Anticancer drug-drug interaction prediction method and system

Abstract

The invention discloses an anti-cancer drug-drug interaction prediction method and system, and relates to the technical field of drug effect prediction, and the method comprises the steps: constructing a known drug pair and a to-be-detected drug pair of a target anti-cancer drug combination, determining a plurality of known coded drug feature pairs of each known drug pair and a plurality of to-be-detected coded drug feature pairs of each to-be-detected drug pair, determining a trained pulse drug prediction model based on each known coded drug feature pair, and inputting each to-be-detected coded drug feature pair; after anti-cancer pulse features and to-be-detected pulse features of each to-be-detected coded drug feature pair are generated based on a neuron multiplexing mechanism through a pulse twinning pulse convolutional network, the anti-cancer pulse features and the to-be-detected pulse features are correspondingly spliced into pulse joint features, and the pulse joint features are input into an IF pulse sensor to output corresponding sensing features; and performing weighted fusion on the sensing features by adopting a feature weighting device, inputting the fused sensing features into a pulse multi-layer sensor to perform drug-drug interaction prediction, and outputting a target drug action result. And the drug-drug interaction prediction reliability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of drug action prediction, and in particular to a method and system for predicting anticancer drug-drug interaction. Background Art

[0002] The human body is a highly complex system. In the process of treating diseases, traditional single drug therapy usually targets specific molecular pathways or physiological mechanisms, and often cannot fully solve the problem when dealing with complex diseases such as cancer or chronic diseases. Combination drug therapy can more effectively deal with complex diseases through drug-drug interactions (DDIs) produced by multi-target and multi-mechanism synergy, and usually uses lower drug doses to reduce drug-induced toxicity and side effects. At the same time, the form of drug combination can effectively reduce or overcome drug resistance. However, not all drug combination DDIs are beneficial. Some DDIs may cause antagonistic effects and reduce the overall efficacy, so it is necessary to predict the DDIs of drug combinations.

[0003] Traditional drug-drug interaction prediction methods use laboratory validation methods, but they require a large number of expensive experiments and clinical trials, and must meet strict regulatory requirements to ensure their safety and effectiveness, which makes it difficult to meet the needs of practical applications. To address the above problems, existing technologies use a substructure extraction module similar to Transformer to obtain a fixed number of representative vectors related to various substructure patterns of drug molecules. Then, the interaction strength between two drug substructures will be captured by a similarity-based interaction module. However, this method relies on a single drug feature, which can easily lead to information loss, reduced model generalization ability, and inability to fully capture complex interactions between drugs, thereby reducing prediction reliability. Summary of the invention

[0004] The present invention provides an anticancer drug-drug interaction prediction method and system, which solves the technical problem that the prior art relies on a single drug feature when predicting drug-drug interaction, which easily reduces the reliability of prediction.

[0005] The first aspect of the present invention provides a method for predicting anticancer drug-drug interaction, comprising:

[0006] The target anticancer drug and the associated known interacting drugs and the interacting drugs to be tested are used, and the corresponding combinations are known drug pairs and drug pairs to be tested;

[0007] Acquire multiple drug features of the target anticancer drug, each of the known interacting drugs, and each of the interacting drugs to be tested, encode them respectively, and construct corresponding encoded drug features;

[0008] Respectively classifying and aggregating the coded drug features of each of the known drug pairs and each of the drug pairs to be tested, and determining a plurality of known coded drug feature pairs of each of the known drug pairs and a plurality of coded drug feature pairs to be tested of each of the drug pairs to be tested;

[0009] Based on the training of each of the known coded drug feature pairs, a trained pulse drug prediction model is determined, and each of the coded drug feature pairs to be tested is input; the trained pulse drug prediction model includes a pulse twin pulse convolutional network, an IF pulse perceptron, a feature weighter and a pulse multilayer perceptron;

[0010] Extracting features of each of the tested coded drug feature pairs through a pulse twin pulse convolutional network based on a neuron multiplexing mechanism to generate anti-cancer pulse features and tested pulse features of each of the tested coded drug feature pairs;

[0011] After each of the anti-cancer pulse features and the corresponding pulse features to be tested are spliced ​​into a pulse joint feature, the feature is input into an IF pulse sensor for dimensionality reduction nonlinear mapping, and the corresponding perception feature is output;

[0012] The feature weighter is used to perform weighted fusion on each of the perception features to determine the fusion feature, and the fusion feature is input into a pulse multilayer perceptron to predict drug-drug interaction and output the target drug action result.

[0013] Optionally, the pulse twin pulse convolution network includes two sub-pulse convolution networks, and the sub-pulse convolution network includes a LIF pulse convolution block and a flattening layer; the pulse twin pulse convolution network extracts features of each of the tested coded drug feature pairs based on a neuron multiplexing mechanism to generate anti-cancer pulse features and pulse features to be tested of each of the tested coded drug feature pairs, including:

[0014] Input any of the tested coded drug feature pairs into the pulse twin pulse convolutional network;

[0015] After extracting the encoded drug features of the target anticancer drug in the encoded drug feature pair to be tested and resetting the membrane potential in real time through the cascaded LIF pulse convolution block in the first sub-pulse convolution network, flattening is performed using a flattening layer to determine the anticancer pulse features;

[0016] The LIF neurons in the first sub-pulse convolution network are multiplexed in real time by the cascaded LIF pulse convolution block in the second sub-pulse convolution network, and the encoded drug features of the interacting drugs to be tested in the encoded drug feature pair to be tested are extracted and the membrane potential is reset in real time, and then flattened by the flattening layer to determine the pulse features to be tested;

[0017] The next pair of coded drug features to be tested is input into the pulse twin pulse convolution network, and the LIF neurons in the second sub-pulse convolution network are multiplexed in real time through the cascaded LIF pulse convolution blocks in the first sub-pulse convolution network until the anti-cancer pulse features and the pulse features to be tested of each pair of coded drug features to be tested are determined.

[0018] Optionally, the LIF pulse convolution block includes a one-dimensional convolution layer, a LIF neuron and a pooling layer; the processing process of the LIF pulse convolution block includes:

[0019] Perform convolution operation on the convolution input features of the input LIF pulse convolution block through a one-dimensional convolution layer to determine the convolution output features;

[0020] After nonlinear mapping of the convolution output features is performed using LIF neurons, pooling dimension reduction is performed based on the pooling layer to determine the pooling output features.

[0021] Optionally, the step of using the feature weighter to perform weighted fusion on each of the perceptual features to determine a fusion feature includes:

[0022] Concatenating the perceptual features to generate a combined feature;

[0023] Performing a linear transformation on the combined features through a fully connected layer to determine a weight matrix;

[0024] A softmax activation function is used to perform a nonlinear transformation on the weight matrix to determine a plurality of weight element values;

[0025] After performing a Hadamard product on each weight element value and the corresponding perceptual feature, the fusion feature is concatenated and output.

[0026] Optionally, the step of determining a trained pulse drug prediction model based on each of the known encoded drug feature pairs includes:

[0027] Input each of the known encoded drug features into the pulse drug prediction model to be trained to predict drug-drug interactions, and output the predicted drug action results;

[0028] Calculating a loss function value based on the predicted drug action result and the corresponding actual drug action result;

[0029] If the loss function value has not converged, the model parameters are updated using the gradient descent method until the loss function value converges, thereby determining the trained pulse drug prediction model.

[0030] Optionally, the calculation process of the loss function value includes:

[0031] ;

[0032] In the formula, is the loss function, For the real drug effect results, To predict the outcome of drug action, The category label for the drug effect , is the total number of categories.

[0033] The second aspect of the present invention provides an anticancer drug-drug interaction prediction system, comprising:

[0034] A drug data acquisition module is used to use the target anticancer drug and the associated known interacting drugs and the interacting drugs to be tested, and the corresponding combinations are known drug pairs and test drug pairs;

[0035] A data preprocessing module, used for obtaining multiple drug features of the target anticancer drug, each of the known interacting drugs and each of the interacting drugs to be tested and encoding them respectively, and constructing corresponding encoded drug features;

[0036] A drug pair combination module, used for classifying and aggregating the coded drug features of each of the known drug pairs and each of the drug pairs to be tested, and determining a plurality of known coded drug feature pairs of each of the known drug pairs and a plurality of coded drug feature pairs to be tested of each of the drug pairs to be tested;

[0037] A model training module, used to train and determine a trained pulse drug prediction model based on each of the known coded drug feature pairs, and input each of the coded drug feature pairs to be tested; the trained pulse drug prediction model includes a pulse twin pulse convolutional network, an IF pulse perceptron, a feature weighter and a pulse multilayer perceptron;

[0038] A pulse extraction module, used for extracting features of each of the tested coded drug feature pairs through a pulse twin pulse convolution network based on a neuron multiplexing mechanism, and generating anti-cancer pulse features and tested pulse features of each of the tested coded drug feature pairs;

[0039] A dimension reduction extraction module is used to splice each of the anti-cancer pulse features with the corresponding pulse features to be tested into a pulse joint feature, input it into an IF pulse sensor for dimension reduction nonlinear mapping, and output the corresponding perception feature;

[0040] The prediction module is used to use the feature weighter to perform weighted fusion on each of the perception features to determine the fusion feature, and input the fusion feature into the pulse multilayer perceptron to predict drug-drug interaction and output the target drug action result.

[0041] A third aspect of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the anticancer drug-drug interaction prediction method as described in any one of the above items.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the anticancer drug-drug interaction prediction method as described in any one of the above items.

[0043] A fifth aspect of the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the anticancer drug-drug interaction prediction method as described in any one of the above.

[0044] It can be seen from the above technical solutions that the present invention has the following advantages:

[0045] The above scheme of the present invention provides a method for predicting anticancer drug-drug interaction, comprising: using a target anticancer drug and associated known interaction drugs and a to-be-tested interaction drug, and correspondingly combining them into known drug pairs and to-be-tested drug pairs; obtaining a plurality of drug features of the target anticancer drug, each known interaction drug and each to-be-tested interaction drug and encoding them respectively, and constructing corresponding coded drug features; classifying and aggregating the coded drug features belonging to each known drug pair and each to-be-tested drug pair, respectively, and determining a plurality of known coded drug feature pairs of each known drug pair and a plurality of to-be-tested coded drug feature pairs of each to-be-tested drug pair; training based on each known coded drug feature pair to determine a trained pulse drug prediction model, and inputting the trained pulse drug prediction model into each to-be-tested drug pair. Test coding drug feature pairs; the trained pulse drug prediction model includes a pulse twin pulse convolution network, an IF pulse sensor, a feature weighter and a pulse multilayer perceptron; the pulse twin pulse convolution network is used to extract features of each coding drug feature pair to be tested based on a neuron multiplexing mechanism to generate anticancer pulse features and test pulse features of each coding drug feature pair to be tested; after splicing each anticancer pulse feature with the corresponding test pulse feature into a pulse joint feature, the feature is input into the IF pulse sensor for dimensionality reduction nonlinear mapping, and the corresponding perception feature is output; the feature weighter is used to perform weighted fusion on each perception feature to determine the fusion feature, and the feature is input into the pulse multilayer perceptron for drug-drug interaction prediction, and the target drug action result is output. Based on the above scheme, after encoding the drug features into binary data based on multiple drug features, the sparseness problem related to the encoded drug features is improved by introducing spiking neurons in the pulse drug prediction model. The pulse twin pulse convolutional network is used to enhance the internal expression ability of drug features, and the model complexity is reduced through the neuron reuse mechanism. Based on the feature weighter, the adaptive weighted fusion of different drug pairs is guided for fine-grained adjustment, which effectively captures the interactions between drugs and improves the reliability of drug-drug interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0047] Figure 1 A flowchart of the steps of a method for predicting anticancer drug-drug interaction provided by an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of the structure of a pulse twin pulse convolutional network provided in an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of a flow chart of a feature weighting device provided in an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of data flow of a method for predicting anticancer drug-drug interaction provided by an embodiment of the present invention;

[0051] Figure 5 A structural block diagram of an anticancer drug-drug interaction prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The embodiments of the present invention provide an anticancer drug-drug interaction prediction method and system, which are used to solve the technical problem that the prior art relies on a single drug feature when predicting drug-drug interaction, which easily reduces the reliability of prediction.

[0053] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] See also Figure 1 , Figure 1 A flowchart of the steps of a method for predicting anticancer drug-drug interaction provided by an embodiment of the present invention.

[0055] The present invention provides a method for predicting anticancer drug-drug interaction, comprising:

[0056] Step 101: Use a target anticancer drug and associated known interacting drugs and interacting drugs to be tested to form corresponding combinations into known drug pairs and test drug pairs.

[0057] It should be noted that after selecting a certain anticancer drug as the target anticancer drug, multiple known interacting drugs with its known drug-drug interaction results and multiple interacting drugs to be tested with its drug-drug interaction results to be predicted are extracted from the Drugbank database, and the target anticancer drug and any known interacting drug are combined into a known drug pair, and the target anticancer drug and any interacting drug to be tested are combined into a test drug pair.

[0058] It can be understood that, in specific implementation, in order to facilitate visual display, multiple known drug pairs can be used to form a known drug pair set, and multiple test drug pairs can be used to form a test drug pair set, and the set is saved in the form of a list, whereby each row in the known drug pair set represents each known drug pair with a known drug-drug interaction result label, including information such as drug name, drug id, interaction type and interaction description, and each row in the test drug pair set represents each test drug pair with a test drug-drug interaction result label, including information such as drug name and drug id.

[0059] Step 102: Obtain multiple drug features of the target anticancer drug, each known interaction drug, and each to-be-tested interaction drug, and encode them respectively to construct corresponding encoded drug features.

[0060] It should be noted that drug features refer to features that reflect the drug's efficacy, pharmacology and other property information. Through databases such as Drugbank, Uniport and KEGG, multiple drug features of each drug in the target anticancer drugs, known interacting drugs and interacting drugs to be tested can be obtained, such as chemical substructure features, target features, pathway features and enzyme features. The drug features are then encoded and converted into binary data, thereby representing encoded drug features. In one implementation, the chemical substructure features are represented by simplifying the molecular linear input specification and encoded with a molecular fingerprint, while the target features, pathway features and enzyme features can be encoded using a one-dimensional embedding vector.

[0061] It is understandable that, in a specific implementation, in order to better display different dimensions of data, the coded drug features of each drug can be stored in the corresponding view matrix according to the type of coded drug features. In the formula Number the view. To encode drug characteristics, For the A drug.

[0062] Step 103: Classify and aggregate the coded drug features of each known drug pair and each drug pair to be tested, and determine a plurality of known coded drug feature pairs of each known drug pair and a plurality of coded drug feature pairs to be tested of each drug pair to be tested.

[0063] It should be noted that, in the present embodiment, taking each drug pair as a unit, the multiple coded drug features associated with the two drugs of each known drug pair are classified and aggregated according to type, such as combining the chemical substructure feature of the target anticancer drug in the known drug pair and the chemical substructure feature of the known interacting drug into a known coded drug feature pair, thereby determining multiple known coded drug feature pairs of each known drug pair. Similarly, the multiple coded drug features associated with the two drugs of any drug pair to be tested are classified and combined, such as combining the target feature of the target anticancer drug in the drug pair to be tested and the target feature of the interacting drug to be tested into a coded drug feature pair to be tested, thereby obtaining multiple coded drug feature pairs to be tested for each drug pair to be tested.

[0064] Step 104, determine a trained pulse drug prediction model based on each known coded drug feature pair, and input each coded drug feature pair to be tested; the trained pulse drug prediction model includes a pulse twin pulse convolutional network, an IF pulse perceptron, a feature weighter and a pulse multilayer perceptron.

[0065] In a specific embodiment, determining a trained pulse drug prediction model based on each known encoded drug feature pair includes:

[0066] Input each known encoded drug feature pair into the pulse drug prediction model to be trained to predict drug-drug interactions, and output the predicted drug action results;

[0067] Calculate the loss function value based on the predicted drug action results and the corresponding actual drug action results;

[0068] If the loss function value has not converged, the gradient descent method is used to update the model parameters until the loss function value converges to determine the trained pulse drug prediction model.

[0069] It should be noted that, since the form of binary data limits the continuous representation of features, there may be a problem that the drug prediction model cannot be optimized using gradient information. In this embodiment, the introduction of pulse neurons to build a pulse drug prediction model is considered. The introduction of pulse neurons can effectively solve the sparsity problem related to binary coding, and nonlinearity is incorporated into the model so that it can approximate any function through learning.

[0070] In the specific implementation, each known coded drug feature pair is input into the pulse drug prediction model to be trained for prediction to generate a predicted drug action result, and a loss function is defined to calculate the loss function value based on the predicted drug action result and the corresponding real drug action result, and the loss function value is judged for convergence. If it does not converge, the gradient descent method is used to update the model parameters with the goal of minimizing the loss function value until the loss function value converges, thereby determining the trained pulse drug prediction model, and then, each coded drug feature pair to be tested is input into the trained pulse drug prediction model for processing;

[0071] In one implementation, the loss function is defined as multi-class cross entropy: , is the loss function, For the real drug effect results, To predict the outcome of drug action, The category label for the drug effect , is the total number of categories.

[0072] Step 105: extract features of each tested coded drug feature pair through a pulse twin pulse convolutional network based on a neuron multiplexing mechanism to generate anti-cancer pulse features and tested pulse features of each tested coded drug feature pair.

[0073] In a specific embodiment, the pulse twin pulse convolution network includes two sub-pulse convolution networks, and the sub-pulse convolution network includes a LIF pulse convolution block and a flattening layer; step 105 includes the following sub-steps:

[0074] S11, inputting any coded drug feature pair to be tested into the pulse twin pulse convolutional network;

[0075] S12, extracting the encoded drug features of the target anticancer drug in the tested encoded drug feature pair and resetting the membrane potential in real time through the cascaded LIF pulse convolution block in the first sub-pulse convolution network, and then flattening the features using a flattening layer to determine the anticancer pulse features;

[0076] S13, multiplexing the LIF neurons in the first sub-pulse convolution network in real time through the cascaded LIF pulse convolution block in the second sub-pulse convolution network, extracting the encoded drug features of the interacting drugs to be tested in the encoded drug feature pair to be tested and resetting the membrane potential in real time, and then flattening through the flattening layer to determine the pulse features to be tested;

[0077] S14. Input the next tested coded drug feature pair into the pulse twin pulse convolution network, and multiplex the LIF neurons in the second sub-pulse convolution network in real time through the cascaded LIF pulse convolution block in the first sub-pulse convolution network until the anti-cancer pulse features and the pulse features to be tested of each tested coded drug feature pair are determined.

[0078] In a more specific embodiment, the LIF spike convolution block includes a one-dimensional convolution layer, a LIF neuron and a pooling layer; the processing process of the LIF spike convolution block includes:

[0079] Perform convolution operation on the convolution input features of the input LIF pulse convolution block through a one-dimensional convolution layer to determine the convolution output features;

[0080] After nonlinear mapping of the convolution output features using LIF neurons, pooling dimensionality reduction is performed based on the pooling layer to determine the pooling output features.

[0081] It should be noted that in the two sub-spiking convolutional networks of the spike twin convolutional network, the parameters of the two are shared, ensuring that two similar input features are converted into two similar feature representations by the corresponding sub-spiking convolutional networks, and the neuron reuse mechanism refers to the LIF neurons in the spike twin convolutional network being alternately reused in the two sub-spiking convolutional networks, thereby reducing the number of neurons, reducing storage requirements and the complexity of hardware implementation;

[0082] like Figure 2 As shown in the figure, each sub-pulse convolution network includes a LIF pulse convolution block and a flattening layer. Each LIF pulse convolution block includes a one-dimensional convolution layer, a LIF neuron and a pooling layer. When the encoded drug features of the target anticancer drug in any of the encoded drug pairs to be tested or the encoded drug features of the interacting drugs to be tested are input into the corresponding sub-pulse convolution network:

[0083] 1) Feature extraction of the encoded drug features of the target anticancer drug as the convolution input features is performed through the one-dimensional convolution layer of the first LIF spike convolution block of the first sub-spiked convolution network: ,in, , where For the drug pair Drugs, For the drug pair The convolution output features of each drug, is the convolution kernel weight, Input features for convolution, is the convolution bias;

[0084] 2) After determining the convolution output features, pulses are generated through LIF neurons: , where is the time step , For the LIF neurons, For the LIF neuron time steps The membrane potential, For the LIF neuron time steps The membrane potential, is the attenuation factor, is the threshold voltage, is the time step The pulse When , otherwise ;

[0085] 3) The impulses generated by LIF neurons After pooling operation: , where For the pooling operation, Output features for pooling;

[0086] 4) Clear the membrane potential in the LIF neuron to complete the reset and restore it to its initial state: ; In the formula, For the The membrane potential of each LIF neuron, is the initial membrane potential;

[0087] 5) On the one hand, the pooled output features in step 3) are continuously input into the next LIF pulse convolution block, and feature processing is performed according to steps 1) to 4), and then input into the flattening layer to perform the flattening operation to obtain the final anti-cancer pulse features. At the same time, the LIF neurons reset in the first LIF pulse convolution block and the second LIF pulse convolution block in the first sub-pulse convolution network are reused by the first LIF pulse convolution block and the second LIF pulse convolution block of the second sub-pulse convolution network in turn, and feature processing is performed according to steps 1) to 5), thereby obtaining the pulse feature to be tested ;

[0088] 6) Input the next tested coding drug feature pair into the pulse twin pulse convolution network, and reuse the LIF neurons after the second sub-pulse convolution network is reset in the previous time step in the first sub-pulse convolution network, and perform feature processing according to the above steps 1) to 5) until the anti-cancer pulse features and the pulse features to be tested of each tested coding drug feature pair are determined.

[0089] Step 106: After each anti-cancer pulse feature is spliced ​​with the corresponding pulse feature to be measured as a pulse joint feature, it is input into the IF pulse sensor for dimensionality reduction nonlinear mapping, and the corresponding perception feature is output.

[0090] It should be noted that the characteristic pairs of each coded drug to be tested Merge into a one-dimensional column vector to obtain the pulse joint feature After that, it is input to the IF neurons with half the number of corresponding elements in a fully connected manner for feature processing, thereby achieving dimensionality reduction, and then the membrane potential is updated in the IF output neurons of the IF pulse sensor: , where For the IF output neurons, For the IF output neuron time step The membrane potential, is the time step, For the The input of each IF output neuron; according to the membrane potential Whether the threshold voltage is reached , the IF output neuron will trigger the corresponding pulse, and the perception feature can be obtained by combining all the pulse results; since simple series connection only connects feature vectors, it lacks the ability to explore the nonlinear relationship between features, and the pulse sensor can extract more abstract and complex features in feature fusion by simulating complex nonlinear dynamics, and because it transmits information based on discrete spike signals, it can better filter noise and focus on meaningful interactive information.

[0091] Step 107: Use a feature weighter to perform weighted fusion on each perception feature to determine a fusion feature, and input the fusion feature into a pulse multilayer perceptron to predict drug-drug interaction and output the target drug action result.

[0092] It should be noted that the self-attention mechanism is used based on the feature weighter to automatically assign weights to multiple perceptual features, thereby performing weighted fusion to determine the fusion features, and the fusion features are used as the input of the spiking multilayer perceptron (Spiking MLP) to predict drug-drug interactions, and the target drug action results are output as the classification results. , is a spiking multilayer perceptron, are the parameters of the pulsed multilayer perceptron, To predict the outcome of drug action;

[0093] It is understandable that if Figure 4 As shown, drug effects include but are not limited to synergistic, toxic, and antagonistic effects.

[0094] Exemplarily, when the spiking multilayer perceptron is a LIF spiking multilayer perceptron using LIF neurons, the specific processing process includes:

[0095] ;

[0096] In the formula, The LIF spike multilayer perceptron Layer LIF neurons, For the LIF spiking multilayer perceptron weights of layer LIF neurons, For the The pulses of layer LIF neurons, For the LIF-spiking multilayer perceptron bias for layer LIF neurons, For the Linear characteristics of layer LIF neurons, For the Layer LIF neuron time step The membrane potential, For the Layer LIF neuron time step The pulse, is the total time step, is the total number of LIF neuron layers in the LIF spike multilayer perceptron; at each time step , the membrane potential of the LIF neuron will be updated according to its input pulses and synaptic weights. The pulse emission of each LIF neuron in the last layer is regarded as a vote for a specific category. By counting the number of pulses emitted by each LIF neuron in all time steps, the total support of each category can be obtained. After normalization, a probability distribution is obtained, which represents the probability that the input sample belongs to each category.

[0097] In a specific implementation, a feature weighter is used to perform weighted fusion on each perceptual feature to determine a fusion feature, including:

[0098] Concatenate each perceptual feature to generate a combined feature;

[0099] Perform linear transformation on the combined features through the fully connected layer to determine the weight matrix;

[0100] The softmax activation function is used to perform nonlinear transformation on the weight matrix to determine the values ​​of multiple weight elements;

[0101] After performing the Hadamard product between each weight element value and the corresponding perceptual feature, the fusion feature is concatenated and output.

[0102] It should be noted that the self-attention mechanism is used based on the feature weighter to automatically assign weights to multiple perceptual features, such as Figure 3 As shown, the feature weighter includes a fully connected layer and a softmax activation function; first, there are multiple perceptual features under each view number, and each perceptual feature is concatenated into a single combined feature. , is a combination of features, For View Then, a fully connected layer is used to transform the combined features into a weight matrix , is the weight matrix, is the fully connected weight, is the full connection bias, and then the softmax activation function is used to calculate the weight element value , Output index , is the view index, For the View output index The weight element value of For input index , and then, the Hadamard product is performed with the perceptual features to re-weight the feature representation of each view , For View The weighted perceptual features of is the weight element value matrix No. Finally, the newly weighted feature representations are aggregated to form the fused output , For fusion features.

[0103] In order to verify the effectiveness of this solution, a performance comparison experiment was conducted with the existing prediction and classification model. It can be understood that there are multiple evaluation indicators for performance comparison:

[0104] ;

[0105] ;

[0106] ;

[0107] In the formula, is the accuracy, True Positive is the number of samples correctly predicted as positive. True Negative is the number of samples correctly predicted as negative. False Positive is the number of samples that are incorrectly predicted as positive. False Negative is the number of samples that are incorrectly predicted as negative. is the accuracy, is the recall rate;

[0108] Taking accuracy as an evaluation indicator, the performance comparison results are shown in Table 1:

[0109] Table 1 Performance comparison results

[0110]

[0111] In this scheme, the LIF neuron parameters used in drug-drug interaction prediction for the anticancer drugs Erlotinib and Gefitnib include membrane time constant set to 20 milliseconds, membrane resistance set to 10 megohms, resting potential set to -70 millivolts, threshold potential set to -50 millivolts, and reset potential set to -65 millivolts. In the spike twin spike convolution network, the parameters of the convolution layer are set to 16 input channels, 32 output channels, 3 convolution kernel size, 1 step size, padding mode "same", and no bias term, and the pooling type is maximum pooling, the pooling window size is 2, the step size is 2, and the padding mode is "valid".

[0112] As shown in Table 1, the anticancer drugs Erlotinib and Gefitnib were selected as target anticancer drugs, wherein 1019 relevant drug interaction data were collected for the drug Erlotinib, including 5 different types of DDIs; 1458 relevant drug interaction data were collected for the drug Gefitnib, including 8 different types of DDIs; According to the results of the accuracy index in Table 1, it can be seen that the accuracy of the method proposed in this embodiment in predicting drug-drug interactions is significantly better than other algorithms.

[0113] In an embodiment of the present invention, after encoding the drug features into binary data based on multiple drug features, the sparsity problem related to the encoded drug features is improved by introducing pulse neurons in the pulse drug prediction model, and the pulse twin pulse convolutional network is used to enhance the internal expression ability of the drug features. The neuron reuse mechanism can also reduce the complexity of the model, and guide the adaptive weighted fusion of different drug pairs based on the feature weighter for fine-grained adjustment, effectively capturing the interactions between drugs, and improving the generalization ability and prediction reliability.

[0114] See also Figure 5 , Figure 5 A structural block diagram of an anticancer drug-drug interaction prediction system provided by an embodiment of the present invention.

[0115] The present invention provides an anticancer drug-drug interaction prediction system, comprising:

[0116] The drug data acquisition module 501 is used to use the target anticancer drug and the associated known interaction drugs and the interaction drugs to be tested, and the corresponding combinations are known drug pairs and drug pairs to be tested;

[0117] A data preprocessing module 502 is used to obtain multiple drug features of the target anticancer drug, each known interaction drug and each interaction drug to be tested and encode them respectively to construct corresponding encoded drug features;

[0118] The drug pair combination module 503 is used to classify and aggregate the coded drug features of each known drug pair and each drug pair to be tested, and determine a plurality of known coded drug feature pairs of each known drug pair and a plurality of coded drug feature pairs to be tested of each drug pair to be tested;

[0119] The model training module 504 is used to train and determine a trained pulse drug prediction model based on each known coded drug feature pair, and input each coded drug feature pair to be tested; the trained pulse drug prediction model includes a pulse twin pulse convolution network, an IF pulse perceptron, a feature weighter and a pulse multilayer perceptron;

[0120] The pulse extraction module 505 is used to extract features of each pair of coded drug features to be tested through a pulse twin pulse convolution network based on a neuron multiplexing mechanism, and generate anticancer pulse features and pulse features to be tested of each pair of coded drug features to be tested;

[0121] Dimensionality reduction extraction module 506, used to splice each anti-cancer pulse feature with the corresponding pulse feature to be tested into a pulse joint feature, input it into the IF pulse sensor for dimension reduction nonlinear mapping, and output the corresponding perception feature;

[0122] The prediction module 507 is used to use a feature weighter to perform weighted fusion on each perception feature to determine a fusion feature, and input the fusion feature into a pulse multilayer perceptron to predict drug-drug interaction and output the target drug action result.

[0123] Optionally, the pulse twin pulse convolution network includes two sub-pulse convolution networks, and the sub-pulse convolution network includes a LIF pulse convolution block and a flattening layer; the pulse extraction module 505 is specifically used for:

[0124] Input any pair of drug features to be tested into the spike twin spike convolutional network;

[0125] After extracting the encoded drug features of the target anticancer drug in the tested encoded drug feature pair and resetting the membrane potential in real time through the cascaded LIF pulse convolution block in the first sub-pulse convolution network, the flattening layer is used to flatten the features to determine the anticancer pulse features;

[0126] The LIF neurons in the first sub-pulse convolution network are multiplexed in real time through the cascaded LIF pulse convolution block in the second sub-pulse convolution network, and the encoded drug features of the interacting drugs to be tested in the encoded drug feature pair to be tested are extracted and the membrane potential is reset in real time, and then flattened through the flattening layer to determine the pulse features to be tested;

[0127] The next tested coded drug feature pair is input into the pulse twin pulse convolution network, and the LIF neurons in the second sub-pulse convolution network are multiplexed in real time through the cascaded LIF pulse convolution blocks in the first sub-pulse convolution network until the anti-cancer pulse features and the pulse features to be tested of each tested coded drug feature pair are determined.

[0128] Optionally, the LIF spike convolution block includes a one-dimensional convolution layer, a LIF neuron and a pooling layer; the processing process of the LIF spike convolution block includes:

[0129] Perform convolution operation on the convolution input features of the input LIF pulse convolution block through a one-dimensional convolution layer to determine the convolution output features;

[0130] After nonlinear mapping of the convolution output features using LIF neurons, pooling dimensionality reduction is performed based on the pooling layer to determine the pooling output features.

[0131] Optionally, a feature weighter is used to perform weighted fusion on each perceptual feature to determine a fusion feature, including:

[0132] Concatenate each perceptual feature to generate a combined feature;

[0133] Perform linear transformation on the combined features through the fully connected layer to determine the weight matrix;

[0134] The softmax activation function is used to perform nonlinear transformation on the weight matrix to determine the values ​​of multiple weight elements;

[0135] After performing the Hadamard product between each weight element value and the corresponding perceptual feature, the fusion feature is concatenated and output.

[0136] Optionally, the trained pulse drug prediction model is determined based on each known encoded drug feature pair, including:

[0137] Input each known encoded drug feature pair into the pulse drug prediction model to be trained to predict drug-drug interactions, and output the predicted drug action results;

[0138] Calculate the loss function value based on the predicted drug action results and the corresponding actual drug action results;

[0139] If the loss function value has not converged, the gradient descent method is used to update the model parameters until the loss function value converges to determine the trained pulse drug prediction model.

[0140] Optionally, the loss function value calculation process includes:

[0141] ;

[0142] In the formula, is the loss function, For the real drug effect results, To predict the outcome of drug action, The category label for the drug effect , is the total number of categories.

[0143] An embodiment of the present invention further provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the anticancer drug-drug interaction prediction method as described in any of the above embodiments.

[0144] An embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the method for predicting anticancer drug-drug interaction as in any of the above embodiments are implemented.

[0145] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the anticancer drug-drug interaction prediction method as described in any of the above embodiments.

[0146] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0147] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0148] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0149] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0150] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0151] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting anticancer drug-drug interaction, characterized in that: include: The target anticancer drug and the associated known interacting drugs and the interacting drugs to be tested are used, and the corresponding combinations are known drug pairs and drug pairs to be tested; Acquire multiple drug features of the target anticancer drug, each of the known interacting drugs, and each of the interacting drugs to be tested, encode them respectively, and construct corresponding encoded drug features; Respectively classifying and aggregating the coded drug features of each of the known drug pairs and each of the drug pairs to be tested, and determining a plurality of known coded drug feature pairs of each of the known drug pairs and a plurality of coded drug feature pairs to be tested of each of the drug pairs to be tested; Based on the training of each of the known coded drug feature pairs, a trained pulse drug prediction model is determined, and each of the coded drug feature pairs to be tested is input; the trained pulse drug prediction model includes a pulse twin pulse convolutional network, an IF pulse perceptron, a feature weighter and a pulse multilayer perceptron; Extracting features of each of the tested coded drug feature pairs through a pulse twin pulse convolutional network based on a neuron multiplexing mechanism to generate anti-cancer pulse features and tested pulse features of each of the tested coded drug feature pairs; After each of the anti-cancer pulse features and the corresponding pulse features to be tested are spliced ​​into a pulse joint feature, the feature is input into an IF pulse sensor for dimensionality reduction nonlinear mapping, and the corresponding perception feature is output; The feature weighter is used to perform weighted fusion on each of the perception features to determine the fusion feature, and the fusion feature is input into a pulse multilayer perceptron to predict drug-drug interaction and output the target drug action result.

2. The method for predicting anticancer drug-drug interaction according to claim 1, characterized in that: The pulse twin pulse convolution network includes two sub-pulse convolution networks, and the sub-pulse convolution network includes a LIF pulse convolution block and a flattening layer; the pulse twin pulse convolution network extracts features of each of the tested coded drug feature pairs based on a neuron multiplexing mechanism to generate anti-cancer pulse features and pulse features to be tested of each of the tested coded drug feature pairs, including: Input any of the tested coded drug feature pairs into the pulse twin pulse convolutional network; After extracting the encoded drug features of the target anticancer drug in the encoded drug feature pair to be tested and resetting the membrane potential in real time through the cascaded LIF pulse convolution block in the first sub-pulse convolution network, flattening is performed using a flattening layer to determine the anticancer pulse features; The LIF neurons in the first sub-pulse convolution network are multiplexed in real time by the cascaded LIF pulse convolution block in the second sub-pulse convolution network, and the encoded drug features of the interacting drugs to be tested in the encoded drug feature pair to be tested are extracted and the membrane potential is reset in real time, and then flattened by the flattening layer to determine the pulse features to be tested; The next pair of coded drug features to be tested is input into the pulse twin pulse convolution network, and the LIF neurons in the second sub-pulse convolution network are multiplexed in real time through the cascaded LIF pulse convolution blocks in the first sub-pulse convolution network until the anti-cancer pulse features and the pulse features to be tested of each pair of coded drug features to be tested are determined.

3. The method for predicting anticancer drug-drug interaction according to claim 2, characterized in that: The LIF pulse convolution block includes a one-dimensional convolution layer, a LIF neuron and a pooling layer; the processing process of the LIF pulse convolution block includes: Perform convolution operation on the convolution input features of the input LIF pulse convolution block through a one-dimensional convolution layer to determine the convolution output features; After nonlinear mapping of the convolution output features is performed using LIF neurons, pooling dimension reduction is performed based on the pooling layer to determine the pooling output features.

4. The method for predicting anticancer drug-drug interaction according to claim 1, characterized in that: The step of using the feature weighter to weight and fuse the perceptual features to determine a fused feature includes: Concatenating the perceptual features to generate a combined feature; Performing a linear transformation on the combined features through a fully connected layer to determine a weight matrix; A softmax activation function is used to perform a nonlinear transformation on the weight matrix to determine a plurality of weight element values; After performing a Hadamard product on each weight element value and the corresponding perceptual feature, the fusion feature is concatenated and output.

5. The method for predicting anticancer drug-drug interaction according to claim 1, characterized in that: The step of determining a trained pulse drug prediction model based on each of the known encoded drug features comprises: Input each of the known encoded drug features into the pulse drug prediction model to be trained to predict drug-drug interactions, and output the predicted drug action results; Calculating a loss function value based on the predicted drug action result and the corresponding actual drug action result; If the loss function value has not converged, the model parameters are updated using the gradient descent method until the loss function value converges, thereby determining the trained pulse drug prediction model.

6. The method for predicting anticancer drug-drug interaction according to claim 5, characterized in that: The calculation process of the loss function value includes: ； In the formula, is the loss function, For the actual drug effect results, To predict the outcome of drug action, The category label for the drug effect , is the total number of categories.

7. An anticancer drug-drug interaction prediction system, characterized in that: include: A drug data acquisition module is used to use the target anticancer drug and the associated known interacting drugs and the interacting drugs to be tested, and the corresponding combinations are known drug pairs and test drug pairs; A data preprocessing module, used for obtaining multiple drug features of the target anticancer drug, each of the known interacting drugs and each of the interacting drugs to be tested and encoding them respectively, and constructing corresponding encoded drug features; A drug pair combination module, used for classifying and aggregating the coded drug features of each of the known drug pairs and each of the drug pairs to be tested, and determining a plurality of known coded drug feature pairs of each of the known drug pairs and a plurality of coded drug feature pairs to be tested of each of the drug pairs to be tested; A model training module, used to train and determine a trained pulse drug prediction model based on each of the known coded drug feature pairs, and input each of the coded drug feature pairs to be tested; the trained pulse drug prediction model includes a pulse twin pulse convolutional network, an IF pulse perceptron, a feature weighter and a pulse multilayer perceptron; A pulse extraction module, used for extracting features of each of the tested coded drug feature pairs through a pulse twin pulse convolution network based on a neuron multiplexing mechanism, and generating anti-cancer pulse features and tested pulse features of each of the tested coded drug feature pairs; A dimension reduction extraction module is used to splice each of the anti-cancer pulse features with the corresponding pulse features to be tested into a pulse joint feature, input it into an IF pulse sensor for dimension reduction nonlinear mapping, and output the corresponding perception feature; The prediction module is used to use the feature weighter to perform weighted fusion on each of the perception features to determine the fusion feature, and input the fusion feature into the pulse multilayer perceptron to predict drug-drug interaction and output the target drug action result.

8. A computer device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the anticancer drug-drug interaction prediction method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method for predicting anticancer drug-drug interaction according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method for predicting anticancer drug-drug interaction according to any one of claims 1 to 6 are implemented.

Images