Multi-modal molecular property prediction method and device, equipment, storage medium and computer program product

Abstract

The invention relates to a multi-mode molecular property prediction method and device, equipment, a storage medium and a computer program product. The method comprises the following steps: inputting a target data set into a molecular representation learning model to obtain one-dimensional sequence features of a target molecule; inputting the target data set into a message passing neural network model to obtain a two-dimensional image feature of the target molecule; feature fusion is carried out according to the one-dimensional sequence features and the two-dimensional image features, drug property prediction is carried out on the target molecule, a drug property prediction result of the target molecule is output, and the drug property prediction result comprises attention weight results of a plurality of visual groups of the target molecule. The attention weight result describes the contribution degree of the corresponding group to the drug property prediction result. By adopting the method, the prediction precision and the prediction reliability of molecular property prediction can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of molecular property prediction, and in particular to a multimodal molecular property prediction method, apparatus, device, storage medium and computer program product. Background Art

[0002] In the long journey of drug development, the absorption, distribution, metabolism, excretion and toxicity (ADMET) of molecules play a vital role.

[0003] In recent years, with the rapid development of artificial intelligence technology, artificial intelligence has been widely used in the field of ADMET property prediction. For artificial intelligence methods, researchers only need to input molecules into the model to achieve rapid prediction of molecular properties. However, most of the existing artificial intelligence-based ADMET property prediction methods only consider single-dimensional features, making it difficult for the model to capture the complex relationship between compound molecules and ADMET properties, resulting in insufficient prediction accuracy.

[0004] In addition, there is a black box problem in the application of artificial intelligence methods, which leads to insufficient prediction reliability. Therefore, how to improve the prediction accuracy and reliability of molecular property prediction is a technical problem that needs to be solved urgently. Summary of the invention

[0005] Based on this, it is necessary to provide a multimodal molecular property prediction method, device, equipment, storage medium and computer program product that can improve the prediction accuracy and prediction reliability of molecular property prediction in response to the above technical problems.

[0006] In a first aspect, the present application provides a method for predicting multimodal molecular properties. The method comprises:

[0007] Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0008] Inputting the target data set into a message passing neural network model to obtain two-dimensional image features of the target molecule;

[0009] Feature fusion is performed based on the one-dimensional sequence features and the two-dimensional image features, and drug property prediction is performed on the target molecule, and a drug property prediction result of the target molecule is output, wherein the drug property prediction result includes attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0010] In one embodiment, the molecular representation learning model includes a self-referential embedding learning neural network model. Inputting the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule includes:

[0011] Converting the simplified molecular linear input specification string of the target molecule into a one-dimensional self-referential embedded string;

[0012] Inputting the self-referential embedded string into the pre-trained self-referential embedding learning neural network model to obtain the one-dimensional sequence features of the target molecule.

[0013] In one embodiment, inputting the self-referential embedded string into the pre-trained self-referential embedding learning neural network model to obtain the one-dimensional sequence features of the target molecule includes:

[0014] Performing word segmentation processing on the self-referential embedded string through the pre-trained self-referential embedding learning neural network model, and splitting the self-referential embedded string into multiple original tokens;

[0015] Randomly selecting a first preset proportion of the original tokens in the first token sequence through the pre-trained self-referential embedding learning neural network model and replacing them with special tokens to generate a masked second token sequence;

[0016] Inputting the masked second token sequence into the encoder through the pre-trained self-referential embedding learning neural network model, and predicting at the positions of the special tokens in the second token sequence to obtain the one-dimensional sequence features of the target molecule.

[0017] In one embodiment, inputting the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule includes:

[0018] Converting the simplified molecular linear input specification string of the target molecule into a two-dimensional molecular image;

[0019] Extracting the atomic features and bond features of the target molecule according to the two-dimensional molecular image;

[0020] Updating and learning the input atomic features and bond features through the message passing neural network model to obtain the updated two-dimensional image features of the target molecule.

[0021] In one embodiment, updating and learning the input atomic features and bond features through the message passing neural network model includes:

[0022] Updating and learning the combined key feature and its neighboring key features through an attention module and an update function, and summing the key hidden states after message passing to obtain a key feature tensor aggregating neighboring information;

[0023] Aggregating and concatenating the key feature tensor and the atomic features, and then updating and learning through a multi-head atomic attention mechanism to update the atomic feature tensor;

[0024] Outputting the molecular representation obtained according to the specified aggregation method as the two-dimensional image feature of the target molecule.

[0025] In one embodiment, the step of aggregating and concatenating the key feature tensor and the atomic features, and then updating and learning through a multi-head atomic attention mechanism to update the atomic feature tensor includes:

[0026] Adding the inter-atomic adjacency matrix, distance matrix, and Coulomb matrix as bias terms to the weights of different attention heads respectively;

[0027] Wherein, the inter-atomic adjacency matrix describes the connectivity information of atoms;

[0028] The distance matrix describes the topological distance information of atom pairs;

[0029] The Coulomb matrix describes the electrostatic interaction information between atoms.

[0030] In a second aspect, the present application also provides a multi-modal molecular property prediction device. The device includes:

[0031] A first feature extraction module, configured to input a target data set into a molecular representation learning model to obtain one-dimensional sequence features of a target molecule;

[0032] A second feature extraction module, configured to input the target data set into a message passing neural network model to obtain two-dimensional image features of the target molecule;

[0033] A property prediction module, configured to perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the properties of the target molecule, and output a property prediction result of the target molecule. The property prediction result includes attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the property prediction result.

[0034] In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0035] Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0036] Input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule;

[0037] Perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0038] In a fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium has a computer program stored thereon, and when the computer program is executed by a processor, the following steps are implemented:

[0039] Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0040] Input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule;

[0041] Perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0042] In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0043] Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0044] Input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule;

[0045] Perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0046] The embodiments of the present application have the following beneficial effects:

[0047] The multi-modal molecular property prediction method, device, equipment, storage medium and computer program product provided by the embodiments of the present application can perform feature fusion on the features of one-dimensional sequence features and two-dimensional image features and predict the drug properties of the target molecule, so as to realize the rapid and accurate prediction of the ADMET drug properties of the molecule by combining information in different dimensions of the molecule; since the drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, the contribution degree of the corresponding group to the drug property prediction result can be described by the attention weights of different visualized groups, so as to help drug researchers intuitively understand the contribution degree of relevant groups and relevant atoms to the drug properties of the target molecule, improve the interpretability of the model, and thus improve the prediction reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic flowchart of the multi-modal molecular property prediction method in one embodiment;

[0049] Figure 2 It is a schematic flowchart of the specific process of obtaining one-dimensional sequence features in one embodiment;

[0050] Figure 3 It is a schematic architecture diagram of the multi-modal molecular property prediction method in another embodiment;

[0051] Figure 4 It is a schematic flowchart of the specific process of obtaining two-dimensional image features in one embodiment;

[0052] Figure 5 It is a structural block diagram of the multi-modal molecular property device in one embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] The multi-modal molecular property prediction method provided by the embodiments of the present application can be applied to a terminal or a server. Among them, the terminal communicates with the server through a network. The data storage system can store the data that the server needs to process. The data storage system can be integrated on the server, or placed in the cloud or other network servers. Among them, the terminal can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0055] Embodiment 1

[0056] In one embodiment, referring to Figure 1 , a multi-modal molecular property prediction method is provided, including the following steps:

[0057] S1. Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0058] S2. Input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule;

[0059] S3. Perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0060] Specifically, first obtain the target data set required for model construction. For the target data set, it can be input into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule. Then input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule. In this way, the updated and learned different-dimensional features of the one-dimensional sequence features of the target molecule and the two-dimensional image features of the target molecule are obtained. Then, the one-dimensional sequence features and the two-dimensional image features are subjected to feature fusion and the drug properties of the target molecule are predicted, so as to be able to jointly realize the rapid and accurate prediction of the ADMET drug properties of the molecule by using information of different dimensions of the molecule. At the same time, since the drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, the contribution degree of the corresponding groups to the drug property prediction result can be described by the attention weights of different visualized groups, so as to help drug researchers intuitively understand the contribution degree of relevant groups and relevant atoms to the drug properties of the target molecule, improve the interpretability of the model, and thus improve the prediction reliability of the model.

[0061] Exemplarily, during the construction of the target dataset, a dataset covering 54 ADMET properties can be obtained from sources such as the molecular machine learning benchmark dataset MoleculeNet and the dataset Tox21, and the collected data can be cleaned using the cheminformatics toolkit RDKit. The target dataset can be divided into a training set, a validation set, and a test set. Among them, the dataset division ratio can be 6:2:2, or 7:1:2, 8:1:1, etc. Specifically, it can be determined according to different data magnitudes or different training division scenarios. It should be noted that the above division ratios are only for exemplary illustration and are not actually limited.

[0062] Exemplarily, the 54 molecular ADMET properties are: Bio20, Bio50, Caco2, HIA, MDCK, PGB-inhibitor, PGB-substrate, BBB, FU, PPBR, VD, CYP1A2-inhibitor, CYP1A2-substrate, CYP2A6-substrate, CYP2B6-substrate, CYP2C19-inhibitor, CYP2C19-substrate, CYP2C8-substrate, CYP2C9-inhibitor, CYP2C9-substrate, CYP2D6-inhibitor, CYP2D6-substrate, CYP2E1-substrate, CYP3A4-inhibitor, CYP3A4-substrate, CL, T1 / 2, Carcinogenicity, Clinical-tox, DILI, EC, EI, FDAMDD, H-HT, hERG, Mutagenicity, Respiratory, ROA, Skin-sensitization, BCF, IGC50, LC50DM, LC50FM, NR-AR, NR-AR-LBD, NR-AhR, NR-Aromatase, NR-ER, NR-ER-LBD, NR-PPAR-gamma, SR-ARE, SR-ATAD5, SR-HSE, SR-MMP, SR-p53.

[0063] Specifically, according to the relevance of tasks, the target dataset can be further divided. Exemplarily, the following seven multi-task datasets can be constructed, namely: Absorption dataset, Distribution dataset, Metabolism dataset, Toxicity dataset, Tox21 dataset, and Regression dataset. Each dataset includes a SMILES (Simplified molecular input line entry system) column, which stores the SMILES strings of molecules, and the remaining columns are the corresponding molecular properties. If there is no experimental data for a certain property of a molecule, it is empty. Exemplarily, for the collected datasets, RDKit can be used to clean them, including the following operations: reading the molecular SMILES strings, deleting the strings that cannot be read as invalid molecules; removing small fragments in the molecules, including solvents or salt ions; removing chiral information and tautomers; standardizing the SMILES strings and converting them into unique Canonical SMILES strings; removing duplicate data based on the standardized SMILES strings. After completing the above data cleaning process, the processed target dataset is obtained.

[0064] In one embodiment, the molecular representation learning model includes a self-referential embedding learning neural network model. Based on this, S1 includes:

[0065] S11: Convert the Simplified molecular input line entry system string of the target molecule into a one-dimensional self-referential embedded string;

[0066] S12: Input the self-referential embedded string into the pre-trained self-referential embedding learning neural network model to obtain the one-dimensional sequence features of the target molecule.

[0067] Specifically, the SMILES string of the target molecule can be converted into a one-dimensional self-referential embedded SELFIES string, and then the SELFIES string is input into the pre-trained self-referential embedding learning neural network SELFormer model to obtain the one-dimensional sequence features of the target molecule and store them. By adopting such a technical solution, the Simplified molecular input line entry system string of the target molecule can be converted into a unified SELFIES string that can be recognized by the pre-trained self-referential embedding learning neural network model, so as to obtain the one-dimensional sequence features of the target molecule, which is convenient for subsequent operations and improves the computational efficiency of the overall model.

[0068] In one embodiment, referring to Figure 2 , S12 includes:

[0069] S121. Perform word segmentation on the self-referential embedded string through a pre-trained self-referential embedding learning neural network model, and split the self-referential embedded string into multiple original tokens;

[0070] S122. Randomly select the original tokens with the first preset ratio in the first token sequence through the pre-trained self-referential embedding learning neural network model, and replace them with special tokens to generate a masked second token sequence;

[0071] S123. Input the masked second token sequence into the encoder through the pre-trained self-referential embedding learning neural network model, and predict the positions of the special tokens in the second token sequence to obtain the one-dimensional sequence features of the target molecule.

[0072] Specifically, referring to Figure 3 , perform word segmentation on the SELFIES string through the pre-trained self-referential embedding learning neural network model SELFormer model, and split the SELFIES string into a series of original tokens (tokens), which will be used as the input of the SELFormer model. The input first token sequence will be embedded into a high-dimensional space to form an embedding vector. The SELFormer model randomly selects the original tokens with the first preset ratio (which can be 15%, 20%, etc., and is not limited here) in the first token sequence, and replaces them with special tokens ( <mask>(token), generate a masked second token sequence. Exemplarily, assume the first token sequence is [t1, t2, t3, t4, t5], and the masked second token sequence may be [t1, <mask>, t3, t4, <mask>. The masked second token sequence is input into the Transformer encoder. The encoder consists of multiple identical layers, each layer containing a self-attention module and a feed-forward neural network, and finally outputs a high-dimensional feature matrix for subsequent tasks. The prediction process can be expressed as:

[0073]

[0074] where is the probability distribution of the predicted tokens, and are the parameters of the classifier, is the softmax function, and the feature vector at each position represents the context representation of the -th token in the input sequence. For the masked positions, the model needs to predict the original tokens based on the context.

[0075] By adopting such a technical solution, it is possible to randomly select the first preset proportion of the original tokens and replace them with special tokens, generate a masked second token sequence, and predict at the positions of the special tokens in the second token sequence, so as to achieve the updated learning of the one-dimensional features of the target molecule, obtain the one-dimensional sequence features of the updated target molecule, and improve the feature accuracy of the one-dimensional sequence features.

[0076] In one embodiment, S2 includes:

[0077] S21. Convert the simplified molecular linear input specification string of the target molecule into a two-dimensional molecular image;

[0078] S22. Extract the atomic features and bond features of the target molecule according to the two-dimensional molecular image;

[0079] S23. Update and learn the input atomic features and bond features through a message passing neural network model to obtain the updated two-dimensional image features of the target molecule.

[0080] Specifically, first, the SMILES string of the target molecule is converted into a two-dimensional molecular graph, and then the atomic features and bond features of the target molecule are extracted based on the two-dimensional molecular image. Exemplarily, the following atomic features and bond features of the target molecule can be calculated with the help of RDKit: a total of 8 types of atomic features and 4 types of bond features are used, where the atomic features include atomic type, number of atomic bonds, atomic charge, atomic chirality, number of hydrogen atoms, atomic hybridization state, atomic aromaticity, and atomic mass; the bond features include chemical bond type, whether it is a conjugated bond, whether it is in a ring, and stereoisomerism information. Finally, the message passing neural network model is used to update and learn the input atomic features and bond features to obtain the updated two-dimensional image features of the target molecule. By adopting such a technical means, the SMILES string of the target molecule can be converted into a two-dimensional molecular graph, the atomic features and bond features of the target molecule can be extracted, and then updated and learned based on the atomic features and bond features of the target molecule, so as to obtain the updated two-dimensional image features of the target molecule, which can improve the feature accuracy of the two-dimensional image features of the target molecule.

[0081] In one embodiment, referring to Figure 4 , S23 includes:

[0082] S231. Update and learn the bond features by combining the neighbor bond features through the attention module and the update function, and sum the bond hidden states after message passing to obtain a bond feature tensor aggregating neighbor information;

[0083] S232. Update and learn the atomic feature tensor through the multi-head atomic attention mechanism after aggregating and splicing the bond feature tensor and the atomic features;

[0084] S233. Output the molecular representation obtained according to the specified aggregation method as the two-dimensional image features of the target molecule.

[0085] Specifically, the message passing network updates the node features of the next layer through the node update function by combining the features of the current layer nodes and the message obtained from the message function. The formula is as follows:

[0086]

[0087] where is the node update function, whose function is to take the original node state and the message as inputs to update and obtain the new node state .

[0088] The readout function of the message passing network can be expressed by the following formula:

[0089]

[0090] wherein is the final output vector, is the readout function, represents the corresponding key, represents the molecular graph.

[0091] When updating and learning the key feature matrix, the attention module used can be expressed by the following formula;

[0092]

[0093] wherein is the query matrix, is the key matrix, is the value matrix, is the hidden dimension.

[0094] When updating and learning the atomic feature matrix, the multi-head atomic attention mechanism used can be expressed by the following formula:

[0095]

[0096]

[0097] wherein represents each attention head, are all learnable weight matrices.

[0098] Exemplarily, a key feature matrix and an atomic feature matrix are initialized. First, the key feature matrix is processed. It updates and learns by combining neighbor key features to obtain a key hidden state that aggregates neighbor information, that is, a key feature tensor that aggregates neighbor information. Then this key feature tensor that aggregates neighbor information is aggregated and concatenated with the atomic features to update the atomic features. Finally, the molecular representation obtained according to the specified aggregation method is output as the two-dimensional image feature of the target molecule. Since the properties exhibited by each atom in the target molecule have a certain correlation with its neighboring atoms and / or bonds, and the message-passing neural network is a general framework for supervised learning of graph-structured data, by updating the molecular features through the message-passing neural network, the correlation between each atom and its neighboring atoms and / or bonds can be better learned, thereby facilitating the improvement of the accuracy of the two-dimensional image features of the target molecule.

[0099] In one embodiment, S232 includes:

[0100] Adding the inter-atomic adjacency matrix, distance matrix, and Coulomb matrix as bias terms to the weights of different attention heads respectively.

[0101] Among them, the inter-atomic adjacency matrix describes the connectivity information of atoms; the distance matrix describes the topological distance information of atom pairs; the Coulomb matrix describes the electrostatic interaction information between atoms.

[0102] Specifically, during the atom feature update learning, the inter-atomic distance matrix, adjacency matrix, and Coulomb matrix are respectively added as bias terms to the weights of different attention heads. Among them The adjacency matrix of the molecule is input to incorporate the connectivity information into the model. The distance matrix is input to introduce the topological distance information of atom pairs. The Coulomb matrix is introduced to consider the electrostatic interaction between atoms. By adopting such a technical solution, different inter-atomic relationships can be taken into account, further improving the accuracy of the two-dimensional image features of the target molecule.

[0103] Specifically, after obtaining the one-dimensional sequence features and two-dimensional image features after update learning, feature fusion is also performed based on the one-dimensional sequence features and two-dimensional image features, and the drug properties of the target molecule are predicted. During feature fusion, the updated one-dimensional sequence features and two-dimensional image features are concatenated according to the first dimension, and then input into a two-layer feedforward layer for property prediction. The binary cross-entropy loss (Binary Cross-Entropy Loss, BCELoss) is used for the classification task of the model, and the mean squared error loss (Mean Squared Error Loss, MSELoss) is used for the regression task, which can be expressed by the following formula:

[0104]

[0105]

[0106] In the formula, is the number of samples. For the classification task, is the true label of the th sample, taking values of 0 or 1; is the predicted probability of the th sample, taking values between 0 and 1. For the regression task, is the true value of the th sample, is the predicted value of the th sample.

[0107] Specifically, the contribution degree of different groups to the prediction result can be finally clarified by visualizing the atomic attention weights. It mainly records the attention weights of each atom and uses the similarity graph module in RDKit to visualize the potential connection between the compound substructure and the prediction result. Visualize the results, and the darker the color area, the greater the influence.

[0108] Exemplarily, the following is an exemplary description of the entire training process:

[0109] First, update and learn the bond feature matrix. Taking the number of chemical bonds as 147 as an example, use a linear layer to perform dimensional transformation on it and map it to a feature with 300 chemical bonds. Use the Linear rectification function (ReLU) activation function to initialize the features of the transformed features to obtain the bond hidden state, and the tensor dimension remains unchanged. A total of 3 rounds of message passing are performed on the initialized features. In each round, the following operations are performed: obtain the messages of neighbors, sum the neighbor messages, obtain the messages of the reverse edges, and then remove the reverse messages to avoid information leakage.

[0110] Exemplarily, for the updated messages, use the attention mechanism to update them again. Use three linear transformation layers to calculate the Query, Key, and Value matrices respectively, with the shape of the number of keys * num_heads * att_size. Then transpose the Query, Key, and Value to adapt to the attention calculation. Multiply each element (Query) by the initialized weight and scale it, and then use Softmax normalization to obtain the weight alpha_weight. Multiply each element (Query) by the normalized weight alpha_weight and sum it to obtain the global Query; process the global Query to match the shape of the Key, then multiply each element (Key) by the processed global Query, multiply the obtained result P by the initialized weight one by one and scale it, and use Softmax normalization to obtain the weight beta_weight. Then multiply the normalized weight beta_weight by P one by one and sum it to obtain the global Key; process the global Key to match the shape of the Value and calculate the interaction between the Key and the Value, and then add the residual link content of the Quey to the result of the interaction. Activate the calculated result using ReLU and perform Dropout processing, where the Dropout value is 0.1. Then transpose the result, reshape it into the form of the number of keys * hidden state, perform a linear transformation again and perform layer normalization, stack the results into a tensor and output.

[0111] Exemplarily, the output after the attention mechanism is concatenated with the original features using a residual link; then the features are linearly mapped and non-linearly transformed using the ReLU activation function followed by Dropout processing. The key hidden state is concatenated with the surrounding atomic features, processed using a linear layer and a non-linear layer and then subjected to Droupout processing. Then all atomic hidden states of the current molecule are extracted and processed using the multi-head atomic attention mechanism.

[0112] Exemplarily, in the multi-head atomic attention mechanism, three linear transformation layers are used to calculate the Query, Key, and Value matrices respectively, with a shape of the number of atoms * num_head * att_size, and they are transposed. The Query is multiplied by the Key to calculate the attention weight att_a_w, and the inter-atomic adjacency matrix is introduced as a bias term into to obtain att_a_w1 and att_a_w2 after normalization; the inter-atomic distance matrix is introduced as a bias term into to obtain att_a_w3 and att_a_w4 after normalization; the inter-atomic Coulomb matrix is introduced as a bias term into to obtain att_a_w5 and att_a_w6 after normalization. The adjusted attention weights are scaled and then the softmax function is applied for normalization to obtain the normalized attention weight att_a_w. Subsequently, the normalized attention weight and the Value are multiplied using matrix multiplication, and after non-linear transformation and Dropout processing, the attention output att_a_h is obtained. The above result is transposed and reshaped into the form of the number of atoms * hidden state and then linearly transformed, and layer normalization is applied using (LayerNorm) and then output. Finally, feature aggregation is performed in different ways and the feature tensor of the molecule is returned.

[0113] Exemplarily, the returned graph features and sequence features are concatenated along the first dimension and fed into a fully connected layer, where there are two linear layers in total. Among them, the first linear layer has an input dimension of 300 and an output dimension of 300, and a Dropout operation is performed. The second linear layer includes a non-linear transformation layer, a Dropout layer, with an input dimension of 300 and an output dimension of 1. Finally, property prediction is performed, and the contribution degree of different groups to the prediction result can be clarified by visualizing the atomic attention weights. The similarity graph module in RDKit can be colored according to the magnitude of different atomic attention weight values, and the larger the value, the darker the color around the corresponding atom; conversely, the lighter the color.

[0114] In this embodiment, it is possible to perform feature fusion on the features of one-dimensional sequence features and two-dimensional image features and predict the drug properties of target molecules, so as to realize the rapid and accurate prediction of the ADMET drug properties of molecules by combining information in different dimensions of the molecules; since the drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, the contribution degree of the corresponding group to the drug property prediction result can be described by the attention weights of different visualized groups, so that it can help drug researchers intuitively understand the contribution degree of relevant groups and relevant atoms to the drug properties of the target molecule, improve the interpretability of the model, and thus improve the prediction reliability of the model.

[0115] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the indications of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or steps in other steps.

[0116] Embodiment 2

[0117] Based on the same inventive concept, the embodiment of the present application also provides a multimodal molecular property prediction device for implementing the multimodal molecular property prediction method involved above. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the multimodal molecular property prediction device provided below can refer to the limitations on the multimodal molecular property prediction method in the above text, and will not be repeated here.

[0118] In one embodiment, as Figure 5 As shown, a multi-modal molecular property prediction device is provided, including: a first feature extraction module, configured to input a target data set into a molecular representation learning model to obtain one-dimensional sequence features of a target molecule; a second feature extraction module, configured to input the target data set into a message passing neural network model to obtain two-dimensional image features of the target molecule; a property prediction module, configured to perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and perform property prediction on the target molecule, and output a property prediction result of the target molecule, where the property prediction result includes attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the property prediction result.

[0119] Further, the molecular representation learning model includes a self-referential embedding learning neural network model. Based on this, the first feature extraction module is further configured to convert a simplified molecular linear input specification string of a target molecule into a one-dimensional self-referential embedded string; and to input the self-referential embedded string into a pre-trained self-referential embedding learning neural network model to obtain one-dimensional sequence features of the target molecule.

[0120] Further, the first feature extraction module is further configured to perform word segmentation processing on the self-referential embedded string through a pre-trained self-referential embedding learning neural network model, and split the self-referential embedded string into multiple original tokens; and to randomly select a first preset proportion of the original tokens in the first token sequence through a pre-trained self-referential embedding learning neural network model and replace them with special tokens to generate a masked second token sequence; and is further configured to input the masked second token sequence into an encoder through a pre-trained self-referential embedding learning neural network model to predict at the positions of the special tokens in the second token sequence to obtain one-dimensional sequence features of the target molecule.

[0121] Further, the second feature extraction module is further configured to convert a simplified molecular linear input specification string of a target molecule into a two-dimensional molecular image; and to extract atomic features and bond features of the target molecule according to the two-dimensional molecular image; and is further configured to perform update learning on the input atomic features and bond features through a message passing neural network model to obtain updated two-dimensional image features of the target molecule.

[0122] Further, the second feature extraction module is further configured to update and learn the key features in combination with neighbor key features through an attention module and an update function, and sum the key hidden states after message passing to obtain a key feature tensor aggregating neighbor information; and is further configured to update and learn the atomic feature tensor through a multi-head atomic attention mechanism after aggregating and splicing the key feature tensor and the atomic feature; and is further configured to output a molecular representation obtained according to a specified aggregation method as the two-dimensional image feature of the target molecule.

[0123] Further, the second feature extraction module is further configured to add an inter-atomic distance matrix, an adjacency matrix, and a Coulomb matrix as bias terms to the weights of different attention heads respectively; wherein, the inter-atomic adjacency matrix describes the connectivity information of atoms; the distance matrix describes the topological distance information of atom pairs; the Coulomb matrix describes the electrostatic interaction information between atoms.

[0124] Each module in the above multi-modal molecular property prediction device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor in the computer device in hardware form or be independent of it, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0125] Embodiment III

[0126] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0127] Input a target data set into a molecular representation learning model to obtain one-dimensional sequence features of a target molecule;

[0128] Input the target data set into a message passing neural network model to obtain two-dimensional image features of the target molecule;

[0129] Perform feature fusion according to the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0130] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0131] Convert the simplified molecular linear input specification string of the target molecule into a one-dimensional self-referential embedded string;

[0132] Input the self-referential embedded string into a pre-trained self-referential embedding learning neural network model to obtain the one-dimensional sequence features of the target molecule.

[0133] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0134] Perform word segmentation processing on the self-referential embedded string through a pre-trained self-referential embedding learning neural network model, and split the self-referential embedded string into multiple original tokens;

[0135] Randomly select a first preset proportion of the original tokens in the first token sequence through a pre-trained self-referential embedding learning neural network model and replace them with special tokens to generate a masked second token sequence;

[0136] Input the masked second token sequence into an encoder through a pre-trained self-referential embedding learning neural network model, predict at the positions of the special tokens in the second token sequence, and obtain the one-dimensional sequence features of the target molecule.

[0137] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0138] Convert the simplified molecular linear input specification string of the target molecule into a two-dimensional molecular image;

[0139] Extract the atomic features and bond features of the target molecule according to the two-dimensional molecular image;

[0140] Perform update learning on the input atomic features and bond features through a message passing neural network model to obtain the updated two-dimensional image features of the target molecule.

[0141] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0142] Update and learn the bond features by combining the neighbor bond features through an attention module and an update function, and sum the bond hidden states after message passing to obtain a bond feature tensor aggregating neighbor information;

[0143] Perform update learning on the aggregated concatenation of the bond feature tensor and the atomic features through a multi-head atomic attention mechanism to update the atomic feature tensor;

[0144] Output the molecular representation obtained according to the specified aggregation method as the two-dimensional image features of the target molecule.

[0145] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0146] The inter-atomic adjacency matrix, distance matrix, and Coulomb matrix are added as bias terms to the weights of different attention heads respectively;

[0147] Among them, the inter-atomic adjacency matrix describes the connectivity information of atoms;

[0148] The distance matrix describes the topological distance information of atom pairs;

[0149] The Coulomb matrix describes the electrostatic interaction information between atoms.

[0150] Example 4

[0151] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0152] Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0153] Input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule;

[0154] Perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0155] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0156] Convert the simplified molecular linear input specification string of the target molecule into a one-dimensional self-referential embedded string;

[0157] Input the self-referential embedded string into the pre-trained self-referential embedding learning neural network model to obtain the one-dimensional sequence features of the target molecule.

[0158] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0159] Perform word segmentation processing on the self-referential embedded string through the pre-trained self-referential embedding learning neural network model, and split the self-referential embedded string into multiple original tokens;

[0160] Randomly select the original tokens with the first preset ratio in the first token sequence through the pre-trained self-referential embedding learning neural network model and replace them with special tokens to generate a masked second token sequence;

[0161] The masked second token sequence is input into an encoder through a pre-trained self-referential embedding learning neural network model to predict at the special token positions in the second token sequence, obtaining one-dimensional sequence features of the target molecule.

[0162] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0163] Convert the simplified molecular linear input specification string of the target molecule into a two-dimensional molecular image;

[0164] Extract the atomic features and bond features of the target molecule according to the two-dimensional molecular image;

[0165] Update and learn the input atomic features and bond features through a message passing neural network model to obtain the updated two-dimensional image features of the target molecule.

[0166] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0167] Update and learn the bond features by combining the bond features with neighbor bond features through an attention module and an update function, and sum the bond hidden states after message passing to obtain a bond feature tensor aggregating neighbor information;

[0168] Update and learn the atomic feature tensor through a multi-head atomic attention mechanism after aggregating and splicing the bond feature tensor with the atomic features;

[0169] Output the molecular representation obtained according to the specified aggregation method as the two-dimensional image features of the target molecule. In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0170] Add the inter-atomic adjacency matrix, distance matrix, and Coulomb matrix as bias terms to the weights of different attention heads respectively;

[0171] Wherein, the inter-atomic adjacency matrix describes the connectivity information of atoms;

[0172] The distance matrix describes the topological distance information of atom pairs;

[0173] The Coulomb matrix describes the electrostatic interaction information between atoms.

[0174] Embodiment Five

[0175] In one embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0176] Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule;

[0177] Input the target data set into the message passing neural network model to obtain the two-dimensional image features of the target molecule;

[0178] Perform feature fusion based on the one-dimensional sequence features and the two-dimensional image features and predict the drug properties of the target molecule, and output the drug property prediction result of the target molecule. The drug property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

[0179] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0180] Convert the simplified molecular linear input specification string of the target molecule into a one-dimensional self-referential embedded string;

[0181] Input the self-referential embedded string into the pre-trained self-referential embedding learning neural network model to obtain the one-dimensional sequence features of the target molecule.

[0182] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0183] Perform word segmentation processing on the self-referential embedded string through the pre-trained self-referential embedding learning neural network model, and split the self-referential embedded string into multiple original tokens;

[0184] Randomly select the original tokens with the first preset ratio in the first token sequence through the pre-trained self-referential embedding learning neural network model and replace them with special tokens to generate a masked second token sequence;

[0185] Input the masked second token sequence into the encoder through the pre-trained self-referential embedding learning neural network model, and predict at the positions of the special tokens in the second token sequence to obtain the one-dimensional sequence features of the target molecule.

[0186] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0187] Convert the simplified molecular linear input specification string of the target molecule into a two-dimensional molecular image;

[0188] Extract the atomic features and bond features of the target molecule according to the two-dimensional molecular image;

[0189] The input atomic features and bond features are updated and learned through a message passing neural network model to obtain the two-dimensional image features of the target molecule after the update and learning.

[0190] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0191] The bond features are combined with neighbor bond features and updated and learned through an attention module and an update function, and the bond hidden states are summed after message passing to obtain a bond feature tensor aggregating neighbor information;

[0192] The atom feature tensor is updated and learned through a multi-head atom attention mechanism after aggregating and splicing the bond feature tensor and the atomic features;

[0193] Output the molecular representation obtained according to the specified aggregation method as the two-dimensional image features of the target molecule. In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0194] The inter-atomic adjacency matrix, distance matrix, and Coulomb matrix are added as bias terms to the weights of different attention heads respectively;

[0195] Among them, the inter-atomic adjacency matrix describes the connectivity information of atoms;

[0196] The distance matrix describes the topological distance information of atom pairs;

[0197] The Coulomb matrix describes the electrostatic interaction information between atoms.

[0198] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0199] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0200] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0201] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.< / mask> < / mask> < / mask>

Claims

1. A multimodal molecular property prediction method, characterized in that: The method comprises: Input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule; Inputting the target data set into a message passing neural network model to obtain two-dimensional image features of the target molecule; Feature fusion is performed based on the one-dimensional sequence features and the two-dimensional image features, and drug property prediction is performed on the target molecule, and a drug property prediction result of the target molecule is output, wherein the drug property prediction result includes attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the drug property prediction result.

2. The method according to claim 1, characterized in that The molecular representation learning model includes a self-reference embedding learning neural network model, and the target data set is input into the molecular representation learning model to obtain a one-dimensional sequence feature of the target molecule, including: Convert the simplified molecular linear input specification string of the target molecule into a one-dimensional self-referencing embedded string; The self-reference embedded string is input into a pre-trained self-reference embedding learning neural network model to obtain a one-dimensional sequence feature of the target molecule.

3. The method according to claim 2, characterized in that The step of inputting the self-reference embedded string into a pre-trained self-reference embedding learning neural network model to obtain a one-dimensional sequence feature of the target molecule includes: Performing word segmentation processing on the self-reference embedded string through a pre-trained self-reference embedding learning neural network model to split the self-reference embedded string into multiple original tokens; Randomly selecting a first preset proportion of original tags in the first tag sequence through a pre-trained self-reference embedding learning neural network model, and replacing them with special tags to generate a masked second tag sequence; The masked second marker sequence is input into an encoder through a pre-trained self-reference embedding learning neural network model, and the special marker position in the second marker sequence is predicted to obtain a one-dimensional sequence feature of the target molecule.

4. The method according to claim 1, characterized in that: The step of inputting the target data set into a message passing neural network model to obtain the two-dimensional image features of the target molecule comprises: Convert a simplified molecular linear input specification string of a target molecule into a two-dimensional molecular image; extracting atomic features and bond features of the target molecule according to the two-dimensional molecular image; The input atomic features and bond features are updated and learned through a message passing neural network model to obtain the two-dimensional image features of the target molecule after updated learning.

5. The method according to claim 4, characterized in that The updating and learning of the input atomic features and the key features by using a message passing neural network model includes: The key feature is combined with the neighbor key feature through the attention module and the update function for update learning, and the key hidden state is summed after the message is passed to obtain the key feature tensor that aggregates the neighbor information; After aggregating and splicing the key feature tensor and the atomic feature, updating and learning are performed through a multi-head atomic attention mechanism to update the atomic feature tensor; The molecular representation obtained according to the specified polymerization method is output as the two-dimensional image feature of the target molecule.

6. The method according to claim 5, characterized in that The updating and learning of the atomic feature tensor is performed through a multi-head atomic attention mechanism after aggregating and splicing the key feature tensor and the atomic feature, and updating the atomic feature tensor includes: The inter-atom adjacency matrix, distance matrix, and Coulomb matrix are added as bias terms to the weights of different attention heads; Wherein, the inter-atom adjacency matrix describes the connectivity information of atoms; The distance matrix describes the topological distance information of the atom pairs; The Coulomb matrix describes the electrostatic interaction information between atoms.

7. A multimodal molecular property prediction device, characterized in that: The device comprises: The first feature extraction module is used to input the target data set into the molecular representation learning model to obtain the one-dimensional sequence features of the target molecule; A second feature extraction module is used to input the target data set into a message passing neural network model to obtain a two-dimensional image feature of the target molecule; A property prediction module is used to perform feature fusion and property prediction on the target molecule based on the one-dimensional sequence features and the two-dimensional image features, and output the property prediction result of the target molecule, wherein the property prediction result includes the attention weight results of multiple visualized groups of the target molecule, and the attention weight results describe the contribution degree of the corresponding groups to the property prediction result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images