Methods, apparatus, equipment and media for training drug target interaction prediction models

Abstract

This invention relates to the field of artificial intelligence technology and discloses a method, apparatus, device, and medium for training a drug-target interaction prediction model. The method includes: inputting source domain training samples into the source feature extraction unit of the source domain prediction model, where the output of the source feature extraction unit is a source domain drug-target interaction vector; inputting target domain training samples into the target feature extraction unit of the target domain prediction model, where the output of the target feature extraction unit is a target domain drug-target interaction vector; inputting the source domain drug-target interaction vector and the target domain drug-target interaction vector into the source classifier unit of the source domain prediction model, respectively; and inputting the target domain drug-target interaction vector into the target classifier unit of the target domain prediction model, thereby training the source domain prediction model and the target domain prediction model. This improves the accuracy of the drug-target interaction prediction model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for training a drug target interaction prediction model. Background Technology

[0002] Drug discovery is the process of identifying new candidate compounds (drug molecules) with potential therapeutic effects. An essential step in drug discovery is the prediction of drug-target interactions (DTIs). Because the efficacy of a drug molecule depends on its affinity for its target protein or receptor, a drug molecule with no interaction or affinity for the target protein will not provide a therapeutic response. For example, in some cases, a developed drug molecule may interact with an unintended protein or receptor, leading to toxicity. Given that experimental determination of drug-target interaction (DTIs) is both time-consuming and resource-intensive, machine learning methods are currently an important tool for assisting in the prediction of drug molecule interactions with target proteins.

[0003] The inventors realized that current methods for training drug target interaction prediction models using machine learning typically rely on a large number of training samples. However, for tasks involving predicting new drug molecules against new target proteins, only a small number of training samples are available. Furthermore, it's difficult to combine multiple drug target interaction (DTI) measurement data from different experimental conditions to create effective training samples. For example, if a deep learning model is trained to predict a drug targeting target A for a specific disease, the same model cannot be used for predicting a drug targeting target B for a new disease, even if closely related to A. It's necessary to start from scratch, collecting a large amount of real-world measurement data for task B as training samples to train an effective prediction model for B. In other words, in the field of targeted drug prediction in AI drug development, a large amount of data from a specific scenario (targeted drug prediction for a specific target) is typically used to train the model; however, when the scenario changes (to a newly discovered disease target), even if they may be related, the model needs to be retrained from scratch. For each targeted drug prediction task, the model needs to be trained from scratch.

[0004] In summary, a limited number of training samples often results in suboptimal accuracy of the drug-target interaction prediction model after training, and the current training methods for prediction models have very limited universality and generalization. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for training a drug target interaction prediction model, in order to solve the technical problem that the accuracy of the prediction model after training is insufficient due to the small amount of training data.

[0006] Firstly, a method for training a drug target interaction prediction model is provided, for training a source domain prediction model and a target domain prediction model, wherein both the source domain prediction model and the target domain prediction model include a feature extraction unit and a classifier unit, and the method includes:

[0007] Obtain source domain training samples and target domain training samples. The source domain training samples include source training drug molecules, source training target proteins, and source labeling tags of the effects of source training drug molecules on source training target proteins. The target domain training samples include target training drug molecules and target training target proteins.

[0008] The source domain training samples are input into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector; the target domain training samples are input into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is the target domain drug-target interaction vector.

[0009] The source domain drug-target interaction vector and the target domain drug-target interaction vector are respectively input into the source classifier unit of the source domain prediction model. The output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein, and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model. The output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0010] The source domain prediction model and the target domain prediction model are trained based on the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source label, the source prediction label, the first prediction label, and the second prediction label.

[0011] In some embodiments, training the source domain prediction model and the target domain prediction model based on the source domain drug-target action vector, the target domain drug-target action vector, the source label, the source prediction label, the first prediction label, and the second prediction label includes:

[0012] Calculate the maximum mean difference loss between the alignment of the source domain drug-target interaction vector and the target domain drug-target interaction vector;

[0013] The comparison loss function of the target classifier unit of the target domain prediction model and the source classifier unit of the source domain prediction model is calculated based on the first predicted label and the second predicted label.

[0014] The classification loss function of the source domain prediction model is calculated based on the source label and the predicted label.

[0015] The source domain prediction model and the target domain prediction model are trained based on the maximum mean difference loss, the contrast loss function, and the classification loss function.

[0016] In some embodiments, obtaining source domain training samples and target domain training samples includes:

[0017] Obtain at least two sets of source domain training samples, each set of source domain training samples corresponds one-to-one with a source domain prediction model; among them, the target proteins of different source domain training samples belong to different categories.

[0018] The step of inputting the source domain training samples into the source domain prediction model's source feature extraction unit includes:

[0019] Each of the source domain training samples is input into the source feature extraction unit of the source domain prediction model corresponding to the source domain training sample.

[0020] In some embodiments, training the source domain prediction model and the target domain prediction model based on the source domain drug-target action vector, the target domain drug-target action vector, the source label, the source prediction label, the first prediction label, and the second prediction label includes:

[0021] Calculate the maximum mean difference loss between the source domain drug-target interaction vector and the target domain drug-target interaction vector output by the source feature extraction unit of each source domain prediction model;

[0022] The source feature extraction unit of the source domain prediction model and the target feature extraction unit of the target domain prediction model are trained by minimizing the maximum mean difference loss.

[0023] In some embodiments, the step of inputting the source domain drug-target interaction vector and the target domain drug-target interaction vector into the source classifier unit of the source domain prediction model includes:

[0024] The source domain drug-target interaction vector corresponding to each source domain training sample is input into the source classifier unit of the source domain prediction model corresponding to the source domain training sample. The source classifier unit outputs the source prediction label of the interaction between the source training drug molecule and the source training target protein.

[0025] The target domain drug-target interaction vector is input into the source classifier unit of each source domain prediction model, and each source classifier unit outputs the first prediction label of the interaction between the target training drug molecule and the target training target protein.

[0026] The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model, and the target classifier unit outputs a second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0027] In some embodiments, training the source domain prediction model and the target domain prediction model based on the source domain drug-target action vector, the target domain drug-target action vector, the source label, the source prediction label, the first prediction label, and the second prediction label includes:

[0028] Based on the first and second predicted labels output by the source classifier units of each source domain prediction model, calculate the comparison loss function between the source classifier units of each source domain prediction model and the target classifier units of the target domain prediction model.

[0029] Based on the source prediction labels output by the source classifier units of each source domain prediction model and the corresponding source label labels of the source domain prediction model, calculate the classification loss function of the source classifier units of the source domain prediction model.

[0030] The source classifier unit of the source domain prediction model and the target classifier unit of the target domain prediction model are trained based on the classification loss function and the contrast loss function.

[0031] In some embodiments, the target feature extraction unit includes a first neural network and a second neural network;

[0032] The step of inputting the target domain training samples into the target domain prediction model's target feature extraction unit includes:

[0033] The target training drug molecule is input into a first neural network, and the target training target protein is input into a second neural network;

[0034] The influence weight of the target training drug molecule on the target training target protein is determined based on the output of each layer of the first neural network and the output of each layer of the second neural network.

[0035] The first neural network outputs a vector sequence of the target training drug molecule based on the action weights, and the second neural network outputs a vector sequence of the target protein target based on the action weights.

[0036] The vector sequence of the target training drug molecule and the vector sequence of the target protein target are concatenated as the output of the target feature extraction unit.

[0037] Secondly, a training device for a drug target interaction prediction model is provided, used to train a source domain prediction model and a target domain prediction model. Both the source domain prediction model and the target domain prediction model include a feature extraction unit and a classifier unit. The device includes:

[0038] The acquisition module is used to acquire source domain training samples and target domain training samples. The source domain training samples include source training drug molecules, source training target proteins, and source labeling tags of the effects of source training drug molecules on source training target proteins. The target domain training samples include target training drug molecules and target training target proteins.

[0039] The feature extraction module is used to input the source domain training samples into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector; and to input the target domain training samples into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is the target domain drug-target interaction vector.

[0040] The label determination module is used to input the source domain drug-target interaction vector and the target domain drug-target interaction vector into the source classifier unit of the source domain prediction model, respectively. The output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model, and the output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0041] The adjustment module is used to train the source domain prediction model and the target domain prediction model based on the source domain drug-target action vector, the target domain drug-target action vector, the source label, the source prediction label, the first prediction label, and the second prediction label.

[0042] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0043] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method.

[0044] In the above-mentioned scheme implemented by the drug-target interaction prediction model training method, device, equipment, and medium, the source domain training samples are input into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit is the source domain drug-target interaction vector; the target domain training samples are input into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is the target domain drug-target interaction vector; the source domain drug-target interaction vector and the target domain drug-target interaction vector are respectively input into the source classifier unit of the source domain prediction model, and the output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively; the target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model, and the output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein; the source domain prediction model and the target domain prediction model are trained based on the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source label, the source prediction label, the first prediction label, and the second prediction label. In this invention, the model obtained through domain adaptation exhibits good generalization ability. When solving drug discovery tasks targeting new targets, it can be directly trained and fine-tuned using a small number of measured sample data points on the new target, leveraging the learning knowledge gained from previous tasks, without having to train from scratch. This avoids training a completely new model from zero when performing drug prediction for each target disease or virus. It allows for end-to-end learning, offering many potential advantages over traditional methods, such as eliminating the need for sequence alignment, ease of operation, and excellent generalization ability. Therefore, it has high potential for DTI prediction tasks. When used for drug prediction tasks targeting multiple diseases or viruses with new protein targets, only a small amount of labeled data is needed to obtain good accuracy results, transforming the original method of manual parameter tuning relying on machine learning engineers and experts into a method applicable to large-scale, replicable industrial expansion. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of an application environment for a prediction model training method according to an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating a prediction model training method in one embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the neighborhood and target domain;

[0049] Figure 4 This is a flowchart illustrating a prediction model training method in one embodiment of the present invention;

[0050] Figure 5 It is a DTI drug target action model;

[0051] Figure 6 It is a protein fragment represented by an amino acid "letter" sequence;

[0052] Figure 7 This is a schematic diagram of the predictive model device in one embodiment of the present invention;

[0053] Figure 8 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0054] Figure 9 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] It is understood that the terms "first" and "second," as well as "source" and "target," as used herein, may be used to describe various elements, but these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this application, a first predicted label may be referred to as a second predicted label, and similarly, a second predicted label may be referred to as a first predicted label.

[0057] Explanation of technical terms:

[0058] Transfer learning, also known as domain adaptation, primarily involves applying knowledge learned in one domain to other related (but not identical) domains to achieve the learning effect required for a target domain task. Traditional supervised learning methods train models using labeled data within a specific domain A. Typically, a reliable and effective network model can be trained when the test data also falls within the same domain. However, when a specific task or label is lacking, the different labels for different tasks often prevent the training of an effective and reliable model. Transfer learning can effectively address this problem to some extent by fully utilizing data from similar domains. In other words, transfer learning attempts to apply knowledge learned from a source domain task to a new target task.

[0059] The basic concepts of transfer learning mainly include the following:

[0060] (1) Field

[0061] A domain refers to the subject of learning. A domain primarily includes the data and the distribution of labels that generated that data. In transfer learning, there are two main types of domains: the source domain and the target domain. The source domain is the domain with knowledge and a large amount of labeled data; it is the object to be transferred. The target domain is the object to which knowledge and labels are ultimately assigned. The transfer task is completed when knowledge is transferred from the source domain to the target domain.

[0062] (2) Task

[0063] A task refers to a learning objective. It mainly includes: labels and the functions corresponding to those labels.

[0064] (3) Transfer learning

[0065] Suppose the source domain D consists of the edge label distribution P(x) and the d-dimensional feature space X, i.e., D = {x, P(x)}, x ∈ X. Transfer learning: Given a labeled source domain... And the learning objective Γ=(y,f(x)),y∈Y, the unlabeled target domain With learning objectives t When D s ≠D t Or Γ s ≠Γ t In this context, the task of transfer learning is to reduce the generalization error of the target domain. The core of transfer learning is to find the similarities between the source and target domains and utilize them appropriately. That is, to learn knowledge from the target domain using existing knowledge.

[0066] (4) Domain Adaptation

[0067] Domain adaptation is a representative method in transfer learning. It is defined as mapping data features from different domains (such as two different datasets) to the same feature space, so that data from other domains can be used to enhance the training of the target domain.

[0068] Domain adaptation is implemented as follows: Given a labeled source domain and an unlabeled target domain, assuming they have the same feature space but different distributed labels, a classifier is learned using labeled source domain data, and then the target domain label is predicted.

[0069] A vivid analogy can be drawn from sentiment analysis of product reviews, considering two different product domains: books and furniture. In the book domain, words like "good paper quality" and "clear font" typically express positive sentiment, while in the furniture domain, words like "sturdy" and "moderate weight" convey positive sentiment. This demonstrates that while knowledge from different domains can be transferred in this task, the lack of overlap between sentiment terms, the existence of domain-specific terms, and the significant differences in the frequency of these terms across domains lead to challenges such as mismatched label distribution between domains.

[0070] The source domain and target domain share the same features and categories, but their feature distributions differ. The challenge is to leverage information-rich source domain samples to improve the performance of the target domain model. The source domain represents a different domain from the test samples and has rich supervised annotation information; the target domain represents the domain of the test samples and has no labels or only a few labels. The source and target domains often belong to the same task category but have different distributions.

[0071] In the case of multi-source domains, is it sufficient to simply train a model by mixing multiple domains (combining various datasets as training data)? Intuitively, a model trained under mixed domains should have better generalization ability, but experimental results show otherwise; in fact, it performs worse than single-source training. Therefore, simply mixing datasets is not a feasible approach. This invention, for drug prediction targeting new targets or across targets, employs the concept of transfer learning, training a target domain prediction model using a neighborhood prediction model (neighborhood training samples).

[0072] The drug target interaction prediction model training method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can receive instructions from the client to train the model. Based on these instructions, the server obtains training samples (source domain training samples and target domain training samples). According to preset training rules, the source domain training samples are input into the source feature extraction unit of the source domain prediction model. By inputting the source domain training samples into the source feature extraction unit of the source domain prediction model and the target domain training samples into the target feature extraction unit of the target domain prediction model, the output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector; the output of the target feature extraction unit is the target domain drug-target interaction vector. The source domain drug-target interaction vector and the target domain drug-target interaction vector are then used to train the model. The input vectors are respectively input into the source classifier unit of the source domain prediction model, and the input vector of the drug-target interaction in the target domain is input into the target classifier unit of the target domain prediction model. The outputs of the source classifier unit are the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. The output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein. Finally, the source domain prediction model and the target domain prediction model are trained based on the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source label, the source prediction label, the first prediction label and the second prediction label. In this invention, when addressing the task of drug discovery for new targets, the problem of training a target domain drug target interaction prediction model is solved by utilizing a source domain drug target interaction prediction model with a large (sufficient) training sample through transfer learning. This involves organizing known target data through a multi-source domain approach, aligning and transferring knowledge from different known domains to the new target. In other words, a small number of measurement sample data points on the new target can be directly used for training and fine-tuning, leveraging knowledge gained from previous tasks instead of training from scratch. This avoids training a completely new model from zero for each target disease or virus drug prediction. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0073] Please see Figure 2 As shown, Figure 2 This is a flowchart illustrating a drug target interaction prediction model training method provided in an embodiment of the present invention. The method is used to train a source domain prediction model and a target domain prediction model. Both the source domain prediction model and the target domain prediction model include a feature extraction unit and a classifier unit, and may include the following steps:

[0074] S10: Obtain source domain training samples and target domain training samples. The source domain training samples include source training drug molecules, source training target proteins, and source labels of the effects of source training drug molecules on source training target proteins. The target domain training samples include target training drug molecules and target training target proteins.

[0075] The source domain training samples contain rich supervised annotation information (labels). These labels indicate whether drug molecules within the domain (in the training samples) can act on the target protein. Specific labels can be set as needed, such as two labels: "useful" and "useless," or multiple labels based on the probability of action. The target domain training samples may have a few labels or none. The target domain training samples and source domain training samples belong to different domains. For example, the target protein in the source domain training samples could be the SARS virus and various drug molecules acting on it; the target domain training samples could be the novel coronavirus and various drug molecules acting on it.

[0076] S20: Input the source domain training samples into the source feature extraction unit of the source domain prediction model. The output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector. Input the target domain training samples into the target feature extraction unit of the target domain prediction model. The output of the target feature extraction unit is the target domain drug-target interaction vector.

[0077] It is understood that the drug-target interaction prediction model to be trained in this invention represents the prediction of drug-molecule and target protein interactions as a binary (or multi-class) task of predicting whether a drug-target pair (drug molecule and target protein) represents an interaction between the drug molecule and the target protein (target protein). Thus, the information of the drug molecule and target protein is represented as feature vectors, and the interaction between the drug molecule and target protein is represented as class labels. For example, a neural network model can be used to automatically learn low-dimensional dense representation vectors of drug molecules and target proteins directly from descriptors that retain a large amount of original structural information.

[0078] In some embodiments, the target feature extraction unit in S20 above includes a first neural network and a second neural network.

[0079] Of course, the source feature extraction units of the source domain prediction model also include the first neural network and the second neural network. In principle, the source domain prediction model and the target domain prediction model have the same model structure; the difference lies only in the specific parameters of the model.

[0080] In S20 above, inputting the target domain training samples into the target domain prediction model's target feature extraction unit may include:

[0081] S20a. Input the target training drug molecule into the first neural network, and input the target training target protein into the second neural network;

[0082] In this embodiment, the drug molecules are feature-extracted using a first neural network to obtain a feature vector; the drug molecules are then feature-extracted using a second neural network to obtain another feature vector.

[0083] S20b: Determine the weight of the effect of the target training drug molecule on the target training target protein based on the output of each layer of the first neural network and the output of each layer of the second neural network.

[0084] S20c: The first neural network outputs a vector sequence of the target training drug molecules based on the action weights, and the second neural network outputs a vector sequence of the target protein target points based on the action weights.

[0085] This embodiment utilizes attention weights to capture drug-target interaction (DTI) binding sites (in biochemistry and molecular biology, a binding site is a region on a protein; binding to a protein often causes changes in protein conformation and leads to alterations in cellular function. Inhibiting protein binding can lead to pathway regulation, homeostasis, and physiological dysfunction in cancer cells or viruses, thereby achieving therapeutic goals): Given a drug molecule carrier D (drug molecule) and hidden vectors of subsequences in a set of proteins C (target proteins), the embodiment calculates which subsequence in the protein is more important to the drug molecule by assigning greater weights to the subsequences. This weighting is calculated using a neural network with attention weights, flexibly simulating the interaction between the compound and protein, rather than a simple summation.

[0086] S20d connects the vector sequence of the target training drug molecule and the vector sequence of the target protein target point as the output of the feature extraction unit.

[0087] In this embodiment, the vector sequence of drug molecules output by the first neural network and the vector sequence of target proteins output by the second neural network are concatenated and output as the input of the subsequent classifier unit.

[0088] In some applications, the first neural network can be a graph neural network, and the second neural network can be a convolutional neural network or a Word2vec model. Details are as follows:

[0089] Graph Neural Networks (GNNs) are a branch of deep learning applied to graph-structured data. The molecular sequence of a drug molecule can be processed into atomic nodes using RDKit. This graph data, connected by chemical bonds, serves as input to the GNN. Common graph-structured data includes nodes and edges. Nodes contain entity information (such as atoms in a compound or individuals in a social network), while edges contain relationship information between entities (chemical bonds between atoms in a graph compound).

[0090] The input to a graph neural network is typically a graph structure with node or edge attributes, as described above, including the graph's adjacency matrix A and corresponding attribute information X. Its final output generally depends on the specific task; for example, node classification outputs node labels, graph classification outputs graph labels, and link prediction outputs the existence of links. Taking graph classification as an example, the graph neural network trains an implicit vector representation for each node in the graph based on the graph structure and input node attributes. The goal is to make this vector representation contain sufficiently powerful expressive information to help each node extract information. Finally, through methods such as average pooling, the information vector representation of the entire graph can be obtained (e.g., extracting molecular-level information representation of the entire molecule compound from the features of atomic nodes and the chemical bond information of the edges between atoms).

[0091] The main learning process of a graph neural network model involves iteratively aggregating and updating the neighbor information of nodes in the graph data. In one iteration, each node updates its information by aggregating the features of its neighbors and its own features from the previous layer, and typically performs a non-linear transformation on the aggregated information. By stacking multiple layers of the network, each node can obtain neighbor node information within a corresponding number of hops.

[0092] If we understand the learning of a graph neural network model in terms of node message passing, it involves two processes: the message passing phase and the readout phase. The message passing phase is the forward propagation phase, which runs T steps in a loop and is processed by the function M. t Information is obtained through function U t Update the nodes; the equation for this stage is as follows.

[0093]

[0094]

[0095] Among them, e vw This represents the feature vector of the edge from node v to w.

[0096] The readout phase computes a feature vector for the representation of the entire graph, implemented using the function R.

[0097]

[0098] Where T represents the total number of time steps, and the function M... t U t R and R can use different model settings.

[0099] Proteins are fundamental macromolecules that control organisms and life itself. Research on proteins contributes to understanding human health and developing disease therapies. The decreasing cost of sequencing technology has enabled the creation of vast natural protein databases, rich in information, allowing for the development of excellent sequence-based deep learning methods. Proteins, composed of amino acid sequences, can be viewed, in a sense, as a language; specifically, proteins are linear chains of amino acids linked by covalent bonds. There are 20 standard amino acids, and these standard amino acid "letters" can label proteins as a sequence, much like encoding English sentences. This discrete sequence representation is called the protein's major structure. Therefore, feature extraction of target proteins by a second neural network can be achieved by using a CNN model to extract features from a text sequence composed of amino acid elements to obtain the corresponding vector sequence.

[0100] Specifically, the basic architecture of this model for processing target protein sequence data is based on n-gram amino acid embedding. The input amino acid sequence is divided into words using the n-gram method. In a specific application scenario, an overlapping partition with n=3 can be used to divide the protein sequence into overlapping 3-amino acid sequences. Then, a pre-trained word2vec model is used to convert the words into real-valued embeddings.

[0101] In particular, given the latent vectors of a molecule and a protein, the algorithm calculates which subsequence of the protein is more important to the drug molecule by assigning greater weights to the sequences. This weighting is calculated using a neural network with an attention weights mechanism, which flexibly simulates the "binding sites" of the interaction between the drug molecule and the protein, rather than a simple summation.

[0102] Of course, the specific process of inputting the source domain training samples into the target feature extraction unit of the source domain prediction model is similar to that described in S20a to S20d above, and will not be repeated here.

[0103] S30: Input the source domain drug-target interaction vector and the target domain drug-target interaction vector into the source classifier unit of the source domain prediction model, respectively. The output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. Input the target domain drug-target interaction vector into the target classifier unit of the target domain prediction model. The output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0104] In principle, different classifiers should make predictions for the same target domain training sample (i.e., the same training sample) as consistent as possible. Therefore, this embodiment utilizes this feature to train the target classifier unit and the source classifier unit simultaneously by inputting the target domain drug-target action vector into the source classifier unit and the target classifier unit respectively, and inputting the source domain drug-target action vector into the source classifier unit.

[0105] S40: Train the source domain prediction model and the target domain prediction model based on the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source label, the source prediction label, the first prediction label, and the second prediction label.

[0106] In some embodiments, the above-described S40 may include:

[0107] S401a. Calculate the maximum mean difference loss between the alignment of the source domain drug-target interaction vector and the target domain drug-target interaction vector.

[0108] Among them, the drug-target interaction vectors of the target domain and the source domain are aligned using MMD, and the maximum mean difference loss of LMMD is calculated.

[0109] S401b: Calculate the contrast loss function between the target classifier unit of the target domain prediction model and the source classifier unit of the source domain prediction model based on the first and second prediction labels.

[0110] The contrastive loss function characterizes the difference between the predicted labels of the source classifier unit and the target classifier unit for the same target domain training samples.

[0111] S40c. Calculate the classification loss function of the source domain prediction model based on the source label and the predicted label.

[0112] In this process, the source domain samples are input into the source classifier unit. Since there are corresponding classification labels (whether multiple drug molecules in this domain can act on the target protein), the classification loss (Lclass, using cross-entropy) can be obtained.

[0113] S40d trains source domain prediction models and target domain prediction models based on maximum mean difference loss, contrastive loss function, and classification loss function.

[0114] The overall loss function of the method is shown below, consisting of three parts: domain alignment loss of MMD, labeled classification loss, and unlabeled classifier prediction consistency comparison loss.

[0115] L_total=L_MMD+L_contrast+L_class.

[0116] In some embodiments, obtaining the source domain training samples and the target domain training samples in S10 above may include:

[0117] Obtain at least two sets of source domain training samples, each set of source domain training samples corresponds one-to-one with a source domain prediction model; among them, the target proteins of different source domain training samples belong to different categories.

[0118] Among them, the target proteins belonging to different categories means that the target proteins in the source domain training samples are different proteins, that is, targeting different diseases. For example, one source domain training sample can be the 2019-nCoV coronavirus, and another source domain training sample can be a variant of the coronavirus (Beta).

[0119] This embodiment uses multiple source domain prediction models to train the target domain prediction model, thereby improving the prediction accuracy of the target domain training model.

[0120] In some embodiments, in S20 above, inputting source domain training samples into the source feature extraction unit of the source domain prediction model includes:

[0121] Each source domain training sample is input into the source feature extraction unit of the source domain prediction model corresponding to the source domain training sample.

[0122] In this embodiment, the source domain training samples are input into the source feature extraction units of their respective source domain prediction models. The feature extraction process of each source feature extraction unit for the source domain training samples is the same as that of S20a and S20d mentioned above, and will not be repeated here.

[0123] In some embodiments, S40 may include:

[0124] S401a. Calculate the maximum mean difference loss between the source domain drug-target action vector and the target domain drug-target action vector output by the source feature extraction unit of each source domain prediction model.

[0125] S401b: Train the source feature extraction unit of the source domain prediction model and the target feature extraction unit of the target domain prediction model by minimizing the maximum mean difference loss.

[0126] Understandably, MMD (Maximum Mean Difference) is one of the most widely used (currently) loss functions in transfer learning, especially in domain adaptation. It primarily measures the distance between two different but related distributions. The LMMD loss calculation works as follows: to achieve domain adaptation, the source and target data need to be mapped onto a reproducing kernel Hilbert space (RKHS) using the same mapping. Then, the mean difference between the two mapped data is calculated. Here, the maximum mean difference (MMD) is used as a metric to measure the difference between cross-domain distributions. If the dataset... and If the data follow probability distributions p and q respectively, then the MMD between datasets X and Y is:

[0127]

[0128] Where sup(·) is the sum of the input sets, H represents the Reproducing Kernel Hilbert Space (RKHS), and φ(·) represents the nonlinear mapping from the original feature space to the RKHS. The nonlinear mappings of the RKHS can be considered sufficiently rich to find a suitable mapping that maximizes the distance between datasets X and Y. The RKHS is generated from the feature kernel using a distributed kernel mean embedding.

[0129] The empirical estimation formula for MMD based on kernel mean embedding is as follows:

[0130]

[0131] Specifically, in this embodiment, the maximum mean difference loss between the source domain drug-target action vector and the target domain drug-target action vector output by the source feature extraction unit of each source domain prediction model is calculated. That is, there is as many maximum mean difference losses as there are source domain prediction models. The source domain prediction model and the target domain prediction model are adjusted by all the maximum mean difference losses.

[0132] In some embodiments, inputting the source domain drug-target interaction vector and the target domain drug-target interaction vector into the source classifier unit of the source domain prediction model in S30 above may include:

[0133] S30a. Input the source domain drug-target interaction vector corresponding to each source domain training sample into the source classifier unit of the source domain prediction model corresponding to the source domain training sample. The source classifier unit outputs the source prediction label of the interaction between the source training drug molecule and the source training target protein.

[0134] S30b: Input the drug-target interaction vector of the target domain into the source classifier unit of each source domain prediction model, and each source classifier unit outputs the first prediction label of the interaction between the target training drug molecule and the target training target protein.

[0135] S30c: Input the target domain drug-target interaction vector into the target domain prediction model's target classifier unit, and the target classifier unit outputs the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0136] In this embodiment, the drug-target interaction vector of the target domain is input into the source classifier unit to obtain the classification label of the target domain training sample by the source domain prediction model. The drug-target interaction vector of the target domain is input into the target classifier unit to obtain the classification label of the target domain training sample by the target domain prediction model. Here, the drug-target interaction vector of the target domain is input into the source classifier unit of each source domain prediction model. That is, each source domain prediction model outputs its own corresponding first prediction label. That is, there are as many first prediction labels as there are source domain prediction models.

[0137] In some embodiments, the above-described S40 may include:

[0138] S402a. Based on the first and second predicted labels output by the source classifier units of each source domain prediction model, calculate the comparison loss function between the source classifier units of each source domain prediction model and the target classifier units of the target domain prediction model.

[0139] S402b: Calculate the classification loss function of the source classifier unit of the source domain prediction model based on the source prediction label output by the source classifier unit of each source domain prediction model and the source label corresponding to the source domain prediction model.

[0140] S402c: Based on the classification loss function and the contrastive loss function, train the source classifier unit of the source domain prediction model and the target classifier unit of the target domain prediction model.

[0141] Specifically, in this embodiment, the contrast loss function is calculated between the first predicted label output by the source classifier unit of each source domain prediction model and the second predicted label output by the target domain classifier unit. That is, there is as many contrast loss functions as there are source domain prediction models. The source domain prediction model and the target domain prediction model are adjusted by using all the contrast loss functions and the classification loss function of each source domain prediction model.

[0142] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0143] In one embodiment, a device based on a drug molecule target interaction prediction model is provided, which corresponds one-to-one with the drug molecule target interaction prediction model training method in the above embodiments.

[0144] Once the model is trained, when used to predict the effects of new, unknown drug targets, only the target protein sequence data and the drug molecule sequence dataset to be predicted need to be input to predict which drugs can have an effective effect on the target target.

[0145] In one embodiment, a drug target interaction prediction model training device is provided for training a source domain prediction model and a target domain prediction model. Both the source domain prediction model and the target domain prediction model include a feature extraction unit and a classifier unit. This drug target interaction prediction model training device corresponds one-to-one with the drug target interaction prediction model training method described in the above embodiments. Figure 7 As shown, the prediction model training device includes an acquisition module 101, a feature extraction module 102, a label determination module 103, and an adjustment module 104. Detailed descriptions of each functional module are as follows:

[0146] The acquisition module 101 is used to acquire source domain training samples and target domain training samples. The source domain training samples include source training drug molecules, source training target proteins, and source annotation tags of the effects of source training drug molecules on source training target proteins. The target domain training samples include target training drug molecules and target training target proteins.

[0147] The feature extraction module 102 is used to input source domain training samples into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector; and to input target domain training samples into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is the target domain drug-target interaction vector.

[0148] The label determination module 103 is used to input the source domain drug-target interaction vector and the target domain drug-target interaction vector into the source classifier unit of the source domain prediction model, respectively. The output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model, and the output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0149] The adjustment module 104 is used to train the source domain prediction model and the target domain prediction model based on the source domain drug-target action vector, the target domain drug-target action vector, the source label, the source prediction label, the first prediction label, and the second prediction label.

[0150] In one embodiment, the adjustment module 104 is specifically used for:

[0151] Calculate the maximum mean difference loss between the alignment of the source domain drug-target interaction vector and the target domain drug-target interaction vector;

[0152] The comparison loss function is calculated based on the first and second prediction labels to compare the target classifier unit of the target domain prediction model and the source classifier unit of the source domain prediction model.

[0153] Calculate the classification loss function of the source domain prediction model based on the source label and the predicted label;

[0154] The source domain prediction model and the target domain prediction model are trained using the maximum mean difference loss, the contrast loss function, and the classification loss function.

[0155] In one embodiment, the acquisition module 101 is specifically used for:

[0156] Obtain source domain training samples and target domain training samples, including:

[0157] Obtain at least two sets of source domain training samples, where the target proteins of different source domain training samples belong to different categories.

[0158] In one embodiment, each set of source domain training samples corresponds one-to-one with a source domain prediction model; the feature extraction module 102 is specifically used for:

[0159] Each source domain training sample is input into the source feature extraction unit of the source domain prediction model corresponding to the source domain training sample.

[0160] In one embodiment, the adjustment module 104 is specifically used for:

[0161] Calculate the maximum mean difference loss between the source domain drug-target interaction vector and the target domain drug-target interaction vector output by the source feature extraction unit of each source domain prediction model;

[0162] The source feature extraction unit of the source domain prediction model and the target feature extraction unit of the target domain prediction model are trained by minimizing the maximum mean difference loss.

[0163] In one embodiment, each set of source domain training samples corresponds one-to-one with a source domain prediction model; the label determination module 103 is specifically used for:

[0164] The source domain drug-target interaction vector corresponding to each source domain training sample is input into the source classifier unit of the source domain prediction model corresponding to the source domain training sample. The source classifier unit outputs the source prediction label of the interaction between the source training drug molecule and the source training target protein.

[0165] The drug-target interaction vector of the target domain is input into the source classifier unit of each source domain prediction model, and each source classifier unit outputs the first prediction label of the interaction between the target training drug molecule and the target training target protein.

[0166] The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model, and the target classifier unit outputs the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0167] In one embodiment, the adjustment module 104 is specifically used for:

[0168] Based on the first and second predicted labels output by the source classifier units of each source domain prediction model, calculate the comparison loss function between the source classifier units of each source domain prediction model and the target classifier units of the target domain prediction model.

[0169] Based on the source prediction labels output by the source classifier units of each source domain prediction model and the corresponding source label labels of the source domain prediction model, calculate the classification loss function of the source classifier units of the source domain prediction model.

[0170] Based on the classification loss function and the contrastive loss function, the source classifier unit of the source domain prediction model and the target classifier unit of the target domain prediction model are trained.

[0171] In one embodiment, the target feature extraction unit includes a first neural network and a second neural network; the feature extraction module 102 is specifically used for:

[0172] Inputting the target domain training samples into the target domain prediction model's target feature extraction unit includes:

[0173] The target training drug molecule is input into the first neural network, and the target training target protein is input into the second neural network;

[0174] Based on the outputs of each layer of the first neural network and the outputs of each layer of the second neural network, the influence weights of the target training drug molecule on the target training target protein are determined.

[0175] The first neural network outputs a vector sequence of target training drug molecules based on the interaction weights, and the second neural network outputs a vector sequence of target protein targets based on the interaction weights.

[0176] The vector sequence of the target training drug molecule and the vector sequence of the target protein target are concatenated and used as the output of the feature extraction unit.

[0177] In one embodiment, the first neural network is a graph neural network, and the second neural network is a convolutional neural network or a Word2vec model.

[0178] This invention provides an intelligent question-answering processing device. First, it obtains a preliminary ranking result of candidate questions through semantic matching. Then, it proposes a scheme to optimize the question-answering engine based on entity alignment. By using entity alignment, the ranking result of candidate questions is re-ranked, so that more matching candidate questions are selected. This can effectively avoid the generalization ability defects of the model, greatly improve the effect of entity matching, and improve the effect of the question-answering engine.

[0179] Specific limitations regarding the intelligent question-answering processing device can be found in the limitations of the intelligent question-answering method described above, and will not be repeated here. Each module in the aforementioned intelligent question-answering processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0180] In one embodiment, a drug target interaction prediction device is provided for predicting the interaction between a drug molecule and a target protein in a target domain, comprising:

[0181] The acquisition module is used to acquire the drug molecule and target protein to be predicted, wherein the target protein belongs to the same category as the target domain target protein mentioned above.

[0182] The result output unit is used to input the drug molecule and protein to be predicted into the trained drug target interaction prediction model to obtain the drug molecule's effect label on the target protein; wherein, the drug target interaction prediction model is trained using the above-mentioned drug target interaction prediction model training method.

[0183] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side intelligent question-answering method based on artificial intelligence.

[0184] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of an artificial intelligence-based intelligent question-answering processing method.

[0185] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0186] Obtain source domain training samples and target domain training samples. The source domain training samples include source training drug molecules, source training target proteins, and source labels of the effects of source training drug molecules on source training target proteins. The target domain training samples include target training drug molecules and target training target proteins.

[0187] The source domain training samples are input into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector; the target domain training samples are input into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is the target domain drug-target interaction vector.

[0188] The source domain drug-target interaction vector and the target domain drug-target interaction vector are respectively input into the source classifier unit of the source domain prediction model. The output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model. The output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0189] The source domain prediction model and the target domain prediction model are trained based on the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source label, the source prediction label, the first prediction label, and the second prediction label.

[0190] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0191] Obtain source domain training samples and target domain training samples. The source domain training samples include source training drug molecules, source training target proteins, and source labels of the effects of source training drug molecules on source training target proteins. The target domain training samples include target training drug molecules and target training target proteins.

[0192] The source domain training samples are input into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit of the source domain prediction model is the source domain drug-target interaction vector; the target domain training samples are input into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is the target domain drug-target interaction vector.

[0193] The source domain drug-target interaction vector and the target domain drug-target interaction vector are respectively input into the source classifier unit of the source domain prediction model. The output of the source classifier unit is the source prediction label of the interaction between the source training drug molecule and the source training target protein and the first prediction label of the interaction between the target training drug molecule and the target training target protein, respectively. The target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model. The output of the target classifier unit is the second prediction label of the interaction between the target training drug molecule and the target training target protein.

[0194] The source domain prediction model and the target domain prediction model are trained based on the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source label, the source prediction label, the first prediction label, and the second prediction label.

[0195] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0196] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0197] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0198] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for training a drug target interaction prediction model, characterized in that, The method is used for training a source domain prediction model and a target domain prediction model, the source domain prediction model and the target domain prediction model both comprise a feature extraction unit and a classifier unit, and the method comprises the following steps: Obtaining source domain training samples and target domain training samples, wherein the source domain training samples comprise source training drug molecules, source training target proteins and source annotation labels of the interaction between the source training drug molecules and the source training target proteins, and the target domain training samples comprise target training drug molecules and target training target proteins; Inputting the source domain training samples into a source feature extraction unit of the source domain prediction model, wherein the output of the source feature extraction unit of the source domain prediction model is a source drug-target interaction vector; inputting the target domain training samples into a target feature extraction unit of the target domain prediction model, wherein the output of the target feature extraction unit is a target drug-target interaction vector; Inputting the source drug-target interaction vector and the target drug-target interaction vector into a source classifier unit of the source domain prediction model respectively, wherein the output of the source classifier unit is a source prediction label of the interaction between the source training drug molecules and the source training target proteins and a first prediction label of the interaction between the target training drug molecules and the target training target proteins respectively; inputting the target drug-target interaction vector into a target classifier unit of the target domain prediction model, wherein the output of the target classifier unit is a second prediction label of the interaction between the target training drug molecules and the target training target proteins; Training the source domain prediction model and the target domain prediction model according to the source drug-target interaction vector, the target drug-target interaction vector, the source annotation label, the source prediction label, the first prediction label and the second prediction label; The training of the source domain prediction model and the target domain prediction model according to the source drug-target interaction vector, the target drug-target interaction vector, the source annotation label, the source prediction label, the first prediction label and the second prediction label comprises the following steps: Calculating a maximum mean discrepancy loss of the alignment between the source drug-target interaction vector and the target drug-target interaction vector; calculating a contrastive loss function of the target classifier unit of the target domain prediction model and the source classifier unit of the source domain prediction model according to the first prediction label and the second prediction label; calculating a classification loss function of the source domain prediction model according to the source annotation label and the prediction label; and training the source domain prediction model and the target domain prediction model according to the maximum mean discrepancy loss, the contrastive loss function and the classification loss function; The inputting of the source drug-target interaction vector and the target drug-target interaction vector into the source classifier unit of the source domain prediction model comprises the following steps: inputting the source domain drug-target interaction vector corresponding to each source domain training sample into a source classifier unit of a source domain prediction model corresponding to the source domain training sample, the source classifier unit outputting a source prediction label of the interaction between the source training drug molecule and the source training target protein; inputting the target domain drug-target interaction vector into a target classifier unit of the target domain prediction model, the target classifier unit outputting a second prediction label of the interaction between the target training drug molecule and the target training target protein; The drug-target interaction prediction model to be trained is represented as a binary classification task of predicting whether a drug-target pair is an interaction between a drug molecule and a target protein. The information of the drug molecule and the target protein is represented as a feature vector, and the interaction between the drug molecule and the target protein is represented as a class label. 2.The method of claim 1, wherein the drug target interaction prediction model is trained using a machine learning algorithm. The source domain training samples and the target domain training samples are obtained, including: At least two groups of source domain training samples are obtained, each group of source domain training samples corresponding to a source domain prediction model; wherein the target proteins of different source domain training samples belong to different categories; The source domain training sample is input into the source feature extraction unit of the source domain prediction model, including: Each source domain training sample is input into the source feature extraction unit of the source domain prediction model corresponding to the source domain training sample. 3.The method of claim 2, wherein the drug target interaction prediction model is trained using a machine learning algorithm. The source domain prediction model and the target domain prediction model are trained according to the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source annotation label, the source prediction label, the first prediction label and the second prediction label, including: The maximum mean difference loss of the alignment of the source domain drug-target interaction vector and the target domain drug-target interaction vector output by the source feature extraction unit of each source domain prediction model is calculated respectively; The source feature extraction unit of the source domain prediction model and the target feature extraction unit of the target domain prediction model are trained by minimizing the maximum mean difference loss.

4. The method for training a drug target interaction prediction model as described in claim 1, characterized in that, The source domain prediction model and the target domain prediction model are trained according to the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source annotation label, the source prediction label, the first prediction label and the second prediction label, including: According to the first prediction label and the second prediction label output by the source classifier unit of each source domain prediction model, the contrast loss function of the source classifier unit of each source domain prediction model and the target classifier unit of the target domain prediction model is calculated; According to the source prediction label output by the source classifier unit of each source domain prediction model and the source annotation label corresponding to the source domain prediction model, the classification loss function of the source classifier unit of the source domain prediction model is calculated; According to the classification loss function and the contrast loss function, the source classifier unit of the source domain prediction model and the target classifier unit of the target domain prediction model are trained. 5.The method of claim 1, wherein the drug target interaction prediction model is trained using a machine learning algorithm. The target feature extraction unit comprises a first neural network and a second neural network; The target feature extraction unit comprises a first neural network and a second neural network; The target feature extraction unit comprises a first neural network and a second neural network; The target feature extraction unit comprises a first neural network and a second neural network; The target feature extraction unit comprises a first neural network and a second neural network; The target feature extraction unit comprises a first neural network and a second neural network; 6.A drug target interaction prediction model training apparatus, characterized by, The target feature extraction unit comprises a first neural network and a second neural network. The source domain prediction model and the target domain prediction model are trained, and the source domain prediction model and the target domain prediction model each comprise a feature extraction unit and a classifier unit, and the device comprises: The source domain training sample and the target domain training sample are obtained, the source domain training sample comprises a source training drug molecule, a source training target protein, and a source annotation label of the source training drug molecule acting on the source training target protein, and the target domain training sample comprises a target training drug molecule and a target training target protein; The source domain training sample is input into the source feature extraction unit of the source domain prediction model, and the output of the source feature extraction unit of the source domain prediction model is a source domain drug-target interaction vector; the target domain training sample is input into the target feature extraction unit of the target domain prediction model, and the output of the target feature extraction unit is a target domain drug-target interaction vector; The source domain drug-target interaction vector and the target domain drug-target interaction vector are input into the source classifier unit of the source domain prediction model, and the output of the source classifier unit is a source prediction label of the source training drug molecule interacting with the source training target protein and a first prediction label of the target training drug molecule interacting with the target training target protein; the target domain drug-target interaction vector is input into the target classifier unit of the target domain prediction model, and the output of the target classifier unit is a second prediction label of the target training drug molecule interacting with the target training target protein; The source domain prediction model and the target domain prediction model are trained according to the source domain drug-target interaction vector, the target domain drug-target interaction vector, the source annotation label, the source prediction label, the first prediction label, and the second prediction label; The adjustment module is specifically configured to: Calculate the maximum mean difference loss of the alignment of the source domain drug-target interaction vector and the target domain drug-target interaction vector; Calculate the contrast loss function of the target classifier unit of the target domain prediction model and the source classifier unit of the source domain prediction model according to the first prediction label and the second prediction label; Calculate the classification loss function of the source domain prediction model according to the source annotation label and the prediction label. The source domain prediction model and the target domain prediction model are trained according to a maximum mean discrepancy loss, a contrastive loss function and a classification loss function; The label determination module is specifically configured to: input the source domain drug-target interaction vector corresponding to each source domain training sample into a source classifier unit of the source domain prediction model corresponding to the source domain training sample, and the source classifier unit outputs a source prediction label of the interaction between the source training drug molecule and the source training target protein; input the target domain drug-target interaction vector into the target classifier unit of the target domain prediction model, and the target classifier unit outputs a second prediction label of the interaction between the target training drug molecule and the target training target protein; The drug-target interaction prediction model to be trained predicts the interaction between a drug molecule and a target protein as a binary classification task of predicting whether a drug-target pair is an interaction between a drug molecule and a target protein, the information of the drug molecule and the target protein is represented as a feature vector, and the interaction between the drug molecule and the target protein is represented as a class label. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 5.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: ​

Images