Molecular toxicity pathway prediction method based on link symbol prediction

Abstract

The application discloses a molecular toxicity path prediction method based on link symbol prediction, utilizes a topological relation of a molecular-toxicity path network, designs a calculation method of node embedding features to perform molecular toxicity prediction, and explores the possibility of network topological structure for assisting drug design.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug development technology, specifically relating to a method for predicting molecular toxicity pathways based on link symbol prediction. Background Technology

[0002] The discovery and development of small molecule drugs requires a long period of time and significant costs. A good drug must not only possess good efficacy but also excellent physicochemical and pharmacokinetic properties. If pharmacokinetic properties are only considered in the later stages of drug development, the risk of drug development failure increases.

[0003] Drug toxicity refers to the ability of a drug to damage the structure or substructure of an organism. To ensure drug safety, toxicity studies are required before clinical trials can begin. Traditional animal experiments have drawbacks such as long experimental cycles, high costs, and difficulty in large-scale screening. Therefore, computational methods for predicting drug toxicity that are rapid, efficient, and cost-effective are rapidly developing.

[0004] Significant progress has been made in computational methods for molecular toxicity prediction. In general, most methods, especially machine learning and deep learning models, have proven effective in analyzing the vast amounts of current molecular toxicity data and predicting new compounds. The effectiveness of traditional machine learning methods often depends on the quality of the extracted features and requires considerable domain knowledge. In recent years, thanks to the rapid development of deep learning algorithms and their successful applications in other fields, the accumulated research in molecular toxicity prediction has also promoted the application prospects of deep learning in drug design. Structural data such as drugs can be automatically extracted using graph neural networks; these structured deep learning models combined with multilayer neural networks have been successfully applied in drug design. However, despite the tremendous efforts and significant achievements made by researchers in toxicity prediction, considerable challenges remain in practical work, mainly in the following aspects:

[0005] 1) Feature extraction requires certain professional knowledge, and it is difficult for non-professionals to carry out related work, requiring a high level of expertise from researchers;

[0006] 2) Insufficient interpretability: Drug molecule embedding representations learned by deep learning or graph representation are always difficult to interpret, lacking the relationship between compound functional groups and toxicity.

[0007] Therefore, it is necessary to design a new method for predicting molecular toxicity. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies and to provide a molecular toxicity prediction method based on link symbol prediction.

[0009] The concept of this invention:

[0010] We propose a molecular toxicity prediction method based on link symbol prediction that does not require prior chemical knowledge. Both molecules and toxic pathways are treated as nodes. The embedding of each node is learned through computation, and then a classification function is used to perform binary classification of the features of node pairs. We explore the feasibility of molecular-toxic pathway network topology for molecular toxicity prediction.

[0011] This invention employs a link symbol prediction approach. First, it obtains molecular toxicity pathway data from the tox21 database, including 12 typical toxicity signaling pathways such as estrogen receptor α and estrogen receptor. The dataset is then organized and divided into commonly used data formats for link prediction. Randomly generated node vectors are used, and the positive and negative sums of each node and its neighbors are calculated for iterative updates (step 2.2). Finally, a classifier (a conventional classifier, also referred to as a predictor or prediction model in this invention) is used to perform binary classification on the updated vectors to predict whether they are toxic. In this invention, logistic regression is used for binary classification.

[0012] In view of the above-mentioned inventive concept, the technical solution provided by the present invention to achieve the purpose of the invention is as follows:

[0013] The molecular toxicity pathway prediction method based on link symbol prediction is characterized by the following steps:

[0014] 1) Convert the data format of the tox21 molecular toxicity pathway into the file format required for link symbol prediction: "Molecular Node Toxicity Pathway Node Node Pair Label"; where the label for toxicity is 1 and the label for non-toxicity is 0.

[0015] The positive and negative samples of the toxicity link prediction data are divided into training and test sets in the same proportion, and the positive and negative samples are merged to form training and test sets respectively.

[0016] 2) Extract and calculate the feature vector of each node and node pair from step 1).

[0017] 2.1) For each molecule-toxicity node pair (u,v), randomly generate a k1-dimensional vector U for nodes u and v respectively. p V p and k2-dimensional vector U N V N ;

[0018] Will U p V p Mapped to the positive space, they serve as the positive space vectors for nodes u and v, respectively;

[0019] Will U N VN Mapped to the negative space, they serve as the negative space vectors for nodes u and v, respectively;

[0020] That is, each node is initialized to generate a positive space vector and a negative space vector respectively; 2.2) For each node u, calculate the cosine similarity z of the positive space vector and the negative space vector of the node pair (u,v) formed by the node u and the node v that are connected to it. + z - The processed cosine similarity p is obtained by nonlinear calculation using the LeakReLU activation function. + p - ,in:

[0021]

[0022] In the formula, α is a constant;

[0023] p calculated each time + p - The positive space vectors U corresponding to node u respectively p and negative space vector U N Multiplication (i.e., p) + The positive space vector U of node u p Multiplication, p - The negative space vector U of node u N (Multiplication), the resulting vectors correspond to U respectively. p and U N Perform summation to obtain the updated U p and U N (i.e.: p) + The positive space vector U of node u p The result of multiplication and U p Perform summation to obtain the updated U p p - The negative space vector U of node u N The result of multiplication and U N Perform summation to obtain the updated U N Perform the same computation as for node u on all nodes v to obtain the updated V. p and V N ;

[0024] After calculating all (u,v) node pairs, we obtain the final positive and negative space vectors updated for each node; that is, the positive space vector of each node is the sum of the positive relationships between the node and its neighboring nodes, and the negative space vector of each node is the sum of the negative relationships between the node and its neighboring nodes.

[0025] 2.3) The last k1-dimensional positive space vector and k2-dimensional negative space vector updated for each node are concatenated horizontally to obtain a (k1+k2)-dimensional vector, which is used as the embedding feature of the node.

[0026] The embedding features of each node in each node pair are horizontally concatenated to obtain the embedding features of that node pair, resulting in a (k1+k2)*2 dimensional concatenated vector x, which is applicable to both the training and test sets.

[0027] 3) Use the concatenated vector x of the node pairs obtained in step 2) to train and validate the classifier.

[0028] The classifier is trained using the training set data, and the logistic regression function h is used. θ (x) Perform binary classification on the concatenated vector x of the node pairs obtained in the training set to obtain the toxicity score of the edge connection for each pair of nodes, where:

[0029]

[0030]

[0031] In the formula, θ T The logistic regression function h obtained during training θ The parameters of (x) are used to apply the loss function to the logistic regression function h. θ The parameters of (x) are updated to minimize the difference between the predicted and actual values, thereby making the logistic regression function more accurate in its predictions.

[0032] The classifier was validated using a test set with an updated logistic regression function h. θ (x) Perform binary classification on the spliced ​​vector x of node pairs in the test set to obtain the toxic pathway prediction effect of the test set and verify the prediction effect of the classifier;

[0033] 4) Use the classifier trained and validated in step 3) to predict molecular toxicity pathways.

[0034] Furthermore, step 1) specifically involves:

[0035] Molecules and toxic pathways are represented by nodes. All molecules are considered as one type of node, and all toxic pathways are considered as another type of node. The nodes are labeled with serial numbers and tags to form a toxicity link prediction data format of "molecule node toxic pathway node node pair tag". Among them, the tag for toxic is 1, and the tag for non-toxic is 0.

[0036] The positive and negative samples of the toxicity link prediction data are first divided into training and test sets in an 8:2 ratio. Then, 8 positive and negative samples are merged to form the training set, and 2 positive and negative samples are merged to form the test set. This ensures that the proportion of positive and negative samples in the training and test sets is consistent, achieving data balance. Data balance is very beneficial to the accuracy of the prediction results in the later stage.

[0037] Furthermore, in step 3), the method for constructing the loss function is as follows:

[0038]

[0039] in, i∈(0,1), y[i]∈[0,1] indicates whether label i belongs to a compound, and

[0040] Meanwhile, the present invention provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0041] An electronic device, characterized in that it includes a processor and a computer-readable storage medium;

[0042] The computer-readable storage medium stores a computer program, which, when executed by the processor, performs the steps of the above method.

[0043] The advantages of this invention are:

[0044] 1. The present invention proposes a toxicity prediction method based on link symbol prediction, which requires no chemical knowledge, reduces the professional requirements of researchers, and does not require conventional feature extraction of drug molecules beforehand. It can achieve prediction results comparable to conventional machine learning methods such as RF and SVM and advanced deep learning methods such as GCN using only the network topology.

[0045] 2. This invention links molecules to toxic pathways, significantly improving the interpretability of prediction results. Attached Figure Description

[0046] Figure 1 This is the overall architecture of the method proposed in this invention;

[0047] Figure 2 This is a demonstration of the positive vector calculation of u2 in this invention. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0049] An embodiment of the molecular toxicity pathway prediction method based on link symbol prediction proposed in this invention is as follows:

[0050] This embodiment uses a metabolic pathway dataset from tox21: this dataset contains 7831 compounds, possessing one or more of 12 toxic pathways, or being completely non-toxic. These 12 toxic pathways are estrogen receptor α, estrogen receptor, aromatic hydrocarbon receptor, androgen receptor, androgen receptor, peroxisome proliferator-activated receptor γ, antioxidant response element, heat shock factor response element, ATAD5, mitochondrial membrane potential (MMP), and p53. The data format of "molecular node toxic pathway node-to-node tag" is divided into training and test sets in an 8:2 ratio.

[0051] In this tox21 dataset, after cleaning, there are 7811 molecular data points and 12 toxic pathways. Each molecule is numbered from 0 to 7810, and each toxic pathway is numbered from 7810 to 7822. The data is processed according to the above format to form a "molecular node toxic pathway" network. For each node, a 32-dimensional vector is randomly generated as the initial positive space vector, and a 32-dimensional vector is randomly generated as the negative space vector. The positive vector's value range is (0, 1), and the negative vector's value range is (-1, 0).

[0052] like Figure 2 As shown, u1, u2, v1, and v2 form a molecular-toxicity pathway network. The following describes the calculation process of its positive space vector using node u2 as an example:

[0053] For node u2, first calculate the cosine similarity z1 of the positive space vectors between u2 and node v1 (u2, v1), which are connected to each other, and then use the LeakReLU activation function to perform nonlinear calculations to obtain the processed cosine similarity p1.

[0054] in:

[0055]

[0056] In the formula, α is a constant, which is taken as 0.15 here;

[0057] Multiplying p1 by the initial positive space vector u2+ of node u2, and then adding the resulting modified vector p1 to the initial positive space vector of u2, we get the following: Figure 2As shown in the diagram, u2+* is used to calculate the cosine similarity z2 between the positive space vectors of node pair (u2, v2) formed by u2 and its connected node v2. This cosine similarity p2, after processing with the LeakReLU activation function, is multiplied by the initial positive space vector u2+ of node u2. The resulting modified vector p2 is then summed with u2+* to obtain the updated positive space vector of node u2. If u2 has multiple adjacent nodes v, the same calculation as for node v1 is performed, and the positive space vector of node u2 is updated by summing the results. The negative space vector of u2 is also updated in the same way. When updating one node, the other nodes remain unchanged. The generation of the positive and negative space vectors and the calculation of the cosine similarity both use known functions.

[0058] After all node pairs have been calculated, the final positive and negative space vectors for each node are obtained; that is, the positive space vector of each node is the sum of the positive relationships between that node and its neighboring nodes, and the negative space vector of each node is the sum of the negative relationships between that node and its neighboring nodes. The feature calculation process for a node is as follows: Figure 2 As shown;

[0059] The last k1-dimensional positive space vector and k2-dimensional negative space vector updated for each node are concatenated horizontally to obtain a (k1+k2)-dimensional vector, which is used as the embedding feature of that node.

[0060] The embedding features of each node in each node pair are horizontally concatenated to obtain the embedding features of that node pair, resulting in a (k1+k2)*2 dimensional concatenation vector x.

[0061] The classifier is trained using the training set data, and the logistic regression function h is used. θ (x) Perform binary classification on the concatenated vector x of the node pairs obtained in the training set to obtain the toxicity score of the edge connection for each pair of nodes, where:

[0062]

[0063]

[0064] To evaluate prediction performance, this invention selected AUC as the basic evaluation metric. A higher AUC value indicates better performance. These metrics were calculated using the scikit-learn package in Python.

[0065] The trained model was tested and validated using test set data. At the same time, the present invention was compared with other advanced methods on the same dataset. The test results are shown in Table 1.

[0066] Table 1 shows the performance of the toxicity prediction method based on link symbol prediction on the tox21 dataset.

[0067]

[0068]

[0069] Among the methods shown in the table, XGraphBoost uses graph neural networks to extract features and Xgboost for classification; Chemception uses image feature extraction methods to predict toxicity; AttentiveFP proposes a new molecular graph representation method to predict molecular toxicity; ECFP+Xgboost uses classic ECFP fingerprints as molecular features and Xgboost for classification. It can be clearly seen from the table that the method of this invention outperforms the other methods mentioned above in terms of performance.

[0070] In summary, this invention can be used for the prediction of molecular toxicity pathways. The well-known implementation methods and characteristics described above are not elaborated upon here. It should be noted that those skilled in the art can make several modifications without departing from this invention, and these modifications should also be considered within the scope of protection of this invention. These modifications will not affect the implementation effect or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of the claims, and the specific embodiments described in the specification are used to interpret the content of the claims.

Claims

1. A method for predicting molecular toxicity pathways based on link symbol prediction, characterized in that, Includes the following steps: 1) Convert the data format of the tox21 molecular toxicity pathway into the file format required for link symbol prediction: "Molecular Node Toxicity Pathway Node Node Pair Label"; where the label for toxicity is 1 and the label for non-toxicity is 0. The positive and negative samples of the toxicity link prediction data are divided into training and test sets in the same proportion, and the positive and negative samples are merged to form training and test sets respectively. 2) Extract and calculate the feature vector of each node and node pair from step 1). 2.1) For each molecule-toxicity node pair (u,v), randomly generate a k1-dimensional vector U for nodes u and v respectively. p V p and k2-dimensional vector U N V N ; Will U p V p Mapped to the positive space, they serve as the positive space vectors for nodes u and v, respectively; Will U N V N Mapped to the negative space, they serve as the negative space vectors for nodes u and v, respectively; 2.2) For each node u, calculate the cosine similarity z of the positive and negative space vectors of the node pair (u,v) formed by nodes u and v that are connected to it. + z - The processed cosine similarity p is obtained by nonlinear calculation using the LeakReLU activation function. + p - ,in: In the formula, α is a constant; p calculated each time + p - The positive space vectors U corresponding to node u respectively p and negative space vector U N Multiplying them, the resulting vectors correspond to U respectively. p and U N Perform summation to obtain the updated U p and U N ; Perform the same computation as for node u on all nodes v to obtain the updated V. p and V N ; After calculating all (u,v) node pairs, we obtain the final positive space vector and negative space vector updated for each node; 2.3) The last k1-dimensional positive space vector and k2-dimensional negative space vector updated for each node are concatenated horizontally to obtain a (k1+k2)-dimensional vector, which is used as the embedding feature of the node. The embedding features of each node in each node pair are horizontally concatenated to obtain the embedding features of that node pair, resulting in a (k1+k2)*2 dimensional concatenation vector x. 3) Use the concatenated vector x of the node pairs obtained in step 2) to train and validate the classifier. The classifier is trained using the training set data, employing the logistic regression function h. θ (x) Perform binary classification on the concatenated vector x of the node pairs obtained in the training set to obtain the toxicity score of the edge connection for each pair of nodes, where: In the formula, θ T The logistic regression function h obtained during training θ The parameters of (x); Using the loss function to apply the logistic regression function h θ The parameters of (x) are updated to minimize the difference between the predicted and actual values; The classifier was validated using a test set with an updated logistic regression function h. θ (x) Perform binary classification on the spliced ​​vector x of node pairs in the test set to obtain the toxic pathway prediction effect of the test set and verify the effect of the classifier; 4) Use the classifier trained and validated in step 3) to predict molecular toxicity pathways.

2. The molecular toxicity pathway prediction method based on link symbol prediction according to claim 1, characterized in that, Step 1) specifically involves: Molecules and toxic pathways are represented by nodes. All molecules are considered as one type of node, and all toxic pathways are considered as another type of node. The nodes are labeled with serial numbers and tags to form a toxicity link prediction data format of "molecule node toxic pathway node node pair tag". Among them, the tag for toxic is 1, and the tag for non-toxic is 0. The positive and negative samples of the toxicity link prediction data are first divided into training and test sets in an 8:2 ratio. Then, 8 positive and negative samples are merged to form the training set, and 2 positive and negative samples are merged to form the test set.

3. The molecular toxicity pathway prediction method based on link symbol prediction according to claim 1 or 2, characterized in that: In step 3), the loss function is constructed as follows: in, i∈(0,1), y[i]∈[0,1] indicates whether label i belongs to a compound, and 4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

5. An electronic device, characterized in that: Includes processors and computer-readable storage media; The computer-readable storage medium stores a computer program that, when executed by the processor, performs the steps of the method according to any one of claims 1 to 3.

Images