Molecular structure prediction system of multimodal fine-grained molecule pre-training model based on prompt learning

Abstract

The invention discloses a molecular structure prediction system of a multimodal fine-grained molecule pre-training model based on prompt learning, and relates to the field of molecular structure prediction, in particular to a molecular structure prediction system. The objective of the invention is to solve the problems of low accuracy and efficiency and low inter-molecular interaction prediction accuracy when complex molecular data is processed due to data scarcity and insufficient task applicability in the existing method. The system comprises a data acquisition module used for acquiring sample data in a multi-modal molecule pre-training data set and sample data in a downstream task data set; the processing module is used for establishing a multi-modal fine-grained molecule pre-training model based on prompt learning and obtaining a trained multi-modal fine-grained molecule pre-training model based on prompt learning; and the prediction module is used for performing attribute and drug interaction relationship prediction on a to-be-detected molecular structure based on the trained prompt learning multi-mode fine-grained molecule pre-training model to obtain a prediction result.

Description

Technical Field

[0001] The present invention relates to the field of molecular structure prediction, and in particular to a molecular structure prediction system. Background Art

[0002] With the continuous advancement of artificial intelligence technology, deep learning has become a key tool to accelerate research and development. Traditional discovery methods are limited by professional skills and experimental conditions and have developed relatively slowly. At present, molecular representation learning methods based on machine learning rely on supervised models, which provide new perspectives for research and development by analyzing features such as molecular fingerprints, SMILES strings, two-dimensional molecular graphs, and three-dimensional structures. However, the scarcity of data and the complexity of annotations limit the development of these models. In order to overcome these limitations, researchers have begun to focus on large-scale pre-trained language models (PTMs) in the field of natural language processing (NLP), such as BERT, GPT series, and T5. These models are unsupervised pre-trained on large-scale text data. Through fine-tuning with a small amount of annotated data, they reduce the dependence on a large amount of annotated data, improve the generalization ability of the model, and enable it to show excellent performance in a variety of tasks.

[0003] Inspired by multimodal models such as CLIP, BLIP2, and LLaVA, researchers use self-supervised learning methods to explore the intrinsic connections between molecules of different modalities. Some studies regard SMILES strings as languages ​​with special grammars, and use masked language model tasks to understand their relationship with molecular structures. Models like the T5 backbone use the attention mechanism of Transformer to learn the serialized representation of molecular structures. This processing method makes the expression of molecular structures more concise and flexible, but the SMILES string representation has limitations, especially the difficulty in expressing the structural space information inside the molecule. In order to fully capture the structural characteristics of molecules, researchers consider representing molecules in other forms. The topological structure of molecules provides an intuitive representation of spatial arrangement, which helps to understand the intrinsic connections and properties of molecules. The researchers simulated the molecular topology by converting SMILES strings into two-dimensional molecular graphs, and used graph neural networks to aggregate and propagate information between atoms and chemical bonds to learn the complex characteristics of molecules. Although these methods have made progress in molecular modeling, they may destroy the integrity of the molecular topology when reconstructing the internal structure of the mask, resulting in the inability to fully capture the unique structural characteristics of the molecule, which in turn affects the accuracy of the model's prediction of the properties and functions of the molecule.

[0004] Multimodal fine-grained molecular pre-training models have made significant progress in molecular property prediction, molecule-to-text generation tasks, and molecular optimization. However, the potential of these models in predicting molecular interactions has not been fully explored. Currently, fine-tuning pre-trained language models through transfer learning has become a common practice to improve the performance of models on specific tasks. Although this strategy has achieved significant performance improvements in natural language processing tasks, as the size of pre-trained models grows, the traditional full-model micro-computation and storage costs are becoming increasingly high. Recently, parameter-efficient fine-tuning methods have been proposed to alleviate this problem by only updating a small part of the model parameters. While maintaining similar performance, they exhibit the advantages of modular design, strong adaptability, and prevention of catastrophic forgetting. Hint adjustment has become a promising method to reduce the amount of parameter adjustment by introducing a trainable hint vector in the input. However, improper initialization will result in the model being unable to effectively utilize the knowledge acquired by the model during the pre-training phase, which in turn affects the model's performance on specific tasks. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of low accuracy and efficiency in processing complex molecular data and low accuracy in predicting intermolecular interactions caused by data scarcity and insufficient task applicability of existing methods, and to propose a molecular structure prediction system based on a multimodal fine-grained molecular pre-training model based on prompt learning.

[0006] The prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning includes: a data acquisition module, a processing module, and a prediction module;

[0007] The data acquisition module is used to acquire sample data in the multimodal molecular pre-training data set and sample data in the downstream task data set;

[0008] The processing module is used to establish a multimodal fine-grained molecular pre-training model based on prompt learning, and obtain the trained multimodal fine-grained molecular pre-training model based on prompt learning;

[0009] The prediction module is used to predict the properties and drug interaction relationship of the molecular structure to be tested based on the multimodal fine-grained molecular pre-training model of trained prompt learning to obtain a prediction result.

[0010] The beneficial effects of the present invention are:

[0011] The present invention proposes a multimodal fine-grained molecular pre-training model MolFinePrompt based on prompt learning, which takes into account the integrity of molecular structures when learning complex structures, accurately predicts the interactions between molecules, and gives full play to the potential of instruction fine-tuning in different tasks to improve the accuracy and efficiency of drug discovery.

[0012] The present invention uses a contrastive learning strategy on a data set containing 316k pairs of molecular texts to effectively align molecular features with text information; by integrating substructures that reflect chemical properties into molecular topology, the multimodal fine-grained molecular pre-training model based on prompt learning improves the accuracy and efficiency of processing complex molecular data;

[0013] In order to enhance the adaptability of the multimodal fine-grained molecular pre-training model based on prompt learning to diversified downstream tasks, the present invention designs a corresponding prompt template for each task, guides the multimodal fine-grained molecular pre-training model based on prompt learning to more effectively identify and utilize task-related features, and enhances the generalization and adaptability of the multimodal fine-grained molecular pre-training model based on prompt learning in diversified tasks.

[0014] By deeply analyzing the interaction patterns between different drug molecules, the multimodal fine-grained molecular pre-training model MolFinePrompt based on prompt learning can more accurately understand the properties of molecules and potential synergistic effects, providing a basis for the development of drug combination therapy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a structural diagram of the system of the present invention. DETAILED DESCRIPTION

[0016] Specific implementation method 1: In this implementation method, the molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning includes: a data acquisition module, a processing module, and a prediction module;

[0017] The data acquisition module is used to acquire sample data in the multimodal molecular pre-training data set and sample data in the downstream task data set;

[0018] The processing module is used to establish a prompt-based learning multimodal fine-grained molecular pre-training model MolFinePrompt, and obtain the trained prompt-based learning multimodal fine-grained molecular pre-training model;

[0019] The MolFinePrompt is Fine-Grained Multimodal Molecular PretrainingLarge Model via Prompt-Learning;

[0020] The prediction module is used to predict the properties and drug interaction relationship of the molecular structure to be tested based on the multimodal fine-grained molecular pre-training model of trained prompt learning to obtain a prediction result.

[0021] The molecular structure to be tested and the data acquired by the data acquisition module come from the same field (the same data type means that there are molecular text pairs, such as: molecular structure CC(=O)O and the corresponding text description: Acetic acid is a product of the oxidation of ethanol and of the destructive distillation of wood. It is used locally, occasionally internally, as a counterirritant and also as areagent. Acetic acid otic (for the ear) is an antibiotic that treats infections caused by bacteria or fungus. (Acetic acid is a product of the oxidation of ethanol and the destructive distillation of wood. It is used locally, occasionally internally, as a counterirritant and reagent. Acetic acid otic (for the ear) is an antibiotic that treats infections caused by bacteria or fungus.)

[0022] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that: the data acquisition module is used to acquire sample data in the multimodal molecular pre-training data set and sample data in the downstream task data set; the specific process is:

[0023] The sample data in the multimodal molecular pre-training dataset are molecule-text pairs;

[0024] The sample data in the downstream task dataset is the molecular structure (only the structure without text, the text is a prompt, and the molecular structure uses the trained GIN);

[0025] The sample data in the dataset and the sample data in the downstream task dataset both include molecular structure G and text description T;

[0026] The molecular structure G and text description T are 316k molecule-text pairs collected from the public dataset Pubchem, where k is in thousand;

[0027] For example: the fine-grained molecular structure diagram G′ and the standardized molecular text description T′ in the pre-training dataset are used as the input of the MolFinePrompt model, and the fine-grained molecular structure diagram G' and the specific task instruction text in the downstream task dataset are used as the input of the pre-training model, and the model outputs the result of the prediction task.

[0028] The other steps and parameters are the same as those in the first embodiment.

[0029] Specific implementation method three: This implementation method is different from specific implementation method one or two in that: the processing module is used to establish a multimodal fine-grained molecular pre-training model based on prompt learning, and obtain the trained multimodal fine-grained molecular pre-training model based on prompt learning; the specific process is:

[0030] The multimodal fine-grained molecular pre-training model based on prompt learning includes: a fine-grained molecular graph construction module, a molecular and text description representation learning module, a self-supervised collaborative comparison optimization module, and an instruction fine-tuning downstream task module;

[0031] The fine-grained molecular graph construction module uses a specific decomposition rule to decompose the molecular structure graph G in the sample data in the multimodal molecular pre-training data set into sub-structure units, and constructs a fine-grained molecular graph G' based on the sub-structure units;

[0032] The molecular and text description representation learning module is used to input the text description T in the sample data in the multimodal molecular pre-training data set into the text encoder, and the text encoder outputs semantic features; input the fine-grained molecular graph G″ corresponding to 1.65 million molecular structures in the PubChem data set into the molecular encoder, and the molecular encoder outputs a molecular feature vector; train the molecular encoder GIN to obtain a pre-trained molecular encoder GIN;

[0033] The self-supervised collaborative contrast optimization module uses a contrastive learning method to train the pre-trained molecular encoder GIN and text encoder to obtain the trained molecular encoder GIN and text encoder; the trained molecular encoder GIN and text encoder constitute a multimodal fine-grained molecular pre-training model; improve the model's ability to understand and represent molecular multimodal data;

[0034] The instruction fine-tuning downstream task module is used to use prompt learning to guide the multimodal fine-grained molecular pre-training model to understand downstream tasks, and obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning.

[0035] Improve its applicability and flexibility in specific application scenarios.

[0036] Based on the above, the present invention constructs a fine-grained molecular structure diagram, combines text information, uses contrastive learning to optimize the multimodal fine-grained molecular pre-training model, and uses prompt learning technology to apply the multimodal fine-grained molecular pre-training model to key specific tasks in the discovery field.

[0037] The other steps and parameters are the same as those in the first or second embodiment.

[0038] Specific implementation method 4: This implementation method is different from any one of specific implementation methods 1 to 3 in that: the fine-grained molecular graph construction module uses a specific decomposition rule to decompose the molecular structure graph G in the sample data in the multimodal molecular pre-training data set into sub-structure units, and constructs a fine-grained molecular graph G' based on the sub-structure units;

[0039] The specific process is:

[0040] 1) Based on the RDKit tool, obtain the chemical element characteristics of each atom in the molecular structure represented by the SMILES string in the sample data in the multimodal molecular pre-training dataset;

[0041] The chemical element characteristics of each atom include the atomic number, connectivity, bond type between atoms (single bond, double bond, triple bond) and atomic ring type (whether ring-forming);

[0042] Use connectivity features as node features;

[0043] The bond type between atoms and the type of atoms forming a ring are used as bond characteristics;

[0044] The connectivity is the number of chemical bonds connecting atoms;

[0045] 2) In order to more comprehensively and completely describe the chemical properties of the molecular structure, the molecular structure represented by the SMILES string in the sample data in the multimodal molecular pre-training dataset is converted into a 2D topological graph G = (V, E) based on the RDKit tool;

[0046] Wherein, V represents the node set of the 2D topological graph G, and E represents the edge set of the 2D topological graph G;

[0047] V represents the corresponding atom in the molecule, and E is the chemical bond between atoms;

[0048] 3) Based on the BRICS decomposition method, the molecular structure in the sample data in the multimodal molecular pre-training data set is decomposed to obtain substructure units; the specific process is as follows:

[0049] Based on the BRICS decomposition method, the molecular structure in the sample data in the multimodal molecular pre-training data set is preliminarily decomposed to obtain multiple substructure units after the preliminary decomposition;

[0050] Decompose each substructure unit after the preliminary decomposition to obtain the smallest multiple substructure units (if it can be decomposed, perform this step; if it cannot be decomposed, take the multiple substructure units after the preliminary decomposition as the smallest multiple substructure units); for example, the SMIELS structure of the molecule is represented by C1=CC=C2C(=C1)N(C(=O)N2CCO)CCO, which can be decomposed into three substructure units of CCO, C1=CC=C2C(=C1)N(C(=O)N2), and CCO by the BRICS decomposition method. According to the decomposition rule, C1=CC=C2C(=C1)N(C(=O)N2) is further decomposed into C1=CC=C2C(=C1), C=O, C1N=CN=C1, and finally the molecule has five substructure units: CCO, C1=CC=C2C(=C1), C=O, C1N=CN=C1, and CCO;

[0051] 4) Use the substructure unit as the new node V f Add to the topology map;

[0052] The connection relationship E between each substructure unit and the nodes it contains f Added to the topology graph as a new edge;

[0053] Construct an empty graph node V g , the graph-level node V g With all new nodes V f Connect to get the connection relationship E g , based on V,V f ,V g ,E,E f ,E g Form a fine-grained molecular graph G' to mine the deep semantic information hidden in the molecules;

[0054] G′=(V′,E′),V′=[V,V f ,V g ],E′=[E,E f ,E g ].

[0055] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0056] Specific implementation method five: This implementation method is different from any one of specific implementation methods one to four in that: the molecular and text description representation learning module is used to input the text description T in the sample data in the multimodal molecular pre-training data set into the text encoder, and the text encoder outputs semantic features; the fine-grained molecular graph G″ corresponding to 1.65 million molecular structures in the PubChem data set is input into the molecular encoder, and the molecular encoder outputs a molecular feature vector; the molecular encoder GIN is trained to obtain a pre-trained molecular encoder GIN;

[0057] The specific process is:

[0058] Step 1), changing the molecule name in the text of the molecule text pair in the sample data in the multimodal molecule pre-training data set to a unified format of "The molecule is";

[0059] The text in the unified format of "The molecule is" is input into the text encoder, and the output semantic features of the text encoder are;

[0060] The text encoder is SciBERT based on the BERT architecture;

[0061] As supplementary knowledge to the molecular graph, the text description is summarized in natural language from the perspectives of molecular-related functions, properties, etc. In order to prevent the leakage of the collected molecular text data and eliminate the bias of molecular names, we changed the text names in the 316k molecule-text pairs to a unified format of "The molecule is" to enhance the generalization and interpretability of the model, so that it can focus on the intrinsic connection between molecular structure and properties instead of relying on specific molecular names.

[0062] SciBERT provides richer and more accurate semantic domain knowledge for molecules by pre-training on a large number of corpora in fields such as biochemistry and medicine.

[0063] Step 2), the fine-grained molecular graph G″ corresponding to the 1.65 million molecular structures in the PubChem dataset and the node features and bond features corresponding to the 1.65 million molecular structures in the PubChem dataset are input into the molecular encoder GIN, and the molecular encoder GIN outputs a feature vector;

[0064] The molecular encoder GIN includes: input layer, hidden layer, hidden layer, hidden layer, hidden layer;

[0065] The molecular encoder GIN is optimized by using the self-supervised learning method of generation task and prediction task until convergence, and a pre-trained molecular encoder GIN is obtained;

[0066] Use the feature vector corresponding to V and E to complete the generation task;

[0067] Using V g ,E g The corresponding feature vector completes the prediction task;

[0068] The generated tasks are the connectivity of atoms, the type corresponding to the atomic number (carbon, hydrogen and oxygen belong to different atomic types) and the bond type between atoms;

[0069] The prediction tasks are the number of atoms in the molecular structure (how many atoms are there in a molecule) and the number of bonds (how many chemical bonds are there in a molecule);

[0070] The first layer of the molecular encoder GIN receives the original node features as input, and the second to fourth layers are hidden layers. Each layer takes the output of the previous layer as input, aggregates the features of neighboring nodes, updates the node representation through a learnable multi-layer perceptron (MLP), and adds residual connections (adding the node representation of the current layer to the node representation of the previous layer) as the output of each layer;

[0071] For molecular structure processing, we adopted a Graph Isomorphism Network (GIN) to encode the topological structure, which is a variant of Graph Neural Network (GNN) widely used in the field of molecular representation learning. GIN stands out for its excellent ability to capture and express the topological features of molecules.

[0072] Based on the RDKit tool, 1.65 million molecular structures in the PubChem dataset were processed to obtain node features and bond features. The node features and bond features were input into a five-layer GIN model, and the GIN model output feature vectors. Two types of self-supervised learning tasks, generation tasks and prediction tasks, were used to optimize the molecular encoder GIN until convergence, and a pre-trained molecular encoder GIN model was obtained.

[0073] Our model is able to learn and generalize the structural information of molecules more accurately.

[0074] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0075] Specific implementation method 6: This implementation method is different from any one of specific implementation methods 1 to 5 in that: the self-supervised collaborative contrast optimization module uses a contrast learning method to train the pre-trained molecular encoder GIN and text encoder to obtain the trained molecular encoder GIN and text encoder; the trained molecular encoder GIN and text encoder constitute a multimodal fine-grained molecular pre-training model; and the model's ability to understand and represent molecular multimodal data is improved;

[0076] The specific process is:

[0077] The training set (molecule-text pairs) is the sample data in the multimodal molecule pre-training dataset;

[0078] The loss function is the InfoNCE loss function;

[0079] The optimization method is contrastive learning method;

[0080] The molecular feature vector V output by the pre-trained molecular encoder GIN g ,E g The corresponding eigenvector is represented as the eigenvector of the numerator.

[0081] The text encoder and the pre-trained molecular encoder GIN are trained until convergence to obtain the trained text encoder and molecular encoder GIN;

[0082] The trained text encoder and molecular encoder GIN constitute a multimodal fine-grained molecular pre-training model to achieve the best alignment between molecular structure features and semantic features;

[0083] The InfoNCE loss function is:

[0084]

[0085] in, Input the pre-trained molecular encoder GIN to the molecules in the i-th pair of samples, and the image feature vector output by the pre-trained molecular encoder GIN;

[0086] The text feature vector output by the text encoder is the text input to the i-th pair of samples;

[0087] Input the text encoder for the text in the jth sample, and output the text feature vector of the text encoder;

[0088] sim() is a function to measure the similarity between two modal features, τ is a temperature parameter, and N is the total number of molecule-text pairs.

[0089] The multimodal contrastive learning method is used to enhance the semantic relevance between molecules and texts. This method improves the matching quality of molecules and texts by continuously optimizing the model, while ensuring that mismatched molecule-text pairs maintain a certain distance in the embedding space. n ,t n ), where g i ,t i are the text descriptions corresponding to the ith molecule and the ith molecule, respectively;

[0090] Verify the multimodal fine-grained molecular pre-training model:

[0091] Zero-shot retrieval task for new molecular structures / texts based on multimodal fine-grained molecular pre-training model; the specific process is as follows:

[0092] Obtain new molecular structure text pair datasets PCDes and MoMu (new molecular structure text pairs);

[0093] Freeze the zero-shot molecular structure / text retrieval task based on the multimodal fine-grained molecular pre-training model;

[0094] The zero-shot molecular structure retrieval task inputs the molecular structure and text of the new molecular structure text pair into a multimodal fine-grained molecular pre-training model to obtain the molecular structure feature h g and semantic features h t , calculate the molecular structure feature h g and the semantic feature vector h t The cosine similarity of is obtained to obtain a cosine similarity matrix, in which each element represents the similarity between a molecular structure feature vector and all semantic feature vectors;

[0095] Determine the index of the text representation that is most similar to each molecular structure representation, and compare it with the index of the correct text representation (the first text corresponding to the first molecular structure, the second text corresponding to the second molecular structure). If the indexes match, the retrieval is considered successful, and the retrieval accuracy is calculated;

[0096] For each molecular structure feature vector, the semantic feature vectors are sorted according to the similarity value and the most similar items are retrieved; if the molecular structure feature vector is in the top 20 most similar items after sorting, the retrieval is successful. The proportion of correct retrieval is calculated as the average recall rate.

[0097] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0098] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that: the instruction fine-tuning downstream task module is used to use prompt learning to guide the multimodal fine-grained molecular pre-training model to understand the downstream task, and obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning;

[0099] The specific process is:

[0100] In order to improve the performance of the model on specific molecular prediction tasks, this project uses tips and parameters guided by professional knowledge to efficiently fine-tune the model to adapt to specific molecular prediction tasks. Specifically, the initialization prompts are first customized for different tasks so that the model can understand the task goals and requirements. By designing special prompts for attribute prediction and drug-drug interaction (DDI) tasks, rich background and target information is provided to help the model understand the relevant biological and chemical principles, ensuring that the model can recognize and utilize key molecular features.

[0101] 1) Build the instruction prompt text; the specific process is:

[0102] 11) Obtain attribute prediction tasks in the training set (structure only without text, text is prompt, and the molecular structure uses the trained GIN) from the sample data in the downstream task dataset;

[0103] The attribute prediction task includes six subtasks: BBBP, Bace, Sider, Tox21, ToxCast, and HIV. Each subtask corresponds to a data set, that is, the attribute prediction task includes six data sets;

[0104] The BBBP is blood-brain barrier penetrating;

[0105] The Bace is a β-amyloid precursor protein cleaving enzyme;

[0106] The Sider is a side effect;

[0107] The Tox21 is the 21st Century Toxicology Test;

[0108] The ToxCast is a multi-year toxicology prediction research project initiated by the U.S. Environmental Protection Agency (EPA);

[0109] The HIV is human immunodeficiency virus;

[0110] The BBBP dataset contains 2039 molecular structures and corresponding attribute labels;

[0111] The Bace dataset contains 1513 molecular structures and corresponding attribute labels;

[0112] The Sider dataset contains 1427 molecular structures and corresponding attribute labels;

[0113] The Tox21 dataset contains 7831 molecular structures and corresponding attribute labels;

[0114] The ToxCast dataset contains 8576 molecular structures and corresponding attribute labels;

[0115] The HIV dataset contains 41,127 molecular structures and corresponding attribute labels;

[0116] 12) Obtain the drug interaction task in the training set (only structure without text, text is prompt, and the molecular structure uses the trained GIN) in the sample data in the downstream task dataset;

[0117] The drug interaction task includes three subtasks: ZhangDDI, ChChMiner, and DeepDDI. Each subtask corresponds to a data set, that is, the drug interaction task includes three data sets.

[0118] The ZhangDDI dataset contains 48,548 molecular structure pairs (two molecular structures) and corresponding drug interaction relationship labels;

[0119] The ChChMiner dataset contains 48,514 molecular structure pairs (two molecular structures) and corresponding drug interaction relationship labels;

[0120] The DeepDDI dataset contains 192,284 molecular structure pairs (two molecular structures) and corresponding drug interaction relationship labels;

[0121] 13) Construct instruction prompt text according to the task objectives, task background, guiding principles and task requirements of each subtask;

[0122] Take the BBBP task as an example:

[0123] For example, the prompt for the BBBP (blood-brain barrier penetrability) task: "ThistaskisBBBP, our objectiveistopredict whether adrugmoleculecanpenetratetheblood-brainbarrier,whichiscomposedofbrain capillaryendothelialcells.Theblood-brainbarrierishighlyselective,allowingonlycertain substancestopassthrough.Weneedtoanalyzethefollowingmolecularstructural characteristics:lipophilicity,molecularweight,chargestate,proteinbindingcapacity,presenceofhydrophobicgroups,activityofmetabolicproducts,andthecyclicstructureofthemolecule. ismorelikelytopenetratetheblood-brainbarrier.Pleaseusetheseguidingprinciplesto determine whether this molecule has the capability to do so. (This task is a BBBP task. We have designed the task goal, task background, guiding principles and task requirements; Task goal: Predict whether a drug molecule can penetrate the blood-brain barrier; Task background: The blood-brain barrier is composed of brain capillary endothelial cells. The blood-brain barrier is highly selective and only allows certain substances to pass through; Guiding principles: If a molecule has high lipophilicity, small molecular weight, lacks charge, has little protein binding, has hydrophobic groups, has no active metabolites, and is a polycyclic compound, it is easier to penetrate the blood-brain barrier; Task requirements: Please use these guiding principles to determine whether the molecule has the ability to penetrate the blood-brain barrier.)

[0124] Bace (β-amyloid precursor protein cleavage enzyme) task prompt: ThistaskisBACE,ourgoalistopredict whetheradrugmoleculecaneffectivelyinhibitBACEenzyme,whichisakeytargetforthe treatment ofAlzheimer'sdisease.TheactivityofBACEenzymeiscloselyrelatedtothe productionofbeta-amyloidprotein.Weneedtoanalyzethefollowingmolecularstructural characteristics:bindingaffinity,molecularweight,lipophilicity,hydrogenbondcapability,stereoisomerism,andmetabolicstability.Ifthemoleculedemonstratesstrongbindingaffinity withtheactivesiteofBACEenzyme,possesskeychemicalgroupstoformnecessaryhydrogenbonds,andisstableunderphysiologicalconditionswithgoodcellmembrane permeability,itmaybeaneffectiveBACEinhibitor.Pleaseusetheseguidingprinciplesto This task is the BACE task. We have designed the task goal, task background, guiding principles and task requirements. Task goal: predict whether a drug molecule can effectively inhibit the BACE enzyme, which is a key target for the treatment of Alzheimer's disease. Task background: The activity of the BACE enzyme is closely related to the production of β-amyloid protein. Guiding principles: We need to analyze the following molecular structure characteristics: binding affinity, molecular weight, lipophilicity, hydrogen bonding ability, stereoisomerism and metabolic stability.If a molecule exhibits strong binding affinity to the active site of the BACE enzyme, possesses key chemical groups that form necessary hydrogen bonds, and is stable under physiological conditions and has good cell membrane permeability, it may be an effective BACE inhibitor. Task requirements: Please use these guidelines to analyze and determine whether the molecule has the potential for BACE inhibition);.

[0125] Sider (side effects) task prompt: ThistaskisSIDE, ourgoalistopredictthepotentialsideeffectsthatadrugmoleculemightcause.Predictingdrugsideeffectsiscrucialforthesafetyassessmentofmedications.Weneedtoanalyzethefollowingmolecularstructuralcharacteristics:pharmacologicalmechanismsofactionsuchasreceptorbindingorenzymeinhibition,propertieslikemolecularweight,lipophilicity,andsolubility,andpharmacokinetic characteristicssuchasabsorption,distribution,metabolism,andexcretion.Pleaseusetheseguidelinestoanalyzeandpredictthepotentialsideeffectsthatthisdrugmoleculemay cause.(This task is SIDE. We have designed the task goal, task background, guidelines and task requirements; Task goal: Predict the potential side effects that drug molecules may cause. Task background: Predicting drug side effects is crucial for drug safety assessment. Guidelines: We need to analyze the following molecular structure characteristics: pharmacological mechanisms such as receptor binding or enzyme inhibition, properties such as molecular weight, lipophilicity and solubility, and pharmacokinetic characteristics such as absorption, distribution, metabolism and excretion. Task requirements: Please use these guidelines to analyze and predict the potential side effects that this drug molecule may cause.)

[0126] Tox21 (21st Century Toxicology Testing) Task Tip: ThistaskisTOX21,ourgoalistopredictwhethera pharmaceuticalmoleculemightinducearangeofbiologicaleffectsassociatedwithtoxicity.TheTOX21projectaimstoidentifythepotentialtoxicityofchemicalsubstances,including hormonedisruption,genotoxicity,andmore.Weneedtoanalyzemolecularstructural characteristics: the presence of aromatic rings, the balance between lipophilicity and hydrophilicity, metabolicstability, and permeability. If a molecule affects the pharmaceutical action of known toxicity-related targets, shows metabolic instability, or is likely to produce toxicmetabolites,or exhibits poor cellular permeability,then may have TOX21 properties. Please use these guiding principles to predict whether the pharmaceutical molecule has TOX21 characteristics. (This task is TOX21, and our goal is to predict whether a drug molecule will induce a range of toxicity-related biological effects. The TOX21 project aims to determine the potential toxicity of chemical substances, including hormone disruption, genotoxicity, etc. We need to analyze molecular structural characteristics: the presence of aromatic rings, the balance between lipophilicity and hydrophilicity, metabolic stability and permeability. If a molecule affects the pharmacology of a known toxicity-related target, exhibits metabolic instability, or may produce toxic metabolites, or exhibits poor cellular permeability, it may have TOX21 properties. Please use these guiding principles to predict whether a drug molecule has TOX21 characteristics.)

[0127] ToxCast (a multi-year toxicology prediction research project started by the U.S. Environmental Protection Agency (EPA)) task prompt: Thistaskis ToxCast,ourgoalistoassessthepotentialtoxicityofchemicalsubstancestobiologicalsystems.TheToxCastprojectisalarge-scale,high-throughputscreeningprojectthatusesavarietyofbiologicalteststopredictthetoxicityofchemicalsubstances.Weneedtoanalyzethefollowing molecularstructuralcharacteristics:molecularweight,solubility,lipophilicity,etc.Ifamoleculehasalargemolecularweight,lowsolubility,highlipophilicity,andactivemetabolites,itis morelikelytoexhibitspecifictoxicity.Pleasejudgewhetherthismoleculehaspotentialtoxicity based on these guidelines. (This task is ToxCast. We designed the task goal, task background, guidelines and task requirements; Task goal: Evaluate the potential toxicity of chemicals to biological systems. Task background: The ToxCast project is a large-scale, high-throughput screening project that uses various biological tests to predict the toxicity of chemicals. Guidelines: We need to analyze the following molecular structure features: molecular weight, solubility, lipophilicity, etc. If the molecule has a large molecular weight, low solubility, high lipophilicity, and strong metabolite activity, it is more likely to exhibit specific toxicity. Task requirements: Please use these guidelines to determine whether the molecule is potentially toxic.)

[0128] HIV (human immunodeficiency virus) task prompt: ThistaskisHIV,ourgoalistoassessthepotential inhibitoryeffectofdrugmoleculesonHIV.HIVisavirusthatattacksthehumanimmune system, particularly affecting CD4+T cells, leading to Acquired Immunodeficiency Syndrome (AIDS). Weneed to analyze the following molecular structural characteristics: lipophilicity, molecular weight, chargestate, protein binding capacity, hydrophobic groups, and the activity of metabolic products. If the molecule has an optimized chemical structure targeting HIV'skey proteins,asmallermolecularweight,strongproteinbindingcapacity,hydrophobicgroups,andgoodmetabolicstability,thenitismorelikelytohaveaninhibitoryeffectonHIV.Pleasejudge whether this molecule has antiviral potential of specific drug molecules. (This task is HIV. We have designed the task goal, task background, guidelines and task requirements; Task goal: Evaluate the potential inhibitory effect of drug molecules on HIV. Task background: HIV is a virus that attacks the human immune system, especially affecting CD4+T cells, leading to acquired immunodeficiency syndrome (AIDS). Guidelines: We need to analyze the following molecular structure characteristics: lipophilicity, molecular weight, charge state, protein binding ability, hydrophobic groups and activity of metabolites. If a molecule has an optimized chemical structure for HIV key proteins, a small molecular weight, strong protein binding ability, hydrophobic groups, and good metabolic stability, then it is more likely to have an inhibitory effect on HIV. Task requirements: Please use these guidelines to predict the antiviral potential of a specific drug molecule.)

[0129] Task prompts for the drug interaction task: This task is Drug-Drug Interaction, which refers to the phenomenon where the simultaneous use of two or more drugs in the body can lead to an enhancement or reduction in their efficacy, or even cause adverse reactions, due to the mutual influence between the drugs. Determine if the interaction between these two drugs is positive or not. (This task is drug interaction. We have designed the task objectives, task background, guiding principles and task requirements; Task objectives: refers to the phenomenon that due to the mutual influence between drugs, the simultaneous use of two or more drugs in the body can lead to an enhancement or reduction in their efficacy, or even cause adverse reactions. Task requirements: Please determine whether there is an interaction between the two drugs.); Drug interactions are different types of drugs. It is just necessary to determine whether there is an interaction between each pair of drugs.

[0130] 2) Based on the instruction prompt text constructed in 1), obtain the feature vector h of the prompt text p ;

[0131] 3) Based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text, the attribute prediction of the molecular structure (without attribute label) is performed;

[0132] 4) Based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text, the drug interaction relationship is predicted for molecular structure pairs (without interaction relationship);

[0133] 5) Repeat 1) to 4) until convergence to obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning.

[0134] The other steps and parameters are the same as those in Specific Implementations 1 to 6-1.

[0135] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that: in 2), based on the instruction prompt text constructed in 1), the feature vector h of the prompt text is obtained. p ; The specific process is:

[0136] Create a hint embedding layer P∈R (l×d), where l is the number of virtual tokens, d is the dimension of embedding; R is a real number;

[0137] Input the instruction prompt text constructed in 1) into the word segmenter corresponding to the trained text encoder, and the word segmenter outputs a token ID sequence;

[0138] The token ID sequence is input into the trained text encoder, and the word embedding layer in the trained text encoder outputs the word embedding vector;

[0139] The word embedding vector is used as the initial weight of the created prompt embedding layer P to obtain the initialized prompt embedding layer;

[0140] The virtual token in the initialized prompt embedding layer is input into the trained text encoder, and the trained text encoder outputs the feature vector h corresponding to the instruction prompt text p .

[0141] Each subtask corresponds to many molecular structures, and each molecular structure corresponds to an h p , h in 1 subtask p are the same.

[0142] The other steps and parameters are the same as those in Specific Implementation 1 to 7-1.

[0143] Specific implementation method 9: This implementation method is different from any one of the specific implementation methods 1 to 8 in that: in the above 3), based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text, the molecular structure (without attribute labels) is predicted; the specific process is:

[0144] 31) Input the molecular structure of BBBP (blood-brain barrier permeability) (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g1 ;

[0145] Based on the instruction prompt text constructed by BBBP in 1), the feature vector h of the prompt text is obtained according to 2). p1 ;

[0146] The feature vector h of the prompt text p1 and the structural feature vector h g1 After fusion, the attribute feature representation corresponding to the molecular structure of BBBP after prompt guidance is obtained;

[0147]

[0148] in,

[0149] h property1It is the attribute feature representation corresponding to the molecular structure of BBBP after prompt guidance;

[0150] α is the weight parameter, For splicing;

[0151] Input the attribute feature representation corresponding to the molecular structure of BBBP after prompt guidance into the classifier for classification (obtaining attributes);

[0152] 32) Input the molecular structure of Bace (β-amyloid precursor protein cleaving enzyme) (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g2 ;

[0153] Based on the instruction prompt text constructed by Bace in 1), the feature vector h of the prompt text is obtained according to 2). p2 ;

[0154] The feature vector h of the prompt text p2 and the structural feature vector h g2 Fusion, to obtain the attribute feature representation corresponding to the molecular structure of Bace guided by the prompt;

[0155]

[0156] in,

[0157] h property2 It is the attribute feature representation corresponding to the molecular structure of Bace after prompt guidance;

[0158] α is the weight parameter, For splicing;

[0159] Input the attribute feature representation corresponding to the molecular structure of Bace after prompt guidance into the classifier for classification (obtaining attributes);

[0160] 33) Input the molecular structure of Sider (side effect) (without attribute label) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g3 ;

[0161] Based on the instruction prompt text constructed by Sider in 1), the feature vector h of the prompt text is obtained according to 2). p3 ;

[0162] The feature vector h of the prompt text p3 and the structural feature vector h g3 Fusion, to obtain the attribute feature representation corresponding to the molecular structure of Sider guided by the prompt;

[0163]

[0164] in,

[0165] h property3 It is the attribute feature representation corresponding to the molecular structure of Sider after prompt guidance;

[0166] α is the weight parameter, For splicing;

[0167] Input the attribute feature representation corresponding to the molecular structure of Sider after prompt guidance into the classifier for classification (obtaining attributes);

[0168] 34) Input the molecular structure of Tox21 (21st Century Toxicology Test) (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g4 ;

[0169] Based on the instruction prompt text constructed by Tox21 in 1), the feature vector h of the prompt text is obtained according to 2). p4 ;

[0170] The feature vector h of the prompt text p4 and the structural feature vector h g4 After fusion, the attribute feature representation corresponding to the molecular structure of Tox21 after prompt guidance is obtained;

[0171]

[0172] in,

[0173] h property4 It is the property feature representation corresponding to the molecular structure of Tox21 after prompt guidance;

[0174] α is the weight parameter, For splicing;

[0175] Input the attribute feature representation corresponding to the molecular structure of Tox21 guided by the prompt into the classifier for classification (obtaining attributes);

[0176] 35) Input the molecular structure (without attribute labels) of ToxCast (a multi-year toxicology prediction research project started by the U.S. Environmental Protection Agency (EPA)) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g5 ;

[0177] Based on the instruction prompt text constructed by ToxCast in 1), the feature vector h of the prompt text is obtained according to 2). p5 ;

[0178] The feature vector h of the prompt text p5 and the structural feature vector h g5 After fusion, the attribute feature representation corresponding to the molecular structure of ToxCast after prompt guidance is obtained;

[0179]

[0180] in,

[0181] h property5 To indicate the property feature representation of the molecular structure of ToxCast after guidance;

[0182] α is the weight parameter, For splicing;

[0183] Input the attribute feature representation corresponding to the molecular structure of ToxCast guided by the prompt into the classifier for classification (obtaining attributes);

[0184] 36) Input the molecular structure of HIV (human immunodeficiency virus) (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g6 ;

[0185] Based on the instruction prompt text constructed by HIV in 1), the feature vector h of the prompt text is obtained according to 2). p6 ;

[0186] The feature vector h of the prompt text p6 and the structural feature vector h g6 Fusion, to obtain the attribute feature representation corresponding to the molecular structure of HIV after prompt guidance;

[0187]

[0188] in,

[0189] h property6 It is a representation of the attribute characteristics corresponding to the molecular structure of HIV after prompting;

[0190] α is the weight parameter, For splicing;

[0191] Input the attribute feature representation corresponding to the molecular structure of HIV guided by the prompt into the classifier for classification (obtaining attributes);

[0192] Each subtask corresponds to a data set, each of which contains multiple molecular structures and corresponding attribute labels; each subtask corresponds to many molecular structures, and each molecular structure corresponds to an h p , h in 1 subtask p are the same.

[0193] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0194] Specific embodiment ten: This embodiment differs from any one of specific embodiments one to nine in that: in said 4), based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text, the drug interaction relationship is predicted for the molecular structure pair (without interaction relationship); the specific process is:

[0195] 41) Input the molecular structure pair of ZhangDDI (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g7 and h g8 ;

[0196] Based on the instruction prompt text constructed by ZhangDDI in 1), the feature vector h of the prompt text is obtained according to 2). p7 ;

[0197] The structural feature vector h g7 、h g8 and the feature vector h of the prompt text p7 After splicing, the molecular structure of ZhangDDI is obtained to obtain the characteristic representation of the corresponding drug interaction relationship;

[0198]

[0199] in,

[0200] h ddi1 To indicate the characteristic representation of drug interaction relationship between the molecular structure of ZhangDDI after guidance;

[0201] For splicing;

[0202] Input the drug interaction relationship features corresponding to the molecular structure of ZhangDDI after prompt guidance into the classifier for classification (obtaining attributes);

[0203] 42) Input the molecular structure pair of ChChMiner (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g9 and hg10 ;

[0204] Based on the instruction prompt text constructed by ChChMiner in 1), the feature vector h of the prompt text is obtained according to 2). p8 ;

[0205] The structural feature vector h g9 、h g10 and the feature vector h of the prompt text p8 After splicing, the molecular structure of ChChMiner guided by hints is obtained, and the corresponding drug interaction relationship feature representation is obtained;

[0206]

[0207] in,

[0208] h ddi2 To indicate the characteristic representation of drug interaction relationship between molecular structure pairs of ChChMiner after guidance;

[0209] The molecular structure of ChChMiner guided by the prompts and the corresponding drug interaction relationship feature representation are input into the classifier for classification (obtaining attributes);

[0210] 43) Input the molecular structure pair of DeepDDI (without attribute labels) into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g11 and h g12 ;

[0211] Based on the instruction prompt text constructed by DeepDDI in 1), the feature vector h of the prompt text is obtained according to 2). p9 ;

[0212] The structural feature vector h g11 and h g12 and the feature vector h of the prompt text p9 After splicing, the molecular structure of DeepDDI guided by hints is obtained, and the corresponding drug interaction relationship feature representation is obtained;

[0213]

[0214] in,

[0215] h ddi3 To indicate the characteristic representation of drug interaction relationship corresponding to the molecular structure of DeepDDI after guidance;

[0216] The molecular structure of DeepDDI guided by the prompt is input into the classifier to classify (obtain attributes) the corresponding drug interaction relationship feature representation.

[0217] Each subtask corresponds to a data set, each of which contains multiple molecular structure pairs (two molecular structures) and corresponding drug interaction relationship labels; each subtask corresponds to many molecular structures, and each molecular structure corresponds to an h p , h in 1 subtask p are the same.

[0218] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 9.

[0219] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A molecular structure prediction system based on a multimodal fine-grained molecular pre-training model with prompt learning, characterized by: The system includes: a data acquisition module, a processing module, and a prediction module; The data acquisition module is used to acquire sample data in the multimodal molecular pre-training data set and sample data in the downstream task data set; The processing module is used to establish a multimodal fine-grained molecular pre-training model based on prompt learning, and obtain the trained multimodal fine-grained molecular pre-training model based on prompt learning; The prediction module is used to predict the properties and drug interaction relationship of the molecular structure to be tested based on the multimodal fine-grained molecular pre-training model of trained prompt learning to obtain a prediction result.

2. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 1, characterized in that: The data acquisition module is used to acquire sample data in the multimodal molecular pre-training data set and sample data in the downstream task data set; The specific process is: The sample data in the multimodal molecular pre-training dataset are molecule-text pairs; The sample data in the downstream task dataset is a molecular structure.

3. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 2, characterized in that: The processing module is used to establish a multimodal fine-grained molecular pre-training model based on prompt learning, and obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning; The specific process is: The multimodal fine-grained molecular pre-training model based on prompt learning includes: a fine-grained molecular graph construction module, a molecular and text description representation learning module, a self-supervised collaborative comparison optimization module, and an instruction fine-tuning downstream task module; The fine-grained molecular graph construction module uses a specific decomposition rule to decompose the molecular structure graph G in the sample data in the multimodal molecular pre-training data set into sub-structure units, and constructs a fine-grained molecular graph G' based on the sub-structure units; The molecular and text description representation learning module is used to input the text description T in the sample data in the multimodal molecular pre-training data set into the text encoder, and the text encoder outputs semantic features; input the fine-grained molecular graph G″ corresponding to 1.65 million molecular structures in the PubChem data set into the molecular encoder, and the molecular encoder outputs a molecular feature vector; train the molecular encoder GIN to obtain a pre-trained molecular encoder GIN; The self-supervised collaborative contrast optimization module uses a contrast learning method to train the pre-trained molecular encoder GIN and text encoder to obtain the trained molecular encoder GIN and text encoder; the trained molecular encoder GIN and text encoder constitute a multimodal fine-grained molecular pre-training model; The instruction fine-tuning downstream task module is used to use prompt learning to guide the multimodal fine-grained molecular pre-training model to understand downstream tasks, and obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning.

4. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 3, characterized in that: The fine-grained molecular graph construction module uses a specific decomposition rule to decompose the molecular structure graph G in the sample data in the multimodal molecular pre-training data set into sub-structure units, and constructs a fine-grained molecular graph G' based on the sub-structure units; The specific process is: 1) Based on the RDKit tool, obtain the chemical element characteristics of each atom in the molecular structure represented by the SMILES string in the sample data in the multimodal molecular pre-training dataset; The chemical element characteristics of each atom include the atomic number, degree of connectivity, type of bonds between atoms, and type of atomic ring formation; Use connectivity features as node features; The bond type between atoms and the type of atoms forming a ring are used as bond characteristics; The connectivity is the number of chemical bonds connecting atoms; 2) Based on the RDKit tool, the molecular structure represented by the SMILES string in the sample data in the multimodal molecular pre-training dataset is converted into a 2D topological graph G = (V, E); Wherein, V represents the node set of the 2D topological graph G, and E represents the edge set of the 2D topological graph G; 3) Based on the BRICS decomposition method, the molecular structure in the sample data in the multimodal molecular pre-training data set is decomposed to obtain substructure units; the specific process is as follows: Based on the BRICS decomposition method, the molecular structure in the sample data in the multimodal molecular pre-training data set is preliminarily decomposed to obtain multiple substructure units after the preliminary decomposition; Decompose each sub-structure unit after the preliminary decomposition to obtain multiple minimum sub-structure units; 4) Use the substructure unit as the new node V f Add to the topology map; The connection relationship E between each substructure unit and the nodes it contains f Added to the topology graph as a new edge; Construct an empty graph node V g , the graph-level node V g With all new nodes V f Connect to get the connection relationship E g , based on V,V f ,V g ,E,E f ,E g Forming a fine-grained molecular graph G'; G=(V′,E'),V=[V,V f ,V g ],E=[E,E f ,E g ]。 5. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 4, characterized in that: The molecular and text description representation learning module is used to input the text description T in the sample data in the multimodal molecular pre-training data set into the text encoder, and the text encoder outputs semantic features; input the fine-grained molecular graph G″ corresponding to 1.65 million molecular structures in the PubChem data set into the molecular encoder, and the molecular encoder outputs a molecular feature vector; train the molecular encoder GIN to obtain a pre-trained molecular encoder GIN; The specific process is: Step 1), changing the molecule name in the text of the molecule text pair in the sample data in the multimodal molecule pre-training dataset to a unified format of "The molecule is"; The text in the unified format of "The molecule is" is input into the text encoder, and the text encoder outputs semantic features; The text encoder is SciBERT based on the BERT architecture; Step 2), the fine-grained molecular graph G″ corresponding to the 1.65 million molecular structures in the PubChem dataset and the node features and bond features corresponding to the 1.65 million molecular structures in the PubChem dataset are input into the molecular encoder GIN, and the molecular encoder GIN outputs a feature vector; The molecular encoder GIN includes: input layer, hidden layer, hidden layer, hidden layer, hidden layer; The molecular encoder GIN is optimized by using the self-supervised learning method of generation task and prediction task until convergence, and a pre-trained molecular encoder GIN is obtained; The generated tasks are the connectivity of atoms, the types corresponding to the atomic numbers, and the bond types between atoms; The prediction task is the number of atoms and the number of bonds in a molecular structure.

6. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 5, characterized in that: The self-supervised collaborative contrast optimization module uses a contrast learning method to train the pre-trained molecular encoder GIN and text encoder to obtain the trained molecular encoder GIN and text encoder; the trained molecular encoder GIN and text encoder constitute a multimodal fine-grained molecular pre-training model; The specific process is: The training set is sample data in the multimodal molecular pre-training data set; The loss function is the InfoNCE loss function; The optimization method is contrastive learning method; The text encoder and the pre-trained molecular encoder GIN are trained until convergence to obtain the trained text encoder and molecular encoder GIN; The trained text encoder and molecular encoder GIN constitute a multimodal fine-grained molecular pre-training model.

7. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 6, characterized in that: The instruction fine-tuning downstream task module is used to use prompt learning to guide the multimodal fine-grained molecular pre-training model to understand the downstream task, and obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning; The specific process is: 1) Build the instruction prompt text; the specific process is: 11) Obtaining attribute prediction tasks in the training set from sample data in the downstream task dataset; The attribute prediction task includes six subtasks: BBBP, Bace, Sider, Tox21, ToxCast, and HIV. Each subtask corresponds to a data set, that is, the attribute prediction task includes six data sets; The BBBP is blood-brain barrier penetrating; The Bace is a β-amyloid precursor protein cleaving enzyme; The Sider is a side effect; The Tox21 is the 21st Century Toxicology Test; The ToxCast is a multi-year toxicology prediction research project initiated by the EPA; The HIV is human immunodeficiency virus; The BBBP dataset contains 2039 molecular structures and corresponding attribute labels; The Bace dataset contains 1513 molecular structures and corresponding attribute labels; The Sider dataset contains 1427 molecular structures and corresponding attribute labels; The Tox21 dataset contains 7831 molecular structures and corresponding attribute labels; The ToxCast dataset contains 8576 molecular structures and corresponding attribute labels; The HIV dataset contains 41,127 molecular structures and corresponding attribute labels; 12) Obtain drug interaction tasks in the training set from sample data in the downstream task dataset; The drug interaction task includes three subtasks: ZhangDDI, ChChMiner, and DeepDDI. Each subtask corresponds to a dataset, that is, the drug interaction task includes three datasets; The ZhangDDI dataset contains 48,548 molecular structure pairs and corresponding drug interaction relationship labels; The ChChMiner dataset contains 48,514 molecular structure pairs and corresponding drug interaction relationship labels; The DeepDDI dataset contains 192,284 molecular structure pairs and corresponding drug interaction relationship labels; 13) Construct instruction prompt text according to the task objectives, task background, guiding principles and task requirements of each subtask; 2) Based on the instruction prompt text constructed in 1), obtain the feature vector h of the prompt text p ; 3) Predict the properties of the molecular structure based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text; 4) Predict drug interaction relationships for molecular structure pairs based on the multimodal fine-grained molecular pre-training model and feature vectors of prompt text; 5) Repeat 1) to 4) until convergence to obtain a trained multimodal fine-grained molecular pre-training model based on prompt learning.

8. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 7, characterized in that: In the above 2), based on the instruction prompt text constructed in 1), the feature vector h of the prompt text is obtained. p ; The specific process is: Create a hint embedding layer P∈R( l×d ), where l is the number of virtual tokens, d is the dimension of embedding; R is a real number; Input the instruction prompt text constructed in 1) into the word segmenter corresponding to the trained text encoder, and the word segmenter outputs a token ID sequence; The token ID sequence is input into the trained text encoder, and the word embedding layer in the trained text encoder outputs the word embedding vector; The word embedding vector is used as the initial weight of the created prompt embedding layer P to obtain the initialized prompt embedding layer; The virtual token in the initialized prompt embedding layer is input into the trained text encoder, and the trained text encoder outputs the feature vector h corresponding to the instruction prompt text p .

9. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 8, characterized in that: In the above 3), based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text, the molecular structure is predicted; The specific process is: 31) Input the molecular structure of BBBP into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g1 ; Based on the instruction prompt text constructed by BBBP in 1), the feature vector h of the prompt text is obtained according to 2). p1 ; The feature vector h of the prompt text p1 and the structural feature vector h g1 After fusion, the attribute feature representation corresponding to the molecular structure of BBBP after prompt guidance is obtained; in, j property1 It is the attribute feature representation corresponding to the molecular structure of BBBP after prompt guidance; α is the weight parameter, For splicing; The attribute feature representation corresponding to the molecular structure of BBBP guided by the prompt is input into the classifier for classification; 32) Input the molecular structure of Bace into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g2 ; Based on the instruction prompt text constructed by Bace in 1), the feature vector j of the prompt text is obtained according to 2). p2 ; The feature vector h of the prompt text p2 and the structural feature vector h g2 Fusion, to obtain the attribute feature representation corresponding to the molecular structure of Bace guided by the prompt; in, h property2 It is the attribute feature representation corresponding to the molecular structure of Bace after prompt guidance; α is the weight parameter, For splicing; The attribute feature representation corresponding to the molecular structure of Bace guided by the prompt is input into the classifier for classification; 33) Input the molecular structure of Sider into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g3 ; Based on the instruction prompt text constructed by Sider in 1), the feature vector h of the prompt text is obtained according to 2). p3 ; The feature vector h of the prompt text p3 and the structural feature vector h g3 Fusion, to obtain the attribute feature representation corresponding to the molecular structure of Sider guided by the prompt; in, h property3 It is the attribute feature representation corresponding to the molecular structure of Sider after prompt guidance; α is the weight parameter, For splicing; The attribute feature representation corresponding to the molecular structure of Sider guided by the prompt is input into the classifier for classification; 34) Input the molecular structure of Tox21 into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g4 ; Based on the instruction prompt text constructed by Tox21 in 1), the feature vector h of the prompt text is obtained according to 2). p4 ; The feature vector h of the prompt text p4 and the structural feature vector h g4 After fusion, the attribute feature representation corresponding to the molecular structure of Tox21 after prompt guidance is obtained; in, h property4 It is the property feature representation corresponding to the molecular structure of Tox21 after prompt guidance; α is the weight parameter, For splicing; The attribute feature representation corresponding to the molecular structure of Tox21 guided by the prompt is input into the classifier for classification; 35) Input the molecular structure of ToxCast into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g5 ; Based on the instruction prompt text constructed by ToxCast in 1), the feature vector h of the prompt text is obtained according to 2). p5 ; The feature vector h of the prompt text p5 and the structural feature vector h g5 After fusion, the attribute feature representation corresponding to the molecular structure of ToxCast after prompt guidance is obtained; in, h property5 To indicate the property feature representation of the molecular structure of ToxCast after guidance; α is the weight parameter, For splicing; The attribute feature representation corresponding to the molecular structure of ToxCast guided by the prompt is input into the classifier for classification; 36) Input the molecular structure of HIV into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g6 ; Based on the instruction prompt text constructed by HIV in 1), the feature vector h of the prompt text is obtained according to 2). p6 ; The feature vector h of the prompt text p6 and the structural feature vector h g6 Fusion, to obtain the attribute feature representation corresponding to the molecular structure of HIV after prompt guidance; in, h property6 It is a representation of the attribute characteristics corresponding to the molecular structure of HIV after prompting; α is the weight parameter, For splicing; The attribute feature representation corresponding to the molecular structure of HIV after prompt guidance is input into the classifier for classification.

10. The molecular structure prediction system of the multimodal fine-grained molecular pre-training model based on prompt learning according to claim 9, characterized in that: In the above 4), based on the multimodal fine-grained molecular pre-training model and the feature vector of the prompt text, the drug interaction relationship of the molecular structure pair is predicted; the specific process is: 41) Input the molecular structure of ZhangDDI into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g7 and h g8 ; Based on the instruction prompt text constructed by ZhangDDI in 1), the feature vector h of the prompt text is obtained according to 2). p7 ; The structural feature vector h g7 、h g8 and the feature vector h of the prompt text p7 After splicing, the molecular structure of ZhangDDI is obtained to obtain the characteristic representation of the corresponding drug interaction relationship; in, h ddi1 To indicate the characteristic representation of drug interaction relationship between the molecular structure of ZhangDDI after guidance; For splicing; The molecular structure of ZhangDDI guided by the prompt is input into the classifier to classify the corresponding drug interaction relationship features; 42) Input the molecular structure pair of ChChMiner into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g9 and h g10 ; Based on the instruction prompt text constructed by ChChMiner in 1), the feature vector h of the prompt text is obtained according to 2). p8 ; The structural feature vector h g9 、h g10 and the feature vector h of the prompt text p8 After splicing, the molecular structure of ChChMiner guided by hints is obtained, and the corresponding drug interaction relationship feature representation is obtained; in, h ddi2 To indicate the characteristic representation of drug interaction relationship between molecular structure pairs of ChChMiner after guidance; The molecular structure pairs of ChChMiner guided by the prompts and the corresponding drug interaction relationship feature representations are input into the classifier for classification; 43) Input the molecular structure pair of DeepDDI into the trained molecular encoder GIN in the multimodal fine-grained molecular pre-training model, and the trained molecular encoder GIN outputs the structural feature vector h g11 and h g12 ; Based on the instruction prompt text constructed by DeepDDI in 1), the feature vector h of the prompt text is obtained according to 2). p9 ; The structural feature vector h g11 and h g12 and the feature vector h of the prompt text p9 After splicing, the molecular structure of DeepDDI guided by hints is obtained, and the corresponding drug interaction relationship feature representation is obtained; in, h ddi3 To indicate the characteristic representation of drug interaction relationship corresponding to the molecular structure of DeepDDI after guidance; The molecular structure of DeepDDI guided by the prompts is input into the classifier for classification.

Images