Method for predicting sensitivity of anticancer drugs based on deep factorization machine

Abstract

The application is based on a method for predicting anticancer drug sensitivity by deep factor decomposition machine, characterized in that it comprises the following steps: step 1, constructing a data set and preprocessing; step 2, constructing a deep factor decomposition machine DeepFM model: the main idea of the deep factor decomposition machine DeepFM model is to construct a factor decomposition machine FM model, construct a deep neural network model DNN, and fuse the factor decomposition machine FM and the deep neural network to construct a deep factor decomposition machine DeepFM model; step 3, training and evaluating the model results of each stage, and dividing the anticancer drug sensitivity data obtained in step 1 into 5 parts, performing 5-fold cross-validation, and training the model results of each stage in this stage. The application separately models the low-order feature combination of the drug, and fuses the high-order feature combination to construct a model that fuses the deep neural network model and the factor decomposition machine model, better extracts the drug feature structure, and makes the constructed model more accurate and higher in precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep model DNN and factorization machine FM feature extraction, and particularly relates to a method for predicting anticancer drug sensitivity based on a deep factorization machine. BACKGROUND

[0002] Cancer refers to a class of diseases caused by uncontrolled proliferation and spread of abnormal cells, which can form tumors and can invade surrounding tissues or spread to other parts of the body, causing body dysfunction or even life-threatening, cancer is closely related to human DNA, it is a disease caused by mutation or abnormal change of cell DNA, research shows that different types of cancer are usually related to specific DNA mutations or abnormalities, by analyzing the tumor tissue of cancer patients in terms of genomics, transcriptomics and epigenomics, etc., the mechanism of cancer can be revealed, and the relationship between cancer and DNA is understood. However, in the clinical treatment of cancer in real life, different cancer patients often show different sensitivities to the same anticancer drugs, that is, different individuals also need individual treatment when facing the same drugs, therefore, cancer patients should be treated individually, that is, the DNA sequence of the cancer patient is used to select the most effective anticancer drug for him, which will greatly promote the development of the field of cancer treatment.

[0003] With the rapid development of artificial intelligence (AI), researchers are applying it to cancer research. Based on big data and machine learning, AI can analyze tumor DNA and transcriptome data to understand tumor pathogenesis, patient characteristics, and drug sensitivity, enabling earlier preventive measures and personalized treatment plans for each individual. Furthermore, AI technology allows for safer experimental environments. Knowledge of machine learning and AI technologies is significantly aiding data prediction, enabling simpler methods to standardize and process extremely large and complex datasets, converting them into usable data in the required formats required for research. Researchers are also applying data prediction to biology, driving rapid development in the field of biological informatics. Several research organizations, including CCLE (Cancer Cell Line Encyclopedia) and TCGA (The Cancer Genome Atlas), have established data analysis platforms based on DNA sequence data from multiple patients. Numerous research groups, both domestically and internationally, have explored various genomic databases and studied the pathogenic genes of certain cancers. For example, Gao Meng, Cheng Zhujun, and Zhang Cheng published "Breast cancer-specific copy number variation leads to DNA repair imbalance." This research also includes studies using the CCLE database, using baseline gene expression to establish a classification model for predicting drug sensitivity, and introducing algorithms such as random forests and support vector machines. Existing research results have also demonstrated high accuracy in predicting drug sensitivity. For example, in the article "Anticancerdrug sensitivity prediction in cell lines from baseline gene expression through recursive feature selection" published by Zuoli Dong et al., baseline gene expression data were used through recursive feature selection to predict the sensitivity of cell lines to anticancer drugs, achieving a prediction accuracy of 80%. In addition, in the article "Prediction of Anticancer Drug Sensitivity Based on Cell Line-Drug K-Nearest Neighbors" published by Wang Bo, Wei Dong, and Li Yushuang on February 23, 2020, the gene expression and gene mutation characteristics of the cell lines were comprehensively considered, a new definition of cell line similarity was given, and combined with the drug similarity measurement method, a "cell line-drug K-nearest neighbor" calculation model was proposed. The obtained anticancer drug sensitivity prediction results were significantly better than the classic model.

[0004] Factorization Machine (FM) is a machine learning model that combines the efficiency of linear models and the ability of factorization models to handle large-scale sparse data and learn interactions between features. FM model represents each feature's embedding as a low-dimensional vector by factorizing features to capture the interaction effects between features. By considering feature interactions, FM can better model the non-linear relationships between features, thereby improving prediction accuracy. FM model can handle high-dimensional sparse data, suitable for large-scale datasets, with high prediction accuracy and efficiency, and can learn the interaction relationships between features, including second-order and higher-order interactions. Deep Neural Network (DNN) is a neural network structure composed of multiple hidden layers, with strong learning and representation capabilities, which can effectively capture complex non-linear relationships in large datasets. In cancer drug sensitivity prediction, DNN can be trained using large-scale biomedical data (such as genomic data, transcriptomic data, protein interaction data, etc.) and learn complex drug-gene or drug-disease interaction patterns. Through deep learning, DNN can automatically extract and learn high-level features in data, thereby more accurately predicting drug sensitivity to cancer.

[0005] The prior art lacks precision in extracting drug feature structures, and the correlation between relevant data is not fully considered in establishing a multi-element network model, which leads to inaccurate drug sensitivity analysis. Therefore, it is necessary to study a method for predicting the sensitivity of anticancer drugs based on deep factorization machine. SUMMARY

[0006] The present application combines deep neural networks and factorization machines to better extract feature vectors and improve the accuracy of models for predicting anticancer drug sensitivity in cell lines by considering the correlation between data such as genes, anticancer drugs, and cell gene expression.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0008] The method for predicting the sensitivity of anticancer drugs based on deep factorization machine, characterized in that it comprises the following steps:

[0009] Step 1: Constructing a dataset and preprocessing;

[0010] 1-1, Obtain gene expression data of cells to construct a dataset;

[0011] 1-2, Obtain drug chemical structure data, and obtain the chemical structure of all drugs in the dataset of step 1-1 from PubChem;

[0012] 1-3, using ActArea as an indicator of drug sensitivity, standardizing and normalizing drug sensitivity data, and determining the final data set;

[0013] Step 2: Building a Deep Factorization Machine DeepFM model: the main idea of the model is to build a factorization machine FM model, build a deep neural network model, and fuse the factorization machine with the deep neural network to build a deep factorization machine DeepFM model;

[0014] Step 3: Train and evaluate the model results of each stage, divide the anti-cancer drug sensitivity data obtained in step one into 5 parts, and perform 5-fold cross-validation. In this stage, in order to train the model results of each stage, the Pearson correlation coefficient and the epoch parameter are introduced for evaluation;

[0015] 3-1, training and analyzing the factorization machine FM result data;

[0016] 3-2, training and analyzing the deep neural network model result data;

[0017] 3-3, training and analyzing the deep factorization machine DeepFM model result data;

[0018] 3-4, analysis and comparison of the results of the three models.

[0019] Further, in step 1-1, the data set comes from the Cancer Cell Line Encyclopedia (CCLE), which provides researchers with genomic data and analysis based on multiple cell lines based on a large number of cancer genomic model genetic performance characteristics and pharmacological characteristics and visualization. From the CCLE database, more than 11000 pieces of drug anti-cancer sensitivity data containing 24 drugs and 504 cell lines were extracted.

[0020] Further, in step 1-2, the chemical structure of the 24 drugs in the data set described in step 1-1 is obtained from PubChem, which is an open database containing a large amount of chemical and physical properties, biological structure, toxicity, drug chemical structure and other data, maintained by the National Center for Biotechnology Information (NCBI) of the National Institutes of Health (NIH).

[0021] Further, in step 1-3, the variable ActArea(Activity Area) describing the active area of a drug on a cell line is used as an indicator of drug sensitivity, which can describe whether a cell line is sensitive to a drug. The initial data of Actarea is between 0 and 7.8, and the value is uniformly distributed around 0. The following formula is used to perform minimum-maximum normalization to make the value uniformly distributed around 0:

[0022]

[0023] wherein, ActArea is the original data, ActArea * is the normalized data, min(ActArea) and max(ActArea) are the minimum and maximum values of the original data, respectively;

[0024] If the drug cell lines with ActArea greater than 0.8 are regarded as cell lines sensitive to the drug, and the drug cell lines with ActArea less than negative 0.8 are regarded as cell lines resistant to the drug, then there are 2215 sensitive drug-cell line pairs, and the number of sensitive samples and resistant samples is basically the same;

[0025] In the embodiment of the present application, after the screening process, 2215 cell lines of CCLE are obtained.

[0026] Further, the step 2 is specifically operated as follows:

[0027] 2-1: Construct a factorization machine FM model, which is mainly used to solve the problem of how to combine features in the case of sparse data. In order to express the correlation information between various feature quantities, a second-order polynomial model is used, and the final data set obtained in step 1 is taken as an input layer. The factorization machine FM is used to process low-order features and extract the cross relationship between the low-order features.

[0028] 2-2: Construct a deep neural network DNN model, and use feature conversion to one-hot form to process discrete features. In order to avoid a large number of redundant network parameters, the features are divided into different regions, and then two fully linked layers are added to combine dense vectors, and then high-order feature combinations are extracted.

[0029] 2-3: Use the fusion mechanism of parallel structure to fuse the prediction results of the models constructed in steps 2-1 and 2-2 into a deep factorization machine DeepFM model.

[0030] Further, in step 2-1, a factorization machine FM model is constructed by using a binomial polynomial, and the correlation information between various feature quantities is expressed. The expression of the second-order polynomial model is as follows:

[0031]

[0032] wherein, y represents an output value, x i represents when the feature variable, w0 represents a constant term of the model, represents when the input feature variable x i is 0, the output value predicted by the model, w i represents a weight parameter of the model, x ix j denotes the feature variable x i and x j is the combined variable, w ij denotes the weight of the intersection part of the variables x i and x j , when i≠j, w ij denotes the degree of influence of the interaction between the variables x i and x j .

[0033] Since the factorization machine FM is a classification function model, after one-hot encoding, the data of the sample becomes very sparse, and the case of two variables x i and x j not being 0 rarely occurs, which will lead to the inability to obtain w ij through training. In order to find w ij , an auxiliary vector U i =(U i , U i1 , …, U i2 ) is introduced for each feature vector x ik , and U j =(U j , U j1 , …, U j2 ) is introduced for each feature vector x jk , and U i U j T is used to solve w ij .

[0034] The matrix composed of w ij can be represented as:

[0035]

[0036] where, denotes the output value of the matrix, U denotes the auxiliary vector of the feature vector, U T denotes the transposition of the row vector in the matrix U into a column vector, u n denotes the nth column in the matrix U, denotes the transpose of the nth column vector u n in the matrix U, i.e. the row vector form.

[0037] Finally, the model is trained using stochastic gradient descent SGD:

[0038]

[0039] where, is a parameter vector, φ represents a parameter, y(x) represents the output value, xi denotes the feature variable, w0 denotes the constant term of the model, w i denotes the weight parameter of the model, U i denotes the feature variable x i The final output value of the model will change according to different values of the parameters, and the low-order features are processed and the low-order feature cross relationship is extracted through the factor decomposition machine FM model framework.

[0040] Further, in step 2-2, a deep neural network is introduced to further extract high-order features. After dividing the input features of the deep neural network DNN into feature regions, two fully connected layers are added to combine the dense vectors and further extract high-order features.

[0041] Further, in step 2-3, a deep factor decomposition machine DeepFM model is obtained by using the parallel structure of the fusion mechanism.

[0042] The structure framework of the deep factor decomposition machine DeepFM is divided into two parts, one part is the structure layer of the factor decomposition machine FM, and the other part is the structure layer of the deep neural network, both of which share the same feature input. The factor decomposition machine FM and the deep neural network DNN are responsible for the extraction of low-order and high-order features respectively. The prediction result of the deep factor decomposition machine DeepFM can be expressed as:

[0043] Y DeepFM = sigmoid(Y FM +Y DNN ) (4)

[0044] Where Y DeepFM is the output result of the deep factor decomposition machine DeepFM model, Y FM is the output result of the factor decomposition machine FM, and Y DNN is the output result of the deep neural network. The outputs of the factor decomposition machine FM part and the deep neural network DNN part are weighted and summed, and then the result is mapped to (0, 1) through the Sigmoid function.

[0045] Further, step 3 uses a 5-fold cross-validation method to divide the training set and the validation set, and finally uses these data to obtain the sensitivity of the anticancer drug predicted by the deep factor decomposition machine DeepFM model. The 5-fold cross-validation is to divide the data into 5 parts, with the training set accounting for 80% and the validation set accounting for 20%. During the entire cross-validation process, the training set and the validation set are replaced in turn, so that each part is used as both the training set and the validation set. The final result is obtained by averaging the results of each fold of the 5-fold cross-validation.

[0046] A series of training results are obtained through 5-fold cross-validation, which are analyzed and evaluated in the following three ways:

[0047] (I), the epoch-pearson linear analysis diagram of three models of factor decomposition machine FM, deep neural network DNN and deep factor decomposition machine DeepFM;

[0048] (II), the drug category and its Pearson correlation coefficient linear analysis diagram;

[0049] (III), epoch-pearson linear contrast diagram analysis and drug-Pearson columnar contrast diagram analysis of three models;

[0050] In the linear analysis diagram of (I), the abscissa is the epoch parameter 35 training times, and the ordinate is the Pearson correlation coefficient, epoch refers to when a complete data set passes through a neural network and returns once, this process is called an epoch, all training samples in the neural network have been forward propagated and backward propagated once, and an epoch is a process of training all training samples once.

[0051] Pearson correlation coefficient refers to when two variables are normal continuous variables and have a linear relationship, the correlation between the two variables is represented by the product difference correlation coefficient, Pearson correlation coefficient is used to measure whether two data sets are on a line, and is used to measure the linear relationship between distance variables, the greater the Pearson correlation coefficient, the stronger the correlation, the closer the Pearson correlation coefficient to 1, the stronger the correlation, and the closer the Pearson correlation coefficient to 0, the weaker the correlation, and the calculation formula is as follows:

[0052]

[0053] Where r represents the Pearson correlation coefficient between two variables (x and y), x i and y i are the i-dimensional features of two nodes x and y, and N is the number of feature dimensions.

[0054] Further, the epoch-pearson linear contrast diagram and the drug-Pearson columnar contrast diagram of the three models are analyzed in step 3-4, and the result shows that the anti-cancer drug sensitivity prediction calculation model constructed by the deep factor decomposition machine DeepFM is more accurate and stable.

[0055] Compared with the prior art, the beneficial effects of the present application are:

[0056] 1. The drug low-order feature combination is modeled alone, and high-order feature combinations are fused, a model of fusion of a deep neural network DNN and a factorization machine FM is constructed, drug feature structures are better extracted, and the model is more accurate and has higher precision.

[0057] 2. The application mainly studies the relationship between gene expression data and anticancer drug sensitivity, establishes a multiple network model, considers the relationship between anticancer drugs, the relationship between cell lines, the relationship between genes and anticancer drug sensitivity, and establishes a model with as high precision as possible on the basis of these relationships. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 It is a flowchart of the application;

[0059] Figure 2 It is a full-link diagram of the deep neural network DNN model in the embodiment;

[0060] Figure 3 It is a diagram for dividing input features of the deep neural network DNN model in the embodiment;

[0061] Figure 4 It is a diagram of the overall structure of the deep neural network DNN model after adding a link layer in the embodiment;

[0062] Figure 5 It is a schematic diagram of the deep factorization machine DeepFM model in the embodiment;

[0063] Figure 6 It is an epoch-pearson linear analysis diagram of the factorization machine FM model in the embodiment;

[0064] Figure 7 It is a drug-pearson linear analysis diagram of the factorization machine FM model in the embodiment;

[0065] Figure 8 It is an epoch-pearson linear analysis diagram of the deep neural network DNN model in the embodiment;

[0066] Figure 9 It is a drug-pearson linear analysis diagram of the deep neural network DNN model in the embodiment;

[0067] Figure 10 It is an epoch-pearson linear analysis diagram of the deep factorization machine DeepFM model in the embodiment;

[0068] Figure 11 It is a drug-pearson linear analysis diagram of the deep factorization machine DeepFM model in the embodiment;

[0069] Figure 12 Epoch-Pearson linear comparison analysis diagram of the three models in the embodiment

[0070] Figure 13 It is a drug-Pearson bar comparative analysis chart of the three models in the embodiment.

[0071] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0072] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0073] like Figures 1-6 As shown, the present application provides a method for predicting anticancer drug sensitivity based on a deep factor decomposition machine, comprising the following steps:

[0074] Step 1: Build and preprocess the dataset

[0075] Step 1-1 Obtain a cell gene expression dataset: The dataset comes from the Cancer Cell Line Encyclopedia (CCLE). This project provides researchers with genomic data and analysis based on multiple cell lines, based on the genetic expression characteristics, pharmacological characterization, and visualization of a large number of cancer genome composition models. The dataset was downloaded from the website containing cell genetic data (https: / / portals.broadinstitute.org / ccle / data) and contains drug anticancer sensitivity data for 24 drugs and 504 cell lines, totaling more than 11,000 data points.

[0076] Step 1-2: Obtain drug chemical structure data: Obtain the chemical structures of the 24 drugs in the data set described in step 1-1 from PubChem.

[0077] PubChem is an open database containing a large number of chemical physical properties, biological structures, toxicity, drug chemistry structure and other data, maintained by the National Center for Biotechnology Information (NCBI) of the National Institutes of Health (NIH), the chemical structure of 24 drugs in step 1-1 is downloaded from the PubChem website (https: / / pubchem.ncbi.nlm.nih.gov / ) in the form of sdf (Structure Data File), and then the PaDEL-Descriptor software (an open source chemical information processing software designed to calculate the molecular descriptors of compounds, which can be downloaded from http: / / padel-descriptor.winsite.com / ) is used to convert the sdf file into the corresponding fingerprint file (a file format representing the molecular structure or characteristics).

[0078] Step 1-3 normalizes the data set obtained in step 1-1 to determine the final data set: the variable ActArea (Activity Area) describing the active area of a certain drug on a cell line is used as an indicator of drug sensitivity, which can describe whether the cell line is sensitive to the drug. The initial data of Actarea is between 0 and 7.8, and the following formula is used to perform minimum-maximum normalization processing to make the value uniformly distributed around 0:

[0079]

[0080] Wherein, ActArea is the original data, ActArea * is the normalized data, min(ActArea) and max(ActArea) are the minimum and maximum values of the original data, respectively;

[0081] The drug cell line with ActArea greater than 0.8 is considered as a cell line sensitive to the drug, and the drug cell line with ActArea less than negative 0.8 is considered as a cell line resistant to the drug, so there are 2215 sensitive drug-cell line pairs, and the number of sensitive samples and resistant samples is basically the same.

[0082] In the embodiment of the present application, after the screening treatment, 2215 cell lines of CCLE are obtained.

[0083] Step 2, construct a DeepFM model, the main idea of the model is to construct a factorization machine model, construct a deep neural network model, and fuse the factorization machine and the deep neural network to construct a deep factorization machine model.

[0084] In step 2-1, a factorization machine FM model is constructed using a binomial polynomial to express the correlation information between various feature quantities. The expression of this second-order polynomial model is:

[0085]

[0086] Among them, y represents the output value, w0 represents the constant term of the model, which means that when the input feature variable x i When both are 0, the output value predicted by the model, w i Represents the weight parameter of the model, x i x j Represents the characteristic variable x i and x j The combined variable, w ij Represents the variable x i and x j The weight of the cross part, when i≠j, w ij Represents the variable x i and x j The degree of interaction between them.

[0087] Since the factorization machine FM is a classification function model, after one-hot encoding, the sample data becomes very sparse, and there are very few cases where two variables x i and x j If it is not 0, it will result in the inability to obtain w through training ij , in order to find w ij , for each eigenvector x i Introducing auxiliary vector U i =(U i1 , U i2 ,…,U ik ), similarly for each eigenvector x j Introducing U j =(U j1 , U j2 ,…,U jk ), use U i U j T Solve for w ij .

[0088] w ij The matrix can be expressed as:

[0089]

[0090] in, Represents the output value of the matrix, U represents the auxiliary vector of the eigenvector, U T Indicates transposing the row vector in matrix U into a column vector, u nrepresents the nth column in the matrix U, Represents the nth column vector u in the matrix U n The transpose of , that is, the row vector form.

[0091] Finally, we use stochastic gradient descent (SGD) to train the model:

[0092]

[0093] in, is a parameter vector, Represents a parameter, y(x) represents the output value, x i represents the characteristic variable, w0 represents the constant term of the model, w i Represents the weight parameter of the model, U i Represents the characteristic variable x i The auxiliary vector of ,depends on the different values ​​of the parameters, the final output value of the model will also change.,Through the factorization machine FM model framework, low-order features are processed and,the cross-relationships of low-order features are extracted.

[0094] Step 2-2, build a deep neural network DNN model: Since the factor decomposition model built in step 2-1 cannot extract the high-order features of the drug, a deep neural network is introduced to extract the high-order features. The full-link structure diagram of the deep neural network DNN is shown in the attached figure. Figure 2 As shown in the attached Figure 2 It can be seen that the number of full-link parameters is too large, which makes the model prone to overfitting. Therefore, in this experiment, the deep neural network DNN input features are divided into regions, and the structure diagram is shown in the attached figure. Figure 3 shown.

[0095] After dividing the feature regions, two fully connected layers are added to combine dense vectors to extract high-order features, such as the following: Figure 4 As shown. Figure 4 It can be seen that although the input variables are divided into feature regions and two fully connected layers are added, the combination of low-order and high-order features is still implicitly reflected in the hidden layer. In order to better extract the characteristics of high-order and low-order features, the deep factor decomposition machine DeepFM is introduced.

[0096] Step 2-3 uses the fusion mechanism of the parallel structure to fuse the prediction results of the two models constructed in step 2-1 and step 2-3 to obtain the DeepFM model. The model principle is as shown in the attached figure. Figure 5 shown.

[0097] The structural framework of DeepFM is divided into two parts, one part is the structural layer of factorization machine FM, and the other part is the structural layer of deep neural network, both of which share the same feature input, and the factorization machine FM and the deep neural network DNN are respectively responsible for the extraction of low-order and high-order features, and the prediction result of DeepFM can be expressed as:

[0098] Y DeepFM = sigmoid(Y FM + Y DNN ) (4)

[0099] Wherein, Y DeepFM is the output result of the DeepFM model, Y FM is the output result of the factorization machine FM, and Y DNN is the output result of the deep neural network, the outputs of the factorization machine FM part and the deep neural network DNN part are weighted and summed, and then the result is mapped to (0, 1) through the Sigmoid function.

[0100] Step 3, training each stage model result and analysis and evaluation

[0101] In the present application, the 5-fold cross-validation method is used to divide the training set and the validation set, and finally the data is used to obtain the sensitivity of the anticancer drug predicted by the DeepFM model. 5-fold cross-validation is to divide the data into 5 parts, the training set accounts for 80%, and the validation set accounts for 20%. In the whole process of cross-validation, the training set and the validation set are replaced in turn, so that each part is used as both the training set and the validation set. The final result is obtained by averaging the results of each fold of 5-fold cross-validation.

[0102] A series of training results are obtained through 5-fold cross-validation, which are analyzed and evaluated by the following three ways:

[0103] (I), the epoch-pearson linear analysis diagram of factorization machine FM, deep neural network DNN and DeepFM three models

[0104] (II), drug species and its Pearson correlation coefficient linear analysis diagram

[0105] (III), epoch-pearson linear contrast diagram analysis and Drug-Pearson column contrast diagram analysis of three kinds of models

[0106] In the linear analysis graph (I), the horizontal axis represents the number of training cycles (epoch parameter, 35), and the vertical axis represents the Pearson correlation coefficient. An epoch is when a complete dataset passes through the neural network once and returns once. All training samples undergo a forward and backward propagation in the neural network. An epoch is the process of training all training samples once.

[0107] The Pearson correlation coefficient refers to the product-moment correlation coefficient that expresses the degree of correlation between two variables when both are normal continuous variables and there is a linear relationship between them. The Pearson correlation coefficient is used to measure whether two data sets are on the same line and to measure the linear relationship between interval variables. The larger the Pearson correlation coefficient, the stronger the correlation. The closer the Pearson correlation coefficient is to 1, the stronger the correlation is, and the closer the Pearson correlation coefficient is to 0, the weaker the correlation is. Its calculation formula is as follows:

[0108]

[0109] Where r represents the Pearson correlation coefficient between two variables (x and y), x i and y i are the i-th dimension features of the two nodes x and y respectively, and N is the number of feature dimensions.

[0110] Step 3-1: Train the FM model and analyze the result data:

[0111] 3-1.1、Epoch-Pearson linear analysis diagram is as attached Figure 6 shown.

[0112] By the attached Figure 6 It can be seen that the overall Pearson correlation coefficients are all above 0.8, among which the relatively high Pearson correlation coefficient can reach 0.9, and most of the Pearson correlation coefficients fluctuate around 0.88, indicating that the predicted value and the true value of the 5-fold data results of the factor decomposition machine FM model are still highly correlated, indicating that the accuracy of the model is relatively reliable. Although the Pearson correlation coefficient values ​​of the 34th group of experimental data show a sharp drop compared with those of the same group, this is because there are certain data errors in the experimental data entry or the experiment has a certain degree of randomness. The data deviation in a very small range does not affect the overall presentation of the experimental data results.

[0113] 3-1.2. The linear analysis chart of drug types and their Pearson correlation coefficients is shown in the attached figure. Figure 7 shown.

[0114] By the attached Figure 7It can be seen that the Pearson correlation coefficient of each drug is above 0.3, most of which can even reach 0.7, basically fluctuating around 0.6, the Pearson correlation coefficient of 4 drugs is even close to 0.9, the Pearson correlation coefficient of 19 drugs is stable between 0.5 and 0.8, but the Pearson correlation coefficient of one drug (LBW242) is lower than that of other drugs in the same group (less than 0.4), and the prediction result of this drug is not very ideal, and the accuracy of the anti-cancer drug sensitivity prediction model needs to be further improved.

[0115] Step 3-2 Training deep neural network model result data and analysis:

[0116] 3-2.1, epoch-pearson linear analysis chart as shown in the accompanying Figure 8 .

[0117] From the accompanying Figure 8 It can be seen that the Pearson correlation coefficient of the training set and the validation set is above 0.86, the higher Pearson correlation coefficient of the training set can reach 0.98, and the higher Pearson correlation coefficient of the validation set can reach 0.92, the Pearson correlation coefficient of the training set is mostly around 0.94, and the Pearson correlation coefficient of the validation set is mostly around 0.9, indicating that the predicted value and the true value of the 5-fold data result of the deep neural network DNN model have high correlation, and the accuracy of the model is reliable.

[0118] 3-2.2, drug type and its Pearson correlation coefficient linear analysis chart as shown in the accompanying Figure 9 .

[0119] From the accompanying Figure 9 It can be seen that the Pearson correlation coefficient of each drug is above 0.3, most of which can even reach 0.7, basically fluctuating around 0.6, the Pearson correlation coefficient of 4 drugs is even close to 0.9, the Pearson correlation coefficient of 19 drugs is stable between 0.5 and 0.8, but the Pearson correlation coefficient of one drug (LBW242) is lower than that of other drugs in the same group (less than 0.4), and the prediction result of this drug is not very ideal, and the accuracy of the anti-cancer drug sensitivity prediction model needs to be further improved.

[0120] Step 3-3 Training deep factor decomposition machine DeepFM model result data and analysis:

[0121] 3-3.1, epoch-pearson linear analysis chart as shown in the accompanyingFigure 10 As shown in the following table.

[0122] As shown in the following table. Figure 10 It can be seen that the Pearson correlation coefficients of the training set and the validation set are all above 0.8, the Pearson correlation coefficients of the training set can reach 0.95, the Pearson correlation coefficients of the validation set can reach 0.92, most of the Pearson correlation coefficients of the training set show an upward trend and fluctuate around 0.92, most of the Pearson correlation coefficients of the validation set fluctuate around 0.9 and show an upward trend, which indicates that the predicted values and the true values of the 5-fold data results of the DeepFM model have high correlation, although the experimental data results of the 16th group show small amplitude unstable fluctuations compared with the Pearson correlation coefficient values of the same group, the overall trend of the experimental results is not affected.

[0123] 3-3.2, drug categories and their Pearson correlation coefficient linear analysis chart as shown in the following table. Figure 11 As shown in the following table.

[0124] As shown in the following table. Figure 11 It can be seen that the Pearson correlation coefficients of each drug are all above 0.3, most of them can even reach above 0.7, and basically fluctuate around 0.7, the Pearson correlation coefficients of 4 kinds of drugs are even close to 0.9, the Pearson correlation coefficients of 19 kinds of drugs are stable and fluctuate between 0.5 and 0.8, but the Pearson correlation coefficient of one drug (LBW242) is lower than that of other drugs in the same group (0.3), the prediction result of this drug is not very ideal, and the accuracy of the anti-cancer drug sensitivity prediction model needs to be further improved.

[0125] Step 3-4, the results of the three models are combined and analyzed:

[0126] 3-4.1, the factorization machine FM model and the deep neural network DNN model are combined through a parallel structure to form a deep factorization machine DeepFM model, the epoch-pearson linear comparison chart of the three models is shown in the following table. Figure 12 As shown in the following table. Figure 12 It can be seen from the following table that the results of the deep factorization machine DeepFM model have higher accuracy compared with the results of the factorization machine FM model, and the "jumping fluctuation" of individual data is reduced, and the results of the deep factorization machine DeepFM model have reduced the difference between the Pearson correlation coefficients of the training set and the validation set compared with the results of the deep neural network DNN model, which improves the accuracy and stability of the prediction to a certain extent.

[0127] 3-4.2, according to the factor decomposition machine FM, the deep neural network DNN and the deep factor decomposition machine DeepFM model, the Pearson correlation coefficient linear relationship between the predicted value and the true value of the drug is divided according to each drug as shown in the following figure: Figure 13 As can be seen from the following figure: Figure 13 The linear curve of the Pearson correlation coefficient obtained by the deep factor decomposition machine DeepFM model on the overall data is basically above the factor decomposition machine FM and the deep neural network DNN model for the Y-axis value of each drug, that is, the absolute value of the correlation between the predicted value and the true value of the deep factor decomposition machine DeepFM model for almost every drug is greater than the absolute value of the correlation obtained by the factor decomposition machine FM and the deep neural network DNN model, and the anti-cancer drug sensitivity prediction calculation model constructed by the deep factor decomposition machine DeepFM is more accurate and stable.

[0128] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the statement "includes a" does not exclude the existence of other identical elements in the process, device, article or method including the element.

[0129] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or directly or indirectly applied to other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for predicting anticancer drug sensitivity based on a deep factorization machine, characterized in that: The following steps are involved: Step 1: Build dataset and preprocess; 1-1, obtain gene expression data of cells to build a dataset; 1-2, obtaining drug chemical structure data, obtaining the chemical structures of all drugs in the data set described in step 1-1 from PubChem; 1-3, using ActArea as an indicator of drug sensitivity, standardize and normalize the drug sensitivity data to determine the final data set; Step 2: Constructing a Deep Factorization Machine (DeepFM) model: The main idea of ​​the Deep Factorization Machine (DeepFM) is to construct a factorization machine (FM) model, a deep neural network (DNN) model, and integrate the factorization machine (FM) with the deep neural network (DNN) to construct a Deep Factorization Machine (DeepFM) model. 2-1: Construct a factorization machine (FM) model. This model is mainly used to solve the problem of how to combine features when data is sparse. In order to express the correlation information between various feature quantities, a second-order polynomial model is used here. The final dataset obtained in step 1 is used as the input layer. The deep neural network (DNN) model is used to process low-order features and extract the cross-relationships between low-order features. 2-2: Build a deep neural network (DNN) model, use feature conversion to process discrete features in one-hot form, divide the features into different regions to avoid a large number of redundant network parameters, then add two fully connected layers, combine dense vectors, and then extract high-order feature combinations; 2-3: Using the parallel structure fusion mechanism, the model prediction results constructed in step 2-1 and step 2-2 are fused into the Deep Factorization Machine DeepFM model; Step 3: Train and evaluate the model results at each stage. The anticancer drug sensitivity data obtained in step 1 were evenly divided into five parts and a 5-fold cross-validation was performed. In this stage, the Pearson correlation coefficient and epoch parameter were introduced to evaluate the model results at each stage. 3-1, training the factorization machine FM model result data and analyzing it; 3-2, training the deep neural network model result data and analyzing it; 3-3, training the DeepFM model result data and analyzing it; 3-4, analysis and comparison of the results of the three models; In step 2-1, a binomial polynomial is used to construct a factorization machine FM model to express the correlation information between various feature quantities. The expression of this second-order polynomial model is: （2） in, Indicates the output value, represents the feature vector, Represents the constant term of the model, which means when the input feature variable When both are 0, the output value predicted by the model is, represents the weight parameter of the model, Represents feature variables and The combined variables, Representing variables and The weight of the cross part, when hour, Representing variables and The degree of interaction between them; Since the factorization machine FM is a classification function model, after one-hot encoding, the sample data becomes very sparse, and two variables rarely appear. and If it is not 0, it will result in failure to obtain , in order to find , for each eigenvector Introducing auxiliary vectors , and similarly for each eigenvector Introduction ,use Solution ; The matrix can be expressed as: （3） in, represents the output value of the matrix, represents the auxiliary vector of the eigenvector, Indicates that the matrix The row vectors in are transposed into column vectors, Representation matrix The List, Representation matrix The column vectors The transpose of , i.e. the row vector form; Finally, we use stochastic gradient descent (SGD) to train the model: （4） in, is a parameter vector, Represents a parameter, Indicates the output value, represents the feature vector, represents the constant term of the model, represents the weight parameter of the model, Represents feature variables Auxiliary vector; Depending on the different values ​​of the parameters, the final output value of the model will also change. The factor decomposition machine FM model framework is used to process low-order features and extract low-order feature cross-relationships.

2. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 1, characterized in that: In step 1-1, the dataset comes from the Cancer Cell Line Encyclopedia (CCLE). This project provides researchers with genomic data and analysis based on multiple cell lines, based on the genetic expression characteristics, pharmacological characterization, and visualization of a large number of cancer genome composition models. Drug anticancer sensitivity data, totaling more than 11,000 data items, including 24 drugs and 504 cell lines, were extracted from the CCLE database.

3. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 2, characterized in that: In step 1-2, the chemical structures of the 24 drugs in the data set described in step 1-1 are obtained from PubChem. PubChem is an open database containing a large amount of chemical and physical properties, biological structures, toxicity, drug chemical structures and other data, maintained by the National Center for Biotechnology Information (NCBI) of the National Institutes of Health (NIH).

4. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 3, characterized in that: In steps 1-3, the variable ActArea, which describes the active area of ​​a certain drug on the cell line, is used as an indicator of drug sensitivity. It can describe whether the cell line is sensitive to the drug. The initial data of Actarea ranges from 0 to 7.

8. The following formula is used to perform minimum-maximum normalization processing on it so that its value is evenly distributed around 0: （1） Among them, ActArea is the original data, is the normalized data, and are the minimum and maximum values ​​of the original data respectively; If drug-treated cell lines with ActArea greater than 0.8 are considered drug-sensitive and drug-treated cell lines with ActArea less than -0.8 are considered drug-resistant, there are 2215 sensitive drug-cell line pairs, with an equal number of sensitive and resistant samples.

5. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 1, characterized in that: In step 2-2, a deep neural network is introduced. After dividing the input features of the deep neural network DNN into feature regions, two fully connected layers are added to extract high-order features.

6. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 1, characterized in that: In steps 2-3, the parallel structure of the fusion mechanism is used to obtain the DeepFM model; The structural framework of the Deep Factorization Machine (DeepFM) model is divided into two parts: one is the structural layer of the Factorization Machine (FM), and the other is the structural layer of the Deep Neural Network (DNN). Both share the same feature input. The Factorization Machine model and the Deep Neural Network (DNN) model are responsible for extracting low-order and high-order features, respectively. The prediction results of the Deep Factorization Machine (DeepFM) can be expressed as: （5） in, is the output of the DeepFM model, is the output of the factorization machine FM, It is the output result of the deep neural network. The output of the factorization machine FM part and the deep neural network DNN part are weighted summed, and then the result is mapped to (0,1) through the Sigmoid function.

7. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 1, characterized in that: Step 3 uses a 5-fold cross-validation method to divide the training set and validation set. Finally, this data is used to obtain the sensitivity of anticancer drugs predicted by the Deep Factorization Machine DeepFM model. The 5-fold cross-validation method is to divide the data into 5 parts, with the training set accounting for 80% and the validation set accounting for 20%. During the entire cross-validation process, the training set and validation set are alternated so that each part is used as both a training set and a validation set. The final result is obtained by taking the average of the results of each fold of the 5-fold cross-validation. After 5-fold cross-validation, a series of training results were obtained, which were analyzed and evaluated in the following three ways: (I) Epoch-Pearson linear analysis of the three models: factorization machine FM, deep neural network DNN and deep factorization machine DeepFM; (II), linear analysis diagram of drug types and their Pearson correlation coefficients; (III) Epoch-Pearson linear comparison chart analysis and Drug-Pearson bar chart analysis of the three models; Among them, the horizontal axis of the linear analysis graph (I) is the number of training times of the epoch parameter 35, and the vertical axis is the Pearson correlation coefficient. Epoch refers to when a complete data set passes through the neural network once and returns once. This process is called an epoch. All training samples undergo a forward propagation and a backward propagation in the neural network. An epoch is the process of training all training samples once. The Pearson correlation coefficient refers to the product-moment correlation coefficient when two variables are normal continuous variables and there is a linear relationship between them. The Pearson correlation coefficient is used to measure whether two data sets are on the same line and to measure the linear relationship between interval variables. The larger the Pearson correlation coefficient, the stronger the correlation. The closer the Pearson correlation coefficient is to 1, the stronger the correlation is. The closer the Pearson correlation coefficient is to 0, the weaker the correlation is. Its calculation formula is as follows: （6） Where r represents the Pearson correlation coefficient between two variables x and y. and Two nodes respectively and No. Dimensional features, N is the number of feature dimensions.

8. The method for predicting anticancer drug sensitivity based on deep factorization machine according to claim 1, characterized in that: In steps 3-4, the epoch-Pearson linear comparison chart and drug-Pearson histogram comparison chart of the three models were analyzed. The results show that the anticancer drug sensitivity prediction calculation model built by the deep factorization machine DeepFM is more accurate and stable.

Images