New pollutant PBT prediction method based on convolutional neural network

Abstract

The invention discloses a new pollutant PBT prediction method based on a convolutional neural network, and the method comprises the following steps: obtaining SMILES data of a compound and PBT data of a new pollutant, carrying out the preprocessing of the SMILES data and the PBT data, and generating an SMILES character embedding vector and a PBT attribute index; constructing a convolutional neural network prediction model by taking the SMILES character embedding vector as an independent variable and the PBT data as a dependent variable, and learning correlation characteristics between the SMILES character embedding vector and the PBT data by adopting the convolutional neural network prediction model; and according to the association features, predicting PBT attribute indexes corresponding to the SMILES data. The method has the beneficial effects that the molecular information acquisition efficiency is relatively high, the interference of human factors on the molecular identification information is avoided, meanwhile, the model prediction capability is improved, and the stability is also relatively high. In addition, due to the fact that text data is adopted for prediction, only a CPU needs to be adopted for operation, the calculation speed is high, GPU resources do not need to participate, and a large amount of cost input is saved.

Description

Technical Field

[0001] The present invention relates to the technical field of new pollutant risk identification, and in particular to a new pollutant PBT prediction method based on convolutional neural network. Background Art

[0002] New pollutants refer to pollutants that have been recently discovered or have been paid attention to, pose risks to the ecological environment or human health, but have not yet been included in management or the existing management measures are not sufficient to effectively prevent and control their risks. New pollutants are of various types, with wide sources, hidden effects, strong persistence, bioaccumulation, and great environmental and health risks. According to statistics, there are currently more than 200 million chemical substances registered with the American Chemical Society, and there are about 45,000 chemical substances in my country. These chemical substances can enter the receiving environment directly or indirectly throughout their production, use, treatment and disposal life cycle. In addition, chemical substances will also undergo degradation and transformation products after entering the environment, thus forming a large amount of pollutants. Environmental persistence (P, persistent), bioaccumulation (B, bioaccumulative) and toxicity (T, toxic) are one of the important bases for risk control of new pollutants. New pollutants undergo biological / abiotic degradation transformation in the environment, and the environmental persistence, bioaccumulation and toxicity of some of their transformation products are even higher than those of the parent substances. Therefore, systematically evaluating the PBT properties of new pollutants and their transformation products is of great significance for the control of new pollutants.

[0003] The most traditional method for identifying new pollutants is to obtain the PBT properties of new pollutants and transformation products through experimental methods, which has the characteristics of low efficiency, long cycle and high cost.

[0004] Thanks to the development of neural networks, it is possible to identify new pollutants and transformation products by building a PBT property prediction model. Common methods based on quantitative structure-activity relationship (QSAR) are widely used in the prediction of the physical and chemical properties and toxicity of new pollutants. However, this method is mainly based on a large number of artificially designed molecular descriptors and molecular fingerprint information, and requires the use of a variety of tools / software to calculate molecular characteristics. Therefore, the prediction results are easily interfered by human factors. In addition, this method system is limited by its own characteristics, requiring a large amount of data for modeling and evaluation, and has high requirements for the distribution of positive and negative samples, making model construction time-consuming. Summary of the invention

[0005] In view of the above problems, the present invention proposes a new pollutant PBT prediction method based on convolutional neural network, which mainly solves the problems of background technology.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A new pollutant PBT prediction method based on convolutional neural network includes the following steps:

[0008] Acquire SMILES data of the compound and PBT data of the new pollutant, wherein the SMILES data and the PBT data are preprocessed respectively to generate SMILES character embedding vectors and PBT attribute indicators;

[0009] A convolutional neural network prediction model is constructed using the SMILES character embedding vector as an independent variable and the PBT data as a dependent variable, and the convolutional neural network prediction model is used to learn the association features between the SMILES character embedding vector and the PBT data;

[0010] According to the association features, the PBT attribute index corresponding to the SMILES data is predicted.

[0011] In some implementations, the preprocessing of the SMILES character embedding vector includes:

[0012] The SMILES data is processed by using a word vector model to generate a SMIELS character dictionary. During the construction of the SMIELS character dictionary, the SMILES data is character segmented and hot-coded according to predefined rules to obtain corresponding SMILES character embedding vectors.

[0013] In some implementations, the SMILES character embedding vector includes corresponding unique character information, length information, character-to-vector information, and vector-to-character information.

[0014] In some embodiments, the preprocessing process of the PBT attribute index includes:

[0015] The PBT attribute data are sorted, and the PBT attribute data indicators are divided into 1 and 0 according to preset conditions, so as to represent the P and non-P attributes, B and non-B attributes, and T and non-T attributes of the new pollutant.

[0016] In some embodiments, the training process of the convolutional neural network prediction model includes: randomly splitting the PBT data into a training set and a test set according to a preset ratio, the training set is used to train to obtain the associated features, and the test set is used to evaluate the prediction effect of the convolutional neural network prediction model.

[0017] The beneficial effects of the present invention are as follows: SMILES data is input as a text sentence, and through the encoding-decoding process, one-dimensional character information is mapped to a three-dimensional space structure and converted into a SMILES character embedding vector. Compared with the acquisition of molecular descriptors, molecular fingerprints and molecular graph data, the molecular information acquisition efficiency of the present method is higher, and the interference of human factors on molecular identification information is avoided. At the same time, the model prediction ability is improved, and the stability is also higher. In addition, since the present scheme uses text data for prediction, only the CPU needs to be used for calculation, the calculation speed is fast, and no GPU resources are required, which saves a lot of cost investment. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of the process of predicting the PBT of new pollutants based on a convolutional neural network according to an embodiment of the present invention;

[0019] Figure 2 Schematic diagram of the architecture of the convolutional neural network prediction model in the present invention;

[0020] Figure 3 The evaluation indicators and 5-fold cross-validation results of the convolutional neural network prediction model in Example 1;

[0021] Figure 4 This is the screening and identification result of the PBT properties of the new pollutant transformation products in Example 2. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the content of the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents.

[0023] This embodiment proposes a new pollutant PBT prediction method based on convolutional neural network, such as Figure 1 As shown, the following steps are included:

[0024] S1, obtaining SMILES data of compounds and PBT data of new pollutants, wherein the SMILES data and PBT data are preprocessed respectively to generate SMILES character embedding vectors and PBT attribute indicators.

[0025] In this embodiment, the SMILES data of the compound can be collected from the PubChem database. PubChem is a database of chemical modules that can obtain biological activity data of various organic small molecules. Figure 2In the SMILES preprocessing module in pubchem, write functions to obtain the corresponding Compound Name, CAS, InChI, InChIKey, ExactMass, Molecular Formula, MolecularWeight and Canonical SMILES through the pubchem API as the components of SMILES data. SMILES data is a linear symbol that describes the molecular structure of a compound using ASCII strings, which is used as the input of training data.

[0026] The preprocessing process of SMILES character embedding vector includes:

[0027] S101, obtain and parse SMILES data through the reticulate toolkit of R language combined with RDKit.

[0028] reticulate is a toolkit in the R language that allows users to call Python code directly in the R environment. reticulate calls Python's powerful libraries (such as TensorFlow, PyTorch, Pandas, NumPy, etc.) for data analysis and machine learning. reticulate bridges the R and Python interpreters to make the interaction between the two simple and direct. RDKit is an open source project designed specifically for the fields of chemical informatics and drug discovery. It provides a comprehensive software framework for processing, analyzing and visualizing molecular data. RDKit uses standards such as SMILES (Simplified Molecular Linear Input Specification) and InChI (Internationally Unique Chemical Identifier) ​​to represent molecules. This allows molecular data to be easily stored, searched and compared. RDKit can also create molecular objects from SMILES strings or molecular files (such as .mol ​​files), and supports various operations on molecular structures, such as access to atoms and bonds, detection of rings, etc.

[0029] By combining the reticulate toolkit and RDKit, the process of acquiring SMILES data is standardized and plays a role in quickly constructing input data.

[0030] S102, using a word vector model to process the SMILES data and generate a SMIELS character dictionary.

[0031] In the present embodiment, SMILES can be processed using the word2vec tool, which is a related model for generating word vectors and was launched by Google in 2013. These models are shallow and double-layer neural networks used for training to reconstruct the word text of linguistics. The training speed of Word2vec is very fast, especially on large-scale corpora. Compared with hot encoding (One-Hot), the word vector generated by Word2vec is low-dimensional, reducing computational complexity and storage cost. Word2vec can capture the semantic relationship between words, so that similar words are closer in the vector space.

[0032] Character dictionary is another type of dictionary used in natural language processing and text processing. Unlike vocabulary, it focuses on single characters rather than complete words or vocabulary. In the character dictionary described in the present embodiment, it contains all characters in SMILES data, which can be letters, numbers, punctuation marks, special symbols, etc.

[0033] S1021, during the construction of the SMIELS character dictionary, character segmentation and hot-single encoding are performed on the SMILES data according to predefined rules to obtain the corresponding SMILES character embedding vector.

[0034] The predefined rules described in S1021, that is, based on the composition information of SMILES data, fully consider the double-letter elements (such as Br, Cl, etc.) and special symbols (such as #, @, / , etc.) therein and define them as a character dictionary.

[0035] In this embodiment, the character dictionary is a crucial component. It is a list of all words or vocabulary that the convolutional neural network prediction model may encounter during training and prediction. Each word or vocabulary has a unique index (ID) in the dictionary, which is used to reference the specific word in the input and output of the convolutional neural network prediction model.

[0036] Before the vocabulary converted from SMILES data is input into the convolutional neural network prediction model, each word or word in the vocabulary needs to be converted to a corresponding index. This ensures that the model can handle any given text without being affected by the diversity of words in the text or spelling differences. By mapping words to indexes, high-dimensional text data can be converted into low-dimensional numerical representations. Using a vocabulary can significantly improve the efficiency of model training and reasoning. Since the model only needs to process a relatively small set of indexes, rather than an infinite set of the entire vocabulary library, the amount of computation and memory usage can be greatly reduced.

[0037] The one-hot encoding described in S1021 is a method for processing categorical data, which is mainly used to convert categorical variables into numerical forms that can be understood by machine learning algorithms. In this embodiment, the input SMILES data is required to be numerical, while the categorical data is in text form, so one-hot encoding is required to convert the SMILES data. Figure 2 The SMILES encoding module in uses one-hot encoding, and each category is converted into a binary number with 1 only in the corresponding position and 0 in the other positions.

[0038] When the SMIELS character dictionary is segmented, it is processed as a whole, and the resulting SMILES character embedding vector includes the corresponding unique character information, length information, character to vector information (char_to_int), and vector to character information (int_to_char).

[0039] For the PBT data described in this embodiment, it is possible to select to collect from public literature. Among them, the persistence data of the new pollutant may include biological half-life data and Biowin1-7 predicted values, the bioaccumulation data of the new pollutant may include BCF data and logKow data, and the toxicity data of the new pollutant may include freshwater fish acute toxicity 96-hr LC50 data, chronic toxicity ChV data, large Daphnia acute toxicity 48-hr LC50 data, chronic toxicity ChV data, green algae acute toxicity 96-hr EC50 data, and chronic toxicity ChV data.

[0040] The attribute evaluation method for PBT data of new pollutants refers to relevant evaluation systems at home and abroad, where attribute P refers specifically to the persistence characteristics of new pollutants. The parameters used to evaluate attribute P mainly include biological half-life (Amot-Biowinhalf_life, day) and Biowin (Biowin 2&3 and Biowin6&3), where Amot-Biowin half_life represents the time taken for the new pollutant to be reduced to 50% in the organism, Biowin 2 represents the probability of the degradation speed of the new pollutant based on the nonlinear biodegradation model, Biowin 3 represents the time taken for the new pollutant to be finally converted into carbon dioxide and water (5 represents a few hours, 4 represents a few days, 3 represents a few weeks, 2 represents a few months, and 1 represents a long time), and Biowin 6 represents the probability of whether the new pollutant is easily degraded based on the nonlinear MITI biodegradation model. With reference to the evaluation methods of P attributes in my country, Europe and the United States, when the biological half-life of a new pollutant is known to be greater than 60 days, it is judged to be persistent; if the biological half-life data of a new pollutant is unknown, the biological half-life value, Biowin2, 3 and 6 values ​​are predicted by the model. If the predicted biological half-life is greater than 60 days, Biowin2 and Biowin6 are less than 0.5, Biowin3 is less than 0.5 and Biowin6 is less than 2.2, then the new pollutant is judged to have potential persistence characteristics.

[0041] Specifically, the collected PBT data were sorted using R language to obtain the data sets of various indicators of the PBT properties of new pollutants.

[0042] The preprocessing process of PBT attribute indicators includes:

[0043] S103, sorting the PBT attribute data, and dividing the PBT attribute data indicators into 1 and 0 according to preset conditions, so as to represent the P and non-P attributes, B and non-B attributes, and T and non-T attributes of the new pollutant.

[0044] S2, a convolutional neural network prediction model is constructed with SMILES character embedding vector as the independent variable and PBT data as the dependent variable, and the convolutional neural network prediction model is used to learn the association features between SMILES character embedding vector and PBT data.

[0045] Combined with the above S1, the present embodiment S2 uses the SMILES character embedding vector after hot unique encoding as the independent variable, and the PBT data of the new pollutant (P: biological half-life, Biowin 2&3 and Biowin6&3; B: BCF; T: acute / chronic toxicity to freshwater organisms, NOEC, LC50 or EC50) as the dependent variable to construct the following Figure 2The convolutional neural network prediction model shown in the figure looks for the functional relationship between input and output. The convolutional neural network prediction model includes: SMILES Embedding layer, multiple one-dimensional convolutional layers, flattening layer, fully connected layer, dropout layer and output layer; the SMILES data is segmented into characters according to predefined rules, and then hot-coded, and the SMILES character embedding vector is processed by the character embedding layer; the processed SMILES character embedding vector is input into the one-dimensional convolutional layer to learn the SMILES character information, complete the training of the new pollutant PBT property prediction model, and predict the PBT properties of new pollutants and transformation products based on the trained convolutional neural network prediction model.

[0046] S3, predict the PBT attribute index corresponding to the SMILES data based on the associated features.

[0047] The training process of the convolutional neural network prediction model includes: randomly splitting the PBT data into a training set and a test set according to a preset ratio (such as 4:1). The training set is used to train and obtain associated features, and the test set is used to evaluate the prediction effect of the convolutional neural network prediction model.

[0048] S3 belongs to the PBT attribute determination process for SMILES data. For the determination of the P attribute, the biological half-life values, Biowin2, 3 and 6 predicted by the model, such as the predicted biological half-life, Biowin2 and Biowin6, Biowin3 and Biowin6 respectively point to the P attribute, then it is judged that the new pollutant has potential persistence characteristics. For the determination of the B attribute, if the BCF and logKow of the new pollutant point to the B attribute, it is judged to have the characteristics of bioaccumulation. For the determination of the T attribute, if the NOEC, LC50 or EC50 of different organisms point to the T attribute, then it is judged that the new pollutant has the characteristics of biological toxicity. Finally, the new pollutants and transformation products are screened for pollutants with PB, BT, PT and PBT attributes.

[0049] Example 1:

[0050] The PBT data of compounds were obtained from online papers. The data were labeled with positive and negative to indicate whether they had PBT attributes, and the data were labeled with P, B and T to indicate the persistence, bioaccumulation and toxicity characteristics. The total number of compounds was 3111, of which 350 were labeled with positive and 2761 were labeled with negative. The data sets were classified and sorted to obtain 2650 P attribute data sets, 1242 B attribute data sets and 2959 T attribute data sets.

[0051] The annotated data set is used as the measured data, and the prediction result of the SMILES2PBT model constructed by the present invention is used as the predicted value. Figure 3 As shown, the AUC curve on the left represents the test set result of the SMILES2PBT model of the present invention, the middle bar graph represents the five-fold cross-validation result of the SMILES2PBT model of the present invention, and the right side is the AUC curve obtained by taking the labeled data as the measured value and the model prediction result as the predicted value.

[0052] The results show that the SMILES2PBT model of the present invention has strong predictive ability and the predicted results are very consistent with the measured results.

[0053] Example 2:

[0054] The 3111 compounds obtained from the online papers were used as the parent, and the pollutant transformation product data collected by the pubchem database were obtained through the traversal method. Among the 3111 compounds, 180 were reported to have degradation transformation behaviors, and a total of 554 transformation products were identified.

[0055] For the above 554 transformation products, the SMILES2PBT model constructed by the present invention was used to predict their PBT properties, such as Figure 4 As shown, a total of 394 transformation products with PBT properties were found, including 17 transformation products with PBT properties, 81 transformation products with PB or PT or BT properties, and 296 transformation products with P or B or T properties. The results showed that among the 394 transformation products with PBT properties, there were 147 corresponding parent compounds, of which 55 had PBT properties themselves, but 94 parent compounds did not have PBT property characteristics themselves, but after degradation and transformation, degradation and transformation products with PBT properties were produced.

[0056] The results show that the SMILES2PBT model of the present invention can be applied to the risk identification of new pollutant transformation products.

[0057] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A new pollutant PBT prediction method based on convolutional neural network, characterized in that: The following steps are involved: Acquire SMILES data of the compound and PBT data of the new pollutant, wherein the SMILES data and the PBT data are preprocessed respectively to generate SMILES character embedding vectors and PBT attribute indicators; A convolutional neural network prediction model is constructed using the SMILES character embedding vector as an independent variable and the PBT data as a dependent variable, and the convolutional neural network prediction model is used to learn the association features between the SMILES character embedding vector and the PBT data; According to the association features, the PBT attribute index corresponding to the SMILES data is predicted.

2. The new pollutant PBT prediction method based on convolutional neural network according to claim 1, characterized in that: The preprocessing process of the SMILES character embedding vector includes: The SMILES data is processed by using a word vector model to generate a SMIELS character dictionary. During the construction of the SMIELS character dictionary, the SMILES data is character segmented and hot-coded according to predefined rules to obtain corresponding SMILES character embedding vectors.

3. The new pollutant PBT prediction method based on convolutional neural network according to claim 1, characterized in that: The SMILES character embedding vector includes corresponding unique character information, length information, character to vector information, and vector to character information.

4. The new pollutant PBT prediction method based on convolutional neural network according to claim 1, characterized in that: The preprocessing process of the PBT attribute index includes: The PBT attribute data are sorted, and the PBT attribute data indicators are divided into 1 and 0 according to preset conditions, so as to represent the P and non-P attributes, B and non-B attributes, and T and non-T attributes of the new pollutant.

5. The new pollutant PBT prediction method based on convolutional neural network according to claim 1, characterized in that: The training process of the convolutional neural network prediction model includes: randomly splitting the PBT data into a training set and a test set according to a preset ratio, the training set is used to train and obtain the associated features, and the test set is used to evaluate the prediction effect of the convolutional neural network prediction model.

Images