JP2020025471A - Toxicity learning device, toxicity learning method, learned model, toxicity prediction device, and program - Google Patents

Abstract

To provide a novel method for predicting toxicity of a compound.SOLUTION: A toxicity learning device 10 comprises: an input part 11 which inputs expression data of a sample on which a compound is exposed and expression data of a control; a comparison part 13 which compares the expression data of the sample and the control by each predetermined gene; an encoding part 14 which encodes the expression data of a gene on the basis of a difference of the expression data; a label imparting part 15 which imparts a toxicity label of the compound to the encoded expression data; and a model learning part 16 which performs learning of a model in which toxicity of the compound is predicted from the expression data of the gene by using teacher data to which the label is imparted.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a toxicity learning device, a toxicity learning method, a learned model, a toxicity prediction device, and a program.

  While drug development involves enormous development costs and long research and development periods, the success rate of successfully launching a drug is by no means high. One of the causes is that human safety cannot be sufficiently ensured. If the toxicity can be determined at an early stage of drug development, the success rate of drug development will increase. Therefore, studies for predicting the safety of pharmaceuticals have been conducted.

  Non-Patent Document 1 discloses that hepatotoxicity evaluation of acetaminophen using a three-dimensional hepatocyte culture system is performed, focusing on a spheroid culture system using primary hepatocytes. In addition, software for predicting toxicity based on structure-activity relationship (QSAR) is also available (Non-Patent Document 2). In addition, a study has been published that seeks to determine the toxicity of a compound by deep learning, focusing on the structure of the compound (Non-Patent Document 3).

Masanobu Sano, "Study on in vitro and in vivo prediction evaluation system for hepatotoxicity and metabolites of pharmaceuticals in humans" YAKUGAKU ZASSHI 135 (11) 1273-1279 (2015) QSAR Toxicity Prediction Software Leadscope (R) Model Applier Genetic Toxicity Suite ITOCHU Techno Solutions Co., Ltd., 2017, Internet <URL: http://ls.ctc-g.co.jp/products/leadscope/files/LeadscopeModelApplier_2017.pdf> Andreas Mayr et al. `` Deep Tox: Toxicity Prediction using Deep Learning '' Frontiers in Environment Science, February 2016

  The present invention aims to provide a new method for predicting the toxicity of a compound.

  The present invention uses expression data from samples exposed to a compound to predict the toxicity of the compound. In the present invention, toxicity prediction is performed using a learned model that has been learned in advance using machine learning. The problem here is how to prepare the teacher data used for learning the model. Non-Patent Document 3 focuses on the structure of a compound, but the present invention employs a different method.

  The toxicity learning device of the present invention includes an input unit for inputting expression data of a sample exposed to a compound and expression data of a control, a comparing unit for comparing the sample and the expression data of the control for each predetermined gene, An encoding unit that encodes the expression data of the gene based on a difference between the data, a labeling unit that labels the encoded expression data with a toxicity label of the compound, and teacher data to which the label is added. And a model learning unit for learning a model for predicting the toxicity of the compound from the gene expression data using

  From the comparison of the expression data of the sample exposed to the compound and the control in this way, the difference in expression data for each gene is encoded, and the labeling is applied using knowledge about the toxicity of the compound to generate teacher data. Can be generated. A model for inferring the toxicity of the compound from the expression data can be generated using the teacher data.

  In the toxicity learning device of the present invention, the encoding unit ranks the genes based on the difference in the expression data, and the rank is “1” for a predetermined number of genes in an upper rank and “1” in a predetermined number of genes in a lower rank. “−1” and “0” may be added to other predetermined numbers. The predetermined number is an arbitrary number. For example, it may be defined by the percentage of the total number of genes, preferably 5%, more preferably 1-2%.

  By encoding the expression data with 1, 0, -1 in this manner, learning by machine learning becomes easy. Further, in the present invention, since the same predetermined number is set to a numerical value other than 0 for the upper order and the lower order, the range in which the numerical value other than 0 is taken in the upper and lower parts is symmetrical, so that the model can be appropriately learned.

In the toxicity learning device of the present invention, the comparing unit may obtain a plurality of comparison results by comparing the expression data for each gene by the following methods (i) to (iii).
(I) Take the difference between the expression data of the sample and the control.
(Ii) Take the ratio of the expression data of the sample and the control.
(Iii) Normalize the expression data of the sample and the control and take the difference.

  In machine learning research based on data from experiments (actual measurements) on living things, it is difficult to obtain a large amount of teacher data. In the present invention, it is possible to increase the data amount of the comparison result by making use of fluctuations caused by a method of acquiring expression data (for example, an analysis method using a microarray) and performing comparisons using different calculation methods (i) to (iii). Can do appropriate learning.

  In the toxicity learning device of the present invention, the labeling unit may use toxicity data read from a database storing the toxicity of the existing compound as the label. Further, the toxicity learning device of the present invention includes a label generation unit that reads data on the side effects of the compound from the database storing the side effects of the existing compound, and obtains data on the toxicity of the compound based on the read side effect data, The labeling unit may use the obtained toxicity data as the label.

  The toxicity learning method of the present invention comprises the steps of: inputting expression data of a sample exposed to a compound and expression data of a control; comparing the expression data of the sample and the control for each gene; Encoding the expression data of the gene, generating the teacher data by labeling the encoded expression data with the toxicity of the compound, and using the teacher data to express the gene. Learning a model for predicting the toxicity of the compound from the data. The toxicity learning method of the present invention may have various features of the toxicity learning device described above.

  In the toxicity learning method of the present invention, in the inputting step, expression data of a plurality of samples to which a known compound has been exposed are input, and in the comparing step, the expression data of the plurality of samples and the control are expressed by a gene. And a plurality of comparison results may be obtained for the one compound. In experiments on living organisms, it is usual to take an average of a plurality of actually measured data in order to reduce fluctuations caused by the experiment.However, according to the configuration of the present invention, a plurality of actually measured data are compared with independent experiment data. Considering this, the data amount of the teacher data can be increased.

  The toxicity prediction device of the present invention is a device for inferring the toxicity of a compound using a learned model learned by the above-described toxicity learning method, and an input unit for inputting expression data of a sample exposed to an unknown compound. A comparing unit that compares the expression data of the sample with the expression data of the control for each predetermined gene; an encoding unit that encodes the expression data of the gene based on a difference between the expression data; and the encoding unit. An inference unit that infers the toxicity of the compound by applying the obtained expression data to the learned model, and an output unit that outputs an inference result by the inference unit. With this configuration, the toxicity of the compound can be predicted using the expression data to which the compound has been exposed.

  The trained model of the present invention is a trained model for operating a computer so as to output a value quantifying the toxicity of a compound based on expression data upon exposure of the compound, and is an input of a neural network. In the layer, expression data encoded based on the difference between the expression data with the control is input, and based on the input encoded data, perform a calculation based on the learned weighting coefficients of the neural network, and from the output layer The computer is operated to output a value quantifying the toxicity of the compound.

  According to the present invention, it is possible to generate teacher data for learning a model for inferring the toxicity of the compound from the expression data of the sample exposed to the compound in this way.

(A) is a figure which shows the learning stage in the whole framework of hepatotoxicity discrimination. (B) It is a figure which shows the inference stage in the whole framework of hepatotoxicity discrimination. It is a figure showing composition of a toxicity learning device of a 1st embodiment. It is a figure showing an example of the result of having compared the expression level of a sample and a control. It is a figure showing the example which compared expression data for every gene. FIG. 7 is a diagram showing an example in which expression data is encoded according to the order given to genes. (A) It is a figure which shows the example which attached the label to each compound by the label provision part. (B) It is a figure which shows the data which encoded the expression data. It is a figure showing operation of a toxicity learning device of a 1st embodiment. It is a figure showing the composition of the toxicity prediction device of a 1st embodiment. It is a figure showing composition of a toxicity learning device of a 2nd embodiment. (A) is a diagram showing an example of data stored in a side effect DB. (B) It is a figure which shows the example which processed the data memorize | stored in the side effect DB.

  Hereinafter, a toxicity learning device and a toxicity prediction device according to an embodiment of the present invention will be described with reference to the drawings. In the embodiment described below, a toxicity prediction device that discriminates hepatotoxicity of an unknown compound and a toxicity learning device that learns a model used in the toxicity prediction device will be described as examples. Note that the toxicity learning device and the toxicity prediction device of the present invention can be applied to a toxicity prediction device for determining toxicity other than hepatotoxicity.

(First Embodiment)
FIG. 1 is a diagram showing an overall framework of hepatotoxicity discrimination by the toxicity learning device 10 and the toxicity prediction device 20 of the present embodiment. As shown in FIG. 1A, in the learning stage, the toxicology learning device 10 learns a model using a large number of teacher data, and generates a learned model. Here, there is a problem as to what should be used as teacher data. In the inference stage, as shown in FIG. 1B, the toxicity prediction device 20 determines the hepatotoxicity of the new compound using the learned model. In the present embodiment, hepatotoxicity is determined in three stages of “Most”, “Less”, and “Non”. “Most” indicates the highest hepatotoxicity, and “Non” indicates no hepatotoxicity.

  Note that a neural network model is used as the learned model. The number of layers of the structure of the neural network and whether or not to provide a convolutional layer and a pooling layer are arbitrary. However, according to experiments by the inventors, inference can be performed with high accuracy by using a convolutional neural network. It turns out that the model can be constructed.

  FIG. 2 is a diagram illustrating a configuration of the toxicity learning device 10 according to the first embodiment. The toxicity learning device 10 includes an input unit 11 for inputting expression data, a teacher data generating unit 12 for generating teacher data from the input data, and a model learning unit 16 for learning a model using the teacher data. are doing.

  The input unit 11 receives input of expression data of a sample to which the compound has been exposed and expression data of a control. Here, the compound to which the expression data of the sample is exposed is a known compound. The expression data may be gene expression data or protein expression data. In the present embodiment, the expression data is data obtained by a microarray.

The teacher data generation unit 12 includes a comparison unit 13, an encoding unit 14, and a label assignment unit 15. The comparing unit 13 has a function of comparing the expression data of the sample with the expression data of the control. The comparing unit 13 obtains a plurality of comparison results by comparing the expression data for each gene by the following methods (i) to (iii).
(I) Take the difference between the expression data of the sample and the expression data of the control.
(Ii) Take the ratio between the expression data of the sample and the expression data of the control.
(Iii) Normalize the expression data of the sample and the expression data of the control, and then take the difference.

  As described above, by performing the comparison using the three methods, the data amount of the comparison result between the sample and the control becomes three times the data amount of the original sample. The toxicology learning device 10 can use three times the amount of input data for learning.

  FIG. 3 is a diagram showing an example of the result of comparing the expression levels of a sample and a control. “Diff” is the result of the difference between the expression data ((i) above), “Ratio” is the result of the ratio of the expression data ((ii)), and “CR” (meaning Change Ratio) "This is the result of taking the difference after normalization ((iii) above).

“Probe” described on the vertical axis indicates the probe number of the microphone array, and Drug on the horizontal axis indicates the compound exposed to the sample. For example, in the sample exposed to “Drug1”, the difference of Probe1 is 10, the difference of Probe2 is 1253, the difference of Probe3 is 324, and so on.

  In the example shown in FIG. 3, “Probe” on the vertical axis indicates a probe of a microarray from which expression data has been acquired, but the comparison unit 13 of the present embodiment compares a sample with a control for each gene. That is, the comparison result for each “Probe” is changed to the comparison result for each gene. In the present embodiment, when a plurality of Probes correspond to one gene, the average value of the expression data of the plurality of Probes is used as the representative value of the expression data of the gene. As a result, in the case of human genes, it is possible to narrow down the comparison results of about 20,000 genes. The comparison unit 13 may narrow down to about 1000 landmark genes that are generally considered to be important among the genes. By narrowing down the comparison results in this way, calculations by machine learning can be made possible.

  FIG. 4 is a diagram showing an example in which expression data is compared for each gene. For example, in the sample exposed to “Drug1”, the difference between GeneID1 genes is 101, the difference between GeneID2 genes is 5387, and the difference between GeneID3 genes is 324.

  The encoding unit 14 encodes the expression data of the gene based on the difference (or ratio) of the expression data obtained as shown in FIG. Specifically, the encoding unit 14 ranks gene IDs based on the difference (or ratio) of the expression data. The encoding unit 14 assigns “1” to a predetermined number of genes with higher ranks, “−1” to a predetermined number of genes with lower ranks, and “0” to other predetermined numbers. Encode the data.

  FIG. 5 is a diagram illustrating an example in which “1”, “0”, and “−1” are assigned to genes according to the order. Since the gene ID1 gene has a large difference in expression data and has a high rank, the expression data is coded as "1", and the GeneID3 gene has a small difference (or a large negative value) in the expression data and has a low rank. Data is encoded as "-1". By performing the data conversion of the expression data by the encoding unit 14 in this manner, machine learning can be appropriately performed.

  In the present embodiment, the expression data of the top 2% of the genes is set to “1”, the expression data of the bottom 2% of the genes is set to “−1”, and the expression data of the other genes are set to “0”. . Here, the upper and lower 2% are set to “1” and “−1”, respectively, but the accuracy of the inference using the trained model depends on the degree of “1” and “−1”. Since it changes, it is preferable to adjust based on the evaluation of the trained model.

  The label assignment unit 15 assigns a label indicating toxicity corresponding to the compound. The toxicity learning device 10 is connected to a hepatotoxicity database (hereinafter, referred to as “hepatotoxicity DB”) 30, and based on the toxicity data of the compound stored in the hepatotoxicity DB 30, “Most” “ Labels “Less” and “Non” are given. One example of the hepatotoxicity DB 30 is the Liver Toxicity Knowledge Base (LTKB) provided by the U.S. Food and Drug Administration (FDA). The labeling unit 15 labels the compound with reference to the LTKB.

  FIG. 6A is a diagram illustrating an example in which a label is assigned to each compound by the label assigning unit 15. It should be noted that FIG. 6A shows the vertical axis and the horizontal axis interchanged as compared with FIGS. 4 and 5. For example, a compound of Drug 1 is labeled with a toxicity of "Most", a compound of Drug 2 is labeled with a toxicity of Most, and a compound of Drug 3 is labeled with a toxicity of "Less". As a result, teacher data in which encoded data of gene expression data and labels are set is obtained. Note that it is not important what the existing compound is (name of “Drug1” in FIG. 6A) as the teacher data. What is needed is data obtained by encoding expression data as shown in FIG. In other words, the gene expression data and the corresponding toxicity label serve as the teacher data.

  The model learning unit 16 learns a model using the teacher data generated by the teacher data generation unit 12. The expression data of the teacher data is input to the input layer of the neural network, and the weight coefficient of the neural network is learned so that the corresponding label can be obtained from the output layer. The model learning unit 16 generates a model for inferring toxicity from the expression data by learning the model using a large amount of teacher data. The model learning unit 16 stores the model obtained by the learning in the learned model storage unit 17.

  FIG. 7 is a flowchart illustrating the operation of the toxicity learning device 10 according to the first embodiment. The toxicity learning device 10 inputs the expression data of the sample to which the compound was exposed and the expression data of the control (S10). The toxicity learning device 10 compares the input sample and the expression data of the control (S11). Here, as described above, the difference, ratio, and normalization of the expression data are used to obtain the difference.

  Next, the toxicity learning device 10 compresses the data (S12). That is, the data of the probe of the microarray is converted into the data of the gene, and the number of data is compressed. At this time, human genes (about 20,000) may be used, or about 1000 genes likely to be related to hepatotoxicity.

  The toxicity learning apparatus 10 ranks the genes based on the comparison result of the expression data (S13), and converts the expression data into encoded data based on the assigned rank (S14). Specifically, the expression data of the gene with the higher rank of 2% is “1”, the expression data of the gene with the lower rank of 2% is “−1”, and the expression data of the other genes are “0”. I do. Subsequently, the toxicity learning device 10 refers to the data in the hepatotoxicity DB 30 and assigns a label to the compound to generate teacher data (S15).

  The toxicology learning device 10 determines whether there is any sample data that has not been processed (S16). If there is another sample data (YES in S16), the above process is repeated (S11 to S15). . If there is no other sample data (NO in S16), the toxicity learning device 10 performs model learning using the generated large amount of teacher data, and stores the model obtained by learning in the learned model storage unit 17. (S17).

  FIG. 8 is a diagram showing a configuration of the toxicity prediction device 20. The toxicity prediction device 20 has a learned model storage unit 28 that stores a learned model generated by learning in the toxicity learning device 10. The toxicity prediction device 20 includes an input unit 21 for inputting expression data of a new compound whose hepatotoxicity is to be examined, a pre-processing unit 22 for performing pre-processing on the input expression data, and a liver processing unit using the pre-processed expression data. It has an inference unit 25 for inferring the presence or absence of toxicity and an output unit 26 for outputting an inference result.

  The preprocessing unit 22 includes a comparison unit 23 and an encoding unit 24. The comparison unit 13 and the encoding unit 14 of the toxicology learning device 10 apply the expression data of the new compound input to the input unit 21 to the data. The same processing as that performed is performed to convert the expression data into encoded data. Note that a control data storage unit 27 is connected to the preprocessing unit 22, and compares the expression data of the control read from the control data storage unit 27 with the expression data of the new compound. Thereby, it is not necessary to input control expression data to the toxicity prediction device 20. The preprocessing unit 22 passes the encoded data of the expression data to the inference unit 25.

  The inference unit 25 reads the learned model from the learned model storage unit 28, and applies the encoded data input from the preprocessing unit 22 to the input layer of the read learned model. Thereby, the inference unit 25 obtains the hepatotoxicity output from the output layer of the learned model.

  The configuration of the toxicity learning device 10 and the toxicity prediction device 20 according to the present embodiment has been described above. Examples of the hardware of the toxicity learning device 10 and the toxicity prediction device 20 include a CPU, a RAM, a ROM, a hard disk, and a display. , A keyboard, a mouse, a communication interface, and the like. By storing a program having modules for realizing the above-described functions in a RAM or a ROM and executing the program by the CPU, the above-described toxicity learning device 10 and toxicity prediction device 20 are realized. Such a program is also included in the scope of the present invention.

  The toxicity learning device 10 of the present embodiment encodes the difference in the expression data for each gene based on the result of comparing the expression data of the sample to which the compound was exposed and the expression data of the control, and also describes the toxicity of the compound. The teacher data can be generated by giving the label of. A model for inferring the toxicity of the compound from the expression data can be generated using the teacher data.

(Second embodiment)
FIG. 9 is a diagram illustrating a configuration of the toxicity learning device 10 according to the second embodiment. The toxicity learning device 10 according to the second embodiment uses side effect data stored in a side effect database (hereinafter, referred to as “side effect DB”) 31 in addition to the hepatotoxicity DB 30. The basic configuration of the toxicology learning device 10 according to the second embodiment is the same as that of the toxicology learning device 10 according to the first embodiment, but a label related to the hepatotoxicity of the compound is obtained from the data read from the side effect DB 31. It further includes a label generation unit 18 for generating. The labeling unit 15 uses the label generated by the label generation unit 18 in addition to the label based on the hepatotoxicity data stored in the hepatotoxicity DB 30.

  One example of the side effect DB 31 is FDA Advertise Event Reporting System (FARES) provided by the U.S. Food and Drug Administration (FDA). FARES is a voluntary reporting system for side effects reports, and contains a huge amount of report data from various reporters, such as medical professionals, patients, and pharmaceutical companies. However, since the hepatotoxicity data of the compounds is not systematically compiled, the toxicity learning device 10 of the present embodiment uses various data stored in the side effect DB 31 in order to use the data of the side effect DB 31. Generate a label for hepatotoxicity based on the data.

FIG. 10A is a diagram illustrating an example of data stored in the side effect DB 31. In the example shown in FIG. 10A, the side effects 1 to m are stored for the medicines 1 to k. Data such as x ₁₁ listed in the matrix at the intersection of drugs and side effects is data indicating whether or not there is a side effect said medicament.

  The label generation unit 18 classifies the medicines stored in the side effect DB 31 into target medicines and other medicines used for the sample input to the toxicity learning device 10, and pays attention to the side effects stored in the side effect DB 31. It is classified into the side effects (ie, hepatotoxic side effects) and other side effects, and the number n of side effects occurring with each drug is counted.

FIG. 10B is a diagram showing an example in which data stored in the side effect DB 31 has been processed. In the example shown in FIG. 10 (b), the number of side effects caused of interest in the pharmaceutical of interest is n _11, the number of side effects of interest in other drugs occurs is n _21, attention The sum of the side effects is n _{+ 1} . Further, the number of other side effects drugs caused the subject is n _12, the number of other side effects caused by other medicines is n _22, the sum of the other side effect is n _+2.

The label generator 18 calculates a reporting odds ratio (ROR) for reporting a specific event in order to determine whether or not the target drug has hepatotoxicity. Specifically, it is calculated by the following equation (1).
ROR = (n ₁₁ / n ₂₁ ) / (n ₁₂ / n ₂₂ ) (1)

The (n ₁₁ / n ₂₁ ) of the molecule indicates the proportion of the side effect of interest occurring in the target drug, as a percentage of other drugs. The denominator (n ₁₂ / n ₂₂ ) represents the ratio of other side effects that occurred in the target drug and the ratio to the other drugs. If the value of the ratio of the numerator to the denominator is close to “1”, it can be interpreted that the report of the side effect of interest is also made by accident, and if this ratio is much larger than “1”, It can be construed that the report of the noted side effect made on the drug in question is not accidental. When the lower limit of the 95% confidence interval of the ROR is greater than 1, the label generation unit 18 determines that there is a side effect of interest in the target drug.

  When counting the number of reported side effects of interest from the data (see FIG. 10A) stored in the side effect DB 31, the label generation unit 18 uses an ontology or a medical glossary (MedDRA). Alternatively, the report contents of side effects may be aggregated.

Since the toxicity learning device 10 of the second embodiment generates a hepatotoxicity label using the data stored in the side effect DB 31, it is possible to generate teacher data using data of many compounds. In the present embodiment, an example of calculating the ROR is used to determine whether there is a signal of hepatotoxicity, but the ratio of the report rate of the specific event calculated by the following equation (2) is used. (Proportional Reporting Rations: PRR) may be used.
_{PRR = (n 11 / n 1+} ) / (n 21 / n 2+) ··· (2)

  As a method of determining whether or not the target drug has a noticeable side effect from the data stored in the side effect DB 31, for example, a method such as principal component analysis, factor analysis, or SVM may be used in addition to the above method. You may.

  As described above, the toxicity learning device and the toxicity prediction device of the present invention have been described in detail with reference to the embodiments. However, the toxicity learning device of the present invention is not limited to the above embodiments. For example, in the above-described embodiment, the expression data of the sample is ranked, and the upper and lower 2% are respectively set to “1” and “−1”. However, the upper 2% is set to “1” and the other 2% is set to “1”. It may be “0”.

  In the above-described embodiment, an example has been described in which the toxicity learning device 10 and the toxicity prediction device 20 are configured as separate devices. However, the toxicity learning device 10 and the toxicity prediction device 20 may be configured as a single device. If the toxicity learning device 10 and the toxicity prediction device 20 are configured as one device, it is easy to correct the learned model based on the inference result by the toxicity prediction device 20. If the accuracy of the inference of the trained model is not good, for example, when encoding is performed by the encoding unit 14, the order of the expression data to be converted to “1” and “−1” (upper%, lower%) ) May be changed.

  In the embodiment described above, in order to amplify the data amount of the teacher data, the comparing unit 13 compares the expression data of the sample and the expression data of the control by three methods and generates three times the teacher data. However, the toxicity learning method of the present invention may increase the amount of teacher data by inputting the obtained raw data when inputting the expression data of the sample. That is, usually, in experiments on living organisms, an average of a plurality of measured data is averaged in order to reduce fluctuations caused by the experiment. By considering them as data and using them as teacher data, the data amount can be increased.

  INDUSTRIAL APPLICABILITY The present invention is useful as a toxicity learning device for a model used in a toxicity prediction device for determining the toxicity of an unknown compound.

Reference Signs List 10 Toxicity learning device 11 Input unit 12 Teacher data generation unit 13 Comparison unit 14 Encoding unit 15 Label assignment unit 16 Model learning unit 17 Trained model storage unit 18 Label generation unit 20 Toxicity prediction device 21 Input unit 22 Preprocessing unit 23 Comparison Unit 24 encoding unit 25 inference unit 26 output unit 27 control data storage unit 28 trained model storage unit 30 hepatotoxicity database

Claims (14)

An input unit for inputting expression data of a sample to which the compound is exposed and expression data of a control,
A comparison unit that compares the expression data of the sample and the control for each predetermined gene,
An encoding unit that encodes the expression data of the gene based on the difference in the expression data,
A labeling unit for labeling the encoded expression data with a toxicity label of the compound,
Using the labeled teacher data, a model learning unit for learning a model for predicting the toxicity of the compound from the expression data of the gene,
Toxicity learning device equipped with. The encoding unit includes:
The genes are ranked based on the difference in the expression data, and the rank is determined to be “1” for a predetermined number of genes with a higher rank, “−1” for a predetermined number of genes with a lower rank, and “0” for other genes. The toxicity learning device according to claim 1, wherein The toxicity learning device according to claim 1, wherein the comparing unit obtains a plurality of comparison results by comparing the expression data for each gene by the following methods (i) to (iii).
(I) taking the difference between the expression data of the sample and the control;
(Ii) taking the expression data of the sample and the control;
(Iii) Normalize the expression data of the sample and the control and take the difference.   The toxicity learning device according to any one of claims 1 to 3, wherein the labeling unit uses toxicity data read from a database storing the toxicity of the existing compound as the label. Reads the data of the side effects of the compound from the database that stores the side effects of the existing compound, comprising a label generating unit that obtains the data of the toxicity of the compound based on the data of the read side effects,
The toxicity learning device according to claim 1, wherein the labeling unit uses the obtained toxicity data as the label. Inputting expression data of a sample exposed to a known compound and expression data of a control,
Comparing the expression data of the sample and the control for each gene,
Encoding the expression data of the gene based on the difference in the expression data,
Generating teacher data by labeling the encoded expression data with the toxicity of the compound;
Using the teacher data, learning a model to predict the toxicity of the compound from the expression data of the gene,
Toxicity learning method comprising. The encoding step includes:
Ranking the genes based on the difference in the expression data;
A step of assigning “1” to a predetermined number of genes whose rank is higher, “−1” to a predetermined number of genes whose rank is lower, and “0” to other genes;
The toxicity learning method according to claim 6, comprising: 8. The toxicity learning method according to claim 6, wherein the comparing step obtains a plurality of comparison results by comparing the expression data for each gene by the following methods (i) to (iii).
(I) taking the difference between the expression data of the sample and the control;
(Ii) taking the expression data of the sample and the control;
(Iii) Normalize the expression data of the sample and the control and take the difference. In the inputting step, input expression data of a plurality of samples exposed to a known compound,
In the comparing step, the expression data of the plurality of samples and the control are compared for each gene, and a plurality of comparison results are obtained for the one compound.
The toxicity learning method according to claim 6. In the step of generating the teacher data,
10. The toxicity learning method according to claim 6, wherein toxicity data read from a database storing the toxicity of the existing compound is used as the label. In the step of generating the teacher data,
10. The toxicity learning according to any one of claims 6 to 9, wherein toxicity data of the compound is obtained based on side effect data read from a database storing the side effects of the existing compound, and the obtained toxicity data is used as the label. Method. A program for generating a model used to infer toxicity of a compound based on expression data of a sample to which the compound has been exposed, comprising:
Inputting expression data of a sample to which the compound is exposed and expression data of a control;
Comparing the expression data of the sample and the control for each gene,
Encoding the expression data of the gene based on the difference in the expression data,
Generating teacher data by labeling the encoded expression data with the toxicity of the compound;
Using the teacher data, learning a model to predict the toxicity of the compound from the expression data of the gene,
A program that executes An apparatus for inferring the toxicity of a compound by using a learned model learned by the toxicity learning method according to claim 6,
An input unit for inputting expression data of a sample exposed to an unknown compound,
A comparison unit that compares the expression data of the sample and the expression data of the control for each predetermined gene,
An encoding unit that encodes the expression data of the gene based on the difference in the expression data,
Applying the encoded expression data to the trained model, an inference unit for inferring the toxicity of the compound,
An output unit that outputs an inference result by the inference unit;
A toxicity prediction device comprising:   A trained model for operating a computer to output a value quantifying the toxicity of a compound based on the expression data when the compound is exposed. Expression data encoded based on the difference is input, and based on the input encoded data, an operation based on the learned weighting factors of the neural network is performed, and the toxicity of the compound is quantified from the output layer. A trained model that lets a computer function to output values.

Images