Drug sensitivity prediction and model training method, storage medium and device

Abstract

A drug sensitivity prediction and model training method, a storage medium and an equipment, the drug sensitivity prediction method comprises: obtaining gene expression information of a to-be-tested cell line, gene mutation information of the to-be-tested cell line, and structure information of a to-be-tested drug; based on a first attention model, calculating first related information between the structure information of the to-be-tested drug and the gene expression information; based on a second attention model, calculating second related information between the structure information of the to-be-tested drug and the gene mutation information; splicing the first related information and the second related information to obtain a splicing result; based on a drug sensitivity prediction model, performing prediction processing on the splicing result to obtain sensitivity information of the to-be-tested cell line to the to-be-tested drug.

Description

Technical Field

[0001] This disclosure relates to, but is not limited to, the field of big data processing technology, and in particular to a method, storage medium, and device for drug sensitivity prediction and model training. Background Technology

[0002] For cancer treatment drugs, low drug sensitivity is a significant reason for treatment failure, and decreased drug sensitivity is also a contributing factor to tumor recurrence. Currently, companion diagnostic products for anti-tumor drugs focus on targeted therapies, and their main mechanism involves detecting the patient's gene mutation type and recommending drugs based on the mutation results. Therefore, drug sensitivity analysis based on large-scale pharmacogenomics data is a current research direction, and the Genomics of Drug Sensitivity in Cancer (GDSC) database and the Broad Institute Cancer Cell Line Encyclopedia (CCLE) database, which contain information on tumor cell line mutations, expression, copy number variations, methylation, and drug dose-response data, have become among the most important tools. With the rise of deep learning methods and the further confirmation of their ability to learn the richest information from raw data, using deep learning to study drug sensitivity is imperative. Summary of the Invention

[0003] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0004] In a first aspect, embodiments of this disclosure provide a method for predicting drug sensitivity, including:

[0005] Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell lines to be tested.

[0006] Based on a first attention model, a first relevant correlation is calculated between the structural information of the drug under test and the gene expression information; based on a second attention model, a second relevant correlation is calculated between the structural information of the drug under test and the gene mutation information.

[0007] The first relevant information and the second relevant information are concatenated to obtain the concatenation result;

[0008] Based on the drug sensitivity prediction model, the splicing results are processed to obtain the sensitivity information of the cell line to the drug.

[0009] Secondly, embodiments of this disclosure also provide a method for training a drug sensitivity prediction model, comprising:

[0010] A training sample set is obtained, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibition concentrations; wherein each baseline half-inhibition concentration corresponds to the structural information of a drug.

[0011] Based on the first attention model, multiple first prediction related information is obtained between the structural information of the multiple drugs and the gene expression information; based on the second attention model, multiple second prediction related information is obtained between the structural information of the multiple drugs and the gene mutation information.

[0012] By splicing together the first and second prediction-related information corresponding to the structural information of the same drug, multiple spliced ​​prediction results are obtained.

[0013] The drug sensitivity prediction model is obtained by training the prediction model to be trained using the multiple spliced ​​prediction results and the multiple baseline half-inhibition concentration information.

[0014] Thirdly, this disclosure also provides a non-transient computer-readable storage medium, which is configured to store computer program instructions, wherein the computer program instructions, when executed, can implement the drug sensitivity prediction method described in any of the above embodiments, or the computer program instructions, when executed, can implement the drug sensitivity prediction model training method described in any of the above embodiments.

[0015] Fourthly, embodiments of this disclosure also provide a drug sensitivity prediction device, including a first memory, a first processor, and a computer program stored in the first memory and executable on the first processor to perform:

[0016] Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell lines to be tested.

[0017] Based on a first attention model, a first relevant correlation is calculated between the structural information of the drug under test and the gene expression information; based on a second attention model, a second relevant correlation is calculated between the structural information of the drug under test and the gene mutation information.

[0018] The first relevant information and the second relevant information are concatenated to obtain the concatenation result;

[0019] Based on the drug sensitivity prediction model, the splicing results are processed to obtain the sensitivity information of the cell line to the drug.

[0020] Fifthly, embodiments of this disclosure also provide a drug sensitivity prediction model training device, including a second memory, a second processor, and a computer program stored in the second memory and executable on the second processor to perform:

[0021] A training sample set is obtained, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibition concentrations; wherein each baseline half-inhibition concentration corresponds to the structural information of a drug.

[0022] Based on the first attention model, multiple first prediction related information is obtained between the structural information of the multiple drugs and the gene expression information; based on the second attention model, multiple second prediction related information is obtained between the structural information of the multiple drugs and the gene mutation information.

[0023] By splicing together the first and second prediction-related information corresponding to the structural information of the same drug, multiple spliced ​​prediction results are obtained.

[0024] The drug sensitivity prediction model is obtained by training the prediction model to be trained using the multiple spliced ​​prediction results and the multiple baseline half-inhibition concentration information.

[0025] Sixthly, embodiments of this disclosure also provide a drug sensitivity prediction device, comprising:

[0026] The acquisition module is configured to acquire gene expression information of the cell line to be tested, gene mutation information of the cell line to be tested, and structural information of the drug to be tested.

[0027] The first feature fusion module is configured to calculate a first relevant correlation between the structural information of the drug to be tested and the gene expression information based on a first attention model.

[0028] The second feature fusion module is configured to calculate a second relevant information between the structural information of the drug under test and the gene mutation information based on a second attention model.

[0029] The splicing module is configured to splice the first relevant information and the second relevant information to obtain the splicing result;

[0030] The drug sensitivity prediction module is configured to perform prediction processing on the splicing results based on a drug sensitivity prediction model to obtain the sensitivity information of the cell line to the drug to be tested.

[0031] Seventhly, embodiments of this disclosure also provide a drug sensitivity prediction model training device, comprising:

[0032] The sample acquisition module is configured to acquire a training sample set, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibition concentrations; wherein each baseline half-inhibition concentration corresponds to the structural information of a drug.

[0033] The first feature fusion prediction module is configured to obtain multiple first prediction related information between the structural information of the multiple drugs and the gene expression information based on the first attention model;

[0034] The second feature fusion prediction module is configured to obtain multiple second prediction related information between the structural information of the multiple drugs and the gene mutation information based on the second attention model;

[0035] The splicing prediction module is configured to splice the first and second prediction related information corresponding to the structural information of the same drug to obtain multiple spliced ​​prediction results.

[0036] The sensitivity prediction module is configured to train the prediction model to be trained using the multiple spliced ​​prediction results and the multiple baseline half-inhibition concentration information to obtain the drug sensitivity prediction model.

[0037] After reading and understanding the accompanying diagrams and detailed descriptions, the other aspects can be understood. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the technical solutions of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the technical solutions of this disclosure and do not constitute a limitation on the technical solutions of this disclosure. The shape and size of each component in the drawings do not reflect actual proportions and are only intended to illustrate the content of this disclosure.

[0039] Figure 1 The diagram shown is a flowchart of a drug sensitivity prediction method provided in an embodiment of this disclosure;

[0040] Figure 2 The diagram shown is a flowchart of a drug sensitivity prediction method provided by an exemplary embodiment of this disclosure;

[0041] Figure 3 The diagram shown is a schematic representation of the coding layer structure of an encoder provided in an exemplary embodiment of this disclosure.

[0042] Figure 4 The diagram shown is a flowchart of a drug sensitivity prediction model training method provided in an embodiment of this disclosure;

[0043] Figure 5 The diagram shown is a structural schematic of an encoder provided by an exemplary embodiment of this disclosure;

[0044] Figure 6 The diagram shown is a schematic representation of a drug sensitivity prediction device provided in an embodiment of this disclosure.

[0045] Figure 7 The diagram shown is a schematic diagram of a drug sensitivity prediction model training device provided in an embodiment of this disclosure;

[0046] Figure 8 The diagram shown is a schematic representation of a drug sensitivity prediction device provided in an embodiment of this disclosure.

[0047] Figure 9 The diagram shown is a schematic diagram of a drug sensitivity prediction model training device provided in an embodiment of this disclosure;

[0048] Figure 10 The diagram shown is a logical structure diagram of a drug sensitivity prediction provided by an exemplary embodiment of this disclosure. Detailed Implementation

[0049] The embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The implementation can be carried out in many different forms. Those skilled in the art will readily understand that the methods and content can be varied in various forms without departing from the spirit and scope of this disclosure. Therefore, this disclosure should not be construed as limited to the content described in the following embodiments. Without conflict, the embodiments and features in the embodiments of this disclosure can be arbitrarily combined with each other. To keep the following description of the embodiments of this disclosure clear and concise, detailed descriptions of some known functions and components have been omitted. The accompanying drawings of the embodiments of this disclosure only relate to the structures involved in the embodiments of this disclosure; other structures can be referred to in general design.

[0050] The ordinal numbers “first,” “second,” and “third” used in this specification are used to avoid confusion among the constituent elements, not to limit their quantity.

[0051] In this specification, "electrical connection" includes the situation where components are connected together by elements having a certain electrical function. There are no particular limitations on the term "elements having a certain electrical function," as long as they enable the transmission and reception of electrical signals between the connected components. Examples of "elements having a certain electrical function" include not only electrodes and wiring, but also switching elements such as transistors, resistors, inductors, capacitors, and other elements having one or more functions.

[0052] The data used for deep learning-based antitumor drug sensitivity prediction includes messenger ribonucleic acid (mRNA) expression information, mutation information, drug chemical structure information, and copy number variation information. Specifically, mRNA expression information is extracted and compressed using autoencoders, then concatenated with other information (which may involve simple network processing) before being input into a fully connected network for prediction. This prediction method is prone to feature sparsity and loss of information between data points. Furthermore, feature extraction and compression using autoencoders only utilizes a portion of the mRNA expression information, resulting in poor drug sensitivity prediction performance.

[0053] This disclosure provides a method for predicting drug sensitivity, such as... Figure 1 As shown in the embodiments of this disclosure, a drug sensitivity prediction method may include:

[0054] Step A1: Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell line to be tested;

[0055] Step A2: Based on the first attention model, calculate the first relevant information between the structural information and gene expression information of the drug under test; based on the second attention model, calculate the second relevant information between the structural information and gene mutation information of the drug under test.

[0056] Step A3: Combine the first relevant information and the second relevant information to obtain the combined result;

[0057] Step A4: Based on the drug sensitivity prediction model, perform predictive processing on the splicing results to obtain the sensitivity information of the cell line to the drug.

[0058] The drug sensitivity prediction method provided in this disclosure obtains a first correlation between the structural information and gene expression information of the drug to be tested using a first attention model, and a second correlation between the structural information and gene mutation information of the drug to be tested using a second attention model. The first and second correlations are then concatenated to obtain a concatenated result. This concatenated result is then processed by a drug sensitivity prediction model to obtain the sensitivity information of the cell line to the drug to be tested. Before the drug sensitivity prediction model performs its prediction, obtaining the correlation between gene expression information, gene mutation information, and the structural information of the drug through an attention mechanism, and predicting drug sensitivity based on this correlation, can improve the prediction effect of the drug sensitivity prediction model and overcome the shortcomings of poor drug sensitivity prediction performance.

[0059] In an exemplary embodiment, step A2, calculating the first relevant information between the structural information and gene expression information of the drug to be tested based on the first attention model, may include:

[0060] Step A201: Multiply the gene expression information by the first weight matrix to obtain the first vector, multiply the drug structure information by the second weight matrix to obtain the second vector, and multiply the drug structure information by the third weight matrix to obtain the third vector;

[0061] Step A202: Normalize the first vector and the second vector to obtain the first processing result, and multiply the first processing result with the third vector to obtain the first relevant information.

[0062] In an exemplary embodiment, step A202 involves normalizing the first vector and the second vector to obtain a first processing result, including: transposing the second vector to obtain a transpose vector of the second vector, multiplying the first vector by the transpose vector of the second vector to obtain a first product, and dividing the first product by a first constant to obtain the first processing result; the first constant is the arithmetic square root of the dimension of the second vector.

[0063] For example, in step A201, Q is set to gene expression information, K and V are set to drug structural information, and the first weight matrix is ​​set to W. Q The second weight matrix is ​​W K The third weight matrix is ​​W V Gene expression information Q is compared with the first weight matrix W. Q Multiplying them together yields the first vector q = Q * W. Q The structural information K of the drug is combined with the second weight matrix W. K Multiplying them together yields the second vector k = K * W K The structural information V of the drug is combined with the third weight matrix W V Multiplying them together yields the third vector v = V * W V Here, the first vector q can be understood as the query vector of the first self-attention model, the second vector k can be understood as the key vector of the first self-attention model, and the third vector v can be understood as the value vector of the first self-attention model.

[0064] In step A202, the first vector and the second vector are normalized to obtain the first processing result. The first processing result is then multiplied by the third vector to obtain the first relevant information. The calculation formula is as follows: Where, k T Let d be the transpose of the second vector (key vector). k Let be the dimension of the key vector, and softmax be the normalization function.

[0065] In an exemplary embodiment, after obtaining the gene expression information of the cell line to be tested and the structural information of the drug to be tested in step A1, the method further includes: performing dimensionality reduction operation on the gene expression information through a first convolutional neural network to obtain dimensionality-reduced gene expression information; and performing dimensionality reduction operation on the drug structural information through a second convolutional neural network to obtain dimensionality-reduced drug structural information.

[0066] In step A201, multiplying the gene expression information with the first weight matrix to obtain the first vector includes: multiplying the dimension-reduced gene expression information with the first weight matrix to obtain the first vector;

[0067] In step A201, the structural information of the drug is multiplied by the second weight matrix to obtain the second vector, including: multiplying the structural information of the drug after dimensionality reduction by the second weight matrix to obtain the second vector;

[0068] In step A201, the structural information of the drug is multiplied by the third weight matrix to obtain the third vector, including: multiplying the structural information of the drug after dimensionality reduction by the third weight matrix to obtain the third vector.

[0069] In an exemplary embodiment, the gene expression information of the cell line to be tested obtained in step A1 has a dimension of 1*500, the structure information of the drug has a dimension of 72*188, the dimension of the reduced gene expression information has a dimension of 1*188, the dimension of the reduced drug structure information has a dimension of 1*188, and the dimensions of the first vector q, the second vector k, and the third vector v obtained in step A201 are all 1*188. In step A202, the formula is used... The dimension of the first relevant information obtained is 1*188.

[0070] In an exemplary embodiment, step A2, calculating the second relevant information between the structural information and gene mutation information of the drug under test based on the second attention model, may include:

[0071] Step A211: Multiply the gene mutation information by the fourth weight matrix to obtain the fourth vector, multiply the drug structure information by the fifth weight matrix to obtain the fifth vector, and multiply the drug structure information by the sixth weight matrix to obtain the sixth vector.

[0072] Step A212: Normalize the fourth and fifth vectors to obtain the second processing result, and multiply the second processing result with the sixth vector to obtain the second relevant information.

[0073] In an exemplary embodiment, step A212, normalizing the fourth and fifth vectors to obtain a second processing result, may include:

[0074] Transpose the fifth vector to obtain the transpose of the fifth vector. Multiply the fourth vector by the transpose of the fifth vector to obtain the second product. Divide the second product by the second constant to obtain the second processing result; the second constant is the arithmetic square root of the dimension of the fifth vector.

[0075] For example, in step A211, Q1 is set to gene mutation information, K1 and V1 are set to drug structural information, and the fourth weight matrix is ​​set to W. Q1 The fifth weight matrix is ​​W. K1 The sixth weight matrix is ​​W. V1 The gene mutation information Q1 is combined with the fourth weight matrix W. Q1 Multiplying them together yields the fourth vector q1 = Q1 * W. Q1 The structural information K1 of the drug is combined with the fifth weight matrix W K1 Multiplying them together yields the fifth vector k1 = K1 * W K The structural information V1 of the drug is combined with the sixth weight matrix W. V1 Multiplying them together yields the sixth vector v1 = V1 * W. V1 In this context, the fourth vector q1 can be understood as the query vector of the second self-attention model, the fifth vector k1 can be understood as the key vector of the second self-attention model, and the sixth vector v1 can be understood as the value vector of the second self-attention model.

[0076] In step A222, the fourth and fifth vectors are normalized to obtain the second processing result. The second processing result is then multiplied by the sixth vector to obtain the second relevant information. The calculation formula is as follows: Where, k T1 Let d be the transpose of the fifth vector (key vector). k Let be the dimension of the key vector, and softmax be the normalization function.

[0077] In an exemplary embodiment, after obtaining the gene mutation information of the cell line to be tested and the structural information of the drug to be tested in step A1, the method further includes: performing dimensionality reduction operation on the gene mutation information through a third convolutional neural network to obtain dimensionality-reduced gene mutation information; and performing dimensionality reduction operation on the drug structural information through a second convolutional neural network to obtain dimensionality-reduced drug structural information.

[0078] In step A211, the gene mutation information is multiplied by the fourth weight matrix to obtain the fourth vector, including: multiplying the dimension-reduced gene mutation information by the fourth weight matrix to obtain the fourth vector;

[0079] In step A211, the structural information of the drug is multiplied by the fifth weight matrix to obtain the fifth vector, including: multiplying the structural information of the drug after dimensionality reduction by the fifth weight matrix to obtain the fifth vector;

[0080] In step A211, the structural information of the drug is multiplied by the sixth weight matrix to obtain the sixth vector, including: multiplying the structural information of the drug after dimensionality reduction by the sixth weight matrix to obtain the sixth vector.

[0081] In an exemplary embodiment, the gene mutation information of the cell line to be tested obtained in step A1 has a dimension of 1*310, and the structural information of the drug has a dimension of 72*188. The dimension-reduced gene mutation information and drug structural information both have a dimension of 1*188. The fourth vector q1, fifth vector k1, and sixth vector v1 obtained in step A211 all have a dimension of 1*188. In step A212, the formula... The dimension of the second relevant information obtained is 1*188.

[0082] In an exemplary embodiment, step A1, obtaining the gene expression information of the cell line to be tested, may include steps A11-A14:

[0083] Step A11: Obtain the raw data of gene expression information, which includes the mean and standard deviation of multiple first gene expression features;

[0084] Step A12: Standardize the mean values ​​of multiple first gene expression features to obtain multiple standardized expression means, and standardize the standard deviations of multiple first gene expression features to obtain multiple standardized expression standard deviations. Input the standardized expression standard deviations and multiple standardized expression means into the encoder. Step A13: Control the encoder to add or subtract the standardized expression standard deviations corresponding to a portion of the standardized expression means to obtain multiple processed standardized expression means. Use the other portion of unprocessed standardized expression means and the multiple processed standardized expression means as multiple encoding input features.

[0085] In an exemplary embodiment, step A13 can be understood as adding or subtracting the standardized expression standard deviation corresponding to the standardized expression mean of multiple standardized expression means with a certain probability.

[0086] In an exemplary embodiment, step A13, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from a portion of the standardized expression mean, may include: controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from all of the standardized expression means. In another exemplary embodiment, step A13, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from a portion of the standardized expression mean, may include: controlling the encoder to add the standardized expression standard deviation corresponding to the standardized expression mean to a portion of the standardized expression mean, and controlling the encoder to subtract the standardized expression standard deviation corresponding to the standardized expression mean from another portion of the standardized expression mean.

[0087] Step A14: Control the encoder to encode multiple input features to obtain multiple second expression features as gene expression information. The number of multiple second gene expression features is less than the number of multiple first gene expression features.

[0088] In an exemplary embodiment, the encoder may include an encoding layer, which may include an input layer and an output layer;

[0089] In step A14, controlling the encoder to encode multiple coded input features may include:

[0090] The encoder performs the following operations on multiple encoded input features to obtain multiple second gene expression features: in, y represents the input feature, y represents the expression feature of the second gene, W represents the link weights from the input layer to the output layer, b represents the bias of the output layer, and s represents the nonlinear function.

[0091] In an exemplary embodiment, the coding layer may further include an intermediate hidden layer located between the input layer and the output layer, and the input layer, the intermediate hidden layer and the output layer constitute a three-layer neural network with the number of neurons gradually decreasing.

[0092] In an exemplary embodiment, the number of neurons in the input layer is 1500 to 2500, the number of neurons in the intermediate hidden layers is 500 to 1500, and the number of neurons in the output layer is 250 to 750. For example, the number of neurons in the input layer is 2000, the number of neurons in the intermediate hidden layers is 1000, and the number of neurons in the output layer is 500.

[0093] In an exemplary implementation, the sensitivity prediction model comprises a four-layer neural network with a progressively decreasing number of neurons.

[0094] In an exemplary embodiment, in a four-layer neural network, the first layer has 400 to 600 neurons, the second layer has 100 to 300 neurons, the third layer has 80 to 120 neurons, and the fourth layer has 1 to 5 neurons. For example, in a four-layer neural network, the first layer has 500 neurons, the second layer has 200 neurons, the third layer has 100 neurons, and the fourth layer has 1 neuron.

[0095] The following describes in detail the methods for predicting drug sensitivity, such as... Figure 2 As shown, it may include the following steps:

[0096] Step 101: Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell line to be tested.

[0097] In an exemplary embodiment, obtaining the gene expression information of the cell line to be tested in step 101 may include steps B12-B14:

[0098] Step B11: Obtain the raw data of gene expression information, which includes the mean and standard deviation of multiple first gene expression features;

[0099] Step B12: Standardize the mean of multiple first gene expression features to obtain a standardized expression mean, standardize the standard deviation of multiple first gene expression features to obtain a standardized expression standard deviation, and input the multiple standardized expression standard deviations and multiple standardized expression means into the encoder;

[0100] In this embodiment of the disclosure, the convergence of the model can be increased by standardizing the mean and standard deviation of the expression features of multiple first genes.

[0101] In an exemplary embodiment, the formula for standardizing the mean of any one of the mean values ​​of multiple first gene expression features is X. norm =(XX) min ) / (X max -X min ), X norm X represents the standardized expression mean, where X is the mean of the first gene expression characteristic. min X is the minimum value among the means of multiple first gene expression characteristics. max It is the largest value among the means of multiple first gene expression characteristics.

[0102] In an exemplary embodiment, standardizing the standard deviation of any one of the standard deviations of multiple first gene expression features may include performing the following calculation on the standard deviation of any one of the standard deviations of multiple first gene expression features: σ norm =(σ / X)*X norm , σ norm The standardized expression standard deviation is σ, where σ is the standard deviation of the first gene expression characteristic, and X is the mean of the first gene expression characteristic. norm This is the standardized expression mean. The standard deviation of a first gene expression characteristic can be obtained by performing the following calculation on multiple first gene expression characteristics and their means: σ is the standard deviation of the first gene expression trait, N is the number of first gene expression trait values, and X is the standard deviation of the first gene expression trait. i Let represent the expression feature of the i-th first gene, and U be the mean of the expression features of multiple first genes.

[0103] Step B13: Control the encoder to add or subtract the standardized expression standard deviation corresponding to a portion of the standardized expression mean to obtain the standardized expression mean of multiple processing, and use the other portion of the unprocessed standardized expression mean and the standardized expression mean of multiple processing as multiple encoding input features;

[0104] For example, encoding input features b represents the deviation of the output layer of the encoder's coding layer.

[0105] In an exemplary embodiment, step B13 can be understood as adding or subtracting the standardized expression standard deviation corresponding to the standardized expression mean from the multiple standardized expression means with a certain probability.

[0106] In an exemplary embodiment, step B13, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from a portion of the standardized expression mean, may include: controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from all of the standardized expression means. In another exemplary embodiment, step B13, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from a portion of the standardized expression mean, may include: controlling the encoder to add the standardized expression standard deviation corresponding to the standardized expression mean to a portion of the standardized expression mean, and controlling the encoder to subtract the standardized expression standard deviation corresponding to the standardized expression mean from another portion of the standardized expression mean.

[0107] Step B14: Control the encoder to encode multiple input features to obtain multiple second gene expression features as gene expression information. The number of multiple second gene expression features is less than the number of multiple first gene expression features.

[0108] In an exemplary implementation, such as Figure 3 As shown, the encoder may include an encoding layer, which may include an input layer and an output layer; controlling the encoder to encode multiple encoded input features may include: controlling the encoder to perform the following operations on the multiple encoded input features to obtain multiple second gene expression features: Let y be the encoding input feature, y be the second gene expression feature, W be the link weights from the input layer to the output layer, b be the output layer bias, and s be a nonlinear function. In an exemplary embodiment, the nonlinear function s can be a sigmoid function.

[0109] In an exemplary implementation, such as Figure 3 As shown, the encoding layer may further include an intermediate hidden layer located between the input layer and the output layer. The input layer, intermediate hidden layer, and output layer constitute a three-layer neural network with the number of neurons decreasing progressively. In an exemplary embodiment, the number of neurons in the input layer is 1500 to 2500, the number of neurons in the intermediate hidden layer is 500 to 1500, and the number of neurons in the output layer is 250 to 750. For example, the number of neurons in the input layer is 1000, the number of neurons in the intermediate hidden layer is 1000, and the number of neurons in the output layer is 500. In this embodiment, the number of neural network layers in the encoder and the number of neurons in each neural network layer can be set according to actual needs during modeling. For example, the number of neurons in the input layer can be set to be greater than or less than 1000 depending on the number of encoded input feature values, the intermediate hidden layer can be set to one layer or more, and the number of neurons in the output layer can be set to less than or greater than 500 according to actual needs. In this embodiment, the encoder can be an autoencoder, which is a neural network model. In this embodiment of the disclosure, since the acquired gene expression information contains many features, encoding multiple first gene expression features using an encoder can reduce dimensionality, resulting in fewer second gene expression features output by the output layer compared to the number of first gene expression features. This saves storage space, improves computation speed, and removes redundant features, such as... Figure 3 As shown, in the encoder's coding layer, the number of neurons in the output layer is less than the number of neurons in the input layer.

[0110] In this embodiment of the disclosure, since the dimension of gene expression information is much larger than that of gene mutation information and drug structure information, and the E-MTAB-3610 dataset includes expression standard deviation data in addition to the expression mean (the individual gene expression level is affected by the time and space characteristics of expression and the interference, and is necessarily a dynamic level, so the standard deviation can better reflect the actual biological significance), the method of integrating expression standard deviation on the basis of Autoencoders in this embodiment of the disclosure can reflect the actual biological characteristics.

[0111] In an exemplary embodiment, the gene expression information and gene mutation information of the cell line to be tested can be obtained through gene detection, and the structural information of the drug to be tested can be obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) database.

[0112] The structural information of the drugs under test is represented by SMILES structures, which use letters, numbers, and special characters to represent a molecule. For example, "C" represents a carbon atom, "=" represents a covalent bond between two atoms, carbon dioxide can be represented as O=C=O, and aspirin can be represented as O=C(C)OC1CCCCC1C(=O)O. The longest SMILES expression for a drug is 188 characters. From the SMILES structural information of 223 antitumor drugs analyzed, there are 72 different characters. Consideration is given to processing them using a one-hot encoding method, that is, the structural information of each drug can be converted into a 72*188 one-hot matrix. For each drug, a value of 1 in the i-th row and j-th column means that the i-th symbol appears in the j-th position in the SMILES format, as shown in Table 1.

[0113] Table 1

[0114]

[0115] In Table 1, the first character of the SMILES structure for drugs is O, the second character is =, and the third character is C.

[0116] For example, carbon dioxide O=C=O is encoded using one-hot encoding as shown in Table 2:

[0117] Table 2

[0118]

[0119] Step 102: Perform dimensionality reduction on the gene expression information through the first convolutional neural network to obtain the dimensionality-reduced gene expression information. Perform dimensionality reduction on the drug structure information through the second convolutional neural network to obtain the dimensionality-reduced drug structure information.

[0120] In an exemplary embodiment, the gene expression information dimension of the cell line to be tested obtained in step 101 is 1*500, the structure information dimension of the drug is 72*188, the gene expression information dimension after dimensionality reduction using a convolutional neural network is 1*188, and the structure information dimension of the drug after dimensionality reduction is 1*188.

[0121] Step 103: Based on the first attention model, calculate the first correlation between the structural information of the drug under test after dimensionality reduction and the gene expression information after dimensionality reduction.

[0122] In an exemplary embodiment, step 103 may include:

[0123] Step A01: Multiply the dimension-reduced gene expression information with the first weight matrix to obtain the first vector; multiply the dimension-reduced drug structure information with the second weight matrix to obtain the second vector; multiply the dimension-reduced drug structure information with the third weight matrix to obtain the third vector.

[0124] Step A02: Normalize the first vector and the second vector to obtain the first processing result, and multiply the first processing result with the third vector to obtain the first relevant information.

[0125] In an exemplary embodiment, step A02 involves normalizing the first vector and the second vector, including: transposing the second vector to obtain the transpose of the second vector, multiplying the first vector by the transpose of the second vector to obtain a first product, and dividing the first product by a first constant to obtain a first processing result; the first constant is the arithmetic square root of the dimension of the second vector.

[0126] For example, in step A01, Q is set to gene expression information, K and V are set to drug structural information, and the first weight matrix is ​​set to W. Q The second weight matrix is ​​W K The third weight matrix is ​​W V The dimensionality-reduced gene expression information Q is compared with the first weight matrix W. Q Multiplying them together yields the first vector q = Q * W. Q The structural information K of the drug after dimensionality reduction is combined with the second weight matrix W. K Multiplying them together yields the second vector k = K * W K The dimensionality-reduced structural information V of the drug is compared with the third weight matrix W. V Multiplying them together yields the third vector v = V * W VHere, the first vector q can be understood as the query vector of the first self-attention model, the second vector k can be understood as the key vector of the first self-attention model, and the third vector v can be understood as the value vector of the first self-attention model.

[0127] In step A02, the first vector and the second vector are normalized to obtain the first processing result. The first processing result is then multiplied by the third vector to obtain the first relevant information. The calculation formula is as follows: Where, k T Let d be the transpose of the second vector (key vector). k Let be the dimension of the key vector, and softmax be the normalization function.

[0128] In an exemplary embodiment, the first vector q, the second vector k, and the third vector v obtained in step A01 all have dimensions of 1*188. In step A02, the vectors are calculated using the formula... The dimension of the first relevant information obtained is 1*188.

[0129] Step 104: Based on the second attention model, calculate the second relevant information between the structural information and gene mutation information of the drug under test.

[0130] In an exemplary embodiment, step 104 may include:

[0131] Step A21: Perform dimensionality reduction on the gene mutation information through the third convolutional neural network to obtain the dimensionality-reduced gene mutation information. Multiply the dimensionality-reduced gene mutation information with the fourth weight matrix to obtain the fourth vector. Multiply the dimensionality-reduced drug structure information with the fifth weight matrix to obtain the fifth vector. Multiply the dimensionality-reduced drug structure information with the sixth weight matrix to obtain the sixth vector.

[0132] Step A22: Normalize the fourth and fifth vectors to obtain the second processing result, and multiply the second processing result with the sixth vector to obtain the second relevant information.

[0133] In an exemplary embodiment, step A22, normalizing the fourth and fifth vectors, may include: transposing the fifth vector to obtain a transpose of the fifth vector, multiplying the fourth vector by the transpose of the fifth vector to obtain a second product, and dividing the second product by a second constant to obtain a second processing result; the second constant is the arithmetic square root of the dimension of the fifth vector.

[0134] For example, in step A21, Q1 is set to gene mutation information, K1 and V1 are set to the structural information of the drug after dimensionality reduction, and the fourth weight matrix is ​​set to W. Q1 The fifth weight matrix is ​​W. K1 The sixth weight matrix is ​​W.V1 The dimensionality-reduced gene mutation information Q1 is combined with the fourth weight matrix W. Q1 Multiplying them together yields the fourth vector q1 = Q1 * W. Q1 The structural information K1 of the drug after dimensionality reduction is combined with the fifth weight matrix W. K1 Multiplying them together yields the fifth vector k1 = K1 * W K The structural information V1 of the drug after dimensionality reduction is compared with the sixth weight matrix W. V1 Multiplying them together yields the sixth vector v1 = V1 * W. V1 In this context, the fourth vector q1 can be understood as the query vector of the second self-attention model, the fifth vector k1 can be understood as the key vector of the second self-attention model, and the sixth vector v1 can be understood as the value vector of the second self-attention model.

[0135] In step A22, the fourth and fifth vectors are normalized to obtain the second processing result. The second processing result is then multiplied by the sixth vector to obtain the second relevant information. The calculation formula can be: Where, k T1 Let d be the transpose of the fifth vector (key vector). k Let be the dimension of the key vector, and softmax be the normalization function.

[0136] In an exemplary embodiment, the gene mutation information dimension of the cell line to be tested obtained in step 101 is 1*310, the structure information dimension of the drug is 72*188, the dimension of the reduced gene mutation information is 1*188, the dimension of the reduced drug structure information is 1*188, and the dimensions of the fourth vector q1, fifth vector k1, and sixth vector v1 obtained in step A21 are all 1*188. In step A22, the calculation formula is used... The dimension of the second relevant information obtained is 1*188.

[0137] Step 105: Combine the first relevant information and the second relevant information to obtain the combined result.

[0138] In an exemplary implementation, concatenating the first relevant information of 1*188 dimensions and the second relevant information of 1*188 dimensions yields a concatenated result of 1*376 dimensions.

[0139] Step 106: Based on the drug sensitivity prediction model, perform predictive processing on the splicing results to obtain the sensitivity information of the cell line to the drug.

[0140] In an exemplary implementation, the sensitivity prediction model comprises a four-layer neural network with a progressively decreasing number of neurons.

[0141] In an exemplary embodiment, in a four-layer neural network, the first layer has 400 to 600 neurons, the second layer has 100 to 300 neurons, the third layer has 80 to 120 neurons, and the fourth layer has 1 to 5 neurons. For example, in a four-layer neural network, the first layer has 500 neurons, the second layer has 200 neurons, the third layer has 100 neurons, and the fourth layer has 1 neuron.

[0142] The drug sensitivity prediction method provided in this disclosure can predict the sensitivity information of a test cell line to a test drug, which can be a drug for treating tumors or other diseases. In this disclosure, the drug sensitivity information can be IC50 (half maximal inhibitory concentration) or log10 (IC50). In antitumor drug-cell line dose data, GDSC uses IC50 to evaluate the therapeutic effect of antitumor drugs. Since the IC50 values ​​of different cell lines and different drugs vary greatly, log10 (IC50) can be used as the drug sensitivity information. IC50 is the concentration at which a cell line achieves 50% inhibition of growth 72 hours after drug treatment.

[0143] This disclosure also provides a method for training a drug sensitivity prediction model, such as... Figure 4 As shown, the drug sensitivity prediction model training method provided in this disclosure embodiment may include steps C1-C4:

[0144] Step C1: Obtain the training sample set, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibitory concentrations; wherein each baseline half-inhibitory concentration corresponds to the structural information of a drug.

[0145] Step C2: Based on the first attention model, obtain multiple first prediction related information between the structural information of multiple drugs and gene expression information; based on the second attention model, obtain multiple second prediction related information between the structural information of multiple drugs and gene mutation information.

[0146] Step C3: Concatenate the first and second prediction-related information corresponding to the structural information of the same drug to obtain multiple concatenated prediction results;

[0147] Step C4: Train the prediction model to be trained using multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information to obtain the drug sensitivity prediction model.

[0148] The drug sensitivity prediction model training method provided in this disclosure obtains multiple first prediction-related information between the structural information and gene expression information of multiple test drugs through a first attention model, and multiple second prediction-related information between the structural information and gene mutation information of multiple test drugs through a second attention model. The first and second prediction-related information containing the structural information of the same drug are concatenated to obtain multiple concatenated prediction results. These multiple concatenated prediction results and multiple baseline half-inhibition concentrations are used to train the prediction model to obtain the drug sensitivity prediction model. Before training the prediction model, obtaining the prediction-related information between gene expression information, gene mutation information, and drug structural information through an attention mechanism, and training the model based on this prediction-related information, can improve the prediction performance of the drug sensitivity prediction model.

[0149] In an exemplary embodiment, step C1 may involve obtaining structural information of multiple drugs and multiple baseline half-inhibitory concentrations from the Genomics of Drug Sensitivity in Cancer (GDSC) database.

[0150] In an exemplary embodiment, the gene mutation information of the cell line to be tested in step C1 can be obtained from the Genetic Features under the Downloads module of the Genomics of Drug Sensitivity in Cancer (GDSC) database. The gene mutation information is a 310-dimensional vector, where 1 represents the corresponding gene has a mutation and 0 represents the absence of a mutation.

[0151] In an exemplary embodiment, obtaining the training sample set in step C1 may include steps C01-C04:

[0152] Step C01: Obtain the raw data of gene expression information, which includes the mean and standard deviation of multiple first gene expression features;

[0153] In an exemplary embodiment, in step C01, raw data of gene expression information of cell lines can be obtained from the Broad Institute Cancer Cell Line Encyclopedia (CCLE) database. Based on this raw data, the original data of gene expression information of the cell lines is obtained. The raw data of gene expression information of the cell lines may include multiple first gene expression features. The mean and standard deviation of these multiple first gene expression features in the raw data of gene expression information of the cell lines are calculated based on these multiple first gene expression features.

[0154] In an exemplary embodiment, the standard deviation σ of the first gene expression characteristic can be calculated using the following formula: N is the number of the first gene expression trait, X i For the expression characteristics of the i-th first gene, This represents the mean of multiple first gene expression characteristics. It is obtained by averaging the expression characteristics of multiple first genes.

[0155] Step C02: Standardize the mean values ​​of multiple first gene expression features to obtain multiple standardized expression means, standardize the standard deviations of multiple first gene expression features to obtain multiple standardized expression standard deviations, and input the multiple standardized standard deviations and multiple standardized expression means into the encoder;

[0156] In an exemplary embodiment, in step C02, the formula for standardizing the mean of any one of the mean values ​​of multiple first gene expression features is X. norm =(XX) min ) / (X max -X min ), X norm X represents the standardized expression mean, where X is the mean of the first gene expression characteristic. min X is the minimum value among the means of multiple first gene expression characteristics. max It is the largest value among the means of multiple first gene expression characteristics.

[0157] In an exemplary embodiment, in step C02, the formula for standardizing the standard deviation of any one of the standard deviations of multiple first gene expression features is σ. norm =(σ / X)*X norm , σ norm The standardized expression standard deviation is σ, where σ is the standard deviation of the first gene expression characteristic, and X is the mean of the first gene expression characteristic. normThis represents the standardized mean expression.

[0158] In this embodiment of the disclosure, the convergence of the model can be increased by standardizing the mean and standard deviation of the expression features of multiple first genes.

[0159] In this embodiment of the disclosure, since the dimension of gene expression information is much larger than that of gene mutation information and drug structure information, and the E-MTAB-3610 dataset includes expression standard deviation data in addition to the expression mean (the individual gene expression level is affected by the time and space characteristics of expression and the interference, and is necessarily a dynamic level, so the standard deviation can better reflect the actual biological significance), the method of integrating expression standard deviation on the basis of Autoencoders in this embodiment of the disclosure can reflect the actual biological characteristics.

[0160] Step C03: Control the encoder to add or subtract the standardized expression standard deviation corresponding to a portion of the standardized expression mean to obtain the standardized expression mean of multiple processes, and use the other portion of the unprocessed standardized expression mean and the standardized expression mean of the multiple processes as multiple encoding input features;

[0161] For example, encoding input features b represents the deviation of the output layer of the encoder's coding layer.

[0162] In an exemplary embodiment, step C03 can be understood as adding or subtracting the standardized expression standard deviation corresponding to the standardized expression mean from the multiple standardized expression means with a certain probability.

[0163] In an exemplary embodiment, step C03, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from a portion of the standardized expression mean, may include: controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from all of the standardized expression means. In another exemplary embodiment, step C03, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean from a portion of the standardized expression mean, may include: controlling the encoder to add the standardized expression standard deviation corresponding to the standardized expression mean to a portion of the standardized expression mean, and controlling the encoder to subtract the standardized expression standard deviation corresponding to the standardized expression mean from another portion of the standardized expression mean.

[0164] Step C04: Control the encoder to encode multiple input features to obtain gene expression information; wherein, the gene expression information includes multiple second gene expression features, and the number of multiple second gene expression features is less than the number of multiple first gene expression features.

[0165] In an exemplary implementation, such as Figure 3 As shown, the encoder includes an encoding layer, which comprises an input layer and an output layer; controlling the encoder to encode multiple encoded input features may include: controlling the encoder to perform the following operations on the multiple encoded input features to obtain multiple second gene expression features: in, Let y be the encoding input feature, y be the second gene expression feature, W be the link weights from the input layer to the output layer, b be the output layer bias, and s be a nonlinear function. In an exemplary embodiment, the nonlinear function s can be a sigmoid function.

[0166] In an exemplary implementation, such as Figure 3 As shown, the encoding layer may further include an intermediate hidden layer located between the input layer and the output layer. The input layer, intermediate hidden layer, and output layer constitute a three-layer neural network with the number of neurons decreasing progressively. In an exemplary embodiment, the number of neurons in the input layer is 1500 to 2500, the number of neurons in the intermediate hidden layer is 500 to 1500, and the number of neurons in the output layer is 250 to 750. For example, the number of neurons in the input layer is 2000, the number of neurons in the intermediate hidden layer is 1000, and the number of neurons in the output layer is 500. In this embodiment, the number of neural network layers in the encoder and the number of neurons in each neural network layer can be set according to actual needs during modeling. For example, the number of neurons in the input layer can be set to be greater than or less than 1000 depending on the number of encoded input feature values, the intermediate hidden layer can be set to one layer or more, and the number of neurons in the output layer can be set to less than or greater than 500 according to actual needs. In this embodiment, the encoder can be an autoencoder, which is a neural network model. In this embodiment of the disclosure, since the acquired gene expression information contains many features, encoding multiple first gene expression features using an encoder can reduce dimensionality, resulting in fewer second gene expression features output by the output layer compared to the number of first gene expression features. This saves storage space, improves computation speed, and removes redundant features, such as... Figure 3 As shown, in the encoder's coding layer, the number of neurons in the output layer is less than the number of neurons in the input layer.

[0167] In an exemplary embodiment, the structural information of any drug to be tested can be represented in the manner described in Table 1 above.

[0168] In an exemplary embodiment, before inputting the multiple standardized standard deviations and multiple standardized expression means into the encoder in step C02, the method further includes:

[0169] Step C0: Obtain the training sample set of the first gene expression features. The training sample set of the first gene expression features includes the mean samples of multiple first gene expression features and the standard deviation samples of the corresponding multiple first gene expression features. Train the encoder based on the mean samples of multiple first gene expression features and the standard deviation samples of multiple first gene expression features to obtain the link weights W, the output layer bias b, and the nonlinear function s.

[0170] Step C0 is performed before the multiple standardized standard deviations and multiple standardized expression means are input into the encoder, and may be performed in step C02 or before step C01.

[0171] In an exemplary embodiment, obtaining the training sample set of the first gene expression feature may include: obtaining the original data of the training sample set of the first gene expression feature from the Broad Institute Cancer Cell Line Encyclopedia (CCLE) database, and calculating the original data of the training sample set of the first gene expression feature based on the original data of the training sample set of the first gene expression feature. The original data of the training sample set of the first gene expression feature may include training samples of multiple first gene expression features, and the original data of the training sample set of the first gene expression feature may include the mean and standard deviation of multiple first gene expression features, which are calculated based on the multiple first gene expression features.

[0172] In an exemplary embodiment, step C0, which trains the encoder based on the mean sample and the standard deviation sample of the expression features of multiple first genes, may include step D0: inputting multiple spliced ​​prediction results into the encoder to be trained multiple times through multiple iterations, optimizing the encoder to be trained based on the results of each iteration, and obtaining a trained encoder.

[0173] In an exemplary embodiment, the encoder further includes a decoding layer; in step D0, multiple concatenated prediction results are input into the encoder to be trained multiple times through multiple iterations, and the encoder to be trained is optimized based on the results of each iteration, which may include steps D01-D05:

[0174] Step D01: Standardize the mean samples of multiple first gene expression features to obtain standardized expression mean samples, standardize the standard deviation samples of multiple first gene expression features to obtain standardized expression standard deviation samples, and input the multiple standardized standard deviation samples and multiple standardized expression mean samples into the encoder to be trained.

[0175] Step D02: Control the encoder to be trained to add or subtract the standard deviation of the expression corresponding to a portion of the standardized expression mean samples to obtain multiple processed standardized expression mean samples, and use another portion of unprocessed standardized expression mean samples and the multiple processed standardized expression mean samples as multiple encoding input feature samples.

[0176] Step D03: Control the encoder to be trained to encode multiple input feature samples to obtain multiple second gene expression feature samples as gene expression information samples. The number of multiple second gene expression feature samples is less than the number of multiple first gene expression feature samples.

[0177] Step D04: Input the gene expression information sample into the decoding layer to obtain the decoded information;

[0178] Step D05: Calculate the loss value based on the decoded information and the mean samples of multiple first gene expression features. Optimize the encoder based on the loss value. Use the optimized encoder as the encoder to be trained in the next iteration. Continue to control the encoder to be trained to add or subtract the expression standard deviation sample corresponding to the standardized expression mean sample from a portion of the standardized expression mean sample.

[0179] In an exemplary embodiment, step D02 can be understood as adding or subtracting the standardized expression standard deviation sample corresponding to the standardized expression mean sample from multiple standardized expression mean samples with a certain probability.

[0180] In an exemplary embodiment, step D02, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean sample from a portion of the standardized expression mean samples, may include: controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean sample from all portions of the standardized expression mean samples. In another exemplary embodiment, step D02, controlling the encoder to add or subtract the standardized expression standard deviation corresponding to the standardized expression mean sample from a portion of the standardized expression mean samples, may include: controlling the encoder to add the standardized expression standard deviation corresponding to the standardized expression mean sample to a portion of the standardized expression mean samples, and controlling the encoder to subtract the standardized expression standard deviation corresponding to the standardized expression mean sample from another portion of the standardized expression mean samples.

[0181] In an exemplary embodiment, step D04, inputting gene expression information into the decoding layer to obtain decoding information, may include: controlling the encoder to be trained to perform the following operation on the gene expression information to obtain the decoding information: z = s(W'y + b'), where s is a nonlinear function, W' is the link weight of the decoding layer, b' is the bias of the decoding layer, y is the feature value in the gene expression information, and z is the feature value of the decoding information;

[0182] In step D05, calculating the loss value based on the decoded information and the mean samples of multiple first gene expression features may include: controlling the encoder to be trained to perform the following operation based on the mean samples of multiple first gene expression features and the decoded information to obtain the loss value: L(x,z)=||xz|| 2 , where L(x,z) is the loss function, x is the feature value in the mean sample of the first gene expression feature, and z is the feature value of the decoded information.

[0183] In the embodiments disclosed herein, such as Figure 5 As shown, the decoding layer and the encoding layer can be symmetrical, and the neurons in the output layer of the encoding layer can serve as the neurons in the input layer of the decoding layer. The information output by the decoding layer is equivalent to verifying the encoding effect of the encoding layer.

[0184] In an exemplary implementation, such as Figure 5 As shown, both the decoding and encoding layers in the encoder are three-layer neural networks.

[0185] In an exemplary embodiment, the encoder's encoding layer comprises an input layer, intermediate hidden layers, and an output layer, forming a three-layer neural network with a progressively decreasing number of neurons. In this embodiment, the input layer has 1500 to 2500 neurons, the intermediate hidden layers have 500 to 1500 neurons, and the output layer has 250 to 750 neurons. For example, the input layer may have 2000 neurons, the intermediate hidden layers 1000 neurons, and the output layer 500 neurons.

[0186] In an exemplary embodiment, the input layer, intermediate hidden layer, and output layer of the decoding layer constitute a three-layer neural network with an increasing number of neurons. In this exemplary embodiment, the number of neurons in the input layer of the decoding layer is 250 to 750, the number of neurons in the intermediate hidden layer is 500 to 1500, and the number of neurons in the output layer is 1500 to 2500. For example, the number of neurons in the input layer of the decoding layer is 500, the number of neurons in the intermediate hidden layer is 1000, and the number of neurons in the output layer is 2000.

[0187] In this embodiment of the disclosure, Figure 5 The encoder shown can be an automatic encoder.

[0188] In an exemplary embodiment, step C2, based on the first attention model, obtains multiple first prediction-related information between the structural information of multiple drugs and gene expression information, which may include:

[0189] M11: Multiply the gene expression information by the first weight matrix to obtain the first vector; multiply the structural information of multiple drugs by the corresponding second weight matrix to obtain multiple second vectors; multiply the structural information of multiple drugs by the corresponding third weight matrix to obtain multiple third vectors.

[0190] M12: Normalize the first and second vectors corresponding to the structural information of the same drug to obtain multiple first processing results;

[0191] M13: Multiply the first processing result and the third vector corresponding to the structural information of the same drug to obtain multiple first related information.

[0192] In an exemplary embodiment, before obtaining multiple first prediction related information between the structural information of multiple drugs and the gene expression information based on the first attention model in step C2, it may further include: training the first attention model using the structural information and gene expression information of multiple drugs to obtain a first weight matrix, a second weight matrix, and a third weight matrix.

[0193] In an exemplary embodiment, step C2, based on the second attention model, obtains multiple second prediction-related information between the structural information of multiple drugs and the gene mutation information, which may include:

[0194] Step M21: Multiply the gene mutation information by the fourth weight matrix to obtain the fourth vector; multiply the structural information of multiple drugs by the corresponding fifth weight matrix to obtain multiple fifth vectors; multiply the structural information of multiple drugs by the corresponding sixth weight matrix to obtain multiple sixth vectors.

[0195] Step M22: Normalize the fourth and fifth vectors corresponding to the structural information of the same drug to obtain multiple second processing results;

[0196] Step M23: Multiply the second processing result and the sixth vector corresponding to the structural information of the same drug to obtain multiple second related information.

[0197] In an exemplary embodiment, before obtaining multiple second prediction related information between the structural information of multiple drugs and the gene mutation information based on the second attention model in step C2, it may further include: training the second attention model using the structural information and gene expression information of multiple drugs to obtain a fourth weight matrix, a fifth weight matrix, and a sixth weight matrix.

[0198] In an exemplary embodiment, step C4, training the prediction model to be trained using multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information to obtain a drug sensitivity prediction model, may include: training the prediction model to be trained multiple times in a multi-iterative manner based on multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information to obtain a drug sensitivity prediction model.

[0199] In each iteration, multiple concatenated prediction results are input into the drug sensitivity model to be trained, resulting in multiple predicted half-inhibition concentrations (WICs). Sensitivity loss information is then derived based on these predicted WICs and multiple baseline WICs. The model is then optimized based on this sensitivity loss information, and the optimized model is used as the training model for the next iteration. Alternatively, in each iteration, multiple concatenated prediction results are input into the model in batches, resulting in multiple predicted WICs for the current batch. Sensitivity loss information for the current batch is derived based on these predicted WICs and multiple baseline WICs for the current batch. The model is then optimized based on this sensitivity loss information, and the optimized model is used as the training model for the next batch or the next iteration. Specifically, if the current batch is the last batch, the optimized model is used as the training model for the next iteration; otherwise, it is used as the training model for the next batch.

[0200] In an exemplary embodiment, obtaining the sensitivity loss information of the current batch based on multiple predicted half-inhibition concentrations and multiple baseline half-inhibition concentrations corresponding to the current batch may include:

[0201] The following calculations are performed on multiple predicted half-inhibition concentrations for the current batch and multiple baseline half-inhibition concentrations for the corresponding current batch to obtain the sensitivity loss information for the current batch: in, For the predicted half-inhibition concentration information, y i RSME represents the baseline half-inhibition concentration, RSME represents the sensitivity loss information for the current batch, and N represents the quantity of the current batch.

[0202] In an exemplary embodiment, step C4 may further include setting the parameters of the prediction model to be trained before training the prediction model using multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information.

[0203] The parameters of the prediction model to be trained may include: the optimizer is set to SGD, the batch size N is set to 32, the number of iterations is set to 100, and the dropout probability is set to 0.001. SGD stands for Stochastic Gradient Descent.

[0204] In an exemplary embodiment, the parameters of the prediction model to be trained may further include: the number of neural network layers of the prediction model to be trained, and the number of neurons in each neural network layer.

[0205] Before training the prediction model using multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information, the process also includes: establishing the prediction model to be trained based on the parameters of the prediction model to be trained.

[0206] In an exemplary embodiment, the neural network of the prediction model to be trained has four layers, and the number of neurons in the four layers decreases sequentially: the number of neurons in the first layer is 400 to 600, the number of neurons in the second layer is 100 to 300, the number of neurons in the third layer is 80 to 120, and the number of neurons in the fourth layer is 1 to 5.

[0207] This disclosure also provides a drug sensitivity prediction device, such as... Figure 6 As shown, it may include:

[0208] The acquisition module 111 is configured to acquire gene expression information of the cell line to be tested, gene mutation information of the cell line to be tested, and structural information of the drug to be tested.

[0209] The first feature fusion module 112 is configured to calculate the first relevant information between the structural information and gene expression information of the drug under test based on the first attention model.

[0210] The second feature fusion module 113 is configured to calculate the second relevant information between the structural information and gene mutation information of the drug under test based on the second attention model.

[0211] The splicing module 114 is configured to splice the first relevant information and the second relevant information to obtain the splicing result;

[0212] The drug sensitivity prediction module 115 is configured to perform prediction processing on the splicing results based on the drug sensitivity prediction model to obtain the sensitivity information of the cell line to the drug.

[0213] This disclosure also provides a drug sensitivity prediction model training device, such as... Figure 7 As shown, it may include:

[0214] The sample acquisition module 211 is configured to acquire a training sample set, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibition concentrations; wherein each baseline half-inhibition concentration corresponds to the structural information of one of the drugs.

[0215] The first feature fusion prediction module 212 is configured to obtain multiple first prediction related information between the structural information of multiple drugs and the gene expression information based on the first attention model.

[0216] The second feature fusion prediction module 213 is configured to obtain multiple second prediction related information between the structural information of multiple drugs and the gene mutation information based on the second attention model.

[0217] The splicing prediction module 214 is configured to splice together the first prediction-related information and the second prediction-related information containing the structural information of the same drug to obtain multiple spliced ​​prediction results.

[0218] The sensitivity prediction module 215 is configured to train the prediction model to be trained using multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information to obtain a drug sensitivity prediction model.

[0219] This disclosure also provides a non-transient computer-readable storage medium configured to store computer program instructions, wherein the computer program instructions, when executed, can implement the drug sensitivity prediction method described in any of the above embodiments.

[0220] This disclosure also provides a non-transient computer-readable storage medium configured to store computer program instructions, wherein the computer program instructions, when executed, can implement the drug sensitivity prediction model training method described in any of the above embodiments.

[0221] This disclosure also provides a drug sensitivity prediction device, such as... Figure 8 As shown, it may include a first memory 311, a first processor 312, and a computer program 3110 stored in the first memory 311 and executable on the first processor 312 to perform:

[0222] Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell lines to be tested.

[0223] Based on the first attention model, the first correlation between the structural information and gene expression information of the drug under test is calculated; based on the second attention model, the second correlation between the structural information and gene mutation information of the drug under test is calculated.

[0224] The first relevant information and the second relevant information are concatenated to obtain the concatenation result;

[0225] Based on the drug sensitivity prediction model, the splicing results are processed to obtain the sensitivity information of the cell line to the drug.

[0226] This disclosure also provides a drug sensitivity prediction model training device, such as... Figure 9 As shown, it may include a second memory 411, a second processor 412, and a computer program 4110 stored on the second memory 411 and executable on the second processor 412 to perform:

[0227] A training sample set is obtained, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibition concentrations; wherein each baseline half-inhibition concentration corresponds to the structural information of one of the drugs.

[0228] Based on the first attention model, multiple first-prediction related information between the structural information of multiple drugs and gene expression information is obtained; based on the second attention model, multiple second-prediction related information between the structural information of multiple drugs and gene mutation information is obtained.

[0229] By splicing together the first and second prediction-related information that contain the structural information of the same drug, multiple spliced ​​prediction results are obtained.

[0230] The drug sensitivity prediction model is obtained by training the prediction model using multiple spliced ​​prediction results and multiple baseline half-inhibition concentration information.

[0231] In the drug sensitivity prediction method provided in this embodiment, the SMILES structural information of the drug is fused with gene expression information to obtain a first relevant information, and the SMILES structural information of the drug is fused with gene mutation information to obtain a second relevant information. The first and second relevant information are concatenated to obtain a concatenated result, which is then input into a fully connected network drug sensitivity prediction model. The logical structure diagram of drug sensitivity prediction is shown below. Figure 10 As shown.

[0232] The drug sensitivity prediction and model training method, storage medium, and device provided in this disclosure include a drug sensitivity prediction method. In this method, a first relevant correlation is obtained between the structural information and gene expression information of the drug to be tested using a first attention model. A second relevant correlation is obtained between the structural information and gene mutation information of the drug to be tested using a second attention model. The first and second relevant correlations are concatenated to obtain a concatenated result. This concatenated result is then processed by a drug sensitivity prediction model to obtain the sensitivity information of the cell line to the drug to be tested. Before the drug sensitivity prediction model performs its prediction, obtaining the relevant correlation between gene expression information, gene mutation information, and the structural information of the drug through an attention mechanism, and then predicting drug sensitivity based on this relevant correlation, can improve the prediction effect of the drug sensitivity prediction model and overcome the shortcomings of poor drug sensitivity prediction performance.

[0233] The accompanying drawings of the embodiments disclosed herein only relate to the structures involved in the embodiments of this disclosure; other structures can be referred to in a general design.

[0234] Where there is no conflict, the features of the embodiments disclosed herein can be combined with each other to obtain new embodiments.

[0235] While the embodiments disclosed herein are as described above, the content is merely for the purpose of facilitating understanding of these embodiments and is not intended to limit them. Any person skilled in the art to which these embodiments pertain may make any modifications and changes to the form and details of the implementation without departing from the spirit and scope disclosed herein; however, the patent protection scope of these embodiments shall still be determined by the scope defined in the appended claims.

Claims

1. A method for predicting drug sensitivity, comprising: Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell lines to be tested. Based on the first attention model, the first relevant information between the structural information of the drug to be tested and the gene expression information is calculated; Based on the second attention model, a second relevant correlation is calculated between the structural information of the drug under test and the gene mutation information; The first relevant information and the second relevant information are concatenated to obtain the concatenation result; Based on the drug sensitivity prediction model, the splicing results are processed to obtain the sensitivity information of the cell line to the drug.

2. The sensitivity prediction method according to claim 1, wherein, The calculation of the first relevant information between the structural information of the drug under test and the gene expression information based on the first attention model includes: The gene expression information is multiplied by the first weight matrix to obtain a first vector, the structure information of the drug is multiplied by the second weight matrix to obtain a second vector, and the structure information of the drug is multiplied by the third weight matrix to obtain a third vector. The first vector and the second vector are normalized to obtain a first processing result. The first processing result is then multiplied by the third vector to obtain the first related information.

3. The sensitivity prediction method according to claim 2, wherein, The normalization process of the first vector and the second vector to obtain the first processing result includes: The second vector is transposed to obtain the transpose of the second vector. The first vector is multiplied by the transpose of the second vector to obtain the first product. The first product is divided by the first constant to obtain the first processing result. The first constant is the arithmetic square root of the dimension of the second vector.

4. The sensitivity prediction method according to claim 2 or 3, further comprising, after obtaining the gene expression information of the cell line to be tested and the structural information of the drug to be tested: The gene expression information is subjected to dimensionality reduction operation through a first convolutional neural network to obtain dimensionality-reduced gene expression information; the drug structure information is subjected to dimensionality reduction operation through a second convolutional neural network to obtain dimensionality-reduced drug structure information. The step of multiplying the gene expression information with the first weight matrix to obtain the first vector includes: multiplying the dimension-reduced gene expression information with the first weight matrix to obtain the first vector; The step of multiplying the structural information of the drug with the second weight matrix to obtain the second vector includes: multiplying the dimensionality-reduced structural information of the drug with the second weight matrix to obtain the second vector; The step of multiplying the structural information of the drug with the third weight matrix to obtain the third vector includes: multiplying the dimensionality-reduced structural information of the drug with the third weight matrix to obtain the third vector.

5. The sensitivity prediction method according to claim 1, wherein, The calculation of the second relevant information between the structural information of the drug under test and the gene mutation information based on the second attention model includes: The gene mutation information is multiplied by the fourth weight matrix to obtain the fourth vector, the drug structure information is multiplied by the fifth weight matrix to obtain the fifth vector, and the drug structure information is multiplied by the sixth weight matrix to obtain the sixth vector. The fourth and fifth vectors are normalized to obtain a second processing result. The second processing result is then multiplied by the sixth vector to obtain the second relevant information.

6. The sensitivity prediction method according to claim 5, wherein, The normalization process for the fourth and fifth vectors to obtain the second processing result includes: The fifth vector is transposed to obtain the transpose of the fifth vector. The fourth vector is multiplied by the transpose of the fifth vector to obtain the second product. The second product is divided by the second constant to obtain the second processing result. The second constant is the arithmetic square root of the dimension of the fifth vector.

7. The sensitivity prediction method according to claim 5 or 6, wherein, After obtaining the gene mutation information of the cell line to be tested and the structural information of the drug to be tested, the method further includes: performing dimensionality reduction operation on the gene mutation information through a third convolutional neural network to obtain dimensionality-reduced gene mutation information; and performing dimensionality reduction operation on the structural information of the drug through a second convolutional neural network to obtain dimensionality-reduced drug structural information. The step of multiplying the gene mutation information with the fourth weight matrix to obtain the fourth vector includes: multiplying the dimension-reduced gene mutation information with the fourth weight matrix to obtain the fourth vector; The step of multiplying the structural information of the drug with the fifth weight matrix to obtain the fifth vector includes: multiplying the dimensionality-reduced structural information of the drug with the fifth weight matrix to obtain the fifth vector; The step of multiplying the structural information of the drug with the sixth weight matrix to obtain the sixth vector includes: multiplying the dimensionality-reduced structural information of the drug with the sixth weight matrix to obtain the sixth vector.

8. The sensitivity prediction method according to claim 1, wherein, The acquisition of gene expression information of the cell line to be tested includes: Obtain the raw data of the gene expression information, which includes the mean of multiple first gene expression features and the standard deviation of multiple first gene expression features; The mean values ​​of the multiple first gene expression features are standardized to obtain multiple standardized expression means, and the standard deviations of the multiple first gene expression features are standardized to obtain multiple standardized expression standard deviations. The multiple standardized expression standard deviations and the multiple standardized expression means are input into the encoder. The encoder is controlled to add or subtract the standardized expression standard deviation corresponding to a portion of the standardized expression mean to obtain a plurality of processed standardized expression means. The other portion of unprocessed standardized expression means and the plurality of processed standardized expression means are used as a plurality of encoded input features. The encoder is controlled to encode the plurality of encoded input features to obtain a plurality of second gene expression features as the gene expression information. The number of the plurality of second gene expression features is less than the number of the plurality of first gene expression features.

9. The sensitivity prediction method according to claim 8, wherein, The encoder includes an encoding layer, which includes an input layer and an output layer; The process of controlling the encoder to encode the plurality of coded input features includes: The encoder is controlled to perform the following operation on multiple encoded input features to obtain multiple second gene expression features: y=s(W +b), where, Let y be the encoded input feature, y be the second gene expression feature, W be the link weight from the input layer to the output layer, b be the deviation of the output layer, and s be a nonlinear function.

10. The sensitivity prediction method according to claim 9, wherein, The encoding layer also includes an intermediate hidden layer located between the input layer and the output layer. The input layer, the intermediate hidden layer, and the output layer constitute a three-layer neural network with the number of neurons decreasing gradually.

11. The sensitivity prediction method according to claim 1, wherein, The sensitivity prediction model comprises a four-layer neural network with a gradually decreasing number of neurons.

12. A method for training a drug sensitivity prediction model, comprising: A training sample set is obtained, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and multiple baseline half-inhibition concentrations; wherein each baseline half-inhibition concentration corresponds to the structural information of a drug. Based on the first attention model, multiple first prediction related information is obtained between the structural information of the multiple drugs and the gene expression information; based on the second attention model, multiple second prediction related information is obtained between the structural information of the multiple drugs and the gene mutation information. By splicing together the first and second prediction-related information corresponding to the structural information of the same drug, multiple spliced ​​prediction results are obtained. The drug sensitivity prediction model is obtained by training the prediction model to be trained using the multiple spliced ​​prediction results and the multiple baseline half-inhibition concentration information.

13. The model training method according to claim 12, wherein, The drug sensitivity prediction model is trained using the multiple concatenated prediction results and the multiple baseline half-inhibition concentration information to obtain the drug sensitivity prediction model, including: Based on the multiple concatenated prediction results and the multiple baseline half-inhibition concentration information, the prediction model to be trained is trained multiple times in a multi-iteration manner to obtain the drug sensitivity prediction model. In each iteration, the multiple concatenated prediction results are input into the drug sensitivity model to be trained to obtain multiple predicted half-inhibition concentrations (WICs). Sensitivity loss information is obtained based on the multiple predicted WICs and the multiple baseline WICs. The prediction model to be trained is optimized based on the sensitivity loss information, and the optimized model is used as the prediction model to be trained in the next iteration. Alternatively, in each iteration, the multiple concatenated prediction results are input into the prediction model to be trained in batches to obtain multiple predicted WICs for the current batch. Sensitivity loss information for the current batch is obtained based on the multiple predicted WICs and the multiple baseline WICs corresponding to the current batch. The prediction model to be trained is optimized based on the sensitivity loss information for the current batch, and the optimized model is used as the prediction model to be trained in the next batch or the next iteration.

14. The model training method according to claim 12, wherein, The method, based on a first attention model, obtains multiple first prediction-related information between the structural information of the multiple drugs and the gene expression information, including: The gene expression information is multiplied by the first weight matrix to obtain a first vector; the structural information of the multiple drugs is multiplied by the corresponding second weight matrix to obtain multiple second vectors; and the structural information of the multiple drugs is multiplied by the corresponding third weight matrix to obtain multiple third vectors. The first vector and the second vector corresponding to the structural information of the same drug are normalized to obtain multiple first processing results; The structural information of the same drug is multiplied by the first processing result and the third vector to obtain the multiple first related information.

15. The model training method according to claim 14, wherein, Before obtaining multiple first prediction-related information between the multiple drug structure information and the gene expression information based on the first attention model, the method further includes: The first attention model is trained using the structural information of the multiple drugs and the gene expression information to obtain the first weight matrix, the second weight matrix, and the third weight matrix.

16. The model training method according to claim 12, wherein, The second attention model yields multiple second prediction-related information pairs between the structural information of the multiple drugs and the gene mutation information, including: The gene mutation information is multiplied by the fourth weight matrix to obtain the fourth vector; the structural information of the multiple drugs is multiplied by the corresponding fifth weight matrix to obtain multiple fifth vectors; and the structural information of the multiple drugs is multiplied by the corresponding sixth weight matrix to obtain multiple sixth vectors. The fourth and fifth vectors corresponding to the structural information of the same drug are normalized to obtain multiple second processing results; The second processing result and the sixth vector corresponding to the structural information of the same drug are multiplied together to obtain the multiple second related information.

17. The model training method according to claim 16, wherein, Before obtaining multiple second prediction-related information between the structural information of the multiple drugs and the gene mutation information based on the second attention model, the method further includes: The second attention model is trained using the structural information of the multiple drugs and the gene expression information to obtain the fourth weight matrix, the fifth weight matrix, and the sixth weight matrix.

18. The model training method according to claim 12, wherein, The acquisition of the training sample set includes: Obtain the raw data of the gene expression information, which includes the mean of multiple first gene expression features and the standard deviation of multiple first gene expression features; The mean values ​​of the multiple first gene expression features are standardized to obtain multiple standardized expression means, and the standard deviations of the multiple first gene expression features are standardized to obtain multiple standardized expression standard deviations. The multiple standardized standard deviations and the multiple standardized expression means are input into the encoder. The encoder is controlled to add or subtract the standardized expression standard deviation corresponding to a portion of the standardized expression mean to obtain a plurality of processed standardized expression mean, and the other portion of unprocessed standardized expression mean and the plurality of processed standardized expression mean are used as a plurality of encoded input features; The encoder is controlled to encode the plurality of coded input features to obtain a plurality of second gene expression features as the gene expression information, wherein the number of the plurality of second gene expression features is less than the number of the plurality of first gene expression features.

19. The model training method according to claim 18, wherein, The encoder includes an encoding layer, which includes an input layer and an output layer; The process of controlling the encoder to encode the plurality of coded input features includes: The encoder is controlled to perform the following operation on multiple encoded input features to obtain multiple second gene expression features: y=s(W +b); where, Let y be the encoded input feature, y be the second gene expression feature, W be the link weight from the input layer to the output layer, b be the deviation of the output layer, and s be a nonlinear function.

20. The model training method according to claim 19, wherein, Before inputting the plurality of standardized standard deviations and the plurality of standardized expression means into the encoder, the method further includes: A training sample set of first gene expression features is obtained, which includes mean samples of multiple first gene expression features and standard deviation samples of corresponding multiple first gene expression features. The encoder is trained based on the mean samples of multiple first gene expression features and the standard deviation samples of multiple first gene expression features to obtain the link weight W, the bias b of the output layer, and the nonlinear function s.

21. The model training method according to claim 20, wherein, The step of training the encoder based on the mean sample and the standard deviation sample of the expression features of the plurality of first genes includes: The multiple concatenated prediction results are input into the encoder to be trained multiple times through multiple iterations, and the encoder to be trained is optimized based on the results of each iteration.

22. The model training method according to claim 21, wherein, The encoder also includes a decoding layer; The step of repeatedly inputting the multiple concatenated prediction results into the encoder to be trained through multiple iterations, and optimizing the encoder to be trained based on the results of each iteration, includes: The mean samples of the expression features of the plurality of first genes are standardized to obtain standardized expression mean samples, and the standard deviation samples of the expression features of the plurality of first genes are standardized to obtain standardized expression standard deviation samples. The plurality of standardized standard deviation samples and the plurality of standardized expression mean samples are input into the encoder to be trained. The encoder to be trained is controlled to add or subtract the standard deviation of the expression corresponding to a portion of the standardized expression mean samples to obtain multiple processed standardized expression mean samples. The other portion of unprocessed standardized expression mean samples and the multiple processed standardized expression mean samples are used as multiple encoding input feature samples. The encoder to be trained is controlled to encode the plurality of encoded input feature samples to obtain a plurality of second gene expression feature samples as the gene expression information samples, wherein the number of the plurality of second gene expression feature samples is less than the number of the plurality of first gene expression feature samples; The gene expression information sample is input into the decoding layer to obtain the decoded information; The loss value is calculated based on the decoded information and the mean samples of the expression features of the multiple first genes. The encoder is optimized based on the loss value. The optimized encoder is used as the encoder to be trained in the next iteration. The encoder to be trained is controlled to add or subtract the expression standard deviation sample corresponding to the standardized expression mean sample from a portion of the standardized expression mean sample.

23. The model training method according to claim 22, wherein, The step of inputting the gene expression information into the decoding layer to obtain decoded information includes: The encoder to be trained is controlled to perform the following operation on the gene expression information to obtain the decoded information: z=s(W'y+b'), where s is a nonlinear function, W' is the link weight of the decoding layer, b' is the bias of the decoding layer, y is the feature value in the gene expression input information, and z is the feature value of the decoded information; The step of calculating the loss value based on the decoded information and the mean sample of the expression features of the plurality of first genes includes: The encoder to be trained is controlled to perform the following operation based on the mean samples of multiple first gene expression features and the decoded information to obtain the loss value: L(x,z)=||xz|| 2 , where L(x,z) is the loss function, x is the feature value in the mean sample of the first gene expression feature, and z is the feature value of the decoded information.

24. A non-transient computer-readable storage medium, the storage medium being configured to store computer program instructions, wherein, When the computer program instructions are executed, they can implement the drug sensitivity prediction method according to any one of claims 1 to 11, or the computer program instructions can implement the drug sensitivity prediction model training method according to any one of claims 12 to 23.

25. A drug sensitivity prediction device, comprising a first memory, a first processor, and a computer program stored in the first memory and executable on the first processor, for performing: Obtain gene expression information, gene mutation information, and structural information of the drug to be tested from the cell lines to be tested. Based on the first attention model, the first relevant information between the structural information of the drug to be tested and the gene expression information is calculated; Based on the second attention model, a second relevant correlation is calculated between the structural information of the drug under test and the gene mutation information; The first relevant information and the second relevant information are concatenated to obtain the concatenation result; Based on the drug sensitivity prediction model, the splicing results are processed to obtain the sensitivity information of the cell line to the drug.

26. A drug sensitivity prediction model training device, comprising a second memory, a second processor, and a computer program stored in the second memory and executable on the second processor, for performing: A training sample set is obtained, which includes gene expression information of cell lines, gene mutation information of cell lines, structural information of multiple drugs, and information on multiple baseline half-inhibitory concentrations; wherein, Each baseline half-inhibitory concentration corresponds to the structural information of a drug. Based on the first attention model, multiple first prediction related information is obtained between the structural information of the multiple drugs and the gene expression information; Based on the second attention model, multiple second prediction-related information between the structural information of the multiple drugs and the gene mutation information is obtained; By splicing together the first and second prediction-related information corresponding to the structural information of the same drug, multiple spliced ​​prediction results are obtained. The drug sensitivity prediction model is obtained by training the prediction model to be trained using the multiple spliced ​​prediction results and the multiple baseline half-inhibition concentration information.

Images