Cooperative double-target combined prediction method for central nervous system diseases

Abstract

The invention relates to the technical field of biological medicine, in particular to a collaborative double-target combined prediction method for central nervous system diseases, which comprises the following steps: S1, data collection; s2, data preprocessing; s3, establishing a model; s4, collaborative double-target combined prediction is carried out; and S5, carrying out target combination collaboration verification. According to the method, the potential anti-Alzheimer disease synergistic double-target combination can be automatically predicted, more target combination selection spaces are provided for designing novel double-target drugs, and after the model is adjusted and optimized, the synergistic double-target combination in complex diseases except the Alzheimer disease can be predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological medicine, and particularly relates to a synergistic double-target combination prediction method for central nervous system diseases. BACKGROUND

[0002] Alzheimer's disease (AD) is the fourth largest high-incidence disease threatening the lives of the elderly after cardiovascular and cerebrovascular diseases, tumors and stroke. The clinical manifestations are mainly memory decline, cognitive and behavioral dysfunction, and various neuropsychiatric symptoms, and eventually lead to death. The pathogenesis of AD is complex and has not been fully elucidated so far. At present, the cholinergic hypothesis, oxidative stress and free radical damage hypothesis, β-amyloid (Aβ) cascade hypothesis, tau protein phosphorylation hypothesis, metal ion hypothesis, inflammation hypothesis and gene mutation hypothesis have been proposed, and a large number of researches on AD treatment drugs have been carried out accordingly, but the results are not satisfactory. In the past few decades, the research of AD treatment drugs has mostly focused on the design or search for high-selectivity therapy acting on single target. But AD is regulated by a complex network and link, and intervention of one target or link often cannot change the overall state of the disease, and when one pathway of the disease is inhibited, the body may activate another related pathway to maintain the stability of the disease. Therefore, the single-target anti-AD drugs currently used in clinic, such as acetylcholinesterase (AChE) inhibitor donepezil, can only relieve the symptoms of patients and appropriately slow down the development of the disease, and cannot completely reverse the progress of the disease, and long-term use also has various side effects. In addition, in recent years, therapies designed for the Aβ cascade hypothesis, such as reducing Aβ production, immunotherapy, and antagonizing Aβ receptor; therapies designed for the Tau protein hypothesis, such as inhibiting Tau protein aggregation, reducing the degree of Tau protein phosphorylation, and immunotherapy; and anti-inflammatory and antioxidant therapies designed for other pathological mechanisms such as inflammation and oxidative stress, etc., have all been unsatisfactory and have announced failure. Therefore, finding drugs that can selectively act on two or more targets at the same time has become the key direction of current AD treatment.

[0003] Currently, the multi-target drugs for treating AD can be divided into three forms according to the different components: (1) a single drug molecule selectively acting on multiple targets; (2) multi-drug combination, that is, a combination of multiple drugs acting on different targets; (3) multi-component complex preparation, that is, a combination of multiple active components into the same dosage form. The truly multi-target drug refers to a single component drug that can selectively act on two or more targets at the same time, which has obvious advantages compared with the multi-drug combination and multi-component drug; as a single component, it is superior to the multi-drug combination and multi-component drug in drug metabolism; it overcomes the adverse reactions caused by the interaction between components; it is convenient to use, and there is no problem of dosage or ratio in combination drug use; it has predictable pharmacokinetics and pharmacodynamics properties; under the premise of the same efficacy, due to the synergistic effect of multi-target regulation, the drug dosage can be reduced, thereby improving the adverse reactions of highly selective single-target drugs; and the occurrence of drug resistance is slowed down. Therefore, using a single compound to act on multiple targets of AD has become a new direction of current research.

[0004] Although such multi-target drugs have many advantages, there are still many challenges in their design, the first of which is the selection of target combinations, which is the premise and key of designing multi-target drugs. A reasonable target combination can produce a synergistic effect to achieve better efficacy, but there is currently no systematic study of target combinations. So far, there have been hundreds of targets reported to be closely related to AD, resulting in a huge number of target combinations. The traditional method of determining synergistic target combinations through clinical experience and phenotypic screening is time-consuming and costly, resulting in a very limited number of anti-AD target combinations entering clinical research or in development, which greatly limits the research space of multi-target anti-AD drugs. Therefore, how to explore synergistic target combinations from the currently reported AD-related targets is crucial for the discovery of multi-target anti-AD drugs. In addition, although artificial intelligence has been widely used in AD disease classification prediction, image diagnosis, biomarker discovery, drug repositioning, and anti-AD combination drug prediction research, there are few reports on target combinations. The improvement of combination drug efficacy can preliminarily verify the feasibility and safety of the target combination, so predicting the combination of drugs for treating AD can reflect the effectiveness of the target combination. However, the existing drugs only target a small part of the proteome, and many AD-related targets currently have no specific active compounds. Therefore, there are certain limitations in predicting synergistic anti-AD target combinations through drug combinations, and a large number of targets have not yet been explored as potential combination therapy targets. Therefore, there is an urgent need for new artificial intelligence algorithms to directly predict the synergism of target combinations, breaking the current bottleneck of multi-target drug development. SUMMARY

[0005] The application provides a synergistic dual-target combination prediction method for central nervous system diseases, which can automatically predict potential AD-resistant synergistic target combinations, provides more target combination selection space for designing new AD-resistant dual-target drugs in the future, and the model can predict synergistic dual-target combinations in complex diseases other than AD after adjustment and optimization.

[0006] To achieve the above-mentioned purpose, the application provides a synergistic dual-target combination prediction method for central nervous system diseases, which comprises the following steps:

[0007] S1, data collection;

[0008] S2, data preprocessing;

[0009] S3, model establishment;

[0010] S4, synergistic dual-target combination prediction;

[0011] S5, verification of synergistic target combination.

[0012] Preferably, in step S1, the data collection is used to collect original data required for dual-target combination prediction, including manually collecting reported dual-target combinations for treating related diseases from PubMed and ClinicalTrials databases; collecting AD-related targets from PandaOmincs target discovery platform; collecting sequences of all human proteins, protein-protein interaction relationships and disease-related signaling pathways from Uniprot, String, KEGG and Reactome databases, which are used as input data sets of the model. 140 AD targets closely related to aging are collected, combined in pairs and excluded from positive target combination pairs as prediction data sets of the model.

[0013] Preferably, in step S2, the data preprocessing is used to convert the original data into an input format that can be received and processed by the model training, and to optimize and preprocess the original data set of the drug target interaction prediction to obtain a standardized interaction prediction data set.

[0014] Preferably, in step S3, a systematic anti-Alzheimer's disease synergistic target combination prediction model LDGCN is constructed by combining the knowledge graph and the graph convolutional neural network, including an LDGCN module, an Attention module, a Concat module and an FC module. The LDGCN module is used to integrate the information of points far away (such as the points in the red circle) in a small number of steps, and can also avoid the problem of Oversmoothing, so as to make each protein in the knowledge graph integrate the information of other proteins associated therewith; the Attention module updates the protein representation by integrating the signal pathway information based on the attention, so that the model actively learns the importance of each signal pathway for synergistic action, which helps the model to better utilize the signal pathway information and provide interpretability; the Concat module combines the obtained protein representations in pairs to represent the target combination; and the FC module outputs the synergistic score of the target combination.

[0015] Preferably, in step S4, the synergistic double-target combination prediction is used to predict the target combination in the prediction set by using the LDGCN model, to obtain a target combination synergistic score, and the prediction score is 0-1. The closer to 1, the greater the synergistic possibility.

[0016] Preferably, in step S5, the specificity of the target double-target is used to verify the synergistic effect of the predicted target combination by evaluating the improvement of animal cognitive impairment after combined administration and double-target gene interference.

[0017] The beneficial effects of the present application are:

[0018] 1. The present application overcomes some problems faced by existing target combination prediction models such as OptiCon and VIPER, such as narrow application, focus on the tumor field, and single data volume. The high-performance LDGCN model constructed by the deep learning technology of the present application starts from multiple dimensional data, which can not only efficiently predict anti-AD synergistic target combinations, but also can be applied to other pathological complex disease researches such as Parkinson's disease and cancer, thereby providing more target combination selection space for clinical research and development of new anti-AD and other disease multi-target drugs, and improving the success rate and efficiency of double-target drug design.

[0019] 2. The present application verifies the synergistic anti-AD activity of the selected new target combination in zebrafish by using the CRISPR / Cas9 technology for single / double gene knockout and double-target specific regulator combined administration, and confirms the application potential of LDGCN in predicting synergistic target combinations. This verification method helps to discover more synergistic double-target combinations in the future, and effectively saves experimental resources. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0021] Figure 1 Flow chart for establishing LDGCN model and verifying synergistic target combination;

[0022] Figure 2 For the construction of knowledge graph: red circles represent relatively v i nodes far away;

[0023] Figure 3 For the LDGCN deep learning model framework;

[0024] Figure 4 For the LDGCN model ten-fold cross-validation test set index;

[0025] Figure 5 For the external test set and the target combination score statistics in the model prediction set, wherein: (A) the target combination prediction results of the external test set (56 pairs). (B) the target combination prediction results of the model prediction set (9732 pairs).

[0026] Figure 6 For the knockout double-target combination AChE+NLRP3 and BACE1+NLRP3 can synergistically improve the cognitive ability of AD zebrafish, wherein: (A) the representative swimming trajectory of each group of zebrafish in the cross maze. (B) the ratio of the swimming distance of each group of zebrafish in the blue area to the total movement distance in the four areas. ***p<0.001versus normal control group, ## p<0.01, ### p<0.001versus model group, & p<0.05, &&& p<0.001versus each KO group, n=20 tails of fish / group, the distance of 5 zebrafish in each group in the figure represents a data point.

[0027] Figure 7Five target combinations were selected to significantly improve cognitive abilities in AD zebrafish. (A) MTC assay for donepezil hydrochloride, MCC950, CNP520, and SRT1720, n = 30 fish / group. (B, D) Swimming trajectories of zebrafish from representative groups in the plus-maze test. (C, E) Ratio of the distance swam in the blue zone to the total distance traveled in the four zones for each group, n = 30 fish / group. The distances of five zebrafish in each group represent one data point. **p < 0.01, ***p < 0.001 versus normal control group. # p<0.05, ## p<0.01, ### p<0.001versus model group; D,& p<0.05versusDon group, D,&&& p<0.001versus Don group, M,&&& p<0.001versus MCC950 group, S,&&& p<0.001versusSRT1720 group, C,&& p<0.01versus CNP520 group, C,&&& p<0.001versus CNP520 group.

[0028] Figure 8 The five target combinations selected significantly improved the response ability of AD zebrafish. (A, B) Average speed changes of zebrafish in each group during the light / dark stimulation phase, n = 10 fish / group. ***p < 0.001 versus normal control group. ### p<0.001versus model group; D,&& p<0.01versus Don group, D,&&& p<0.001versus Don group, M,& p<0.05versusMCC950 group, M,&&& p<0.001versus MCC950 group, S,&& p<0.01versus SRT1720 group, S,&&& p<0.001versusSRT1720 group, C,& p<0.05versus CNP520 group, C,&& p<0.01versus CNP520 group. DETAILED DESCRIPTION

[0029] The technical solutions of the present application will be described clearly and completely below in combination with the drawings of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0030] A synergistic dual-target combination prediction method for central nervous system diseases, comprising the following steps:

[0031] S1, data collection, for collecting the original data required for dual-target combination prediction, including manually collecting the reported dual-target combination for treating the related diseases from PubMed and ClinicalTrials databases; collecting the AD-related targets from PandaOmincs target discovery platform; collecting the sequences of all human proteins, protein-protein interaction relationships and disease-related signaling pathways from Uniprot, String, KEGG and Reactome databases, which are used as the input data set of the model. 140 AD targets closely related to aging are collected, combined two by two and excluded as the prediction data set of the model.

[0032] S2, data preprocessing, for converting the original data into an input format that can be received and processed by the model training, and optimizing and preprocessing the original data set of the drug target interaction prediction to obtain a standardized interaction prediction data set.

[0033] To obtain sufficient positive target combination datasets, we first manually collected reported disease treatment target combinations associated with AD, PD, depression, schizophrenia, epilepsy, and tumors from PubMed and ClinicalTrials databases. The effectiveness of these combinations is more reflected by the therapeutic effect of the dual / multi-target ligands or drug combinations of the target combinations in actual experiments, rather than verified by clinical data. Although the targets in the combinations come from multiple diseases, these targets have also been found to be associated with AD pathology, and thus can also be potential anti-AD synergistic target combinations. Due to the many challenges faced in the design of multi-target drugs, more research has focused on dual-target drugs. Moreover, the positive target combinations we collected are mostly dual-target pairs (a total of 369 pairs). Therefore, the artificial intelligence model established in this study is for the prediction of dual-target synergistic combinations. We used 313 pairs of positive target combinations as the positive dataset for model input, and another 56 pairs of target combinations as the external test set. Second, we used literature databases and the PandaOmincs target discovery platform of Insilico Medicine to collect AD-related targets that are also involved in the occurrence of PD and tumors, and obtained 1314 negative target combination pairs (excluding positive target pairs) by negative sampling after two-by-two combination. Subsequently, we converted these proteins into the ID form in String and performed ten-fold cross-validation on the 1627 positive and negative target combination datasets. In addition, we collected all human protein sequences, PPI relationships, and disease-related signaling pathways from Uniprot, String, KEGG, and Reactome databases as input datasets for the model. Finally, since AD is an aging-related disease, we used the PandaOmincs platform, TTD database, and Aging Atlas website to collect 140 AD targets closely related to aging, two-by-two combined and excluded positive target combination pairs, and obtained 9732 pairs of target combinations as the prediction dataset for the model.

[0034] S3, model establishment, combined with knowledge graph and graph convolutional neural network, a systematic anti-alzheimer's disease collaborative target combination prediction model LDGCN is constructed, including LDGCN module, Attention module, Concat module and FC module. The LDGCN module is used to integrate the information of points far away (such as the points in the red circle) in a small number of steps, which can also avoid the problem of oversmoothing, so as to make each protein in the knowledge graph integrate the information of other proteins associated with it; the Attention module updates the protein representation by integrating the signal pathway information based on attention, so that the model can actively learn the importance of each signal pathway for synergistic effect, which helps the model to better utilize the signal pathway information and provide interpretability; the Concat module combines the obtained protein representation to represent the target combination; the FC module outputs the synergistic score of the target combination.

[0035] Among them, the graph convolutional network (GCN) in deep learning can iteratively aggregate the attribute information of neighbor nodes, and the representation vector of each node contains the attribute of the node itself and the features of its neighbors, so each node propagates the information of the entire graph network, not just the attribute information of the node itself. This method can learn multiple input information at the same time, and has been applied in drug combination prediction. Therefore, this study plans to use the classic GCN as the model framework for target combination prediction, and collect the target combinations of AD, Parkinson's disease, depression, schizophrenia, epilepsy and tumor, as well as the sequence information of all human proteins in Uniprot, the interaction relationship of all human proteins in String database and the signal pathway in KEGG and other databases as the input data set of the model, to construct a deep learning model to predict the currently reported AD-related target synergistic combination.

[0036] Considering that the interaction between proteins and the position of proteins in the signal pathway play a great role in protein-protein synergistic interaction, this study adopts a deep learning method based on knowledge graph to construct the prediction model. The mathematical symbols are described as follows:

[0037]

[0038] Construction of knowledge graph: G=(V, E): The shortest path between two proteins is calculated using the PPI relationship of String database, and the shortest path is taken as the edge E={e ij |(i, j)∈V×V}, the node V in the knowledge graph is protein (V={v1, v2,..., vn}, n is the number of proteins in the knowledge graph), and the edge E is the shortest path between two proteins. |v|} is denoted by v iTo represent the i-th node in G) Figure 2 As shown. The node feature consists of two parts: signal path P = {p1, p2, ..., p |V|} and ESM (Evolutionary scale modeling) characterizes M={m1,m2,...,m |v|}, where m i represents the ESM representation of the i-th node. ESM representation is a protein sequence representation method based on deep learning, developed by Meta AI (formerly Facebook AI Research). It uses evolutionary information and multiple sequence alignments of protein sequences to train deep neural networks to generate high-quality protein representations. In this study, we directly use the protein representation generated by the trained ESM model. i ={path_1, path_2, ..path_Q} represents the pathway sequence of the i-th node / protein, which includes all signal pathways in which this protein participates. Q represents the participation in Q signal pathways, and path_q represents the q-th signal pathway.

[0039] LDGCN module such as Figure 3 As shown in , its role is to transmit and update messages to protein nodes in the knowledge graph. Although this function can be directly implemented using a graph neural network (GNN), GNN has difficulties in processing long-distance node transmission, such as Figure 2 The nodes in the circle in v i Farther, v i It is difficult to integrate the information of the node. The idea of ​​LDGCN is to randomly sample multiple subgraphs G′ from the large Graph (G) in each epoch during the training process. n =(V′ n , E′ n )in n represents the nth subgraph (Formula 1). In this way, the large graph is randomly divided into multiple subgraphs, and the GCN (Formula 2) operation is performed in each subgraph. It is possible to integrate the information of distant points (such as the points in the circle) in a fewer number of steps, and it can also avoid the problem of oversmoothing. It should be noted that the message passing and updating in this module is only for ESM representation. The signal pathway characteristics are used in the next module (Attention module) to obtain the attention weight of the signal pathway. After the information passing and updating of the LDGCN module, the initial ESM representation of each protein is composed of m i became In addition, during the training process, we design new random sampling for each epoch.

[0040] G′ 1 , G′ 2 ,...,G′ n =RS(G) (1)

[0041] G L =[GCN(G′ 1 ), GCN(G′ 2 ),...,(GCN(G′ n )] (2)

[0042] The role of the Attention module is to obtain the representation of each protein after integrating the signal pathway information (Formula 3), here we use TransformerDecoder, which is the decoder layer of Transformer, Equivalent to trg, p in Transformer i Equivalent to src in Transformer.

[0043]

[0044] where p i ={path_1, path_2,...path_Q}, Represents the i-th protein / node v i ESM characterization Integrate signaling pathway information i The subsequent protein / node representation. A is the abbreviation for Attention Module. Note that the for loop here is for ease of understanding; in the actual code, it is implemented in parallel.

[0045] Concat module: Each protein representation after obtaining the final integrated signal pathway information Afterwards, the characterization of the two proteins concatenate together to obtain any protein target combination v i and v j Characterization (Formula 4)

[0046]

[0047] Concat represents the concatenate operation.

[0048] FC module: any combination of protein targets (vi and v j ) are characterized After several fully connected layers and activation functions (simplified as FC here), the output protein target combination v i and v j synergy score score ij (formula 2), score ij The range of score is 0~1, the closer to 1, the greater the possibility of synergy.

[0049]

[0050] S4, synergistic double-target combination prediction, for predicting the target combination in the prediction set using the LDGCN model to obtain the target combination synergy score, the prediction score is 0~1, the closer to 1, the greater the possibility of synergy.

[0051] Using the collected multi-dimensional information of target combination pairs, protein sequences, PPI and disease signaling pathways, an LDGCN prediction model is constructed. Considering that most of the disease targets under research do not have reliable specific active molecules, in order to avoid the situation of data imbalance, the active compound data of the target target is not referenced when modeling. After training and optimization, the LDGCN model reached an accuracy of 89%, a recall rate of 75%, a precision rate of 73%, and an F1 score of 0.74 as shown in Figure 4 In addition, on an external data set containing 56 pairs of target combinations, the LDGCN model can correctly predict 41 pairs, as shown in Figure 5 Part B, the accuracy is as high as 73%, and the recall rate is 73%. Based on this, we use the model to predict the synergy of 9732 pairs of target combinations composed of 140 AD key proteins, as shown in Figure 5 Part A. The prediction score is 0~1, the higher the score, the stronger the synergy of the target combination, and part of the prediction results are shown in the table below:

[0052]

[0053] S5, using the specific active compounds of the target double-target combination for combined administration and interfering with the target target gene, and verifying the synergy of the predicted target combination by evaluating the improvement of animal cognitive impairment after combined administration and double gene interference.

[0054] Since AChE inhibitors are the most clinically proven effective AD treatment drugs, the anti-AD target combination we collected is mostly based on AChE combined with some targets such as NMDAR, BACE1, GSK3β and HDAC6 in the key pathological mechanisms of AD. Based on the fact that NLRP3, BACE1 and SIRT1 are the key targets of the neuroinflammation hypothesis, the Aβ hypothesis and cell aging in AD pathology, we focused on the performance of the five new target combinations composed of these three targets and AChE in LDGCN.

[0055] First, when using CRISPR / Cas9 technology to intervene in the target gene to verify the synergistic effect of the target combination, the anti-AD effect of SIRT1 activation is difficult to simulate by gene overexpression. Therefore, this part of the experiment only evaluates whether the A+N and B+N combination knockout has the ability to synergistically improve the cognitive impairment of AD zebrafish by knocking out AChE (A), NLRP3 (N) and BACE1 (B) through single / double gene knockout. The synergistic effect of the other three target combinations involving SIRT1 (A+S, B+S and N+S) is further evaluated by subsequent co-administration.

[0056] CRISPR / Cas9 gene editing:

[0057] Three gRNA targets were designed and synthesized for the target genes (ache, bace1 and nlrp3), which were mixed with Cas9 protein and injected into zebrafish fertilized eggs. After randomly selecting 20 2dpf embryos, the genome was extracted, PCR was used to detect whether the target was cut, TA cloning was used to detect the cutting efficiency of the target, and the efficient target was mixed for gene knockout.

[0058] 1) Target gene design: Obtain zebrafish target genome sequence, mRNA sequence, exon information, etc. from Ensembl database. Design according to CRISPR target site prediction website (http: / / www.e-crisp.org / E-CRISP / designcrispr.html).

[0059] 2) Primer design and synthesis: Enter the genomic sequence into the primer design website primer 3.0, select appropriate primers by setting primer length, target fragment size and position, melting temperature, etc. Design identification primers with target DNA sequence as template, and detect primer specificity by BLAST. The designed primers were synthesized by General Biological (Anhui) Co., Ltd. The synthesized primer sequences are as follows:

[0060]

[0061] 3) gRNA in vitro transcription: Wild zebrafish genome was amplified with synthetic detection primers, and the target forward primers were synthesized according to the actual sequence after comparing the database downloaded sequence with the sequencing results. Then the target forward primers and T7gRNA-R primers were used to amplify and recover the transcription templates, and the E2040s transcription kit was used for in vitro transcription. The transcribed target gRNA was recovered with AM1561 kit, and the concentration was detected using the Nano-500 microspectrophotometer of ALLSHENG. First, wash with ddH2O twice, set the sample type as RNA, and use the solvent of the solution to be tested as a blank control for verification. Detect the concentration of gRNA sample, dilute to 2000 ng / μL after measuring the concentration, and store in a -80℃ refrigerator. Eight tubes of each target were amplified, and the amplified products were recovered. The PCR reaction conditions were as follows: 95℃ for 2 min; 95℃ for 15 s, 55℃ for 15 s, 72℃ for 10 s, 45 cycles; 72℃ for 5 min. The reaction system was as follows:

[0062]

[0063] 4) Zebrafish embryo microinjection

[0064] After mixing gRNA (final concentration 320 ng / μL) and Cas9 protein (final concentration 800 ng / μL), they were injected into zebrafish embryos, 1 nL per embryo, and about 100 embryos were injected for each target. Randomly selected 20 2dpf larvae in each group were extracted for genomic DNA and PCR amplification, and sent for sequencing.

[0065] 5) Target cleavage efficiency detection: The PCR recovery product of the previous step was TA cloned, 15 clones were selected for each sequencing, and the sequencing results without signal and non-single clone were removed to calculate the target cleavage efficiency.

[0066] Gene knockout verification of target combination synergy: The final concentration of Cas9 in the AChE, NLRP3 and BACE1 knockout group was 800 ng / μL, the final concentration of gRNA was 320 ng / μL, the rest was supplemented with fish water, and 1.00 nL was injected; the final concentration of Cas9 in the AChE knockout+NLRP3 knockout group and the BACE1 knockout+NLRP3 knockout group was 800 ng / μL, two gRNA stock solutions were mixed, the final concentration of single gRNA was 320 ng / μL, the rest was supplemented with fish water, and 1.00 nL was injected. After the injection solution was prepared, it was placed in a 37°C water bath for 5 min, and then microinjection was performed. Randomly select 1-2 cell period wild type AB strain zebrafish in a 6-well plate, 30 tails per group. Except for the normal control group and the model group, the rest of each experimental group was injected with Cas9 and gRNA for gene knockout. After 2 days of treatment at 28°C, each experimental group was given 150 μM AlCl3 and 40 mg / mL D-gal to establish an AD model of zebrafish (Comput Struct Biotechnol J. 2024 May 22; 23:2230-2239) except for the normal control group. During this period, the liquid was changed and dead fish were removed every day. After 72 h of continuous administration at 28°C, the behavior of each group of zebrafish was observed, and the cognitive ability changes of AD zebrafish after single / double gene knockout were detected by cross maze experiment.

[0067] Figure 6 Results: Compared with the model group, the swimming distance of AChE, NLRP3 and BACE1 knockout group zebrafish in the blue area increased significantly; after AChE / NLRP3 double knockout, the swimming distance of AD zebrafish in the blue area further increased significantly, and was more obvious than the AChE and NLRP3 single knockout groups; similarly, the BACE1 / NLRP3 double knockout group increased the swimming distance of AD zebrafish in the blue area more obviously than the NLRP3 and BACE1 single knockout groups. The above shows that the combination of A+N and B+N knockout can synergistically improve the cognitive impairment of AD zebrafish, which further verifies the accuracy of the LDGCN prediction model.

[0068] Since the anti-AD activity of the combination administration can reflect the synergy of the corresponding target combination, we further verified the synergistic anti-AD activity of five pairs of target combinations by combination administration of specific active compounds of the four targets. First, the maximum tolerance concentration (MTC) of Don (AChE specific inhibitor), MCC950 (NLRP3 specific inhibitor), CNP520 (BACE1 specific inhibitor) and SRT1720 (SIRT1 specific activator) was detected respectively, and it was found that the MTC of Don, MCC950, CNP520 and SRT1720 was 31.2, 15.6, 7.81, 7.81 μg / mL respectively, see Part A of Figure 7 Secondly, the learning and memory abilities of the four compounds at the gradient of their respective safe concentrations were detected, and the results showed that Don, MCC950, CNP520 and SRT1720 could significantly improve the learning and memory abilities of zebrafish at the concentrations of 15.6, 7.81, 7.81, 7.81 μg / mL respectively, see Parts B, C of Figure 7 On this basis, the synergies of five pairs of target combinations A+N, S+N, B+S, A+S and B+N were evaluated by combination administration of specific active compounds of the target targets, and the results showed that Don (7.81 μg / mL) + MCC950 (3.91 μg / mL), SRT1720 (3.91 μg / mL) + MCC950 (3.91 μg / mL), CNP520 (3.91 μg / mL) + SRT1720 (3.91 μg / mL), Don (7.81 μg / mL) + SRT1720 (3.91 μg / mL), CNP520 (3.91 μg / mL) + MCC950 (3.91 μg / mL) could significantly improve the cognitive ability of zebrafish, and the efficacy was significantly enhanced than that of each single drug, see Parts D, E of Figure 7

[0069] The light-dark box test results showed that Don, MCC950, CNP520 and SRT1720 could significantly improve the response ability of zebrafish at the concentrations of 31.2, 7.81, 7.81, 7.81 μg / mL respectively, see Part A of Figure 8 The selected five pairs of target combinations could significantly improve the response ability of zebrafish, and the efficacy was significantly enhanced than that of each single drug, which was consistent with the results observed in the cross maze, see Part B of Figure 8

[0070] ​​In summary, the combination of synergistic targets is crucial for the development of multi-target drugs. In this study, we innovatively used deep neural networks to construct the LDGCN model to predict the combination of synergistic targets against AD. After training and optimization, the model achieved an accuracy of 73% and a recall rate of 73% on the external validation set, demonstrating good predictive performance. Based on the previously established zebrafish AD model, we used gene knockout and combined drug administration to find that the selected five pairs of target combinations (A+N, S+N, B+S, A+S, and B+N) could synergistically improve the cognitive and response ability disorders of AD zebrafish, preliminarily confirming the predictive accuracy of the LDGCN model. With the support of more experimental data in the future, the model prediction accuracy can be further optimized and improved. In addition, although this study predicts and verifies the synergistic target combination against AD, it can be transferred to other complex disease systems for further exploration by adjusting the prediction set and model parameters, providing more theoretical basis and selection space for the design of dual-target therapeutic drugs.

[0071] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A synergistic dual-target combination prediction method for central nervous system diseases, characterized in that The following steps are involved: S1, data collection; S2, data preprocessing; S3, model building; S4, prediction of synergistic dual-target combinations; S5. Verification of target combination synergy.

2. The synergistic dual-target combination prediction method for central nervous system diseases according to claim 1, characterized in that: In step S1, data collection is used to acquire the raw data required for dual-target combination prediction. This includes manually collecting reported dual-target combinations for related disease treatments from the PubMed and ClinicalTrials databases; using the PandaOmincs target discovery platform to collect AD-related targets; and collecting sequences of all human proteins, protein-protein interactions, and disease-related signaling pathways from databases such as Uniprot, String, KEGG, and Reactome. These serve as the input datasets for the model. 140 AD targets closely associated with aging were collected, paired together, and positive target combination pairs were excluded as the prediction dataset for the model.

3. The synergistic dual-target combination prediction method for central nervous system diseases according to claim 1, characterized in that: In step S2, the data preprocessing is used to convert the original data into an input format that can be received and processed by model training, and to optimize and preprocess the original data set for drug-target interaction prediction to obtain a standardized interaction prediction data set.

4. The synergistic dual-target combination prediction method for central nervous system diseases according to claim 1, characterized in that: In step S3, a systematic anti-Alzheimer's disease synergistic target combination prediction model LDGCN was constructed by combining knowledge graph and graph convolutional neural network, including LDGCN module, Attenti on module, Concat module and FC module; The LDGCN module is used to integrate information from distant points within a smaller number of steps, thus avoiding the oversmoothing problem and enabling each protein in the knowledge graph to integrate information from other related proteins. In the Attention module, the updated protein representation integrates signaling pathway information in an attention-based manner, enabling the model to actively learn the importance of each signaling pathway for synergy, helping the model to better utilize signaling pathway information and provide interpretability. The Concat module combines the obtained protein representations in pairs to characterize the target combination; The FC module outputs a target combination synergy score.

5. The synergistic dual-target combination prediction method for central nervous system diseases according to claim 4, characterized in that: In step S4, the synergistic dual-target combination prediction is used to predict the target combination in the prediction set using the LDGCN model to obtain a target combination synergy score. The prediction score is 0 to 1, and the closer to 1, the greater the possibility of synergy.

6. The synergistic dual-target combination prediction method for central nervous system diseases according to claim 1, characterized in that: In step S5, the synergy of the predicted target combination is verified by combining the administration of specific active compounds of the dual targets and interfering with the target genes, and evaluating the improvement of the cognitive impairment of animals after combined administration and dual gene interference.

Images