CN114678141A

CN114678141A - Method, apparatus and medium for predicting drug-pair interaction relationship

Info

Publication number: CN114678141A
Application number: CN202210263559.8A
Authority: CN
Inventors: 唐继军; 窦明亮; 郗文辉; 郭菲
Original assignee: Shenzhen Technology University
Current assignee: Shenzhen Technology University
Priority date: 2022-03-17
Filing date: 2022-03-17
Publication date: 2022-06-28
Also published as: WO2023173823A1

Abstract

The embodiment of the application is applicable to the technical field of biomedicine, and provides a method, equipment and a medium for predicting interaction relation of a drug pair, wherein the method comprises the following steps: acquiring a target text; the target text comprises a plurality of medicines, and each medicine appears at least once in the target text; respectively determining a comprehensive entity representation corresponding to each medicine, wherein the comprehensive entity representation is used for describing semantic information of the corresponding medicine at each position in the target text; determining a fused entity representation of a drug pair from the integrated entity representation of the two drugs in the drug pair for any drug pair of the plurality of drugs; and predicting the interaction relation of the drug pairs according to the fused entity representation of the drug pairs. By adopting the method, the prediction accuracy of the interaction relation of each drug pair in the document can be improved.

Description

Method, apparatus and medium for predicting drug-pair interaction relationship

Technical Field

The application belongs to the technical field of biomedicine, and particularly relates to a method, equipment and medium for predicting interaction relation of a drug pair.

Background

Drug-Drug Interaction (DDI) prediction is an important research area in Drug surveillance, playing a crucial role in virtual Drug screening, patient treatment regimens, treatment outcomes, and patient safety studies.

All DDI predictions at present are based on sentence-level relationships. Firstly, the abstract or the text part of each scientific literature about the medicine description is divided into a plurality of sentences, and each sentence contains a plurality of medicines. Therefore, it is known to input each drug pair into a specific network for processing, and to predict the interaction relationship between each drug pair.

However, the above-mentioned DDI prediction methods all deal with sentence-level drug pairs, and in practice, the interaction relationship between a drug pair is often determined by a plurality of sentences. Therefore, in the prior art, the prediction accuracy of the interaction relation of each drug pair in the document is low.

Disclosure of Invention

The embodiment of the application provides a method, equipment and a storage medium for predicting interaction relation of drug pairs, and can solve the problem of low accuracy rate of predicting the interaction relation of each drug pair in a document.

In a first aspect, the present embodiments provide a method for predicting drug-pair interaction relationship, the method including:

acquiring a target text; the target text comprises a plurality of medicines, and each medicine appears at least once in the target text;

Respectively determining a comprehensive entity representation corresponding to each medicine, wherein the comprehensive entity representation is used for describing semantic information of the corresponding medicine at each position in the target text;

for any drug pair in the plurality of drugs, determining a fused entity representation of the drug pair from the synthetic entity representation of the two drugs in the drug pair;

and predicting the interaction relation of the drug pairs according to the fused entity representation of the drug pairs.

In a second aspect, the present application provides a device for predicting drug pair interaction relationship, the device comprising:

the acquisition module is used for acquiring a target text; the target text comprises a plurality of medicines, and each medicine appears at least once in the target text;

the comprehensive entity representation determining module is used for respectively determining a comprehensive entity representation corresponding to each medicine, and the comprehensive entity representation is used for describing semantic information of the corresponding medicine at each position in the target text;

a fused entity representation determining module for determining a fused entity representation of a drug pair according to a comprehensive entity representation of two drugs in the drug pair for any drug pair in the plurality of drugs;

and the prediction module is used for predicting the interaction relation of the drug pairs according to the fused entity representation of the drug pairs.

In a third aspect, an embodiment of the present application provides a terminal device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the method according to the first aspect is implemented.

In a fourth aspect, embodiments of the present application provide a computer-readable storage medium, in which a computer program is stored, and the computer program, when executed by a processor, implements the method of the first aspect.

In a fifth aspect, an embodiment of the present application provides a computer program product, which, when run on a terminal device, causes the terminal device to execute the method of the first aspect.

Compared with the prior art, the embodiment of the application has the beneficial effects that: by acquiring the corresponding comprehensive entity representation of each drug in the target text, the terminal device can adopt one comprehensive entity representation to comprehensively describe semantic information of the corresponding drug at each position in the target text. Then, for each target drug pair, the terminal device may perform fusion according to the comprehensive entity representation of each drug to generate a fused entity representation of the obtained drug pair. Furthermore, the terminal device can accurately predict the interaction relation of the drug pair based on the fusion entity representation.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for the embodiments or the prior art descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.

Fig. 1 is a schematic diagram illustrating an implementation of an embodiment of sentence level generation according to an embodiment of the present application;

FIG. 2 is a flowchart illustrating an implementation of a method for predicting drug-pair interactions according to an embodiment of the present application;

FIG. 3 is a schematic diagram illustrating one implementation of generating a synthetic entity representation in a method for predicting drug pair interactions according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a model structure for generating a comprehensive entity representation of a drug according to an embodiment of the present application;

FIG. 5 is a schematic flow chart illustrating a process for predicting drug-pair interaction relationships using a drug relationship prediction model according to an embodiment of the present application;

FIG. 6 is a schematic diagram illustrating an implementation of generating a training set in a method for predicting drug-pair interactions according to an embodiment of the present application;

FIG. 7 is a schematic diagram illustrating an implementation of determining relationship labels in a method for predicting drug pair interactions according to another embodiment of the present application;

FIG. 8 is a schematic diagram illustrating an apparatus for predicting drug pair interaction relationships according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a terminal device according to an embodiment of the present application.

Detailed Description

In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular system structures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

It will be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

Furthermore, in the description of the present application and the appended claims, the terms "first," "second," "third," and the like are used for distinguishing between descriptions and not necessarily for describing a relative importance or importance.

The method for predicting the interaction relationship of the drug provided by the embodiment of the application can be applied to terminal devices such as a tablet computer, a wearable device, a notebook computer, an ultra-mobile personal computer (UMPC), a netbook and the like, and the embodiment of the application does not limit the specific types of the terminal devices.

As described in the background, all current DDI predictions are based on sentence-level relationships. Among other things, in training sentence-level drug relationship prediction models, it is often necessary to use large data sets. Specifically, after the data set is acquired, a summary or a part of text about drug descriptions of each scientific literature in the data set is taken as a document and stored in a separate XML file. At this time, each XML file is composed of at least one or more sentences. Each sentence contains several drugs and a relationship label between every two drugs.

Therefore, there has been work to convert each sentence into several instances (each instance is called instance, focusing on only two drugs). Specifically, see fig. 1) for details, fig. 1 is a schematic diagram of an implementation manner of an embodiment of generating a sentence level according to an embodiment of the present application. Specifically, assuming that there is one Sentence (sequence) containing three non-repeating drugs (Drug1, Drug2, Drug3), the terminal device needs to convert it into three instances (Instance _ a, Instance _ B, Instance _ C), each of which focuses only on one Drug pair (e1 and e 2).

Based on this, the number of instances of sentence generation can be determined by the number of drugs contained in the sentence (permutation and combination, for example, if 4 drugs are contained in a sentence, it is necessary to convert them into C² ₄6 instances). That is, training the drug relationship prediction model based on sentence level requires processing a number of instances that is several times the number of sentences contained in the original text. Therefore, the sentence-level drug relationship prediction model requires a larger memory space in the training process, and the network training is more time-consuming. In the embodiment, the same medicine in one text is described by using one integrated entity representation, and one medicine pair corresponds to one fused entity representation. That is, the drug relationship prediction model is trained on a sentence level to divide the text into a plurality of sentences, and the drugs and drug pairs in each sentence are expressed differently using vectors, respectively. Therefore, the terminal equipment can greatly reduce the memory space required to be used during model training.

In addition, because the sentence-level drug relationship prediction model only pays attention to certain two drugs in the sentence during training, the prediction process is also only based on semantic information between certain two drugs in the sentence. However, in practice, the relationship between one drug is often determined by a plurality of sentences in common. Therefore, the existing method for predicting the interaction relationship of the drug pairs has the problem that semantic information of more complete interaction between the drug pairs cannot be mined from the text related to biomedicine.

Based on this, in order to solve the problem that semantic information of more complete interaction between drug pairs cannot be mined from the biomedical related texts, in the embodiment, the terminal device processes the target text in the following manner S101-S104, so as to generate semantic information that can be used for describing the interaction between two drugs in the target text, so as to improve the accuracy of predicting the drug interaction relationship.

The method for predicting the drug-pair interaction relationship provided by the present application is described below with reference to specific examples.

Referring to fig. 2, fig. 2 is a flowchart illustrating an implementation of a method for predicting drug-pair interaction relationship according to an embodiment of the present application, where the method includes the following steps:

S201, terminal equipment acquires a target text; the target text includes a plurality of drugs, each drug appearing at least once in the target text.

In one embodiment, the target text is generally text related to the pharmaceutical field, including but not limited to text in the form of periodicals, papers, and the like. The target text may be a text in chinese, english, or other languages, which is not limited herein.

In one embodiment, the target text at least includes more than two drugs for predicting drug pair interaction, otherwise, the interaction relationship between the two drugs cannot be predicted from the target text.

It will be appreciated that for any drug in the target text, it may appear at different locations in the target text multiple times. Therefore, in the present embodiment, there is no limitation on the occurrence of the target drug at this time.

In an embodiment, the terminal device may specifically obtain the target text through the following steps, which are detailed as follows:

the method comprises the steps that terminal equipment obtains an initial text, wherein the initial text comprises a plurality of medicines, and each medicine appears in the initial text at least once; and if the initial text has the medicine name using the medicine sharing suffix, expanding the medicine name to obtain a target text.

In an embodiment, the initial text is an unprocessed text, which may be a text crawled from a network by the terminal device based on a drug name, or a text pre-stored in the terminal device.

In one embodiment, the entity name is a drug name of the drug. Wherein the drug sharing suffix is: when the names of the medicines have the same part name, a case may occur in which a letter is written in the initial text so that the medicines share a suffix. The shared suffix is the same part name.

For example, the entity names of two drugs may be: 1) diagnostic monoclonal antibodies; 2) therapeutic monoclonal antibodies. And the initial text containing both drugs may be "… where treated with other diagnostic or therapeutic monoclonal antibodies". That is, the two drug names share the "monoclonal antibodies" as a shared suffix. For this situation, the terminal device needs to expand the sentence in the initial text to obtain the target text. Namely, the above statement is changed into: "… where branched with other diagnostic monoclonal antibodies or therapeutic monoclonal antibodies".

S202, the terminal device determines a comprehensive entity representation corresponding to each medicine respectively, and the comprehensive entity representation is used for describing semantic information of the corresponding medicine at each position in the target text.

In an embodiment, the synthetic entity is configured to synthesize semantic information describing respective locations of the corresponding drug in the target text. Specifically, when a drug appears at multiple positions in the target text, if only the sentence containing the drug at a certain position is processed to obtain semantic information indicating that the drug is in the sentence, the accuracy of prediction may be inaccurate when performing subsequent drug-to-interaction relationship prediction based on the semantic information. That is, the semantic information of the drug is obtained only by processing a sentence containing the drug in the target text, and the semantic information cannot replace the semantic information corresponding to the drug at the other positions in the target text.

Based on this, in this embodiment, the terminal device may participate in the subsequent processing through the comprehensive entity representation corresponding to the drug, so as to improve the accuracy rate of the prediction of the interaction relationship by the drug.

In a specific embodiment, referring to fig. 3, in S202, the terminal device may specifically be implemented by the following sub-steps S301 to S303, which are detailed as follows:

S301, for any kind of medicine, the terminal device respectively determines a plurality of positions of the medicine in the target text.

S302, the terminal device respectively generates a text sequence corresponding to each position according to the plurality of positions.

And S303, carrying out vector processing on the text sequence corresponding to each position by the terminal equipment to obtain comprehensive entity representation.

In one embodiment, the location is the location information of the drug in the target text. Wherein the text sequence is a sequence generated based on the location. Specifically, the terminal device may determine the location information of the drug in the target text according to the entity name of the drug. And, the terminal device can identify the start position and the end position of each occurrence of the medicine using "[" and "]" in order to mark the position of the medicine. In the case where one drug has a plurality of position information (that is, the drug appears in the target text a plurality of times), the order information corresponding to each position may be given according to the order in which the drug appears in the target text.

Illustratively, the terminal device may represent the text sequence of the medication in the following manner:

X＝{x₁，x₂，…x_ndenotes the text sequence of the target text as a whole, X_nRepresenting the nth character in the target text, n also representing the total number of characters in the target text. Assuming that for a given Drug- α, which consists of k characters and occurs 2 times, its text sequence can be: p1 ═ { x ═ x _i，x_i+1，…x_i+k-1}，P2＝{x_j，x_j+1，…x_j+k-1}. Wherein 1 and 2 in P1 and P2 represent the order in which the drugs appear in the target text, respectively. x is a radical of a fluorine atom_iThe ith character, which is represented as the first occurrence of the drug in the target text; since the drug name consists of k characters, then x_i+k-1I.e. the end position in the text after the first appearance of the drug. It will be appreciated that if the number of times the drug occurs is multiple, the text sequence will also correspond to the multiple.

In one embodiment, vector processing the text sequence is: the text sequence is represented as a process recognizable by the terminal device.

Specifically, the terminal device may respectively perform vector representation on the text sequence corresponding to each position, and obtain a plurality of text vectors correspondingly; each text vector is used for describing semantic information of the medicine at the corresponding position; and performing vector integration on each text vector to generate a comprehensive entity representation of the medicine.

The terminal device performs vector representation on each text sequence, and specifically, the text sequence can be generated by processing through a model. The text sequence is vector processed, for example by the BioBERT (named entity recognition model), resulting in a synthetic entity representation. For example, the text vector generated by the BioBERT processing the two text sequences P1 and P2 may be: drug 1 ═ vp1_1, vp1_2, … vp1_ k, and drug 2 ═ vp2_1, vp2_2, … vp2_ k. Wherein, drug P1 is represented as a text vector corresponding to the P1 text sequence; vp1-k is represented as a vector corresponding to the kth character in the P1 text sequence.

It will be appreciated that each text vector can only describe semantic information of the drug at the corresponding location in the target text at this time. Based on this, in order to obtain the comprehensive entity representation of the medicine, the terminal device needs to integrate a plurality of text vectors of the medicine. Specifically, the terminal device may integrate the text vectors according to the following formulas 1 and 2:

wherein, Drug_e1Representing an integration vector obtained after the integration processing is carried out on the first text vector; that is, for drug P1, after summing each vector in the representation drug P1, the average value is calculated, and the average value is the integration vector. Then, averaging each integration vector corresponding to the Drug again to generate a Drug integrated entity representation_a。

Specifically, reference may be made to fig. 4, and fig. 4 is a schematic structural diagram of a model for generating a comprehensive entity representation of a drug according to an embodiment of the present application. Wherein Drug- α at the lowermost layer in fig. 4 represents the Drug name of the Drug; { x_i，...，x_i+k-1Expressing the text sequence of Drug-alpha, and then performing vector expression processing through a BioBERT model to generate a text vector (Drug in the figure)_e1And Drug_e2). Then, the text vector is processed by the above formula 1 and formula 2 to generate the top layer Drug _a. I.e. to generate a comprehensive entity representation. It should be noted that this process is for the terminal device to obtain the final composite entity representation by integrating all parts of a single drug and integrating a single text vector for all the same drugs.

S203, aiming at any drug pair in the multiple drugs, the terminal equipment determines the fusion entity representation of the drug pair according to the comprehensive entity representation of two drugs in the drug pair.

In one embodiment, the integrated entity of the single drug represents semantic information describing the respective drug at various locations in the target text. Thus, the above-mentioned fusion entity can be considered to represent semantic information describing the interaction of two drugs in the target text.

Wherein, the terminal device can process the comprehensive entity expression through the following formula 3 to obtain the fusion

The entity represents: h₁＝W₁*tanh(Drug_α)]+b₁,H₂＝W₂[tanh(Drug_β)]+b₂ (3)

Wherein H₁And H₂Respectively representing the target vectors, W, obtained after processing the synthetic entity representation₁And W₂Representing the known parameter matrix, b1 and b2 representing the known offset terms; tanh represents hyperbolic tangent processing of the composite entity representation.

In the presence of hydrogen to obtain H₁And H₂The terminal device may then send H₁And H₂And splicing, and then inputting the spliced entity into the formula 4 again to obtain a fused entity expression.

H₀＝W₃[concat(H₁,H₂)]+b₃ (4)

Wherein H₀I.e. a fused entity representation of the drug pair, W₃Representing the known parameter matrix, b3 known offset terms; concat is expressed as a function of concatenating multiple strings (i.e., for H)₁And H₂Splicing is performed).

It should be noted that, the above formula is a calculation formula for processing only the medicine with the number of occurrences of the target text being two, and when the number of occurrences is multiple, the formula should be adapted accordingly.

And S204, the terminal equipment predicts the interaction relation of the drug pairs according to the fusion entity representation of the drug pairs.

In one embodiment, S203 above has illustrated that the fused entity represents semantic information that may be used to describe the interaction of two drugs in the target text. Based on the method, the terminal device has higher prediction accuracy when predicting the interaction relation of the drug pair based on the fusion entity representation.

Specifically, the terminal device may predict the interaction relationship of the drug pair according to the following formula 5:

Type＝softmax(H₀) (5)

wherein softmax represents a classification function for H₀Processing is performed and the probability value that the drug pair belongs to each interaction relationship is output. And then, determining the interaction relation corresponding to the maximum value of the probability value as the finally predicted interaction relation of the drug pair.

In this embodiment, by acquiring the corresponding integrated entity representation of each drug in the target text, the terminal device may use one integrated entity representation to comprehensively describe semantic information of the corresponding drug at each position in the target text. Then, for each target drug pair, the terminal device may perform fusion according to the comprehensive entity representation of each drug to generate a fused entity representation of the obtained drug pair. Furthermore, the terminal device can accurately predict the interaction relation of the drug pair based on the fusion entity representation.

In an embodiment, the above S202-S204 may all be performed by a drug relationship prediction model in the terminal device to process the target text. That is, after executing S201, the terminal device may input the acquired target text into the drug relationship prediction model to predict the interaction relationship of each drug pair in the plurality of drugs.

Specifically, the drug relationship prediction model may include a first activation layer, a second activation layer, a first fully-connected layer, and a second fully-connected layer. Wherein the first activation layer is configured to perform a tanh function process in equation 3 on the synthetic entity representation. The first full-link layer is used for executing W in formula 3 on the vector processed by the tanh function ₁[]+b₁Or W₂[]+b₂And processing to obtain a target vector. Then, the terminal device may splice the target vectors of the two drugs and input the spliced target vectors into the second active layer for processing, where the second active layer is configured to perform concat function processing in formula 4 on the spliced target vectors, and input the concat function processed vectors into W₃[]+b₃Processing to obtain the fused entity representation.

It should be noted that the above example only illustrates the model structure of the drug relationship prediction model processing the integrated entity representation of two drugs in the drug pair to generate the fused entity representation of the drug pair. That is, only the model structure for processing S203 is described. The drug relationship prediction model should further include a model structure for executing the processes of S202 and S204, and this embodiment is not explained one by one.

In an embodiment, please refer to fig. 5, and fig. 5 is a schematic flow chart illustrating a process of predicting drug-pair interaction relationships by a drug relationship prediction model according to an embodiment of the present application. The data processing is specifically to convert the sentence-level DDI2013 data set (sense-level DDI Extraction 2013) into a document-level DDI2013 data set (document-level DDI Extraction 2013). Thereafter, a load critical information process is performed on the document set DDI2013 data set. Specifically, a text sequence is established (articule seq) for each text in the data set, which includes, but is not limited to, the establishment of a text sequence for the entire text, and the establishment of a text sequence for each drug; drug Pairs (Pairs) are determined and Drug information (Drug info) is generated. Thereafter, for the determined drug pairs, a Document-entity Embedding process (Document-entity Embedding) is performed separately for each drug in the drug pairs. Specifically, a synthetic entity representation (i.e., a Drug emb is generated) is performed for each Drug (Drug), respectively. Thereafter, the tanh + fully-connected process is performed on the integrated entity representations of the drugs, respectively. Respectively inputting the comprehensive entity representation to the first activation layer and the first full-connection layer in turn and processing to obtain a target vector (H) of each drug ₁And H₂). Then, splicing the two target vectors, and inputting the spliced target vectors into a second activation layer and a second full-connection layer to obtain a fusion entity representation (H)₀). Finally, inputting the fused entity representation into a Sofmax layer for classification prediction to obtain the interaction relationship (Type) of the drug pair.

In an embodiment, the drug relationship prediction model is a model trained in advance. Illustratively, the drug relationship prediction model can be a model such as BERT, SciBERT, and BioBERT. In this embodiment, the drug relationship prediction model may be specifically a BioBERT.

In an embodiment, regarding the existing problem that the sentence-level drug relationship prediction model requires a larger memory space and takes longer time for network training in the training process, referring to fig. 6, the terminal device may specifically process the original data set through the following steps S601 to S604, so as to reduce the data amount of the original data set and improve the efficiency of network training, which is detailed as follows:

s601, the terminal device obtains an original data set, wherein the original data set comprises a plurality of original texts.

S602, the terminal device respectively counts the number of the medicines contained in each original text.

S603, the terminal device screens original texts containing at least two medicines in the original data set to obtain an original data subset.

And S604, the terminal equipment performs label processing on each original text in the original data subset to obtain a training set.

In an embodiment, the manner of acquiring the original text may be similar to that of acquiring the original text, and a comparison will not be described. It should be noted that, if the raw data set is directly used for model training, a large amount of training time is consumed.

It will be appreciated that there may be text in the original text that does not contain both drugs. Such raw text cannot be used directly for training. Based on this, the terminal device can count the number of drugs contained in each original text, respectively. Thereafter, the original text including only one drug is deleted, and preprocessing is performed on the original text that is not deleted.

In one embodiment, the preprocessing includes at least a process of augmenting the medication with a medication sharing suffix. Here, the above processing procedure is already explained in the above S201, and will not be described again. It should be added that the above pre-treatments also include, but are not limited to: the method includes the steps of carrying out lowercase writing on Chinese and English characters in the original text, removing punctuations, and converting all numbers in the original text into NUM for replacement, and is not limited.

It can be understood that the original text after the preprocessing is the text that can be used for training the drug relationship prediction model, so that the redundancy of the training data can be reduced.

In an embodiment, the tag processing on the original text specifically includes: and respectively labeling the drug pairs appearing at each position in the original text so as to participate in model training.

However, based on the above description of the prior art sentence-level drug relationship prediction model, it can be seen that when one document is divided into a plurality of sentences, there may be some plurality of sentences containing the same drug pairs. However, when the same drug pair is at different locations in the document, its corresponding relationship label may be different. Meaning that the same drug pair corresponds to different semantic information in the document, respectively. If the first drug pair or the relationship label of a certain drug pair is used to participate in model training, the prediction accuracy of the finally generated drug relationship prediction model will also be reduced. If each identical drug pair in a document uses a different relationship label, the problem of confusion of relationship labels in the data set will be caused.

Based on this, in this embodiment, referring to fig. 7, the terminal device may further process the label relationship of each drug pair in the original text in the following manners S701-S702, so that the same drug pair may also use the optimal relationship label, thereby solving the problem of confusion of relationship labels:

S701, the terminal device obtains each medicine pair contained in the original text and relationship labels among the various medicine pairs.

S702, if the medicine pair with the multiple relation labels exists, the terminal equipment determines the relation label with the higher priority in the multiple relation labels as the relation label of the medicine pair according to the preset label priority.

In one embodiment, the relationship labels are used to represent the action relationship between drug pairs for participating in an iterative process in a drug relationship prediction model. In the training process, the relation label of each medicine pair is usually labeled by a worker in advance, so that the terminal device can directly acquire each medicine pair contained in the original text and the relation label between each medicine pair.

It should be noted that, if a same drug pair appears at different positions in the original text, the context and semantics may be different, and therefore, the corresponding relationship labels may also be different. In this embodiment, the drug pair is a synthetic entity representation based on two drugs, and a fused entity representation is generated. That is, when a drug pair has a plurality of relationship labels, only one relationship label should be used to correspond to the drug pair for model training.

In an embodiment, the tag priority is a preset priority. Illustratively, the labels may be: false, Int, Advise, Effect, Mechanism. The priority may be: false < Int < Advise < Effect < Mechanism

Wherein, the label mecanism has the highest priority, namely, more pharmacokinetic information is contained between two medicines; the label Effect indicates that there is some degree of reaction between the two drugs, but not to the same extent as mecanism; the label Advise shows that the two medicines have interaction, and the degree is lower than that of Effect; the label Int represents that the interaction degree between the two drugs is low and is not as good as Advise; the label False indicates that there is no drug interaction between the two drugs.

In this embodiment, when a drug pair has a plurality of relationship labels, the multi-relationship label can be converted into a single-relationship label through the rule of label priority as described above, so that the converted single-relationship label can better represent the interaction relationship of the drug pair in the original text.

In one embodiment, the method for predicting drug-pair interactions in the present application is a document-level based prediction method, and has the following advantages compared to the sentence-level method for predicting drug-pair interactions:

In practical applications, sentence-level drug pair interaction prediction must convert one sentence into multiple instances containing only two drug entities. In contrast, the prediction of document-level drug-pair interactions can be for multiple drug entities simultaneously. Therefore, the prediction of the interaction relation by the document-level medicine can simplify the operation of data preprocessing and reduce the text input into the medicine relation prediction model. To reflect this advantage more intuitively, the number of sentences recorded (which need to be input to the drug relationship prediction model) in the recent sentence-level drug-drug pair interaction relationship prediction method is collected and compared with the number of sentences contained herein. See table 1 below for details:

TABLE 1 number of sentences in different methods

As can be seen from Table 1, the amount of text contained in the original DDI Extraction 2013 is the highest. After preprocessing, 27792 sentences were in the training set, 5716 sentences, 33508 sentences were in the testing set. In this embodiment, the number after the preprocessing is the smallest: there were 3784 sentences in the training set and 790 sentences in the test set for a total of 4574 sentences.

(2) Comparison of different BERT models: in text processing, there are three commonly used BERT pre-training models, BERT, SciBERT, and BioBERT. To observe the effect of the three pre-trained models in the prediction of document-level drugs on the interaction relationships, the BioBERT in this approach was replaced with BERT and SciBERT. However, in practical experiments, it was found that the proposed method would not work properly after replacing BioBERT with BERT or ScIBERT. The method specifically comprises the following steps: document-level drug pair interaction prediction methods do not blind the drug and most drugs consist of complex drug nouns. Of the three pre-trained models, only the BioBERT was trained on a large-scale biomedical corpus, and therefore, only the BioBERT was able to accurately express entity representations of complex drugs. In order to further obtain the characterization effects of the three pre-training models, a method for extracting the sentence-level drug pair interaction relation on a DDI corpus is adopted. Other experimental settings were completely consistent in order to analyze which pre-trained model could better express the textual data of drug pair interaction relationships. See table 2 below for details

TABLE 2 results using different BERT models

Pre-training model	macro-P(％)	macro-R(％)	macro-F1(％)
				BERT	78.78	73.27	75.92
SciBERT	81.71	74.80	78.10
				BioBERT	85.89	73.46	79.19

As shown in Table 2, the performance of the method using BERT is the lowest, and macro-P (macro-average precision, one evaluation index of the model) reaches 78.78%, macro-R (macro-average recall, the other evaluation index of the model) reaches 73.27%, and macro-F1 (macro-average harmonic mean, the other evaluation index of the model) reaches 75.92%. The Scibert method gave modest results, with macro-R being the highest, reaching 74.80%. This is because SciBERT is trained on a large-scale scientific literature corpus, and therefore has a much improved performance compared to BERT. Best results were obtained using the BioBERT method with 85.89% macro-P and 79.19% macro-F1. This indicates that the BioBERT trained on the biomedical corpus can more accurately express textual data of drug-pair interactions.

(3) Integrated entity representation (imbedding) performance of document level drugs: to verify the above-described actual performance, an experiment can be designed to compare the effects of embedding without and with document-level drugs. Specifically, the former was labeled Without DEE (for each drug, only embedding was performed at the first appearance) and compared with the method using DEE proposed in this example.

TABLE 3 Effect of using DEE and not using DEE

Method	macro-P(％)	macro-R(％)	macro-F1(％)
				Without DEE	60.07	56.32	58.43
Use DEE	65.60	59.71	62.51

As shown in Table 3, in the absence of the DEE method, the macro-P reached 60.07%, the macro-R reached 56: 32%, and the macro-F1 reached 58.43%. In the case of DEE, macro-P reached 65.60%, macro-R reached 59.71%, and macro-F1 reached 62.51%, which were 5.53%, 3.39%, and 4.08% higher, respectively, than those without DEE. The reason is that: without DEE, contextual semantic information of the same drug at different locations in the document is not considered. Therefore, the method can obtain the complete comprehensive entity representation of the medicine in the document by using the embedding of the document-level medicine so as to obtain a more accurate prediction result.

(4) Comparison of different neural network model structures: the embodiment carries out special pretreatment on the DDI Extraction 2013 data set for the first time, and realizes the prediction of the interaction relation of the document-level drugs. Currently, there is no work on document level DDI datasets. To verify the validity of the proposed method, it was compared to methods using CNN and BiLSTM (two of the most common neural network models). These two methods also employ imbedding of document-level drugs, but after obtaining a comprehensive entity representation of the drug, use different neural network model structures. However, in practical applications, it was found that the method using only the BilSTM network model does not work. Therefore, the terminal device changes the method into a method combining the CNN and the BiLSTM neural network model, and represents the method as "CNN + BiLSTM".

TABLE 4 results using different neural network model structures

As can be seen from Table 4, although macro-P in the neural network model structure of CNN + BilSTM reaches 66.98%, which is the highest of the three methods, macro-R only reaches 50.19%, and macro-F1 only reaches 57.38%, so that the overall performance of the network structure is the lowest. The macro-P of the CNN method reaches 56.75 percent, the macro-R reaches 59.97 percent and the macro-F1 reaches 58.32 percent. The overall performance of CNN is slightly higher than CNN + BilSTM. However, using the structure of the drug relationship prediction model in this application, macro-P was only 1.38% lower than CNN + BilSTM and macro-R was 0.26% lower than CNN. Macro-P and Macro-R are almost the highest and therefore the overall performance is the best. The reason is that: the input is a composite entity representation of two drugs, rather than a complete sentence. Therefore, fitting sentence-level neural network model structures does not achieve the same performance in document-level neural network model structures (especially in BilSTM).

In summary, in the embodiment, compared with the method for predicting the interaction relationship of the document-level drug pair in the prior art, the method for predicting the interaction relationship of the document-level drug pair can greatly reduce the data amount input into the drug relationship prediction model, and can comprehensively perform accurate semantic expression on the drugs at a plurality of different positions, so that the drug relationship prediction model can extract the real semantic information of the drugs in the document, and the model prediction accuracy is improved.

Referring to fig. 8, fig. 8 is a block diagram illustrating a device for predicting drug pair interaction according to an embodiment of the present disclosure. The device for predicting the interaction relationship of a drug pair in this embodiment includes modules for executing the steps in the embodiments corresponding to fig. 2, fig. 3, fig. 6 and fig. 7. Please refer to fig. 2, fig. 3, fig. 6 and fig. 7, and the related descriptions in the embodiments corresponding to fig. 2, fig. 3, fig. 6 and fig. 7. For convenience of explanation, only the portions related to the present embodiment are shown. Referring to fig. 8, a device 800 for predicting drug pair interaction relationships may include: an acquisition module 810, a synthetic entity representation determination module 820, a fused entity representation determination module 830, and a prediction module 840, wherein:

an obtaining module 810, configured to obtain a target text; the target text includes a plurality of drugs, each drug appearing at least once in the target text.

An integrated entity representation determining module 820 for determining an integrated entity representation corresponding to each drug, respectively, the integrated entity representation describing semantic information of the corresponding drug at various positions in the target text.

A fused entity representation determining module 830 for determining a fused entity representation of a drug pair from a composite entity representation of two drugs in the drug pair for any drug pair of the plurality of drugs.

The prediction module 840 is configured to predict an interaction relationship of the drug pair according to the fused entity representation of the drug pair.

In an embodiment, the obtaining module 810 is further configured to:

acquiring an initial text, wherein the initial text comprises a plurality of medicines, and each medicine appears in the initial text at least once; and if the initial text has the medicine name using the medicine sharing suffix, expanding the medicine name to obtain a target text.

In an embodiment, the synthetic entity representation determination module 820 is further configured to:

respectively determining a plurality of positions of the medicines in the target text aiming at any medicine; respectively generating a text sequence corresponding to each position according to the plurality of positions; and carrying out vector processing on the text sequence corresponding to each position to obtain comprehensive entity representation.

respectively carrying out vector representation on the text sequence corresponding to each position to correspondingly obtain a plurality of text vectors; each text vector is used for describing semantic information of the medicine at the corresponding position; and performing vector integration on each text vector to generate a comprehensive entity representation of the medicine.

In one embodiment, the device 800 for predicting drug pair interaction relationship further comprises:

And the input module is used for inputting the target text into the pre-trained medicine relation prediction model for processing to obtain the interaction relation of each medicine pair in the multiple medicines.

In one embodiment, the drug relationship prediction model includes a first activation layer, a second activation layer, a first fully-connected layer, and a second fully-connected layer; the fused entity representation determining module 830 is further configured to:

inputting the comprehensive entity expression of the two medicines into the first activation layer and the first full-connection layer in sequence to obtain target vectors corresponding to the two medicines respectively; and splicing the two target vectors, and sequentially inputting the spliced target vectors into a second activation layer and a second full-connection layer to obtain a fusion entity representation.

In one embodiment, the drug relationship prediction model is trained according to a training set; the device 800 for predicting drug pair interaction relationship further comprises the following modules for obtaining a training set:

the original data set acquisition module is used for acquiring an original data set, and the original data set comprises a plurality of original texts.

And the counting module is used for counting the number of the medicines contained in each original text respectively.

And the screening module is used for screening the original text containing at least two drugs in the original data set to obtain an original data subset.

And the label processing module is used for carrying out label processing on each original text in the original data subset to obtain a training set.

In one embodiment, the tag processing module is further configured to:

acquiring each drug pair contained in an original text and relationship labels among the drug pairs; and if the medicine pair with the multiple relation labels exists, determining the relation label with the higher priority in the multiple relation labels as the relation label of the medicine pair according to the preset label priority.

It should be understood that, in the structural block diagram of the device for predicting drug pair interaction relationship shown in fig. 8, each module is used to execute each step in the embodiments corresponding to fig. 2, fig. 3, fig. 6, and fig. 7, and each step in the embodiments corresponding to fig. 2, fig. 3, fig. 6, and fig. 7 has been explained in detail in the above embodiments, and specific reference is made to the description in the embodiments corresponding to fig. 2, fig. 3, fig. 6, and fig. 7 and fig. 2, fig. 3, fig. 6, and fig. 7, and no further description is repeated here.

Fig. 9 is a block diagram of a terminal device according to an embodiment of the present application. As shown in fig. 9, the terminal apparatus 900 of this embodiment includes: a processor 910, a memory 920 and a computer program 930, such as a program for a method of predicting a drug-pair interaction relationship, stored in the memory 920 and executable at the processor 910. The processor 910, when executing the computer program 930, implements the steps in the embodiments of the method for predicting drug-pair interaction relationship described above, such as S101 to S104 shown in fig. 1. Alternatively, the processor 910, when executing the computer program 930, implements the functions of the modules in the embodiment corresponding to fig. 8, for example, the functions of the modules 810 to 840 shown in fig. 8, please refer to the related description in the embodiment corresponding to fig. 8.

Illustratively, the computer program 930 may be divided into one or more modules, which are stored in the memory 920 and executed by the processor 910 to implement the drug pair interaction relationship prediction methods provided by the embodiments of the present application. One or more of the modules may be a series of computer program instruction segments capable of performing certain functions, which are used to describe the execution of computer program 930 in terminal device 900. For example, computer program 930 may implement the methods for predicting drug-pair interactions provided in embodiments of the present application.

Terminal device 900 can include, but is not limited to, a processor 910, a memory 920. Those skilled in the art will appreciate that fig. 9 is merely an example of a terminal device 900 and is not intended to limit terminal device 900 and may include more or fewer components than those shown, or some of the components may be combined, or different components, e.g., the terminal device may also include input output devices, network access devices, buses, etc.

The processor 910 may be a central processing unit, but may also be other general purpose processors, digital signal processors, application specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic, discrete hardware components, or the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The storage 920 may be an internal storage unit of the terminal device 900, such as a hard disk or a memory of the terminal device 900. The memory 920 may also be an external storage device of the terminal device 900, such as a plug-in hard disk, a smart card, a flash memory card, etc. provided on the terminal device 900. Further, the memory 920 may also include both internal and external memory units of the terminal device 900.

The embodiments of the present application provide a computer-readable storage medium, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the method for predicting drug-pair interaction relationship in the above embodiments.

The embodiment of the present application provides a computer program product, which, when running on a terminal device, causes the terminal device to execute the method for predicting a drug pair interaction relationship in the above embodiments.

The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present application and are intended to be included within the scope of the present application.

Claims

1. A method for predicting drug-pair interaction relationships, the method comprising:

determining, for any drug pair of the plurality of drugs, a fused entity representation of the drug pair from the integrated entity representations of both drugs of the drug pair;

predicting an interaction relationship of the drug pair based on the fused entity representation of the drug pair.

2. The method of claim 1, wherein obtaining the target text comprises:

acquiring an initial text, wherein the initial text comprises the plurality of medicines, and each medicine appears in the initial text at least once;

and if the initial text has the medicine name using the medicine sharing suffix, expanding the medicine name to obtain the target text.

3. The method of claim 1, wherein said separately determining a composite entity representation corresponding to each of said drugs comprises:

For any one of the medicines, respectively determining a plurality of positions of the medicine in the target text;

respectively generating a text sequence corresponding to each position according to the positions;

and carrying out vector processing on the text sequence corresponding to each position to obtain the comprehensive entity representation.

4. The method of claim 3, wherein the vector processing the text sequence corresponding to each position to obtain the synthetic entity representation comprises:

respectively carrying out vector representation on the text sequence corresponding to each position to correspondingly obtain a plurality of text vectors; each text vector is used for describing semantic information of the medicine at a corresponding position;

and performing vector integration on each text vector to generate a comprehensive entity representation of the medicine.

5. The method according to any one of claims 1-4, wherein said determining a synthetic entity representation corresponding to each of said drugs respectively, said synthetic entity representation describing semantic information of the corresponding said drug at each position in said target text; determining, for any drug pair of the plurality of drugs, a fused entity representation of the drug pair from the integrated entity representations of the two drugs of the drug pair; predicting an interaction relationship of the drug pair based on the fused entity representation of the drug pair, comprising:

And inputting the target text into a pre-trained drug relationship prediction model for processing to obtain the interaction relationship of each drug pair in the multiple drugs.

6. The method of claim 5, wherein the drug relationship prediction model comprises a first activation layer, a second activation layer, a first fully-connected layer, and a second fully-connected layer; the determining a fused entity representation of the drug pair from the integrated entity representation of the two drugs in the drug pair comprises:

inputting the comprehensive entity representation of the two medicines into the first activation layer and the first full-connection layer in sequence to obtain target vectors corresponding to the two medicines respectively;

and splicing the two target vectors, and sequentially inputting the spliced target vectors into the second activation layer and the second full-connection layer to obtain the fused entity representation.

7. The method of claim 5, wherein the drug relationship prediction model is trained according to a training set, the training set being obtained by:

acquiring an original data set, wherein the original data set comprises a plurality of original texts;

Respectively counting the quantity of the medicines contained in each original text;

screening original texts containing at least two drugs in the original data set to obtain an original data subset;

and performing label processing on each original text in the original data subset to obtain the training set.

8. The method of claim 7, wherein said tagging each of said original texts in said original data subset to obtain said training set comprises:

acquiring each drug pair contained in the original text and relationship labels among the drug pairs;

and if the medicine pair with the multiple relation labels exists, determining the relation label with the higher priority in the multiple relation labels as the relation label of the medicine pair according to the preset label priority.

9. A terminal device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the method according to any of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the method according to any one of claims 1 to 8.