A small sample-based drug inhibition ability prediction method

Abstract

This invention discloses a method for predicting drug inhibitory efficacy based on small samples. The original feature vector is processed through an attention matrix to obtain a smaller, more representative feature vector. These feature vectors are then concatenated pairwise, with new labels defined as 0 or 1 based on the original labels. The concatenated feature vectors are then passed through a classifier model consisting of convolutional and fully connected layers. The feature vectors of the compound whose inhibitory efficacy needs to be detected, along with the feature vectors of other compounds with known inhibitory efficacy values ​​from a table, are fed into the trained model. By comparing the prediction results, the relative magnitudes are obtained, and the range of inhibitory efficacy values ​​for the compound to be detected is determined by looking up the table. This invention provides a simpler way to obtain the range of inhibitory efficacy values ​​than directly predicting the precise value of inhibitory efficacy through regression, and requires fewer samples.

Description

Technical Field

[0001] This invention relates to the field of computation-assisted drug research technology, and in particular to a method for predicting drug inhibition capacity based on small samples. Background Technology

[0002] Tumor cells exhibit significant metabolic differences compared to normal cells. Even under aerobic conditions, they rely on glycolysis to generate adenosine triphosphate (ATP) for cell division, proliferation, and invasion. Consequently, the consumption of nicotinamide adenine dinucleotide (NAD+), a crucial coenzyme in the glycolytic pathway, is significantly increased in tumor cells. NAD in tumor cells primarily originates from major rescue pathways. Nicotinamide phosphoribosyltransferase (NAMPT), the rate-limiting enzyme in the NAD rescue synthesis pathway, has been found to be upregulated in cancer cells, increasing NAD+ levels and promoting cellular biological processes such as glycolysis. It can also indirectly promote the binding of reduced glutathione (GSH) to reactive oxygen species (ROS) by influencing and upregulating the activity of NAD-dependent enzymes, thereby enabling tumor cells to adapt to relatively energy-deficient environments and enhancing their resistance to adverse conditions. Therefore, NAMPT, as a potential anti-tumor therapeutic target, and NAMPT inhibitors may offer excellent opportunities for cancer treatment.

[0003] Despite advances in biotechnology and understanding of biological systems, drug design remains a lengthy, expensive, difficult, and inefficient process. One of the most labor-intensive tasks in drug discovery is selecting compounds from a library for experimental evaluation. Computer-aided drug design plays a crucial role in accelerating the discovery of potential lead compounds and optimizing their structures. In computer-aided drug design, machine learning is widely used to train models to predict target properties, including their potency and toxicity: for example, using support vector machines combined with pharmacophore modeling to distinguish compound activity; using random forests combined with open-source data to predict the activity of related enzymes; using Bayesian algorithms to process and organize large molecular datasets, and utilizing Bayesian classifiers for mutant prediction. Various deep learning methods also discover implicit relationships in compound data through general learning processes: for example, using deep convolutional neural networks to predict molecular chemical properties; using multi-task autoencoder deep neural networks to predict the inhibitory potency of cytochromes; using deep learning methods and random forests to compute predicted bioactivity; and training two-layer neural networks to predict model quality using structural features, physicochemical properties, and residual contact.

[0004] Classification or regression using machine learning or deep learning often requires a large amount of data to obtain good results, and the same is true for drug prediction. Since compound molecules typically have thousands of molecular properties corresponding to thousands of dimensions of feature space, the inhibitory efficacy of a target may only be related to a few features. An excessively large feature space can lead to the model failing to grasp the key points during training and resulting in excessively long training times. Furthermore, experimental data from different institutions or using different experimental methods can vary significantly, making it impossible to aggregate them into a large sample for predicting molecular properties. Therefore, the number of samples in datasets used to predict the inhibitory efficacy of a specific molecule is usually very small. Additionally, due to differences in measurement tools and the biological cellular environment in which the compound molecules exist, the datasets used for training may contain errors. Therefore, small sample sizes, high feature space dimensions, and systematic errors during measurement often result in low accuracy for drug prediction using machine learning or deep learning. Experiments using regression methods to accurately predict drug inhibitory efficacy also suffer from the inability to distinguish compounds with similar inhibitory efficiencies. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of the present invention is to provide a method for predicting drug inhibition ability based on small samples. Regression prediction using machine learning or deep learning methods suffers from problems such as small sample size, large feature dimension, and regression model not emphasizing the relationship between compounds.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting drug inhibition capacity based on small samples, comprising the following steps:

[0007] S1: Reconstruct the feature vector set ω in the small-scale suppression effectiveness dataset into a feature vector set ω′ by dimensionality reduction using a parameter matrix;

[0008] S2: Expand the feature vector set ω′ obtained in S1 into a feature vector set ω″, and at the same time expand the label set ι in the small-scale suppression effectiveness dataset into data label ι';

[0009] S3: Put the expanded feature vector set ω″ and data label ι' into a classifier consisting of a three-layer perceptron, and use the cross-entropy loss function to perform supervised binary classification training on the classifier;

[0010] S4: Optimize the formula for the three-layer perceptron using the mini-batch gradient descent method;

[0011] S5: Create a table of N compound molecules with known inhibition efficacy values ​​and their feature vectors. Put the feature vectors of the compound molecules whose inhibition efficacy needs to be tested and the feature vectors of other compound molecules with known inhibition efficacy values ​​in the table into the trained model. By comparing the prediction results, the relative magnitude of the inhibition efficacy of the compound molecule to be tested and other compound molecules with known inhibition efficacy values ​​can be obtained. Then, by looking up the table, the range of the inhibition efficacy value of the compound to be tested can be known.

[0012] The method for dimensionality reduction and reconstruction of S1 includes the following steps:

[0013] The feature space is reconstructed as

[0014]

[0015] Transform the feature space dimension of the original dataset from Δ dimension to ε1 dimension, where ε1 < Δ.

[0016] in Let ω' represent the attention matrix of size Δ×ε1, and let ω' represent the new feature space after the feature space transformation. This indicates that the reconstructed feature space has ε samples, each sample consisting of a vector of length ε1, where each element is a real number.

[0017] The step of expanding the feature vector set ω′ in S2 includes:

[0018] The feature vector set is expanded to

[0019] Data tags expanded to

[0020]

[0021]

[0022] The feature vectors are processed using the `concat` operation: `concat` represents a concatenation operation, which concatenates each element in the feature vector set pairwise to form a new feature vector; the labels are processed by comparing the label values ​​of two elements: when the value of the τ-th element in the label set is less than or equal to the value of the χ-th element, the new label is... The value is 0 if it is 0, otherwise it is 1.

[0023] ω τ ′ represents the τ-th element in ω', ω χ ′ represents the x-th element in ω', ι τ Represents the τ-th element in ι.χ Let represent the x-th element in ι, and T represent the transpose of the matrix. The represented feature space set has (ε×(ε-1)) elements, each of which is a vector of length (2×(ε1)), and each element in the vector is a real number. This indicates that the expanded label set has (ε×(ε-1)) elements, each of which is a real number.

[0024] The formula for the three-layer perceptron in S3 is as follows:

[0025]

[0026] The cross-entropy loss function is

[0027]

[0028] The feature vector is passed through a three-layer perceptron to obtain the predicted label value. Then, the predicted label value and the true label value are used to calculate the loss through the cross-entropy loss function to obtain the predicted loss value of the model.

[0029] Where Φ represents the training formula for the classifier, ω τχ "" represents the τχ-th element in ω, softmax and sigmoid represent activation functions in deep learning, and π (1) ξ represents the weight parameters of the first-layer perceptron. (1) π represents the bias parameter of the first-layer perceptron. (2) ξ represents the weight parameters of the second-layer perceptron. (2) π represents the bias parameter of the second-layer perceptron. (3) ξ represents the weight parameters of the third-layer perceptron. (3) The parameters represent the bias parameters of the third-layer perceptron; Loss represents the loss value, log represents the logarithmic function, and ι represents the bias parameter of the third-layer perceptron. τχ This represents the τx-th element in ι'.

[0030] The gradient calculation formulas for the weight and bias values ​​of the classifier in S4 are as follows:

[0031]

[0032]

[0033] The update formulas for the weights and biases are:

[0034]

[0035]

[0036] Where i = 1, 2, 3, Δξ(i) ξ represents the gradient of the bias value of the i-th layer perceptron. (i) Δπ represents the bias value of the i-th layer perceptron. (i) π represents the gradient of the weights of the i-th perceptron layer. (i) Represents the weight values ​​of the i-th layer perceptron. This represents the operation of taking partial derivatives, and α represents the learning rate, which is used to control the speed of network training.

[0037] In each round of iterative training, the partial derivative of the loss function with respect to each parameter in the network is calculated. Then, the partial derivative of each parameter in the network is subtracted from the learning rate to obtain the updated parameters. The network is then trained until it converges and the loss value tends to stabilize.

[0038] Implementing the embodiments of the present invention has the following beneficial effects: The present invention can expand the drug dataset to a square multiple of the original dataset size, and can also transform the regression problem of predicting the accurate value of drug inhibitory efficacy into a classification problem of predicting the relative size of drug inhibitory efficacy, thereby enabling the accuracy of predicting the relative size of inhibitory efficacy to reach a high level, thus obtaining a more accurate range of inhibitory efficacy values. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0041] This invention provides a method for predicting drug inhibition capacity based on small samples. The small-scale inhibition efficacy dataset used consists of a feature vector set ω and a label set ι. ω is a set of feature vectors from the dataset. ω i Let represent the feature vector of the i-th sample in the dataset, ε represent the number of samples in the dataset, Δ represent the dimension of the feature vector, T represent the transpose of the matrix, and R represent the set of real numbers. ι is the set of labels (inhibition power) in the dataset. ε represents the number of samples in the dataset, ι i Represents the label value (suppression effectiveness) of the i-th sample in the dataset, T represents the transpose of the matrix, and R represents the set of real numbers.

[0042] S1: Since small sample high-dimensional datasets often have the characteristic of data sparsity, that is, the sample points are often very sparse and only a few features are non-zero, this invention designs to reduce the dimensionality of the original high-dimensional feature vectors.

[0043] The feature space is reconstructed as

[0044]

[0045] This step transforms the feature space dimension of the original dataset from Δ dimensions to ε1 dimensions, where ε1 < Δ.

[0046] in Let ω' represent the attention matrix of size Δ×ε1, and let ω' represent the new feature space after the feature space transformation. This indicates that the reconstructed feature space has ε samples, each sample consisting of a vector of length ε1, where each element is a real number.

[0047] S2: Due to the small number of samples, this invention designs a data augmentation method, which is carried out in the following way.

[0048] The feature vector set is expanded to Data tags expanded to

[0049]

[0050]

[0051] The feature vectors are processed using the `concat` operation: `concat` represents a concatenation operation, which concatenates each element in the feature vector set pairwise to form a new feature vector. Labels are processed by comparing the label values ​​of two elements: when the value of the `τ`-th element in the label set is less than or equal to the value of the `χ`-th element, the new label is `ι`. τχ The value is 0 if it is 0, otherwise it is 1.

[0052] ω τ ′ represents the τ-th element in ω', ω χ ′ represents the x-th element in ω'. ι τ Represents the τ-th element in ι. χ represents the x-th element in ι. T represents the transpose of the matrix. The represented feature space set has (ε×(ε-1)) elements, each of which is a vector of length (2×(ε1)) and each element in the vector is a real number. This indicates that the expanded label set has (ε×(ε-1)) elements, each of which is a real number.

[0053] S3: Feed the features and labels into a classifier consisting of a three-layer perceptron, and use the cross-entropy loss function to perform supervised binary classification training on the classifier.

[0054] Φ(ω τχ ") = softmax(ν (3) +π (3) (sigmoid(ξ (2) +π (2) (ξ (1) +π (1) ω τχ ")))),#(4)

[0055]

[0056] The feature vector is passed through a three-layer perceptron to obtain the predicted label value. Then, the predicted label value and the true label value are used to calculate the loss using the cross-entropy loss function to obtain the model's predicted loss value.

[0057] Equation (4) is the formula for a three-layer perceptron, and equation (5) is the formula for the cross-entropy loss function. Φ represents the training formula for the classifier, and ω τχ "" represents the τχ-th element in ω, softmax and sigmoid represent activation functions in deep learning, and π (1) ξ represents the weight parameters of the first-layer perceptron. (1) π represents the bias parameter of the first-layer perceptron. (2) ξ represents the weight parameters of the second-layer perceptron. (2) π represents the bias parameter of the second-layer perceptron. (3) ξ represents the weight parameters of the third-layer perceptron. (3) This represents the bias parameters of the third-layer perceptron. Loss represents the loss value, log represents the logarithmic function, and ι represents the bias parameter. τχ This represents the τx-th element in ι'.

[0058] S4: Optimize equation (4) using mini-batch gradient descent.

[0059]

[0060]

[0061]

[0062]

[0063] The gradient calculation formulas for the weights and biases of the classifier are (6) and (7), and the update formulas for the weights and biases are (8) and (9). In each round of iterative training, the partial derivative of the loss function with respect to each parameter in the network is calculated, and then the partial derivative of each parameter in the network is multiplied by the learning rate to obtain the updated parameters. Then the network continues to be trained until the network converges and the loss value tends to stabilize.

[0064] i = 1, 2, 3, Δξ (i) ξ represents the gradient of the bias value of the i-th layer perceptron. (i) Δπ represents the bias value of the i-th layer perceptron. (i) π represents the gradient of the weights of the i-th perceptron layer. (i) This represents the weight value of the i-th layer perceptron. This represents the operation of calculating partial derivatives. α represents the learning rate, which controls the speed of network training. The larger α is, the faster the perceptron weights and biases are updated.

[0065] S5: When it is necessary to check whether the inhibitory efficacy of a compound molecule meets the experimental requirements, a table can be designed. This table has N rows, representing N compound molecules with known inhibitory efficacy values. The table has two columns: the first column represents the feature vector of the compound molecule, and the second column represents the inhibitory efficacy value of the compound molecule. Next, the feature vector of the compound molecule whose inhibitory efficacy needs to be tested, along with the feature vectors of other compound molecules with known inhibitory efficacy values ​​in the table, are fed into a trained model. By observing the model's prediction results, the relative magnitude of the inhibitory efficacy of the compound molecule to be tested compared to other compound molecules with known inhibitory efficacy values ​​can be determined. Then, by looking up the table, the range of the inhibitory efficacy value of the compound to be tested can be determined, thereby determining whether the compound can be used as an experimental sample.

[0066] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for predicting drug inhibition capacity based on small samples, characterized in that, Includes the following steps: S1: Reconstruct the feature vector set ω in the small-scale suppression effectiveness dataset into a feature vector set ω′ by dimensionality reduction using a parameter matrix; S2: Expand the feature vector set ω′ obtained in S1 into a feature vector set ω″, and at the same time expand the label set ι in the small-scale suppression effectiveness dataset into data label ι'; S3: Put the expanded feature vector set ω″ and data label ι' into a classifier consisting of a three-layer perceptron, and use the cross-entropy loss function to perform supervised binary classification training on the classifier; S4: Optimize the formula for the three-layer perceptron using the mini-batch gradient descent method; S5: Create a table of N compound molecules with known inhibition efficacy values ​​and their feature vectors. Put the feature vectors of the compound molecules whose inhibition efficacy needs to be tested and the feature vectors of other compound molecules with known inhibition efficacy values ​​in the table into the trained model. By comparing the prediction results, the relative magnitude of the inhibition efficacy of the compound molecule to be tested and other compound molecules with known inhibition efficacy values ​​can be obtained. Then, by looking up the table, the range of the inhibition efficacy value of the compound to be tested can be obtained.

2. The method for predicting drug inhibition capacity based on small samples according to claim 1, characterized in that, The dimensionality reduction and reconstruction method of S1 Includes the following steps: The feature space is reconstructed as Transform the feature space dimension of the original dataset from Δ dimension to ε1 dimension, where ε1 < Δ. in Let ω' represent the attention matrix of size Δ×ε1, and let ω' represent the new feature space after the feature space transformation. This indicates that the reconstructed feature space has ε samples, each sample consisting of a vector of length ε1, where each element is a real number.

3. The method for predicting drug inhibition capacity based on small samples according to claim 1, characterized in that, The step of expanding the feature vector set ω′ in S2 includes: The feature vector set is expanded to Data tags expanded to oh τχ ″=concat(ω τ ′,ω χ ′); The feature vectors are processed using the `concat` operation: `concat` represents a concatenation operation, which concatenates each element in the feature vector set pairwise to form a new feature vector; the labels are processed by comparing the label values ​​of two elements: when the value of the τ-th element in the label set is less than or equal to the value of the χ-th element, the new label is ι. τχ The value is 0 if it is 0, otherwise it is 1. ω τ ′ represents the τ-th element in ω', ω χ ′ represents the x-th element in ω', ι τ Represents the τ-th element in ι. χ Let represent the x-th element in ι, and T represent the transpose of the matrix. The represented feature space set has (ε×(ε-1)) elements, each of which is a vector of length (2×(ε1)), and each element in the vector is a real number. This indicates that the expanded label set has (ε×(ε-1)) elements, each of which is a real number.

4. The method for predicting drug inhibition capacity based on small samples according to claim 1, characterized in that, The formula for the three-layer perceptron in S3 is: F(ω τχ ″)=softmax(ξ (3) +p (3) (sigmoid(ξ (2) +p (2) (x) (1) +p (1) oh τχ ″)))); The cross-entropy loss function is The feature vector is passed through a three-layer perceptron to obtain the predicted label value. Then, the predicted label value and the true label value are used to calculate the loss through the cross-entropy loss function to obtain the predicted loss value of the model. Where Φ represents the training formula for the classifier, ω τχ "" represents the τχ-th element in ω, softmax and sigmoid represent activation functions in deep learning, and π (1) ξ represents the weight parameters of the first-layer perceptron. (1) π represents the bias parameter of the first-layer perceptron. (2) ξ represents the weight parameters of the second-layer perceptron. (2) π represents the bias parameter of the second-layer perceptron. (3) ξ represents the weight parameters of the third-layer perceptron. (3) The parameters represent the bias parameters of the third-layer perceptron; Loss represents the loss value, log represents the logarithmic function, and ι represents the bias parameter of the third-layer perceptron. τχ This represents the τx-th element in ι'.

5. The method for predicting drug inhibition capacity based on small samples according to claim 4, characterized in that, The gradient calculation formulas for the weight and bias values ​​of the classifier in S4 are as follows: The update formulas for the weights and biases are: Where i = 1, 2, 3, Δξ () ξ represents the gradient of the bias value of the i-th layer perceptron. () Δπ represents the bias value of the i-th layer perceptron. () π represents the gradient of the weights of the i-th perceptron layer. () Represents the weight values ​​of the i-th layer perceptron. This represents the operation of taking partial derivatives, and α represents the learning rate, which is used to control the speed of network training. In each round of iterative training, the partial derivative of the loss function with respect to each parameter in the network is calculated. Then, the partial derivative of each parameter in the network is subtracted from the learning rate to obtain the updated parameters. The network is then trained until it converges and the loss value tends to stabilize.

Images