to be the first synthesizable molecule. For unsynthesized molecules with a similarity value less than the first threshold, a machine learning model is used to score them. If the score is greater than or equal to the second threshold, they are judged as second synthesizable molecules; otherwise, they are judged as unsynthesizable molecules. Based on the first and second synthesizable molecules, the unsynthesized molecules in the optimal synthesis path are replaced to obtain the optimized optimal synthesis path; The enhanced search specifically includes: The initial molecular reaction process and corresponding initial score are determined based on the MCTS algorithm; Based on the initial score, select the initial response path with the highest UCB score; In real time, it is determined whether a new reaction pathway exists. If so, a new score is given to it, and it is determined whether the new score is higher than the high UCB score. If so, the new reaction pathway is replaced with the initial molecular reaction process. The new reaction pathway is trained and analyzed based on a neural network model to predict the optimal path and replace the initial molecular reaction process. Based on the real-time filtering of invalid actions according to reaction feasibility, the scoring process of molecular reaction process is optimized, and the initial molecular reaction process is replaced.

Images