Application of AI-driven target screening technology in anti-obesity treatment of drug delivery system

Abstract

The invention discloses application of an AI-driven target screening technology in anti-obesity treatment of a drug delivery system, and belongs to the field of cross fusion of biological medicine and artificial intelligence. The screening method provided by the invention comprises the following steps: selecting a signal channel target and active small molecules related to lipid metabolism; processing the signal channel target and the active small molecules through the data set to obtain an input matrix; and inputting the input matrix into a deep learning model to obtain the interaction intensity between the active small molecules and the signal path target. According to the method, an artificial intelligence technology is introduced, a deep learning model is constructed to efficiently predict the affinity between a drug and a protein target, the problems that a traditional target screening method depends on experience and is low in efficiency are solved, and target expression is subjected to experimental verification through a molecular biology method, so that the target screening efficiency is improved. A verification closed loop from calculation prediction to molecular demonstration is realized, and the biological credibility of a target screening result is proved.

Description

Technical Field

[0001] The present invention belongs to the field of cross-integration of biomedicine and artificial intelligence, and specifically relates to an application of AI-driven target screening technology in anti-obesity treatment of drug delivery systems. Background Art

[0002] Obesity, a metabolic disease, continues to rise in incidence and has become a key contributor to numerous chronic diseases, including diabetes, cardiovascular disease, and non-alcoholic fatty liver disease. Although lifestyle interventions (such as diet and exercise) are effective in weight management, most patients struggle to maintain long-term adherence, making drug therapy an important adjunct.

[0003] In recent years, natural products have attracted widespread attention due to their strong bioactivity and minimal side effects. Among them, fucoxanthin (Fx) has shown promising application prospects in regulating lipid metabolism, promoting fatty acid oxidation, and anti-inflammatory effects. However, Fx suffers from problems such as poor water solubility, low bioavailability, and poor in vivo stability, which severely limit its clinical translation for oral treatment of obesity. To enhance its in vivo efficacy, a variety of delivery systems (such as liposomes, nanoparticles, and polymer carriers) have been developed. These systems have improved stability to a certain extent, but their retention, targeting, and long-term efficacy in the gastrointestinal environment are still significantly limited, and some carriers have the risk of immunogenicity. In addition, traditional drug development relies on empirical screening and high-throughput experiments, which are inefficient and time-consuming, making it difficult to systematically analyze the targeting mechanism of natural products.

[0004] With the development of artificial intelligence (AI) technology, deep learning-based molecular docking and target prediction models can efficiently identify drug-protein interactions, significantly accelerate the discovery of drug targets, and provide theoretical support for drug efficacy mechanism research. This method breaks through the inefficient model of traditional reliance on experimental screening and literature inference, and can identify potential targets based on large-scale biological data. It is particularly suitable for predicting the target of natural product drug candidates with complex structures and unknown mechanisms. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of low efficiency, poor prediction accuracy, and reliance on manual experience in the existing drug target screening process, and to provide an artificial intelligence-driven target screening and verification technology, combined with a delivery system to enhance the targeting and clinical application potential of natural products in anti-obesity treatment.

[0006] To achieve the above objectives, the present invention provides the following solutions, including the following contents:

[0007] The present invention provides a method for target screening combined with artificial intelligence, which is used for efficiently targeting and intervening in obesity-related metabolic pathways, and the method comprises:

[0008] Select signaling pathway targets and active small molecules related to lipid metabolism, and represent them in the form of FASTA amino acid sequences and SMILES strings, respectively;

[0009] The signaling pathway targets and active small molecules are processed through the data set to obtain the input matrix;

[0010] The input matrix is ​​fed into the deep learning model to obtain the interaction strength between active small molecules and signaling pathway targets.

[0011] Furthermore, the datasets are Davis and KIBA.

[0012] Furthermore, the deep learning model includes a protein representation model, a small molecule representation model and a binding strength prediction model.

[0013] Furthermore, the protein representation model consists of a convolutional neural network (CNN) and a Transformer layer.

[0014] Furthermore, the CNN consists of three one-dimensional convolutional layers and one maximum pooling layer; wherein the number of convolutional filters in the second and third one-dimensional convolutional layers is twice and three times that of the first layer, respectively.

[0015] Furthermore, the Transformer layer consists of four encoder layers, a dropout layer and a fully connected layer.

[0016] Furthermore, the hyperparameters in the Transformer layer include input dimension of 64, hidden layer size of 1024, number of attention heads of 8, and training and prediction are performed through five-fold cross validation.

[0017] In the protein representation model, CNN is used to extract the local chemical features of proteins, and the Transformer layer is used to extract the global chemical features of proteins.

[0018] Furthermore, the small molecule representation model is composed of a convolutional neural network (CNN) and a recurrent neural network (RNN) modeling consisting of four layers of LSTM.

[0019] Furthermore, the CNN consists of three one-dimensional convolutional layers and one maximum pooling layer; wherein the number of convolutional filters in the second and third one-dimensional convolutional layers is twice and three times that of the first layer, respectively.

[0020] In the small molecule representation model, CNN extracts the local structural features of small molecules, and LSTM captures long-range dependencies, thereby achieving a comprehensive characterization of molecular inputs.

[0021] Furthermore, the combined strength prediction model contains three fully connected layers, with a dropout layer between every two fully connected layers; the first two fully connected layers contain 1024 neurons each, and the last fully connected layer contains 512 neurons, with a dropout rate of 0.3.

[0022] Furthermore, ReLU is set as the activation function, mean square error (MSE) as the loss function, the learning rate is 0.002, the batch size is 256, and the maximum training rounds are 100 in the combined strength prediction model.

[0023] Furthermore, in the above method, the FASTA amino acid sequence of the signaling pathway target is represented by a position-specific scoring matrix, and an array is generated by PSI-BLAST, and the array is input into the data set to obtain an input matrix; the input matrix is ​​input into the convolutional neural network in the protein representation model.

[0024] Furthermore, in the above method, the FASTA amino acid sequence of the signaling pathway target is encoded into a 64-dimensional word embedding into the dataset through Prot2Vec to obtain an input matrix; the input matrix is ​​input into the Transformer layer in the protein representation model.

[0025] Furthermore, in the above method, each character in the SMILES string of the active small molecule is mapped to an integer to encode the SMILES string, and then an input matrix is ​​generated through the data set; after the input matrix is ​​input into the convolutional neural network of the small molecule representation model, it is processed by the convolutional neural network and the recurrent neural network to obtain the compound sequence representation.

[0026] Furthermore, in the above method, the data obtained from the protein representation model and the small molecule representation model are input into the binding strength prediction model, and the mean square error (MSE) and consistency index (CI) are used as evaluation indicators. Finally, the binding affinity of the active small molecule to the signal pathway target is evaluated by regression prediction.

[0027] Furthermore, binding affinity is expressed using a continuous affinity score; higher scores indicate stronger intermolecular binding.

[0028] Furthermore, after obtaining the interaction strength, the corresponding binding energy between the signal pathway target and the active small molecule can be obtained through docking simulation for verification.

[0029] Furthermore, during docking simulation, it is necessary to obtain the three-dimensional structure of the signal pathway target, remove water molecules and small molecule ligands, add polar hydrogen and assign atom types, generate a three-dimensional molecular model of the active small molecule through the SMILES string, and use the molecular docking tool to set the grid box of the binding site to cover the potential active area of ​​the target.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The present invention introduces artificial intelligence technology to construct a deep learning model to efficiently predict the affinity between drugs and protein targets, overcoming the problems of traditional target screening methods that rely on experience and are inefficient.

[0032] (2) The present invention uses molecular docking technology to verify the AI ​​screening results at the structural level to ensure the rationality of the binding between the candidate small molecules and the target protein and the reliability of the prediction.

[0033] (3) The present invention experimentally verifies target expression through molecular biological methods, realizing a closed-loop verification from computational prediction to molecular evidence, and proving the biological credibility of the target screening results. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 :Schematic diagram of predicting protein-small molecule binding affinity through deep learning AI.

[0035] Figure 2 : Deep learning model architecture for predicting small molecule-protein binding affinities.

[0036] Figure 3 : Deep learning model training process curve based on DAVIS dataset.

[0037] Figure 4 : Training process curve of deep learning model based on KIBA dataset.

[0038] Figure 5 : Molecular docking conformation diagram of Fx and AMPKα1.

[0039] Figure 6 Analysis of target protein expression and lipid accumulation in HepG2 cells after different treatments. Immunofluorescence staining images and Oil Red O staining results of AMPK and FAS (A), and quantitative analysis of AMPKα1 and FAS fluorescence intensity (B).

[0040] Figure 7 Analysis of the effects of LCEV@Fx on body weight and body fat accumulation in mice induced by a high-fat diet. Small animal magnetic resonance imaging (MRI) was used to visualize the body fat distribution of mice in the control, model, and treatment groups at weeks 0, 4, and 8 (A). Final body weights of mice in each group were also statistically analyzed at the end of the experiment (B).

[0041] Figure 8Histological analysis of target protein expression and lipid accumulation in mouse liver tissue. AMPKα1 expression in liver tissue was detected by immunofluorescence, and hematoxylin-eosin (HE) and Oil Red O staining were used to assess hepatocyte histology and lipid deposition (A). Quantitative analysis of AMPKα1 immunofluorescence intensity was performed (B). DETAILED DESCRIPTION

[0042] In order to better understand the technical problems, technical solutions and technical effects solved by the present invention, the present invention is further described in detail below with reference to the embodiments and drawings, but this should not be construed as limiting the present invention.

[0043] The vesicle-loaded fucoxanthin (LCEV@Fx) used in the following examples was prepared with reference to the description in patent application number 2025106158267. The specific method is as follows:

[0044] S1: Preparation of LCEVs: Lactobacillus casei LC39 was first inoculated into MRS broth and incubated aerobically at 37°C for 16 h. The culture was then passaged twice, each time in fresh MRS broth. After the second passage, the culture was incubated for 48 h until stationary phase was reached. Bacterial viability, as determined by plate counts, was maintained at 9 log CFU / g. The MRS culture was collected and centrifuged at 8000 g for 30 min at 4°C to remove cells and other debris. Following centrifugation, the supernatant was filtered sequentially through 0.45 μm and then 0.22 μm PVDF filters. The supernatant was then concentrated using a tangential flow filtration system equipped with an ultrafiltration membrane with a molecular weight cutoff of 10 kDa, reducing the volume to one-fifth of its original volume. Finally, the supernatant was ultracentrifuged at 130,000 g for 2 h at 4°C. The pellet was resuspended in sterile PBS and washed twice by ultracentrifugation at 100,000 g for 60 minutes at 4°C. The collected pellet was the Lactobacillus casei LC39-derived vesicles (LCEVs), which were snap-frozen in liquid nitrogen and stored at -80°C before testing.

[0045] S2: Preparation of vesicle delivery system (LCEV@Fx): The microfluidic chip was placed in an ultrasonic cleaner and enough water was added to ensure that the chip was completely immersed. The water temperature was 25°C. Subsequently, the vesicle solution (concentration of 1 mg / mL, solvent is 0.01M PBS buffer) and the fucoxanthin solution (concentration of 1 mg / mL, solvent is ethanol) were respectively delivered into the microfluidic chip through two syringe pumps. The flow rate of the vesicle solution was fixed at 10μL / min, and the flow rate of the fucoxanthin solution was 1.25μL / min. Once the two phases intersected, ultrasound was started immediately with an ultrasonic power of 240W, and the sample was collected from the outlet.

[0046] S3: The ethanol in the sample was removed by rotary evaporation, and the sample was then incubated at room temperature for 2 h to restore the stability of the vesicles to obtain vesicle nanoparticles (LCEV@Fx).

[0047] The Lactobacillus casei LC39 has a preservation number of CCTCC NO: M2018886 and is purchased from Hehe Biotechnology. The microfluidic chip is provided with two input ports, namely a first input port and a second input port, and an output port. The first input port and the second input port are located on the same side, and the output port is located on the other side. The first input port and the second input port are connected through a microchannel. The output port is connected to a microchannel connecting the two input ports through the microchannel. The size of the microchannel in the microfluidic chip is 100 to 250 μm. The length of the microchannel connecting the output port is 5.5 cm.

[0048] The measurement methods involved in the following embodiments are:

[0049] (I) Treatment and molecular detection of vesicle-loaded fucoxanthin (LCEV@Fx) in HepG2 cells: DMSO stock solutions of 78 mM and 354 mM palmitic acid and oleic acid were prepared, respectively. The palmitic acid and oleic acid stock solutions were then mixed in DMEM medium at a molar ratio of 1:2, with a final free fatty acid (palmitic acid and oleic acid) concentration of 1 mM. HepG2 cells were cultured at 1×10 5 HepG2 cells were seeded at a density of 100 μg / mL in a 12-well microplate and incubated at 37°C and 5% CO₂ for 24 hours. The culture medium in the wells was then replaced with 1 mL of DMEM medium containing LCEV@Fx (LCEV@Fx concentration was 64 μg / mL, equivalent to an initial 8 μg / mL free Fx, as determined by protein concentration) and cultured for an additional 24 hours. After the 24-hour incubation, the LCEV@Fx-containing medium was removed and replaced with DMEM medium containing free fatty acids, and incubated again for 24 hours. The control group received only fresh culture medium, while the model group was stimulated with FFA. The treated HepG2 cells were washed three times with PBS, fixed with 4% paraformaldehyde for 15 minutes, and permeabilized with 0.2% Triton X-100 for 10 minutes. Nonspecific sites were then blocked with 5% BSA for 1 hour at room temperature. A mixed solution of primary antibodies against AMPK and FAS was added, and the cells were incubated overnight at 4°C. The next day, fluorescently labeled secondary antibodies (FITC and Cy3) corresponding to the two primary antibodies were added and incubated at room temperature in the dark for 1 hour. Subsequently, nuclei were stained with DAPI for 5 minutes, washed with PBS, and mounted. Fluorescence signals of AMPK and FAS were sequentially acquired using a laser confocal microscope at different excitation wavelengths. Image analysis software was used to semi-quantitatively analyze the fluorescence intensity of each channel to assess the expression levels of the two proteins and their localization relationship under different treatment conditions.

[0050] (II) Treatment and target validation of LCEV@Fx in a high-fat diet-induced animal model: Male C57BL / 6J mice (6 weeks old) were purchased from Liaoning Changsheng Biotechnology Co., Ltd. (Benxi, China). Unless otherwise specified, mice had free access to food and water. The laboratory environment was maintained at 22°C, 50% humidity, and a 12-h light / dark cycle. To establish the obesity model, a high-fat diet was formulated with 14% protein, 26% carbohydrates, and 60% fat. After a one-week adaptation period, the mice were randomly divided into three groups of six mice each and received the following treatments:

[0051] (1) High-fat diet group (HFD): high-fat diet + PBS (200 μl, gavage);

[0052] (2) Normal diet group (Control): normal diet + PBS (200 μl, gavage);

[0053] (3) LCEV@Fx group (LCEV@Fx): high-fat diet + LCEV@Fx (400 μg LCEV@Fx in 200 μl PBS; quantified by protein concentration, this is equivalent to 50 μg Fx before assembly, gavage).

[0054] Mice were treated orally every other day for an 8-week experimental period. Liver tissue samples were fixed in 4% paraformaldehyde for 24 hours, dehydrated routinely, and embedded in paraffin. Serial sections were sectioned at a thickness of 5–8 μm. Sections were deparaffinized in xylene and rehydrated with graded ethanol. Antigen retrieval was performed (citrate buffer, 95°C, 15–20 minutes). After cooling to room temperature, nonspecific sites were blocked with 5% BSA for 1 hour. Primary antibodies against AMPK were added, incubated overnight at 4°C, and the following day, fluorescently labeled secondary antibodies were incubated for 1 hour. Nuclei were counterstained with DAPI, and the sections were mounted. Images were acquired and observed using a laser confocal microscope. Fluorescence signals were semi-quantitatively analyzed using image analysis software to assess changes in AMPK expression in liver tissue. Separate sections of liver tissue were cryosectioned at a thickness of 6–10 μm. Sections were dried at room temperature, pretreated with 60% isopropanol for 1 minute, then stained with pre-prepared Oil Red O solution for 10–15 minutes, and then differentiated with 75% ethanol. After washing with PBS, the cell nuclei were counterstained with hematoxylin for 1 minute and the sections were finally mounted.

[0055] (III) Analysis of mouse weight changes and fat distribution assessed by MRI: During the entire experiment, the weight of each mouse was measured daily. In order to dynamically monitor changes in fat content, MRI examinations were performed on mice 1 day before high-fat diet intervention (day 0), at the end of the 4th week, and at the end of the 8th week. Before the test, the mice were anesthetized by inhalation of isoflurane and positioned prone in a MiniMR-60 MRI scanner (model: MesoQMR23-060H, Niumag, Shanghai) equipped with a 0.5T permanent magnet. This device is used to non-invasively obtain the spatial distribution and relative content of adipose tissue, realize a visual assessment of fat accumulation in the body, and provide a quantitative basis for subsequent judgment of the effect of drug intervention.

[0056] Example 1 (Molecular Docking and In Vitro and In Vitro Verification of Artificial Intelligence Screening Target AMPKα1)

[0057] S1. Target selection and dataset construction: Based on two publicly available datasets, Davis and KIBA. The KIBA dataset integrates kinase inhibitor bioactivity data from multiple sources and converts them into a unified KIBA score. After data processing, the KIBA dataset contains 229 unique proteins, 2111 unique drug molecules, and 118,254 interaction terms. The maximum and average lengths of protein amino acid sequences are 4128 and 728, respectively, while the maximum and average lengths of compound SMILES sequences are 590 and 58, respectively. The Davis dataset is based on K d The maximum and average lengths of protein sequences were 2549 and 788, respectively, and the maximum and average lengths of compound SMILES sequences were 103 and 64, respectively.

[0058] In terms of molecular representation, AMPKα1 and Fx were input into the dataset in the form of FASTA amino acid sequence and SMILES string, respectively.

[0059] Small molecule SMILES sequences are represented by 64 unique chemical structure characters, each mapped to an integer, such as "C" corresponding to 1, "[" corresponding to 14, "@" corresponding to 18, and "H" corresponding to 25. For example, the SMILES sequence "C[C@@H" is encoded as [1, 14, 1, 18, 18, 25]. The embedding dimension of small molecules is set to 128, resulting in input matrices of dimensions 85 × 12885 × 12885 × 128 and 100 × 128100 × 128100 × 128 in the Davis and KIBA datasets, respectively.

[0060] For protein target sequences, the position-specific scoring matrix (PSSM) is used. The matrix is ​​generated by PSI-BLAST and forms an n×20n\times 20n×20 array, in which each element represents the probability of a certain amino acid at a specific position.

[0061] In addition, to capture the global chemical information of proteins, the DeepGS method was used to encode amino acid sequences into 64-dimensional word embeddings through Prot2Vec. The input dimensions of proteins in the Davis and KIBA datasets are 1200×128(64)1200\times 128(64)1200×128(64) and 1000×128(64)1000\times 128(64)1000×128(64), respectively.

[0062] In actual prediction, the structures of Fx and AMPKα1 are used as inputs respectively, and data processing is performed in the dataset to obtain the input matrix, which is then input into the deep learning model to output its binding affinity score to evaluate the potential binding ability of the small molecule with the target protein.

[0063] S2. Construction and training of deep learning models: For protein representation, the model uses a convolutional neural network (CNN) and Transformer layers. The CNN consists of three one-dimensional convolutional layers and a maximum pooling layer. The number of convolutional filters in the second and third layers is twice and three times that of the first layer, respectively. The Transformer module includes four encoder layers, a dropout layer, and a fully connected layer (FCC) to extract local and global chemical features of proteins. The Transformer hyperparameters include an input dimension of 64, a hidden layer size of 1024, and the number of attention heads of 8. Training and prediction are performed using five-fold cross-validation.

[0064] For small molecule representation, the model uses a CNN with the same architecture as proteins, followed by a recurrent neural network (RNN) model with four LSTM layers. The CNN extracts local structural features of the small molecule, while the LSTM captures long-range dependencies, achieving a comprehensive representation of the molecular input.

[0065] After combining protein and small molecule features, the data was passed through three fully connected layers, separated by a dropout layer. The first two fully connected layers each contained 1024 neurons, and the final layer contained 512 neurons. The dropout rate was 0.3, ReLU was used as the activation function, mean squared error (MSE) was used as the loss function, the learning rate was 0.002, the batch size was 256, and the maximum number of training epochs was 100.

[0066] S3. Model evaluation: Finally, the mean square error (MSE) and consistency index (CI) are used as evaluation indicators. The former is used to measure the size of the model prediction error, and the latter is used to verify the unbiasedness and consistency of the model learning process. The binding affinity between small molecules and protein targets is evaluated by regression prediction.

[0067] S4. Molecular Docking Verification: The three-dimensional structure of AMPKα1 was obtained from the PDB database and preprocessed, including removing water molecules and small molecule ligands, adding polar hydrogens, and assigning atom types. A three-dimensional molecular model of the active small molecule was generated using a SMILES string, and the molecular conformation was optimized to ensure docking accuracy. Using the AutoDock molecular docking tool, a grid box was set for the binding site, covering the potential active region of the target. Docking simulations were run, and the binding mode of Fx with the target and the corresponding binding energy (ΔG) were recorded.

[0068] Comparative Example 1 (Verification of the Binding of Other Candidate Targets to Fx)

[0069] In Comparative Example 1, AMPKα1 in the above embodiment was replaced with protein targets related to lipid metabolism, such as PKA, PKCθ, GSK-3, S6K, JNK, ROCK, MAPK and BTK (see Table 1 for details), and the same artificial intelligence prediction and molecular docking verification were used.

[0070] Table 1 AI-predicted binding affinity scores between Fx and targets and their relative ranking results

[0071]

[0072]

[0073] Table 2 Binding energy analysis between protein target and Fx based on molecular docking

[0074]

[0075] Figure 1 The flowchart of protein-small molecule binding affinity prediction based on deep learning in the present invention is shown in Figure 1. The construction and training process of the deep learning model for predicting small molecule-protein binding affinity includes several key steps, such as Figure 2 shown.

[0076] The training performance of the proposed AI model on the DAVIS and KIBA datasets is as follows: Figure 3 and Figure 4 As shown in the following two figures, it can be clearly seen that with the increase of training rounds, the average loss function value of the training set (loss curve) and the average loss function value of the validation set (val_loss curve) continue to decline, the fluctuation gradually decreases, and the model tends to be stable.

[0077] Lipid metabolism regulation involves multiple complex signaling pathways. In order to clarify the potential mechanism of action of Fx, we selected several key targets related to lipid metabolism to cover the core links of lipid synthesis, decomposition and metabolic regulation. The introduction of AI screening significantly improved the efficiency and accuracy of target selection. Through binding affinity prediction, targets with high potential were quickly screened out. The results of AI screening showed that among all candidate targets, AMPKα1 showed the best binding affinity (prediction score of -0.875), identifying it as the key regulatory target of Fx in lipid metabolism (Table 1). As a core regulatory factor of energy metabolism, the activation of AMPK is of great significance for the inhibition of lipid synthesis and the promotion of lipid decomposition. Therefore, the identification of AMPKα1 provides a clear direction for subsequent molecular verification and mechanism research. The accuracy of AI screening was further verified by molecular docking. The results showed that Fx had three potential binding modes on the α1β2γ1 complex of AMPK, and the calculated binding energies were -8.07, -6.84 and -6.05 kcal / mol, respectively. Figure 5 Among them, the binding site of –6.05kcal / mol is located in the AMPKα1 subunit, forming a bond with the VAL440 residue. The binding affinity of AMPKα1 to AMPKα1 was found to be the lowest, with approximately 100 hydrogen bonds, and the overall stability was enhanced by significant van der Waals interactions. Furthermore, a comparison of the docking results with other candidate targets revealed that AMPKα1 exhibited the lowest binding energy and highest affinity (Table 2). These interaction patterns were consistent with the binding affinity predicted by AI, further supporting the possibility that AMPKα1 may be an important target for Fx in regulating lipid metabolism.

[0078] To verify the biological relevance of AI screening and molecular docking predictions, a lipid metabolism disorder model was first constructed at the cellular level, and the activation effect of the carrier LCEV@Fx on the key energy metabolism target AMPKα1 and its regulatory effect on the lipid synthesis pathway after intervention were evaluated. AMPK is a core sensor of intracellular energy status. It is activated when energy is insufficient, and then inhibits anabolism and promotes catabolism by phosphorylating downstream targets. FAS is the rate-limiting enzyme in fatty acid synthesis and is negatively regulated by AMPK. Activation of AMPK can directly or indirectly inhibit the expression of FAS, thereby blocking the synthesis and accumulation of lipids. Therefore, AMPKα1 and FAS constitute an important pathway for the coupled regulation of energy metabolism and lipid metabolism. Immunofluorescence staining results showed that the expression of AMPKα1 in the model group constructed by FFA treatment was significantly downregulated (reduced to 0.28 times that of the control group), and the expression of FAS was significantly enhanced (increased to 3.38 times that of the control group), indicating that the energy sensing pathway was inhibited and lipid synthesis was active; after LCEV@Fx intervention, the level of AMPKα1 rebounded significantly and the expression of FAS decreased synchronously, indicating that the negative regulation of FAS was achieved through the activation of AMPK, thereby inhibiting the lipid synthesis process ( Figure 6 ). Oil red O staining further showed that LCEV@Fx significantly reduced the accumulation of lipid droplets in the model cells, confirming its intervention effect on abnormal lipid accumulation. The above results indicate that LCEV@Fx intervention significantly enhanced the expression of AMPKα1 and indirectly inhibited the activity of its downstream FAS, ultimately reducing intracellular lipid accumulation. This biological effect is highly consistent with the result predicted by the artificial intelligence model that Fx has a high binding potential with AMPKα1. By observing the significant upregulation of AMPKα1 after Fx treatment, it indirectly supports its biological rationality as a target of Fx, further illustrating that AI-assisted screening has practical guiding value in identifying potential drug targets.

[0079] To further verify whether the above mechanism is established at the whole animal level, we used a high-fat diet-induced mouse model to evaluate the intervention effect of LCEV@Fx on weight control, fat content and tissue lipid deposition in vivo to confirm its potential application in anti-obesity treatment. Adipose tissue appears as a high signal intensity area in magnetic resonance imaging and appears as a white area in the image. It can be clearly observed that at the end of the experiment, the white high signal area in the mice in the high-fat diet group (model group) increased significantly, indicating that the fat accumulation in their bodies increased significantly. In contrast, there was almost no significant white in the images of the mice in the control group. Although there was also fat accumulation in the mice in the LCEV@Fx-treated group, the white was significantly reduced ( Figure 7A). In addition, in terms of weight changes, the model group mice significantly increased their weight under high-fat diet intervention, with the average weight reaching 30.9g at the end of the experiment, significantly higher than the control group (26.33g). However, the weight gain of mice treated with LCEV@Fx was significantly suppressed, with the final weight being 27.22g, a statistically significant decrease compared to the model group (p<0.05) ( Figure 7 B). The expression of AMPKα1 and lipid deposition were further analyzed at the liver tissue level. Immunofluorescence staining results showed that the expression of AMPKα1 in the liver tissue of mice in the model group was significantly downregulated ( Figure 8 A), quantitative data showed that its fluorescence intensity was only 0.23 times that of the control group ( Figure 8 B), suggesting that a high-fat diet disrupts the normal activation of the energy-sensing pathway. In the LCEV@Fx-treated group, AMPKα1 expression levels significantly recovered to 0.49-fold that of the control group, a significant increase compared to the model group, demonstrating that this intervention strategy can also activate the AMPK signaling axis in vivo. Furthermore, HE staining revealed that the model group exhibited disrupted liver cell structure, significant vacuolar degeneration, and prominent steatosis. LCEV@Fx treatment partially alleviated these morphological abnormalities, resulting in improved hepatocyte structure. Oil Red O staining further confirmed a significant decrease in lipid accumulation in the treated group's liver tissue, with a lower area of ​​red lipid droplets than in the model group, suggesting improved lipid metabolism. Combined results on changes in body weight, body fat distribution, and liver tissue suggest that LCEV@Fx intervention can exert a significant anti-obesity effect in vivo.

[0080] Lipid metabolism regulation involves multiple target pathways, and the selection of targeted interventions is highly dependent on the precise identification of the mechanism of action. The present invention uses an artificial intelligence model to predict the binding affinity of Fx with multiple candidate proteins, and the results show that AMPKα1 is its potential high-affinity target. On this basis, through in vitro and in vivo experimental system verification, Fx intervention can significantly activate the expression of AMPKα1 and inhibit its downstream fatty acid synthesis pathway, ultimately achieving effective relief of lipid accumulation. This result confirms the biological validity of the AI ​​screening results, clarifies the key target status of AMPKα1 in the lipid-lowering effect of Fx, and further verifies the applicability and translational value of the artificial intelligence-assisted target screening strategy proposed in the present invention in the treatment of metabolic diseases.

[0081] Comparative Example 2

[0082] This comparative example uses other deep learning models for evaluation. AMPKα1 and Fx are input into the data set in the form of FASTA amino acid sequences and SMILES strings, respectively, to obtain input matrices. These matrices are then input into the deep learning model to output their binding affinity scores. The mean square error (MSE) and consistency index (CI) are used as evaluation indicators, as shown in Table 3 below:

[0083] Table 3

[0084]

[0085] Among them, SimBoost can be found in the article https: / / doi.org / 10.1186 / s13321-017-0209-z; DeepDTA can be found in the article http: / / doi.org / 10.1093 / bioinformatics / bty593; DeepCDA can be found in the article https: / / doi.org / 10.1093 / bioinformatics / btaa544; MATT-DTI can be found in the article https: / / doi.org / 10.1093 / bib / bbab117; FusionDTA can be found in the article https: / / doi.org / 10.1093 / bib / bbab506.

[0086] According to the results in Table 3, it can be found that the binding affinity results output by the model in Example 1 of the present invention are the best and most consistent with the detection results of the above-mentioned mouse experiments, indicating that the model in the present invention is suitable for screening targets for anti-obesity treatment.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for target screening combined with artificial intelligence, characterized in that: The method is used for efficiently targeting and intervening in obesity-related metabolic pathways, and the method comprises: Select signaling pathway targets and active small molecules related to lipid metabolism, and represent them in the form of FASTA amino acid sequences and SMILES strings, respectively; The signaling pathway targets and active small molecules are processed through the data set to obtain the input matrix; Input the input matrix into the deep learning model to obtain the interaction strength between the active small molecule and the signaling pathway target; The deep learning model includes a protein representation model, a small molecule representation model and a binding strength prediction model; The protein representation model consists of a convolutional neural network and a Transformer layer; the convolutional neural network consists of three one-dimensional convolutional layers and a maximum pooling layer; the Transformer layer consists of four encoder layers, a dropout layer, and a fully connected layer; The small molecule representation model is composed of a convolutional neural network and a recurrent neural network model including a four-layer LSTM; The binding strength prediction model includes three fully connected layers, and there is a dropout layer between every two fully connected layers.

2. The method according to claim 1, wherein In the protein representation model and small molecule representation model, the number of convolutional filters in the second and third one-dimensional convolutional layers of the convolutional neural network is twice and three times that of the first layer, respectively.

3. The method according to claim 1, wherein The hyperparameters in the Transformer layer include input dimension of 64, hidden layer size of 1024, number of attention heads of 8, and training and prediction are performed through five-fold cross validation.

4. The method according to claim 1, wherein In the combined strength prediction model, the first two fully connected layers each contain 1024 neurons, the last fully connected layer contains 512 neurons, and the dropout rate is 0.

3.

5. The method according to claim 1, wherein In the combined strength prediction model, ReLU is set as the activation function, mean square error is set as the loss function, the learning rate is 0.002, the batch size is 256, and the maximum training rounds are 100.

6. The method according to claim 1, wherein The FASTA amino acid sequence of the signaling pathway target is represented by a position-specific scoring matrix, and an array is generated by PSI-BLAST. The array is input into the data set to obtain an input matrix; the input matrix is ​​input into the convolutional neural network in the protein representation model.

7. The method according to claim 1, wherein The FASTA amino acid sequence of the signaling pathway target is encoded into a 64-dimensional word embedding into the dataset using Prot2Vec to obtain an input matrix; the input matrix is ​​input into the Transformer layer in the protein representation model.

8. The method according to claim 1, wherein Each character in the SMILES string of the active small molecule is mapped to an integer to encode the SMILES string, and then an input matrix is ​​generated through the data set; the input matrix is ​​input into the convolutional neural network of the small molecule representation model, and then processed by the convolutional neural network and the recurrent neural network to obtain the compound sequence representation.

9. The method according to claim 1, wherein The data obtained from the protein representation model and the small molecule representation model are input into the binding strength prediction model, and the mean square error and consistency index are used as evaluation indicators. Finally, the binding affinity of the active small molecule and the signal pathway target is evaluated by regression prediction.

10. The method according to claim 1, wherein After obtaining the interaction strength, the corresponding binding energy between the signaling pathway target and the active small molecule is obtained through docking simulation for verification; During docking simulation, it is necessary to obtain the three-dimensional structure of the signal pathway target, remove water molecules and small molecule ligands, add polar hydrogen and assign atom types, generate a three-dimensional molecular model of the active small molecule through the SMILES string, and use the molecular docking tool to set the grid frame of the binding site to cover the potential active area of ​​the target.

Images