High-throughput automatic method for drug screening

Abstract

The method comprises the following steps: acquiring compound data, biological target point data and biological activity data between a compound and a biological target point, establishing a compound and target point data set, performing feature extraction on the compound data and the biological target point data respectively, and obtaining a target point data set; performing feature fusion on the obtained compound feature vector and the biological target feature vector, and constructing a deep learning model; a prediction error of a focus loss function measurement model is defined, a deep learning model is trained, the trained deep learning model is evaluated through a Marius correlation coefficient, a compound and biological target data set is predicted, and candidate drugs are screened out; according to the method, the recognition rate of the model to real interaction is improved by defining the focus loss function, the robustness of the model is enhanced, in addition, a traditional Adam optimizer is improved, and the training efficiency and prediction precision of the model are improved through a quadratic regularization term and an adaptive learning rate strategy.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a high-throughput automated method for drug screening. Background Art

[0002] In recent years, the application of deep learning technology in new drug development has attracted widespread attention, especially in the field of compound-target interaction (CTI) prediction. Traditional CTI prediction methods, such as molecular docking and quantitative structure-activity relationship (QSAR) models, usually rely on a large amount of experimental data and simplified assumptions about molecular interactions, and thus have limitations in scalability and accuracy. Among them, the rise of deep learning has made it possible to build more complex models that can capture the nonlinear relationships inherent in biological systems. For example, technologies such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph neural networks (GNNs) have been used to simulate molecular structures and biological targets, improving prediction performance. Among them, GNNs perform well in molecular graph representation and can effectively encode atomic-level features and molecular topology. In addition, integrating multi-source heterogeneous data, such as molecular structure, protein sequence, and biological activity information, has become a key strategy to enhance model robustness and generalization capabilities. The use of sequence embedding and structural information in protein databases has promoted more accurate representation of biological targets. Advances in optimization algorithms and loss functions have also facilitated the efficient training of deep learning models, enabling their application in large-scale virtual screening and accelerating the process of new drug discovery.

[0003] Despite the above progress, existing deep learning methods still face several challenges in the prediction of CTI (Clinical Trial Information), which limits their effectiveness. First, many models only consider the sequence or structural information of proteins separately, resulting in limited prediction accuracy. This prediction method often ignores the complex relationship between sequence motifs and three-dimensional conformations, which is crucial for understanding protein functions and interactions. Second, traditional graph convolutional networks use a fixed normalization method in molecular representation, which does not consider the difference in node degrees in the molecular graph and may ignore important structural details. In addition, the inherent data imbalance problem in biological datasets, that is, the positive samples (interactions exist) are far less than the negative samples (no interactions), which causes the model to tend to the majority class and reduces the ability to identify true interactions. For example, traditional loss functions, such as binary cross entropy, cannot effectively solve this problem, resulting in poor model performance. Furthermore, commonly used optimization algorithms such as Adam, although they have good convergence characteristics, may lead to overfitting and lack a mechanism to effectively adjust the learning rate during training. These limitations hinder the practical application of existing models in high-throughput drug screening, and it is difficult to meet the high requirements for accuracy and efficiency. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] In view of the above existing problems, the present invention is proposed. Therefore, the present invention provides a high-throughput automated method for drug screening to solve the problems raised in the background technology.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a high-throughput automated method for drug screening, comprising:

[0007] Obtain compound data, biological target data, and biological activity data between compounds and biological targets, and establish compound and target data sets;

[0008] According to the compound and target data sets, feature extraction is performed on the compound data and the biological target data respectively to obtain a compound feature vector and a biological target feature vector, and feature fusion is performed according to the compound feature vector and the biological target feature vector to construct a deep learning model;

[0009] A focal loss function is defined to measure the prediction error of the model, and the deep learning model is trained to minimize the focal loss function. The trained deep learning model is evaluated using the Matthews correlation coefficient, and the compound and biological target data sets are predicted to screen out candidate drugs.

[0010] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, the step of obtaining compound data, biological target data, and biological activity data between the compound and the biological target comprises:

[0011] The chemical structure of the acquired compound data is represented in the form of a molecular graph;

[0012] Divide the acquired biological target data into amino acid sequence and three-dimensional structure information;

[0013] The biological activity data between the acquired compounds and biological targets are divided into quantitative activity data and qualitative activity data.

[0014] As a preferred embodiment of the high-throughput automated method for drug screening of the present invention, wherein: according to the compound and target data set, feature extraction is performed on the compound data to obtain a compound feature vector, including:

[0015] Features are extracted from compound data, features are updated for nodes in the compound data, and an adaptive normalization factor is introduced in the feature update of the nodes.

[0016] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, it further comprises:

[0017] By assigning an initial feature vector to each node, updating the node features, iterating k times, gradually integrating the information of neighboring nodes, and updating the feature representation of each node;

[0018] After obtaining the final feature representation of each node, a pooling operation is performed on the entire molecular graph to obtain the global feature vector of the compound.

[0019] As a preferred embodiment of the high-throughput automated method for drug screening of the present invention, wherein: according to the compound and target data set, feature extraction is performed on the biological target data to obtain a biological target feature vector, including:

[0020] Performing feature extraction on the amino acid sequence in the biological target data, and representing each amino acid residue in the amino acid sequence as a binary vector with a length of 20;

[0021] Set a One-Hot encoding matrix for the amino acid sequence and map the One-Hot encoding matrix to d through a learnable embedding matrix e dimensional continuous vector space, d e Represented as the dimensions of the vector in the embedding matrix;

[0022] Extract the deep features of the amino acid sequence, apply one-dimensional convolution and pooling operations on the embedded amino acid sequence to obtain the local amino acid sequence;

[0023] Extracting features of the three-dimensional structural information in the biological target data, dividing the three-dimensional space into three-dimensional grids of a fixed size, and applying a 3D convolution operation to the three-dimensional grid to extract spatial structural features;

[0024] The spatial structure feature is reconstructed, and a pooling operation is performed on the three-dimensional structure information feature to obtain a local feature vector of the three-dimensional structure information.

[0025] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, it further comprises:

[0026] The amino acid sequence and three-dimensional structure information in the biological target data are combined with local feature vectors to obtain the global feature vector of the biological target.

[0027] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, wherein: feature fusion is performed according to the compound feature vector and the biological target feature vector to construct a deep learning model, including:

[0028] The constructed deep learning model includes feature fusion layer and output layer.

[0029] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, a focal loss function is defined to measure the prediction error of the model, and the deep learning model is trained to minimize the focal loss function, and the trained deep learning model is evaluated by the Matthews correlation coefficient, including:

[0030] The steps in the model training process are as follows:

[0031] Calculate the predicted value of the model output layer;

[0032] Calculate the model loss value using the defined focal loss function;

[0033] Calculate the model iteration value in each iteration process, establish the AdamX update rule through the AdamX optimizer, and update the parameters in the optimizer;

[0034] According to the parameters in the updated optimizer, the iteration value of the next model is obtained, and it is determined whether the iteration value of the next model meets the model convergence condition;

[0035] Repeat the model training process until the model converges and the model training is completed, and the model prediction value is output;

[0036] The predicted values ​​were evaluated using the Matthews correlation coefficient based on the model output.

[0037] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, it further comprises:

[0038] In the AdamX optimizer, a quadratic regularization term and an adaptive learning rate adjustment strategy are introduced.

[0039] As a preferred embodiment of the high-throughput automated method for drug screening described in the present invention, wherein: predicting the compound and biological target data set to screen out candidate drugs includes:

[0040] The compound and biological target data sets are input into the model. If the model output prediction value is positive, that is, there is an interaction between the compound and the biological target data, the current compound is marked as a candidate drug. If the model output prediction value is negative, that is, there is no interaction between the compound and the biological target data, the current compound is marked as a non-candidate drug and discarded.

[0041] Compared with the prior art, the invention has the following beneficial effects:

[0042] 1. The present invention introduces an adaptive normalization factor into the graph convolutional network, fully considers the influence of the node degree in the molecular graph, and enhances the sensitivity to the molecular topological structure; ensures that the model can accurately capture key features when processing complex molecular structures, improves the accuracy of compound feature representation, and thus improves the performance of interaction prediction;

[0043] 2. The amino acid sequence and three-dimensional structure information of biological targets are integrated. On the one hand, the deep learning model can capture the sequence pattern and spatial conformation characteristics of proteins at the same time. On the other hand, it can make the model's description of biological targets more comprehensive and accurate, thereby improving the accuracy of compound-target interaction prediction.

[0044] 3. By defining the focal loss function and dynamically adjusting the weights of difficult and easy samples, the model's learning ability for minority classes (positive samples) is enhanced. The focal loss function can reduce the loss weight of easy-to-classify samples, focusing on the training of difficult samples, improving the model's recognition rate of real interactions, and enhancing the model's robustness and generalization ability.

[0045] 4. Aiming at the problem that the traditional Adam optimizer may cause overfitting and lack an effective learning rate adjustment mechanism, the present invention has made improvements on it. The quadratic regularization term helps prevent the model from overfitting and maintains the stability of the model parameters. In addition, the adaptive learning rate strategy is combined to dynamically adjust the learning rate according to the number of iterations, so that the model has a faster convergence speed in the early stage of training and can finely adjust the parameters in the later stage, thereby improving the training efficiency and prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0047] Figure 1 The overall flow chart of a high-throughput automated method for drug screening according to one embodiment of the present invention;

[0048] Figure 2 A flow chart of model training for a high-throughput automated method for drug screening according to an embodiment of the present invention;

[0049] Figure 3 A diagram showing the influence of the focal loss function on the model robustness of a high-throughput automated method for drug screening according to an embodiment of the present invention;

[0050] Figure 4 This is a diagram showing the influence of the optimizer of the high-throughput automated method for drug screening described in one embodiment of the present invention on the convergence speed of model training. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0054] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0055] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0056] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0057] Example 1

[0058] Reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a high-throughput automated method for drug screening, comprising:

[0059] S1. Acquire compound data, biological target data, and biological activity data between compounds and biological targets, and establish compound and target data sets;

[0060] Specifically, compound data is obtained through a compound database, and the chemical structure of the obtained compound data is represented in the form of a SMILES (Simplified Molecular Linear Input Specification) string or a molecular graph. The molecular graph is used in the solution of the present invention;

[0061] Specifically, biological target data is obtained through the protein database, wherein the biological target data includes amino acid sequence and three-dimensional structure information, wherein proteins are composed of amino acids, and their sequence information is obtained from the PDB database; the three-dimensional structure information is the three-dimensional spatial conformation of the protein, obtained through X-ray crystallography, nuclear magnetic resonance or cryo-electron microscopy technology; for targets without structural data, AlphaFold is used to predict the structure;

[0062] It should be explained that biological target data refers to data that can interact specifically with drugs, compounds or other molecules in the body, thereby affecting molecules or structures of biological processes; biological target data are usually proteins, receptors, enzymes, DNA sequences or RNA fragments, etc.

[0063] Furthermore, the molecular graph in the compound data is represented as G c =(V c ,E c ), the amino acid sequence in the biological target data is represented as S t ={a 1 ,a 2 ,a 3 ,…,a L}, the three-dimensional structure information in the biological target data is expressed as: P t ={(x i ,y i ,z i ,f i )|i=1,2,3,…,N};

[0064] Among them, V c is the set of atomic nodes, E c is the bond edge set; a i Represents the ith amino acid residue; x i ,y i ,z i is the spatial coordinate of the ith atom, f i is the property corresponding to the i-th atom (such as atomic type, charge, etc.);

[0065] It should be noted that by representing molecules as graph structures, graph neural networks (GNNs) can be used to learn the characteristics of molecules;

[0066] Furthermore, the biological activity data between the compound and the biological target include quantitative activity data and qualitative activity data, wherein the quantitative activity data include IC 50 K is the concentration at which the compound reduces the enzyme activity by 50%. i represents the equilibrium constant between the ith inhibitor and the enzyme or receptor, EC 50It indicates the concentration at which the compound causes 50% of the maximum effect; qualitative activity data include activity and inactivity, where activity and inactivity refer to whether a certain compound or drug has a significant biological effect on the target biological target under specific experimental conditions;

[0067] Specifically, the compound and biological target dataset is represented as: {(G c ,S t ,P t ,M)};

[0068] Wherein, M represents the biological activity data between the compound and the biological target;

[0069] S2. Based on the compound and target data sets, feature extraction is performed on the compound data and the biological target data respectively to obtain the compound feature vector and the biological target feature vector, and feature fusion is performed based on the compound feature vector and the biological target feature vector to construct a deep learning model;

[0070] Furthermore, the compound data is subjected to feature extraction, and the features of the nodes in the compound data are updated. The formula is as follows:

[0071]

[0072] in, Represents the feature representation of node v at the kth layer (feature vector of the compound); Represents the set of neighbor nodes of node v, that is, the nodes (atoms) directly connected to node v; represents the feature representation of neighbor node u at the (k-1)th layer; W (k) represents the learnable weight matrix of the kth layer; b (k) represents the bias vector of the kth layer; Z vu Normalization factor between node v and its neighbor node u; ReLU(·) represents the Rectified Linear Unit activation function;

[0073] It needs to be explained that traditional graph convolutional networks usually use a fixed normalization method and do not consider the adaptive normalization of node degrees;

[0074] Specifically, deg(v) represents the degree of node v, that is, the number of edges connected to node v. Similarly, deg(u) represents the degree of neighbor node u, that is, the number of edges connected to neighbor node u.

[0075] It should be noted that by introducing an adaptive normalization factor into the node feature update formula The influence of node degree is taken into account, which enhances the sensitivity of the model to the molecular structure. In addition, the introduction of the normalization factor can prevent the influence of node degree on feature update, ensuring that the subsequent model is robust to molecules with different structures.

[0076] Furthermore, each node (atom) is assigned an initial feature vector Through the above formula Iterate k times (i.e., there are k graph convolution layers), gradually fuse the information of neighboring nodes, and update the feature representation of each node;

[0077] Furthermore, after obtaining the final feature representation of each node After that, a pooling operation is performed on the entire molecular graph, such as global average pooling or global maximum pooling, to obtain the global feature vector h of the compound. c :

[0078]

[0079] Furthermore, feature extraction is performed on the amino acid sequence in the biological target data, and each amino acid residue is represented as a binary vector with a length L of 20 (the number of standard amino acid species), where the position corresponding to the amino acid is 1 and the other positions are 0;

[0080] Furthermore, is the One-Hot encoding matrix of the amino acid sequence, where:

[0081]

[0082] Going a step further, using a learnable embedding matrix Map the One-Hot encoded sequence to d e dimensional continuous vector space;

[0083] Specifically, the mapping formula process is expressed as:

[0084] E s =E one-hot ×W e

[0085] in, is the embedded amino acid sequence representation;

[0086] Furthermore, in order to extract the deep features of the amino acid sequence, it is necessary to apply one-dimensional convolution and pooling operations on the embedded amino acid sequence to capture the local amino acid sequence, which can be expressed as:

[0087] H s =RNN(CNN(E s ))

[0088] in, is the feature vector of the local amino acid sequence;

[0089] Furthermore, the three-dimensional structure information in the biological target data is extracted and the three-dimensional space is divided into three-dimensional grids of fixed size. Among them, d v is the voxel feature dimension, where each voxel contains the feature aggregation of atoms within the grid;

[0090] Furthermore, a 3D convolution operation is applied to the three-dimensional grid V to extract the spatial structure features. The formula is expressed as:

[0091] H p =3D_CNN(V)

[0092] in, Represented as a three-dimensional structural feature vector;

[0093] Specifically, the voxel feature d p It is obtained by aggregating the atomic features located within this voxel;

[0094] Furthermore, the spatial structure features are reconstructed and the pooling operation of the three-dimensional structure information features is performed to obtain the local feature vector of the three-dimensional structure information:

[0095] H p =Pooling({H p |v∈P t ,S t})

[0096] Furthermore, the amino acid sequence and three-dimensional structure information in the biological target data are combined to construct the global feature vector h of the biological target. t ,get:

[0097] h t =(W p ×H p +W s ×H s )

[0098] in, are all represented as learnable weight matrices;

[0099] Furthermore, in the feature fusion layer, the compound feature vector and the biological target feature vector are fused to construct a deep learning model h, and the following is obtained:

[0100] h=tanh(W f [h t ⊙h c ]+b f )

[0101] Among them, W f ∈W p ,W s is the weight matrix of the feature fusion layer, ⊙ is the element-wise product, indicating the multiplication of corresponding elements, and b f is the bias vector of the feature fusion layer; tanh is the hyperbolic tangent activation function;

[0102] Furthermore, at the output layer, the output formula of the deep learning model h is expressed as:

[0103]

[0104] in, is the model prediction output value, which indicates the probability of interaction between the compound and the target. is the weight matrix of the output layer, is the bias of the output layer;

[0105] It should be noted that for deep learning models, it is necessary to define a loss function to measure the prediction error of the model, and use an optimization algorithm to adjust the model parameters to minimize the loss function;

[0106] S3. Define the focus loss function to measure the prediction error of the model, and train the deep learning model to minimize the focus loss function. The trained deep learning model is evaluated by the Matthews correlation coefficient. At the same time, the compound and biological target data sets are predicted to screen out candidate drugs.

[0107] Furthermore, the focal loss function is defined as:

[0108]

[0109] Among them, y i is the true label of the i-th sample, is the predicted probability of the model for the i-th sample, and γ is the control factor used to control the weights of positive and negative samples;

[0110] It should be noted that defining the focal loss function can effectively address the imbalance in the ratio of positive and negative samples and enhance the model's learning ability for the minority class (positive samples);

[0111] It should be explained that the traditional Adam optimizer has the advantages of fast convergence and adaptive learning rate, but it may have problems of overfitting or insufficient generalization ability in the drug screening process. For this reason, the Adam optimizer is improved by introducing a quadratic regularization term and an adaptive learning rate adjustment strategy;

[0112] Furthermore, the AdamX update rule is expressed as:

[0113]

[0114] Among them, θ t is the model parameter of the tth iteration, η t is an adaptive learning rate adjustment strategy, and denote the bias-corrected estimates of the first-order and second-order momentum, respectively, and ∈ is a small constant to prevent division by zero, which is taken as 10 -8 ; is the quadratic regularization term, λ is the regularization strength coefficient;

[0115] Furthermore, the adaptive learning rate adjustment strategy η t It is expressed as:

[0116]

[0117] Among them, τ is the learning rate decay exponent;

[0118] For further reference, Figure 2 , the training process is as follows:

[0119] S301. Calculate the predicted value of the model output layer

[0120] S302, using the focus loss function L to calculate the model loss value;

[0121] S303, calculate the gradient in each iteration process and update and η t ;

[0122] S304, according to the updated and η t , and get the updated θ t+1 Value, judge whether the value meets the model convergence condition;

[0123] S305, repeat S301 to S304 until the model converges and the model training is completed, and the predicted value is output

[0124] It should be noted that the adaptive learning rate strategy can make the model learning rate larger in the early stage of training and gradually decrease in the later stage to prevent the model from overfitting the training data;

[0125] Furthermore, the predicted value is output according to the model Assessed by Matthews correlation coefficient;

[0126] Specifically, the MCC evaluation formula is as follows:

[0127]

[0128] Among them, TP represents true positive examples, which are the number of samples correctly predicted as positive by the model; TN represents true negative examples, which are the number of samples correctly predicted as negative by the model; FP represents false positive examples, which are the number of samples incorrectly predicted as positive by the model; FN represents false negative examples, which are the number of samples incorrectly predicted as negative by the model;

[0129] It should be explained that the Matthews Correlation Coefficient (MCC) is an indicator used to evaluate the performance of a binary classification model and is suitable for use when positive and negative samples are unbalanced.

[0130] Furthermore, the compound and biological target data sets are input into the model. If the model output prediction value is positive, that is, there is an interaction between the compound and the biological target data, the current compound is marked as a candidate drug. If the model output prediction value is negative, that is, there is no interaction between the compound and the biological target data, the current compound is marked as a non-candidate drug and discarded.

[0131] It needs to be explained that when the model predicts an interaction between a compound and a biological target, it indicates that the compound has a biological effect on the target and can therefore be screened as a candidate drug.

[0132] Example 2

[0133] Reference Figure 3 and Figure 4 , which is the second embodiment of the present invention, and which provides a high-throughput automated method for drug screening, comprising: in order to verify the effectiveness of the high-throughput automated drug screening method proposed in the present invention, a group of known compound-target pairs are selected for experiment;

[0134] The data of 5000 compounds, including their chemical structure and biological activity information, were obtained from the public database PubChem. The chemical structure was represented in the form of a molecular graph, where nodes represent atoms and edges represent chemical bonds. For each node, an initial feature vector was assigned, including properties such as atom type and electronegativity. The data of 20 biological targets, including their amino acid sequence and three-dimensional structure information, were obtained from the protein database PDB. The amino acid sequence was represented as a binary vector of length 20 (One-Hot encoding), and then mapped to a 128-dimensional continuous vector space through a learnable embedding matrix. One-dimensional convolution and pooling operations were applied to extract deep features of the sequence. For three-dimensional structural information, the three-dimensional space was divided into three-dimensional grids of fixed size 32×32×32, and 3D convolution operations were applied to extract spatial structural features, and pooling operations were performed to obtain local feature vectors of the three-dimensional structure. The compound feature vector and the biological target feature vector were fused to construct a deep learning model. The focus loss function was defined as the loss function of the model, the regulation factor γ was set to 2, and the AdamX optimizer was used for model training. The quadratic regularization term λ=0.001 and the adaptive learning rate adjustment strategy were introduced, and the initial learning rate η 0 =0.001, learning rate decay exponent τ = 0.5; during the training process, the prediction error of the model gradually decreases, and the focal loss function tends to be minimized; after the training is completed, the Matthews correlation coefficient (MCC) is used to evaluate the model;

[0135] In order to verify the effectiveness of the method of the present invention, a traditional deep learning model (without the introduction of adaptive normalization factor, focal loss function and improved optimizer) was selected as the control group, and the same data set was trained and evaluated to compare the performance differences between the two methods. The results are shown in Table 1.

[0136] Table 1

[0137] parameter Traditional Model Model of the present invention Improvement Accuracy 0.85 0.91 7.06％ Matthews correlation coefficient (MCC) 0.60 0.75 25.00％ AUC value 0.88 0.94 6.82％ Number of drug candidates 120 180 50.00％

[0138] From Table 1, it can be seen that the MCC index is improved from 0.60 of the traditional model to 0.75 of the model of the present invention, with an improvement of 25%. Since MCC has a good evaluation ability for data with imbalanced categories, it shows that the model of the present invention has obvious advantages in dealing with the problem of imbalanced positive and negative samples. In addition, the AUC value (area under the curve) of the model of the present invention is improved from 0.88 to 0.94, an increase of 6.82%. It should be noted that the closer the AUC value is to 1, the stronger the discrimination ability of the model. The improvement of the AUC value shows that the model of the present invention has a stronger ability to distinguish positive and negative samples.

[0139] Combination Figure 3, through the F1 score, we can see that the defined focal loss function enhances the model's learning ability for minority classes (positive samples), making the model's robustness improved by 5% to 12%; reference Figure 4 , setting the number of model training rounds to 50 rounds and introducing the improved AdamX optimizer can enable the model to converge faster in the initial stage of training and fine-tune the model parameters in the later stages of training, thereby improving the model's training efficiency and prediction accuracy, and providing a guarantee for drug screening.

[0140] Those skilled in the art will appreciate that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program codes. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0141] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0142] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0143] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0144] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0145] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A high-throughput automated method for drug screening, characterized in that: include: Obtain compound data, biological target data, and biological activity data between compounds and biological targets, and establish compound and target data sets; According to the compound and target data sets, feature extraction is performed on the compound data and the biological target data respectively to obtain a compound feature vector and a biological target feature vector, and feature fusion is performed according to the compound feature vector and the biological target feature vector to construct a deep learning model; A focal loss function is defined to measure the prediction error of the model, and the deep learning model is trained to minimize the focal loss function. The trained deep learning model is evaluated using the Matthews correlation coefficient, and the compound and biological target data sets are predicted to screen out candidate drugs.

2. The high-throughput automated method for drug screening according to claim 1, characterized in that: Obtain compound data, biological target data, and biological activity data between compounds and biological targets, including: The chemical structure of the acquired compound data is represented in the form of a molecular graph; Divide the acquired biological target data into amino acid sequence and three-dimensional structure information; The biological activity data between the acquired compounds and biological targets are divided into quantitative activity data and qualitative activity data.

3. The high-throughput automated method for drug screening according to claim 2, characterized in that: According to the compound and target data set, feature extraction is performed on the compound data to obtain a compound feature vector, including: Features are extracted from compound data, features are updated for nodes in the compound data, and an adaptive normalization factor is introduced in the feature update of the nodes.

4. The high-throughput automated method for drug screening according to claim 3, characterized in that: Also includes: By assigning an initial feature vector to each node, updating the node features, iterating k times, gradually integrating the information of neighboring nodes, and updating the feature representation of each node; After obtaining the final feature representation of each node, a pooling operation is performed on the entire molecular graph to obtain the global feature vector of the compound.

5. The high-throughput automated method for drug screening according to claim 2 or 4, characterized in that: According to the compound and target data set, feature extraction is performed on the biological target data to obtain a biological target feature vector, including: Performing feature extraction on the amino acid sequence in the biological target data, and representing each amino acid residue in the amino acid sequence as a binary vector with a length of 20; Set a One-Hot encoding matrix for the amino acid sequence and map the One-Hot encoding matrix to d through a learnable embedding matrix e dimensional continuous vector space, d e Represented as the dimensions of the vector in the embedding matrix; Extract the deep features of the amino acid sequence, apply one-dimensional convolution and pooling operations on the embedded amino acid sequence to obtain the local amino acid sequence; Extracting features of the three-dimensional structural information in the biological target data, dividing the three-dimensional space into three-dimensional grids of a fixed size, and applying a 3D convolution operation to the three-dimensional grid to extract spatial structural features; The spatial structure feature is reconstructed, and a pooling operation is performed on the three-dimensional structure information feature to obtain a local feature vector of the three-dimensional structure information.

6. The high-throughput automated method for drug screening according to claim 5, characterized in that: Also includes: The amino acid sequence and three-dimensional structure information in the biological target data are combined with local feature vectors to obtain the global feature vector of the biological target.

7. The high-throughput automated method for drug screening according to claim 4 or 6, characterized in that: Performing feature fusion according to the compound feature vector and the biological target feature vector to construct a deep learning model, including: The constructed deep learning model includes feature fusion layer and output layer.

8. The high-throughput automated method for drug screening according to claim 7, characterized in that: A focal loss function is defined to measure the prediction error of the model, and the deep learning model is trained to minimize the focal loss function. The trained deep learning model is evaluated by the Matthews correlation coefficient, including: The steps in the model training process are as follows: Calculate the predicted value of the model output layer; Calculate the model loss value using the defined focal loss function; Calculate the model iteration value in each iteration process, establish the AdamX update rule through the AdamX optimizer, and update the parameters in the optimizer; According to the parameters in the updated optimizer, the iteration value of the next model is obtained, and it is determined whether the iteration value of the next model meets the model convergence condition; Repeat the model training process until the model converges and the model training is completed, and the model prediction value is output; The predicted values ​​were evaluated using the Matthews correlation coefficient based on the model output.

9. The high-throughput automated method for drug screening according to claim 8, characterized in that: Also includes: In the AdamX optimizer, a quadratic regularization term and an adaptive learning rate adjustment strategy are introduced.

10. The high-throughput automated method for drug screening according to claim 9, characterized in that: Predict compound and biological target data sets to screen candidate drugs, including: The compound and biological target data sets are input into the model. If the model output prediction value is positive, that is, there is an interaction between the compound and the biological target data, the current compound is marked as a candidate drug. If the model output prediction value is negative, that is, there is no interaction between the compound and the biological target data, the current compound is marked as a non-candidate drug and discarded.

Images