阅读说明：本技术 基于深度学习和基因表达数据的化合物肝毒性早期预测方法 (Compound hepatotoxicity wind agitation method for early prediction based on deep learning and gene expression data ) 是由 冯春来 陈恒巍 季薇 芮蒙杰 于 2019-06-24 设计创作，主要内容包括：本发明涉及计算机辅助药物筛选领域,具体地说是涉及一种基于深度学习和基因表达数据的化合物肝毒性的早期预测方法,其包括以下步骤：(1)化合物毒理学基因表达数据的挖掘与预处理；(2)肝毒性特征基因的挑选；(3)基于肝毒性特征基因的表达数据建立深度学习预测模型；(4)模型的参数优化与性能提升。本方法将药物基因组学与人工智能深度学习算法充分结合,克服了传统化合物肝毒性预测方法的局限性,实现了通过基因水平系统地对化合物迟发性肝毒性进行早期预测,从而为新药研发过程中的临床前毒性安全评估及临床合理用药提供一种高效、准确、快速的化合物肝毒性预测方法。(The present invention relates to area of computer aided drug screening fields, concretely relate to a kind of hepatotoxic method for early prediction of the compound based on deep learning and gene expression data comprising following steps: (1) excavation and pretreatment of compound toxicology gene expression data；(2) hepatotoxicity wind agitation characterizing gene is selected；(3) the expression data based on hepatotoxicity wind agitation characterizing gene establish deep learning prediction model；(4) parameter optimization and performance boost of model.This method is sufficiently combined pharmacogenomics with artificial intelligence deep learning algorithm, overcome the limitation of conventional compounds hepatotoxicity wind agitation prediction technique, realize and early prediction systematically carried out to compound Delayed onset hepatotoxicity wind agitation by gene level, thus for during new drug development preclinical toxicity security evaluation and clinical rational drug use a kind of efficient, accurately and rapidly compound hepatotoxicity wind agitation prediction technique is provided.)

1. a kind of compound hepatotoxicity wind agitation method for early prediction based on deep learning and gene expression data, which is characterized in that packet Include following steps:

Step 1:, will by being excavated from public Service functions to the gene expression data after compound effects Obtained initial data carries out data cleansing and standardization, constructs modeling data sample；

Step 2: by analysis of gene differential expression and feature weight calculating sifting hepatotoxicity wind agitation characterizing gene, as final mask Sample characteristics；

Step 3: prediction model is constructed based on deep learning algorithm, the expression data for the hepatotoxicity wind agitation characterizing gene that screening is obtained Training and study for model；

Step 4: the key parameter of model is optimized by grid-search algorithms and cross validation, improves the prediction of model Performance.

2. a kind of compound hepatotoxicity wind agitation early prediction side based on deep learning and gene expression data as described in claim 1 Method, which is characterized in that

The step 1 includes the following steps:

1.1 by from collecting gene after the compound effects that same gene chip measures in public Service functions Express data；

1.2 gene expression data samples according to hepatotoxicity wind agitation lesion rank are divided into Severe by 5 point-scores, and (serious, extent of disease is [75%, 100%]), Moderately Severe (moderate is serious, extent of disease be [50%, 75%)), Moderate (in Degree, extent of disease be [25%, 50%)), Slight (it is slight, extent of disease be [1%, 25%)), Minimal (slight, lesion Range be [0%, 1%))；

1.3 any toxicity level will occur for Severe, Moderately Severe, Moderate and Slight toxicity Compound label be positive compound, the medication group sample under all time points, all dosage is hepatotoxicity wind agitation positive sample This；The gene expression data sample and all control group samples of the compound of toxicity lesion rank Minimal are as hepatotoxicity wind agitation yin Property sample；

1.4 are based on R language and Bioconductor R packet, to the hepatotoxicity wind agitation positive sample of collection and the gene expression of negative sample Data are pre-processed；Wherein, it is filled by missing values and invalid value of the Impute packet to gene expression, passes through Limma Packet is standardized gene expression data；

1.5 according to genetic chip comment file, by the probe I D of standardized gene expression data and corresponding gene Symbol is matched one by one；

All modeling sample data are randomly divided into training set and test set according to 80%:20% by 1.6, and training set is for model Training study, test set are used for the Performance Evaluation of model.

3. a kind of compound hepatotoxicity wind agitation early prediction side based on deep learning and gene expression data as claimed in claim 2 Method, which is characterized in that in step 1.1, the public Service functions include ArrayExpress, Gene Expression and Omnibus (GEO).

4. a kind of compound hepatotoxicity wind agitation early prediction side based on deep learning and gene expression data as described in claim 1 Method, which is characterized in that

The step 2 includes the following steps:

The 2.1 limma R packets based on bayesian algorithm carry out Differential expression analysis to pretreated gene expression data, choose Wherein the absolute value of differential expression multiple is greater than or equal to gene of the 2 and adjust-P value less than or equal to 0.05 as feature Gene；

2.2, which further calculate simultaneously keeping characteristics weighted value by weight of the deep learning algorithm to characterizing gene, is greater than 0.1 The characterizing gene that is constructed as final mask of gene, i.e. hepatotoxicity wind agitation characterizing gene.

5. a kind of compound hepatotoxicity wind agitation early prediction side based on deep learning and gene expression data as described in claim 1 Method, which is characterized in that

The step 3 includes the following steps:

3.1 models choose sequential (Sequential) model interface, and model structure includes input layer (Input layer), hides Layer (Hidden layer) and output layer (Output layer), wherein hidden layer includes full articulamentum (Dense layer) With Dropout layers (Dropout layer)；

The input of 3.2 models is gene expression data, wherein each characterizing gene is used as a node of input layer；

The output of 3.3 models is two classification results 0 and 1, wherein 0 represents hepatotoxicity wind agitation feminine gender, 1 represents the hepatotoxicity wind agitation positive；

In the hidden layer of 3.4 models, input layer is activated by Rectified Linear Unit (ReLU) activation primitive Value is passed to full articulamentum, the formula of the activation primitive in turn are as follows:

Y=ReLU (Wx+b)

Wherein, x is the value of input data, and y is the value after data activation, and W is weight matrix, and b is deviation；

In the output layer of 3.5 models, the value of hidden layer is activated to spread out of as finally defeated by Sigmoid activation primitive Out as a result, the formula of the activation primitive are as follows:

Z=sigmoid (W ' y+b')

Wherein, y is the value after the activation of hidden layer outflow, and z is model output as a result, W ' is transposition weight matrix, and b ' is transposition Deviation；

In the training process of 3.6 models, configured using learning process of the compile module to model, parameter is set respectively Be set to: optimizer (optimizer) is set as Root Mean Square prop (RMSprop), and index list (metrics) is set For accuracy and loss function (loss function) it is set as binary_crossentropy, wherein the loss function Calculation formula are as follows:

Wherein, L_H(x, z) is the difference size (losing) of predicted value and true value, and x is the corresponding true value of sample, and z is sample Corresponding predicted value, d are epoch number；

3.7 step 2 is handled after finally obtained hepatotoxicity wind agitation characterizing gene expression data as the input of model, wherein 80% It is used to training pattern as training set, 20% is used as test set to be used to test model performance.

6. a kind of compound hepatotoxicity wind agitation early prediction side based on deep learning and gene expression data as described in claim 1 Method, which is characterized in that

The step 4 includes the following steps:

4.1 setting parameter optimization ranges, wherein epoch number is [10,50,100,200,500], and batch size is [10,32,64,128], learning rate are [0.01,0.001,0.00001], and dropout rate is [0,0.2,0.5], Node number is [50,100,300,500,1000]；

4.2 carry out optimizing to a model in 900 (5 × 4 × 3 × 3 × 5) of building by grid-search algorithms；

4.3 evaluate the estimated performance of model by 10 folding cross validation modes and evaluation index, and wherein performance evaluation refers to Mark includes: susceptibility (Sensitivity, SEN)；Specific (Specificity, SPE)；Accuracy (Accuracy, ACC)； Ma Xiusi related coefficient (Matthews correlation coefficient, MCC)；Area (the area under ROC curve Under the Receiver Operating Characteristic (ROC) curve, AUC)；Wherein, susceptibility, specificity Area shows the model prediction closer to 1 under 100%, Ma Xiusi related coefficient and ROC curve with accuracy It can be better；On the contrary, susceptibility, specificity and the accuracy area under 0, Ma Xiusi related coefficient and ROC curve more connect It is bordering on 0.5, shows that the model prediction performance is poorer；

Wherein, TP represents true positives；TN represents true negative；FP represents false positive；FN represents false negative.

Technical field

The present invention relates to area of computer aided drug screening fields, concretely relate to a kind of based on deep learning and gene The hepatotoxic method for early prediction of compound of data is expressed, is suitable for carrying out compound hepatotoxicity wind agitation according to gene expression data Early prediction.

Background technique

Use in medicament-induced hepatotoxicity is a key factor for causing new drug research failure and clinical medicine to remove city.According to statistics in new drug The ratio to lead to the failure in R&D process by drug candidate hepatotoxicity wind agitation is 37%, is led in clinical application because of use in medicament-induced hepatotoxicity Causing drug to remove the ratio in city is 18%, therefore, is predicted in medicament research and development early stage and clinical use use in medicament-induced hepatotoxicity It is of great significance for improving research and development success rate and the rational use of medicines.Since use in medicament-induced hepatotoxicity mechanism is complicated, how to improve The accuracy and applicability of use in medicament-induced hepatotoxicity prediction, especially still face great choose to the prediction of Delayed onset use in medicament-induced hepatotoxicity War.

Traditional use in medicament-induced hepatotoxicity prediction technique is mainly based upon inside and outside experimental model and carries out safely to the toxicity of drug Assessment, this method need a large amount of living animal to carry out experiment basis, have the shortcomings that experimental period is long, it is high to expend.With peace Entirely, the laws and regulations requirement of environmental protection, animal protection etc. is increasingly strict, and the market competition of globalization also requires drug development cycle big Width reduction, therefore how to overcome disadvantage present in conventional medicament hepatotoxicity wind agitation prediction technique, it develops efficient use in medicament-induced hepatotoxicity and comments The method of estimating is of great significance.With the development of information technology, research both at home and abroad begins trying to utilize machine learning and compound Structure feature establishes computation model to carry out use in medicament-induced hepatotoxicity prediction, but it often faces following problems: 1) vulnerable to single chemical combination The limitation of object structure predicts that often accuracy rate is lower for the hepatotoxicity wind agitation of the compound of structure diversity；2) prediction result lacks Biological significance, can not from biological effect mechanism system interpretation prediction result；3) can not use in medicament-induced hepatotoxicity to Delayed onset into Row early prediction；4) traditional machine learning method can not from big data automatic learning characteristic information, need a large amount of artificial Feature is selected.Therefore, for the limitation of above method, the present invention provides one kind to be based on deep learning and gene expression data Compound hepatotoxicity wind agitation prediction technique, can by gene level systematically to compound Delayed onset hepatotoxicity wind agitation carry out early stage it is pre- It surveys.

Summary of the invention

The present invention overcomes the shortcomings of the prior art, discloses a kind of based on deep learning algorithm (Deep Learning Algorithm, DL) and gene expression data the hepatotoxic method for early prediction of compound, this method is by pharmacogenomics It is sufficiently combined with artificial intelligence deep learning algorithm, overcomes the limitation of conventional compounds hepatotoxicity wind agitation prediction technique, realize Early prediction systematically is carried out to compound Delayed onset hepatotoxicity wind agitation by gene level, thus for the toxicity during new drug development Security evaluation and clinical rational drug use provide it is a kind of efficiently, accurately and rapidly compound hepatotoxicity wind agitation prediction technique.

The purpose of the present invention can pass through following technology path (Fig. 1) Lai Shixian:

1. a kind of compound hepatotoxicity wind agitation method for early prediction based on deep learning and gene expression data, feature exist In including the following steps:

Step 1: by being dug from public Service functions to the gene expression data after compound effects Obtained initial data is carried out data cleansing and standardization, constructs modeling data sample by pick；

Step 2: by analysis of gene differential expression and feature weight calculating sifting hepatotoxicity wind agitation characterizing gene, as final Model sample feature；

Step 3: prediction model is constructed based on deep learning algorithm, the expression for the hepatotoxicity wind agitation characterizing gene that screening is obtained Data are used for the training and study of model；

Step 4: the key parameter of model is optimized by grid-search algorithms and cross validation, improves model Estimated performance.

2. the step 1 includes the following steps:

1.1 from public Service functions by collecting after the compound effects that same gene chip measures Gene expression data；

1.2 gene expression data samples are divided into Severe (serious, extent of disease by 5 point-scores according to hepatotoxicity wind agitation lesion rank For [75%, 100%]), Moderately Severe (moderate is serious, extent of disease be [50%, 75%)), Moderate (in Degree, extent of disease be [25%, 50%)), Slight (it is slight, extent of disease be [1%, 25%)), Minimal (slight, lesion Range be [0%, 1%))；

1.3 any toxicity level will occur for Severe, Moderately Severe, Moderate and Slight The compound label of toxicity is positive compound, and the medication group sample under all time points, all dosage is hepatotoxicity wind agitation sun Property sample；The gene expression data sample and all control group samples of the compound of toxicity lesion rank Minimal are as liver poison Property negative sample.

1.4 are based on R language and Bioconductor R packet, to the hepatotoxicity wind agitation positive sample of collection and the gene of negative sample Expression data are pre-processed.Wherein, it is filled, is passed through by missing values and invalid value of the Impute packet to gene expression Limma packet is standardized gene expression data；

1.5 according to genetic chip comment file, by the probe I D of standardized gene expression data and corresponding gene Symbol is matched one by one；

All modeling sample data are randomly divided into training set and test set according to 80%:20% by 1.6, and training set is used for mould The training study of type, test set are used for the Performance Evaluation of model.

In step 1.1, the public Service functions include ArrayExpress, Gene Expression and Omnibus(GEO)。

The step 2 includes the following steps:

The 2.1 limma R packets based on bayesian algorithm carry out Differential expression analysis to pretreated gene expression data, The absolute value for choosing wherein differential expression multiple is greater than or equal to the gene conduct that 2 and adjust-P value is less than or equal to 0.05 Characterizing gene；

2.2 are further calculated by weight of the deep learning algorithm to characterizing gene and keeping characteristics weighted value is greater than The characterizing gene that 0.1 gene is constructed as final mask, i.e. hepatotoxicity wind agitation characterizing gene.

The step 3 includes the following steps:

3.1 models choose sequential (Sequential) model interface, model structure include input layer (Input layer), Hidden layer (Hidden layer) and output layer (Output layer), wherein hidden layer includes full articulamentum (Dense ) and Dropout layers (Dropout layer) layer；

The input of 3.2 models is gene expression data, wherein each characterizing gene is used as a node of input layer；

The output of 3.3 models is two classification results 0 and 1, wherein 0 represents hepatotoxicity wind agitation feminine gender, 1 represents the hepatotoxicity wind agitation positive；

In the hidden layer of 3.4 models, input is activated by Rectified Linear Unit (ReLU) activation primitive The value of layer is passed to full articulamentum, the formula of the activation primitive in turn are as follows:

Y=ReLU (Wx+b)

Wherein, x is the value of input data, and y is the value after data activation, and W is weight matrix, and b is deviation；

In the output layer of 3.5 models, it is final for activating the value of hidden layer to spread out of by Sigmoid activation primitive Output as a result, the activation primitive formula are as follows:

Z=sigmoid (W ' y+b ')

Wherein, y is the value after the activation of hidden layer outflow, and z is model output as a result, W ' is transposition weight matrix, and b ' is Transposition deviation；

It in the training process of 3.6 models, is configured using learning process of the compile module to model, parameter point It is not arranged are as follows: optimizer (optimizer) is set as Root Mean Square prop (RMSprop), index list (metrics) it is set as accuracy and loss function (loss function) is set as binary_crossentropy, wherein The calculation formula of the loss function are as follows:

Wherein, L_H(x, z) is the difference size (losing) of predicted value and true value, and x is the corresponding true value of sample, z For the corresponding predicted value of sample, d is epoch number.

3.7 step 2 is handled after finally obtained hepatotoxicity wind agitation characterizing gene expression data as the input of model, wherein 80% is used as training set to be used to training pattern, and 20% is used as test set to be used to test model performance；

The step 4 includes the following steps:

4.1 setting parameter optimization ranges, wherein epoch number is [10,50,100,200,500], batch size For [10,32,64,128], learning rate is [0.01,0.001,0.00001], dropout rate be [0,0.2, 0.5], node number is [50,100,300,500,1000]；

4.2 carry out optimizing to a model in 900 (5 × 4 × 3 × 3 × 5) of building by grid-search algorithms；

4.3 evaluate the estimated performance of model by 10 folding cross validation modes and evaluation index, and wherein performance is commented Valence index includes: susceptibility (Sensitivity, SEN)；Specific (Specificity, SPE)；Accuracy (Accuracy, ACC)；Ma Xiusi related coefficient (Matthews correlation coefficient, MCC)；Area (the under ROC curve Area under the Receiver Operating Characteristic (ROC) curve, AUC).Wherein, susceptibility, Specificity and accuracy area under 100%, Ma Xiusi related coefficient and ROC curve show the model closer to 1 Estimated performance is better；On the contrary, susceptibility, specificity and accuracy are below 0, Ma Xiusi related coefficient and ROC curve Product shows that the model prediction performance is poorer closer to 0.5.

Wherein, TP represents true positives；TN represents true negative；FP represents false positive；FN represents false negative；

Compared with prior art, as follows using advantages of the present invention:

1. early prediction can be carried out by gene level system, accurately to Delayed onset use in medicament-induced hepatotoxicity；

2. the prediction technique compared to tradition based on compound structure feature, this prediction technique can be with Accurate Prediction difference knot The hepatotoxicity wind agitation of structure compound has preferable robustness；

It, can be from big data certainly 3. the deep learning algorithm that this method uses has stronger automatic learning characteristic ability Dynamic study important feature information, avoids a large amount of characteristic of human nature and selects；

4. there is the compound hepatotoxicity wind agitation prediction model of this method building preferably estimated performance, forecasting accuracy to compare Have in traditional prediction technique and is significantly promoted；

It can be the preclinical of new drug development since it has the characteristics that early prediction 5. the applicability of this method is preferable Toxicity safety evaluation and clinical rational drug use provide theoretical foundation and technical support.

Detailed description of the invention:

Fig. 1 is total stream of the compound hepatotoxicity wind agitation method for early prediction the present invention is based on deep learning and gene expression data Cheng Tu；

Fig. 2 is the structural schematic diagram of the method for the present invention compound hepatotoxicity wind agitation prediction model；

Fig. 3 is the prediction result of the compound hepatotoxicity wind agitation method for early prediction based on deep learning and gene expression data Figure；

Fig. 4 is the estimated performance of the compound hepatotoxicity wind agitation method for early prediction based on deep learning and gene expression data Figure；

Specific embodiment:

In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawing to of the invention Technical solution is further described again.

A kind of particular technique side of the compound hepatotoxicity wind agitation method for early prediction based on deep learning and gene expression data Case is:

1. coming from Affymetrix Gene Chip Rat Genome by collecting from Array Express database Gene expression data under 87 compound effects of 230 2.0 chips measurement.The gene expression data sample collected according to Toxicity lesion rank is divided into (serious, extent of disease is [75%, 100%]) by 5 point-scores, and (moderate is tight by Moderately Severe Weight, extent of disease be [50%, 75%)), Moderate (moderate, extent of disease be [25%, 50%)), Slight (slight, disease Become range be [1%, 25%)), Minimal (it is slight, extent of disease for [0%, 1%)).To enable constructed model early stage Predict Delayed onset toxicity, and applicability with higher, this research will occur any toxicity level for Severe, Moderately Severe, Moderate and Slight toxicity compound label be positive compound, all time points, Medication group sample under all dosage is positive sample；The gene expression data of the compound of toxicity lesion rank Minimal Sample and all control group samples are as toxicity negative sample.By arranging, 988 cdna sample data are finally obtained, wherein Positive sample 496, negative sample 492.All modeling sample data are randomly divided into training set and survey according to 80%:20% Examination collection, training study of 790 sample datas as training set for model, 198 sample datas are used for mould as test set The Performance Evaluation of type.

2. it is clear to carry out a series of statistical datas to raw gene expression data based on R language and Bioconductor R packet It washes, constructs final gene expression profile as modeling data.Wherein, first by Impute packet to the missing values and nothing of gene expression Valid value is filled, and is standardized by Limma packet to gene expression data, further according to the annotation text of genetic chip Part matches probe I D with gene symbol, carries out differential expression point to gene expression data finally by DESeq2 packet Analysis, constructs final characterizing gene expression matrix.Wherein, first by the limma R packet based on bayesian algorithm to gene table Differential expression analysis is carried out up to data, the absolute value for choosing wherein differential expression multiple is greater than or equal to 2, adjust-P value and is less than Or the characterizing gene that the gene equal to 0.05 is obtained as preliminary screening, further preliminary screening is obtained by deep learning algorithm To characterizing gene carry out the gene that corresponding weight calculation and keeping characteristics weighted value be greater than 0.1, finally screen to obtain 1574 A characterizing gene is used for the building of final mask.

3. this programme is platform with the Python 3.6 of Anaconda5.1, by the depth based on Python and Theano Learning framework Keras builds deep learning prediction model.The model chooses sequential (Sequential) model interface, builds two points Class prediction model.Model structure includes input layer (Input layer), hidden layer (Hidden layer) and output layer (Output layer), wherein hidden layer includes full articulamentum (Dense layer) and Dropout layers (Dropout layer) (Fig. 2).In the hidden layer of model, input layer is activated by Rectified Linear Unit (ReLU) activation primitive Value is passed to full articulamentum, the formula of the activation primitive in turn are as follows:

Y=ReLU (Wx+b)

Wherein, x is the value of input data, and y is the value after data activation, and W is weight matrix, and b is deviation；

In the output layer of model, it is final for activating the value of hidden layer to spread out of by Sigmoid activation primitive Export the formula as a result, the activation primitive are as follows:

Z=sigmoid (W ' y+b ')

Wherein, y is the value after the activation of hidden layer outflow, and z is model output as a result, W ' is transposition weight matrix, and b ' is Transposition deviation；

In the training process of model, configured using learning process of the compile module to model, parameter is set respectively Be set to: optimizer (optimizer) is set as Root Mean Square prop (RMSprop), and index list (metrics) is set For accuracy and loss function (loss function) it is set as binary_crossentropy, wherein the loss function Calculation formula are as follows:

Wherein, L_H(x, z) is the difference size (losing) of predicted value and true value, and x is the corresponding true value of sample, z For the corresponding predicted value of sample, d is epoch number；

Then characterizing gene expression data processing obtained are as the input of model, wherein 80% is used to as training set Training pattern, 20% is used as test set to be used to test model performance.

4. the estimated performance of trained model is assessed by using the test set sample in modeling data, wherein Susceptibility (Sensitivity, SEN) is respectively adopted；Specific (Specificity, SPE)；Accuracy (Accuracy, ACC)； Ma Xiusi related coefficient (Matthews correlation coefficient, MCC)；Area (the area under ROC curve Under the Receiver Operating Characteristic curve, AUC) etc. relevant evaluations index carry out performance It judges.Further, by using grid-search algorithms and ten folding cross validations to 900 (5 × 4 × 3 × 3 × 5) a moulds of building The key parameter (epoch number, batch size, learning rate, dropout rate and node number) of type It optimizes, so that model be made to obtain optimum prediction performance.

Wherein, TP represents true positives；TN represents true negative；FP represents false positive；FN represents false negative；

Finally, by parameter optimization, the design parameter of optimal models be set as hidden layer be 2 layers and every layer be 50 section Point, dropout rate are set as 0.5 to avoid the over-fitting of model, and learning rate is that 0.001, batch size is 128, epoch number are 50.By test set to the Performance of optimal models, predictablity rate 97.1%, AUC It is 0.989, sensibility 97.4%, it is 0.942 that specificity, which is 96.8%, Ma Xiusi related coefficient, most compared to both at home and abroad The compound hepatotoxicity wind agitation prediction model that base learns in conventional machines, the model have preferably estimated performance (Fig. 3, Fig. 4).

Examples detailed above is only the specific embodiment of the present invention, is also being sent out its simple transformation, replacement etc. In bright protection scope.