Drug synergistic effect prediction method, device and system based on multiple views

Abstract

The invention discloses a method for training a combined learning prediction model of a drug synergistic effect, and the method comprises the steps: obtaining a synergistic data triple, a real synergistic score and a cell line expression profile, the synergistic data triple comprising a drug A, a drug B and a cell line C; the data of the collaborative data triple is preprocessed, to-be-trained drug feature data is obtained, and the to-be-trained drug feature data comprises a combined molecular graph of a drug combination AB of a drug A and a drug B, Morgan fingerprint features of the drug A and Morgan fingerprint features of the drug B; taking a combined molecular map of a drug combination AB of a drug A and a drug B, Morgan fingerprint characteristics of the drug A, Morgan fingerprint characteristics of the drug B, an expression profile of a cell line C and a real collaborative score as input characteristics of a combined learning prediction model, and taking collaborative scores of the drug A and the drug B on the cell line C as output characteristics; and training the combined learning prediction model. The combined learning prediction model can predict the drug synergistic effect more comprehensively and more accurately.

Description

Technical Field

[0001] The present application relates to the field of bioinformatics and deep learning technology, and in particular to a multi-view based drug synergy prediction method, device and system. Background Art

[0002] Traditional and modern medicine have long utilized drug combinations to better treat diseases. Compared to single drugs, drug combinations can improve efficacy, reduce toxicity and side effects, and lower drug resistance. However, in addition to synergistic effects, drug combinations can also produce antagonistic or additive effects. For example, the combination of statins with other chemotherapy drugs can compromise the therapeutic efficacy of chemotherapy drugs and lead to chemotherapy resistance in patients. Therefore, the precise discovery of synergistic drug combinations tailored to specific cancer cell lines is crucial for cancer treatment. With the increasing number of anticancer drugs being developed, the search space for drug combinations is exponentially expanding, making experimental screening difficult. Therefore, using computational methods to discover new drug combinations is a key task in precision cancer drug treatment research. Using computational tools to more efficiently predict drug synergies and identify synergistic drug combinations is highly valuable.

[0003] In recent years, high-throughput drug screening technology has developed rapidly, and the release of several large-scale drug combination datasets has made it possible to develop machine learning or deep learning models to predict drug synergy. O'Neil et al. published a large-scale drug pair synergy study, including 22,737 experimental measurements of 38 drugs on 39 cancer cell lines (O'Neil, J., Benita, Y., Feldman, I., Chenard, M., Roberts, B., Liu, Y., Li, J., Kral, A., Lejnine, S., Loboda, A., et al. (2016). An Unbiased Oncology Compound Screen to Identify Novel Combination Strategies. Molecular Cancer Therapeutics 15, 1155-1162.10.1158 / 1535-7163.MCT-15-0843). NCI published a larger drug pair collaboration study that included 304,549 experimental measurements of 104 drugs on 60 cancer cell lines (Holbeck, SL, Camalier, R., Crowell, JA, Govindharajulu, JP, Hollingshead, M., Anderson, LW, Polley, E., Rubinstein, L., Srivastava, A., Wilsker, D., et al. (2017). The National Cancer Institute ALMANAC: A Comprehensive Screening Resource for the Detection of Anticancer Drug Pairs with Enhanced Therapeutic Activity. Cancer Research 77, 3564-3576.10.1158 / 0008-5472.CAN-17-0489).On this basis, DrugComb collected and organized data sets from NCI-ALMANAC and O'Neil studies, and used a unified calculation method to calculate the Loewe additivity (LOEWE) synergy score for each experimental measurement of synergy (Zagidullin, B., Aldahdooh, J., Zheng, S., Wang, W., Wang, Y., Saad, J., Malyutina, A., Jafari, M., Tanoli, Z., Pessia, A., et al. (2019). DrugComb: an integrative cancer drug combination data portal. Nucleic Acids Research 47, W43-W51.10.1093 / nar / gkz337). Furthermore, thanks to advances in high-throughput sequencing technology, the Cancer Cell Line Encyclopedia project provides comprehensive characterizations of 1,072 cancer cell lines, including RNA sequencing, microRNA expression profiles, and global histone modification profiles (Ghandi, M., Huang, FW, Jané-Valbuena, J., Kryukov, GV, Lo, CC, McDonald, ER, Barretina, J., Gelfand, ET, Bielski, CM, Li, H., et al. (2019). Next-generation characterization of the Cancer Cell Line Encyclopedia. Nature 569, 503-508.10.1038 / s41586-019-1186-3). These resources provide a solid data foundation for predicting drug combination synergy.

[0004] Existing methods for predicting drug synergy are primarily based on machine learning and deep learning, but they still suffer from the following shortcomings: Existing models only model and concatenate individual drugs when embedding them, without jointly embedding the global information of drug combinations. Existing models also use only one drug feature in a single view to model, failing to consider drug synergy from multiple perspectives.

[0005] Therefore, there is an urgent need in this field to develop a multi-view based drug synergy prediction method, device and system, which develops a joint learning prediction model for drug synergy, which can predict drug synergy more comprehensively and accurately. Summary of the Invention

[0006] The purpose of this application is to provide a multi-view based drug synergy prediction method, device and system, which develops a joint learning prediction model for drug synergy, which can predict drug synergy more comprehensively and accurately.

[0007] A first aspect of the present application provides a method for training a joint learning prediction model for drug synergy, comprising the following steps:

[0008] Step S1: Obtaining a synergy data triplet, a true synergy score, and a cell line expression profile, wherein the synergy data triplet includes drug A, drug B, and cell line C, the cell line expression profile includes the expression profile of cell line C, and the true synergy score refers to the experimentally measured synergy score of drug A and drug B on cell line C;

[0009] Step S2: Preprocessing the data of the collaborative data triples to obtain drug feature data to be trained, wherein the drug feature data to be trained includes 1) a joint molecular graph of the drug combination AB of the drugs A and B, 2) a Morgan fingerprint feature of the drug A, and 3) a Morgan fingerprint feature of the drug B;

[0010] Step S3: Using 1) the joint molecular graph of the drug combination AB of the drug A and drug B, 2) the Morgan fingerprint feature of the drug A, 3) the Morgan fingerprint feature of the drug B, 4) the expression spectrum of the cell line C, and 5) the true synergy score as input features of the joint learning prediction model, and using the synergy score of the drug A and the drug B on the cell line C as the output feature, the joint learning prediction model is trained.

[0011] In another preferred embodiment, in step S2, the molecular graph of drug A and the molecular graph of drug B are connected by a hyperedge method, thereby constructing a joint molecular graph of the drug combination AB of drug A and drug B.

[0012] In another preferred example, the hyperedge method refers to adding a virtual edge as a hyperedge between the most important node in the molecular graph of drug A and the most important node in the molecular graph of drug B, thereby constructing a joint molecular graph of the drug combination AB of drug A and drug B based on the "hyperedge".

[0013] In another preferred embodiment, the most important nodes in the molecular graph of drug A and drug B are the two nodes with the highest betweenness centrality between drug A and drug B. The betweenness centrality is defined as follows:

[0014]

[0015] Among them, σ stis the number of shortest paths from node s to node t, and σ st (v) is the number of these shortest paths that pass through node v.

[0016] In another preferred example, the molecular graph of drug A and the molecular graph of drug B are obtained by converting the SMILES formula of drug A and the SMILES formula of drug B; the molecular graph of drug A and the molecular graph of drug B respectively include multiple nodes and edges, wherein the nodes represent atoms and the edges represent chemical bonds.

[0017] In another preferred embodiment, the SMILES formula of drug A and the SMILES formula of drug B are obtained from the PubChem database.

[0018] In another preferred embodiment, the O'Neil and NCI-ALMANAC benchmark datasets are selected from the DrugComb database to obtain the synergy data triples and true synergy scores, and the cell line expression profiles are obtained from the CCLE database.

[0019] In another preferred embodiment, in step S3, training the joint learning prediction model includes the following steps:

[0020] Step S31: inputting the joint molecular graph of the drug combination AB and the expression profile of the cell line C into the first branch network of the joint learning prediction model, thereby obtaining embedded features of the drug combination AB on the cell line C;

[0021] Step S32: Inputting the expression profile of the cell line C, the Morgan fingerprint features of drug A, and the Morgan fingerprint features of drug B into the second branch network of the joint learning prediction model, thereby obtaining a combined feature of the embedding of drug A and the embedding of drug B in the cell line C;

[0022] Step S33: Input the embedded features of the drug combination AB on the cell line C obtained in step S31 and the combined features of the representation of drug A and the embedded drug B on the cell line C obtained in step S32 into the dual-view integration network of the joint learning prediction model, thereby obtaining a representation of drug synergy.

[0023] In another preferred embodiment, the dual-view integrated classification network is an MLP model.

[0024] In another preferred example, the first branch network and the second branch network include a graph attention network layer and a fully connected network layer.

[0025] The second aspect of the present application provides a method for predicting drug synergy, comprising the following steps:

[0026] Providing a joint learning prediction model for drug synergy, wherein the joint learning prediction model is obtained by training using the above method;

[0027] Providing expression profiles of a first drug to be predicted, a second drug to be predicted, and a cell line to be predicted;

[0028] Preprocessing the first drug to be predicted and the second drug to be predicted to obtain 1) a joint molecular graph of the drug combination of the first drug to be predicted and the second drug to be predicted, 2) a Morgan fingerprint feature of the first drug to be predicted, and 3) a Morgan fingerprint feature of the second drug to be predicted

[0029] 1) a joint molecular graph of the drug combination of the first drug to be predicted and the second drug to be predicted, 2) the Morgan fingerprint feature of the first drug to be predicted, 3) the Morgan fingerprint feature of the second drug to be predicted, and 4) the expression profile of the cell line to be predicted are input into the joint learning prediction model to obtain a predicted synergy score of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted.

[0030] In another preferred example, if the synergy scores of the first drug to be predicted and the second drug to be predicted on the cell line C to be predicted are known, the synergy scores of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted output by the joint learning prediction model can be used to evaluate the performance of the joint learning prediction model.

[0031] In another preferred example, if the synergistic scores of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted C are unknown, the output of the joint learning prediction model is the predicted synergistic scores of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted.

[0032] In another preferred example, the classification task of the joint learning prediction model outputs whether the first drug to be predicted and the second drug to be predicted are synergistic or antagonistic in the cell line to be predicted; the regression task of the joint learning prediction model outputs the synergistic score value of the first drug to be predicted and the second drug to be predicted in the cell line to be predicted.

[0033] In another preferred embodiment, in the classification task, Kappa and F1 are used as main evaluation indicators.

[0034] In another preferred embodiment, multiple hyperparameters are included in steps S1 to S3, including but not limited to loss function, activation function, learning rate, regularization method, number of neural network layers and number of hidden layer neurons.

[0035] A second aspect of the present application provides a device for training a joint learning prediction model for drug synergy, comprising:

[0036] an acquisition module configured to acquire a synergy data triplet, a true synergy score, and a cell line expression profile, wherein the synergy data triplet includes drug A, drug B, and cell line C, the cell line expression profile includes an expression profile of cell line C, and the true synergy score refers to an experimentally measured synergy score of drug A and drug B on cell line C;

[0037] a data preprocessing module configured to preprocess the data of the collaborative data triples to obtain drug feature data to be trained, wherein the drug feature data to be trained includes 1) a joint molecular graph of the drug combination AB of the drugs A and B, 2) a Morgan fingerprint feature of the drug A, and 3) a Morgan fingerprint feature of the drug B;

[0038] The training module uses 1) the joint molecular graph of the drug combination AB of the drug A and drug B, 2) the Morgan fingerprint feature of the drug A, 3) the Morgan fingerprint feature of the drug B, 4) the expression profile of the cell line C, and 5) the true synergy score as input features of the joint learning prediction model, and uses the synergy score of the drug A and the drug B on the cell line C as an output feature to train the joint learning prediction model.

[0039] A third aspect of the present application provides a device for training a joint learning prediction model for drug synergy, comprising:

[0040] a memory for storing computer-executable instructions; and

[0041] A processor, coupled to the memory, configured to implement the steps of the method according to any one of claims 1 to 8 when executing the computer-executable instructions.

[0042] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps in the method according to any one of claims 1 to 8 are implemented.

[0043] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features described in detail below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be listed here one by one. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. It should be understood that the drawings described below are merely some implementation examples of the present invention, and those skilled in the art can also derive other implementation examples based on these drawings without inventive effort.

[0045] Figure 1 1 is a flow chart of a method for training a joint learning prediction model for drug synergy according to the first embodiment of the present application;

[0046] Figure 2 A framework diagram of a joint learning prediction model for drug synergy according to one embodiment of the present application is shown;

[0047] Figures 3A-3D Four examples of methods for building a joint graph of drug combinations are shown, where Figure 3A Schematic diagram showing the use of hyperedge method to build a joint molecular graph of drug combinations, Figure 3B Schematic diagram showing the construction of a joint molecular graph for drug combinations using the all-nodes approach, Figure 3C A schematic diagram showing the use of a supernode method to build a joint molecular graph of drug combinations, and Figure 3D A schematic diagram showing the use of a 3-supernode approach to build a joint molecular graph of drug combinations;

[0048] Figure 4 4-A shows the performance of each model on the O'Neil benchmark dataset in the classification task according to an embodiment of the present application, 4-B shows the performance of each model on the O'Neil benchmark dataset in the regression task, 4-C shows the performance of each model on the NCI-ALMANAC benchmark dataset in the classification task, and 4-D shows the performance of each model on the NCI-ALMANAC benchmark dataset in the regression task;

[0049] Figure 5 A schematic diagram of a joint molecular map of the drug combination of erlotinib and vorinostat according to Example 1 of the present application is shown, wherein red represents vorinostat and blue represents erlotinib;

[0050] Figure 6 A schematic diagram of a combined molecular diagram of the erlotinib and dactolisib drug combination according to Example 1 of the present application is shown, wherein red represents dactolisib and blue represents erlotinib;

[0051] Figure 7A schematic diagram of a joint molecular diagram of the drug combination of erlotinib and MK-2206 according to Example 1 of the present application is shown, wherein red represents MK-2206 and blue represents erlotinib;

[0052] Figure 8A shows an estimated spectrum of a synergistic drug combination for lung cancer according to Example 1 of the present application;

[0053] Figure 8B shows a collaborative scoring heat map according to Example 1 of the present application;

[0054] Figure 8C The figure shows the predicted synergy scores of "erlotinib and vorinostat" in NSCLC cell lines and all cell lines according to Example 1 of the present application, and grouped by whether or not there is EGFR mutation, EGFR_MUT indicates EGFR mutation, and EGFR_WT indicates no EGFR mutation;

[0055] Figure 8D The figure shows the predicted synergy scores of "erlotinib and dactolisib" in NSCLC cell lines and all cell lines according to Example 1 of the present application, and grouped by whether or not there is EGFR mutation, EGFR_MUT indicates EGFR mutation, and EGFR_WT indicates no EGFR mutation;

[0056] Figure 8E The figure shows the predicted synergistic scores of "erlotinib and MK-2206" in NSCLC cell lines and all cell lines according to Example 1 of the present application, and grouped by whether or not there is EGFR mutation, EGFR_MUT indicates EGFR mutation, and EGFR_WT indicates no EGFR mutation. DETAILED DESCRIPTION

[0057] Through extensive and in-depth research, the inventors have developed for the first time a novel deep learning method, JointSyn, which characterizes the synergistic effect of drug combinations from two perspectives: the embedding of drug combination on cellline and the combination of drug embedding on cell line. In addition, we use hyperedges to jointly characterize the global graph information of drug combinations. On the O'Neil and NCI-ALMANAC benchmark datasets, we compared JointSyn with classic and state-of-the-art models in the prior art (such as DeepSynergy, AuDNNsynergy, DeepDDS, DTSyn and HypergraphSynergy, the five most advanced deep learning methods, and XGBoost, the classic machine learning method). The results of the study on a wide range of benchmark datasets show that JointSyn outperforms other methods in eight cross-validation tasks, especially in unseen drug combination tasks.

[0058] the term

[0059] As used herein, the terms "joint molecular graph" and "global molecular graph" are used interchangeably;

[0060] As used herein, the terms "drug A" and "drug B" are used merely to distinguish one drug from another and do not have a limiting effect;

[0061] As used herein, the terms “representation” and “embedding” mean the same thing and can be used interchangeably.

[0062] The term “JointSyn” refers to “Joint Learning Prediction Model for Drug Synergy”;

[0063] It should be noted that, in this application, relational terms such as first and second are merely used to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or device. Without further restriction, an element defined by the phrase "comprising a" does not exclude the presence of other identical elements in the process, method, article, or device comprising the element. In this application, if it is mentioned that an action is performed according to a certain element, it means that the action is performed at least according to that element, including two situations: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "multiple," and "multiple" include 2, 2 times, 2 kinds, and more than 2, more than 2 times, and more than 2 kinds.

[0064] This application has at least one of the following advantages

[0065] (a) The joint learning prediction model for drug synergy in this application provides a new method for drug synergy prediction, providing a complete input-to-output process that can be applied to drug synergy screening experiments as a more efficient computational tool;

[0066] (b) The joint molecular graph of the drug combination AB of drug A and drug B input to the joint learning prediction model of drug synergy of the present application is a joint molecular graph (global molecular graph) of the drug combination AB containing global information constructed for drug A and drug B by using the "hyperedge" method;

[0067] (c) The joint learning prediction model for drug synergy of the present application models drug synergy from two perspectives: the embedding of drug combinations on cell lines and the combination of drug embeddings on cell lines. These two perspectives can predict drug synergy more comprehensively and accurately.

[0068] (d) The joint learning prediction model of drug synergy of the present application uses the relevant features of drug A, drug B and cell line C. The system can predict the synergy score of the combined use of drug A and drug B on cell line C, and then infer whether the drug combination is synergistic or antagonistic on the cell line;

[0069] (e) Experimental conclusions show that the joint learning prediction model of drug synergy (JointSyn) of this application can output qualitative (whether the drug combination is synergistic or antagonistic in the cell line) and quantitative (the synergy score value of the drug combination in the cell line) results, and is an important tool for achieving precise identification of cancer drug combinations.

[0070] In the following description, many technical details are provided to help readers better understand this application. However, those skilled in the art will understand that even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can be implemented.

[0071] A method for training a joint learning prediction model for drug synergy

[0072] The first embodiment of the present application provides a method for training a joint learning prediction model (JointSyn) of drug synergy, comprising the following steps:

[0073] Step S1: Obtaining a synergy data triplet, a true synergy score, and a cell line expression profile, wherein the synergy data triplet includes drug A, drug B, and cell line C, the cell line expression profile includes the expression profile of cell line C, and the true synergy score refers to the experimentally measured synergy score of drug A and drug B on cell line C;

[0074] Step S2: Preprocessing the data of the collaborative data triples to obtain drug feature data to be trained, wherein the drug feature data to be trained includes 1) a joint molecular graph of the drug combination AB of the drugs A and B, 2) a Morgan fingerprint feature of the drug A, and 3) a Morgan fingerprint feature of the drug B;

[0075] Step S3: Using 1) the joint molecular graph of the drug combination AB of the drug A and drug B, 2) the Morgan fingerprint feature of the drug A, 3) the Morgan fingerprint feature of the drug B, 4) the expression spectrum of the cell line C, and 5) the true synergy score as input features of the joint learning prediction model, and using the synergy score of the drug A and the drug B on the cell line C as the output feature, the joint learning prediction model is trained.

[0076] Preferably, in step S3, training the joint learning prediction model includes the following steps:

[0077] Step S31: inputting the joint molecular graph of the drug combination AB and the expression profile of the cell line C into the first branch network of the joint learning prediction model, thereby obtaining the embedding features of the drug combination AB on the cell line C;

[0078] Step S32: inputting the expression profile of the cell line C, the Morgan fingerprint features of drug A, and the Morgan fingerprint features of drug B into the second branch network of the joint learning prediction model, thereby obtaining a combined feature of the representation of drug A and the representation of drug B on the cell line C;

[0079] Step S33: Input the features of the characterization of the drug combination AB on the cell line C obtained in step S31 and the features of the combination of the characterization of drug A and the characterization of drug B on the cell line C obtained in step S32 into the dual-view integrated classification network of the joint learning prediction model, thereby obtaining the characterization of drug synergy.

[0080] Preferably, the dual-view integrated classification network is an MLP model.

[0081] Preferably, the first branch network and the second branch network include a graph attention network layer and a fully connected network layer.

[0082] Preferably, in step S31, the method further includes: inputting the joint molecular graph of the drug combination AB into the first sub-network model, thereby obtaining a representation / embedding of the drug combination AB,

[0083] The expression profile of cell line C is input into the second sub-network model to obtain the embedding of the expression profile of cell line C.

[0084] The embedding of the drug combination AB and the embedding of the expression profile of the cell line C are spliced and input into the third sub-network, thereby obtaining the characteristics representing the drug combination AB on the cell line C.

[0085] Preferably, in step S32, the method further includes: inputting the Morgan fingerprint feature of the drug A into the fourth sub-network model, thereby obtaining the embedding of the Morgan fingerprint feature of the drug A.

[0086] The moiré fingerprint feature of the drug B is input into the fifth sub-network model to obtain the moiré fingerprint feature embedding of the drug B.

[0087] The moiré fingerprint feature embedding of the drug A and the embedding of the expression profile of the cell line C are concatenated and input into the fifth sub-network model to obtain the representation / embedding of the drug A on the cell line C.

[0088] The Morgan fingerprint feature embedding of the drug B and the embedding of the expression profile of the cell line C are spliced and input into the sixth sub-network model to obtain the representation / embedding of the drug B on the cell line C.

[0089] The representation of drug A on the cell line C and the representation of drug B on the cell line C are spliced and input into the seventh sub-network model of the joint learning prediction model, thereby obtaining the combined features of the representations of drug A and drug B on the cell line C.

[0090] Preferably, the first sub-network model is a graph attention network layer, and the second to seventh sub-network models are all fully connected network layers.

[0091] A method for predicting drug synergy

[0092] The second embodiment of the present application provides a method for predicting drug synergy, comprising:

[0093] Providing a joint learning prediction model for drug synergy, wherein the joint learning prediction model is obtained by training using the above method;

[0094] Obtaining expression profiles of a first drug to be predicted, a second drug to be predicted, and a cell line to be predicted;

[0095] Preprocessing the first drug to be predicted and the second drug to be predicted to obtain 1) a joint molecular graph of the drug combination of the first drug to be predicted and the second drug to be predicted, 2) a Morgan fingerprint feature of the first drug to be predicted, and 3) a Morgan fingerprint feature of the second drug to be predicted

[0096] 1) a joint molecular graph of the drug combination of the first drug to be predicted and the second drug to be predicted, 2) the Morgan fingerprint feature of the first drug to be predicted, 3) the Morgan fingerprint feature of the second drug to be predicted, and 4) the expression profile of the cell line to be predicted are input into the joint learning prediction model to obtain a predicted synergy score of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted.

[0097] Preferably, if the synergy scores of the first drug to be predicted and the second drug to be predicted on the cell line C to be predicted are known, the synergy scores of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted output by the joint learning prediction model can be used to evaluate the performance of the joint learning prediction model.

[0098] Preferably, if the synergy scores of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted are unknown, the output of the joint learning prediction model is the predicted synergy scores of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted.

[0099] A device for training a joint learning prediction model for drug synergy

[0100] The third embodiment of the present application provides a device for training a joint learning prediction model for drug synergy, comprising:

[0101] an acquisition module configured to acquire a synergy data triplet, a true synergy score, and a cell line expression profile, wherein the synergy data triplet includes drug A, drug B, and cell line C, the cell line expression profile includes an expression profile of cell line C, and the true synergy score refers to an experimentally measured synergy score of drug A and drug B on cell line C;

[0102] a data preprocessing module configured to preprocess the data of the collaborative data triples to obtain drug feature data to be trained, wherein the drug feature data to be trained includes 1) a joint molecular graph of the drug combination AB of the drugs A and B, 2) a Morgan fingerprint feature of the drug A, and 3) a Morgan fingerprint feature of the drug B;

[0103] The training module uses 1) the joint molecular graph of the drug combination AB of the drug A and drug B, 2) the Morgan fingerprint characteristics of the drug A, 3) the Morgan fingerprint characteristics of the drug B, the expression profile of the cell line C, and 5) the true synergy score as input features of the joint learning prediction model, and uses the synergy score of the drug A and the drug B on the cell line C as an output feature to train the joint learning prediction model.

[0104] A drug synergy prediction system based on multiple views

[0105] The fourth embodiment of this application provides a multi-view drug synergy prediction system. During operation, this system requires drug and cell line features. Drug SMILES formulas from PubChem can be converted into molecular graphs and Morgan fingerprint features as drug features; expression profiles of cell lines from CCLE serve as cell line features.

[0106] The system uses datasets from two studies on drug synergy, NCI-ALMANAC and O'Neil, as the original training data. Each piece of data can be represented as the synergy score between drug A and drug B on cell line C. Using the "hyperedge" method, the molecular graphs of drug A and drug B are connected to form a global molecular graph of drug combination AB. Using the global molecular graph of drug combination AB and the expression profile of cell line C as input, the embedding of the drug combination on the cell line can be obtained after passing through a graph neural network and a fully connected neural network. Using the Morgan fingerprint of drug A and the expression profile of cell line C as input, and splicing the Morgan fingerprint of drug B and the expression profile of cell line C at the same time, a combination of drug embeddings on the cell line can be obtained. Through the embedding of these two views, the final synergy score can be obtained through a classifier composed of a fully connected neural network.

[0107] During the training phase, the reverse gradient propagation method is used to train the neural network based on the original triplet data and the corresponding synergy scores. During the inference phase, only the names of drug A, drug B, and cell line C need to be input. The corresponding features can be obtained from the library and input into the neural network. The system can then predict the synergy score of drug A and drug B on cell line C.

[0108] The system consists of the following parts: data preprocessing module, data conversion module, joint molecular graph construction module, feature extraction module (two views), and collaborative score output module.

[0109] Data preprocessing module: drug SMILES formulas are collected from the PubChem database, cell line expression profiles are collected from the CCLE database, and triplets (drug A, drug B, and cell line C) and synergy scores are collected from the DrugComb database.

[0110] Data conversion module: converts the drug SMILES formula into molecular graphs and Morgan fingerprint features as drug features, and normalizes the expression profile data of the cell line. At the same time, the data is normalized;

[0111] Joint molecular graph construction module: Using the "hyperedge" method, the molecular graphs of drug A and drug B are connected to form the global molecular graph of drug combination AB.

[0112] Feature extraction module: The global molecular graph of drug combination AB and expression profile of cell line C are used as visual Figure 1 The input of drug A and expression profile of cell line C can be used as the visual input to extract the embedded features of drug combination on cell line. Figure 2 As an input, the embedding of drug A on cell line C can be extracted, and the Morgan fingerprint of drug B and the expression profile of cell line C are used as visual Figure 2As an input, we can extract the embedding of drug B on cell line C. Combining the two embeddings, we can extract the combined features of the drug embeddings on the cell line.

[0113] Synergy score output module: During the training phase, this module outputs the data-annotated synergy scores (synergy scores experimentally measured in publicly available datasets). Backpropagation is used to converge the entire network, allowing it to be used for inference. During the inference phase, the model's predicted synergy scores for drug A and drug B on cell line C are output.

[0114] In order to make the purpose, technical solution and advantages of the present invention more clear, an embodiment of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that these are only some examples that the present invention may take, but are not intended to limit the scope of the present invention.

[0115] Joint learning prediction model for drug synergy (deep learning model JointSyn)

[0116] This application proposes a deep learning model JointSyn (Joint Synergy) to predict the synergistic effect of drug combinations. First, we divide the synergistic effect of drug combinations into two sources, namely the representation of drug combinations on cell lines (embedding of drug combination on cell line) and the combination of drug representations on cell lines (combination of drug embedding on cell line). Next, in order to better represent the global graph information of drug combinations, we propose to form a global molecular graph (joint molecular graph) of drug combinations through hyperedges of two drug molecular graphs. The joint representation of drug combinations is obtained using a graph convolutional network, and integrated with the gene expression profile information of cell lines to form the characteristics of the representation of drug combinations on cell lines. Afterwards, we embed the Morgan fingerprint features of drugs and the gene expression profile information of cell lines into vectors to form the representation of drugs on cell lines, and splice the drug representations of the two drugs on cell lines to obtain the combination of drug representations on cell lines (for example, the characteristics of the combination of the representation of drug A and the representation of drug B on cell line C). Finally, we use DNN to predict the synergistic effect of drug combinations for specific cell lines based on the embeddings from the two sources. We use two benchmark datasets, NCI-ALMANAC and O'Neil, and show that JointSyn outperforms other state-of-the-art deep learning methods in eight cross-validated tasks.

[0117] Method for building the deep learning model JointSyn

[0118] This application aims to extract global information of drug combinations through "hyperedges", and on this basis, model drug synergy through multiple views, make more accurate predictions of drug synergy, and more efficiently discover synergistic drug combinations and cell lines.

[0119] The following describes the process of establishing a method for the deep learning model JointSyn provided by this application.

[0120] 1.1 Data

[0121] 1.1.1 Synergy Data

[0122] Synergy data consists of drug-drug-cell line triplets from DrugComb. We used data from the O'Neil and NCI-ALMANAC studies in DrugComb. The O'Neil dataset consists of 38 drugs and 39 cancer cell lines, including 24 FDA-approved drugs and 14 experimental drugs (O'Neil, J., Benita, Y., Feldman, I., Chenard, M., Roberts, B., Liu, Y., Li, J., Kral, A., Lejnine, S., Loboda, A., et al. (2016). An Unbiased Oncology Compound Screen to Identify Novel Combination Strategies. Molecular Cancer Therapeutics 15, 1155-1162.10.1158 / 1535-7163.MCT-15-0843). The NCI-ALMANAC dataset consists of 104 FDA-approved drugs and 60 cancer cell lines (Holbeck, SL, Camalier, R., Crowell, JA, Govindharajulu, JP, Hollingshead, M., Anderson, LW, Polley, E., Rubinstein, L., Srivastava, A., Wilsker, D., et al. (2017). The National Cancer Institute ALMANAC: A Comprehensive Screening Resource for the Detection of Anticancer Drug Pairs with Enhanced Therapeutic Activity. Cancer Research 77, 3564-3576.10.1158 / 0008-5472.CAN-17-0489).

[0123] It is worth noting that the original data contain experimental data from repeated measurements of the same triplets. Therefore, to filter out experimental noise, we preprocess the synergy scores. If the synergy scores of a triplet are inconsistent in sign across repeated experimental measurements, that is, the synergy score is positive in one experiment and negative in another, we delete this triplet. If the synergy scores of a triplet are consistent in sign across repeated experimental measurements and the coefficient of variation is greater than 0.5, we delete this triplet. If the synergy scores of a triplet are consistent in sign across repeated experimental measurements and the coefficient of variation is less than 0.5, we take the median of the multiple measurements as the synergy score for this triplet.

[0124] 1.1.2 Drug features

[0125] To better represent the molecular structure and physicochemical properties of drugs, we use drug molecular graphs and Morgan fingerprints as drug features. Drug SMILES were obtained from PubChem. First, we used RDKit to convert the drug SMILES into a molecular graph, where vertices represent atoms and edges represent chemical bonds (Landrum, G.RDKitDocumentation). For the features of each atomic node in the drug molecule graph, we use DeepChem (Ramsundar, B., Eastman, P., Walters, P., and Pande, V. (2019). Deep Learning for the Life Sciences: Applying Deep Learning to Genomics, Microscopy, Drug Discovery, and More (O'Reilly Media, Inc.)) and follow the work of Wang et al. (Wang, J., Liu, X., Shen, S., Deng, L., and Liu, H. (2022). DeepDDS: deep graph neural network with attention mechanism to predict synergistic drug combinations. Briefings in Bioinformatics 23,bbab390.10.1093 / bib / bbab390), and calculated the atomic features (nodefeatures); secondly, we used RDKit to calculate the Morgan fingerprint and descriptors based on the SMILES expression of the drug, and drew on the work of Preuer et al. (Preuer, K., Lewis, RPI, Hochreiter, S., Bender, A., Bulusu, KC, and Klambauer, G. (2018). DeepSynergy: predicting anti-cancer drug synergy with Deep Learning. Bioinformatics 34, 1538-1546.10.1093 / bioinformatics / btx806), and generated a Morgan fingerprint with a radius of 6 for each drug and represented it as a 1039-dimensional binary vector (Landrum, G. RDKit Documentation).

[0126] 1.1.3 Cell line features

[0127] In this application, expression profiles of cell lines were extracted from the Cancer Cell Line Encyclopedia (CCLE), and the transcripts per million (TPM) values of gene expression were log2 transformed and z-scores were standardized as cell line information (Barretina, J., Caponigro, G., Stransky, N., Venkatesan, K., Margolin, AA, Kim, S., Wilson, CJ, Lehár, J., Kryukov, GV, Sonkin, D., et al. (2012). The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drugs ensitivity. Nature 483, 603-607.10.1038 / nature11003). We only considered the 2128 genes screened by PaccMann based on expression profiles and PPI networks (Manica, M., Oskooei, A., Born, J., Subramanian, V., Sáez-Rodríguez, J., and Rodríguez Martínez, M. (2019). Toward Explainable Anticancer Compound Sensitivity Prediction via Multimodal Attention-Based Convolutional Encoders. Mol. Pharmaceutics 16, 4797-4806.10.1021 / acs.molpharmaceut.9b00520).

[0128] 1.2 Joint Learning Prediction Model for Drug Synergy (JointSyn)

[0129] 1.2.1 Model Architecture

[0130] In this application, we propose a novel deep learning method called JointSyn to predict the synergy scores of drug combinations in cell lines. The framework of JointSyn is as follows: Figure 2As shown in Figure 2, JointSyn takes as input the joint molecular graph of the drug combination, the drug's Morgan fingerprint, and the expression profile of the cell line. JointSyn consists of two views: a graph neural network view that extracts the embedding of the drug combination on the cell line, and a Morgan fingerprint network view that extracts the drug combination embedding on the cell line. The prediction network then uses the embeddings learned from these two views to predict the synergy score of the drug combination on the cell line. More details about JointSyn are provided below.

[0131] 1.2.2 Embedding of drug combination on cell line

[0132] As mentioned earlier, we use RDKit to convert SMILES into molecular graphs, where nodes are atoms and edges are chemical bonds. Specifically, the graph of a drug is defined as G = (V, E), where V is a set of N nodes represented by a C-dimensional vector, and the i-th atom can be represented as v i ∈V; E is the set of edges, the adjacency matrix is A, and the chemical bond between the i-th and j-th atoms can be expressed as e i,j ∈E, which can also be expressed as <v i ,v j >∈E.

[0133] In this view, we want to get the embedding of drug combination on cell line, so the first task is to represent the drug combination with a joint graph. In the drug combination, we have drug AG A =(V A ,E A ) and drug BG B =(V B ,E B ), we want to get the joint graphG of drug A and drug B AB =(V AB ,E AB ).like Figures 3A-3D As shown in Figure 3, we propose four methods to build a joint graph of drug combinations.

[0134] Super Edge:

[0135] like Figure 3A As shown in , we added a "hyperedge" so that atomic information can be transferred between drugs, and the neural network can capture cross-drug interactions. We use Betweenness Centrality to measure the importance of nodes connecting to other nodes in the network. Betweenness centrality is defined as follows:

[0136]

[0137] Among them, σ st is the number of shortest paths from node s to node t, and σ st (v) is the number of these shortest paths that pass through node v.

[0138] Select two nodes with the highest drug intermediary centrality, namely and Add an edge between these two nodes, which we call a "super edge". AB =V A ∪V B ,E AB =E A ∪E B ∪{(v1,v2)}

[0139] All Nodes:

[0140] like Figure 3B As shown in , we based on drug AG A and drug BG B Construct a bipartite graph between two drugs Specifically, drug G A Each atom in the drug G B Each atom in V is connected by an edge. AB =V A ∪V B , One Supernode:

[0141] like Figure 3C As shown in , we define a super node (SUP) to collect information about two drugs, and the super node is connected to each atom in the two drugs. At this time:

[0142] V AB =V A ∪V B ∪{v Sup},

[0143]

[0144] where n A represents the number of atoms in drug A, n B Represents the number of atoms in drug B.

[0145] Three Super nodes:

[0146] like Figure 3D As shown in , we define three supernodes to collect information about two drugs. Supernode 1 (SUP1) is connected to all atoms in drug A, and supernode 2 (SUP2) is connected to all atoms in drug B. After aggregating the information of drug A and drug B, SUP1 and SUP2 are aggregated by supernode 3 (SUP3). At this time:

[0147] V AB =V A ∪V B ∪{v Sup1 ,v Sup2 ,v Sup3},

[0148]

[0149] where n A represents the number of atoms in drug A, n B Represents the number of atoms in drug B.

[0150] Extracting drug combination features based on graph attention network

[0151] We use the GAT model based on the multi-head attention architecture to transform the joint graphG AB =(V AB ,E AB ) as input, its atomic node feature matrix is X, and its adjacency matrix is A. The output features of the node after each layer of iterative propagation are as follows:

[0152]

[0153] Among them, k is the number of attention heads, || concatenates the output results of multiple attention mechanisms, W (k) is a learnable weight matrix. The attention coefficient between each atom i and its neighbor atom j The calculation is as follows:

[0154]

[0155] Among them, W (k) It is the weight matrix shared with the above, the purpose is to enhance the features of the vertex, || splicing the enhanced features of vertices i and j, is a learnable weight vector, T is a transpose operation, and LeakyReLU is a nonlinear activation function.

[0156] Our model is based on a three-layer GAT network, with the three GAT layers connected by ReLU activation functions. Through the three-layer GAT, each atom can see its three-hop neighbors, and through the joint graph we constructed, atomic information can be transferred between drugs. We add a Global Pooling layer after the last GAT layer to aggregate the learned multiple atomic features to obtain the embedding of the drug combination. Extracting cell line features based on MLP

[0157] Although we only considered the 2128 genes selected by PaccMann based on expression profiles and PPI networks, the current cell line data is still high-dimensional and sparse. We used the 2128 genes selected by PaccMann and the 2087 genes in the intersection of CCLE gene expression profiles as feature inputs for the cell lines and used a two-layer MLP to construct a low-dimensional dense embedding for the cell lines. The calculation process is as follows:

[0158]

[0159] Among them, W cell and b cell Is a learnable weight matrix, in the first layer, It is the expression profile data of 2087 genes, and ReLU is a nonlinear activation function. Finally, the embedding of the cell line is obtained. Embedding of drug combination on cell line

[0160] The drug combination embedding is obtained through GAT, and the cell line embedding is obtained through MLP. After connecting them and passing through the fully connected layer, the embedding of drug combination on cell line can be obtained. The calculation process is as follows:

[0161]

[0162] 1.2.3 Combination of drug embedding on cell line: Extracting drug features based on MLP

[0163] We take the Morgan fingerprint feature X of the drug's 1039-dimensional binary vector drug Embedding with cell lines Splicing, input into the MLP model, get drug embedding on cell line The calculation process is as follows:

[0164]

[0165] We will use two drugs and Splicing, through the fully connected layer, we can get the combination of drug embedding on cell line The calculation process is as follows:

[0166]

[0167] 1.2.4 Predicting Drug Synergy Based on Dual Views

[0168] Through the above two methods, we can obtain the dual-view embedding of drug synergy. By concatenating these two embeddings and inputting them into the three-layer MLP, we can obtain the final embedding of drug synergy.

[0169]

[0170] Among them, W l and b l Is a learnable weight matrix, in the first layer,

[0171] Embedding through drug synergy We can get the final prediction value:

[0172]

[0173] Among them, W out and b out is a learnable weight matrix. In classification tasks, σ is the softmax activation function; in regression tasks, σ is the linear activation function.

[0174] 1.2.5 Model Hyperparameters

[0175] In the process of building the JointSyn model, there are many hyperparameters (see Table 1 for detailed hyperparameter settings for JointSyn), including but not limited to the loss function, activation function, learning rate, regularization method, number of neural network layers, and number of hidden layer neurons. We use grid search to adjust hyperparameters and also perform ten five-fold cross-validations to increase the robustness of the grid search. The cross entropy loss function is used as the loss function for classification tasks, and the MSE loss function is used as the loss function for regression tasks. The output layer uses the softmax activation function in classification tasks, and the linear activation function in regression tasks. The hidden layers all use the relu activation function. We use the mini batch method for training, with each batch size of 128 and a maximum training epoch of 1000.

[0176] Table 1 Detailed hyperparameter settings of JointSyn

[0177]

[0178] 1.3 Method Evaluation

[0179] 1.3.1 Method Comparison

[0180] We compared JointSyn with several state-of-the-art drug synergy prediction methods. In addition to five advanced deep learning methods, DeepSynergy, AuDNNsynergy, DeepDDS, DTSyn, and HypergraphSynergy, we also selected the classic XGBoost machine learning method. All experimental results were obtained from the same data input as JointSyn and based on two benchmark datasets, O'Neil and NCI-ALMANAC. XGBoost is a classic and recognized effective machine learning model; DeepSynergy is the first deep learning model to predict drug combination synergy; AuDNNsynergy is a deep learning model pre-trained on a large number of cell lines; DeepDDS, DTSyn, and HypergraphSynergy are all state-of-the-art models published in 2022. DeepDDS is a deep learning model based on graph neural networks and attention mechanisms; DTSyn is a Dual Transformer encoder model based on a multi-head attention mechanism; HypergraphSynergy is a multi-way relation-enhanced hypergraph representation learning deep learning model.

[0181] 1.3.2 Evaluation Metrics

[0182] In the classification task, we adopt Kappa and F1 as the main evaluation indicators. We also report other widely used classic performance indicators, including Area under the ROC Curve (ROC AUC), Precision-Recall Curve (PR AUC), Balanced Accuracy (BACC), Precision and Recall.

[0183] In the regression task, we adopt Pearson Correlation Coefficient (PCC) as the main evaluation metric, and we also report other widely used classic performance indicators, including Mean Squared Error (MSE), Root Mean Square Error (RMSE), and R-Squared (R2).

[0184] result

[0185] 2.1 Joint Method

[0186] On the O'Neil benchmark dataset, we used the branch network embedding of drug combination on cell line to compare the performance of four Joint Methods through ten random five-fold cross validations. The performance of the four Joint Methods on the classification task is shown in Table 2, and the performance on the regression task is shown in Table 3. Surprisingly, the simplest method, Sup Edge, achieved the best performance, while the complex method, All Node, performed the worst, regardless of classification or regression tasks. This may be because too many "virtual edges" were added in All Node, resulting in a large gap between the same drug in different drug combinations, and the model was unable to learn the drug representation well. Based on the performance of different Joint Methods on classification and regression tasks and the number of added "virtual edges", we chose Super Edge as the Joint Method.

[0187] Table 2

[0188] method ROC AUC BACC Kappa F1 Super Edge 0.964(0.963,0.964) 0.9(0.898,0.902) 0.8(0.797,0.803) 0.865(0.863,0.867) One Super node 0.962(0.961,0.963) 0.898(0.895,0.9) 0.795(0.791,0.799) 0.862(0.859,0.865) Three Super nodes 0.962(0.96,0.963) 0.897(0.894,0.9) 0.795(0.791,0.799) 0.862(0.858,0.865) All Node 0.903(0.899,0.907) 0.813(0.809,0.816) 0.624(0.617,0.631) 0.747(0.743,0.752)

[0189] Table 3

[0190] method RMSE R2 PCC Super Edge 10.128(10.067,10.19) 0.705(0.701,0.708) 0.851(0.848,0.853) One Super node 10.304(10.244,10.363) 0.694(0.691,0.698) 0.843(0.842,0.845) Three Super nodes 10.29(10.249,10.331) 0.695(0.693,0.698) 0.844(0.843,0.845) All Node 13.699(13.652,13.745) 0.46(0.456,0.464) 0.687(0.685,0.69)

[0191] 2.2 Performance Comparison

[0192] To evaluate JointSyn's performance, we compared it with six leading or state-of-the-art methods using ten randomized five-fold cross-validation runs. We selected classic machine learning methods and recent deep learning methods for comparison: XGBoost, DeepSynergy, AuDNNsynergy, DeepDDS, DTSyn, and HypergraphSynergy. To ensure a more comprehensive and convincing evaluation, we compared these models on both regression and classification tasks using the O'Neil and NCI-ALMANAC benchmark datasets. Tables S2 and S3 show the performance of all models on the two benchmark datasets for classification and regression, respectively.

[0193] For classification tasks, Figure 4 -A shows the performance of each model on the O'Neil benchmark dataset, Figure 4Figure -C shows the performance of each model on the NCI-ALMANAC benchmark dataset. JointSyn surpassed all other models, achieving state-of-the-art performance across all evaluation metrics on both benchmark datasets. We note that DeepDDS and XGBoost also achieved significant performance, but were still inferior to JointSyn. Furthermore, JointSyn achieved Kappa improvements of 3.7% and 3.1% and F1 improvements of 2.4% and 1.9% over the best baseline, DeepDDS. These results demonstrate JointSyn's robust predictive power for drug combination synergy.

[0194] For regression tasks, Figure 4 -B shows the performance of each model on the O'Neil benchmark dataset, Figure 4 Figure -D shows the performance of each model on the NCI-ALMANAC benchmark dataset. JointSyn achieved the best performance across all evaluation metrics on both benchmark datasets. We note that the performance of the XGBoost model, which performed well in the classification task, declined. DeepDDS still achieved significant performance, but still fell short of JointSyn. This may be because regression tasks are more complex than classification tasks, and XGBoost, as a machine learning model, cannot learn such complex patterns. We also note that JointSyn achieved a 6.8% higher R² and a 4.4% higher PCC than the best baseline, DeepDDS, further widening the gap between DeepDDS and JointSyn. These results further demonstrate JointSyn's strong predictive power for drug combination synergy (for detailed results, see Tables 4 and 5 below: Table 4 reports the results for regression and classification tasks on the O'Neil dataset; Table 5 reports the results for regression and classification tasks on the NCI-ALMANAC dataset). The values are the means of ten five-fold cross-validation results, and the values in parentheses are the confidence intervals of the ten five-fold cross-validation results.

[0195] Table 4 Results of regression and classification tasks on the O'Neil dataset

[0196]

[0197] Table 5. Results of regression and classification tasks on the NCI-ALMANAC dataset

[0198]

[0199] In order to better understand the technical solution of the present application, a specific example is provided below for illustration. The details listed in the example are mainly for ease of understanding and are not intended to limit the scope of protection of the present application.

[0200] Example 1

[0201] See also Figure 5 - Figure 8. In this example, to further demonstrate the utility of JointSyn in personalized drug synergy prediction, we applied JointSyn to predict synergy scores for 996 cell lines from 7 tumor lineages on 703 drug combinations, generating an estimated map of pan-cancer synergistic drug combinations (see Figure 8A ). The predicted drug synergy matrix divides tumors into 4 clusters (C1-C4) and drug combinations into 5 clusters (D1-D5). The two drug combination clusters D1 and D2 have no synergistic effects in almost all tumors, and the other drug combination clusters show heterogeneous synergistic score patterns among different tumors. The C1 tumor cluster is mainly composed of blood cancers and lymphomas, with no synergistic drug combinations, which reflects the huge difference between blood lymphomas and solid tumors; the C3 tumor cluster is mainly skin cancer; other types of solid tumors are mixed in C2 and C4.

[0202] We used lung cancer as an example to illustrate the heterogeneity between cancers of the same lineage. We selected 108 drug combinations with synergy ratios greater than 5% in 188 lung cancer cell lines and compared them with the control group. Figure 8B A synergy score heat map was plotted in the Figure 1. The two major types of lung cancer, non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), showed significant differences in the distribution of synergy scores. Lung_C1, composed primarily of SCLC, had few synergistic drug combinations. Lung_C3, composed primarily of NSCLC, showed synergistic effects in the Lung_C3 cell line with multiple drug combinations. This is consistent with many previous studies that have shown high drug resistance and poor prognosis in SCLC, with the predominant gene mutations TP53 and RB1 being associated with no specific drug targets.

[0203] Some drug combinations have synergistic effects on specific NSCLC cell lines, so we further explored whether the differences in drug synergy are related to certain somatic mutations. EGFR is a frequently mutated driver gene in NSCLC, and its common inhibitor is erlotinib. We examined the combination of erlotinib with other drugs. "erlotinib and vorinostat", "erlotinib and dactolisib", and "erlotinib and MK-2206" had significantly higher synergy scores in EGFR-mutated cell lines than in cell lines without EGFR mutations, both in NSCLC cell lines and in all cell lines (see Figure 8C 、8D and 8E). Previous studies have demonstrated that vorinostat can enhance the therapeutic potential of erlotinib in lung cancer cells; PI3K / Akt / mTOR signaling is a major mechanism of EGFR resistance, and dactolisib is a dual PI3K / mTOR inhibitor; MK-2206 combined with erlotinib can synergistically inhibit cell proliferation in human cancer cell lines; these studies confirm the reliability of JointSyn predictions.

[0204] in, Figure 5 A schematic diagram of a joint molecular diagram of the drug combination of erlotinib and vorinostat used in this example is shown, wherein red represents vorinostat and blue represents erlotinib; Figure 6 A schematic diagram of the combined molecular diagram of the erlotinib and dactolisib drug combination used in this example is shown, wherein red represents dactolisib and blue represents erlotinib; Figure 7 A schematic diagram of the joint molecular map of the drug combination of erlotinib and MK-2206 used in this example is shown, wherein red represents MK-2206 and blue represents erlotinib.

[0205] It should be noted that those skilled in the art should understand that the implementation functions of each module shown in the embodiment of the device for training the joint learning prediction model of drug synergy can be understood with reference to the relevant description of the method for training the joint learning prediction model of drug synergy. The functions of each module shown in the embodiment of the device for training the joint learning prediction model of drug synergy can be implemented by a program (executable instruction) running on a processor. If the device for training the joint learning prediction model of drug synergy described in the embodiment of the present application is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application 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 and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk. Thus, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0206] Accordingly, the embodiments of the present application also provide a computer-readable storage medium having computer-executable instructions stored therein, which implement the various method embodiments of the present application when executed by a processor. Computer-readable storage media include permanent and non-permanent, removable and non-removable media that can implement information storage by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable storage media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0207] In addition, the embodiments of the present application also provide a device for training a joint learning prediction model of drug synergy, which includes a memory for storing computer executable instructions, and a processor; the processor is used to implement the steps in the above-mentioned method embodiments when executing the computer executable instructions in the memory. Wherein, the processor can be a central processing unit (Central Processing Unit, referred to as "CPU"), a graphic processing unit (Graphic Processing Unit, referred to as "GPU"), a digital signal processor (Digital Signal Processor, referred to as "DSP"), a microcontroller unit (Microcontroller Unit, referred to as "MCU"), a neural network processor (referred to as "NPU"), an application specific integrated circuit (Application Specific Integrated Circuit, referred to as "ASIC"), a field programmable gate array (Field Programmable Gate Array, referred to as "FPGA") or other programmable logic devices, etc. The aforementioned memory can be a read-only memory (read-only memory, referred to as "ROM"), a random access memory (random access memory, referred to as "RAM"), a flash memory (Flash), a hard disk or a solid-state drive, etc. The steps of the method disclosed in each embodiment of the present invention can be directly embodied as being executed by a hardware processor, or executed by a combination of hardware and software modules in the processor.

[0208] The serial numbers used in describing the steps of a method do not themselves constitute any limitation on the order of these steps. For example, a step with a larger serial number does not necessarily have to be executed after a step with a smaller serial number. The step with a larger serial number can be executed first and then the step with a smaller serial number, or they can be executed in parallel, as long as this execution order is reasonable to those skilled in the art. For another example, having multiple steps with consecutive serial numbers (such as step S1, step S2, step S3, etc.) does not limit other steps that can be executed in between. For example, there can be other steps between step S1 and step S2.

[0209] This specification includes combinations of the various embodiments described herein. Individual references to embodiments (e.g., "one embodiment" or "some embodiments" or "preferred embodiments") are not mutually exclusive unless indicated as such or clear to one skilled in the art. It should be noted that the word "or" is used in this specification in a non-exclusive sense unless the context clearly indicates or requires otherwise.

[0210] All documents mentioned in this specification are considered to be included in their entirety in the disclosure of this application so that they can be used as a basis for modification when necessary. In addition, it should be understood that the above description is only a preferred embodiment of this specification and is not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of this specification should be included in the scope of protection of one or more embodiments of this specification.

[0211] In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the accompanying drawings do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for training a joint learning prediction model for drug synergy, characterized in that: The following steps are involved: Step S1: Obtaining a synergy data triplet, a true synergy score, and a cell line expression profile, wherein the synergy data triplet includes drug A, drug B, and cell line C, the cell line expression profile includes the expression profile of cell line C, and the true synergy score refers to the experimentally measured synergy score of drug A and drug B on cell line C; Step S2: Preprocessing the data of the collaborative data triples to obtain drug feature data to be trained, wherein the drug feature data to be trained includes 1) a joint molecular graph of the drug combination AB of the drugs A and B, 2) a Morgan fingerprint feature of the drug A, and 3) a Morgan fingerprint feature of the drug B; Step S3: Using 1) the joint molecular graph of the drug combination AB of the drug A and drug B, 2) the Morgan fingerprint feature of the drug A, 3) the Morgan fingerprint feature of the drug B, 4) the expression spectrum of the cell line C, and 5) the true synergy score as input features of the joint learning prediction model, and using the synergy score of the drug A and the drug B on the cell line C as the output feature, the joint learning prediction model is trained.

2. The method according to claim 1, wherein In step S2, the molecular graph of drug A and the molecular graph of drug B are connected by a hyperedge method, thereby constructing a joint molecular graph of the drug combination AB of the drug A and drug B.

3. The method according to claim 2, wherein The hyperedge method refers to adding a virtual edge as a hyperedge between the most important nodes in the molecular graph of drug A and the most important nodes in the molecular graph of drug B, thereby constructing a joint molecular graph of the drug combination AB of the drug A and drug B based on the "hyperedge".

4. The method according to claim 3, wherein The most important nodes in the molecular graph of drug A and drug B are the two nodes with the highest betweenness centrality between drug A and drug B. The betweenness centrality is defined as follows: Among them, σ st is the number of shortest paths from node s to node t, and σ st (v) is the number of these shortest paths that pass through node v.

5. The method according to claim 4, wherein The molecular graph of drug A and the molecular graph of drug B are obtained by converting the SMILES formula of drug A and the SMILES formula of drug B; the molecular graph of drug A and the molecular graph of drug B respectively include multiple nodes and edges, wherein the nodes represent atoms and the edges represent chemical bonds.

6. The method according to claim 1, wherein In step S3, training the joint learning prediction model includes the following steps: Step S31: inputting the joint molecular graph of the drug combination AB and the expression profile of the cell line C into the first branch network of the joint learning prediction model, thereby obtaining embedded features of the drug combination AB on the cell line C; Step S32: Inputting the expression profile of the cell line C, the Morgan fingerprint features of drug A, and the Morgan fingerprint features of drug B into the second branch network of the joint learning prediction model, thereby obtaining a combined feature of the embedding of drug A and the embedding of drug B in the cell line C; Step S33: Input the embedded features of the drug combination AB on the cell line C obtained in step S31 and the combined features of the representation of drug A and the embedded drug B on the cell line C obtained in step S32 into the dual-view integration network of the joint learning prediction model, thereby obtaining a representation of drug synergy.

7. A method for predicting drug synergy, characterized in that: The following steps are involved: Providing a joint learning prediction model for drug synergy, wherein the joint learning prediction model is obtained by training using the method according to any one of claims 1 to 6; Providing expression profiles of a first drug to be predicted, a second drug to be predicted, and a cell line to be predicted; Preprocessing the first drug to be predicted and the second drug to be predicted to obtain 1) a joint molecular graph of the drug combination of the first drug to be predicted and the second drug to be predicted, 2) a Morgan fingerprint feature of the first drug to be predicted, and 3) a Morgan fingerprint feature of the second drug to be predicted 1) a joint molecular graph of the drug combination of the first drug to be predicted and the second drug to be predicted, 2) the Morgan fingerprint feature of the first drug to be predicted, 3) the Morgan fingerprint feature of the second drug to be predicted, and 4) the expression profile of the cell line to be predicted are input into the joint learning prediction model to obtain a predicted synergy score of the first drug to be predicted and the second drug to be predicted on the cell line to be predicted.

8. The method according to claim 7, wherein The classification task of the joint learning prediction model outputs whether the first drug to be predicted and the second drug to be predicted are synergistic or antagonistic in the cell line to be predicted; the regression task of the joint learning prediction model outputs the synergistic score value of the first drug to be predicted and the second drug to be predicted in the cell line to be predicted.

9. A device for training a joint learning prediction model for drug synergy, characterized in that: include: an acquisition module configured to acquire a synergy data triplet, a true synergy score, and a cell line expression profile, wherein the synergy data triplet includes drug A, drug B, and cell line C, the cell line expression profile includes an expression profile of cell line C, and the true synergy score refers to an experimentally measured synergy score of drug A and drug B on cell line C; a data preprocessing module configured to preprocess the data of the collaborative data triples to obtain drug feature data to be trained, wherein the drug feature data to be trained includes 1) a joint molecular graph of the drug combination AB of the drugs A and B, 2) a Morgan fingerprint feature of the drug A, and 3) a Morgan fingerprint feature of the drug B; The training module uses 1) the joint molecular graph of the drug combination AB of the drug A and drug B, 2) the Morgan fingerprint feature of the drug A, 3) the Morgan fingerprint feature of the drug B, 4) the expression profile of the cell line C, and 5) the true synergy score as input features of the joint learning prediction model, and uses the synergy score of the drug A and the drug B on the cell line C as an output feature to train the joint learning prediction model.

10. A device for training a joint learning prediction model for drug synergy, characterized in that: include: a memory for storing computer-executable instructions; as well as, A processor, coupled to the memory, configured to implement the steps of the method according to any one of claims 1 to 8 when executing the computer-executable instructions.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the steps of the method according to any one of claims 1 to 8.

Images