A disease treatment target discovery and drug prediction method based on multi-omics network and deep learning model

Abstract

This invention relates to a method for disease therapeutic target discovery and drug prediction based on multi-omics networks and deep learning models, belonging to the interdisciplinary field of bioinformatics and artificial intelligence drug discovery. The method includes: integrating genomic expression profiles and common molecular interaction data from disease and control groups to construct a candidate whole-genome network; refining the network based on expression profile data through systematic modeling and the AIC criterion to obtain the real molecular interaction network; extracting the core network using the master network projection method and identifying key targets through pathway analysis; predicting candidate drugs interacting with the targets using a pre-trained deep neural network model; and finally screening potential therapeutic drugs based on multi-dimensional criteria such as regulatory ability, sensitivity, and toxicity. This invention achieves a complete integration from disease mechanism analysis to drug prediction, and is particularly suitable for complex diseases such as atopic dermatitis. It can systematically discover precise targets and efficiently predict repositionable drugs, significantly improving R&D efficiency.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics, computational systems biology, and artificial intelligence drug discovery. Specifically, it relates to a computational method that integrates whole-genome multi-omics data with deep learning models to systematically discover disease therapeutic targets and predict potential therapeutic drugs. Background Technology

[0002] Atopic dermatitis (AD) is a common, chronic, relapsing, inflammatory skin disease with a global prevalence of approximately 2.4%, making it the third most prevalent skin disease. AD typically manifests in infancy and early childhood, clinically characterized by intense itching, dry skin, eczematous lesions, and marked inflammation, accompanied by impaired epidermal barrier function. In addition to physical discomfort, AD patients often experience significant psychological burden due to the appearance of their skin lesions and the unbearable itching; the incidence of depression and anxiety is significantly higher than in the general population.

[0003] The pathogenesis of Alzheimer's disease (AD) is complex, involving the combined effects of genetic susceptibility, immune system dysregulation (especially Th2 immune response dominance), skin barrier defects, and environmental factors. Currently, AD treatment primarily employs a stepwise approach. For mild to moderate cases, topical therapy is the first-line option, including topical corticosteroids (TCS) and topical calcineurin inhibitors (TCIs). However, long-term use of TCS may cause local side effects such as skin atrophy and telangiectasia; while TCIs do not carry the risk of skin atrophy, their long-term safety (such as the potential risk of lymphoma) still requires further data evaluation. For moderate to severe cases, phototherapy and systemic therapies (such as cyclosporine, dupilumab, and baricitinib) are used. Phototherapy carries potential carcinogenic risks, while systemic immunosuppressants or biologics may cause systemic side effects such as infections and hematological abnormalities, and some drugs are expensive. These existing therapies all have limitations: they are mostly "broad-spectrum" immunomodulatory, lacking precision targeting specific key nodes in the complex pathogenesis of AD, and their efficacy is limited for some refractory patients.

[0004] New drug development faces severe challenges, including high costs, lengthy cycles, and high failure rates. Against this backdrop, drug repositioning has become a highly attractive strategy. This strategy aims to discover new therapeutic uses for existing drugs, significantly shortening development time, reducing development costs, and lowering clinical risks. For example, the successful repositioning of the antiviral drug remdesivir during the COVID-19 pandemic is a well-known case.

[0005] In recent years, the rapid development of systems biology and artificial intelligence has provided new paradigms for understanding complex diseases and accelerating drug discovery. Systems biology, by integrating multi-omics data and constructing biological network models (such as protein-protein interaction networks and gene regulatory networks), can analyze disease mechanisms and identify key regulatory nodes (biomarkers) at a holistic level. Deep learning, especially deep neural networks (DNNs), has shown powerful advantages in processing high-dimensional, nonlinear biomedical big data (such as drug-target interaction prediction).

[0006] However, current research often applies network pharmacology and AI prediction models in relatively isolated ways, lacking a computational framework that seamlessly integrates the entire process from disease mechanism discovery to precise drug prediction. Specifically, how to accurately reconstruct disease-specific whole-genome genetic and epigenetic networks using real patient data (such as whole-genome microarrays), how to systematically and quantitatively screen out the core biomarkers driving the disease, and then use high-performance AI models to predict and screen drugs targeting these specific targets remain key problems to be solved.

[0007] Therefore, there is an urgent need to develop a method that integrates systems biology and deep learning to systematically reveal the pathogenic network of Alzheimer's disease (AD), accurately locate therapeutic targets, and efficiently predict potential therapeutic agents that can be relocated, thereby providing AD patients with more safe and effective treatment options. This invention is proposed against this backdrop. Summary of the Invention

[0008] This invention designs a method for disease therapeutic target discovery and drug prediction based on multi-omics networks and deep learning models. The technical problems it addresses are: in existing technologies, disease mechanism analysis is often limited to isolated pathways, making it difficult to systematically reveal network perturbations at the whole genome level; the false positive rate in therapeutic target screening is high, and there is a lack of methods for accurately identifying core driving molecules; the drug discovery process is disconnected from mechanism research, especially lacking an efficient computational framework for drug relocation and integration targeting specific targets. This leads to problems such as broad-spectrum treatments, significant side effects, or insufficient efficacy for complex diseases like atopic dermatitis.

[0009] A method for disease therapeutic target discovery and drug prediction based on multi-omics networks and deep learning models includes the following steps: S1. Data Integration and Candidate Network Construction: Integrate genomic expression profile data, common protein-protein interaction data, and gene regulation data of the target disease and healthy controls to construct candidate whole-genome genetic and epigenetic networks; S2. Real network identification: Based on the expression profile data, the candidate network is refined through system modeling and parameter estimation, and false positive connections are eliminated according to the model selection criteria to obtain the real molecular interaction networks for disease-specific and health-specific networks respectively. S3. Core Network Extraction: Perform dimensionality reduction analysis on the real network obtained in step S2 to extract key nodes and connections that characterize the core control structure of the network, thus forming the core network. S4. Therapeutic target identification: Based on the core network, pathway and function analysis is performed. By comparing disease and health status, the core signaling pathways and hub molecules that drive the disease are identified and determined as candidate therapeutic targets. S5. Drug Prediction and Screening: The candidate therapeutic targets are input into a pre-trained deep neural network drug-target interaction model to predict potential interacting drugs, and the drugs are screened according to multi-dimensional efficacy and safety standards, and finally recommended therapeutic drugs are output.

[0010] Preferably, in step S2, the system modeling and parameter estimation specifically involves: constructing system dynamic equations for the candidate network and using constrained least squares method for parameter estimation; the model selection criterion is the Akaike Information Criterion, used to determine the optimal number of interacting partners for each node to prune the network.

[0011] Preferably, in step S3, the dimensionality reduction analysis is specifically a principal network projection method, including: constructing the system matrix of the real network and performing singular value decomposition, selecting the top k principal component subspaces with a cumulative energy contribution of 85% as the core space, calculating the projection value of the nodes in this space, and selecting the top 6000 nodes with the highest projection values ​​to form the core network.

[0012] Preferably, in step S5, the input of the deep neural network drug-target interaction model is the feature vector of the drug and the target. Its architecture includes an input layer, multiple hidden layers with ReLU activation function and Dropout regularization, and an output layer using Sigmoid activation function; the model is trained using the binary cross-entropy loss function.

[0013] Preferably, the architecture of the model is: a 1000-dimensional input layer, a 512-dimensional first hidden layer, a 256-dimensional second hidden layer, a 128-dimensional third hidden layer, a 64-dimensional fourth hidden layer, and a 1-dimensional output layer.

[0014] Preferably, in step S5, the multi-dimensional efficacy and safety specifications include at least: the ability to regulate the target, drug sensitivity in cells or models, and predicted toxicity indicators.

[0015] Preferably, the target disease is atopic dermatitis.

[0016] Preferably, the candidate therapeutic targets identified in step S4 include one or more of interleukin-1β, GATA3, Akt, and nuclear factor-κB.

[0017] Preferably, the recommended treatment drugs finally screened out in step S5 include one or more of metformin, allantoin, and U-0126.

[0018] A computing system for performing the above method.

[0019] Compared with existing technologies, this method for disease treatment target discovery and drug prediction based on multi-omics networks and deep learning models has the following advantages: (1) This invention integrates multi-omics data to construct whole-genome genetic and epigenetic networks, which can systematically analyze disease mechanisms from a holistic perspective, overcome the limitations of traditional methods that isolate and analyze single pathways, and provide a panoramic view for understanding the networked pathogenesis of complex diseases.

[0020] (2) The present invention adopts a model screening strategy that combines master network projection and AIC criterion, which can effectively eliminate false positives from massive molecular interactions and accurately identify key pivot molecules that drive the occurrence and development of diseases as therapeutic targets, thus significantly improving the reliability of target discovery.

[0021] (3) The deep neural network drug-target interaction model constructed in this invention has an average AUC value of 0.981 after five-fold cross-validation, which shows excellent prediction performance and can efficiently and accurately screen out candidate drugs that interact with specific targets from the existing drug library.

[0022] (4) This invention realizes the full-process automated calculation from pathogenic mechanism analysis to potential drug prediction. In particular, through the drug repositioning strategy, it can quickly discover new therapeutic uses of existing drugs, greatly shorten the drug development cycle, and reduce development costs and clinical risks.

[0023] (5) Taking atopic dermatitis as an example, this invention successfully identified key targets such as IL-1β and Akt, and screened out existing drugs with therapeutic potential such as metformin and allantoin, providing new and more precise treatment candidates for complex diseases, and has a clear prospect for translational medicine. Attached Figure Description

[0024] Figure 1 This invention provides a flowchart for the study of the pathogenesis of atopic dermatitis (AD) and the identification of potential molecular drugs.

[0025] Figure 2 The present invention provides the actual GUGEN images of AD patients and healthy control groups.

[0026] (a) represents the true gugenucleus of AD patients. (b) represents the true gugenucleus of healthy controls. The red line represents the regulatory relationships in GRN, and the yellow line represents the interactions in PPI; the numbers represent the number of different molecules.

[0027] Figure 3 The core GUGEN diagram of AD patients and healthy control groups in this invention.

[0028] (a) Represents the core GUGEN of AD patients; (b) Represents the core GUGEN of healthy controls. The red line represents the regulatory relationships in GRN, and the yellow line represents the interactions in PPI; the numbers represent the number of different molecules.

[0029] Figure 4 : Comparison of core signaling pathways between AD and healthy control groups in this invention.

[0030] Figure 5 The training and validation accuracy trend chart of the DNN-DTI model of this invention (five-fold cross-validation).

[0031] Figure 6 The training and validation loss trend diagram of the DNN-DTI model in this invention (five-fold cross-validation).

[0032] Figure 7 ROC curve of the prediction performance of the DNN-DTI model in this invention.

[0033] Figure 8 The flowchart of DNN model training and drug prediction based on the DTI database in this invention. Detailed Implementation

[0034] The following is combined Figures 1 to 8 The technical solution of the present invention will be described in detail and completely so that those skilled in the art can understand and implement the present invention.

[0035] 1. Overall process overview; like Figure 1 As shown, the implementation of this invention follows a systematic process from data to discovery. The entire process can be divided into five logically interconnected stages: data integration and candidate network construction, real network identification, core network extraction, therapeutic target identification, and drug prediction and screening. This process uses atopic dermatitis (AD) as an example, but its methodological framework is universal.

[0036] 2. Data preparation and construction of candidate whole-genome genetic and epigenetic networks (GWGEN); This stage corresponds to Figure 1 The initial part of the process, namely step S1: Data Integration and Candidate Network Construction. First, two types of data are collected and integrated: Public database knowledge: Protein-protein interaction (PPI) data: Downloaded from authoritative databases such as BioGRID, DIP, BIND, IntAct, and MINT to construct a candidate PPI network (PPIN).

[0037] Gene regulatory relationship (GRN) data: Regulatory relationships between transcription factors (TFs), microRNAs (miRNAs), long non-coding RNAs (lncRNAs) and target genes were obtained from databases such as StarBase, ITFP, TRANSFAC, TargetScanHuman, and HTRIdb to construct candidate GRNs.

[0038] Experimental sample data: A specific dataset (e.g., GSE193309) for the target disease (atopic dermatitis, AD in this example) was obtained from the Gene Expression Organization (GEO) database of the National Center for Biotechnology Information (NCBI). This dataset contains whole-genome microarray expression profiles of skin lesions from AD patients (disease group) and normal skin tissues from healthy individuals (healthy control group).

[0039] Subsequently, candidate PPINs and candidate GRNs were integrated to form an initial candidate genome-wide genetic and epigenetic network (candidate GUGEN) containing a large number of potential interactions. This network is represented in matrix form, with matrix elements being Boolean values ​​(1 or 0), which preliminarily indicate the possible interactions or regulatory relationships between molecules.

[0040] 3. System identification and acquisition of "real GUGEN"; Because the candidate GUGEN integrates data from different sources, it contains a large number of false positive connections. This step refines it using real sample expression data. Figure 1 The "System Identification and System Order Detection" module.

[0041] (1) Mathematical modeling (system modeling): Establish system dynamic equations for proteins (Equation 1), genes (Equation 2), miRNAs (Equation 3), and lncRNAs (Equation 4) in GUGEN to describe the quantitative relationship between their expression levels and the regulation by other molecules.

[0042] Equation 1: Protein-Protein Interaction Network (PPIN) equation, describing how the expression level of the t-th protein is affected by the interactions of other proteins in the network.

[0043]

[0044] and : The expression levels of the t-th and k-th proteins in the n-th sample.

[0045] : The number of proteins that interact with the t-th protein.

[0046] : The interaction coefficient between the t-th and k-th proteins (the interaction is represented by a product term and may reflect complex formation).

[0047] : The baseline level of the t-th protein (used to capture undescribed mechanisms such as phosphorylation or ubiquitination).

[0048] Noise item.

[0049] Applicable to n=1 to N (number of samples), t=1 to T (total number of proteins).

[0050] Note: The equation is implicit (it exists on both the left and right sides). The parameter estimation is solved by reorganizing the solution using linear regression.

[0051] Equation 2: Gene Regulatory Network (GRN) Equation, describing the first... e How gene expression is regulated by transcription factors (TFs), long non-coding RNAs (lncRNAs), and microRNAs (miRNAs).

[0052] Gene expression levels are regulated by transcription factors (TFs), lncRNAs, and miRNAs, as shown in the model:

[0053] • , , , : The expression levels of the e-th gene, the f-th TF, the y-th lncRNA, and the u-th miRNA in the n-th sample.

[0054] • , , : Regulates the number of TF, lncRNA and miRNA in the e-th gene.

[0055] • , , : Corresponding regulatory capacity (miRNA regulation is usually negative, hence the negative sign, and ).

[0056] • : Baseline level of the e-th gene (capturing mechanisms such as methylation).

[0057] • Noise item.

[0058] • Applicable to n=1 to N, e=1 to E (total number of genes).

[0059] • The miRNA term represents negative regulation, through multiplication. express.

[0060] Equation 3: MicroRNA (miRNA) regulatory equation, describing the first... r How the expression of a miRNA is regulated by transcription factors, lncRNAs, and other miRNAs.

[0061] The expression level of miRNA is regulated by TF, lncRNA, and other miRNAs, and the model is as follows:

[0062] • , , , : The expression levels of the r-th miRNA, the f-th TF, the y-th lncRNA, and the u-th miRNA in the n-th sample.

[0063] • , , : Regulates the number of TF, lncRNA and miRNA of the r-th miRNA.

[0064] • , , : Corresponding regulatory capacity (the regulation between miRNAs is usually negative, hence the negative sign, and ).

[0065] • : Baseline level of the r-th miRNA (capturing mechanisms such as acetylation).

[0066] • Noise item.

[0067] • Applicable to n=1 to N, r=1 to R (total number of miRNAs).

[0068] Equation 4: Regulation equation of long non-coding RNA (lncRNA), describing the first... c How the expression of a lncRNA is regulated by transcription factors, other lncRNAs, and miRNAs.

[0069] The expression level of lncRNAs is regulated by TF, other lncRNAs, and miRNAs, and the model is as follows:

[0070] • , , , : Expression levels of the c-th lncRNA, f-th TF, y-th lncRNA, and u-th miRNA in the n-th sample.

[0071] • , , : Regulates the number of TF, lncRNA, and miRNA in the c-th lncRNA.

[0072] • , , : Corresponding regulatory capacity (miRNA regulation is usually negative, hence the negative sign, and ).

[0073] • : Baseline level of the c-th lncRNA (capturing mechanisms such as acetylation, phosphorylation, or methylation).

[0074] • Noise item.

[0075] • Applicable to n=1 to N, c=1 to C (total number of lncRNAs).

[0076] These equations are based on candidate networks mined from large databases, and real interactions / regulatory relationships are identified from microarray data using least-squares parameter estimation methods (combined with the AIC criterion to prune false positives). Negative regulatory terms (such as miRNAs) ensure biological validity, and the equations are restructured into linear regression forms to solve for the parameters.

[0077] (2) Parameter estimation: Substitute the microarray data of the AD group and the healthy control group into the above equation system, and use the constrained least squares method (the specific solution can be obtained with the help of optimization toolboxes such as lsqlin in MATLAB) to estimate the regulatory parameters (such as interaction strength ρ, regulatory ability γ, etc.) in the equation. The estimation process introduces biological constraints (such as the miRNA inhibitory effect parameter should be positive).

[0078] (3) Network pruning (false positive elimination): System order detection is performed based on the model selection criterion—Akaike Information Criterion (AIC) (Equations 17-20). AIC strikes a balance between model fit and complexity. By minimizing the AIC value (Equations 21-24), the actual number of interacting partners for each molecule is automatically determined. Accordingly, the weakest connections exceeding this number are removed from the candidate GUGEN, resulting in a more concise and realistic GUGEN that is closer to the biological reality.

[0079] Formulas 17-20: Akaike Information Criterion (AIC) equations, which are used to evaluate and screen the optimal complexity (system order) of biomolecular regulatory models.

[0080] Formula 17 (AIC of proteins):

[0081] Formula 18 (AIC of a gene):

[0082] Formula 19 (AIC of miRNA) and Formula 20 (AIC of lncRNA) have similar structures and are used to evaluate the AIC of miRNA, respectively. r The regulatory model complexity for the miRNA and the cc lncRNA is defined similarly to that described above.

[0083] Equations 21-24: Equations for determining the optimal system order. These equations determine the actual number of regulatory partners at each node in a biomolecular network by minimizing the AIC value.

[0084] Formula 21 (Optimal number of protein partners):

[0085] Formula 22 (Number of Optimal Gene Regulators):

[0086] Formula 23 (the optimal number of miRNA regulators) and Formula 24 (the optimal number of lncRNA regulators) have the same structure, and respectively determine the actual number of regulatory partners for miRNA and lncRNA.

[0087] Network visualization: Figure 2 The resulting visualizations of “real GUGEN” in AD patients and healthy controls are presented. Figure 2 The nodes represent various biomolecules, the red line represents the regulatory relationships in GRN, and the yellow line represents the interactions in PPI. Statistical comparisons in Tables 1 and 2 show that after AIC pruning, the number of network nodes and edges is significantly reduced, effectively eliminating false positives.

[0088] Table 1: Statistics on the Number of Candidate GUGEN and Actual GUGEN Nodes

[0089] Table 2: Number of edges in candidate GUGGEN and actual GUGGEN

[0090] 4. Core network extraction and pathogenesis mechanism analysis; This step corresponds to steps S3 and S4, namely core network extraction and therapeutic target identification. Because the "real GUGEN" is too large (over 16,000 nodes), direct pathway-level functional analysis is not possible, and pathway annotation tools such as KEGG typically have processing limits (e.g., 6000 molecules). This step aims to extract the most core parts of the network, corresponding to... Figure 1 The modules for "Principal Network Projection (PNP)" and "KEGG Path Annotation" are included.

[0091] (1) Implementation of the PNP method (core network extraction): Matrix construction: Arrange all estimated parameters in the “real GUGEN” by category to construct a system matrix Y (Equation 25) that characterizes the interaction strength of the entire system.

[0092] Formula 25: The equation for constructing the true GUGEN system matrix. This formula describes how to integrate all estimated biomolecular interaction parameters into a single system matrix.

[0093]

[0094] Singular Value Decomposition: Perform singular value decomposition (SVD) on matrix Y to decompose it into multiple orthogonal subspaces spanned by singular vectors, and sort them according to the size of the singular values. The size of the singular values ​​represents the "energy" contribution of the subspace to the network structure.

[0095] Core Subspace Selection: According to Formula 27, the subspace corresponding to the top k singular values ​​whose cumulative energy accounts for 85% of the total energy is selected as the principal component space representing the core structure of the network.

[0096] Formula 27 (Energy Ratio Calculation):

[0097] Core node selection involves calculating the projection value proj_i (Equation 28) of each node in the "real GUGEN" (corresponding to each row of matrix Y) onto the aforementioned core subspace. The larger the projection value, the more important the node is in the core structure of the network.

[0098] Formula 28 (Calculation of nodal projection values):

[0099] The 6000 nodes with the highest projection value proj_i and the connections between them are selected to form the "core GUGEN".

[0100] Figure 3A visualization of the "core GUGEN" was presented for the AD group and the healthy control group, with the network size compressed to 6,000 nodes, focusing on key molecules.

[0101] (2) Pathway annotation and mechanism analysis (therapeutic target identification): The molecular list in the "core GUGEN" was submitted to the KEGG pathway database for annotation to obtain the core signaling pathway sets for the AD group and the healthy control group, respectively.

[0102] like Figure 4 As shown, by comparing the core signal pathways of the two groups, aberrant pathways specific to AD were identified. Figure 4 (Red area on the left). These pathways are mainly involved in immune dysregulation and chronic inflammation driven by microenvironmental factors such as IL-4, IL-17A, and IL-1β. Meanwhile, the LEP / Akt / NF-κB axis and TNF-α pathway in common pathways (yellow area) are abnormally activated in AD, leading to increased inflammation and abnormal epidermal cell proliferation. Pathways specific to the healthy group (green area), such as the IGF-1 pathway, mainly maintain normal epithelial growth.

[0103] The target identification process is as follows: Based on the above-mentioned abnormal signaling pathways and the downstream cell dysfunctions they cause (immune disorders, inflammation, excessive proliferation), key molecules that play a pivotal role in them are screened out: IL-1β (pro-inflammatory factor), GATA3 (Th2 differentiation key transcription factor), Akt (signal transduction central protein kinase) and NF-κB (core inflammatory transcription factor), and they are identified as candidate biomarkers (drug targets) for AD treatment.

[0104] 5. Training and Application of Drug Prediction Models Based on Deep Neural Networks

[0105] This stage corresponds to the drug prediction step in step S5, which aims to identify potential therapeutic drugs for the aforementioned targets. Figure 1 The "DTI Database Training" and "DNN-based DTI Model" modules are detailed in the following sections: Figure 8 .

[0106] (1) Data preparation and preprocessing: Feature extraction is as follows: Known interaction data are collected from drug-target interaction (DTI) databases such as DrugBank, ChEMBL, and STITCH. Using the PyBioMed toolkit, the structure and sequence information of the drug (small molecule) and target (protein) are converted into numerical feature vectors (Equation 29). The drug and target feature vectors are concatenated to form the model input.

[0107] Formula 29: Drug-target feature vector construction equation. This formula defines the combined features input to the deep learning model.

[0108]

[0109] Data balancing and partitioning are as follows: Negative samples with "no interaction" are randomly downsampled to balance the number of positive and negative samples (e.g., 80,291 each). The total data is then randomly divided into training and test sets in a 75%:25% ratio.

[0110] Standardization and dimensionality reduction are performed as follows: Features are standardized (Z-score), and principal component analysis (PCA) is used to reduce the feature dimension from 1359 to 1000 to match the model input layer.

[0111] (2) Construction and training of DNN-DTI model: like Figure 8 As shown in the middle, a deep neural network is constructed with the following architecture: Input layer (1000 nodes) → Hidden layer 1 (512 nodes, ReLU activation) → Hidden layer 2 (256 nodes, ReLU activation) → Hidden layer 3 (128 nodes, ReLU activation) → Hidden layer 4 (64 nodes, ReLU activation) → Output layer (1 node, Sigmoid activation). Dropout (ratio 0.5) is applied during hidden layer training to prevent overfitting.

[0112] Training was performed using the Adam optimizer (learning rate 0.001, β1=0.9, β2=0.999) with a loss function of binary cross-entropy (Equations 32-34). Five-fold cross-validation was used during training.

[0113] Formulas 32-34: Training equations for deep learning models. These formulas define the loss function and optimization objective of the DNN-DTI model.

[0114] Formula 32 (Binary Cross-Entropy Loss):

[0115] Formula 33 (Overall Empirical Error):

[0116] Formula 34 (Optimization Objective):

[0117] The loss on the validation set is monitored. When the loss stops decreasing for several consecutive periods (e.g., 10 periods), an early stopping method is triggered to save the optimal model. In this example, the model converges around the 66th period.

[0118] The model performance is verified as follows: like Figure 5 and Figure 6 As shown, the model's accuracy steadily increased and eventually stabilized at approximately 0.922 during training and validation, while the loss value continued to decrease, indicating that the training process was stable and effective.

[0119] like Figure 7 As shown, the model achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.981 on the test set, demonstrating its strong ability to distinguish whether a drug interacts with a target, significantly outperforming random guessing.

[0120] The feature vectors of the AD biomarkers (IL-1β, GATA3, Akt, NF-κB) identified in step 4 are input into the trained DNN-DTI model. The model outputs a probability value between 0 and 1. Molecular drugs with output values ​​greater than 0.5 are considered to potentially interact with the target, forming a preliminary "candidate drug list".

[0121] 6. Potential drug screening based on multi-dimensional standards; After obtaining the list of candidate drugs, selection should be made according to reasonable drug design specifications. Figure 1 The "Drug Design Guidelines" module.

[0122] Standard definitions and data sources: (1) Regulatory capacity: Evaluate the regulatory effect of drugs on target expression (upregulation or downregulation). Data are obtained from the L1000 Level 5 gene expression perturbation database.

[0123] (2) Sensitivity: Assess the consistency of drug efficacy across different individuals or cell models. Data were obtained from the PRISM drug sensitivity screening database.

[0124] (3) Toxicity: The potential toxic side effects of the drug were assessed using the median lethal concentration (LC50) as an indicator. Data were obtained through prediction using the ADMETlab 2.0 platform.

[0125] (4) Screening process and results: Candidate drugs for each target were comprehensively evaluated and ranked according to the above three criteria. Finally, drugs that simultaneously met the requirements of effective regulation (clear inhibitory effect on the target), good sensitivity, and low toxicity were selected. As shown in Table 4, the potential molecular drugs for treating AD that ultimately stood out were: metformin, allantoin, and U-0126. Table 5 further clarifies the inhibitory relationship between these drugs and AD targets.

[0126]

[0127] Table 4: List of molecular drugs screened for different target molecules (biomarkers) in Alzheimer's disease (AD)

[0128] Table 5: Interactions between potential drugs and AD target molecules (biomarkers)

[0129] In summary, the present invention achieves this through... Figure 1 The systematic process shown comprehensively utilizes Figure 2-4 The network biology analysis results presented and Figure 5-8 The verified predictive capabilities of the deep learning model enable efficient and accurate discovery of potential therapeutic drugs from AD genomic data. This embodiment focuses on AD, but the method is equally applicable to research on other complex diseases. Any equivalent substitutions or improvements made based on the inventive concept to the data source, specific network modeling algorithm, deep learning model structure, or drug screening specifications fall within the protection scope of this invention.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for disease therapeutic target discovery and drug prediction based on multi-omics networks and deep learning models, characterized in that, Includes the following steps: S1. Data Integration and Candidate Network Construction: Integrate genomic expression profile data, common protein-protein interaction data, and gene regulation data of the target disease and healthy controls to construct candidate whole-genome genetic and epigenetic networks; S2. Real network identification: Based on the expression profile data, the candidate network is refined through system modeling and parameter estimation, and false positive connections are eliminated according to the model selection criteria to obtain the real molecular interaction networks for disease-specific and health-specific networks respectively. S3. Core Network Extraction: Perform dimensionality reduction analysis on the real network obtained in step S2 to extract key nodes and connections that characterize the core control structure of the network, thus forming the core network. S4. Therapeutic target identification: Based on the core network, pathway and function analysis is performed. By comparing disease and health status, the core signaling pathways and hub molecules that drive the disease are identified and determined as candidate therapeutic targets. S5. Drug Prediction and Screening: The candidate therapeutic targets are input into a pre-trained deep neural network drug-target interaction model to predict potential interacting drugs, and the drugs are screened according to multi-dimensional efficacy and safety standards, and finally recommended therapeutic drugs are output.

2. The method according to claim 1, characterized in that, In step S2, the system modeling and parameter estimation specifically involves: constructing system dynamic equations for the candidate network and using constrained least squares method for parameter estimation; the model selection criterion is the Akaike Information Criterion, which is used to determine the optimal number of interacting partners for each node in order to prune the network.

3. The method according to claim 1, characterized in that, In step S3, the dimensionality reduction analysis is specifically a principal network projection method, including: constructing the system matrix of the real network and performing singular value decomposition, selecting the top k principal component subspaces with a cumulative energy contribution of 85% as the core space, calculating the projection value of the nodes in the space, and selecting the top 6000 nodes with the highest projection values ​​to form the core network.

4. The method according to claim 1, characterized in that, In step S5, the input of the deep neural network drug-target interaction model is the feature vector of the drug and the target. Its architecture includes an input layer, multiple hidden layers with ReLU activation function and Dropout regularization, and an output layer using Sigmoid activation function. The model is trained using the binary cross-entropy loss function.

5. The method according to claim 4, characterized in that, The model architecture consists of a 1000-dimensional input layer, a 512-dimensional first hidden layer, a 256-dimensional second hidden layer, a 128-dimensional third hidden layer, a 64-dimensional fourth hidden layer, and a 1-dimensional output layer.

6. The method according to claim 1, characterized in that, In step S5, the multi-dimensional efficacy and safety specifications include at least: the ability to regulate the target, drug sensitivity in cells or models, and predicted toxicity indicators.

7. The method according to any one of claims 1 to 6, characterized in that, The target disease is atopic dermatitis.

8. The method according to claim 7, characterized in that, The candidate therapeutic targets identified in step S4 include one or more of interleukin-1β, GATA3, Akt, and nuclear factor-κB.

9. The method according to claim 8, characterized in that, The recommended treatment drugs in the final screening output of step S5 include one or more of metformin, allantoin, and U-0126.

10. A computing system for performing the method according to any one of claims 1 to 9.

Images