Molecular generation method and device, storage medium, equipment and program product

Abstract

The invention discloses a molecule generation method and device, a storage medium, equipment and a program product, which are applied to a molecule generation scene. The method comprises the following steps: obtaining known modal protein information corresponding to a target protein; performing feature processing on the known modal protein information based on a feature processing module of the trained multi-modal language model to obtain multi-modal standardized features corresponding to the target protein; based on a molecule generation module of the trained multi-modal language model, performing molecule generation processing on the multi-modal standardized features to obtain a target molecule aiming at the target protein; the trained multi-modal language model is obtained by training based on a multi-modal data set, and the multi-modal data set comprises a sequence modal sample, a structure modal sample and a pocket modal sample of a sample protein corresponding to a sample molecule. According to the method, high-quality drug molecules can be generated under the condition that the target protein lacks certain modal information, and the drug design efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a molecule generation method, device, storage medium, equipment and program product. Background Art

[0002] In the field of drug design and molecular biology, generating molecules that bind to specific proteins is a key step in achieving targeted therapy and drug development. Currently, this process mainly relies on three technical methods: pocket-based structure-based drug design (SBDD), sequence-based methods, and diffusion and flow model-based methods. These methods have their own characteristics, but they all face some common technical challenges and limitations, such as dependence on the three-dimensional structure of proteins and the lack of learning ability for unknown structural protein-molecule interactions.

[0003] Therefore, current technologies have problems with their reliance on protein three-dimensional structure and their ability to learn protein-molecule interactions with unknown structures. Summary of the invention

[0004] The embodiments of the present application provide a molecule generation method, apparatus, storage medium, device and program product, which can generate multimodal standardized features using known modal protein information of at least one modality of the target protein, thereby generating high-quality drug molecules when the target protein lacks certain modal information, thereby improving the efficiency of drug design.

[0005] In one aspect, the present invention provides a method for generating a molecule, the method comprising:

[0006] Acquire known modality protein information corresponding to the target protein, wherein the known modality of the known modality protein information includes at least one modality in a modality set, and the modality set includes sequence modality, structural modality, and pocket modality;

[0007] Performing feature processing on the known modal protein information based on the feature processing module of the trained multimodal language model to obtain multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structure standardized features, and pocket standardized features having the same dimension;

[0008] A molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain a target molecule for the target protein;

[0009] The trained multimodal language model is obtained based on a multimodal dataset, wherein the multimodal dataset includes sequence modal samples, structural modal samples and pocket modal samples of sample proteins corresponding to sample molecules, and the multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

[0010] On the other hand, an embodiment of the present application provides a molecule generation device, the device comprising:

[0011] An acquisition unit, used for acquiring known modality protein information corresponding to the target protein, wherein the known modality of the known modality protein information includes at least one modality in a modality set, and the modality set includes sequence modality, structural modality and pocket modality;

[0012] A feature processing unit, configured to perform feature processing on the known modal protein information based on the feature processing module of the trained multimodal language model, and obtain multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structure standardized features, and pocket standardized features having the same dimension;

[0013] A molecule generation unit, configured to perform molecule generation processing on the multimodal standardized features based on the molecule generation module of the trained multimodal language model to obtain a target molecule for the target protein;

[0014] The trained multimodal language model is obtained based on a multimodal dataset, wherein the multimodal dataset includes sequence modal samples, structural modal samples and pocket modal samples of sample proteins corresponding to sample molecules, and the multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

[0015] In some embodiments, the feature processing module includes a feature encoding module, a projector and a modality recovery module. The feature processing unit is used to perform feature processing on the known modality protein information based on the feature processing module of the trained multimodal language model to obtain the multimodal standardized features corresponding to the target protein, specifically for:

[0016] If the known modalities of the known modal protein information include some modalities in the modal set, determining the missing modality corresponding to the target protein according to the known modal protein information and the modal set;

[0017] Based on the feature encoding module, feature encoding processing is performed on the known modal protein information to obtain known modal source features corresponding to the target protein, wherein the known modal source features include one or two of sequence source features, structure source features and pocket source features;

[0018] Based on the projector, the known modality source features are standardized to obtain known modality standardized features corresponding to the target protein;

[0019] Based on the modality recovery module, modality recovery processing is performed on the known modality standardized features and the first target modality prompt information to obtain the missing modality recovery standardized features corresponding to the target protein, and the first target modality prompt information is used to prompt the recovery of the missing modality;

[0020] According to the known modality normalization feature and the missing modality recovery normalization feature, a multimodality normalization feature corresponding to the target protein is obtained, and the multimodality normalization feature includes a sequence normalization feature, a structure normalization feature and a pocket normalization feature.

[0021] In some embodiments, the feature processing unit is used to perform feature processing on the known modal protein information and the modal set based on the feature processing module of the trained multimodal language model to obtain the multimodal standardized features corresponding to the target protein, and is also used to:

[0022] If the known modalities of the known modal protein information include all modalities in the modality set, then based on the feature encoding module, feature encoding processing is performed on the known modal protein information to obtain known modal source features corresponding to the target protein, wherein the known modal source features include sequence source features, structure source features and pocket source features;

[0023] Based on the projector, each source feature in the known modal source features is standardized to obtain a multimodal standardized feature corresponding to the target protein, wherein the multimodal standardized feature includes a sequence standardized feature, a structure standardized feature and a pocket standardized feature.

[0024] In some embodiments, the molecule generation module includes a modality embedding module, an encoder and a decoder. When the molecule generation module is used to perform molecule generation processing on the multimodal standardized features based on the trained multimodal language model to obtain a target molecule for the target protein, the molecule generation unit is specifically used to:

[0025] Based on the modality embedding module, each standardized feature in the multimodal standardized features is subjected to modality embedding processing to obtain a multimodal embedding feature, wherein the multimodal embedding feature includes a sequence embedding feature, a structure embedding feature and a pocket embedding feature;

[0026] Based on the encoder, feature fusion processing is performed on the multimodal embedding features to obtain fusion features corresponding to the target protein;

[0027] The fusion feature is subjected to feature decoding processing based on the decoder to obtain a target molecule for the target protein.

[0028] In some embodiments, when the molecular processing unit is used to perform modality embedding processing on each standardized feature in the multimodal standardized feature based on the modality embedding module to obtain the multimodal embedding feature, it is specifically used to:

[0029] Based on the modality embedding module, the sequence normalization feature in the multimodal normalization feature is added to the sequence modality embedding to obtain the sequence embedding feature in the multimodal embedding feature, wherein the sequence modality embedding is used to enhance the feature representation of the sequence modality;

[0030] Based on the modality embedding module, the structural normalized features in the multimodal normalized features are added to the structural modality embedding to obtain the structural embedding features in the multimodal embedding features, wherein the structural modality embedding is used to enhance the feature representation of the structural modality;

[0031] Based on the modality embedding module, the pocket standardized features in the multimodal standardized features are added to the pocket modality embedding to obtain the pocket embedding features in the multimodal embedding features, and the pocket modality embedding is used to enhance the feature representation of the pocket modality.

[0032] In some embodiments, the apparatus further comprises a training unit for:

[0033] Acquire a multimodal data set, wherein the multimodal data set includes sequence modal samples, structure modal samples, and pocket modal samples of sample proteins corresponding to sample molecules, and the multimodal data set is divided into a sequence modal sample subset, a pocket modal sample subset, and a complete modal sample subset;

[0034] Training the feature processing module in the multimodal language model according to the multimodal data set to determine a modality recovery loss function;

[0035] Training the molecule generation module in the multimodal language model according to the multimodal data set to determine a molecule generation loss function;

[0036] The model parameters of the multimodal language model are updated based on the modality recovery loss function and the molecule generation loss function to obtain the trained multimodal language model.

[0037] In some embodiments, when the training unit is used to train the feature processing module in the multimodal language model according to the multimodal data set to determine the modality recovery loss function, it is specifically used to:

[0038] Based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modality sample subset to obtain multimodal standardized feature true values ​​corresponding to the sample protein, wherein the multimodal standardized feature true values ​​include sequence standardized feature true values, structure standardized feature true values, and pocket standardized feature true values;

[0039] Based on the feature encoding module, the projector and the modality recovery module in the feature processing module, feature processing is performed on the sequence modality sample subset or the pocket modality sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, wherein the multimodal standardized feature prediction value includes a sequence standardized feature prediction value, a structure standardized feature prediction value and a pocket standardized feature prediction value;

[0040] The modality restoration loss function is calculated according to the difference between the true value of the multimodal standardized feature and the predicted value of the multimodal standardized feature.

[0041] In some embodiments, when the training unit is used to perform feature processing on the complete modality sample subset based on the feature encoding module and the projector in the feature processing module to obtain the multimodal standardized feature true value corresponding to the sample protein, it is specifically used to:

[0042] Based on the feature encoding module, feature encoding processing is performed on the complete modality sample subset to obtain the known modality source feature true value corresponding to the sample protein, wherein the known modality source feature true value includes the sequence source feature true value, the structure source feature true value and the pocket source feature true value;

[0043] Based on the projector, each source feature true value in the known modality source feature true value is standardized to obtain a multimodal standardized feature true value corresponding to the sample protein.

[0044] In some embodiments, the training unit is used to perform feature processing on the sequence modality sample subset or the pocket modality sample subset based on the feature encoding module, the projector and the modality recovery module in the feature processing module to obtain the multimodal standardized feature prediction value corresponding to the sample protein, specifically for:

[0045] Based on the feature encoding module, feature encoding processing is performed on the sequence modal sample subset to obtain the sequence source feature true value corresponding to the sample protein; based on the projector, the sequence source feature true value is standardized to obtain the sequence standardized feature true value corresponding to the sample protein; and based on the modality recovery module, modality recovery processing is performed on the sequence standardized feature true value and the second target modality prompt information to obtain the structure standardized feature prediction value and the pocket standardized feature prediction value corresponding to the sample protein, and the second target modality prompt information is used to prompt the recovery of the structure modality and the pocket modality; or

[0046] Based on the feature encoding module, feature encoding processing is performed on the pocket modal sample subset to obtain the pocket source feature true value corresponding to the sample protein; based on the projector, the pocket source feature true value is standardized to obtain the pocket standardized feature true value corresponding to the sample protein; and based on the modality recovery module, modality recovery processing is performed on the pocket standardized feature true value and the third target modality prompt information to obtain the sequence standardized feature prediction value and the structure standardized feature prediction value corresponding to the sample protein, and the third target modality prompt information is used to prompt the recovery of the sequence modality and the structure modality.

[0047] In some embodiments, when the training unit is used to train the molecule generation module in the multimodal language model according to the multimodal data set and determine the molecule generation loss function, it is specifically used to:

[0048] Based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modality sample subset to obtain multimodal standardized feature true values ​​corresponding to the sample protein, wherein the multimodal standardized feature true values ​​include sequence standardized feature true values, structure standardized feature true values, and pocket standardized feature true values;

[0049] Get the real token at position i in the real embedding string sequence of the sample molecule;

[0050] Given all tokens before position i in the true embedded string sequence, performing molecule generation processing on the true value of the multimodal standardized feature based on the molecule generation module to obtain a predicted token at position i in the predicted embedded string sequence of the sample molecule;

[0051] The molecule generation loss function is calculated based on the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

[0052] In some embodiments, when the training unit is used to train the molecule generation module in the multimodal language model according to the multimodal data set and determine the molecule generation loss function, it is specifically used to:

[0053] Based on the feature encoding module, the projector and the modality recovery module in the feature processing module, feature processing is performed on the sequence modality sample subset or the pocket modality sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, wherein the multimodal standardized feature prediction value includes a sequence standardized feature prediction value, a structure standardized feature prediction value and a pocket standardized feature prediction value;

[0054] Get the real token at position i in the real embedding string sequence of the sample molecule;

[0055] Given all tokens before position i in the true embedded string sequence, performing molecule generation processing on the true value of the multimodal standardized feature based on the molecule generation module to obtain a predicted token at position i in the predicted embedded string sequence of the sample molecule;

[0056] The molecule generation loss function is calculated based on the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

[0057] In some embodiments, the model parameters include projection parameters of the projector, modal recovery parameters of the modal recovery module, and molecule generation parameters of the molecule generation module; when the training unit is used to update the model parameters of the multimodal language model based on the modal recovery loss function and the molecule generation loss function, it is specifically used to

[0058] Updating a modal restoration parameter of the modal restoration module based on the modal restoration loss function;

[0059] The projection parameters of the projector and the molecule generation parameters of the molecule generation module are updated based on the molecule generation loss function.

[0060] On the other hand, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the molecule generation method described in any of the above embodiments.

[0061] On the other hand, an embodiment of the present application provides a computer device, which includes a processor and a memory, wherein a computer program is stored in the memory, and the processor is used to execute the molecule generation method described in any of the above embodiments by calling the computer program stored in the memory.

[0062] On the other hand, an embodiment of the present application provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the molecule generation method described in any of the above embodiments.

[0063] The embodiment of the present application obtains known modal protein information corresponding to the target protein, wherein the known modalities of the known modal protein information include at least one modality in a modal set, and the modal set includes sequence modalities, structural modalities, and pocket modalities; a feature processing module based on a trained multimodal language model performs feature processing on the known modal protein information to obtain multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structural standardized features, and pocket standardized features having the same dimension; a molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain a target molecule for the target protein; wherein the trained multimodal language model is obtained by training based on a multimodal data set, and the multimodal data set includes sequence modal samples, structural modal samples, and pocket modal samples of the sample protein corresponding to the sample molecule, and the multimodal data set is divided into a sequence modal sample subset, a pocket modal sample subset, and a complete modal sample subset. The embodiment of the present application can generate multimodal standardized features using known modal protein information of at least one modality of the target protein, thereby generating high-quality drug molecules in the absence of certain modal information of the target protein, and improving the efficiency of drug design. This method reduces the dependence on the complete three-dimensional structure of the protein, thereby speeding up the drug design process. Through the feature processing module of the multimodal language model, the multimodal language model of the present application is trained based on a multimodal data set containing sequence modal samples, structural modal samples, and pocket modal samples. This multimodal training method enables the model to more fully learn the associations and complementary information between different modalities, thereby improving the model's ability to learn the interactions between unknown structural proteins and target molecules. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of the architecture of a molecule generation system provided in an embodiment of the present application.

[0065] Figure 2 A schematic diagram of the process of generating molecules provided in an embodiment of the present application.

[0066] Figure 3 A schematic diagram of the architecture of a multimodal language model provided in an embodiment of the present application.

[0067] Figure 4 A flowchart of a method for training a multimodal language model provided in an embodiment of the present application.

[0068] Figure 5A schematic diagram of the training process of the modal recovery module provided in an embodiment of the present application.

[0069] Figure 6 A schematic diagram of the training process of the projector and molecule generation module provided in an embodiment of the present application.

[0070] Figure 7 A comparison chart of experimental results of the multimodal language model provided in the embodiment of the present application and other models on the PDBBind database.

[0071] Figure 8 A comparison chart of experimental results of the multimodal language model provided in the embodiment of the present application and other models on the CrossDocked database.

[0072] Fig. 9 A comparison chart of experimental results of the multimodal language model provided in the embodiment of the present application and other models on the BindingDB database.

[0073] Fig.10 A graph showing the experimental results of an ablation experiment on the PDB-Bind database using the multimodal language model provided in an embodiment of the present application.

[0074] Fig.11 A diagram showing the experimental results of ablation experiments on the BindingDB database and the CrossDocked database using the multimodal language model provided in an embodiment of the present application and the multimodal language model without the modal recovery module.

[0075] Fig.12 A schematic diagram of the structure of a molecule generation device provided in an embodiment of the present application.

[0076] Fig.13 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0077] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0078] The present application embodiment provides a kind of molecule generation method, device, storage medium, equipment and program product. Exemplarily, the molecule generation method of the present application embodiment can be performed by computer equipment, wherein the computer equipment can be a terminal or server and other equipment. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart TV, smart speaker, wearable smart device, personal computer (Personal Computer, PC), smart car terminal and other equipment, and the terminal can also include a client, which can be a molecular design client, a deep learning model client, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (Content Delivery Network, CDN) and basic cloud computing services such as big data and artificial intelligence platforms.

[0079] The embodiments of the present application can be applied to scenarios such as molecular design and generation, molecular optimization, drug screening and evaluation. For example, in the process of drug discovery, the molecular generation method provided by the present application can obtain multimodal standardized features including sequence modal protein information, structural modal protein information and pocket modal protein information by processing the known modal protein information corresponding to the complex target protein, and further generate high-affinity drug molecules that can bind to the target protein based on the multimodal standardized features. The molecular generation method provided by the embodiments of the present application can still generate multimodal standardized features containing complete modal protein information using the acquired partial modal protein information even in the absence of partial modal protein information, thereby improving the efficiency of drug molecule generation. In addition, through the feature processing module of the multimodal language model, the multimodal language model of the present application is trained based on a multimodal data set containing sequence modal samples, structural modal samples and pocket modal samples. This multimodal training method enables the model to more fully learn the association and complementary information between different modalities, thereby improving the model's ability to learn the interaction between unknown structural proteins and target molecules.

[0080] The scheme provided in the embodiments of the present application involves technologies such as molecular generation, which are specifically described by the following embodiments. The following are described in detail. It should be noted that the order of description of the following embodiments is not intended to limit the priority order of the embodiments.

[0081] See also Figure 1 , Figure 1The schematic diagram of the architecture of the molecule generation system provided in the embodiment of the present application. The system can implement the molecule generation method. The molecule generation system may include a terminal 1000, a server 2000 and a network 3000, the terminal 1000 and the server 2000 may exchange data through the network 3000, and the network 3000 may be a wide area network or a local area network, or a combination of the two.

[0082] In some embodiments, the terminal 1000 is used to generate a molecule generation request. For example, the user inputs information indicating the target protein in the graphical interface of the terminal 1000, and the terminal 1000 generates a corresponding target molecule generation request for the target protein and sends the generation request to the server 2000; the server 2000 is used to obtain the known modal protein information corresponding to the target protein according to the target molecule generation request for the target protein. The server 2000 further performs feature processing on the known modal protein information based on the feature processing module of the trained multimodal language model to obtain the multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structure standardized features and pocket standardized features with the same dimension. The server 2000 then performs molecule generation processing on the multimodal standardized features based on the molecule generation module of the trained multimodal language model to obtain the target molecule for the target protein, wherein the trained multimodal language model is obtained by training the server 2000 based on the multimodal data set. Finally, the server 2000 sends the generated information including the target molecule for the target protein to the terminal 1000, and the terminal 1000 receives and displays the information.

[0083] It should be noted that Figure 1 Taking the server 2000 executing the molecule generation method provided in the embodiment of the present application as an example, in another embodiment, the method may also be executed by the terminal, or may be executed jointly by the terminal and the server.

[0084] Figure 2 This is a flow chart of the method for generating molecules provided in the embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 2 The method may include the following steps 110 to 130:

[0085] Step 110, obtaining known modal protein information corresponding to the target protein, wherein the known modalities of the known modal protein information include at least one modality in a modality set, and the modality set includes sequence modalities, structural modalities, and pocket modalities.

[0086] The computer device obtains known modal protein information corresponding to the target protein, where the known modal protein information includes one or more of sequence modal protein information, structure modal protein information, and pocket modal protein information.

[0087] The target protein can be any protein, for example, a protein involved in a specific physiological or pathological process in an organism. The goal of molecular generation for the target protein is to design and generate molecules that can interact with the target protein. These molecules have drug-like properties, that is, they have the potential to become drugs for treating a disease related to the target protein, thereby achieving effective treatment of the disease.

[0088] The protein information of the target protein can be displayed through three modes in the mode set, including sequence mode, structural mode and pocket mode. Different modes correspond to different properties of the target protein, which can independently or synergistically affect the function or characteristics of the target protein.

[0089] Among them, sequence modality focuses on the amino acid sequence information of proteins, which refers to the linear arrangement order of amino acids in proteins. The sequence of a protein is a linear chain composed of a series of amino acid residues, and the arrangement order of these amino acid residues determines the structure and function of the protein. Through sequence modality, basic information such as molecular weight, isoelectric point, amino acid composition, hydrophilicity and hydrophobicity of proteins can be analyzed. In addition, sequence modality can also be used for protein comparison analysis to help identify homologous proteins and their functions.

[0090] Structural modality focuses on the three-dimensional structural information of proteins, namely their secondary, tertiary and even quaternary structures. The structure of a protein is the basis of its function. Structural modality analyzes the three-dimensional folding of proteins to display the secondary structure (such as α-helix and β-fold) and tertiary structure (such as globular domain) of proteins. Studying the structure of a protein helps reveal the location of its active site, understand its mechanism of action in biochemical processes, and design drug molecules to intervene in specific protein activities.

[0091] Pocket modality focuses on specific cavities or grooves (i.e., "pockets") formed on the surface or inside of proteins. These pockets are key sites for proteins to interact with other molecules (such as ligands, substrates, inhibitors, drug molecules, etc.), such as catalytic sites or regulatory sites. By studying the characteristics of these pockets (such as shape, size, charge distribution, etc.), targeted drugs can be developed and small molecules that can bind to specific protein pockets with high specificity can be designed to treat diseases or regulate biological processes.

[0092] Sequence mode, structural mode and pocket mode together constitute a multidimensional view of protein information. Sequence mode provides basic coding information of protein; structural mode shows the three-dimensional conformation of protein; and pocket mode reveals the potential sites of protein interaction with other molecules. These three modes are interrelated and complementary, and together determine the function and characteristics of protein.

[0093] For example, a complete modality of a protein can be defined as: comp ={p seq , p str , p poc}, where p seq , p str , p poc They represent sequence modal protein information, structural modal protein information and pocket modal protein information respectively. Define A = {seq, str, poc} as the modal set.

[0094] Specifically, the sequence modal protein information corresponding to the target protein includes the linear arrangement order information of the amino acids in the target protein, which provides a basic feature description of the target protein. seq It can be expressed by the following formula:

[0095] Formula 1: p seq =(r1, r2, ... r Nseq );

[0096] Among them, r i represents the ith amino acid residue that constitutes the target protein as a basic unit, and Nseq is the sequence length corresponding to the target protein.

[0097] The structural modal protein information corresponding to the target protein includes the arrangement position information of the main chain atoms of the target protein in three-dimensional space, which determines the function of the target protein. str It can be expressed by the following formula 2:

[0098] Formula 2:

[0099] Among them, a i represents the i-th main chain atom in the target protein, x i Indicates a i The three-dimensional coordinates of Nstr represent the number of main chain atoms included in the target protein.

[0100] The pocket modal protein information corresponding to the target protein includes the structural information of the local binding site in the target protein, which is used to represent the arrangement position information of the atoms corresponding to the local binding site of the target protein in three-dimensional space, and is very critical for executing the function of the target protein. In a specific embodiment provided in the present application, the pocket modal protein information corresponding to the target protein p poc It can be expressed by the following formula three:

[0101] Formula 3:

[0102] Among them, a i represents the i-th atom in the local binding site of the target protein, x i Indicates a i The three-dimensional coordinates of Npoc represent the number of atoms included in the local binding site of the target protein.

[0103] Step 120, based on the feature processing module of the trained multimodal language model, feature processing is performed on the known modal protein information to obtain the multimodal standardized features corresponding to the target protein, and the multimodal standardized features include sequence standardized features, structure standardized features and pocket standardized features with the same dimension.

[0104] Figure 3 For a schematic diagram of the architecture of the multimodal language model provided in the embodiment of the present application, see Figure 3 The multimodal language model includes a feature processing module 10 and a molecule generation module 20, the input data is the known modal protein information corresponding to the target protein, and the output data is the target molecule for the target protein.

[0105] The multimodal language model (denoted as MmDrugGen) is a model that can process data from different modalities. Specifically, the multimodal language model provided in the embodiment of the present application includes a feature processing module 10, which is used to process the known modal protein information corresponding to the target protein to obtain the multimodal standardized features corresponding to the target protein.

[0106] In some embodiments, the feature processing module 10 includes a feature encoding module 11, a projector 12 and a modality recovery module 13. In S120, the feature processing module based on the trained multimodal language model performs feature processing on the known modality protein information to obtain the multimodal standardized features corresponding to the target protein, including the following S1201-S1205:

[0107] S1201. If the known modality of the known modality protein information includes some modalities in the modality set, determining the missing modality corresponding to the target protein according to the known modality protein information and the modality set;

[0108] Specifically, if the known modalities of the known modal protein information include some modalities in the modality set, other modalities in the modality set that are not included in the known modalities are taken as missing modalities corresponding to the target protein.

[0109] For example, if the known modalities of the known modal protein information include the sequence modalities in the modal set, the structural modalities and pocket modalities in the modal set are determined as the missing modalities corresponding to the target protein; for another example, if the known modalities of the known modal protein information include the sequence modalities and pocket modules in the modal set, the structural modalities in the modal set are determined as the missing modalities corresponding to the target protein.

[0110] S1202, performing feature encoding processing on known modal protein information based on a feature encoding module to obtain known modal source features corresponding to the target protein, where the known modal source features include one or two of sequence source features, structure source features, and pocket source features;

[0111] The feature encoding module 11 is used to process known modal protein information and output corresponding known modal source features.

[0112] In some embodiments, the feature encoding module 11 includes three encoders, and different encoders are used to process different modal information of the target protein to obtain source features corresponding to different modal information.

[0113] Then in S1202, feature encoding processing is performed on the known modal protein information based on the feature encoding module to obtain the known modal source features corresponding to the target protein, including:

[0114] For various known modal protein information, the encoder corresponding to each known modal protein information in the feature encoding module is used to extract the modal source features corresponding to the known modal protein information to obtain the known modal source features of the target protein.

[0115] Specifically, for each protein modal information (p seq 、p str 、p poc ), the modal source characteristics corresponding to various protein modal information can be obtained using the following formulas 4, 5, and 6 respectively:

[0116] Formula 4: X seq =f seq (p seq );

[0117] Among them, X seq is the source feature corresponding to the sequence modal protein information, f seq is an encoder for extracting source features corresponding to sequence modal protein information.

[0118] Formula 5: X str =f str (p str );

[0119] Among them, X str is the source feature corresponding to the structural modal protein information, f str is an encoder for extracting source features corresponding to structural modal protein information.

[0120] Formula 6: X poc =f poc (p poc );

[0121] Among them, X poc is the source feature corresponding to the pocket mode protein information, f poc is an encoder for extracting source features corresponding to pocket modal protein information.

[0122] Among them, f seq 、f str and f poc The specific implementation methods can be implemented using relevant technologies respectively. For example, the relevant technologies may include deep learning models for analysis and prediction of protein sequences, universal protein language models enhanced by protein structure skeleton coordinates, and universal large models based on molecular three-dimensional structures, etc. This application does not impose any restrictions on this.

[0123] S1203, performing standardization processing on known modal source features based on a projector to obtain known modal standardized features corresponding to the target protein;

[0124] Since different modal information of the target protein corresponds to different dimensions and representations, it is not conducive to the subsequent integration of different information of the target protein. Therefore, the known modal source features of the target protein including multiple different modalities can be standardized to obtain the known modal standardized features corresponding to the target protein, so as to unify the feature representations of different modalities into the same dimension.

[0125] In some embodiments, the projector 12 includes three sub-projectors, each of which is used to perform standardization processing on source features of different modalities. According to each sub-projector, the known modality standardized features corresponding to the target protein are determined by the following formula 7:

[0126] Formula 7: H k =p k (X k θ proj ), k∈{seq,str,poc};

[0127] Among them, H k Represents the standardized features of each modal source feature, pk is the sub-projector, θ proj is the learnable parameter of the sub-projector, which can also be called the projection parameter.

[0128] In some embodiments, the sub-projectors are implemented by an encoder-decoder, and according to each sub-projector, the known modality normalized features corresponding to the target protein can also be determined by the following formula eight:

[0129] Formula 8: H k =Dec(Enc(X k ),Q k ), k∈{seq,str,poc};

[0130] Among them, Q k For the sub-projector p k Learnable queries in .

[0131] Specifically, Formula 8 indicates that an encoder is first used to encode the known modal source features, and then a decoder is used to decode the encoded known modal source features to obtain known modal standardized features.

[0132] S1204, performing modality recovery processing on the known modality standardized features and the first target modality prompt information based on the modality recovery module to obtain the missing modality recovery standardized features corresponding to the target protein, and the first target modality prompt information is used to prompt the recovery of the missing modality;

[0133] The modality recovery module 13 is used to obtain the missing modality recovery normalization feature of the target protein according to the known modality normalization feature.

[0134] Specifically, after obtaining the missing modality according to the first target modality prompt information, the modality recovery module 13 uses the known modality standardized features to predict the representation of the missing modality and obtain the missing modality standardized features. More specifically, the modality recovery module 13 assigns a specific predictor r to each source modality (i.e., known modality) s , which is used to accurately obtain the information in the standardized features of each modality and assign a specific prompter P to each target modality (i.e., missing modality) t , which is used to follow a specific pattern when generating the restored standardized features of the target modality, helping to understand and generate data that meets the requirements of the target modality. According to the specific presets and prompters of each modality, the missing modality restored standardized features are obtained through the following formulas 9 and 10:

[0135] Formula 9:

[0136] Formula 10:

[0137] Among them, Hs and P t can be in vector format. More specifically, H s and P t are all vectors of dimension m×n, is a vector of dimension 2m×d, s is any mode among the known modes, t is any mode among the missing modes, represents the output of the modality recovery module (i.e., the normalized features of missing modality recovery), H s represents the known modal normalized features, θ rec represents the learnable parameters in the modal recovery module, which can also be called modal recovery parameters. For The prompt P corresponding to the missing modality extracted from t The corresponding part.

[0138] By using the modality recovery module 13 provided in the present application to predict the standardized features corresponding to the missing modality based on the known modality standardized features, the missing modality information can be supplemented when only partial modality information is available, ensuring a comprehensive understanding of the target protein, thereby improving the accuracy of generating target molecules for the target protein.

[0139] S1205. Obtain multimodal standardized features corresponding to the target protein based on known modal standardized features and missing modal restored standardized features. The multimodal standardized features include sequence standardized features, structure standardized features and pocket standardized features.

[0140] Specifically, the computer device uses the known modality source features and the missing modality recovery normalization features as the multimodal normalization features corresponding to the target protein.

[0141] In some other embodiments, in S120, a feature processing module based on a trained multimodal language model performs feature processing on known modal protein information and a modal set to obtain multimodal standardized features corresponding to the target protein, further comprising the following S1206-S1207:

[0142] S1206. If the known modalities of the known modal protein information include all modalities in the modality set, feature encoding processing is performed on the known modal protein information based on the feature encoding module to obtain known modal source features corresponding to the target protein, where the known modal source features include sequence source features, structure source features, and pocket source features;

[0143] Specifically, for each known modal protein information, the computer device extracts the modal source features corresponding to each known modal protein information through the encoder corresponding to each known modal protein information in the feature encoding module 11, and obtains the known modal source features corresponding to the target protein. For details, please refer to the above formulas 4, 5 and 6.

[0144] S1207. Based on the projector, each source feature in the known modal source features is standardized to obtain a multimodal standardized feature corresponding to the target protein. The multimodal standardized feature includes a sequence standardized feature, a structure standardized feature and a pocket standardized feature.

[0145] In some embodiments, the projector 12 includes three sub-projectors, each of which is used to standardize source features of different modalities to obtain known modality standardized features corresponding to the target protein.

[0146] The specific method of using each specific sub-projector to process the source features of different modalities to obtain the known modality standardized features corresponding to the target protein can be found in the above formulas 7 and 8.

[0147] Step 130, based on the molecular generation module of the trained multimodal language model, a molecular generation process is performed on the multimodal standardized features to obtain a target molecule for the target protein; wherein the trained multimodal language model is trained based on a multimodal dataset, and the multimodal dataset includes sequence modal samples, structural modal samples, and pocket modal samples of the sample protein corresponding to the sample molecule, and the multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset, and a complete modal sample subset.

[0148] In some embodiments, the molecule generation module 20 includes a modality embedding module 21, an encoder 22, and a decoder 23. The molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain a target molecule for the target protein, including the following S1301-S1303:

[0149] S1301, performing modality embedding processing on each standardized feature in the multimodal standardized feature based on the modality embedding module to obtain a multimodal embedding feature, where the multimodal embedding feature includes a sequence embedding feature, a structure embedding feature, and a pocket embedding feature;

[0150] Each known modality standardized feature or missing modality standardized feature only focuses on the characteristics of a single modality, without establishing sufficient connections and integration between these modalities. In addition, there is a lack of sufficient differentiation between the standardized features of these different modalities, which easily leads to information confusion and feature redundancy. In order to solve these problems, this application proposes a solution, which specifically includes: for the standardized features of each modality, the standardized features of each modality are modally embedded to obtain multimodal embedded features.

[0151] This approach can capture and emphasize specific information of each mode of the target protein, thereby helping to distinguish different modes and reduce information confusion.

[0152] Among them, in addition to the solution provided in the embodiment of the present application, other methods can also be used to perform modal embedding processing on the standardized features of each modality.

[0153] In some embodiments, in S1301, each standardized feature in the multimodal standardized feature is subjected to modal embedding processing based on the modal embedding module to obtain the multimodal embedding feature, including the following S13011-S13013:

[0154] S13011. Based on the modality embedding module, the sequence normalization feature in the multimodal normalization feature is added to the sequence modality embedding to obtain the sequence embedding feature in the multimodal embedding feature, wherein the sequence modality embedding is used to enhance the feature representation of the sequence modality;

[0155] Specifically, for the sequence standardization feature, the original sequence standardization feature H seq and the corresponding sequence modal embedding E seq Add them together to get the sequence embedding feature.

[0156] According to the original sequence normalization feature H seq , the sequence embedding feature H′ is obtained by the following formula 11 seq :

[0157] Formula 11: H′ seq =H seq +E seq .

[0158] S13012. Based on the modal embedding module, the structural normalized features in the multimodal normalized features are added to the structural modal embedding to obtain the structural embedded features in the multimodal embedded features, wherein the structural modal embedding is used to enhance the feature representation of the structural modality;

[0159] Specifically, for the structural standardization feature, the original structural standardization feature H str and the corresponding structural mode embedding E str Add them together to get the sequence embedding feature.

[0160] According to the original structure normalization feature H str , the structural embedding feature H′ is obtained by the following formula str :

[0161] Formula 12: H′ str =H str +E str .

[0162] S13013. Based on the modal embedding module, the pocket standardized features in the multimodal standardized features are added to the pocket modal embedding to obtain the pocket embedding features in the multimodal embedding features. The pocket modal embedding is used to enhance the feature representation of the pocket modality.

[0163] Specifically, for the pocket normalization feature, the original pocket normalization feature H poc and the corresponding pocket modal embedding E poc Add them together to get the sequence embedding feature.

[0164] According to the original pocket normalization feature Hpoc, the pocket embedding feature H′ is obtained by the following formula 13: poc :

[0165] Formula 13: H′ poc =H poc +E poc .

[0166] S1302, performing feature fusion processing on the multimodal embedding features based on the encoder to obtain fusion features corresponding to the target protein;

[0167] Since the feature descriptions of sequence embedding features, structural embedding features, and pocket embedding features ignore the interactions and dependencies between different modalities, the global perspective and comprehensive understanding of the characteristics of the target protein are limited. In order to fully integrate the features of all modalities of the target protein, the multimodal embedding features are fused to obtain the fusion features of the target protein that fully integrate the sequence modal protein information, structural modal protein information, and pocket modal protein information of the target protein.

[0168] In some embodiments, multimodal embedding features are fused based on the encoder to obtain fusion features corresponding to the target protein, including: generating a fusion weight for the sequence embedding feature, the structure embedding feature and the pocket embedding feature respectively; aggregating the sequence embedding feature, the structure embedding feature and the pocket embedding feature based on the fusion weights of the embedding features of each modality to obtain an initial fusion feature corresponding to the target protein; the initial fusion feature corresponding to the target protein is input into the encoder to obtain the fusion feature corresponding to the target protein, and the encoder is used to extract feature information in the initial fusion feature.

[0169] In some embodiments, sequence embedding features, structural embedding features, and pocket embedding features are aggregated based on the fusion weights of the embedding features of each modality, including: weighted summing up the sequence embedding features, structural embedding features, and pocket embedding features based on the fusion weights of the embedding features of each modality.

[0170] In some embodiments, the values ​​of the fusion weights corresponding to each modality may be reasonably set by relevant personnel.

[0171] In other embodiments, generating fusion weights corresponding to sequence embedding features, structure embedding features and pocket embedding features, respectively, includes: dynamically generating fusion weights corresponding to sequence embedding features, structure embedding features and pocket embedding features, respectively, according to an adaptive module, wherein the adaptive module is trained based on the quality of the generated molecules (such as binding affinity, chemical stability, etc.).

[0172] By using the method for dynamically adjusting the weights of each modality embedding feature provided in this embodiment, the fusion feature corresponding to the target protein can be more suitable for the actual application scenario.

[0173] S1303, performing feature decoding processing on the fusion feature based on the decoder to obtain a target molecule for the target protein.

[0174] Specifically, the fusion feature is input into a decoder, and the decoder is used to decode the fusion feature to obtain a target molecule of the target protein.

[0175] In some embodiments, the process corresponding to S1302-S1303, i.e., performing feature fusion processing on the multimodal embedded features based on the encoder to obtain the fusion features corresponding to the target protein, and performing feature decoding processing on the fusion features based on the decoder to obtain the target molecule for the target protein, can be specifically expressed by the following formula 14:

[0176] Formula 14:

[0177] in, is the target molecule generated for the target protein, g is a module composed of the encoder 22 and the decoder 23 in the molecule generation module 20, θ g is the learnable parameter in module g.

[0178] Figure 4 A flowchart of a method for training a multimodal language model provided in an embodiment of the present application is shown in FIG. 1 . The method of training a multimodal language model provided in an embodiment of the present application is performed by a computer device. Figure 4 , the method may include the following steps S41 to S44:

[0179] S41, obtaining a multimodal dataset.

[0180] The multimodal dataset includes sequence modal samples, structure modal samples and pocket modal samples of sample proteins corresponding to sample molecules. The multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

[0181] The molecule generation module of the multimodal language model is used to generate the target molecule corresponding to the target protein according to the multimodal standardized features of the target protein. The multimodal language model is trained based on the multimodal dataset. During the training process, the multimodal dataset is divided into a sequence modality sample subset D seq , pocket modal sample subset D poc and the complete modal sample subset D comp .

[0182] In a specific embodiment provided in this application, the sequence modality sample subset D seq Can be derived from the public database BindingDB, D seq The sample data in indicates the sequence modal protein information of the sample protein; the pocket modal sample subset D poc Can be derived from the public database CrossDocked, D poc The sample data in indicates the pocket mode protein information of the sample protein; the complete mode sample subset D comp Can be derived from the public database PDBBind, D comp The sample data in includes sample proteins and the sequence mode protein information, structure mode protein information and pocket mode protein information corresponding to the sample proteins.

[0183] Among them, BindingDB is a comprehensive database that provides detailed information on the interaction between small molecules and proteins, including affinity, binding mode, etc.; CrossDocked is a database specifically used to predict protein-ligand binding modes, including more than 22 million docked protein-ligand complexes; PDBBind is a database containing experimentally measured protein-ligand binding data.

[0184] In some embodiments, in order to ensure the independence between the training set and the test set so as to accurately evaluate the performance of the model, the sequence modality sample subset, the pocket modality sample subset and the complete modality sample subset are divided into the test set and the training set respectively.

[0185] Specifically, the pocket modal sample subset or the sequence modal sample subset can be first divided into a training set and a test set, and then the sample proteins in the undivided sample subset whose similarity with the sample proteins in the test set in the divided sample subset is greater than a preset threshold are divided into the test set, wherein the amount of sample data included in the test set can be freely set.

[0186] More specifically, D poc The corresponding database CrossDocked has completed the division of training set and test set. Therefore, in this application, D poc Divided into training set and test set, further, from D seq and D comp Randomly select a preset number of poc The sample proteins in the corresponding test set whose sample protein similarity exceeds the preset threshold are used as the test set, wherein the present application does not limit the specific values ​​of the preset number and the preset threshold. In a specific embodiment provided in the present application, the preset number is set to 100 and the preset threshold is set to 30%.

[0187] The data set partitioning method provided in the above embodiment ensures the independence between the test set and the training set, prevents data leakage (i.e., the training set contains data that is too similar to the test set, causing the model to perform exceptionally well during testing but unable to generalize to new data), and ensures the generalization ability of the model.

[0188] S42, training a feature processing module in a multimodal language model according to the multimodal data set, and determining a modal recovery loss function.

[0189] The feature processing module includes a feature encoding module, a projector and a modal recovery module. The computer device trains one or more of the feature encoding module, the projector and the modal recovery module according to the multimodal data set.

[0190] Figure 5 A schematic diagram of a training process of a modality recovery module provided in an embodiment of the present application, it should be noted that: Figure 5 Only the structure related to the training process of the modal recovery module is shown, and the entire structure of the multimodal language model is not shown.

[0191] In some embodiments, in S42, a feature processing module in a multimodal language model is trained according to a multimodal data set to determine a modality recovery loss function, including the following S421-S423:

[0192] S421, based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modality sample subset to obtain the multimodal standardized feature true value corresponding to the sample protein,

[0193] Among them, the multimodal standardized feature true values ​​include sequence standardized feature true values, structure standardized feature true values ​​and pocket standardized feature true values.

[0194] Specifically, since the complete modal sample subset includes information corresponding to each modality of the sample protein, the sequence modal protein information, pocket module protein information and structural modal protein information of the sample protein in the complete modal sample subset are feature processed respectively through the feature encoding module and the projector, so as to obtain the true value of the multimodal standardized feature corresponding to the sample protein.

[0195] In some embodiments, in S421, feature processing is performed on the complete modality sample subset based on the feature encoding module and the projector in the feature processing module to obtain the multimodal standardized feature true value corresponding to the sample protein, including the following S4211-S4212:

[0196] S4211, performing feature encoding processing on the complete modal sample subset based on the feature encoding module to obtain the known modal source feature true value corresponding to the sample protein, the known modal source feature true value includes the sequence source feature true value, the structure source feature true value and the pocket source feature true value;

[0197] Specifically, for each sample data indicating sequence modal protein information, pocket modal protein information, and structural modal protein information in the complete modal sample subset, each sample data is input into a feature encoding module to obtain the true value of the known modal source feature corresponding to the sample protein.

[0198] See also Figure 5 , using f in the feature encoding module seq 、f str 、f poc By processing the sequence modal protein information, the structure modal protein information and the pocket modal protein information respectively, the true value of the sequence source feature, the true value of the structure source feature and the true value of the pocket source feature can be obtained respectively, and the true value of the sequence source feature, the true value of the structure source feature and the true value of the pocket source feature are further used as the input of the projector, wherein f in the feature encoding module seq 、f str 、f poc In frozen state.

[0199] Among them, the specific method for the feature encoding module to perform feature processing on the sample data indicating each sequence modal protein information, pocket modal protein information, and structural modal protein information is the same as the process of the above-mentioned step S1201, and will not be repeated here.

[0200] S4212. Based on the projector, each source feature true value in the known modal source feature true value is standardized to obtain the multimodal standardized feature true value corresponding to the sample protein.

[0201] Specifically, for each known modal source feature true value, each known modal source feature true value is input into the projector to obtain the multimodal standardized feature true value corresponding to the sample protein.

[0202] See also Figure 5 The projector includes a sub-projector for processing the true value of the sequence source feature, a sub-projector for processing the true value of the structure source feature, and a sub-projector for processing the true value of the pocket source feature.

[0203] The f in the feature encoding module seq 、f str 、f poc The output data sequence source feature true value, structure source feature true value and pocket source feature true value are respectively input into the sub-projector for processing the sequence source feature true value, the sub-projector for processing the structure source feature true value and the sub-projector for processing the pocket source feature true value to obtain the sequence standardized feature true value, the structure standardized feature true value and the pocket standardized feature true value.

[0204] Among them, the specific method for the projector to standardize the true value of each known modal source feature is the same as the process of the above step S1202, which will not be repeated here.

[0205] S422. Based on the feature encoding module, projector and modal recovery module in the feature processing module, feature processing is performed on the sequence modal sample subset or the pocket modal sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein.

[0206] Among them, the multimodal standardized feature prediction value includes the sequence standardized feature prediction value, the structure standardized feature prediction value and the pocket standardized feature prediction value.

[0207] In some embodiments, in S422, based on the feature encoding module, the projector and the modality recovery module in the feature processing module, feature processing is performed on the sequence modality sample subset or the pocket modality sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, including the following S4221-S4222:

[0208] S4221. Based on the feature encoding module, feature encoding processing is performed on the sequence modal sample subset to obtain the sequence source feature true value corresponding to the sample protein; based on the projector, the sequence source feature true value is standardized to obtain the sequence standardized feature true value corresponding to the sample protein; and based on the modal recovery module, the sequence standardized feature true value and the second target modal prompt information are modally restored to obtain the structure standardized feature prediction value and the pocket standardized feature prediction value corresponding to the sample protein, and the second target modal prompt information is used to prompt the recovery of the structure modality and the pocket modality; or

[0209] S4222. Based on the feature encoding module, feature encoding processing is performed on the pocket modal sample subset to obtain the pocket source feature true value corresponding to the sample protein; based on the projector, the pocket source feature true value is standardized to obtain the pocket standardized feature true value corresponding to the sample protein; and based on the modal recovery module, modal recovery processing is performed on the pocket standardized feature true value and the third target modal prompt information to obtain the sequence standardized feature prediction value and the structure standardized feature prediction value corresponding to the sample protein, and the third target modal prompt information is used to prompt the recovery of the sequence modality and the structure modality.

[0210] See also Figure 5 , the sequence modal sample subset can be input into f seq , get the true value of the sequence source feature corresponding to the sample protein, then input the true value of the sequence source feature into the projector, get the true value of the sequence standardization feature corresponding to the sample protein, further input the true value of the sequence standardization feature into the modal recovery module, get the predicted value of the structure standardization feature and the predicted value of the pocket standardization feature. Similarly, the pocket modal sample subset can be input into f poc , the pocket source feature true value corresponding to the sample protein is obtained, and then the pocket source feature true value is input into the projector to obtain the pocket standardized feature true value corresponding to the sample protein, and the pocket standardized feature true value is further input into the modal recovery module to obtain the sequence standardized feature prediction value and the structure standardized feature prediction value.

[0211] Among them, the data processing process of the feature encoding module, projector and modal recovery module is the same as the above-mentioned S1201-S1205 process, which will not be repeated here.

[0212] S423. Calculate a modal recovery loss function according to the difference between the true value of the multimodal standardized feature and the predicted value of the multimodal standardized feature.

[0213] See also Figure 5 , according to the difference between the true value of the multimodal standardized feature and the predicted value of the multimodal standardized feature, the modal recovery loss function is calculated, including:

[0214] The difference between the true value of the sequence modal standardized feature and the predicted value of the sequence modal standardized feature is compared to obtain the sequence modal recovery loss. The difference between the true value of the structural modal standardized feature and the predicted value of the structural modal standardized feature is compared to obtain the structural modal recovery loss. The difference between the true value of the pocket modal standardized feature and the predicted value of the pocket modal standardized feature is compared to obtain the pocket modal recovery loss. The modal recovery loss is determined according to the sequence modal recovery loss, the structural modal recovery loss and the pocket modal recovery loss.

[0215] Among them, in addition to calculating the modal recovery loss using the method provided in the embodiment of the present application, other loss calculation methods can also be used to calculate the modal recovery loss, and the present application does not impose any restrictions on this.

[0216] In a specific embodiment provided in this application, the following formula 15 can be used as the recovery loss function:

[0217] Formula 15:

[0218] Among them, L rec represents the recovery loss function, s is the modality type corresponding to the true value of the standardized feature, including sequence modality and pocket modality, A is a set consisting of sequence modality, pocket modality and structural modality, t represents other modality types in set A except modality s, H t represents the true value of the standardized feature corresponding to mode t, Represents the standardized feature prediction value corresponding to mode t recovered from the sample protein information corresponding to mode s.

[0219] S43, training a molecule generation module in a multimodal language model according to the multimodal data set, and determining a molecule generation loss function.

[0220] In some embodiments, in S43, a molecule generation module in a multimodal language model is trained according to the multimodal data set to determine a molecule generation loss function, including the following S431-S434:

[0221] S431, performing feature processing on the complete modality sample subset based on the feature encoding module and the projector in the feature processing module to obtain multimodal standardized feature true values ​​corresponding to the sample protein, the multimodal standardized feature true values ​​including sequence standardized feature true values, structure standardized feature true values ​​and pocket standardized feature true values;

[0222] Figure 6 The schematic diagram of the training process of a projector and a molecule generation module provided in the embodiment of the present application needs to be explained. Figure 6 Only the structures related to the training process of the projector and molecule generation module are shown, and the entire structure of the multimodal language model is not shown.

[0223] See also Figure 6 The feature encoding module and the projector in the feature processing module process the complete modal sample subset including the sequence modal sample data, the structure modal sample data and the pocket modal sample data in turn, and obtain the multimodal standardized feature true value including the sequence standardized feature true value, the structure standardized feature true value and the pocket standardized feature true value. The specific implementation method of this step is the same as the corresponding process of S1206 and S1207 above, which will not be repeated here.

[0224] S432, obtaining the real token at position i in the real embedded string sequence of the sample molecule;

[0225] In the embodiment of the present application, the real embedded string sequence corresponding to a sample molecule can be represented by M, where M=(S1, ..S Nmol ), S i is the real token at position i. Then, the real token at position i in the real embedded string sequence of the sample molecule is obtained, that is, the i-th token in the real embedded string sequence M corresponding to the sample molecule is obtained.

[0226] S433, given all tokens before position i in the real embedded string sequence, perform molecule generation processing on the real value of the multimodal standardized feature based on the molecule generation module to obtain the predicted token at position i in the predicted embedded string sequence of the sample molecule;

[0227] Specifically, given all tokens before position i in the real embedded string sequence, that is, given a prefix M<i, the prefix M<i is input into the model. Further, the true value of the multimodal standardized feature is input into the molecule generation module for molecule generation processing. The molecule generation module calculates the predicted token at position i in the predicted embedded string sequence of the sample molecule according to the current parameters of the molecule generation module, the given prefix M<i and the true value of the multimodal standardized feature.

[0228] S434. Calculate the molecule generation loss function based on the difference between the real token at position i in the real embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

[0229] See also Figure 6 , respectively add the true value of the sequence modal standardized feature to the sequence modal embedding, add the true value of the structural modal standardized feature to the structural modal embedding, and add the true value of the pocket modal standardized feature to the pocket modal embedding, and further input the obtained sequence embedding features, structural embedding features, and pocket embedding features into the encoder and decoder in turn to obtain the predicted embedded string sequence corresponding to the sample molecule.

[0230] By quantifying the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence, the molecule generation loss function is calculated, which can drive the molecule generation module to improve its prediction ability during the training process.

[0231] Among them, the difference between the real token at position i in the real embedded string sequence and the predicted token at position i in the predicted embedded string sequence, the method for calculating the molecule generation loss can use other methods besides the method provided in this application.

[0232] In a specific embodiment provided in the present application, the following formula 16 is used to calculate the molecular generation loss:

[0233] Formula 16: L gen =-∑ i logP(s i |M ＜i );

[0234] Among them, s i is the real token at position i in the real embedded string sequence, M ＜i represents all token sequences that have been generated before position i of the real embedded string sequence, P(s i |M ＜i ) is the molecular generation module in the generated sequence M ＜i Generate tokens under the condition i probability.

[0235] By minimizing the above molecule generation loss function, the molecule generation module can gradually learn how to predict the next token based on the previous token, and finally generate a complete prediction embedding string sequence.

[0236] In some other embodiments, in S43, a molecule generation module in a multimodal language model is trained according to the multimodal data set to determine a molecule generation loss function, including the following S435-S438:

[0237] S435, based on the feature encoding module, the projector and the modal recovery module in the feature processing module, feature processing is performed on the sequence modal sample subset or the pocket modal sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, where the multimodal standardized feature prediction value includes a sequence standardized feature prediction value, a structure standardized feature prediction value and a pocket standardized feature prediction value;

[0238] See also Figure 6The data input to the feature processing module may only include sequence modal sample data in the sequence modal sample subset and pocket modal sample data in the pocket modal sample subset. After the sequence modal sample subset or the pocket modal sample subset is processed by the feature encoding module, the projector and the modal recovery module in turn, and then processed by the missing modal recovery module, the multimodal standardized feature prediction value corresponding to the sample protein including the sequence standardized feature prediction value, the structure standardized feature prediction value and the pocket standardized feature prediction value can be obtained.

[0239] Among them, the specific implementation method of obtaining the multimodal standardized feature prediction value corresponding to the sample protein according to the sequence modality sample subset or the pocket modality sample subset is the same as the process corresponding to the above S1201-S1205, which will not be repeated here.

[0240] S436, obtaining the real token at position i in the real embedded string sequence of the sample molecule;

[0241] Get the real token S at position i in the real embedding string sequence M of the sample molecule i .

[0242] S437, given all tokens before position i in the real embedded string sequence, perform molecule generation processing on the real value of the multimodal standardized feature based on the molecule generation module to obtain the predicted token at position i in the predicted embedded string sequence of the sample molecule;

[0243] Specifically, given all tokens before position i in the real embedded string sequence, that is, given prefix M ＜i , prefix M ＜i Input model, further, the true value of the multimodal standardized feature is input into the molecule generation module for molecule generation processing, and the molecule generation module generates the molecule according to the current parameters of the molecule generation module and the given prefix M ＜i and the true value of the multimodal normalized features, and the predicted token at position i in the predicted embedding string sequence of the sample molecule is calculated.

[0244] S438. Calculate the molecule generation loss function based on the difference between the real token at position i in the real embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

[0245] By quantifying the difference between the real token at position i in the real embedded string sequence and the predicted token at position i in the predicted embedded string sequence, the molecule generation loss function is calculated, which can drive the molecule generation module to improve its prediction ability during the training process. The specific calculation method of the molecule generation loss can be found in the above formula 16.

[0246] S44, updating the model parameters of the multimodal language model based on the modal recovery loss function and the molecule generation loss function to obtain a trained multimodal language model.

[0247] In some embodiments, when training a multimodal language model, two loss functions, namely a modality recovery loss function and a molecule generation loss function, are jointly optimized, that is, all learnable parameters in the multimodal language model are adjusted simultaneously.

[0248] In some embodiments, the modal recovery loss function and the molecule generation loss function are jointly optimized, which may specifically include:

[0249] A weight is assigned to each of the modal recovery loss function and the molecule generation loss function, and the weighted sum of the modal recovery loss function and the molecule generation loss function is taken as the total loss. By minimizing the total loss, the trained multimodal language model is obtained, wherein the modal recovery loss function and the molecule generation loss function can be reasonably set according to task requirements.

[0250] In some embodiments, the model parameters include projection parameters of the projector, modal recovery parameters of the modal recovery module, and molecule generation parameters of the molecule generation module; in S44, when the model parameters of the multimodal language model are updated based on the modal recovery loss function and the molecule generation loss function, the multimodal language model is trained by a staged training method, that is, each module to be trained in the multimodal language model is trained in stages, specifically including the following S441-S442:

[0251] S441, updating the modal recovery parameters of the modal recovery module based on the modal recovery loss function;

[0252] See also Figure 5 , Figure 5 The true value of the sequence standardized feature, the true value of the structure standardized feature and the true value of the pocket standardized feature correspond to the true value of the multimodal standardized feature, the sequence standardized feature prediction value, the structure standardized feature prediction value and the pocket standardized feature prediction value correspond to the multimodal standardized feature prediction value, the sequence modal recovery loss, the structure modal recovery loss and the pocket modal recovery loss correspond to the recovery loss function. It should be noted that Figure 5 The sequence modal recovery loss, structural modal recovery loss and pocket modal recovery loss are displayed in parallel, which is only to illustrate the determination principle of each recovery loss, and does not mean that each recovery loss is generated in parallel or at the same time.

[0253] Specifically, the modal recovery parameters of the modal recovery module (i.e., the learnable parameters θ in the modal recovery module in the above embodiment) are updated based on the modal recovery loss. rec), the projection parameters of the projector (i.e., the learnable parameters θ of the projector in the above embodiment) proj ), and the molecular generation parameters of the molecular generation module (i.e., the learnable parameters θ in the module g composed of the encoder and the decoder in the above embodiment) g ) are in a frozen state (i.e. unchanged).

[0254] In some embodiments, updating the modal restoration parameters of the modal restoration module based on the modal restoration loss function includes:

[0255] Through the back-propagation algorithm, the loss value calculated according to the modal recovery loss function is back-propagated along the network structure of the modal recovery module to calculate θ rec The gradient of each parameter in , and get the gradient information;

[0256] Use the optimization algorithm to update the parameters θ of the modal recovery module according to the gradient information rec ; The optimization algorithm may be a gradient descent method, an Adam optimizer or other optimization algorithms;

[0257] Repeat the above process until the loss value converges or the preset number of iterations is reached.

[0258] S442. Update the projection parameters of the projector and the molecule generation parameters of the molecule generation module based on the molecule generation loss function.

[0259] See also Figure 6 :Update the projection parameter θ of the projector according to the molecular generation loss function proj , and the molecular generation parameter θ of the molecular generation module g When the modal recovery parameter θ of the modal recovery module is rec In frozen state.

[0260] In some embodiments, updating the projection parameters of the projector and the molecule generation parameters of the molecule generation module based on the molecule generation loss function includes:

[0261] Through the back-propagation algorithm, the loss value calculated according to the molecular generation loss function is back-propagated along the molecular generation module and the network structure of the projector to calculate θ g and θ proj The gradient of each parameter in , and get the gradient information;

[0262] Use the optimization algorithm to update the parameters θ of the molecule generation module according to the gradient information g and the projector parameters θ proj ; The optimization algorithm may be a gradient descent method, an Adam optimizer or other optimization algorithms;

[0263] Repeat the above process until the loss value converges or the preset number of iterations is reached.

[0264] In some embodiments, during the training process of the multimodal language model, a dynamic generation method can also be introduced during the training process of the multimodal language model. Specifically: during the molecular generation process, the intermediate state of the molecule is detected in real time to determine whether the currently generated molecular fragment is stable and whether it meets the drug activity requirements; further, the dynamic adjustment module adjusts the generation strategy according to the intermediate state information detected in real time and the current intermediate state information; and the subsequent generation steps are corrected according to the evaluation results of the intermediate state information through the feedback mechanism.

[0265] In addition to training the multimodal language model with the method provided in the above embodiment, the multimodal language model can also be trained based on the reinforcement learning strategy. Specifically, the generated target molecule is regarded as the "action" in reinforcement learning (RL), and the environment is the binding situation of protein and molecule; in the design of the reward mechanism, a reward function based on the characteristics of molecular drugs is designed, for example, it can be a reward function based on binding affinity, toxicity prediction, synthesizability or drug metabolism and kinetics (ADME / T) properties, which is used to evaluate the generated target molecules; in strategy learning, the strategy gradient method or Q learning and its variants are used to optimize the generation strategy. During the training process, the agent gradually learns how to generate high-quality molecules that meet the specific drug characteristics by continuously generating molecules, observing environmental feedback (i.e., rewards) and adjusting its generation strategy accordingly; as the training proceeds, the multimodal language model can gradually accumulate knowledge about effective molecular structures and further improve the quality of generated molecules through iterative optimization.

[0266] In some embodiments, a multi-task learning method can also be used to generate target molecules for target proteins, specifically:

[0267] First, define multiple related tasks;

[0268] For example, protein function prediction, the purpose of this task is to predict the functional category or specific function of a given target protein, and molecule generation.

[0269] Furthermore, the model architecture is constructed, including shared layers and task-specific layers;

[0270] Among them, the first few layers of the model (such as convolutional layers, recurrent layers or graph neural network layers) can be designed as shared layers, which are used to learn common feature representations between two tasks. The features include basic properties of chemical structures, conservative patterns of protein sequences, etc. After the shared layers, dedicated network layers are designed for each task, and the network layers corresponding to each specific task are used to learn task-specific features. For example, the fully connected layer or the attention mechanism layer can be used as the network layer corresponding to the target protein function prediction, and variants of generative adversarial networks (GANs), variational autoencoders (VAEs) or graph neural networks (GNNs) can be designed as the network layer corresponding to molecule generation.

[0271] Going further, design a multi-task loss function;

[0272] The multi-task loss function combines the loss values ​​of each task and balances the importance of each task through a weight coefficient. The loss function of each task can be determined according to the specific nature of the task. For example, the classification cross entropy loss is used for protein function prediction, while the loss of the generative model may include reconstruction error, adversarial loss, etc.

[0273] Finally, perform model training;

[0274] Model training specifically includes: preparing the input data of the two tasks to ensure that they are suitable for their respective input layers; forward propagating the input data through the shared layer and the task-specific layer to calculate the output of each task; calculating the loss value based on the output of each task and the true label; propagating the total loss function back to the network through the backpropagation algorithm to update the parameters of the shared layer and the task-specific layer; repeating the above process until the performance of the model on the validation set is no longer significantly improved or the preset training rounds are reached.

[0275] A molecular generation method and model based on multi-task learning. Specifically, by sharing representation, the model can learn more generalized features, thereby improving the performance on each task. In addition, multi-task learning can be used as a regularization method to reduce the risk of overfitting of the model on a single task by introducing additional tasks. Moreover, training multiple tasks simultaneously can save computing resources and time because the shared layer only needs to be calculated once.

[0276] In some embodiments, a generative adversarial network (GANs) method can also be used to generate target molecules for target proteins, specifically including:

[0277] First, prepare a dataset containing real drug molecule data;

[0278] This dataset should contain a large number of drug molecules with known structures, drug activities, and other relevant properties, which will be used to train the discriminator so that it can learn the characteristics of real molecules.

[0279] Second, design the generator and discriminator;

[0280] The design goal of the generator is to generate new drug molecular structures, and the task of the discriminator is to distinguish whether the input molecule is real or generated by the generator.

[0281] Third, confrontation training;

[0282] During the training process, the generator and the discriminator update their parameters alternately. First, the parameters of the generator are fixed, and the discriminator is trained to distinguish between real molecules and generated molecules. Then, the parameters of the discriminator are fixed, and the generator is trained to generate molecules that can deceive the discriminator. This process is repeated until a certain stopping criterion is reached (such as reaching a certain number of iterations, the quality of the generated molecules meets the requirements, etc.). Among them, the generator and the discriminator have their own loss functions. The loss function of the discriminator usually includes two parts: one is the loss of the real molecules being correctly classified (real molecules should be judged as real), and the other is the loss of the generated molecules being misclassified (generated molecules should be judged as false, but here we hope that it is as close to the real as possible, so we actually hope that the discriminator's judgment on the generated molecules is close to the judgment of the real molecules). The loss function of the generator attempts to minimize the probability that the discriminator judges its generated molecules as false.

[0283] Fourth, evaluation and optimization;

[0284] During and after the training, the generated molecules need to be evaluated. Based on the evaluation results, the structure, parameters, training strategy, etc. of the generator and discriminator can be adjusted.

[0285] Fifth, application;

[0286] Trained GANs can be used to generate a large number of new drug molecule candidates, which can be further used in drug discovery, drug design and other fields.

[0287] In some embodiments, graph neural networks (GNNs) can also be used for molecular generation: first, graph structure data is constructed: the structure of proteins and molecules is represented as a graph, nodes represent atoms or amino acids, and edges represent chemical bonds or interactions between them; then, the features of nodes and edges are initialized: each node and edge is assigned an initial feature vector. These feature vectors can contain information such as the physical and chemical properties of atoms (such as atom type, charge, chirality, etc.) and the type and length of chemical bonds; then, information is passed between nodes in the graph through a message passing mechanism, and the feature representation of the nodes is gradually updated. In this step, each node collects information from its neighbor nodes (nodes connected by edges) and updates its own feature representation; then a multi-layer graph convolutional network is designed to perform layer-by-layer convolution operations on the molecular structure to extract multi-level structural information; after obtaining the high-order representation of the graph, these representations are used for training downstream tasks, such as drug generation or protein function prediction; further, the parameters of the model are continuously adjusted through training data to optimize the structural rationality and drug activity of the generated molecules, so that the generated molecules meet the requirements of drug design.

[0288] All of the above technical solutions can be arbitrarily combined to form optional embodiments of the present application, which will not be described in detail here.

[0289] Figures 7 to 11 All of them are experimental result diagrams of the multimodal language model MmDrugGen provided in the embodiments of the present application. Figures 7 to 11 It can be concluded that the multimodal language model provided by this application shows excellent performance. For example, Figure 7 This is a comparison chart of the experimental results of the multimodal language model provided in the embodiment of the present application and other models on the PDBBind database, see Figure 7 In the experiment of the public database PDBBind, compared with the molecular generation methods based on sequence modal protein information such as EncDec, BioT5, DrugGPT in the related art or the molecular generation methods based on pocket modal protein information such as TargetDiff, IPDIFF, MolCRAFT, the multimodal language model provided in this application has better binding ability index (Vina Score), high affinity index (High Affinity), drug similarity score (QED), synthetic accessibility score (SA), success rate and other indicators, showing a significant improvement; Figure 8 This is a comparison chart of the experimental results of the multimodal language model provided in the embodiment of the present application and other models on the CrossDocked database, see Figure 8 In the experiment of the public database CrossDocked, the success rate of the multimodal language model provided by this application reached 46.1%, while the traditional model was only 25%; Fig. 9This is a comparison chart of the experimental results of the multimodal language model provided in the embodiment of the present application and other models on the BindingDB database, see Fig. 9 In the experiment of the public database BindingDB, the average value of the binding ability index of the multimodal language model provided by this application is -8.72, which is a significant improvement compared with the score of -8.08 of the baseline model. At the same time, the High Affinity reaches 57.9%, which is significantly higher than the 45.2% of the baseline model. Fig.10 The experimental results of the ablation experiment of the multimodal language model provided in the embodiment of the present application on the PDB-Bind database are shown in FIG. Fig.10 ,Compared with using one modality alone, combining the three modalities, sequence, structure, and pocket, can significantly improve the performance. Fig.11 The experimental results of the ablation experiment on the BindingDB database and the CrossDocked database for the multimodal language model provided in the embodiment of the present application and the multimodal language model without the modal recovery module are shown in FIG. Fig.11 , the molecule generation effect of MmDrugGen without the modal recovery module on the BindingDB database and the CrossDocked database is significantly reduced compared with MmDrugGen.

[0290] Joint training on multimodal datasets significantly improved the performance of the model. In contrast, joint training on multiple datasets of a single modality failed to bring the same improvement, indicating that multimodal fusion is crucial to prevent the model from overfitting to a single modality, thereby improving the effectiveness and compatibility of drug design.

[0291] MmDrugGen uses a modal recovery module to perform modal recovery and an encoder to perform multimodal feature fusion, which significantly improves the diversity and innovation of the generated drug molecules. In experiments, the generated molecular structures are more diverse and have richer chemical properties than traditional methods. In practical applications, this means that drug molecules generated by MmDrugGen can cover more chemical space and increase the possibility of new drug discovery.

[0292] The embodiment of the present application obtains known modal protein information corresponding to the target protein, wherein the known modalities of the known modal protein information include at least one modality in a modal set, and the modal set includes sequence modalities, structural modalities, and pocket modalities; a feature processing module based on a trained multimodal language model performs feature processing on the known modal protein information to obtain multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structural standardized features, and pocket standardized features having the same dimension; a molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain a target molecule for the target protein; wherein the trained multimodal language model is obtained by training based on a multimodal data set, and the multimodal data set includes sequence modal samples, structural modal samples, and pocket modal samples of the sample protein corresponding to the sample molecule, and the multimodal data set is divided into a sequence modal sample subset, a pocket modal sample subset, and a complete modal sample subset. The embodiment of the present application can generate multimodal standardized features using known modal protein information of at least one modality of the target protein, thereby generating high-quality drug molecules in the absence of certain modal information, and improving the efficiency of drug design. This method reduces the dependence on the complete three-dimensional structure of the protein, thereby speeding up the drug design process. Through the feature processing module of the multimodal language model, the multimodal language model of the present application is trained based on a multimodal data set containing sequence modal samples, structural modal samples, and pocket modal samples. This multimodal training method enables the model to more fully learn the associations and complementary information between different modalities, thereby improving the model's ability to learn the interactions between unknown structural proteins and target molecules.

[0293] In order to better implement the molecule generation method of the embodiment of the present application, the embodiment of the present application also provides a molecule generation device. Fig.12 , Fig.12 This is a schematic diagram of the structure of a molecule generation device 200 provided in an embodiment of the present application. The molecule generation device 200 may include:

[0294] An acquisition unit 210 is used to acquire known modality protein information corresponding to the target protein, wherein the known modality of the known modality protein information includes at least one modality in a modality set, and the modality set includes a sequence modality, a structural modality, and a pocket modality;

[0295] A feature processing unit 220 is used to perform feature processing on known modal protein information based on the feature processing module of the trained multimodal language model to obtain multimodal standardized features corresponding to the target protein, where the multimodal standardized features include sequence standardized features, structure standardized features, and pocket standardized features having the same dimensions;

[0296] A molecule generation unit 230 is used to perform molecule generation processing on the multimodal standardized features based on the molecule generation module of the trained multimodal language model to obtain a target molecule for the target protein;

[0297] Among them, the trained multimodal language model is trained based on a multimodal dataset, which includes sequence modal samples, structural modal samples and pocket modal samples of sample proteins corresponding to sample molecules. The multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

[0298] In some embodiments, the feature processing module includes a feature encoding module, a projector and a modality recovery module. The feature processing unit 220 performs feature processing on known modality protein information based on the feature processing module of the trained multimodal language model to obtain multimodal standardized features corresponding to the target protein, specifically for:

[0299] If the known modality of the known modality protein information includes some modalities in the modality set, determining the missing modality corresponding to the target protein according to the known modality protein information and the modality set;

[0300] Based on the feature encoding module, feature encoding processing is performed on the known modal protein information to obtain the known modal source features corresponding to the target protein, where the known modal source features include one or two of sequence source features, structure source features and pocket source features;

[0301] Based on the projector, the known modality source features are normalized to obtain the known modality normalized features corresponding to the target protein;

[0302] Based on the modality recovery module, modality recovery processing is performed on the known modality standardized features and the first target modality prompt information to obtain the missing modality recovery standardized features corresponding to the target protein, and the first target modality prompt information is used to prompt the recovery of the missing modality;

[0303] According to the known modality normalization features and the missing modality recovery normalization features, the multimodal normalization features corresponding to the target protein are obtained. The multimodal normalization features include sequence normalization features, structure normalization features and pocket normalization features.

[0304] In some embodiments, the feature processing unit 220 is used to perform feature processing on the known modal protein information and the modal set based on the feature processing module of the trained multimodal language model to obtain the multimodal standardized features corresponding to the target protein, and is also used to:

[0305] If the known modalities of the known modal protein information include all the modalities in the modal set, feature encoding processing is performed on the known modal protein information based on the feature encoding module to obtain the known modal source features corresponding to the target protein, and the known modal source features include sequence source features, structure source features and pocket source features;

[0306] Based on the projector, each source feature in the known modal source features is standardized to obtain the multimodal standardized features corresponding to the target protein. The multimodal standardized features include sequence standardized features, structure standardized features and pocket standardized features.

[0307] In some embodiments, the molecule generation module includes a modality embedding module, an encoder, and a decoder. The molecule generation unit 230 is used to perform molecule generation processing on the multimodal standardized features based on the molecule generation module of the trained multimodal language model to obtain a target molecule for the target protein, specifically for:

[0308] Based on the modality embedding module, each standardized feature in the multimodal standardized feature is modally embedded to obtain a multimodal embedding feature, which includes a sequence embedding feature, a structure embedding feature and a pocket embedding feature;

[0309] Based on the encoder, feature fusion processing is performed on the multimodal embedding features to obtain the fusion features corresponding to the target protein;

[0310] The decoder is used to decode the fusion features and obtain target molecules for the target protein.

[0311] In some embodiments, when the molecule processing unit 230 is used to perform modality embedding processing on each standardized feature in the multimodal standardized feature based on the modality embedding module to obtain the multimodal embedding feature, it is specifically used to:

[0312] Based on the modality embedding module, the sequence normalization features in the multimodal normalization features are added to the sequence modality embedding to obtain the sequence embedding features in the multimodal embedding features. The sequence modality embedding is used to enhance the feature representation of the sequence modality.

[0313] Based on the modal embedding module, the structural normalized features in the multimodal normalized features are added to the structural modal embedding to obtain the structural embedded features in the multimodal embedded features. The structural modal embedding is used to enhance the feature representation of the structural modality.

[0314] Based on the modality embedding module, the pocket standardized features in the multimodal standardized features are added to the pocket modality embedding to obtain the pocket embedding features in the multimodal embedding features. The pocket modality embedding is used to enhance the feature representation of the pocket modality.

[0315] In some embodiments, the apparatus further comprises a training unit for:

[0316] A multimodal data set is obtained, the multimodal data set includes sequence modal samples, structure modal samples and pocket modal samples of the sample protein corresponding to the sample molecule, and the multimodal data set is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset;

[0317] Training a feature processing module in a multimodal language model according to a multimodal data set and determining a modality recovery loss function;

[0318] Training a molecule generation module in a multimodal language model according to a multimodal data set and determining a molecule generation loss function;

[0319] Based on the modal recovery loss function and the molecule generation loss function, the model parameters of the multimodal language model are updated to obtain the trained multimodal language model.

[0320] In some embodiments, when the training unit is used to train the feature processing module in the multimodal language model according to the multimodal data set and determine the modality recovery loss function, it is specifically used to:

[0321] Based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modal sample subset to obtain the multimodal standardized feature true value corresponding to the sample protein, and the multimodal standardized feature true value includes the sequence standardized feature true value, the structure standardized feature true value and the pocket standardized feature true value;

[0322] Based on the feature encoding module, the projector and the modal recovery module in the feature processing module, feature processing is performed on the sequence modal sample subset or the pocket modal sample subset to obtain the multimodal standardized feature prediction value corresponding to the sample protein, and the multimodal standardized feature prediction value includes the sequence standardized feature prediction value, the structure standardized feature prediction value and the pocket standardized feature prediction value;

[0323] The modality restoration loss function is calculated based on the difference between the true value of the multimodal normalized features and the predicted value of the multimodal normalized features.

[0324] In some embodiments, when the training unit is used to perform feature processing on the complete modality sample subset based on the feature encoding module and the projector in the feature processing module to obtain the multimodal standardized feature true value corresponding to the sample protein, it is specifically used to:

[0325] Based on the feature encoding module, feature encoding processing is performed on the complete modal sample subset to obtain the known modal source feature true value corresponding to the sample protein, and the known modal source feature true value includes the sequence source feature true value, the structure source feature true value and the pocket source feature true value;

[0326] Based on the projector, each source feature true value in the known modal source feature true value is standardized to obtain the multimodal standardized feature true value corresponding to the sample protein.

[0327] In some embodiments, the training unit is used to perform feature processing on a sequence modality sample subset or a pocket modality sample subset based on a feature encoding module, a projector, and a modality recovery module in a feature processing module to obtain a multimodal standardized feature prediction value corresponding to a sample protein, specifically for:

[0328] Based on the feature encoding module, feature encoding processing is performed on the sequence modal sample subset to obtain the sequence source feature true value corresponding to the sample protein; based on the projector, the sequence source feature true value is standardized to obtain the sequence standardized feature true value corresponding to the sample protein; and based on the modal recovery module, modal recovery processing is performed on the sequence standardized feature true value and the second target modal prompt information to obtain the structure standardized feature prediction value and the pocket standardized feature prediction value corresponding to the sample protein, and the second target modal prompt information is used to prompt the recovery of the structure modality and the pocket modality; or

[0329] Based on the feature encoding module, feature encoding processing is performed on the pocket modal sample subset to obtain the pocket source feature true value corresponding to the sample protein; based on the projector, the pocket source feature true value is standardized to obtain the pocket standardized feature true value corresponding to the sample protein; and based on the modal recovery module, modal recovery processing is performed on the pocket standardized feature true value and the third target modal prompt information to obtain the sequence standardized feature prediction value and the structure standardized feature prediction value corresponding to the sample protein, and the third target modal prompt information is used to prompt the recovery of the sequence modality and the structure modality.

[0330] In some embodiments, when the training unit is used to train the molecule generation module in the multimodal language model according to the multimodal dataset and determine the molecule generation loss function, it is specifically used to:

[0331] Based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modal sample subset to obtain the multimodal standardized feature true value corresponding to the sample protein, and the multimodal standardized feature true value includes the sequence standardized feature true value, the structure standardized feature true value and the pocket standardized feature true value;

[0332] Get the real token at position i in the real embedding string sequence of the sample molecule;

[0333] Given all tokens before position i in the true embedded string sequence, the molecule generation module is used to generate molecules for the true value of the multimodal standardized feature to obtain the predicted token at position i in the predicted embedded string sequence of the sample molecule;

[0334] The numerator generation loss function is calculated based on the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

[0335] In some embodiments, when the training unit is used to train the molecule generation module in the multimodal language model according to the multimodal dataset and determine the molecule generation loss function, it is specifically used to:

[0336] Based on the feature encoding module, the projector and the modal recovery module in the feature processing module, feature processing is performed on the sequence modal sample subset or the pocket modal sample subset to obtain the multimodal standardized feature prediction value corresponding to the sample protein, and the multimodal standardized feature prediction value includes the sequence standardized feature prediction value, the structure standardized feature prediction value and the pocket standardized feature prediction value;

[0337] Get the real token at position i in the real embedding string sequence of the sample molecule;

[0338] Given all tokens before position i in the true embedded string sequence, the molecule generation module is used to generate molecules for the true value of the multimodal standardized feature to obtain the predicted token at position i in the predicted embedded string sequence of the sample molecule;

[0339] The numerator generation loss function is calculated based on the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

[0340] In some embodiments, the model parameters include projection parameters of the projector, modal recovery parameters of the modal recovery module, and molecule generation parameters of the molecule generation module; when the training unit is used to update the model parameters of the multimodal language model based on the modal recovery loss function and the molecule generation loss function, it is specifically used to

[0341] updating a modal restoration parameter of a modal restoration module based on a modal restoration loss function;

[0342] The projection parameters of the projector and the molecule generation parameters of the molecule generation module are updated based on the molecule generation loss function.

[0343] It should be noted that the functions of each unit in the molecule generation device 200 in the embodiment of the present application can correspond to the specific implementation methods of any embodiment in the above-mentioned method embodiments, and will not be repeated here.

[0344] Each unit in the above device can be implemented in whole or in part by software, hardware or a combination thereof. Each unit can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each unit.

[0345] For example, the molecule generating device 200 may be integrated in a terminal or a server that has a storage device and a processor installed therein and has computing capabilities, or the molecule generating device 200 may be the terminal or the server.

[0346] In some embodiments, the present application further provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0347] Fig.13 A schematic structural diagram of a computer device provided in an embodiment of the present application, such as Fig.13 As shown, the computer device 300 may include: a communication interface 301, a memory 302, a processor 303 and a communication bus 304. The communication interface 301, the memory 302, and the processor 303 communicate with each other through the communication bus 304. The communication interface 301 is used for the device 300 to communicate data with external devices. The memory 302 can be used to store software programs and modules, and the processor 303 runs the software programs and modules stored in the memory 302, such as the software programs of the corresponding operations in the aforementioned method embodiments.

[0348] In some embodiments, the processor 303 may call software programs and modules stored in the memory 302 to perform the following operations:

[0349] Acquire known modality protein information corresponding to the target protein, wherein the known modality of the known modality protein information includes at least one modality in a modality set, and the modality set includes sequence modality, structural modality, and pocket modality;

[0350] The feature processing module based on the trained multimodal language model performs feature processing on the known modal protein information to obtain the multimodal standardized features corresponding to the target protein. The multimodal standardized features include sequence standardized features, structure standardized features and pocket standardized features with the same dimensions.

[0351] The molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain the target molecule for the target protein;

[0352] Among them, the trained multimodal language model is trained based on a multimodal dataset, which includes sequence modal samples, structural modal samples and pocket modal samples of sample proteins corresponding to sample molecules. The multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

[0353] In some embodiments, the computer device 300 may be integrated in a terminal or server that has a storage device and a processor installed therein and has computing capabilities, or the computer device 300 is the terminal or server.

[0354] The present application also provides a computer-readable storage medium for storing a computer program. The computer-readable storage medium can be applied to a computer device, and the computer program enables the computer device to execute the corresponding processes in the above-mentioned methods in the embodiments of the present application, which will not be described in detail for the sake of brevity.

[0355] The present application also provides a computer program product, which includes computer instructions, which are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the corresponding processes in the above-mentioned methods in the embodiments of the present application, which will not be described here for the sake of brevity.

[0356] The present application also provides a computer program, which includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the corresponding processes in the above-mentioned methods in the embodiments of the present application, which will not be described here for the sake of brevity.

[0357] It should be understood that the processor of the embodiment of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by the hardware integrated logic circuit or software instructions in the processor. The above processor can be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor are combined and performed. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, and other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0358] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0359] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0360] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0361] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0362] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0363] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0364] In addition, each functional unit in the embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0365] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer or a server) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include various media that can store program codes, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks, or optical disks.

[0366] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for generating a molecule, characterized in that: The method comprises: Acquire known modality protein information corresponding to the target protein, wherein the known modality of the known modality protein information includes at least one modality in a modality set, and the modality set includes sequence modality, structural modality, and pocket modality; Performing feature processing on the known modal protein information based on the feature processing module of the trained multimodal language model to obtain multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structure standardized features, and pocket standardized features having the same dimension; A molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain a target molecule for the target protein; The trained multimodal language model is obtained based on a multimodal dataset, wherein the multimodal dataset includes sequence modal samples, structural modal samples and pocket modal samples of sample proteins corresponding to sample molecules, and the multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

2. The method for generating a molecule according to claim 1, wherein The feature processing module includes a feature encoding module, a projector and a modal recovery module. The feature processing module based on the trained multimodal language model performs feature processing on the known modal protein information to obtain the multimodal standardized features corresponding to the target protein, including: If the known modalities of the known modal protein information include some modalities in the modal set, determining the missing modality corresponding to the target protein according to the known modal protein information and the modal set; Based on the feature encoding module, feature encoding processing is performed on the known modal protein information to obtain known modal source features corresponding to the target protein, wherein the known modal source features include one or two of sequence source features, structure source features and pocket source features; Based on the projector, the known modality source features are standardized to obtain known modality standardized features corresponding to the target protein; Based on the modality recovery module, modality recovery processing is performed on the known modality standardized features and the first target modality prompt information to obtain the missing modality recovery standardized features corresponding to the target protein, and the first target modality prompt information is used to prompt the recovery of the missing modality; According to the known modality normalization feature and the missing modality recovery normalization feature, a multimodality normalization feature corresponding to the target protein is obtained, and the multimodality normalization feature includes a sequence normalization feature, a structure normalization feature and a pocket normalization feature.

3. The method for generating a molecule according to claim 2, wherein: The feature processing module based on the trained multimodal language model performs feature processing on the known modal protein information and the modal set to obtain the multimodal standardized features corresponding to the target protein, and further includes: If the known modalities of the known modal protein information include all modalities in the modality set, then based on the feature encoding module, feature encoding processing is performed on the known modal protein information to obtain known modal source features corresponding to the target protein, wherein the known modal source features include sequence source features, structure source features and pocket source features; Based on the projector, each source feature in the known modal source features is standardized to obtain a multimodal standardized feature corresponding to the target protein, wherein the multimodal standardized feature includes a sequence standardized feature, a structure standardized feature and a pocket standardized feature.

4. The method for generating a molecule according to claim 2, wherein: The molecule generation module includes a modality embedding module, an encoder and a decoder. The molecule generation module based on the trained multimodal language model performs molecule generation processing on the multimodal standardized features to obtain a target molecule for the target protein, including: Based on the modality embedding module, each standardized feature in the multimodal standardized features is subjected to modality embedding processing to obtain a multimodal embedding feature, wherein the multimodal embedding feature includes a sequence embedding feature, a structure embedding feature and a pocket embedding feature; Based on the encoder, feature fusion processing is performed on the multimodal embedding features to obtain fusion features corresponding to the target protein; The fusion feature is subjected to feature decoding processing based on the decoder to obtain a target molecule for the target protein.

5. The method for generating a molecule according to claim 4, wherein: The performing modality embedding processing on each standardized feature in the multimodal standardized features based on the modality embedding module to obtain the multimodal embedding features includes: Based on the modality embedding module, the sequence normalization feature in the multimodal normalization feature is added to the sequence modality embedding to obtain the sequence embedding feature in the multimodal embedding feature, wherein the sequence modality embedding is used to enhance the feature representation of the sequence modality; Based on the modality embedding module, the structural normalized features in the multimodal normalized features are added to the structural modality embedding to obtain the structural embedding features in the multimodal embedding features, wherein the structural modality embedding is used to enhance the feature representation of the structural modality; Based on the modality embedding module, the pocket standardized features in the multimodal standardized features are added to the pocket modality embedding to obtain the pocket embedding features in the multimodal embedding features, and the pocket modality embedding is used to enhance the feature representation of the pocket modality.

6. The method for generating a molecule according to claim 4, wherein: The step of training the multimodal language model comprises: Acquire a multimodal data set, wherein the multimodal data set includes sequence modal samples, structure modal samples, and pocket modal samples of sample proteins corresponding to sample molecules, and the multimodal data set is divided into a sequence modal sample subset, a pocket modal sample subset, and a complete modal sample subset; Training the feature processing module in the multimodal language model according to the multimodal data set to determine a modality recovery loss function; Training the molecule generation module in the multimodal language model according to the multimodal data set to determine a molecule generation loss function; The model parameters of the multimodal language model are updated based on the modality recovery loss function and the molecule generation loss function to obtain the trained multimodal language model.

7. The method for generating a molecule according to claim 6, wherein: The step of training the feature processing module in the multimodal language model according to the multimodal data set to determine a modality recovery loss function includes: Based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modality sample subset to obtain multimodal standardized feature true values ​​corresponding to the sample protein, wherein the multimodal standardized feature true values ​​include sequence standardized feature true values, structure standardized feature true values, and pocket standardized feature true values; Based on the feature encoding module, the projector and the modality recovery module in the feature processing module, feature processing is performed on the sequence modality sample subset or the pocket modality sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, wherein the multimodal standardized feature prediction value includes a sequence standardized feature prediction value, a structure standardized feature prediction value and a pocket standardized feature prediction value; The modality restoration loss function is calculated according to the difference between the true value of the multimodal standardized feature and the predicted value of the multimodal standardized feature.

8. The method for generating a molecule according to claim 7, wherein: The feature encoding module in the feature processing module and the projector perform feature processing on the complete modality sample subset to obtain the multimodal standardized feature true value corresponding to the sample protein, including: Based on the feature encoding module, feature encoding processing is performed on the complete modality sample subset to obtain the known modality source feature true value corresponding to the sample protein, wherein the known modality source feature true value includes the sequence source feature true value, the structure source feature true value and the pocket source feature true value; Based on the projector, each source feature true value in the known modality source feature true value is standardized to obtain a multimodal standardized feature true value corresponding to the sample protein.

9. The method for generating a molecule according to claim 7, wherein: The feature encoding module, the projector and the modality recovery module in the feature processing module perform feature processing on the sequence modality sample subset or the pocket modality sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, including: Based on the feature encoding module, feature encoding processing is performed on the sequence modal sample subset to obtain the sequence source feature true value corresponding to the sample protein; based on the projector, the sequence source feature true value is standardized to obtain the sequence standardized feature true value corresponding to the sample protein; and based on the modality recovery module, modality recovery processing is performed on the sequence standardized feature true value and the second target modality prompt information to obtain the structure standardized feature prediction value and the pocket standardized feature prediction value corresponding to the sample protein, and the second target modality prompt information is used to prompt the recovery of the structure modality and the pocket modality; or Based on the feature encoding module, feature encoding processing is performed on the pocket modal sample subset to obtain the pocket source feature true value corresponding to the sample protein; based on the projector, the pocket source feature true value is standardized to obtain the pocket standardized feature true value corresponding to the sample protein; and based on the modality recovery module, modality recovery processing is performed on the pocket standardized feature true value and the third target modality prompt information to obtain the sequence standardized feature prediction value and the structure standardized feature prediction value corresponding to the sample protein, and the third target modality prompt information is used to prompt the recovery of the sequence modality and the structure modality.

10. The method for generating a molecule according to claim 6, wherein: The step of training the molecule generation module in the multimodal language model according to the multimodal data set to determine a molecule generation loss function includes: Based on the feature encoding module and the projector in the feature processing module, feature processing is performed on the complete modality sample subset to obtain multimodal standardized feature true values ​​corresponding to the sample protein, wherein the multimodal standardized feature true values ​​include sequence standardized feature true values, structure standardized feature true values, and pocket standardized feature true values; Get the real token at position i in the real embedding string sequence of the sample molecule; Given all tokens before position i in the true embedded string sequence, performing molecule generation processing on the true value of the multimodal standardized feature based on the molecule generation module to obtain a predicted token at position i in the predicted embedded string sequence of the sample molecule; The molecule generation loss function is calculated based on the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

11. The method for generating a molecule according to claim 6, wherein: The step of training the molecule generation module in the multimodal language model according to the multimodal data set to determine a molecule generation loss function includes: Based on the feature encoding module, the projector and the modality recovery module in the feature processing module, feature processing is performed on the sequence modality sample subset or the pocket modality sample subset to obtain a multimodal standardized feature prediction value corresponding to the sample protein, wherein the multimodal standardized feature prediction value includes a sequence standardized feature prediction value, a structure standardized feature prediction value and a pocket standardized feature prediction value; Get the real token at position i in the real embedding string sequence of the sample molecule; Given all tokens before position i in the true embedded string sequence, performing molecule generation processing on the true value of the multimodal standardized feature based on the molecule generation module to obtain a predicted token at position i in the predicted embedded string sequence of the sample molecule; The molecule generation loss function is calculated based on the difference between the true token at position i in the true embedded string sequence and the predicted token at position i in the predicted embedded string sequence.

12. The method for generating a molecule according to claim 6, wherein: The model parameters include projection parameters of the projector, modal recovery parameters of the modal recovery module, and molecule generation parameters of the molecule generation module; the updating of the model parameters of the multimodal language model based on the modal recovery loss function and the molecule generation loss function includes: Updating a modal restoration parameter of the modal restoration module based on the modal restoration loss function; The projection parameters of the projector and the molecule generation parameters of the molecule generation module are updated based on the molecule generation loss function.

13. A molecular generation device, characterized in that: The device comprises: An acquisition unit, used for acquiring known modality protein information corresponding to the target protein, wherein the known modality of the known modality protein information includes at least one modality in a modality set, and the modality set includes sequence modality, structural modality and pocket modality; A feature processing unit, configured to perform feature processing on the known modal protein information based on the feature processing module of the trained multimodal language model, and obtain multimodal standardized features corresponding to the target protein, wherein the multimodal standardized features include sequence standardized features, structure standardized features, and pocket standardized features having the same dimension; A molecule generation unit, configured to perform molecule generation processing on the multimodal standardized features based on the molecule generation module of the trained multimodal language model to obtain a target molecule for the target protein; The trained multimodal language model is obtained based on a multimodal dataset, wherein the multimodal dataset includes sequence modal samples, structural modal samples and pocket modal samples of sample proteins corresponding to sample molecules, and the multimodal dataset is divided into a sequence modal sample subset, a pocket modal sample subset and a complete modal sample subset.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the molecule generation method according to any one of claims 1 to 12.

15. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein a computer program is stored in the memory, and the processor is used to execute the molecule generation method according to any one of claims 1 to 12 by calling the computer program stored in the memory.

16. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the molecule generation method according to any one of claims 1 to 12 is implemented.

Images