Drug screening method, device and storage medium

Abstract

The present invention discloses a drug screening method, device, and storage medium. The method comprises obtaining a number of drugs corresponding to a virus to be tested; for each of the drugs, inputting the drug molecule of the drug and the base sequence of the virus to be tested into a trained drug screening model, and determining the effective value corresponding to the drug using the drug screening model; and determining a target drug corresponding to the virus to be tested based on the effective value, wherein the target drug is one or more of the drugs. The present invention aims to use deep learning based on a set of drug-virus pairs to screen the drug using a trained drug screening model. This can improve the computational performance of the drug screening model, significantly shorten the drug screening cycle, and thus shorten the research and development cycle of drugs for viral infections, thereby improving the efficiency of drug screening.

Description

Technical Field

[0001] The present invention relates to the field of auxiliary drug technology, and in particular to a drug screening method, device and storage medium. Background Art

[0002] Drug development for novel viruses is a long and costly process. Early drug screening typically relied on manual screening, requiring significant manpower and resources and suffering from a series of drawbacks, including long experimental cycles. With the continuous advancement of science and technology in recent years, artificial intelligence (AI) technology has emerged, playing a key role in drug screening and discovery. However, research on AI-based drug screening and discovery is still in its infancy, requiring extensive computational effort and cumbersome operations. This protracted drug screening process significantly impacts the effectiveness of mitigating viral infections.

[0003] Therefore, the existing technology needs to be improved and developed. Summary of the Invention

[0004] Based on this, the present invention provides a drug screening method, device and storage medium to solve the problem of long time consumption in drug screening methods.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A drug screening method comprising the following steps:

[0007] Obtain several drugs corresponding to the virus to be tested;

[0008] For each of the plurality of drugs, the drug molecule of the drug and the base digital sequence of the virus to be tested are input into a trained drug screening model, and the effective value corresponding to the drug is determined by the drug screening model;

[0009] The target drug corresponding to the virus to be tested is determined according to the effective value, wherein the target drug is one or more of several drugs.

[0010] Optionally, the drug screening model is obtained based on training of a training set, the training set including multiple groups of training data pairs, each group of training data pairs including the drug molecules of the training drug and the base digital sequences of the training virus, and different groups of training data have the same training virus.

[0011] Optionally, the training process of the drug screening model includes:

[0012] For each of the plurality of training data pairs, the drug molecule of the training drug and the base sequence of the training virus are input into the drug screening model, and the drug screening model outputs a predicted value of the training drug-training virus; wherein the predicted value is used to indicate whether the training drug effectively inhibits the training virus;

[0013] The parameters of the drug screening model are modified according to the predicted value until the predicted value output by the drug screening model reaches the expected value, and then the training is stopped to obtain a trained drug screening model.

[0014] Optionally, the drug screening model includes a long short-term memory network model, a graph convolutional neural network model, a feature fusion unit, and a multi-layer perceptron model. For each of the multiple sets of training drug-training virus pairs, the drug molecule of the training drug and the base sequence of the training virus are input into the drug screening model, and the output of the predicted value of the training drug-training virus by the drug screening model specifically includes:

[0015] Inputting the base digital sequence of the training virus into the long short-term memory network model, and outputting the viral gene sequence characteristics of the training virus through the long short-term memory network model;

[0016] Inputting a drug molecule graph of a training drug into a graph convolutional neural network model, and outputting drug molecule features of the training drug through the graph convolutional neural network model;

[0017] Inputting the drug molecular features and the virus gene sequence features into a feature fusion unit, and outputting the fusion features of the training drug-training virus pair through the feature fusion unit;

[0018] The fusion features are input into a multi-layer perceptron model, and the multi-layer perceptron model outputs the predicted value of the training drug-training virus.

[0019] Optionally, the long short-term memory network model includes a plurality of long short-term memory network units connected in sequence, and inputting the base sequence of the training virus into the long short-term memory network model, and outputting the viral gene sequence characteristics of the training virus through the long short-term memory network model specifically includes:

[0020] The first number in the base number sequence is used as the target number and input into the first long short-term memory network unit, and the intermediate viral gene sequence feature is output by the first long short-term memory network unit, and the intermediate viral gene sequence feature is used as the target gene sequence feature;

[0021] Inputting the next digit of the first digit as the target digit and the target gene sequence feature into the next long short-term memory network unit of the first long short-term memory network unit, and obtaining the next target gene feature through the next long short-term memory network unit;

[0022] Repeat the steps until the last digit of the base number sequence is input as the target digit and the previous target gene sequence feature into the last long short-term memory network unit, and the next digit of the first digit is output as the target digit and the first target gene feature into the viral gene sequence feature of the training virus through the output of the last long short-term memory network unit.

[0023] Optionally, the interval time between inputting the target digit into the long short-term memory network unit is the same each time.

[0024] Optionally, the graph convolutional neural network model includes an input layer, a multi-cascade hidden layer, and an output layer. Inputting the drug molecules of the training drug into the graph convolutional neural network model, and outputting the drug molecular features of the training drug through the graph convolutional neural network model specifically include:

[0025] Inputting drug molecules of training drugs into the input layer, and obtaining a drug molecule graph through the input layer;

[0026] The drug molecule graph is input into the first hidden layer, and the first drug molecule characteristic graph is obtained after the output of the first hidden layer is activated by an activation function;

[0027] Inputting the first drug molecular characteristic graph as the target drug molecular graph into the next hidden layer of the first hidden layer, and obtaining the next drug molecular characteristic graph after the output of the next hidden layer is subjected to the same activation function, until the previous drug molecular characteristic graph is input as the target drug molecular graph into the last hidden layer, and the last drug molecular characteristic graph is output through the last hidden layer;

[0028] The last drug molecular feature map is input to the output layer, and the drug molecular features of the training drug are output through the output layer.

[0029] Optionally, the multilayer perceptron model includes an input layer, a multi-cascade hidden layer, and an output layer. Inputting the fusion feature into the multilayer perceptron model and outputting the predicted value of the training drug-training virus through the multilayer perceptron model specifically includes:

[0030] Inputting the fused features into the input layer, and obtaining the initial fused features through the input layer;

[0031] Inputting the initial fusion feature into a first-level hidden layer, and outputting a first target feature through the first-level hidden layer;

[0032] Inputting the first target feature as a target into the next hidden layer of the first hidden layer, outputting the next target feature through the next hidden layer, until inputting the previous target feature as a target into the last hidden layer, and outputting the last target feature through the last hidden layer;

[0033] The last target feature is input into the output layer, and the predicted value of the training drug-training virus is output through the input layer.

[0034] Optionally, the trained drug screening model includes a long short-term memory network model, a graph convolutional neural network model, a feature fusion unit, and a multilayer perceptron model. For each of the plurality of drugs, the drug molecule of the drug and the base sequence of the virus to be tested are input into the trained drug screening model, and determining the effective value corresponding to the drug by the drug screening model specifically includes:

[0035] Obtaining the drug molecule of each drug in each group and the base digital sequence of the virus to be tested; wherein each drug in the same group corresponds to the same virus to be tested;

[0036] Inputting each of the drug molecules into the graph convolutional neural network model in sequence, and simultaneously inputting the base sequence of the virus to be tested into the long short-term memory network model;

[0037] After being processed by the graph convolutional neural network model and the long short-term memory network model, they are jointly output to the feature fusion unit. After being processed by the feature fusion unit, the feature fusion unit outputs to the search multi-layer perceptron model. After being processed by the multi-layer perceptron model, the effective value corresponding to each of the drugs is output in sequence.

[0038] Optionally, obtaining the base sequence of the virus to be detected specifically includes:

[0039] Obtaining the base letter sequence of the virus to be tested;

[0040] The base letter sequence is mapped into a base number sequence, wherein identical letters are converted to the same number.

[0041] Based on the above method, the present invention also provides a drug screening device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned drug screening method when executing the computer program.

[0042] Based on the above method, the present invention also provides a computer-readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps of the drug screening method described above.

[0043] Beneficial effects:

[0044] Compared to the prior art, the present invention provides a drug screening method, device, and storage medium. The method includes obtaining a number of drugs corresponding to a virus to be tested; for each of the several drugs, inputting the drug molecule of the drug and the base sequence of the virus to be tested into a trained drug screening model, and determining the effective value corresponding to the drug through the drug screening model; and determining the target drug corresponding to the virus to be tested based on the effective value, wherein the target drug is one or more of the several drugs. The present invention aims to perform deep learning based on a set of drug-virus pairs to screen a trained drug screening model. This can improve the computational performance of the drug screening model, greatly shorten the drug screening cycle, and thus shorten the research and development cycle of drugs for viral infectious diseases, thereby improving the efficiency of drug screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The present invention provides a flow chart of a drug screening method.

[0046] Figure 2 This is a structural diagram of the drug screening model provided by the present invention.

[0047] Figure 3 This is a structural block diagram of the long short-term memory network model in the drug screening model provided by the present invention.

[0048] Figure 4 This is a schematic diagram of the state operation of the long short-term memory network model in the drug screening model provided by the present invention.

[0049] Figure 5 This is a diagram of the internal structure of the forget gate in the long short-term memory network unit in the long short-term memory network model provided by the present invention.

[0050] Figure 6 This is a block diagram of the internal structure of the input gate in the long short-term memory network unit in the long short-term memory network model provided by the present invention.

[0051] Figure 7 This is a block diagram of the internal structure of the output gate in the long short-term memory network unit in the long short-term memory network model provided by the present invention.

[0052] Figure 8This is a schematic diagram of the structure of the graph convolutional neural network model in the drug screening model provided by the present invention.

[0053] Figure 9 This is a schematic diagram of the structure of the multilayer perceptron model in the drug screening model provided by the present invention.

[0054] Figure 10 This is a structural block diagram of a drug screening device provided by the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] First, the relevant terms of the present invention are explained to facilitate understanding of the technical solution of the present invention.

[0057] The most basic component of a neural network is the neuron, often referred to as a node or unit. A node receives input from other nodes or from external sources and computes an output. Each input is assisted by a weight (w), which depends on the relative importance of the other inputs.

[0058] The role of the activation function is to introduce nonlinearity into the output of the neuron, so that the neuron learns the nonlinear function representation. Each activation function receives a number and performs a specific, fixed mathematical calculation. The activation functions used in this embodiment are as follows:

[0059] Sigmoid (S-type activation function): input a real value and output a value between 0 and 1;

[0060] Tanh (hyperbolic tangent function): input a real value and output a value in [-1, 1];

[0061] ReLu: stands for Rectified Linear Unit. It outputs a real value and has a threshold of 0, meaning that the function converts negative values ​​to zero.

[0062] The Long Short-Term Memory (LSTM) model is a type of recurrent neural network (RNN) designed to address the vanishing and exploding gradient problems associated with long training sequences. Simply put, compared to standard RNNs, LSTMs perform better on longer sequences. LSTM-based systems can learn to translate languages, control robots, analyze images, summarize documents, perform speech and image recognition, recognize handwriting, control chatbots, predict diseases, and synthesize music.

[0063] Graph Convolutional Networks (GCN) extracts features from graphs with points and edges. Simply put, it performs node classification, graph classification, and link prediction on graph data, and can also obtain graph embedding.

[0064] The Multilayer Perceptron (MLP) model, also known as a multilayer perceptron, is a feedforward artificial neural network model that maps multiple input data sets to a single output data set. It includes at least one hidden layer (in addition to an input layer and an output layer). While a single-layer perceptron can only learn linear functions, a multilayer perceptron can learn not only linear functions but also nonlinear functions.

[0065] The input nodes are called input layers. No calculations are performed in the input nodes, and information is only passed to the hidden nodes.

[0066] Hidden nodes need to perform calculations and pass information from input nodes to output nodes, and are also called hidden layers.

[0067] The output node is also called the output layer, which is responsible for calculation and transmitting information out of the network.

[0068] Overfitting means making the hypothesis overly strict in order to obtain a consistent hypothesis. Simply put, if a hypothesis can obtain a better fit than other hypotheses on the training data, but cannot fit the data well on the data set outside the training data, then the hypothesis is considered to be overfitting.

[0069] The loss function refers to the average of the sum of squares of the differences between the model's predicted results and the true values. The more accurate the prediction, the smaller the loss function.

[0070] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, which are as follows:

[0071] See also Figure 1 , Figure 1 This is a flow chart of a drug screening method provided by the present invention. It should be noted that the drug screening method of the embodiment of the present invention is not limited to Figure 1 The steps and order in the flowchart shown can be added, removed or changed in order according to different needs.

[0072] like Figure 1As shown, the drug screening method provided by the present invention comprises the following steps:

[0073] S10. Obtain several drugs corresponding to the virus to be tested.

[0074] In this embodiment, the multiple drugs are one or more drugs that are believed to be able to treat the virus to be tested, which are searched from a large-scale drug database. Each of the multiple drugs may be an effective inhibitor, that is, the drug can effectively inhibit the virus to be tested, or may be an ineffective inhibitor, that is, the drug cannot inhibit the virus to be tested. Therefore, whether the multiple drugs obtained are effective inhibitors of the same virus to be tested needs to be determined through a trained drug screening model.

[0075] In order to screen out drugs that are effective in inhibiting the virus to be tested, it is necessary to obtain sample pairs, that is, to test each of several drugs in combination with the virus to be tested.

[0076] S20. For each of the plurality of drugs, the drug molecule of the drug and the base digital sequence of the virus to be tested are input into a trained drug screening model, and the effective value corresponding to the drug is determined by the drug screening model.

[0077] In this embodiment, the drug screening model refers to a drug screening model with generalization ability. The effective value refers to the degree to which the drug effectively inhibits the test virus. In other words, the effective value can be used to determine whether the drug effectively inhibits the test virus or not.

[0078] This drug screening model can quickly identify drugs that effectively inhibit the test virus through a primary screening. It then conducts a secondary screening of all the drugs identified in the primary screening, ultimately identifying one or more drugs that effectively inhibit the test virus. This initial screening using this drug screening model significantly shortens the drug screening and R&D cycle, greatly reducing the probability of missing a target drug during the drug data collection phase and improving the success rate of drug screening and R&D.

[0079] Furthermore, the drug screening model is trained based on a training set, which includes multiple sets of training data pairs, each set of training data pairs including the drug molecules of the training drugs and the base digital sequences of the training viruses, and different sets of training data have the same training viruses. The training drugs are several drugs in a large-scale drug database that may effectively inhibit the same virus to be tested, and the training viruses are pathogens. To improve efficiency and reduce complexity, the data input to the drug screening model to be trained is the original data, i.e., the training viruses and the training drugs corresponding to the training viruses, after processing the base digital sequences of the training viruses and the drug molecules of the training drugs.

[0080] For example, the training virus is Ribonucleoside-diphosphate reductase large subunit, and its base sequence is:

[0081] ATGCATGTGATCAAGCGAGATGGCCGCCAAGAACGAGTCATGTTTGACAAAATTA CATCTCGAATCCAGAAGCTTTGTTATGGACTCAATATGGATTTTGTTGATCCTGCTCAGATCACCATGAAAGTAATCCAAGGCTTGTACAGTGGGGTCACCACAGTGGAACTAGATAC TTTGGCTGCTGAAACAGCTGCAACCTTGACTACTAAGCACCCTGACTATGCTATCCTGGCAGCCAGGATCGCTGTCTCTAACTTGCACAAAGAAACAAAGAAAGTGTTCAGTGATGTG ATGGAAGACCTCTATAACTACATAAATCCACATAATGGCAAACACTCTCCCATGGTGGCCAAGTCAACATTGGATATTGTTCTGGCCAATAAAGATCGCCTGAATTCTGCTATTATCT ATGACCGAGATTTCTCTTACAATTACTTCGGCTTTAAGACGCTAGAGCGGTCTTATTTGTTGAAGATCAATGGAAAAGTGGCTGAAAGACCACAACATATGTTGATGAGAGTATCTGT TGGGATCCACAAAGAAGACATTGATGCAGCAATTGAAACATATAATCTTCTTTCTGAGAGGTGGTTTACTCATGCTTCGCCCACTCTCTTCAATGCTGGTACCAACCGCCCACAACTT TCTAGCTGTTTTCTTCTGAGTATGAAAGATGACAGCATTGAAGGCATTTATGACACTCT AAAGCAATGTGCATTGATTTCTAAGTCTGCTGGAGGAATTGGTGTTGCTGTGAGTTGTATTCGGGCTACTGGCAGCTACATTGCTGGGACTAATGGCAATTCCAATGGCCTTGTACCGATGCTGAGAGTATATAACAACACAGCTCGATATGTGGATCAAGGTGGGAACAAGCGTCCTGGGGCATTTGCTATTTACCTGGAGCCTTGGCATTTAGACATCTTTGAATTCCTTGATTTAAAGAAGAACACAGGAAAGGAAGAGCAGCGTGCCAGAGATCTTTTCTTTGCTCTTTGG ATTCCGGATCTCTTCATGAAACGAGTGGAGACTAATCAGGACTGGTCTTTGATGTGTCCAAATGAGTGTCCTGGTCTGGATGAGGTTTGGGGAGAGGAATTTGAGAAACTATATGCAA GTTATGAGAAACAAGGTCGTGTCCGCAAAGTTGTAAAAGCTCAGCAGCTTTGGTATGCC ATCATTGAGTCTCAGACGGAAACAGGCACCCCGTATATGCTCTACAAAGATTCCTGTAATCGAAAGAGCAACCAGCAGAACCTGGGAACCATCAAATGCAGCAACCTGTGCACAGAAA TAGTGGAGTACACCAGCAAAGATGAGGTTGCTGTTTGTAATTTGGCTTCCCTGGCCCTGAATATGTATGTCACATCAGAACACACATACGACTTTAAGAAGTTGGCTGAAGTCACTAA AGTCGTTGTCCGAAACTTGAATAAAATTATTGATATAAACTACTATCCTGTACCAGAGGCATGCCTATCAAATAAACGCCATCGCCCCATTGGAATTGGGGTACAAGGTCTGGCAGAT GCTTTTATCCTGATGAGATACCCTTTTGAGAGTGCAGAAGCCCAGTTACTGAATAAGCAGATCTTTGAAACTATTTATTATGGTGCTCTGGAAGCCAGCTGTGACCTTGCCAAGGAGC AGGGCCCATACGAAACCTATGAGGGCTCTCCAGTTAGCAAAGGAATTCTTCAGTATGAT ATGTGGAATGTTACTCCTACAGACCTATGGGACTGGAAGGTTCTCAAGGAGAAGATTGCAAAGTATGGTATAAGAAACAGTTTACTTATTGCCCCGATGCCTACAGCTTCCACTGCTCAGATCCTGGGGAATAATGAGTCCATTGAACCTTACACCAGCAACATCTATACTCGCAGAGTCTTGTCAGGAGAATTTCAGATTGTAAATCCTCACTTATTGAAAGATCTTACCGAGCG GGGCCTATGGCATGAAGAGATGAAAAACCAGATTATTGCATGCAATGGCTCTATTCAGA GCATACCAGAAATTCCTGATGACCTGAAGCAACTTTATAAAACTGTGTGGGAAATCTCTCAGAAAACTGTTCTCAAGATGGCAGCTGAGAGAGGTGCTTCATTGATCAAAGCCAATC TTTGAACATCCACATTGCTGAGCCTAACTATGGCAAACTCACTAGTATGCACTTCTACGGCTGGAAGCAGGGTTTGAAGACTGGGATGTATTATTTAAGGACAAGACCAGCGGCTAAT CCAATCCAGTTCACTCTAAATAAGGAGAAGCTAAAAGATAAAGAAAAGGTATCAAAAGAGGAAGAAGAGAAGGAGAGGAACACAGCAGCCATGGTGTGCTCTTTGGAGAATAGAGATG AATGTCTGATGTGTGGATCCTGA

[0082] The converted base number sequence (wherein the letter A in the base letter sequence is mapped to the number 1, the letter T is mapped to the number 2, the letter G is mapped to the number 3, and the letter C is mapped to the number 4) is: 12341232312411343131233443441131143132412322231411112214124 2431124413113422232212331424112123312222322312442342413124144123111321124411334223214132333324144141323311421312142223 34234231114134234114422314214211341444231421234212442334134413312434232424211422341411131114111311132322413231232312331131442421211421412111244141211233411141424244412332334411324114122331212232242334411211131243442311224234212212421231 4431312224242214112214224334222113143421313433242212223223113124112331111323342311131441411412123223123131321242322333 1244141113113141223123413411223111412121124224222423131332332221424123422434441424242241123423321441144344414114222421 3423222242242313212311131231413412231133412221231414242111341123234122312224211324234233133112233232234232313223212243 33421423341342141223423331421123341122441123344223214431234231313212121141141413424312123233124113323331141134324423333412223421222144233134422334122213141242223112244223122211131131141413311133113134134323441313124222242223424222331224 43312424224123111431323313142112413314233242223123232441112 3132324423324233123133222333313133112223131114212123411322123131114113324323244341113223211113424134134222332123441241 2231324241314331114133414444321212342421411131224423211243111313411441341311442333114412411123413411442323414131112132 3313214144134111312313322342322232112223342244423344423112123212324141241311414141214314222113113223342311324142111324 32232443111422311211112212231212111421421244232144131334123 4421241112111434412434444122331122333321411332423341312342222124423123131214442222313132341311344413221423112113413124 2223111421222122123323424233113441342323144223441133134133344412143111442123133342424413221341113311224224132123121232 33112322142442141314421233314233113322424113313113122341113212332121131114132221422122344443123442141342244142342413124423333112112313244122311442214144134114124212142434131324223241331311222413122321112442414221223111312422144313433334 42123341231131312311111441312212234123411233424212241313412 1441311122442312314423113411422212111142323233311124242413111142322424113123341342313131332342224122312411134411242223 1141244141223423134421142123341114241421321234142242143342331134133322231131423331232122122211331411314413433421124411 2441322414242111211331311342111131211131111332124111131331131131313131313311414134134412332323424222331311213131231123 242312323331244231

[0084] One of the training drugs used to treat the training virus is ID: DB00441, named: Gemcitabine, and its drug molecule is:

[0085]

[0086] The training process of the drug screening model is described in detail below.

[0087] See also Figure 2 , Figure 2 The structural diagram of the drug screening model is illustrated.

[0088] The training process of the drug screening model is:

[0089] M10, for each of the plurality of training data pairs, inputting the drug molecule of the training drug and the base sequence of the training virus into the drug screening model, and outputting a predicted value of the training drug-training virus pair by the drug screening model; wherein the predicted value indicates whether the training drug is effective against the training virus;

[0090] In this embodiment, the drug screening model is pre-constructed. Figure 2 As shown, the drug screening model includes a long short-term memory network model, a graph convolutional neural network model, a feature fusion unit and a multi-layer perceptron model, wherein the output end of the long short-term memory network model and the output end of the graph convolutional neural network model are connected to the input end of the feature fusion unit, and the output end of the feature fusion unit is connected to the input end of the multi-layer perceptron model.

[0091] The long-short-term memory network model and graph convolutional neural network model are used to extract features of the training virus and training drug, respectively. The feature fusion unit is used to fuse the features of the training virus and training drug. The multi-layer perceptron model is used to screen and calculate the fused features to output a prediction value of whether the training drug is effective against the training virus. The prediction value is divided into effective prediction value and ineffective prediction value. The ratio of effective prediction value to ineffective prediction value is used to test whether the training of the drug screening model meets the requirements.

[0092] (1) In specific implementation, during the feature extraction process, for the training virus:

[0093] The base digital sequence of the training virus is input into the long short-term memory network model in the drug screening model, and the viral gene sequence characteristics of the training virus are output through the long short-term memory network model.

[0094] In this embodiment, research has revealed that viral genes are single-stranded (RNA viruses) or double-stranded (DNA viruses) base sequences, composed of various combinations of the letters AGCT, also known as base letter sequences. Depending on the type of virus, the length of these sequences can range from several thousand to several hundred thousand. For example, the base letter sequence of a training virus can be represented by GATAAAAACCATGTGG... Thus, by representing the complex structure of the training virus as a sequence of letter combinations, not only is the training computation speed increased, but the complex structure is simplified, saving space and significantly reducing computational complexity.

[0095] To further simplify the process and speed up screening, the training virus's base letter sequence is mapped to a base numeric sequence. This involves converting each letter in the base letter sequence into a number, with identical letters converted to the same number, to obtain the training virus's base numeric sequence. Therefore, by extracting features from the base numeric sequence, the viral gene sequence characteristics of the training virus can be obtained.

[0096] Furthermore, in this embodiment, the long short-term memory network model is an LSTM model, which includes a plurality of long short-term memory network units connected in sequence, and the parameters and structure of each long short-term memory network unit are the same.

[0097] Specifically, the step of inputting the base sequence of the training virus into the long short-term memory network model and outputting the viral gene sequence features of the training virus through the long short-term memory network model specifically includes:

[0098] The first number in the base number sequence is used as the target number and input into the first long short-term memory network unit, and the intermediate viral gene sequence feature is output by the first long short-term memory network unit, and the intermediate viral gene sequence feature is used as the target gene sequence feature;

[0099] Inputting the next digit of the first digit as the target digit and the target gene sequence feature into the next long short-term memory network unit of the first long short-term memory network unit, and obtaining the next target gene feature through the next long short-term memory network unit;

[0100] Repeat the steps until the last digit of the base number sequence is input as the target digit and the previous target gene sequence feature into the last long short-term memory network unit, and the next digit of the first digit is output as the target digit and the first target gene feature into the viral gene sequence feature of the training virus through the output of the last long short-term memory network unit.

[0101] Furthermore, the interval time between inputting the target digit into the LSTM network unit is the same each time.

[0102] That is, the numbers in the base sequence are represented by X, and each time t corresponds to a number Xt. The number Xt corresponding to the current time t and the output of the previous long short-term memory network unit are used as the input of the next long short-term memory network unit, and the output of the last long short-term memory network is used as the final viral gene sequence feature of the training virus.

[0103] The specific operation process is as follows Figure 4 As shown, the number Xt corresponding to the current moment t and the Ht-1 passed down from the previous state are spliced ​​together to obtain four states Z f , Z i ,Z,Z 0 , where Z f , Z i , Z 0The concatenated vector is multiplied by the weight matrix and then converted to a value between 0 and 1 through a sigmoid activation function, which acts as a gating state. Z represents the result converted to a value between -1 and 1 through a tanh activation function.

[0104] in, Figure 4 In the matrix, ⊙ is used to represent the multiplication of corresponding elements in the operation matrix; Used to represent matrix addition; W represents the parameters of the model, which are randomly initialized and updated according to the model training process.

[0105] The formulas involved are as follows:

[0106] c t =z f ⊙c t-1 +z i ⊙z

[0107] h t =z 0 ⊙tanh(c t )

[0108] y t =σ(W′h t )

[0109] Furthermore, the internal structure of each LSTM network unit is as follows Figure 3 As shown, Figure 3 only illustrates three sequentially connected long short-term memory network units, A is used to represent a long short-term memory network unit, Ht is used to represent the intermediate viral gene sequence features output by each long short-term memory network unit, and σ is used to represent the nonlinear activation function.

[0110] The LSTM network unit is a gate structure, and its internal structure includes input gate, output gate and forget gate. Figure 4-Figure 6 , Figure 4 The schematic diagram of the forget gate structure in the long short-term memory network unit is illustrated. Figure 5 The schematic diagram of the input gate structure in the long short-term memory network unit is illustrated. Figure 6 An example is shown of a schematic diagram of the output gate structure in the long short-term memory network unit.

[0111] like Figure 4 As shown in the figure, the forget gate is used to determine which information in the previous moment is discarded from the unit state. That is, a vector ft is calculated based on the current input Xt and the output ht-1 of the previous moment, and its value in each dimension is in the range of (0, 1). Then, the ft vector Ct-1 of the previous moment is bitwise multiplied by the ft vector. In this way, the information in the dimensions where the ft value is close to 0 will be forgotten, while the information in the dimensions where the ft value is close to 1 will be retained.

[0112] like Figure 5 As shown in Figure 2, the input gate is used to determine what new information is stored in the cell state. The input gate determines what information is added to Ct-1 to generate the new state Ct based on xt and ht-1.

[0113] The input gate has two steps:

[0114] first step( Figure 6 (Middle left image): The sigmoid activation function determines which information needs to be updated; the tanh activation function generates a vector that maps values ​​to (-1, 1), which is the candidate content for update. The two parts together form the input gate:

[0115] Step 2 Figure 6 Middle right image): Superimposing the old state and the new state (+) gives us the new state (forget what we decided to forget and add the new state).

[0116] like Figure 7 As shown, the output gate is used to decide what to output.

[0117] In this way, in order to solve the problem that the base sequence of viral genes is too long and it is difficult to extract features, the long short-term memory network LSTM, which is specially used to process sequence data, is used to model the viral gene sequence, thereby effectively extracting viral features.

[0118] (2) During the feature extraction process, for drug molecules:

[0119] Research has found that drug molecules can be represented as graph data. The nodes in the graph represent each atom in the drug molecule, and the edges in the graph represent the chemical bonds connecting each atom. Specifically, N represents the number of nodes, each with its own unique features, and D represents the dimension of the features. Therefore, the node features form an N×D-dimensional matrix X. The relationships between the nodes also form an N×N-dimensional matrix A, also known as the adjacency matrix. This simplifies the structure, facilitates subsequent calculations, reduces computational complexity, and thus improves feature extraction efficiency. Therefore, training drugs are converted into drug molecule graphs as input data.

[0120] During specific implementation, the drug molecule graph of the training drug is input into a graph convolutional neural network model, and the drug molecule characteristics of the training drug are output through the graph convolutional neural network model.

[0121] The graph convolutional neural network model constructed in the present invention is a GCN model, which includes an input layer, a multi-cascade hidden layer and an output layer. Figure 8 shown.

[0122] The step of inputting the drug molecule graph of the training drug into the graph convolutional neural network model and outputting the drug molecule features of the training drug through the graph convolutional neural network model specifically includes:

[0123] Inputting drug molecules of training drugs into the input layer, and obtaining a drug molecule graph through the input layer;

[0124] The drug molecule graph is input into the first hidden layer, and the first drug molecule characteristic graph is obtained after the output of the first hidden layer is activated by an activation function;

[0125] Inputting the first drug molecular characteristic graph as the target drug molecular graph into the next hidden layer of the first hidden layer, and obtaining the next drug molecular characteristic graph after the output of the next hidden layer is subjected to the same activation function, until the previous drug molecular characteristic graph is input as the target drug molecular graph into the last hidden layer, and the last drug molecular characteristic graph is output through the last hidden layer;

[0126] The last drug molecular feature map is input to the output layer, and the drug molecular features of the training drug are output through the output layer.

[0127] In simple terms, Figure 8 In the two cascade hidden layers shown in the example, the drug molecule graph is input into the input layer, which is then output to the first hidden layer (hidden) through the input layer. The first drug molecule characteristic graph is output through the first hidden layer. The first drug molecule characteristic graph is input into the second hidden layer after passing through the ReLU function. The second drug molecule characteristic graph is output through the second hidden layer. The second drug molecule characteristic graph is input into the output layer (output) after passing through the ReLU function again. The data output by the output layer needs to undergo a linear change to obtain the drug molecular characteristics, which is the final drug molecular characteristics of the training drug.

[0128] In this embodiment, the ReLU function is used as the activation function in order to convert the negative values ​​in the data output by the hidden layer into 0, but the positive values ​​remain unchanged, thereby achieving unilateral inhibition and making the data in the neural network sparsely activated, thereby greatly reducing the error.

[0129] The propagation formula involved between the graph convolutional neural network models is:

[0130]

[0131] Adding the adjacency matrix A to the identity matrix I introduces the feature of node self-connection.

[0132] The matrix D is defined as follows:

[0133] The matrix D is a diagonal matrix, and the values ​​on its diagonal positions are the degrees of the corresponding nodes. yes The degree matrix of each layer. H is the feature of each layer. For the input layer, H is X; σ is the nonlinear activation function.

[0134] In this way, in order to solve the problem that the characteristics of large molecules and complex drug molecules are difficult to learn, graph networks can be used to effectively analyze and obtain their intrinsic characteristics.

[0135] (3) Feature fusion processing

[0136] After extracting the drug molecular features and the virus gene sequence features, the drug molecular features and the virus gene sequence features need to be input into a feature fusion unit, and the feature fusion unit outputs the fusion features of the training drug-training virus pair; that is,

[0137] (IV) Features after training fusion

[0138] The fusion features are input into a multi-layer perceptron model, and the multi-layer perceptron model outputs the predicted value of the training drug-training virus.

[0139] In this embodiment, the multilayer perceptron model is an MLP model, which includes an input layer, multiple cascaded hidden layers, and an output layer. The dimensions of each hidden layer can be the same or different, and are set according to actual needs. Of course, the number of hidden layers is also set according to actual needs. Inputting the fused features into the multilayer perceptron model and outputting the predicted value of the training drug-training virus through the multilayer perceptron model specifically includes:

[0140] Inputting the fused features into the input layer, and obtaining the initial fused features through the input layer;

[0141] Inputting the initial fusion feature into a first-level hidden layer, and outputting a first target feature through the first-level hidden layer;

[0142] Inputting the first target feature as a target into the next hidden layer of the first hidden layer, outputting the next target feature through the next hidden layer, until inputting the previous target feature as a target into the last hidden layer, and outputting the last target feature through the last hidden layer;

[0143] The last target feature is input into the output layer, and the predicted value of the training drug-training virus is output through the input layer.

[0144] In simple terms, Figure 9 As shown in the figure, the first layer is the input layer, with a dimension of 128. For simplicity, only 6 nodes are drawn in the figure; the second layer is the first hidden layer, with a dimension of (128, 256); the third layer is the second hidden layer, with a dimension of (256, 5); and the fourth layer is the output layer, with a dimension of (5, 2). The training process is:

[0145] The fused features are input to the input layer, output to the first hidden layer through the output layer, and then filtered and output to the second hidden layer. The filtering operation is repeated again, and finally the output layer outputs a predicted value, leaving only the category values ​​01 and 10. During the screening process, the first and second hidden layers perform calculations to filter out the fused features of 11 and 00, leaving only the outputs of 01 and 10. Based on the predicted values, the probability of effectiveness and ineffectiveness is calculated. The predicted values ​​are then compared with the expected values ​​to modify the parameters of the drug screening model, so that the predicted values ​​output by the trained drug screening model meet the expected values.

[0146] M11, modifying the parameters of the drug screening model according to the predicted value until the predicted value output by the drug screening model reaches the expected value and then stopping the training to obtain a drug screening model with generalization ability.

[0147] In the present embodiment, in order to ensure that the trained drug screening model has generalization ability, the trained drug screening model is in a moderate fit and does not overfit. That is to say, the parameters of the entire drug screening model are adjusted by the predicted value, and then after adjustment, the adjusted drug screening model is verified by the test set to determine whether the predicted value output by each group of test drugs-test viruses in the test set is determined to be an effective predicted value. When the predicted value output by the drug screening model for each group of test drugs-test viruses in the test set indicates that the test drug is effective for the test virus, the drug screening model is the trained drug screening model. Of course, if the predicted value output by the drug screening model for each group of test drugs-test viruses in the test set does not reach the expected value (i.e., error probability), training continues.

[0148] Once the trained drug screening model has generalization capabilities, it means that the drug screening model meets the requirements. In this case, the trained drug screening model includes a long short-term memory network model, a graph convolutional neural network model, a feature fusion unit, and a multi-layer perceptron model. For each of the multiple drugs, the drug molecule and the base sequence of the virus to be tested are input into the trained drug screening model. Determining the effective value corresponding to the drug using the drug screening model specifically includes:

[0149] S21, obtaining the drug molecule of each drug in each group and the base digital sequence of the virus to be tested; wherein each drug in the same group corresponds to the same virus to be tested;

[0150] Wherein, obtaining the base digital sequence of the virus to be tested specifically includes:

[0151] S211, obtaining a base sequence of the virus to be tested;

[0152] S212, mapping the base letter sequence into a base number sequence, wherein the same letter is converted into the same number.

[0153] S22, inputting each of the drug molecules into the graph convolutional neural network model in sequence, and simultaneously inputting the base sequence of the virus to be tested into the long short-term memory network model;

[0154] S23, after being processed by the graph convolutional neural network model and the long short-term memory network model, they are jointly output to the feature fusion unit, after being processed by the feature fusion unit, the feature fusion unit outputs to the search multi-layer perceptron model, and after being processed by the multi-layer perceptron model, the effective value corresponding to each of the drugs is output in turn.

[0155] The effective value refers to an indication of whether the drug to be screened is effective against the virus to be tested.

[0156] In this way, the drug screening model is continuously trained through the training set to obtain a model with generalization ability. When a new virus appears, the trained drug screening model can be used to quickly screen out a set of candidate drugs that are effective against the new virus from a large number of drug libraries, thereby improving the drug screening speed and greatly shortening the drug screening time.

[0157] Experimental data:

[0158] Sample data: drug-virus pairs, 1200 positive samples (associated samples) and 5000 negative samples (unrelated samples)

[0159] The drug screening model was trained and evaluated multiple times. The multi-fold cross-validation method was used to perform multiple training cycles. The relationship between the number of training cycles and the test set error rate is as follows:

[0160] Number of training rounds Test set error rate 30 epoch 35.2％ 70 epoch 16.3％ 120 epoch 10.2％ 180 epoch 5.6％ 230 epoch 8.4％

[0161] When the number of training rounds is 180 epochs, the test set error rate is the lowest point of 5.6%, and the model has good generalization ability at this time.

[0162] S30. Determine a target drug corresponding to the virus to be tested according to the effective value, wherein the target drug is one or more of several drugs.

[0163] In this example, the effective value indicates whether the drug being screened is effective against the virus being tested. When a new virus emerges, a large database of candidate drugs is initially screened for potentially effective drugs based on the genetic sequence of the newly discovered virus. This screening process is performed automatically and rapidly on a computer, generating a small set of candidate drugs. This small set of candidate drugs is then subjected to secondary screening by professionals to identify drugs effective against the virus being tested. This significantly reduces the drug screening cycle, and thus the development cycle for drugs targeting viral infections.

[0164] In this way, based on steps S10-S30, the present invention uses artificial intelligence technology to directly mine the intrinsic correlation between pathogen gene sequences and drug molecular formulas; it quickly and automatically performs preliminary screening of drugs computationally, greatly reducing the size of the candidate drug database that requires professional screening, thereby significantly reducing the time for drug screening; using a larger-scale drug database retrieved by a computer for preliminary screening can more comprehensively cover various drugs, reduce the probability of target drugs being missed during the drug data collection stage, improve the success rate of screening and research and development of drugs for viral infectious diseases, improve screening efficiency, and greatly facilitate users.

[0165] Based on the above method, the present invention also provides a drug screening device, such as Figure 10 As shown, the device 100 includes a processor 11 and a memory 22 connected to the processor 11. Figure 10 Only some of the components of the device 100 are shown, but it should be understood that implementing all of the shown components is not a requirement, and more or fewer components may alternatively be implemented.

[0166] In some embodiments, the memory 22 can be an internal storage unit of the device 100, such as the internal memory of the device 100. In other embodiments, the memory 22 can also be an external storage device of the device 100, such as a plug-in USB flash drive, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the device 100. Further, the memory 22 can also include both an internal storage unit of the device 100 and an external storage device. The memory 22 is used to store application software and various types of data installed on the device 100, such as the drug screening program code, etc. The memory 22 can also be used to temporarily store data that has been output or is about to be output. In one embodiment, a drug screening program is stored on the memory 22, and the drug screening program can be executed by the processor 11, thereby realizing the drug screening method in the present application, as described in the above method.

[0167] In some embodiments, the processor 11 can be a central processing unit (CPU), a microprocessor, a mobile phone baseband processor or other data processing chip, used to run the program code or process data stored in the memory 22, such as executing the drug screening method, etc., as described in the above method.

[0168] Based on the above method, the present invention also provides a computer-readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps of the drug screening method described above.

[0169] Those skilled in the art will understand that Figure 3 The block diagram shown is merely a partial block diagram of the structure related to the present invention and does not limit the computer device to which the present invention is applied. A specific intelligent terminal may include more or fewer components than shown in the diagram, or may combine certain components or have a different component arrangement. When the processor executes the computer program, the steps of the drug screening method described above are implemented as described above.

[0170] In summary, the present invention discloses a drug screening method, device and storage medium, which includes obtaining several drugs corresponding to the virus to be tested; for each of the several drugs, the drug molecule of the drug and the base digital sequence of the virus to be tested are input into a trained drug screening model, and the effective value corresponding to the drug is determined by the drug screening model; according to the effective value, the target drug corresponding to the virus to be tested is determined, wherein the target drug is one or more of the several drugs. The present invention aims to perform deep learning based on the drug-virus pair set to screen the trained drug screening model, which can improve the computing performance of the drug screening model, greatly shorten the drug screening cycle, and thus shorten the research and development cycle of drugs for viral infectious diseases, thereby improving the efficiency of drug screening.

[0171] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (e.g., a processor, a controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0172] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A drug screening method, characterized in that: The drug screening method comprises the following steps: Obtain several drugs corresponding to the virus to be tested; For each of the plurality of drugs, the drug molecule of the drug and the base digital sequence of the virus to be tested are input into a trained drug screening model, and the effective value corresponding to the drug is determined by the drug screening model; The trained drug screening model includes a long short-term memory network model, a graph convolutional neural network model, a feature fusion unit, and a multi-layer perceptron model. For each of the plurality of drugs, the drug molecule of the drug and the base sequence of the virus to be tested are input into the trained drug screening model, and determining the effective value corresponding to the drug by the drug screening model specifically includes: Obtaining the drug molecule of each drug in each group and the base digital sequence of the virus to be tested; wherein each drug in the same group corresponds to the same virus to be tested; Inputting each of the drug molecules into the graph convolutional neural network model in sequence, and simultaneously inputting the base sequence of the virus to be tested into the long short-term memory network model; After being processed by the graph convolutional neural network model and the long short-term memory network model, the data are jointly output to the feature fusion unit. After being processed by the feature fusion unit, the feature fusion unit outputs the data to the multilayer perceptron model. After being processed by the multilayer perceptron model, the effective value corresponding to each drug is output in sequence. The target drug corresponding to the virus to be tested is determined according to the effective value, wherein the target drug is one or more of several drugs.

2. The drug screening method according to claim 1, characterized in that The drug screening model is obtained based on training of a training set, which includes multiple groups of training data pairs. Each group of training data pairs includes a drug molecule of a training drug and a base digital sequence of a training virus, and different groups of training data have the same training virus.

3. The drug screening method according to claim 2, characterized in that The training process of the drug screening model includes: For each of the plurality of training data pairs, the drug molecule of the training drug and the base sequence of the training virus are input into the drug screening model, and the drug screening model outputs a predicted value of the training drug-training virus; wherein the predicted value is used to indicate the inhibitory effect of the training drug on the training virus; The parameters of the drug screening model are modified according to the predicted value until the predicted value output by the drug screening model reaches the expected value, and then the training is stopped to obtain a trained drug screening model.

4. The drug screening method according to claim 3, characterized in that The drug screening model includes a long short-term memory network model, a graph convolutional neural network model, a feature fusion unit, and a multi-layer perceptron model. For each of the multiple training drug-training virus pairs, the drug molecule of the training drug and the base sequence of the training virus are input into the drug screening model. Outputting the predicted value of the training drug-training virus by the drug screening model specifically includes: Inputting the base digital sequence of the training virus into the long short-term memory network model, and outputting the viral gene sequence characteristics of the training virus through the long short-term memory network model; Inputting drug molecules of training drugs into a graph convolutional neural network model, and outputting drug molecular features of the training drugs through the graph convolutional neural network model; Inputting the drug molecular features and the virus gene sequence features into a feature fusion unit, and outputting the fusion features of the training drug-training virus pair through the feature fusion unit; The fusion features are input into a multi-layer perceptron model, and the multi-layer perceptron model outputs the predicted value of the training drug-training virus.

5. The drug screening method according to claim 4, characterized in that The long short-term memory network model includes a plurality of long short-term memory network units connected in sequence. Inputting the base sequence of the training virus into the long short-term memory network model and outputting the viral gene sequence characteristics of the training virus through the long short-term memory network model specifically include: The first number in the base number sequence is used as the target number and input into the first long short-term memory network unit, and the intermediate viral gene sequence feature is output by the first long short-term memory network unit, and the intermediate viral gene sequence feature is used as the target gene sequence feature; Inputting the next digit of the first digit as the target digit and the target gene sequence feature into the next long short-term memory network unit of the first long short-term memory network unit, and obtaining the next target gene feature through the next long short-term memory network unit; Repeat the steps until the last number of the base number sequence is used as the target number and the previous target gene sequence feature is input into the last long short-term memory network unit, and the viral gene sequence feature of the training virus is output through the last long short-term memory network unit.

6. The drug screening method according to claim 5, characterized in that The interval between each input of the target number into the LSTM network unit is the same.

7. The drug screening method according to claim 4, characterized in that The graph convolutional neural network model includes an input layer, a multi-cascade hidden layer, and an output layer. Inputting the drug molecules of the training drug into the graph convolutional neural network model and outputting the drug molecular features of the training drug through the graph convolutional neural network model specifically include: Inputting drug molecules of training drugs into the input layer, and obtaining a drug molecule graph through the input layer; The drug molecule graph is input into the first hidden layer, and the output of the first hidden layer is processed by an activation function to obtain a first drug molecule characteristic graph; Inputting the first drug molecular characteristic graph as the target drug molecular graph into the next hidden layer of the first hidden layer, and processing the output of the next hidden layer with the same activation function to obtain the next drug molecular characteristic graph, until the previous drug molecular characteristic graph is input as the target drug molecular graph into the last hidden layer, and the last drug molecular characteristic graph is output through the last hidden layer; The last drug molecular feature map is input to the output layer, and the drug molecular features of the training drug are output through the output layer.

8. The drug screening method according to claim 4, characterized in that The multilayer perceptron model includes an input layer, a multi-cascade hidden layer, and an output layer. Inputting the fusion feature into the multilayer perceptron model and outputting the predicted value of the training drug-training virus through the multilayer perceptron model specifically includes: Inputting the fused features into the input layer, and obtaining the initial fused features through the input layer; Inputting the initial fusion feature into a first-level hidden layer, and outputting a first target feature through the first-level hidden layer; Inputting the first target feature as a target into the next hidden layer of the first hidden layer, outputting the next target feature through the next hidden layer, until inputting the previous target feature as a target into the last hidden layer, and outputting the last target feature through the last hidden layer; The last target feature is input into the output layer, and the predicted value of the training drug-training virus is output through the output layer.

9. The drug screening method according to claim 1, wherein The step of obtaining the base sequence of the virus to be detected specifically includes: Obtaining the base letter sequence of the virus to be tested; The base letter sequence is mapped into a base number sequence, wherein identical letters are converted to the same number.

10. A drug screening device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the drug screening method according to any one of claims 1 to 9 when executing the computer program.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the drug screening method according to any one of claims 1 to 9.

Images