A method for predicting drug inhibitory potency based on small data sets

Abstract

The embodiment of the present invention discloses a method for predicting the strength of drug inhibitory efficacy based on a small data set, which uses adaptive data expansion, reconstructs feature space, and relabels labels as relative size relationships of inhibitory efficacy, thereby converting the predicted accurate numerical value of inhibitory efficacy into a direct prediction of the relative size of inhibitory efficacy, and converting the accurate value prediction regression task into a simpler classification task. At the same time, the adaptive feature space reconstruction preserves the relative size relationship of the physicochemical properties between compounds, and more importantly, significantly weakens the inherent similarity between compounds with similar inhibitory efficacy. Using the data set processed by the present invention, the machine learning model can more easily distinguish similar compounds, and thereby screen out samples with better inhibitory efficacy.

Description

Technical Field

[0001] The present invention relates to the technical field of drug inhibitory efficacy screening, and in particular to a method for predicting drug inhibitory efficacy intensity based on a small data set. Background Art

[0002] As a rate-limiting enzyme in the salvage synthesis pathway of nicotinamide adenine dinucleotide (NAD), nicotinamide phosphoribosyltransferase (NAMPT) can effectively regulate cell biological processes in cancer cells, which makes cancer cells more susceptible to the cytotoxic effects of NAMPT inhibitors. Therefore, as a potential experimental target for anti-tumor therapy, NAMPT inhibitors play a vital role in the treatment of cancer diseases. Using drug screening methods to efficiently screen NAMPT inhibitors with better inhibitory efficacy has become one of the development directions of cancer disease treatment.

[0003] Traditionally, the quantification of inhibitory potency used in drug design is usually obtained by measuring the proliferation activity of specific cells or by using in vivo pharmaceutical experiments such as time-resolved fluorescence energy transfer (TR-FRET). This method achieves quantification of inhibitory potency by exciting fluorescent molecules with long decay times, such as lanthanides such as europium, samarium, and terbium, and detecting the fluorescence values. This method has the advantage of specificity. At the same time, the use of fluorescent molecules provides long-term stability and achieves immunity to interference from the autofluorescence of the compound itself.

[0004] Many in vitro computational prediction methods aim to prioritize chemical drug experiments for experimental research, using machine learning and other methods for model training and prediction. The basic idea of ​​this type of method is to predict physicochemical properties through regression models. These methods are based on data sets generated by biological interaction information generated by experimental verification, and use the physicochemical and structural properties of compounds as input to train computational models, which greatly speeds up the process of finding potential lead compounds for further pharmacological testing. Therefore, many regression prediction methods, such as random forests, support vector machines, and ensemble learning, have been widely used in the prediction of adverse drug reactions; at the same time, models based on fully connected deep neural networks, convolution-based models, attention-based models, multimodal-based methods, and generative deep learning methods have been widely used in the prediction of compound-protein relationships and physicochemical properties.

[0005] The drug screening process based on pharmaceutical experiments directly quantifies the inhibitory efficacy through biochemical experimental methods, but the process is challenging, time-consuming, lengthy and demanding. For example, the average enzyme competitive inhibition constant (Ki) is often measured by the TR-FRET method. This method requires the use of a chelate of europium atoms and antibodies, but this chelate becomes very unstable after being interfered by some complexes such as ethylenediaminetetraacetic acid (EDTA), and the experimental results will be seriously affected.

[0006] At the same time, many in vitro machine learning models for predicting compound-protein physicochemical relationships have been widely studied. However, using traditional regression models to solve such inhibitor prediction problems is difficult, and usually faces problems such as low prediction accuracy, difficulty in handling outliers, and difficulty in distinguishing compounds with similar inhibitory potency. Specifically, the data set used for training in inhibitory potency prediction has the following characteristics: 1) The number of samples is very small, usually less than 50; 2) The dimension of the feature space is very high, usually around 3000; 3) There are widespread unpredictable systematic errors from statistical or measurement tools in in vivo experiments in the data set, which can usually be regarded as non-Gaussian white noise. Based on the above three points, the traditional regression model dedicated to predicting the exact value of inhibitory potency cannot obtain enough training data, and is easily affected by the huge redundant information in the feature space, resulting in high training difficulty. Secondly, the trained model is easily affected by non-Gaussian noise, which makes the model more inclined to fit the distribution of noise, resulting in a decrease in the model prediction accuracy. Finally, this type of regression model has difficulty distinguishing compounds with similar inhibitory potency. The traditional regression model first predicts the exact value of the inhibitory potency, and then manually compares and screens. Therefore, a small regression error may lead to completely opposite results in distinguishing the magnitude of inhibitory potency, resulting in errors in drug screening tasks. Summary of the invention

[0007] The technical problem to be solved by the embodiment of the present invention is to provide a method for predicting the strength of drug inhibitory efficacy based on a small data set, which can effectively solve the problem of distinguishing the magnitude of drug inhibitory efficacy in a small data set.

[0008] In order to solve the above technical problems, an embodiment of the present invention provides a method for predicting the drug inhibitory efficacy intensity based on a small data set, comprising the following steps:

[0009] S1: Dataset expansion, which is to expand the small-scale inhibition effectiveness dataset X and its corresponding inhibition effectiveness Y;

[0010] S2: Build a model and define the robust loss function as a mixed generalized correlation entropy:

[0011]

[0012] in represents the generalized Gaussian density, represents the mathematical expectation, represents the equilibrium parameter of the two cores, and is a random variable; calculate the weight matrix of norm and adding the loss function, the optimization objective of the model can be summarized as:

[0013] ,in Represents a basic error function;

[0014] S3: Model training process, the data set obtained in S1 is passed to the model of S2 for training, and the training process is optimized using gradient descent;

[0015] S4: Prediction process: for a given compound, the relative magnitude of the inhibitory strength is determined. C Features p C and the corresponding inhibition strength y C Reconstruct and send the new features into the trained model to obtain the predicted target label; for the predicted label ,like , then the model predicts m Among the compounds i The inhibitory strength of the compound was higher than that of the compound C High; on the contrary, if , then the model predicts m Among the compounds i The inhibitory strength of the compound was higher than that of the compound C Low.

[0016] Wherein, the step S1 specifically includes:

[0017] Each data label pair in the original dataset ,in i and j The value ranges from 1 to n . Use the following formula to calculate the new label value

[0018]

[0019] in represents a relative preference relationship, Represents the new label after reconstruction, and Represents the original data set. i andj The labels of the samples;

[0020] Traverse the sample feature pairs in the original data set , then traverse the features and No. k components, calculate the following formula to obtain the number ij The new features of k Quantity:

[0021]

[0022] in Represents the reconstructed features. k Quantity, and Represents the original data set. i and j The characteristics of the samples, and Represents the original data set. i and j The features of the sample k components, after looping through and calculating , get each new feature , and further obtain a new feature space containing new features:

[0023]

[0024] is a local similarity measure for two random variables. and , the similarity measure is defined as:

[0025]

[0026] in, Represents two random variables and The dimension of is the kernel parameter, and are random variables and No. i A value;

[0027] Substituting the above formula into the traditional classification learning model, we can get the general learning framework after adaptive data expansion:

[0028]

[0029] in Represents the data set expansion operation function, represents the target mapping learned by the model, represents the weight vector, represents the weight coefficient, , as well as Represents the original data set. i and j The features of each sample and the reconstructed labels.

[0030] Wherein, in step S2:

[0031] The generalized Gaussian density is defined as:

[0032]

[0033] in The representative model is i The error at each sample;

[0034] Calculate the weight matrix of norm and add the loss function, for a matrix , and each element in the matrix ,That The norm is defined as:

[0035] .

[0036] Wherein, the step S3 further includes:

[0037] Using the hinge loss function as the basic error function, the hinge loss function is defined as:

[0038]

[0039] in is the estimated value of the model output, t is the true value of the data, i.e. the label.

[0040] The label obtained in step S1 And features The new data set is fed into the model proposed in the present invention for training, and the training process is optimized using gradient descent.

[0041] Wherein, in step S4, for a given compound for which the relative magnitude of the inhibition strength is to be determined, C Features p C and the corresponding inhibition strength y C The steps to perform a rebuild include:

[0042] Follow step S1 to first construct a new target label ; Then construct new features , the new features Send it to the trained model to get the predicted target label .

[0043] Implementing the embodiments of the present invention has the following beneficial effects:

[0044] 1. Traditional inhibitory potency quantification is time-consuming and laborious, and the calculation method has low accuracy due to insufficient training data, and it is difficult to distinguish compounds with similar inhibitory potency. The present invention proposes an adaptive data set expansion method, which effectively expands the size of the data set and greatly weakens the similarity between similar samples, transforming the difficult initial accurate value regression problem into a simple classification problem.

[0045] 2. Traditional classification models are very sensitive to outliers and tend to fit the distribution of outliers. The robust classification model proposed in the present invention greatly reduces the adverse effects of outliers on the model and realizes feature selection during training. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the process structure of the method of the present invention. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0048] like Figure 1 As shown, a drug inhibitory efficacy intensity prediction method based on a small data set of the present invention is based on a small-scale inhibitory efficacy data set. and its corresponding inhibitory efficacy As an example, the present invention implements the relative magnitude prediction of the inhibition strength according to the following steps:

[0049] S1: Dataset expansion process. For a small-scale inhibitory effectiveness dataset and its corresponding inhibitory efficacy , firstly, an adaptive data set expansion operation is performed. The expansion operation is divided into data label reconstruction and feature space reconstruction.

[0050] First, reconstruct the data labels. Traverse each data label pair in the original data set ,in i and j The value ranges from 1 to n . Use the following formula to calculate the new label value:

[0051] ,#(1)

[0052] in represents a relative preference relationship, Represents the new label after reconstruction, and Represents the original data set. i and j The label of each sample is the suppression strength. By calculating formula (1) in the loop traversal process, the continuous label suppression strength in the original dataset is Reconstructed into a new dataset In particular, the labels in the new dataset are discrete or .

[0053] Next, traverse the sample feature pairs in the original data set , then traverse the features and No. k components, calculate the following formula to obtain the number ij The new features of k Quantity:

[0054] ,#(2)

[0055] in Represents the reconstructed features. k Quantity, and Represents the original data set. i and j The characteristics of each sample are called molecular descriptors. and Represents the original data set. i and j The features of the sample k After looping through and calculating formula (2), each new feature is obtained , and further obtain a new feature space containing new features In formula (2), is a local similarity measure for two random variables. and , the similarity measure is defined as:

[0056] ,#(3)

[0057] in, Represents two random variables and The dimension of is the kernel parameter, and are random variables and No. i The local similarity metric only considers the relationship between a pair of compounds, so for each pair of samples in the data set, the feature space reconstruction method proposed in the present invention is adaptive.

[0058] Substituting formula (2) into the traditional classification learning model (see formula (11)), we get the general learning framework after adaptive data expansion:

[0059] ,#(4)

[0060] in Represents the data set expansion operation function, represents the target mapping learned by the model, represents the weight vector, represents the weight coefficient, , as well as Represents the original data set. i and j The features of each sample and the labels reconstructed using formula (1).

[0061] The present invention uses formula (2) as the operation function. The framework disclosed by formula (4) can be used for various small sample inhibition strength prediction tasks. Depending on the classifier used, the performance of the framework is also different. Therefore, the present invention proposes a high-performance robust classifier as a better choice for predicting inhibition strength in the framework of formula (4).

[0062] S2: Model building process. The general learning framework of formula (4) in step 1 has the form of "loss function + regularization term". In this step, according to this form, the present invention proposes a new combined learning form of robust loss function and an embedded real-time feature selection regularization term. First, the robust loss function is defined as a mixed generalized correlation entropy:

[0063] ,#(5)

[0064] in represents the generalized Gaussian density, represents the mathematical expectation, represents the equilibrium parameter of the two kernels. Given two random variables and , the generalized Gaussian density is defined as:

[0065] ,#(6)

[0066] in represents the gamma function, is the scale parameter, is the kernel parameter. For a finite data set , by using the sample mean estimate, formula (5) can be rewritten as:

[0067]

[0068] in The representative model is i The error at each sample.

[0069] The present invention utilizes the hinge loss function ( ) as the basic error function, the hinge loss function can be defined as:

[0070]

[0071] in is the estimated value of the model output, t is the true value of the data, i.e. the label.

[0072] Finally, the label obtained in step S1 And features As a new data set, it is fed into the model proposed by the present invention for training, and the training process is optimized by gradient descent. The next step will summarize the prediction of the relative size of the inhibition strength using the model proposed by the present invention.

[0073] S5: Prediction process. After the trained model is obtained, for a given compound whose relative inhibitory strength is to be determined, C ,in C Has characteristics And the corresponding inhibition strength , first reconstruct the features. Similar to step S1, C Features With a number of ( m Characteristics of compounds with known inhibitory strength (for example) Specifically, according to step S1, a new target tag is first constructed. ,in It can be obtained by formula (1); then construct a new feature ,in It can be obtained by formula (2). Send it to the trained model to get the predicted target label Then, the compound can be obtained based on the prediction result and formula (1): C and mSpecifically, for the predicted labels ,like , then the model predicts m Among the compounds i The inhibitory strength of the compound was higher than that of the compound C High; on the contrary, if , then the model predicts m Among the compounds i The inhibitory strength of the compound was higher than that of the compound C So far, the model proposed by the present invention has completed the in vitro prediction of the relative inhibitory strength of compounds by computational methods, and realized drug screening based on the inhibitory strength.

[0074] Compared with the existing methods, the present invention has the following advantages:

[0075] The existing pharmaceutical experimental methods for measuring inhibitory efficacy are very time-consuming and costly. The present invention gives full play to the advantages of the computational model and greatly saves experimental expenses.

[0076] The use of traditional regression models to predict inhibitory efficacy faces the problems of small data sets and high training difficulty. The data set size obtained by expanding the method proposed in this invention is , much larger than the original dataset size , the expansion of training data greatly reduces the difficulty of model training. At the same time, as shown in formula (2), the adaptive feature space reconstruction based on similarity measurement retains and alleviates the differences between the molecular descriptors of the compounds, weakens the intrinsic similarity between samples with similar inhibitor potency, makes it easier for the model to distinguish sample pairs with similar inhibitor potency, and greatly improves the ability to screen drugs based on inhibitor potency.

[0077] The traditional classification model can be summarized as the following optimization problem:

[0078] ,#(11)

[0079] Traditional classification models use mean square error as the loss function. However, the minimum mean square error is very sensitive to outliers, which means it is difficult to achieve the effect of being immune to noise. The classification model based on mutual entropy aims to solve the problem of sensitivity to outliers. This type of classification model can be summarized as optimizing the following problems:

[0080] ,#(12)

[0081] As shown in formula (12), as a local statistic, the second-order mutual entropy has a certain ability to be insensitive to outliers. However, due to the fixed order, mutual entropy is difficult to obtain better noise immunity. The classification model based on generalized mutual entropy can be defined as:

[0082] ,#(13)

[0083] This type of model gains the ability to measure statistics of any order by changing the order of the Gaussian kernel. However, the performance of this type of model depends largely on the kernel parameters. , fixed kernel parameters will make the model lose flexibility. Therefore, the mixed generalized mutual entropy proposed in this invention, as shown in formula (7), obtains a certain kernel parameter by introducing two generalized Gaussian kernels. Due to the insensitive ability, the two kernel parameters are able to approach some single kernel parameters that are difficult to involve artificially, thus achieving better classification performance.

[0084] As shown in formula (11), the traditional classification model uses As a regularization term. However, in the inhibitory potency prediction task, each compound has thousands of molecular descriptor features, and the high feature dimension makes the training task very difficult. Therefore, the present invention proposes to use The norm is used as the regularization term in the “loss function + regularization term” framework to obtain real-time feature selection capabilities. As shown in formula (8), different from norm, The norm has a stricter constraint on the number of parameters. This makes the optimization process more The model with the norm is more likely to produce sparse solutions, thereby achieving the purpose of feature selection.

[0085] The above disclosure is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for predicting drug inhibitory potency based on a small data set, characterized in that: The following steps are involved: S1: Dataset expansion: perform data set expansion operations on the small-scale inhibition effectiveness dataset X and its corresponding inhibition effectiveness Y; each data label in the original dataset ,in i and j The value ranges from 1 to n , use the following formula to calculate the new label value in represents a relative preference relationship, Represents the new label after reconstruction, and Represents the original data set. i and j The labels of the samples; Traverse the sample feature pairs in the original data set , then traverse the features and No. k components, calculate the following formula to obtain the number ij The new features of k Quantity: in Represents the reconstructed features. k Quantity, and Represents the original data set. i and j The characteristics of the samples, and Represents the original data set. i and j The features of the sample k components, after looping through and calculating , get each new feature , and further obtain a new feature space containing new features: is a local similarity measure for two random variables. and , the similarity measure is defined as: in, Represents two random variables and The dimension of is the kernel parameter, and are random variables and No. i A value; Will Substituting into the traditional classification learning model, we get the general learning framework after adaptive data expansion: in Represents the data set expansion operation function, represents the target mapping learned by the model, represents the weight vector, represents the weight coefficient, , as well as Represents the original data set. i and j The features of each sample and the reconstructed labels; S2: Build a model and define the robust loss function as a mixed generalized correlation entropy: in represents the generalized Gaussian density, represents the mathematical expectation, represents the equilibrium parameter of the two cores, and is a random variable; calculate the weight matrix of norm and adding the loss function, the optimization objective of the model can be summarized as: ,in Represents a basic error function; S3: Model training process, the data set obtained in S1 is passed to the model of S2 for training, and the training process is optimized using gradient descent; S4: Prediction process: for a given compound, the relative magnitude of the inhibitory strength is determined. C Features p C and the corresponding suppression strength yC are reconstructed, and the new features are sent to the trained model to obtain the predicted target label; for the predicted label ,like =1, the model predicts m Among the compounds i The inhibitory strength of the compound was higher than that of the compound C High; on the contrary, if = 0, the model predicts m Among the compounds i The inhibitory strength of the compound was higher than that of the compound C Low.

2. The method for predicting drug inhibitory potency based on a small data set according to claim 1, characterized in that: In step S2: The generalized Gaussian density is defined as: in The representative model is i The error at each sample; Calculate the weight matrix of norm and add the loss function, for a matrix , and each element in the matrix ,That The norm is defined as: 。 3. The method for predicting drug inhibitory potency based on a small data set according to claim 2, characterized in that: The step S3 further comprises: Using the hinge loss function as the basic error function, the hinge loss function is defined as: in is the estimated value of the model output, t is the true value of the data, i.e. the label; The label obtained in step S1 And features As a new data set is passed into the model for training, the training process is optimized using gradient descent.

4. The method for predicting drug inhibitory potency based on a small data set according to claim 3, characterized in that: In step S4, for a given compound for which the relative magnitude of the inhibition strength is to be determined, C Features p The steps of reconstructing C and the corresponding inhibition strength yC include: Follow step S1 to first construct a new target label ; Then construct new features , the new features Send it to the trained model to get the predicted target label .

Images