Construction and prediction method and device of multi-mode comparison drug collaborative prediction model

Abstract

The invention relates to a construction method, a prediction method and a device of a multi-mode contrast drug collaborative prediction model. The construction method comprises the following steps: constructing a multi-mode contrast drug collaborative prediction model according to entity embedding characteristics and phenotypic characteristics of multiple drugs and cell lines; obtaining a first modal feature and a second modal feature of two different modals under the same dimension, wherein the first modal feature and the second modal feature are represented by the drug-drug-cell line triple information; determining sample positive alignment, sample negative alignment and sample semi-positive alignment according to an effectiveness prediction result relative to a preset real sample after drug synergistic effect prediction is performed on each feature sample in the first modal feature and the second modal feature, and further determining a target loss function; and a plurality of models are used according to the loss function. By means of the method, the influence of invalid features which do not have work or have negative influence on drug synergistic effect prediction in any mode on drug prediction can be effectively reduced, and the problem that the accuracy of a prediction result is affected due to the fact that each mode adopts an equal processing strategy in the related technology is avoided.

Description

Technical Field

[0001] The present application relates to the field of drug synergy prediction, and specifically to the construction, prediction method and device of a multi-modal comparative drug synergy prediction model. Background Art

[0002] In the treatment of cancer, the complex compensatory mechanisms exhibited by cancer cells significantly increase the possibility of acquiring resistance when treated with a single drug, which becomes the biggest obstacle in the process of single-drug treatment. In view of this, drug combination therapy can reduce toxicity and reduce drug resistance, becoming a better way to treat cancer.

[0003] Combination drug therapy refers to the use of two or more drugs to enhance the therapeutic effect on cancer. However, not all drug combinations are synergistic, and drugs with antagonistic relationships may cause unnecessary harm to patients. In order to optimize treatment plans, improve clinical efficacy and ensure patient safety, it is necessary to systematically identify and verify synergistic drug combinations.

[0004] Multimodal data provides information about different aspects of drugs, and integrating multimodal data can further improve the performance of prediction models. SynergyX, an existing multimodal method in the field of drug synergy prediction, introduces the mutual attention mechanism into the drug synergy prediction process to introduce multi-dimensional cross-modal interactions. The network is characterized by a "sandwich"-like module consisting of multiple mutual attention modules and self-attention modules, allowing the extraction of drug-cell and drug-drug interactions. These methods focus on integrating multiple phenotypic characteristics of drugs and cell lines to form a more effective joint representation for drug synergy prediction. However, this method usually tends to integrate features of different modalities with equal weights and extract multimodal representations through techniques such as aggregation or contrast fusion, but often ignores the negative effects that invalid single-modal features may bring. Summary of the invention

[0005] The present application provides a construction, prediction method and device of a multi-modal comparative drug synergy prediction model, which can solve the related technical problems existing in the above-mentioned prior art.

[0006] In a first aspect, the present application provides a method for constructing a multi-modal comparative drug synergy prediction model, using the following technical solution:

[0007] A method for constructing a multi-modal comparative drug synergy prediction model, the method comprising:

[0008] Obtaining phenotypic characteristics of the first drug, the second drug, and the cell line;

[0009] According to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the first drug, the second drug, and the cell line and a preset first model, entity embedding features corresponding to the first drug, the second drug, and the cell line are obtained;

[0010] According to the entity embedding feature and the phenotype feature, a first modality feature and a second modality feature of two different modalities representing drug-drug-cell line triple information in the same dimension are obtained;

[0011] Obtaining the effectiveness prediction results of the preset prediction model after predicting the drug synergy for each feature sample in the first modal feature and the second modal feature relative to the preset real samples;

[0012] If the prediction results of the two feature samples corresponding to the same real sample in the first modal feature and the second modal feature are both correct, then the two are determined to be sample positive pairs; if both are wrong, then the two are determined to be sample negative pairs; if only one of them is correct, then the two are determined to be sample semi-positive pairs;

[0013] Determining a target loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs;

[0014] Determine whether the target loss function meets the preset requirements. If not, adjust the multiple models used in the previous step and calculate the target loss function again. If it meets the requirements, determine that the multiple model combinations used this time constitute a multi-modal comparative drug synergistic prediction model.

[0015] In combination with the first aspect, in one implementation, in determining the target loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs,

[0016] A target loss function is determined according to the cosine similarity of the sample positive pair, the sample negative pair and the sample semi-positive pair and a preset temperature coefficient.

[0017] In combination with the first aspect, in one implementation, the target loss function is determined according to the cosine similarity of the sample positive pair, the sample negative pair and the sample semi-positive pair and a preset temperature coefficient, and the following formula is used:

[0018]

[0019] Wherein, n is a paired feature sample of the first modal feature and the second modal feature, P is a positive pair set, S is a semi-positive pair set, and B is a complete set of sample positive pairs, sample negative pairs, and the sample semi-positive pairs. is the first modal feature, is the second mode characteristic and τ is the temperature coefficient.

[0020] In combination with the first aspect, in one implementation, determining a target loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs comprises the following steps:

[0021] Determining a first loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs;

[0022] Determine a second loss function according to the effectiveness prediction results of each feature sample in the first modal feature and the second modal feature after predicting the drug synergy using the prediction model relative to a preset real sample;

[0023] Determine a third loss function according to the effectiveness prediction result of the drug synergy prediction relative to the preset real sample after the fusion feature obtained by concatenating the first modal feature and the second modal feature using the prediction model;

[0024] The target loss function is determined based on at least the first loss function, the second loss function and the third loss function.

[0025] In combination with the first aspect, in one embodiment, the second loss function is determined based on the effectiveness prediction result of the drug synergy prediction relative to the preset real sample after using the prediction model for each feature sample in the first modal feature and the second modal feature, and the following formula is used:

[0026]

[0027] Among them, y k is the label of the kth sample in the first modality feature and the second modality feature, is the softmax output of the kth sample in the corresponding modality feature i.

[0028] In combination with the first aspect, in one embodiment, the third loss function is determined based on the fusion feature obtained by concatenating the first modal feature and the second modal feature in series, and the prediction result of the effectiveness of the drug synergy prediction relative to the preset real sample after using the prediction model to predict the drug synergy, and the following formula is used:

[0029]

[0030] Among them, y k is the predicted value of the kth sample, p k is the softmax output of the kth sample in the fusion feature.

[0031] In combination with the first aspect, in one embodiment, obtaining, according to the entity embedding feature and the phenotypic feature, a first modality feature and a second modality feature of two different modalities representing drug-drug-cell line triple information in the same dimension comprises the following steps:

[0032] Different types of drugs and cell line features are dimensionally aligned according to the preset projection equation;

[0033] According to the entity embedding feature and the phenotype feature, a first basic modality feature and a second basic modality feature representing drug-drug-cell line triple connection information are obtained;

[0034] According to the first basic modal feature, the second basic modal feature and a preset second model, potential features that have an impact on drug synergy are determined from both the first basic modal feature and the second basic modal feature, and the first modal feature and the second modal feature are obtained based on the potential features of the two.

[0035] In a second aspect, the present application embodiment provides a multi-modal comparative drug synergy prediction method based on the model constructed by the above construction method, using the following technical solution:

[0036] A multi-modal comparative drug synergy prediction method, the prediction method comprising the following steps:

[0037] Obtaining phenotypic characteristics of target drug 1, target drug 2, and target cell line 3;

[0038] According to the heterogeneous graph consisting of drugs, targets, and cell lines involved in the target drug one, the target drug two, and the target cell line and the first model, obtaining entity embedding features corresponding to the target drug one, the target drug two, and the target cell line;

[0039] According to the entity embedding feature and the phenotype feature, a modal feature 1 and a modal feature 2 for representing drug-drug-cell line triple information in two different modalities in the same dimension are obtained;

[0040] Connecting the modal feature 1 and the modal feature 2 in series to form a fusion feature;

[0041] According to the prediction model, drug synergy prediction is performed on each feature sample in the fusion feature to determine whether the target drug 1 and the target drug 2 have synergistic effect in the target cell line.

[0042] In a third aspect, the present application provides a device for constructing a multi-modal comparative drug synergy prediction model, which adopts the following technical solution:

[0043] A device for constructing a multi-modal comparative drug synergy prediction model, the device comprising:

[0044] A phenotypic characteristic acquisition module configured to acquire phenotypic characteristics of the first drug, the second drug, and the cell line;

[0045] an entity embedding feature acquisition module, configured to acquire entity embedding features corresponding to the first drug, the second drug, and the cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the first drug, the second drug, and the cell line and a preset first model;

[0046] a single-modal feature representation module configured to obtain, according to the entity embedding feature and the phenotypic feature, a first modal feature and a second modal feature for representing information about a drug-drug-cell line triplet in two different modalities in the same dimension;

[0047] A single-modal feature guided learning module is configured to obtain the effectiveness prediction results of the preset prediction model after predicting the drug synergy of each feature sample in the first modal feature and the second modal feature relative to the preset real sample; if the prediction results of the two feature samples corresponding to the same real sample in the first modal feature and the second modal feature are both correct, then both are determined to be sample positive pairs; if both are wrong, then both are determined to be sample negative pairs; if only one of them is correct, then both are determined to be sample semi-positive pairs; according to the sample positive pairs, sample negative pairs and sample semi-positive pairs, a target loss function is determined; it is judged whether the target loss function meets the preset requirements. If not, the target loss function is calculated again after adjusting the multiple models used in the previous step. If it meets the requirements, it is determined that the combination of multiple models used this time constitutes a multi-modal comparative drug synergy prediction model.

[0048] In a fourth aspect, the present application provides a prediction device for a multi-modal comparative drug synergy prediction method, which adopts the following technical solution:

[0049] A prediction device for implementing the above-mentioned multi-modal comparative drug synergy prediction method, comprising:

[0050] A phenotypic characteristic module, which is configured to obtain phenotypic characteristics of target drug one, target drug two, and target cell line;

[0051] An entity embedding feature module, which obtains entity embedding features corresponding to the target drug one, the target drug two, and the target cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the target drug one, the target drug two, and the target cell line and the first model;

[0052] A multimodal fusion prediction module is configured to obtain modal feature one and modal feature two representing drug-drug-cell line triple information in two different modalities in the same dimension according to the entity embedding feature and the phenotypic feature; concatenate the modal feature one and the modal feature two to form a fusion feature; predict drug synergy for each feature sample in the fusion feature according to the prediction model, and determine whether the target drug one and the target drug two have synergy in the target cell line.

[0053] The beneficial effects brought by the technical solution provided by the embodiments of the present application include:

[0054] By obtaining the phenotypic characteristics and entity embedding characteristics corresponding to the first drug, the second drug and the cell line, the feature representation of the drug-drug-cell line connection information in two modalities is obtained, namely the first modal feature and the second modal feature. Subsequently, the two modal features are further used to predict the drug synergy separately to confirm whether each sample in the first modal feature and the second modal feature is valid, and the sample positive pairs that can be correctly predicted in both modalities, the sample semi-positive pairs that can be correctly predicted in only one modality, and the sample negative pairs that cannot be correctly predicted in both modalities are clearly identified. The sample positive pairs, sample semi-positive pairs and sample negative pairs are further calculated to obtain the target loss function for adjusting the model used. The target loss function obtained can effectively reduce the impact of invalid features that have no effect or have a negative impact on the prediction of drug synergy in any modality on drug prediction in the process of training and constructing the final prediction model, avoiding the problem of affecting the accuracy of the prediction results due to the equal treatment strategy of each modality in the related technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of an embodiment of a method for constructing a multi-modality comparative drug synergy prediction model of the present application;

[0056] Figure 2 This is a schematic diagram of the functional modules of an embodiment of a device for constructing a multi-modality comparative drug synergy prediction model of the present application;

[0057] Figure 3 This is a flow chart of an embodiment of a multi-model comparison drug synergy prediction method of the present application;

[0058] Figure 4 This is a schematic diagram of the functional modules of an embodiment of a multi-model comparative drug synergy prediction device of the present application. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0060] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0061] Combination drug therapy refers to the use of two or more drugs to enhance the therapeutic effect on cancer. However, not all drug combinations are synergistic, and drugs with antagonistic relationships may cause unnecessary harm to patients. In order to optimize treatment plans, improve clinical efficacy and ensure patient safety, it is necessary to systematically identify and verify synergistic drug combinations.

[0062] Multimodal data provides information about different aspects of drugs, and integrating multimodal data can further improve the performance of prediction models. SynergyX, an existing multimodal method in the field of drug synergy prediction, introduces the mutual attention mechanism into the drug synergy prediction process to introduce multi-dimensional cross-modal interactions. The network is characterized by a "sandwich"-like module consisting of multiple mutual attention modules and self-attention modules, allowing the extraction of drug-cell and drug-drug interactions. These methods focus on integrating multiple phenotypic characteristics of drugs and cell lines to form a more effective joint representation for drug synergy prediction. However, this method usually tends to integrate features of different modalities with equal weights and extract multimodal representations through techniques such as aggregation or contrast fusion, but often ignores the negative effects that invalid single-modal features may bring.

[0063] Based on the above problems, the present application provides a method for constructing a multi-modal comparative drug synergy prediction model, a prediction method and a device. The key point of the invention is that a multi-modal comparative drug synergy prediction model that can effectively predict drug synergy is constructed by using the provided method for constructing a multi-modal comparative drug synergy prediction model. In the process of constructing the method, by obtaining the phenotypic characteristics and entity embedding characteristics corresponding to the first drug, the second drug and the cell line, the characteristic representation of the drug-drug-cell line connection information in two modalities is obtained, namely, the first modal characteristics and the second modal characteristics. Subsequently, the two modal characteristics are further used to predict the drug synergy separately to confirm whether each sample in the first modal characteristics and the second modal characteristics is valid, and the sample positive pairs that can be correctly predicted in both modalities, the sample semi-positive pairs that can be correctly predicted in only one modality, and the sample negative pairs that cannot be correctly predicted in both modalities are clearly identified. The sample positive pairs, sample semi-positive pairs and sample negative pairs are further calculated to obtain the target loss function for adjusting the model used. The target loss function obtained can effectively reduce the influence of invalid features that have no effect or have a negative impact on the prediction of drug synergy in any modality on drug prediction in the process of training and constructing the final prediction model, avoiding the problem of affecting the accuracy of the prediction results due to the equal treatment strategy of each modality in the related technology.

[0064] In a first aspect, an embodiment of the present application provides a method for constructing a multi-modal comparative drug synergy prediction model.

[0065] In one embodiment, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the method for constructing a multi-modality comparative drug synergy prediction model of this application. Figure 1 As shown, the method for constructing a multi-modal comparative drug synergy prediction model includes the following steps:

[0066] S100, obtaining phenotypic characteristics of the first drug, the second drug, and the cell line;

[0067] Specifically, the process of obtaining the phenotypic characteristics of the first drug, the second drug, and the cell line will first obtain the SMILES strings of the two drugs, and then use the method in RDKit to calculate the molecular fingerprint. This method represents the molecular structure as a series of atom pairs and hashes each atom pair to generate a unique fingerprint. These fingerprints are converted into bit vectors to represent the characteristics of the drug molecules, marked as For the phenotypic characteristics of the cell lines, the gene expression characteristics are directly used, marked as

[0068] S200, obtaining entity embedding features corresponding to the first drug, the second drug, and the cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the first drug, the second drug, and the cell line and a preset first model;

[0069] The specific process of using the first model to obtain entity embedding features is as follows:

[0070] Get a heterogeneous graph consisting of two drugs, targets, and cell lines, represented as It contains multiple entity types and multiple relationship types. represent the object set, edge set, node type set, and edge type set respectively.

[0071] The heterogeneous graph consists of three nodes, namely drugs (D), proteins (P), and cell lines (C), and five edges. and For each of the five subgraphs with different relationships, a graph attention network (GATConv) is used to learn node embedding features for each subgraph. HeteroConv in the graph attention network allows different convolution operations to be applied to different types of edges, thereby performing effective message passing and feature aggregation in heterogeneous graphs.

[0072]

[0073] Each biological entity node v i (1≤i≤N n ) is represented as a vector Where N n =|V| represents the number of biological entity nodes; d represents the feature dimension. Thus, a biological entity node feature matrix is ​​obtained h i =H[i,:]. GATCov is one of the core components of GAT, which improves the learning ability of the model by introducing an attention mechanism to dynamically assign weights to the neighbors of each node.

[0074] For each node v, GATCnv calculates the attention coefficient α between it and its neighbor node u v，u . These coefficients are calculated using the following formula:

[0075]

[0076] Where W is the learnable weight matrix, a is the parameter vector in the attention mechanism, and h v and h u are the feature vectors of nodes v and u respectively.

[0077] The new representation of a node is obtained by weighting the features of its neighboring nodes:

[0078]

[0079] Where σ is the activation function, is the neighboring node of node v.

[0080] For each subgraph, after two learnings, the corresponding feature matrix H″ will be obtained D-p , H″ P-D , H″ P-P ,H″ C-P and H″ P-C Finally, the features obtained for the corresponding nodes in each subgraph are summed up as the final entity embedding features of each node in the heterogeneous graph.

[0081] H″=H″ D-P +H″ P-D +H″ P-P +H″ C-P +H″ P-C (4)

[0082] Finally, the drug representation is extracted from H″ and cell line expression They are respectively used as the entity embedding features of the two drugs and the entity embedding features of the cell lines.

[0083] S300, obtaining a first modality feature and a second modality feature of two different modalities representing drug-drug-cell line triple information in the same dimension according to the entity embedding feature and the phenotypic feature;

[0084] Specifically, in this embodiment, step S300 includes the following steps:

[0085] S310, dimensionally aligning drug and cell line features of different types, i.e., phenotypic feature modalities and entity embedded feature modalities, according to a preset projection equation;

[0086] The projection equation used is:

[0087] g(x)=σ(Wx+b) (5)

[0088] Among them, W is the weight, b is the bias, and σ is the activation function ReLU. Through the projection equation, we can get: Represents phenotypic characteristics of drugs and cell lines. Entity embedding features representing the feature modality of the corresponding entity.

[0089] S320, obtaining a first basic modality feature and a second basic modality feature representing drug-drug-cell line triple connection information according to the entity embedding feature and the phenotype feature;

[0090] Among them, after obtaining the phenotypic features and real embedding features, we obtain the triple information representing the drug-drug-cell line triple connection information in the two modalities in their respective dimensions: and They are connected in series to form the first basic modal feature and the second basic modal feature respectively.

[0091]

[0092] S330. According to the first basic modal feature, the second basic modal feature and a preset second model, potential features that have an impact on drug synergy are determined from the first basic modal feature and the second basic modal feature, and the first modal feature and the second modal feature are obtained based on the potential features of the two.

[0093] In most previous works, the features of drug A, drug B, and cell lines are directly sent to the classifier for prediction after being concatenated. However, such features simply concatenated from different entity features cannot capture potential connection information. In this embodiment, in order to obtain more effective joint features, Highway Network is used for processing in step S330. The core of this is the introduction of two gating mechanisms. These two gating mechanisms determine the way in which information is transmitted and modified when passing through the network layer. The mathematical formula is as follows:

[0094]

[0095] Among them, non_linear uses the activation function ReLU. T represents the transformation gate, σ represents the activation function sigmoid. C=1-T represents the carrying gate. It captures the combined features The combined features of the potential features. Finally, the first modality features and the second modality features under two modalities are formed.

[0096] S400, obtaining the effectiveness prediction results of the preset real samples after the preset prediction model predicts the drug synergy for each feature sample in the first modal feature and the second modal feature;

[0097] Specifically, the first modal features and the second modal features under the two modes obtained in the above steps are used As the input of the unimodal prediction task, the preset prediction model in this embodiment specifically uses a multi-layer perceptron (MLP) as a unimodal classifier, and the mathematical formula is as follows:

[0098]

[0099] When the prediction results When the relative real sample is correct, It is considered as an effective feature that has a positive effect on the prediction results. It is considered an invalid feature.

[0100] S500, if the prediction results of the two feature samples corresponding to the same real sample in the first modal feature and the second modal feature are both correct, then both are determined to be sample positive pairs; if both are wrong, then both are determined to be sample negative pairs; if only one of them is correct, then both are determined to be sample semi-positive pairs;

[0101] Specifically, after step S400, three situations may occur: When all are valid features, they are considered as positive pairs, and the set of all positive pairs is denoted as P. and When both are invalid features, we regard them as negative pairs and record them as N. When one of them is a valid feature and the other is an invalid feature, we regard it as a semi-positive pair and record it as S.

[0102] S600, determining a target loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs;

[0103] Specifically, in this embodiment, the target loss function is determined according to the cosine similarity of the sample positive pair, the sample negative pair and the sample semi-positive pair and a preset temperature coefficient, using the following formula:

[0104]

[0105] Wherein, n is a paired feature sample of the first modal feature and the second modal feature, P is a positive pair set, S is a semi-positive pair set, and B is a complete set of sample positive pairs, sample negative pairs, and the sample semi-positive pairs. is the first modal feature, is the second mode characteristic and τ is the temperature coefficient.

[0106] S700, determine whether the target loss function meets the preset requirements. If not, adjust the multiple models used in the previous step and recalculate the target loss function. If it meets the requirements, determine that the multiple model combinations used this time constitute a multi-modal comparative drug synergistic prediction model.

[0107] Specifically, by determining the target loss function according to the sample positive pairs, the sample negative pairs and the sample semi-positive pairs, and then using the target loss function to adjust the multiple models used in the aforementioned steps, that is, the first model, the second model and the prediction model used in the previous text, during the adjustment process, for the positive pairs, we can align them with each other. For the semi-positive pairs, the adjusted models can be aligned to a more effective mode to reduce the negative impact of the invalid mode on the prediction results. For the negative pairs, the adjusted models can maximize the difference between the two modes and capture the complementary information of the multi-modality as much as possible.

[0108] Further, in some embodiments, the process of determining the target loss function according to the sample positive pairs, the sample negative pairs and the sample semi-positive pairs in step S600 includes the following steps:

[0109] S610. Determine a first loss function according to the sample positive pairs, the sample negative pairs and the sample semi-positive pairs; wherein the first loss function is the loss function calculated by the formula (10) described above.

[0110] S620, determining a second loss function according to the effectiveness prediction results of the drug synergy prediction performed by the prediction model on each feature sample in the first modal feature and the second modal feature relative to the preset real sample; specifically, the following formula is used:

[0111]

[0112] Among them, y k is the label of the kth sample in the first modality feature and the second modality feature, is the softmax output of the kth sample in the corresponding modality feature i.

[0113] S630, determining a third loss function according to the effectiveness prediction result of the drug synergy prediction relative to a preset real sample after the fusion feature obtained by concatenating the first modal feature and the second modal feature using the prediction model;

[0114] Specifically, all the multimodal features that have passed the comparison module, namely the first modal features and the second modal features, are connected in series to obtain the fusion features, which are expressed as Then, based on the prediction model, that is, the formula (9) mentioned above, the fusion feature is predicted, and the third loss function is calculated for the prediction result, using the following formula:

[0115]

[0116] Among them, y k is the predicted value of the kth sample, p kis the softmax output of the kth sample in the fusion feature. S640, determine the target loss function at least according to the first loss function, the second loss function and the third loss function. That is:

[0117]

[0118] In addition, since the dropout mechanism randomly discards some neurons during each pass, in order to improve the robustness of the model, we use the model to output two different prediction values and Regularized dropout (R-drop) regularizes the output predictions by minimizing the Kullback-Leibler (KL) divergence between the two output distributions, which is calculated as follows:

[0119]

[0120] Finally, the expression of the objective loss function is further optimized as follows:

[0121]

[0122] Finally, in order to prove the superiority of the relevant model finally obtained by this construction method, the model obtained by our construction method (hereinafter referred to as MCDSP) is compared with the following advanced deep learning models:

[0123] HypergraphSynergy: Using the biochemical features of drugs and cell lines as node attributes, the hypergraph is then used to capture and explain the multi-directional relationships between drug combinations and cell lines, and finally a multi-layer perceptron is used to predict drug synergy.

[0124] MatchMaker: Two parallel subnetworks, one for each drug, are designed to learn drug-specific representations on specific cell lines. The joint representation of drug pairs becomes the input to the third subnetwork, which learns to predict the synergistic effects of drug pairs.

[0125] KGANSynergy uses the hierarchical propagation of KG to find the neighbors of drugs / cell lines, and then uses the attention network to obtain the corresponding neighborhood representation of drugs or cell lines. A well-defined scoring function is then used to obtain the prediction score of potential synergy.

[0126] About the experimental setup:

[0127] In this embodiment, the collaborative data set is first randomly shuffled and divided into a cross-validation set and an independent test set at a ratio of 9:1. For the cross-validation set, a five-fold cross-validation is used to train the prediction model. Then, the five best results in the cross-validation stage are tested on the independent test set to evaluate the performance of the prediction model.

[0128] In addition, three practical scenarios on the cross-validation set are considered.

[0129] 1 Random division: The cross-validation set is randomly divided into 5 equal parts. This simulates the prediction of drug combinations that may have synergistic effects when anticancer drugs and cancer cell lines are known.

[0130] 2. Cell line level division: Random division is performed at the cell line level. Some cell lines in the collaborative dataset are used as validation sets to ensure that such cell lines are not included in the training set. The model's response to the emergence of new cell lines is simulated to evaluate its predictive ability for unseen cell lines.

[0131] 3. Drug combination level division: Random division at the drug combination level. Some drug pairs in the synergy dataset are used as validation sets to ensure that such drug combinations do not exist in the training set. The adaptability and accuracy of the model when encountering new drug combinations are examined.

[0132] In order to obtain more objective results, we use independent test sets to evaluate the performance of the model and the baseline in all three partitioning methods. We use the following five indicators to measure the performance of the model, including accuracy (ACC), area under the receiver operating characteristic curve (ROC-AUC), precision (Prec), recall (Rec), and F1 score. The calculation method of the evaluation indicators is as follows:

[0133]

[0134] Where TP is true positive, TN is true negative, FP is false positive, and FN is false negative. Through these evaluation indicators, we can accurately and comprehensively evaluate our prediction model.

[0135] We evaluated the performance of the model in three practical scenarios on two datasets, DrugCombDB and OncologyScreen. As shown in Table 1, our method outperforms the other three advanced methods in all performance indicators in the random partitioning scenario, and surpasses the best baseline model HypergraphSynergy, with an ACC score 1.69% higher and a ROC-AUC score 2.44% higher, which shows the powerful ability of our method in predicting synergistic drug combinations on large datasets. In terms of cell line level and drug combination level partitioning, most of the performance of our model is also due to the other baseline models, which shows that our model also has the best predictive performance when encountering unknown cell lines and drug combinations. Table 2 shows the performance of our model on small-scale datasets. The results show that KGANSynergy is more adaptable to small-scale datasets and has the best prediction effect. However, the overall performance of our model is second only to the baseline model.

[0136] Table 1: Performance comparison of MCDSP and baselines on DrugCombDB dataset

[0137]

[0138] Among them, the first, second, and third rows of each method correspond to randomization division, cell line level division, and drug combination level division, respectively.

[0139] Table 2: Performance comparison of MCDSP and baselines on the OncologyScreen dataset

[0140]

[0141] Among them, the first, second, and third rows of each method correspond to randomization division, cell line level division, and drug combination level division, respectively.

[0142] In a second aspect, the present application also provides a device for constructing a multi-modal comparative drug synergy prediction model, the device comprising:

[0143] A phenotypic characteristic acquisition module configured to acquire phenotypic characteristics of the first drug, the second drug, and the cell line;

[0144] an entity embedding feature acquisition module, configured to acquire entity embedding features corresponding to the first drug, the second drug, and the cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the first drug, the second drug, and the cell line and a preset first model;

[0145] a single-modal feature representation module configured to obtain, according to the entity embedding feature and the phenotypic feature, a first modal feature and a second modal feature for representing information about a drug-drug-cell line triplet in two different modalities in the same dimension;

[0146] A single-modal feature guided learning module is configured to obtain the effectiveness prediction results of the preset prediction model after predicting the drug synergy of each feature sample in the first modal feature and the second modal feature relative to the preset real sample; if the prediction results of the two feature samples corresponding to the same real sample in the first modal feature and the second modal feature are both correct, then both are determined to be sample positive pairs; if both are wrong, then both are determined to be sample negative pairs; if only one of them is correct, then both are determined to be sample semi-positive pairs; according to the sample positive pairs, sample negative pairs and sample semi-positive pairs, a target loss function is determined; it is judged whether the target loss function meets the preset requirements. If not, the target loss function is calculated again after adjusting the multiple models used in the previous step. If it meets the requirements, it is determined that the combination of multiple models used this time constitutes a multi-modal comparative drug synergy prediction model.

[0147] Among them, the functional implementation of each module in the above-mentioned multi-modal comparative drug synergistic prediction model construction device corresponds to the various steps in the above-mentioned multi-modal comparative drug synergistic prediction model construction method embodiment, and its functions and implementation processes will not be repeated here one by one.

[0148] In a third aspect, an embodiment of the present application further provides a multi-modal comparative drug synergy prediction method based on the model constructed by the method for constructing the multi-modal comparative drug synergy prediction model as described above, the prediction method comprising the following steps:

[0149] F100, obtain the phenotypic characteristics of target drug 1, target drug 2 and target cell line;

[0150] F200, obtaining entity embedding features corresponding to the target drug one, the target drug two, and the target cell line according to the heterogeneous graph consisting of drugs, targets, and cell lines involved in the target drug one, the target drug two, and the target cell line and the first model;

[0151] F300, according to the entity embedding feature and the phenotypic feature, obtaining a modal feature 1 and a modal feature 2 for representing drug-drug-cell line triple information in two different modalities in the same dimension;

[0152] Specifically, the following steps are included:

[0153] F310, dimensionally align different types of drugs and cell line features according to the preset projection equation;

[0154] F320, obtaining a first basic modality feature and a second basic modality feature representing drug-drug-cell line triple connection information according to the entity embedding feature and the phenotypic feature;

[0155] F330. Based on the first basic modal features, the second basic modal features and the trained second model, determine potential features that have an impact on drug synergy from both the first basic modal features and the second basic modal features, and obtain the first modal features and the second modal features based on the potential features of the two.

[0156] F400, connecting the modal feature 1 and the modal feature 2 in series to form a fusion feature;

[0157] F500. Predict drug synergy for each feature sample in the fusion feature according to the prediction model, and determine whether the target drug 1 and the target drug 2 have synergy in the target cell line.

[0158] In a fourth aspect, an embodiment of the present application further provides a prediction device for implementing the above-mentioned multi-modal comparison drug synergy prediction method, the prediction device comprising:

[0159] A phenotypic characteristic module, which is configured to obtain phenotypic characteristics of target drug one, target drug two, and target cell line;

[0160] An entity embedding feature module, which obtains entity embedding features corresponding to the target drug one, the target drug two, and the target cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the target drug one, the target drug two, and the target cell line and the first model;

[0161] A multimodal fusion prediction module is configured to obtain modal feature one and modal feature two representing drug-drug-cell line triple information in two different modalities in the same dimension according to the entity embedding feature and the phenotypic feature; concatenate the modal feature one and the modal feature two to form a fusion feature; predict drug synergy for each feature sample in the fusion feature according to the prediction model, and determine whether the target drug one and the target drug two have synergy in the target cell line.

[0162] Among them, the functional implementation of each module in the above-mentioned prediction device corresponds to the various steps in the above-mentioned multi-modality comparison drug synergistic prediction method embodiment, and its functions and implementation processes will not be repeated here one by one.

[0163] It should be noted that the serial numbers of the above-mentioned embodiments of the present application are only for description and do not represent the advantages or disadvantages of the embodiments.

[0164] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit "first", "second" and "third" to different types.

[0165] In the description of the embodiments of the present application, "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a specific way.

[0166] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; the “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.

[0167] In some processes described in the embodiments of the present application, multiple operations or steps that appear in a specific order are included, but it should be understood that these operations or steps may not be executed in the order in which they appear in the embodiments of the present application or in parallel, and the sequence number of the operation is only used to distinguish the different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed in sequence or in parallel, and these operations or steps may be combined.

[0168] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD) as described above, and includes a number of instructions for a terminal device to execute the methods described in each embodiment of the present application.

[0169] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for constructing a multi-modal comparative drug synergy prediction model, characterized in that: The construction method comprises: Obtaining phenotypic characteristics of the first drug, the second drug, and the cell line; According to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the first drug, the second drug, and the cell line and a preset first model, entity embedding features corresponding to the first drug, the second drug, and the cell line are obtained; According to the entity embedding feature and the phenotype feature, a first modality feature and a second modality feature of two different modalities representing drug-drug-cell line triple information in the same dimension are obtained; Obtaining the effectiveness prediction results of the preset prediction model after predicting the drug synergy for each feature sample in the first modal feature and the second modal feature relative to the preset real samples; If the prediction results of the two feature samples corresponding to the same real sample in the first modal feature and the second modal feature are both correct, then the two are determined to be sample positive pairs; if both are wrong, then the two are determined to be sample negative pairs; if only one of them is correct, then the two are determined to be sample semi-positive pairs; Determining a target loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs; Determine whether the target loss function meets the preset requirements. If not, adjust the multiple models used in the previous step and calculate the target loss function again. If it meets the requirements, determine that the multiple model combinations used this time constitute a multi-modal comparative drug synergistic prediction model.

2. The AAAA method according to claim 1, characterized in that In determining the target loss function according to the sample positive pair, the sample negative pair and the sample semi-positive pair, A target loss function is determined according to the cosine similarity of the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs and a preset temperature coefficient.

3. The AAAA method according to claim 2, characterized in that: In determining the target loss function according to the cosine similarity of the sample positive pair, the sample negative pair and the sample semi-positive pair and the preset temperature coefficient, the following formula is used: Wherein, n is a paired feature sample of the first modal feature and the second modal feature, P is a positive pair set, S is a semi-positive pair set, and B is a complete set of sample positive pairs, sample negative pairs, and the sample semi-positive pairs. is the first modal feature, is the second mode characteristic and τ is the temperature coefficient.

4. The AAAA method according to claim 1, wherein: Determining a target loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs comprises the following steps: Determining a first loss function according to the sample positive pairs, the sample negative pairs, and the sample semi-positive pairs; Determine a second loss function according to the effectiveness prediction results of each feature sample in the first modal feature and the second modal feature after predicting the drug synergy using the prediction model relative to a preset real sample; Determine a third loss function according to the effectiveness prediction result of the drug synergy prediction relative to the preset real sample after the fusion feature obtained by connecting the first modal feature and the second modal feature in series using the prediction model; The target loss function is determined based on at least the first loss function, the second loss function and the third loss function.

5. The AAAA method according to claim 4, characterized in that: The following formula is used to determine the second loss function based on the effectiveness prediction results of the preset real samples after the prediction of drug synergy is performed using the prediction model according to each feature sample in the first modal feature and the second modal feature: Among them, y k is the label of the kth sample in the first modality feature and the second modality feature, is the softmax output of the kth sample in the corresponding modal feature o.

6. The AAAA method according to claim 4, characterized in that: The third loss function is determined by using the following formula to predict the effectiveness of the drug synergy relative to the preset real sample after the fusion feature obtained by concatenating the first modal feature and the second modal feature is used by the prediction model: Among them, y k is the predicted value of the kth sample, p k is the softmax output of the kth sample in the fusion feature.

7. The AAAA method according to claim 1, wherein: The method of obtaining, according to the entity embedding feature and the phenotype feature, a first modality feature and a second modality feature of two different modalities representing drug-drug-cell line triple information in the same dimension comprises the following steps: Different types of drugs and cell line features are dimensionally aligned according to the preset projection equation; According to the entity embedding feature and the phenotype feature, a first basic modality feature and a second basic modality feature representing drug-drug-cell line triple connection information are obtained; According to the first basic modal feature, the second basic modal feature and a preset second model, potential features that have an impact on drug synergy are determined from both the first basic modal feature and the second basic modal feature, and the first modal feature and the second modal feature are obtained based on the potential features of the two.

8. A multi-modality comparative drug synergy prediction method based on the model constructed by the method for constructing a multi-modality comparative drug synergy prediction model according to any one of claims 1 to 7, characterized in that: The prediction method comprises the following steps: Obtaining phenotypic characteristics of target drug 1, target drug 2, and target cell line 3; According to the heterogeneous graph consisting of drugs, targets, and cell lines involved in the target drug one, the target drug two, and the target cell line and the first model, obtaining entity embedding features corresponding to the target drug one, the target drug two, and the target cell line; According to the entity embedding feature and the phenotype feature, a modal feature 1 and a modal feature 2 for representing drug-drug-cell line triple information in two different modalities in the same dimension are obtained; Connecting the modal feature 1 and the modal feature 2 in series to form a fusion feature; According to the prediction model, drug synergy prediction is performed on each feature sample in the fusion feature to determine whether the target drug 1 and the target drug 2 have synergistic effect in the target cell line.

9. A device for constructing a multi-modal comparative drug synergy prediction model, characterized in that: The construction device comprises: A phenotypic characteristic acquisition module configured to acquire phenotypic characteristics of the first drug, the second drug, and the cell line; an entity embedding feature acquisition module, configured to acquire entity embedding features corresponding to the first drug, the second drug, and the cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the first drug, the second drug, and the cell line and a preset first model; a single-modal feature representation module configured to obtain, according to the entity embedding feature and the phenotypic feature, a first modal feature and a second modal feature for representing information about a drug-drug-cell line triplet in two different modalities in the same dimension; A single-modal feature guided learning module is configured to obtain the effectiveness prediction results of the preset prediction model after predicting the drug synergy of each feature sample in the first modal feature and the second modal feature relative to the preset real sample; if the prediction results of the two feature samples corresponding to the same real sample in the first modal feature and the second modal feature are both correct, then both are determined to be sample positive pairs; if both are wrong, then both are determined to be sample negative pairs; if only one of them is correct, then both are determined to be sample semi-positive pairs; according to the sample positive pairs, sample negative pairs and sample semi-positive pairs, a target loss function is determined; it is judged whether the target loss function meets the preset requirements. If not, the target loss function is calculated again after adjusting the multiple models used in the previous step. If it meets the requirements, it is determined that the combination of multiple models used this time constitutes a multi-modal comparative drug synergy prediction model.

10. A prediction device for implementing the multi-modality comparative drug collaborative prediction method as claimed in claim 8, characterized in that: It includes: A phenotypic characteristic module, which is configured to obtain phenotypic characteristics of target drug one, target drug two, and target cell line; An entity embedding feature module, which obtains entity embedding features corresponding to the target drug one, the target drug two, and the target cell line according to a heterogeneous graph consisting of drugs, targets, and cell lines involved in the target drug one, the target drug two, and the target cell line and the first model; A multimodal fusion prediction module is configured to obtain modal feature one and modal feature two representing drug-drug-cell line triple information in two different modalities in the same dimension according to the entity embedding feature and the phenotypic feature; concatenate the modal feature one and the modal feature two to form a fusion feature; predict drug synergy for each feature sample in the fusion feature according to the prediction model, and determine whether the target drug one and the target drug two have synergy in the target cell line.

Images