Trustworthy drug-target association prediction method

Abstract

The application discloses a kind of reliable drug-target correlation prediction methods, comprising: drug-target structure feature representation, according to the initial character sequence of drug-target, utilize convolution representation learning module, obtain the feature vector related to its structure characteristics;Utilize deep evidence module, reliably predict drug-target correlation coefficient.This application can be based on the huge drug-target correlation pair database formed by a large number of chemical experiment results, train stable and reliable prediction model.Before entering drug experiment, screen drug-target pair, discover "reliable" active correlation pair in advance, shorten the development cycle of drug.At the same time, a large number of "reliable" non-active correlation pairs are screened out, so that the experimental cost on these non-active correlation pairs is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a reliable method for predicting drug-target correlation. Background Technology

[0002] Drug discovery that meets pharmaceutical industry standards requires collaboration across various research fields. Validating drug-target associations using traditional experimental methods is often time-consuming and expensive. In recent years, with the development of machine learning and deep neural networks, large databases of drug-target associations based on extensive chemical experimental results have been developed. These databases allow for screening of drug-target pairs before initiating drug trials, identifying "active" associations in advance, thus significantly reducing experimental costs and shortening drug development cycles. However, these methods still have limitations. Traditional machine learning drug-target association prediction models can only determine whether a drug-target pair has an association, but cannot provide the reliability of the result. For some drug-target pairs predicted as "active," chemical experiments often fail to provide relevant activity verification, wasting experimental costs. Conversely, simply ignoring drug-target pairs judged as "inactive" may prevent the discovery of some truly "active" drug-target pairs. Therefore, models need to assess the reliability of association predictions while providing them, further reducing experimental costs on "unreliable" data points. Summary of the Invention

[0003] To address the problems existing in the prior art, the purpose of this invention is to provide a reliable method for predicting drug-target correlation. This invention solves the problem that existing drug-target correlation prediction models cannot determine the reliability of the results, and improves the screening quality of drug-targets through reliability indicators, thereby further shortening the drug discovery cycle.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a reliable method for predicting drug-target correlation, comprising the following steps:

[0005] (1) Obtain the feature vector representing the structural features of the drug-target.

[0006] (2) The feature vectors are fused and used as input to the deep evidence classification model, and the output is the evidence vector e of the drug-target pair. i , used to determine its predicted Dirichlet distribution Beta(p i |α i ,β i The parameter α i ,β i Based on the Dirichlet distribution, a loss optimization function is constructed. And with the objective of minimizing the loss optimization function, the parameters of the deep evidence classification model are updated, where y i As a binomial distributed random variable used for prediction, p describes the predicted label. i As a random variable with a Dirichlet distribution on a binomial distribution, it describes the prediction uncertainty, and BCE(·) represents the cross-entropy function;

[0007] (3) Input the new drug-target pair to be predicted back into the deep evidence classification model, perform feature extraction and fusion using step (1), and then use step (2) to obtain the predicted Dirichlet distribution Beta(p) of the drug-target pair. i |α i ,β i The system performs reliable correlation predictions and returns the correlation coefficient of the drug-target pair and the uncertainty coefficient of the prediction as the prediction results to the user.

[0008] As a further improvement of the present invention, in step (1), a feature vector representing the structural features of the drug-target is obtained. Specifically, the following steps are included:

[0009] (1-1) The Morgan fingerprinting algorithm is used to scan the original drug structure sequence to obtain the molecular fingerprint of its SMILES sequence binary vector; then, all molecular fingerprint vectors encoded as "1" are collected to form the drug C. i The encoding vector of the drug C; i The encoded vector is fed into the drug structure information query matrix. Choose matrix E C Different rows of the drug form a feature map matrix;

[0010] (1-2) Drug C i The feature map matrix is ​​fed into the convolutional neural network C. C Ultimately, this convolutional neural network C... C The output will be flattened into a column vector and its dimensionality will be transformed by a fully connected layer, serving as a feature vector representing the structural features of the drug.

[0011] As a further improvement of the present invention, in step (1), a feature vector representing the structural features of the drug-target is obtained. Specifically, the following steps are included:

[0012] (1-3) Using a character encoding table, target point P of the initial protein amino acid sequence. i Perform a scan to obtain its initial encoding vector; input the initial encoding vector of the target point into the structure information query matrix. In the middle, matrix E is selected according to the index of the encoding vector. P Different rows are used to obtain the initial feature map matrix of the target point;

[0013] (1-4) Target point P i The feature map matrix is ​​fed into the convolutional neural network C. P The output feature map will be flattened and its dimensions will be transformed by a fully connected layer to serve as the target point P. i eigenvectors

[0014] As a further improvement of the present invention, in step (2), the deep evidence classification model includes two fully connected layers and uses ReLU as the activation function to convert the output into a non-negative value, specifically as follows: Among them, e i This is the output evidence vector. This represents the parameters of the deep evidence classification model.

[0015] As a further improvement to the present invention, evidence vector e is used. i Determine the predicted beta distribution of the drug-target Beta (p i |α i ,β i The specific method is as follows: According to subjective logic theory, the predicted activity and inactivity probabilities of this drug-target combination are described as follows: The overall uncertainty of this forecast can be described as follows: Where K=2 represents a binary classification prediction problem, i.e., predicting whether the object is active or inactive.

[0016] As a further improvement of the present invention, in step (2), during the training phase of the deep evidence classification model, a Beta loss optimization function is constructed for each drug-target in the training set. Where BCE(·) represents the cross-entropy function; meanwhile, to further limit the number of pieces of evidence appearing in the wrong category, the KL divergence is... As a regularization term, it works together with the Beta distribution as a loss optimization function, as follows:

[0017]

[0018] in, and From α i ,β i Beta distribution parameters after removing non-erroneous evidence; λ t∈[0,1] is the regularization annealing coefficient.

[0019] As a further improvement of the present invention, in step (2), during the training phase of the deep evidence classification model, the goal is to minimize the loss optimization function. Query matrix E for structural information of drug and target respectively C E P And the parameters of the evidence module The model parameters are then calculated by taking the derivative and updating synchronously using gradient descent to gradually reduce the loss value to a certain threshold. Finally, the model parameters obtained at this point are saved.

[0020] As a further improvement of the present invention, step (3) specifically includes the following steps:

[0021] (3-1) The initial SMILES sequence and protein amino acid sequence of the drug-target to be predicted were obtained from the PubChem database and the RCSB PDB database, respectively.

[0022] (3-2) Use step (1) to extract and fuse features, and use step (2) to perform a "credible correlation prediction" for the drug-target pair, and calculate the correlation coefficient of the drug-target pair. The uncertainty coefficient u of the prediction is returned to the user as the prediction result.

[0023] The beneficial effects of this invention are:

[0024] This invention utilizes a vast database of drug-target association pairs derived from numerous chemical experiments to train a stable and reliable prediction model. By screening drug-target pairs before drug trials, "reliable" active association pairs are identified early, shortening the drug development cycle. Simultaneously, a large number of "reliable" inactive association pairs are screened out, significantly reducing experimental costs associated with these inactive pairs. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the framework of an embodiment of the present invention;

[0026] Figure 2 This is a flowchart illustrating the detailed process of reliable prediction in an embodiment of the present invention;

[0027] Figure 3 This is a graph showing the relationship between classification accuracy (ACC) and uncertainty threshold (u) in an embodiment of the present invention. Detailed Implementation

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] Example 1

[0030] A reliable method for predicting drug target association includes the following steps:

[0031] Step (1): Obtain the drug-target association dataset for training, where drug C i With target P i The original structural information was obtained using the SMILES sequence of the compound. and protein amino acid sequence Let L represent this, where L c and L p These represent the lengths of the SMILES sequence and the protein amino acid sequence, respectively.

[0032] Step (2): Construct a representation of drug C i eigenvectors of structural features

[0033] Step (2-1): For drug C i The Morgan fingerprinting algorithm was used to scan the original structural sequence to obtain the molecular fingerprint (binary vector) of its SMILES sequence; then, the position indices of all molecular fingerprint vectors encoded as "1" were collected to form the drug C. i The encoding vector. Note that different compounds have different Morgan fingerprints, which are expressed by encoding vectors of different lengths; therefore, this embodiment limits drug C. i The maximum length of the encoded vector. Encoded vectors exceeding this length will be truncated, and encoded vectors shorter than this length will be padded with 0s.

[0034] Step (2-2): Input the drug coding vector into the drug structure information query matrix. In the middle, matrix E is selected according to the index of the encoding vector. C Different rows are used to obtain the initial feature map matrix of the drug. Structural information query matrix E C The parameters are generated through random initialization. Intuitively, if the structural encoding vectors of two drugs are highly similar, then their feature map matrices will also be highly similar.

[0035] Step (2-3): Place drug C i The embedded feature map matrix is ​​fed into the convolutional neural network C. C (Each convolutional network layer consists of 64 convolutional kernels, one average pooling layer, and one normalization layer.) Ultimately, this convolutional neural network C... C The output feature map will be "flattened" into a column vector, and then its dimensions will be transformed by a fully connected layer, serving as the C of the drug. i eigenvectors of structural features

[0036] Query matrix E for structural information in steps (2-2) and (2-3). C and Convolutional Neural Networks C C Represented as θ C =(E C C C ), then, from drug C i The initial SMILES sequence is used to obtain its feature vector. The process can be described as follows:

[0037] Step (3): Construct a representation of target point P i eigenvectors of structural features

[0038] Step (3-1): For target point P i Using a character encoding table, the initial protein amino acid sequence P i Perform a scan to obtain its encoding vector. Note that different target points P i Targets P have different amino acid sequence lengths and are represented by coding vectors of different lengths; therefore, this embodiment restricts the target to P. i The maximum length of the encoded vector is specified. Encoded vectors that exceed or fall below this length will be processed in the same way as in step (2-1).

[0039] Step (3-2): Input the target encoding vector into the structural information query matrix. In the context of a protein's amino acid sequence (which can consist of a maximum of 25 different characters), matrix E is selected based on the index of the encoding vector. P Different rows are used to obtain the initial feature map matrix of the target point.

[0040] Step (3-3): Place the target point P i The feature map matrix is ​​fed into the convolutional neural network C. P Similar to steps (2-3), the output feature map will be "flattened" and its dimensions will be transformed by a fully connected layer to serve as the target point P. i eigenvectors of structural features

[0041] Query matrix E for structural information in steps (3-2) and (3-3). P and Convolutional Neural Networks C P Represented as θ P =(E C C C ), then, from target point P i The initial protein amino acid sequence is used to obtain its feature vector. The process can be described as follows:

[0042] Step (4): Fuse drug feature vectors With target feature vector

[0043] Step (5): Combine the fusion feature vector h obtained in step (4) i The input consists of an "evidence module" composed of two fully connected layers. The ReLU activation function is used to transform the output of the evidence module into a non-negative value, which serves as the evidence vector for the drug-target.

[0044] Step (6): Using the evidence vector from step (5), determine the Beta distribution (Pa) for the drug-target uncertainty prediction. i |α i ,β i ).

[0045] Step (7): Based on the predicted Beta distribution in step (6), construct a loss optimization function for the drug-target association dataset.

[0046] Step (8): With the goal of minimizing the loss optimization function, query the structural information matrix E for both the drug and the target. C E P Evidence module The parameters are differentiated, and the gradient descent method is used to update them synchronously, so that the loss value is gradually reduced to a certain threshold, and the model parameters obtained at this time are saved.

[0047] Step (9): For the new drug-target pair to be predicted, use steps (2), (3), and (4) to extract and fuse features. Finally, use steps (5) and (6) to perform "credible correlation prediction" on the drug-target pair. Return the correlation coefficient of the drug-target pair and the uncertainty coefficient of the prediction as the prediction result to the user.

[0048] Example 2

[0049] See Figure 1 , Figure 1 This is a schematic diagram illustrating the framework of the reliable drug-target association prediction method provided in this embodiment. Specifically, this embodiment will first train the deep evidence classification model on the Human and C. elegans datasets (data statistics are shown in Table 1). It should be noted that the "active" association pairs in the dataset of this embodiment are all selected from widely used drug-target datasets such as DrugBank, Matador, and STITCH, and highly reliable "inactive" association pairs are selected through statistical testing.

[0050] Table 1

[0051]

[0052] In the prediction experiment, this embodiment randomly divides the Human and C. elegans datasets into training and test sets, with a ratio of 80% and 20%, respectively. For each dataset, this embodiment uses 5-fold cross-validation to ensure the reliability of the model and saves the optimal model. In the field of novel drug discovery, this embodiment directly feeds the drug-target pairs to be screened into the aforementioned saved deep evidence classification model, outputs the "activity" evidence of the drug-target pair, and conducts further clinical trials to verify the correlation for drug-target pairs with a high number of "activity" evidence.

[0053] This embodiment details the process of drug-target reliability prediction as follows: Figure 2 As shown, the specific process is as follows:

[0054] Step (1): Obtain drug C i and target P i The initial feature representation, C i = [C, (, C, ..., S] and protein amino acid sequence P i = [M, M, A, ..., U].

[0055] Step (2): Construct a representation of drug C i eigenvectors of structural features

[0056] Step (2-1): In this embodiment, the Morgan fingerprint generation algorithm with a radius of 2 is first used to scan the initial SMILES sequence C of the drug. i When a specific molecular structure is found at a certain location, it is marked with "1"; otherwise, it is marked with "0". This generates a binary vector of length 2048 [1, 0, 1, 1, ..., 1, ..., 0]. This binary vector is then compressed into a more compact encoding vector [1, 3, 4, ..., 34, 0, ...] based on the index of the position where the "1" appears. Note that different compounds have different Morgan fingerprints, expressed by encoding vectors of different lengths; therefore, this embodiment limits drug C... i The maximum length of the encoded vector is 150. Encoded vectors exceeding this length will be truncated, and encoded vectors shorter than this length will be padded with 0s.

[0057] Step (2-2): Select the drug structure information query matrix based on the index of the encoded vector. Different rows are used to obtain the initial feature map matrix of the drug. For example, the encoding vector of the current drug will be selected from E. CRows 1, 3, 4, and 34. Note that the element 0 in the encoded vector will select E. C In row 2049 of the matrix, all elements are 0. Drug structure information query matrix E C The parameters will be randomly initialized and updated during training using stochastic gradient descent.

[0058] Step (2-3): Place drug C i The embedded feature map matrix is ​​fed into the convolutional neural network C. C (Except for the last layer, each convolutional network layer consists of 64 convolutional kernels, one average pooling layer, and one normalization layer, as shown in Table 2.) Finally, this convolutional neural network C... C The output feature map with a dimension of 17×23 will be "flattened" into a 391×1 vector, and then its dimension will be transformed by a 391×256 fully connected layer, which will serve as the vector for the drug C. i eigenvectors of structural features

[0059] Step (3): Construct a representation of target point P i eigenvectors of structural features

[0060] Step (3-1): For target point P i Using a character encoding table, the initial protein amino acid sequence P i Perform a scan and obtain the initial encoding vector based on the character markers encountered (e.g., 'M': 11, 'A': 1, 'U': 19). Note that different target points P... i Targets P have different amino acid sequence lengths and are represented by coding vectors of different lengths; therefore, this embodiment restricts the target to P. i The maximum length of the encoding vector is 1000. Encoding vectors that exceed or fall below this length will be processed in the same way as in step (2-1). In this example, the encoding vector of the target is [11, 11, 1, ..., 19, 0, ...].

[0061] Step (3-2): Input the target encoding vector into the structure information query matrix. In the middle, matrix E is selected according to the index of the encoding vector. P Different rows are used to obtain the initial embedding feature map matrix of the drug.

[0062] Step (3-3): Place the target point P i The embedded feature map matrix is ​​fed into the convolutional neural network C. P (The specific structure is shown in Table 2). The output feature map will be "flattened" and its dimensions will be transformed by a 240×256 fully connected layer, serving as the target point P.i eigenvectors of structural features

[0063] Table 2

[0064]

[0065] Step (4): Fuse drug feature vectors With target feature vector Specifically as follows:

[0066]

[0067] It should be noted that the fusion method used in this embodiment is "addition". In fact, for other specific embodiments, such as "multiplication of elements row by row" and "weighted average" can also be used as feature fusion methods.

[0068] Step (5): Combine the fusion feature vector h obtained in step (4) i The input consists of an "evidence module" composed of two fully connected layers (256×128 and 128×2). The ReLU activation function is then used to transform the output of the evidence module into a non-negative value, which serves as the evidence vector for the drug-target, as follows:

[0069]

[0070] Step (6): Using the evidence vector from step (5), determine the Beta distribution (Pa) for the drug-target uncertainty prediction. i |α i ,β i The details are as follows:

[0071]

[0072] According to the theory of subjective logic, the predicted activity and inactivity probabilities of this drug-target combination are described as follows:

[0073]

[0074] The overall uncertainty of this forecast can be described as follows:

[0075]

[0076] Here, K=2 represents a binary classification prediction problem (predicting activity or inactivity). As more classification evidence is obtained for the drug-target, the overall uncertainty of this prediction will continue to decrease.

[0077] Step (7): Based on the classification prediction Beta distribution obtained in step (6), construct a loss optimization function for the drug-target association dataset;

[0078] Step (7-1): This step involves constructing the Beta loss optimization function for the drug-target correlation training set. The specific function is as follows:

[0079]

[0080] Where BCE(·) represents the cross-entropy function.

[0081] Step (7-2): Simultaneously, in order to further limit the amount of evidence appearing in the wrong category, this embodiment uses the KL divergence... As a regularization term, it is used together with the Beta loss optimization function as the loss optimization function, as follows:

[0082]

[0083] in, and From α i ,β i The Beta distribution parameters after removing "non-false evidence". For example, when the data label is "inactive", λ t ∈[0,1] is the regularization annealing coefficient.

[0084] Step (8): With the goal of minimizing the loss optimization function, Query matrix E for structural information of drug and target respectively C E P Evidence module The parameters are differentiated, and the batch gradient descent method is used to update them synchronously, so that the loss value is gradually reduced to a certain threshold, and the model parameters obtained at this time are saved.

[0085] Step (9): For the new drug-target pair to be predicted, perform feature extraction and fusion using steps (2), (3), and (4), and finally use steps (5) and (6) to perform "credible correlation prediction" on the drug-target pair, and calculate the correlation coefficient of the drug-target pair. The uncertainty coefficient u of the prediction is returned to the user as the prediction result.

[0086] The reliable drug-target correlation prediction method provided in this embodiment is based on the open-source code library PyTorch and RDKit. The specific experimental parameter settings are shown in Table 3.

[0087] Table 3

[0088]

[0089]

[0090] Performance testing:

[0091] The following section evaluates the performance of this embodiment. For the two datasets, Human and C. elegans, the deep evidence model is trained using the training set and predictions are performed on the test set. It should be noted that the prediction result (step (9)) for a certain drug-target pair in this embodiment includes not only the correlation coefficient but also the uncertainty coefficient of the current prediction result. Therefore, this embodiment first compares with traditional correlation prediction methods such as TransformerCPI, MDL-CPI, BACPI, SVM, and NetLapRLS, and uses ACC, AUC, and AUPRC as basic classification performance metrics to illustrate the effectiveness of this embodiment in predicting the accuracy of the correlation coefficient. In addition, this embodiment provides a series of experimental results to verify that the deep evidence model has good uncertainty prediction capabilities.

[0092] The classification performance of the reliable drug-target association prediction method provided in this embodiment compared with other traditional association prediction methods is shown in Table 4 (Human dataset) and Table 5 (C.elegans dataset).

[0093] Table 4

[0094]

[0095] Table 5

[0096]

[0097] From the experimental results in the table, it can be seen that on the Human and C.elegans datasets, the prediction performance of this embodiment is better than or close to that of other baseline methods in terms of the AUC and AUPRC metrics. In particular, this embodiment also compares the prediction performance on the positive and negative example imbalanced datasets existing in the real application scenarios. On the Human dataset, this embodiment achieves an ACC of 0.948 on the positive and negative example balanced dataset, which is 3.8% - 9.5% higher than other methods. For the AUPRC metric, whether it is a balanced dataset or an imbalanced dataset (positive example: negative example = 1:3, 1:5), this embodiment can obtain the best performance. On the C.elegans dataset, although there is still a slight gap between this embodiment and the BACPI method in terms of the AUPRC metric, when the training set data is imbalanced, the classification performance of this embodiment is better. Through the above comparison, this embodiment has very good performance in terms of basic prediction accuracy. Especially in the case of data imbalance, this embodiment shows better robustness compared to other benchmark methods.

[0098] Table 6

[0099]

[0100] Compared with the traditional drug-target association prediction method, this embodiment can not only predict the association coefficient of the drug-target pair, but also provide an uncertainty evaluation coefficient of the prediction result. In the real application scenario, the prediction model may be "very uncertain" about the prediction results of some drug-target pairs. Therefore, it is necessary to verify whether the prediction model of this embodiment can accurately judge "which data points are unreliable in prediction". The experimental results are as Figure 3 shown. First, use the uncertainty u in step (6) as the threshold to filter those "reliable data points" and only give the prediction results of these data points. Intuitively, as this embodiment continuously reduces the uncertainty threshold u, the credibility of the prediction results of the selected data points should be higher, and the prediction accuracy ACC should also be correspondingly higher. For example, the prediction accuracy ACC of the data points with uncertainty u in the interval 0 < u < 0.1 (low uncertainty) should be higher than those in the interval 0.8 < u < 0.9 (high uncertainty). It can be seen from the results that as the uncertainty threshold decreases, on the Human and C.elegans datasets, the prediction accuracy ACC of the selected data points is continuously increasing. That is to say, this embodiment, a credible drug-target association prediction method, can accurately judge "which data points are relatively reliable".

[0101] To further verify the effectiveness of the reliable drug-target association prediction method in real-world drug discovery scenarios, the following example attempts to screen 82 antiviral drugs and 3 unrelated drugs (aspirin, penicillin, and amoxicillin) for the SARS-CoV23C-Like target (a major protease found in coronaviruses). This embodiment first obtains the initial SMILES sequences of the aforementioned drugs and the protein amino acid sequences of the SARS-CoV23C-Like target from the PubChem and RCSB PDB databases, respectively. These sequences are then used as input to a deep evidence classification model, outputting the association coefficient and prediction uncertainty for each drug-target pair. The results are shown in Table 6. Among the top 20 drugs with the highest association coefficients, 5 drugs (Cobistat, Ritonavir, Darunavir, Tipranavir, and Glecaprevir) have been clinically proven to inhibit the replication function of the SARS-CoV23C-Like target. It should be noted that this embodiment not only outputs the correlation coefficient for each drug-target pair, but also the uncertainty of the prediction results. For example, the probability of the drug Cobicistat having a functional association with the SARS-CoV23C-Like target is 100%, with extremely high confidence (i.e., uncertainty u = 0.08). However, for the drug Glecaprevir, the prediction probability is 90.5%, and the confidence of this result is relatively low (i.e., uncertainty u = 0.21). In addition, this embodiment has high confidence in the prediction results for drugs that do not interact with the SARS-CoV23C-Like target (aspirin, penicillin, and amoxicillin, with uncertainties u of 0.05, 0.04, and 0.02, respectively), which is of great guiding significance for reducing the experimental cost of these "inactive" drugs in this embodiment.

[0102] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A reliable method for predicting drug-target correlation, characterized in that, Includes the following steps: (1) Obtain the feature vector representing the structural features of the drug-target. In step (1), the feature vector representing the structural features of the drug-target is obtained. Specifically, the following steps are included: (1-1) The Morgan fingerprinting algorithm is used to scan the original drug structure sequence to obtain the molecular fingerprint of its SMILES sequence binary vector; then, all molecular fingerprint vectors encoded as "1" are collected to form the drug C. i The encoding vector of the drug C; i The encoded vector is fed into the drug structure information query matrix. Choose matrix E C Different rows of the drug form a feature map matrix; (1-2) Drug C i The feature map matrix is ​​fed into the convolutional neural network C. C Ultimately, this convolutional neural network C... C The output will be flattened into a column vector and its dimensionality will be transformed by a fully connected layer, serving as a feature vector representing the structural features of the drug. In step (1), the feature vector representing the structural features of the drug-target is obtained. Specifically, the following steps are included: (1-3) Using a character encoding table, target point P of the initial protein amino acid sequence. i Perform a scan to obtain its initial encoding vector; input the initial encoding vector of the target point into the structure information query matrix. In the middle, matrix E is selected according to the index of the encoding vector. P Different rows are used to obtain the initial feature map matrix of the target point; (1-4) Target point P i The feature map matrix is ​​fed into the convolutional neural network C. P The output feature map will be flattened and its dimensions will be transformed by a fully connected layer to serve as the target point P. i eigenvectors (2) The feature vectors are fused and used as input to the deep evidence classification model, and the output is the evidence vector e of the drug-target pair. i , used to determine its predicted Dirichlet distribution Beta(p i |α i ,β i The parameter α i ,β i Based on the Dirichlet distribution, a loss optimization function is constructed. And with the objective of minimizing the loss optimization function, the parameters of the deep evidence classification model are updated, where y i As a binomial distributed random variable used for prediction, p describes the predicted label. i As a random variable with a Dirichlet distribution on a binomial distribution, it describes the prediction uncertainty, and BCE(·) represents the cross-entropy function; (3) Input the new drug-target pair to be predicted back into the deep evidence classification model, perform feature extraction and fusion using step (1), and then use step (2) to obtain the predicted Dirichlet distribution Beta(p) of the drug-target pair. i |α i ,β i The system performs reliable correlation predictions and returns the correlation coefficient of the drug-target pair and the uncertainty coefficient of the prediction as the prediction results to the user.

2. The reliable drug-target correlation prediction method according to claim 1, characterized in that, In step (2), the deep evidence classification model includes two fully connected layers and uses ReLU as the activation function to transform the output into a non-negative value, specifically as follows: Among them, e i This is the output evidence vector. This represents the parameters of the deep evidence classification model.

3. The reliable drug-target correlation prediction method according to claim 2, characterized in that, Using evidence vector e i Determine the predicted beta distribution of the drug-target Beta (p i |α i ,β i The specific method is as follows: According to subjective logic theory, the predicted activity and inactivity probabilities of this drug-target combination are described as follows: The overall uncertainty of this forecast can be described as follows: Where K=2 represents a binary classification prediction problem, i.e., predicting whether the object is active or inactive.

4. The reliable drug-target correlation prediction method according to claim 1 or 3, characterized in that, In step (2), during the training phase of the deep evidence classification model, a Beta loss optimization function is constructed for each drug-target in the training set. Where BCE(·) represents the cross-entropy function; meanwhile, to further limit the number of pieces of evidence appearing in the wrong category, the KL divergence is... As a regularization term, it works together with the Beta distribution as a loss optimization function, as follows: in, and From α i ,β i Beta distribution parameters after removing non-erroneous evidence; λ t ∈[0,1] is the regularization annealing coefficient.

5. The reliable drug-target correlation prediction method according to claim 4, characterized in that, In step (2), during the training phase of the deep evidence classification model, the goal is to minimize the loss optimization function. Query matrix E for structural information of drug and target respectively C E P And the parameters of the evidence module The model parameters are then calculated by taking the derivative and updating synchronously using gradient descent to gradually reduce the loss value to a certain threshold. Finally, the model parameters obtained at this point are saved.

6. The reliable drug-target correlation prediction method according to claim 1, characterized in that, Step (3) specifically includes the following steps: (3-1) The initial SMILES sequence and protein amino acid sequence of the drug-target to be predicted were obtained from the PubChem database and the RCSB PDB database, respectively. (3-2) Use step (1) to extract and fuse features, and use step (2) to perform "credible correlation prediction" on the drug-target pair, and calculate the correlation coefficient of the drug-target pair. The uncertainty coefficient u of the prediction is returned to the user as the prediction result.

Images