In silico safety evaluation method, in silico safety evaluation system, and in silico safety evaluation program

Abstract

The aim is to provide an in silico safety evaluation technique that offers statistical reliability similar to multiple sampling in biological trials. [Solution] The machine learning method uses two or more different seed values ​​to output multiple prediction values ​​regarding safety for each seed value, and evaluates safety based on these multiple prediction values.

Description

[Technical Field]

[0001] This invention relates to an in silico safety evaluation technology that uses machine learning techniques to evaluate the safety of chemical substances, including whether or not they are toxic. [Background technology]

[0002] To achieve accurate and robust safety assessments of chemical substances, the Integrated Approach on Testing and Assessment (IATA) approach, which integrates assessment methods based on the Adverse Outcome Pathway (AOP), is recommended. Some validated IATA approaches can be used by certain regulatory authorities as one of the basis for safety assessments of chemical substances.

[0003] As an example of IATA, the Defined Approach (DA) for skin sensitization assessment has been published. The ITS(v1, v2)DA, included in OECD Guideline No. 497, defines a judgment criterion that integrates in chemico, in vitro, and in silico assessments. In the ITS DA, Derek Nexus and QSAR Toolbox are used for in silico safety assessment. These in silico assessments yield results of positive, negative, or out-of-applicability for skin sensitization based on chemical structure information. In addition to such binary classification models, several in silico models have recently been developed to predict LLNA EC3 as a quantitative skin sensitization risk index. LLNA is an in vivo study using mice corresponding to Key Event 4 (KE) of skin sensitization. Models predicting EC3 have been reported by multiple teams, including Givaudan and Shiseido (Non-Patent Literature 1), and are still being updated. Here, it is important to note that many existing qualitative and quantitative skin sensitization prediction models output only one prediction result for a given chemical substance.

[0004] The release of ChatGPT (Generative Pre-Trained Transformer) in 2022 has spurred rapid social implementation of generative AI technology. Such generative AIs perform tasks with astonishing flexibility. This flexibility is influenced by specific parameters, such as "temperature" in GPT, in terms of the reproducibility and diversity of the output. The formula representing the probability distribution involving "temperature" mimics the Boltzmann distribution. The higher the "temperature," the greater the diversity of the output in repeated trials, while the lower the consistency of the output.

[0005] Such parameters related to reproducibility and diversity also exist in existing machine learning models. Some machine learning models used as in silico models have randomness in their algorithms, so if the parameters are not fixed, different prediction results will be obtained from trial to trial. Fixing the seed value is one way to make predictions reproducible by fixing the parameters. For example, in LightGBM, "bagging_seed" functions as a numerical value corresponding to the random number that is the source of the randomness in the bagging process during learning. By fixing such a seed value, it becomes possible to make algorithms with random processes reproducible.

[0006] In contrast, for example, the DPRA (Direct Peptide Reactivity Assay), an in chemico test corresponding to the skin sensitization KE1, involves repeating HPLC measurements three times. The average value is used for the determination. Thus, in many in vivo and in vitro biological tests, stable results are obtained by setting a fixed number of sampling steps, such as n=3. In other words, biological tests assume that the same procedure will be repeated to obtain the same results. In evaluations using in silico models, a single output is often obtained by detecting a fixed alert structure or fixing internal parameters, making it difficult to obtain statistical reliability for the results. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] M. Hirota, et al., Development of an artificial neural network model for risk assessment of skin sensitization using human cell line activation test, direct peptide reactivity assay, KeratinoSens and in silico structure alert parameter. J. Appl. Toxicol., 38 (2018), pp. 514-526 [Overview of the project] [Problems that the invention aims to solve]

[0008] Therefore, in view of the above circumstances, the present invention aims to provide an in silico safety evaluation technique that can obtain statistical reliability, similar to multiple sampling in biological tests. [Means for solving the problem]

[0009] In light of the current situation, the inventors, after diligent study, focused on the seed value, which has conventionally been set to a fixed value in in silico models to fix random numbers, and conceived the idea of ​​obtaining multiple predicted values ​​by providing multiple different seed values. Providing multiple different seed values ​​in this way is expected to yield different results even with the same algorithm, and the fundamental nature of reproducibility differs from the results of multiple samplings in biological tests. However, the inventors believed that obtaining calculation results with multiple different seed values ​​in an in silico model functions similarly in demonstrating that the obtained values ​​are not outliers. As shown in the "Examples" described below, the inventors confirmed that the results of calculating predicted values ​​by setting multiple different seed values ​​in an in silico model showed statistically sufficient reliability, thus completing the present invention.

[0010] In other words, the present invention encompasses the following inventions. (1) An in silico safety evaluation method for predicting the safety of a chemical substance using a machine learning method, characterized in that the machine learning method uses two or more different seed values ​​to output multiple predicted values ​​regarding safety for each seed value, and evaluates safety based on the multiple predicted values.

[0011] (2) The in silico safety evaluation method according to (1), wherein the seed value is the seed value in the learning process of a machine learning model with fixed parameters.

[0012] (3) The in silico safety evaluation method according to (1) or (2), wherein the two or more different seed values ​​are a combination of seed values ​​such that the multiple predicted values ​​output follow a normal distribution.

[0013] (4) The in silico safety evaluation method according to (1) or (2), wherein the safety evaluation is based on the average value of the multiple predicted values.

[0014] (5) An in silico safety evaluation system for predicting the safety of chemical substances by a machine learning method, comprising: a seed value setting unit that sets seed values of two or more different values; a predicted value output unit that uses the set seed values and outputs a predicted value regarding the safety by machine learning means; and an evaluation output unit that outputs an evaluation of the safety based on a plurality of predicted values output for each of the seed values of the different values. The in silico safety evaluation system is characterized by the above.

[0015] (6) An in silico safety evaluation program for predicting the safety of chemical substances by a machine learning method, which causes a computer to function as: a seed value setting unit that sets seed values of two or more different values; a predicted value output unit that uses the set seed values and outputs a predicted value regarding the safety by machine learning means; and an evaluation output unit that outputs an evaluation of the safety based on a plurality of predicted values output for each of the seed values of the different values. The in silico safety evaluation program is characterized by the above. [Effect of the Invention]

[0016] According to the present invention configured as above, it is possible to provide an in silico safety evaluation technology in which statistically sufficient reliability can be obtained for prediction results, similar to the results of multiple samplings in a biological test. [Brief Description of the Drawings]

[0017] [Figure 1] It is a block diagram showing a schematic configuration of an in silico safety evaluation system. [Figure 2] It is a diagram explaining that in Verification 1, 100 seed values of Optuna are set and 100 prediction models and 100 predicted values are obtained. [Figure 3]This is a figure showing the distribution of 100 hyperparameters obtained by Optuna. (a) shows the distribution of 'Log(lamda_l1)', (b) shows the distribution of 'Log(lamda_l2)', (c) shows the distribution of 'num_leaves', (d) shows the distribution of 'feature_fraction', (e) shows the distribution of 'bagging_fraction', (f) shows the distribution of 'bagging_freq', and (g) shows the distribution of'min_child_samples'. [Figure 4] This is a figure explaining the variation in prediction accuracy for all 22 substances for testing by Optuna_seed. [Figure 5] This is a figure explaining the variation in predicted values by 100 LightGBMs by Optuna_seed for 22 substances for testing. [Figure 6] This is a figure explaining the Max / min ratio of 100 predicted EC3(%) by Optuna_seed for 22 substances for testing. [Figure 7] This is a figure explaining that in Verification 2, the seed values of LightGBM are set to 100, and 100 prediction models and 100 predicted values are obtained. [Figure 8] This is a figure explaining the variation in prediction accuracy for all 22 substances for testing according to the seed values of LightGBM. [Figure 9] This is a figure explaining the variation in predicted values by 100 LightGBMs by the seed for 22 substances for testing. [Figure 10] This is a figure explaining the Max / min ratio of 100 predicted EC3(%) by the seed for 22 substances for testing. [Figure 11] This is an explanatory diagram about the required number of samplings according to the allowable prediction error range. [Figure 12] This is a figure explaining that in Verification 4, the average value of multiple samplings is used as the predicted value.

Mode for Carrying Out the Invention

[0018] Next, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0019] The in silico safety evaluation system 1 of the present invention is an information processing system that predicts the safety of chemical substances using computer machine learning techniques, and consists of one or more information processing devices, as shown in the block diagram of Figure 1, comprising a processing device 2, a storage means 3, an input means 4, and an information display unit 5. Specifically, it is a computer device centered on the processing device 2, comprising a storage means 3, an input means 4 such as a pointing device, keyboard, or touch panel, an information display unit 5 such as a display, and other communication control units (not shown).

[0020] The processing unit 2 is mainly composed of a CPU such as a microprocessor and has a storage unit consisting of RAM and ROM (not shown) where programs and processing data that define the procedures for various processing operations are stored. The storage means 3 consists of memory or hard disks inside or outside the information processing system 1. The contents of some or all of the storage unit may be stored in the memory or hard disk of another computer connected to the information processing system 1 via communication.

[0021] The processing unit 2 is equipped with machine learning means 20, and functionally comprises a seed value setting unit 21 that sets two or more different seed values, a prediction value output unit 22 that outputs the predicted values ​​related to safety using the set seed values ​​and the machine learning means 20, and an evaluation output unit 23 that outputs a safety evaluation based on the multiple prediction values ​​output for each of the different seed values, and these functions according to the program described above.

[0022] The machine learning method 20 consists of a machine learning model that predicts the intensity of toxicity risk of a chemical substance, using molecular descriptors based on chemical structure and in chemico or in vitro test data as explanatory variables, and the intensity of toxicity risk as the dependent variable. A specific example of safety assessment of a chemical substance is in silico skin sensitization prediction, which predicts the LLNA EC3 value.

[0023] As data sources used in machine learning means 20, for example, in the case of in silico skin sensitization prediction, data for 195 chemical substances can be obtained from existing literature, including LLNA results, as well as data from DPRA, KeratinoSens®, and h-CLAT, which are in chemico / in vitro test methods corresponding to KE1-KE3. However, "Log" refers to the common logarithm. These can be managed by the storage means 3 described above.

[0024] Further data organization is preferable. Specifically, for example, all 195 substances can be sorted in descending order by EC3 (%) and assigned sequential numbers from SN001 to SN195. A predetermined number of randomly selected substances, for example 22 substances, can be used as test data, and the remaining 173 substances can be set as training data. The machine learning means 20 then uses the training data to train a machine learning model, for example, using k-fold cross-validation. The prediction accuracy of the constructed model can be objectively evaluated using the test data.

[0025] Mordred (ver. 1.2.0) allows calculations for all 195 substances and can obtain molecular descriptors that do not show the same value for all substances. Furthermore, Boruta (ver. 0.3), a method that uses the variable importance of Random Forest to select variables, can be used to narrow down the variables to be used. For this purpose, prediction of LLNA EC3 using Random Forest is performed on only the training data using the obtained descriptors. The descriptors selected here can be used as part of the explanatory variables. In addition to the selected molecular descriptors, Log(MIT(μM)), Log(CV75(μM)), Log(Adjusted Cys depletion), Log(Adjusted Lys depletion), and Log(KEC1.5) can be used as explanatory variables.

[0026] As a machine learning model, LightGBM, a commonly used GBDT algorithm that improves computational efficiency through leaf-wise tree and exclusive feature bundling, can be adopted. All model construction is done using Python (registered trademark) for Windows v3.12.4 and lightgbm (4.1.0).

[0027] Parameter tuning can be performed using, for example, the parameter tuning tool Optuna (3.4.0) based on the Bayesian optimization algorithm, and tuning can be performed on 'lamda_l1', 'lamda_l2', 'num_leaves', 'feature_fraction', 'bagging_fraction', 'bagging_freq', and 'min_child_samples'. The seed value setting unit 21, in this parameter tuning, sets the seed value of LightGBMTunerCV to two or more different values, and obtains prediction results equal to the number of seed values.

[0028] Here, the seed value is preferably the seed value in the learning process of the machine learning model after the parameters have been fixed. Furthermore, preferably, the seed value combination is such that the multiple predicted values ​​output by the predicted value output unit 22 follow a normal distribution. Such combinations of seed values ​​that follow a normal distribution can be found in advance and stored for management.

[0029] Log(EC3(μmol / cm 2 The performance of the model that predicts )) is R 2 It can be evaluated by its value, RMSE, and is defined as shown in formula (1). 2 A higher value, with 1 as the upper limit, indicates higher prediction accuracy, while a lower value, with RMSE as the lower limit, indicates smaller prediction error. Specifically, the prediction value output unit 22 uses the machine learning means 20 to analyze these R values. 2 Outputs the value and RMSE.

[0030]

number

[0031] The safety evaluation by the evaluation output unit 23 is preferably based on the average of multiple predicted values. For example, the average value itself and results that are more than 2 standard deviations away from the average value are excluded as outliers, and the average of the remaining predicted values ​​is used as the final predicted value. While there is no need to question existing prediction methods that handle a single predicted value if a certain level of error is acceptable, by setting two or more different seed values ​​and using multiple predicted values ​​as in the present invention, it is possible to avoid fluctuations in prediction results that depend on the seed value, thereby contributing to an improvement in the reliability of the in silico safety evaluation method.

[0032] The present invention is not limited in any way to the embodiments described above, and can of course be implemented in various forms without departing from the spirit of the invention. [Examples]

[0033] The following describes the results of Verifications 1-4, which investigated whether different predicted values ​​based on seed values ​​correspond to multiple samplings in a biological trial using an in silico skin sensitization prediction model utilizing machine learning technology. Specifically, we explain the results of verifying the statistical reliability of prediction results by comparing multiple calculation results with different seed values ​​to determine whether the results aggregate within a certain range even with different seed values ​​in an in silico model that uses a machine learning algorithm.

[0034] In this embodiment, 100 possible seed values ​​(0-99) are set for both Optuna and the LightGBM model. However, seed values ​​other than 0-99 are also acceptable, and they may be discontinuous rather than consecutive. Furthermore, while 100 seed values ​​are used for verification, the system is not limited to 100.

[0035] (Verification 1) Verification 1, as shown in Figure 2, involves setting the Optuna seed value (Optune_seed) to 100 different values ​​from 0 to 99, thereby obtaining parameters 1 to 100. Then, for each parameter obtained by Optuna, 100 different prediction models (prediction model 1 to prediction model 100) are obtained using a machine learning algorithm with a fixed seed value. Furthermore, prediction values ​​1 to 100 are obtained from prediction models 1 to 100.

[0036] <Adding Randomness> This study analyzes the distribution of parameters specified by Optuna on the same data, and the distribution of model predictions for each parameter configuration. In Optuna, set the 'optuna_seed' in LightGBMTunerCV to one of 100 values ​​from 0 to 99. In LightGBMTunerCV, fix the splits using folds=KFold(n_splits=5, shuffle=True, random_state=42) to prevent randomness in the split cross-validation of the model construction.

[0037] In constructing the Lightgbm.cv model using parameters obtained from Optuna, the splitting is similarly specified by folds=KFold(n_splits=5, shuffle=True, random_state=42). Furthermore, to ensure reproducibility according to optuna_seed, numpy.random.seed(42), random.seed(42), and os.environ['PYTHONHASHSEED']=str(42) are set, and 'deterministic' and 'force_col_wise' in LightGBM are set to True, and 'seed' is set to 42.

[0038] (Results of Verification 1) <Variable Selection> Using Mordred, we obtained 701 molecular descriptors for all 195 substances that are calculable and have a non-zero variance. LLNA Log(EC3(μmol / cm 2 Variable selection using Boruta was performed on the training data only, with )) as the dependent variable and all molecular descriptors as independent variables. As shown in Table 1, seven descriptors were selected as useful variables.

[0039] [Table 1]

[0040] <Construction of a machine learning model to predict skin sensitization intensity and verification of randomness using Optuna_seed> Using data from 173 substances for learning, the skin sensitization strength of a chemical substance was defined as LLNA Log(EC3(μmol / cm²)). 2 A LightGBM model was constructed with )) as the dependent variable. The 7 descriptors selected above and the 5 variables from the collected in chemico / in vitro test results, for a total of 12 variables, were used as independent variables.

[0041] In Optuna, we set the LightGBMTunerCV seed value to 0-99 for 100 different values ​​and obtained 100 prediction results. As shown in Figure 3 and Table 2, the variation in hyperparameters obtained by the 100 parameter tunings was large in regularization terms such as lambda_l1 and lambda_l2, which take continuous values. Note that num_leaves was set to the default value of 31 for all seed values.

[0042] [Table 2]

[0043] Furthermore, 100 LightGBM models were constructed using the 100 hyperparameter settings obtained.

[0044] The variation in prediction accuracy across all 22 test substances is R 2The mean value was 0.517 and the standard deviation was 0.00761, while the mean value for the RMSE was 0.965 and the standard deviation was 0.00754 (Figure 4). Thus, the variation in prediction accuracy across all 22 test substances was very small.

[0045] On the other hand, focusing on individual substances, as shown in Figure 5, the variation in predicted values ​​for the 22 test substances ranged from a minimum of 0.028 to a maximum of 0.242 in standard deviation. As shown in Figure 6, the average maximum / minimum value of the predicted EC3 (%) was 2.5, with a maximum of 4.88 (EC3 6.41% / 1.31%).

[0046] Furthermore, the skin sensitization potential of chemical substances is classified into "sub-category 1A," "sub-category 1B," or "not applicable" in the GHS classification, with EC3 2% as the threshold. Based on the predicted maximum and minimum values, variations in the GHS classification of skin sensitization were observed for 4 out of 22 substances.

[0047] [Table 3]

[0048] (Verification 2) Next, in validation 2, as shown in Figure 7, we analyze the variation in predicted values ​​that occurs when only the 'seed' is changed in a LightGBM model with the same parameters on the same data. We specify the parameters obtained when optuna_seed=42 in LightGBM's params, and set only 'seed' to 100 different values ​​from 0 to 99. At this time, the fixing of the split in the split cross-validation is performed as described above.

[0049] (Results of Verification 2) <Construction of a LightGBM model with fixed parameters and verification of randomness using seeds> Using data from 173 substances for learning, the same 12 variables as above were used as explanatory variables, and LLNA Log(EC3(μmol / cm³)) were used. 2 A LightGBM model was constructed with )) as the dependent variable. In the model, except for ’seed’, the parameters obtained with Optuna_seed = 42 in the above verification were set. ’seed’ was set to 100 values from 0 to 99, and 100 sets of prediction results were obtained.

[0050] As shown in Fig. 7, the variation in prediction accuracy for the entire 22 test substances has an average value of 0.521 and a standard deviation of 0.01308 as R, and an average value of 0.961 and a standard deviation of 0.01319 as RMSE. 2 The average value was 0.521, the standard deviation was 0.01308 as R, and the average value was 0.961, the standard deviation was 0.01319 as RMSE. Similar to the verification of Optuna_seed, the variation in prediction accuracy for the entire 22 test substances was very small.

[0051] Looking at individual substances, as shown in Fig. 9, the variation in predicted values for the 22 test substances had a minimum standard deviation of 0.052 and a maximum of 0.159. As shown in Fig. 10, the maximum value / minimum value of the predicted EC3(%) was on average 2.64, and the maximum was 5.91 (EC3 2.34% / 0.40%). Also, as shown in Table 4, due to the predicted maximum and minimum values, there were variations in the GHS classification of skin sensitization in 3 out of 22 substances. The variation in predicted values due to the LightGBM seed and the variation in predicted values due to Optuna_seed were of the same degree overall.

[0052]

Table 4

[0053] (Verification 3) <Normality test and estimation of required number of samplings> Next, to evaluate whether each of the 100 sets of predicted values obtained due to the variation of Optuna_seed and seed follows a normal distribution, the Shapiro-Wilk test, which is a normality test, was conducted. Also, when the data follows a normal distribution, its 95% confidence interval is the sample mean ± 1.96×SE. If the 95% confidence interval is arbitrarily set as the allowable error range, the required number of samplings n can be calculated from the mean value and SD.

[0054] The standard error SE is given by equation (2) below. Here, since we are using the variability of the population rather than the sample, we use the unbiased standard deviation for SD.

[0055]

number

[0056] For each of the 22 test substances, calculate the number of samples required to correspond to the acceptable error range. The error range is calculated using the predicted Log(EC3(μmol / cm²)). 2 By setting Log k (k=1.25, 1.5, 2.0) to ), we can obtain EC3 (μmol / cm³). 2 This represents a range from k times to 1 / k.

[0057] (Results of Verification 3) <Normality of predicted values ​​and required number of samples> For each of the 22 test substances, we used the Shapiro-Wilk test to verify whether the 100 predicted values ​​obtained by Optuna_seed and seed followed a normal distribution. As shown in Table 5, the variation in Optuna_seed resulted in p<0.05 for 21 out of 22 substances, indicating that the predicted values ​​for most DF substances do not follow a normal distribution. On the other hand, when the parameters were fixed, the variation in seed resulted in p>0.05 for 20 out of 22 substances, leading to the conclusion that the predicted values ​​for most substances do not follow a normal population.

[0058] [Table 5]

[0059] We focused on the variation in predicted values ​​due to variations in the LightGBM seed, which is closer to a normal distribution. When a prediction error of Log k was allowed, and the range was mapped to a 95% confidence interval, the median number of required samples was 0.281 when k=2, as shown in Figure 11(a). The maximum value was 1.083, and all but one substance required less than one sample. When k=1.5, the median was 0.820, as shown in Figure 11(b), and when k=1.25, the median was 2.707, as shown in Figure 11(c). Therefore, it was shown that one sample is sufficient for most substances if an error range of ±Log 1.5 is allowed.

[0060] (Verification 4) <Multiple sampling and outlier removal> In biological studies, results that are more than 2 standard deviations (SDs) away from the mean across multiple samplings are sometimes treated as outliers. As an example of treating multiple predicted values ​​corresponding to a LightGBM seed in a similar manner to multiple samplings in a biological study, we excluded results that were more than 2 SD away from the mean of n predicted values ​​corresponding to a seed value, and adopted the mean of the remaining predicted values ​​as the final predicted value. The n predicted values ​​were obtained using seeds from 0 to n-1. Outliers were excluded after a certain number of samples, and as shown in Table 6, a maximum of 7 predicted values ​​were outliers among 100 predicted values.

[0061] [Table 6]

[0062] (Results of Verification 4) The variation in prediction accuracy due to the variation in the LightGBM seed shown in Figure 8 is R 2 Compared to the previous values ​​of 0.521 (mean) and 0.01308 (standard deviation), the R value when the number of samples is increased is... 2 It improved slightly. However, when the number of samples is only one, R 2The result was below the mean + 2SD. Similarly, the RMSE was within the range of mean ± 2SD.

[0063] From the above, it was shown that multiple samplings and outlier removal do not significantly improve prediction accuracy.

[0064] On the other hand, in the cases where no predicted values ​​were excluded and the number of samples was 3 or 5, 100 different predictions were obtained by randomly selecting 3 or 5 seeds from the 100 predicted values ​​and using their average as the predicted value. For example, in the case of 3 samples, as shown in Figure 12, 100 different predicted values ​​are obtained by randomly selecting 3 predicted values ​​from the 100 predicted values ​​and sampling their average as the average predicted value.

[0065] For cases where no predicted values ​​were excluded, such as 3 or 5 sampling cycles, 100 variations were performed by randomly selecting 3 or 5 seeds from 100 predicted values ​​and using their average as the predicted value. As shown in Table 7, the standard deviation of the prediction accuracy for each of the 100 variations was smaller than the standard deviation of the 100 predicted values ​​when the sampling cycle was 1 cycle.

[0066] [Table 7]

[0067] From the above, it was shown that performing multiple samples reduces the variability of the predicted value due to the seed value.

[0068] (Consideration) In silico models, a single predicted value is often obtained for each test substance. However, many biological studies exhibit variability in results, requiring a certain number of sampling cycles to ensure stable results. To address this difference, the seed value controls the randomness of the in silico model's predictions. We investigated whether this randomness could be used to obtain multiple predicted values, similar to biological studies.

[0069] The variability of predicted values ​​resulting from the randomness of parameter tuning based on Optuna_seed did not exhibit a normal distribution. On the other hand, the variability of predicted values ​​based on the seed in the LightGBM model was shown to be close to a normal distribution. Therefore, multiple samples in a biological experiment and multiple predicted values ​​obtained with different seed values ​​can be treated similarly in that they are expected to follow a normal distribution.

[0070] By associating the 95% confidence interval with an acceptable error range, the number of samples required to obtain stable results within a specified range was calculated. The results showed that, with an error range of approximately ±Log 1.5, one sample was sufficient for most substances. Therefore, if a certain error range is acceptable, the in silico model, as with conventional methods, only requires obtaining a single predicted value. (LLNA EC3 (μmol / cm³)) 2 An error of ±Log 1.5 in ) means that EC3(%) is 1.5 times or 1 / 1.5 times. When determining the error range, care must be taken with the units and scales, as the range will differ significantly depending on whether EC3(%) is 10% (6.7%-15%) or 2% (1.3%-3%). Furthermore, in the examination of the required number of samples shown in Figure 11, the only test substance that showed a large value was SN163 (CAS No. 50-53-3, Chlorpromazine), which was the same as the only substance with Max / min of EC3(%) > 5 in Figure 10. This substance has GHS 1A skin sensitization properties, while showing completely negative results with DPRA and KeratinoSens(registered trademark). Such properties were not observed in other test substances. Skin sensitizing substances that are difficult to detect by existing evaluation methods may cause fluctuations in in silico predictions.

[0071] As an example of treating multiple predicted values ​​obtained in accordance with a seed value in the same way as the results of a biological test, results that were more than 2SD away from the mean were excluded as outliers, and the average of the remaining predicted values ​​was adopted as the final predicted value. These predicted values ​​did not exceed the average prediction accuracy of a single predicted value + 2SD. Therefore, it was shown that prediction accuracy does not improve significantly by multiple sampling and exclusion of outliers. On the other hand, results from 100 tests each with 3 and 5 sampling cycles showed that multiple sampling cycles resulted in less variation in predicted values ​​due to the seed value. From the above, it can be said that while a method of obtaining multiple predicted values ​​according to the seed value and excluding outliers cannot be expected to improve prediction accuracy, it can prevent the prediction results from becoming outliers according to the seed value.

[0072] This study involved 100 trials for each pattern, and it is possible that different results could be obtained by increasing the number of trials. However, even with only 100 trials, clear differences in distribution trends and changes in prediction variability were observed, so it is quite possible that the trends observed in this study will become even clearer with an increase in the number of trials. [Explanation of symbols]

[0073] 1. Information Processing System 2 Processing Unit 3 Memory means 4. Input means 5 Information display section 20 Machine Learning Methods 21 Seed value setting section 22 Predicted Value Output Unit 23 Evaluation Output Unit

Claims

1. An in silico safety assessment method that predicts the safety of chemical substances using machine learning techniques, An in silico safety evaluation method characterized by using two or more different seed values ​​in the machine learning method to output multiple predicted values ​​regarding safety for each seed value, and evaluating safety based on the multiple predicted values.

2. The in silico safety evaluation method according to claim 1, wherein the seed value is the seed value in the learning process of a machine learning model with fixed parameters.

3. The in silico safety evaluation method according to claim 1 or 2, wherein the two or more different seed values ​​are a combination of seed values ​​such that the plurality of predicted values ​​to be output follow a normal distribution.

4. The in silico safety evaluation method according to claim 1 or 2, wherein the safety evaluation is based on the average value of the plurality of predicted values.

5. An in silico safety assessment system that predicts the safety of chemical substances using machine learning methods, A seed value setting unit that sets two or more different seed values, A prediction value output unit that outputs a prediction value related to safety using machine learning means based on the set seed value, An evaluation output unit outputs a safety evaluation based on multiple predicted values ​​output for each of the different seed values ​​mentioned above. An in silico safety evaluation system characterized by comprising the following features.

6. An in silico safety assessment program that predicts the safety of chemical substances using machine learning methods, Computers, A seed value setting unit that sets two or more different seed values. A prediction value output unit that uses a set seed value to output a prediction value related to safety using machine learning means. An in silico safety evaluation program characterized by functioning as an evaluation output unit that outputs a safety evaluation based on a plurality of predicted values ​​output for each of the different seed values.

Images