A sleep-aiding active peptide intelligent screening method and system based on a multi-modal fusion network

Abstract

The application provides a sleep-aiding active peptide intelligent screening method and system based on a multi-modal fusion network, and belongs to the technical field of bioengineering and artificial intelligence. The multi-modal fusion network model of the application is obviously better than the traditional QSAR model in the performance of accuracy, precision, recall rate, F1 score, AUC-ROC, AUC-PR and other indexes, and has good generalization ability and the accuracy of active peptide prediction, and can be widely applied to the discovery and optimization of sleep-aiding active peptides in the research and development of functional foods, health foods and peptide drugs.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of biotechnology and artificial intelligence, specifically to a method and system for intelligent screening of sleep-aiding active peptides based on multimodal fusion networks. Background Technology

[0002] Bioactive peptides are short-chain polypeptides composed of 2 to 50 amino acid residues. They have a variety of physiological regulatory functions and have received widespread attention in the research and development of functional foods, health foods and peptide drugs.

[0003] In recent years, with the continuous increase in the number of people suffering from sleep disorders, the discovery of functional peptides with sleep-aiding activities from natural proteins has become a research hotspot. Existing studies have shown that enzymatic hydrolysates derived from food-derived proteins such as milk protein, soy protein, and tortoise shell protein contain active peptides that can regulate the γ-aminobutyric acid (GABAergic) or serotonergic (5-HTergic) nervous systems. These peptides exert sedative, anti-anxiety, or sleep-promoting biological effects through interactions with receptor proteins related to sleep regulation. Currently, the screening of sleep-aiding active peptides mainly relies on traditional experimental methods. A typical screening process includes: first, enzymatically hydrolyzing the target protein to obtain a complex mixture of peptides; then, fractionating the mixture using chromatographic separation techniques; next, performing activity detection on each component at the cellular or animal level; and finally, using mass spectrometry to identify the sequences of peptides in the active components. While this traditional "enzymatic hydrolysis-separation-activity determination-identification" strategy can identify active peptides, it has significant limitations. On the one hand, the experimental screening cycle is long and costly, usually requiring months or even years to complete the entire discovery process from protein raw materials to active peptides; on the other hand, the screening throughput is limited by experimental conditions, making it difficult to comprehensively evaluate a large number of candidate peptides, and a large number of potential active peptides may be missed as a result.

[0004] To improve screening efficiency, researchers have begun to explore the introduction of computational methods into the discovery of bioactive peptides. Early computational prediction methods were primarily based on quantitative structure-activity relationship (QSAR) models, which established statistical associations between descriptors such as amino acid composition and physicochemical properties of peptides and their activity. However, QSAR methods rely on manually designed features, making it difficult to capture complex patterns in sequences and lacking universality for different types of bioactive peptides. In recent years, with the rapid development of deep learning technology, peptide activity prediction models based on recurrent neural networks or convolutional neural networks have been proposed, achieving some progress in areas such as antimicrobial peptides and anticancer peptides.

[0005] However, existing deep learning prediction models still have several shortcomings. First, most models only utilize the one-dimensional amino acid sequence information of peptides for prediction, ignoring the decisive influence of the peptide's three-dimensional spatial structure on its biological activity. The classic protein science assertion that "structure determines function" also applies to short peptides; peptides with the same sequence may exhibit drastically different receptor-binding abilities in different conformational states. Second, existing models typically do not consider the interaction information between peptides and target receptor proteins, which is precisely the molecular basis for the function of bioactive peptides. Sleep-aiding bioactive peptides need to bind to specific targets such as GABAA receptors or 5-HT receptors to produce physiological effects; predicting solely from a sequence or structural perspective makes it difficult to accurately assess their targeted regulatory capabilities. Third, the interpretability of existing models is generally insufficient, failing to reveal which amino acid sites play a key role in activity and thus unable to provide clear guidance for subsequent peptide optimization and targeted modification.

[0006] Furthermore, there is currently a lack of dedicated predictive models for this specific functional type of sleep-aiding bioactive peptides. Sleep regulation involves complex neurotransmitter systems and receptor pathways, and the mechanism of action of sleep-aiding bioactive peptides differs fundamentally from that of other types of functional peptides such as antimicrobial peptides and antioxidant peptides. Therefore, it is necessary to construct a dedicated predictive system that takes into account sleep-related target information. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent screening of sleep-aiding active peptides based on a multimodal fusion network. This invention achieves high-throughput prediction, ranking, screening, and interpretable analysis of key sites for sleep-aiding active peptides through multimodal characterization learning and fusion modeling of peptide sequence information, three-dimensional spatial structure information, and peptide-sleep-related receptor protein interaction information. It is applicable to the discovery and optimization of sleep-aiding active peptides in the development of functional foods, health foods, and peptide drugs.

[0008] To achieve the above objectives, the specific technical solution of the present invention is as follows:

[0009] A first aspect of the present invention provides a method for intelligent screening of sleep-aiding active peptides based on a multimodal fusion network, comprising:

[0010] An active peptide dataset consisting of positive and negative samples is constructed and redundancy is removed; wherein, the positive samples include peptides with sleep-aiding, sedative or anti-anxiety activities, and the negative samples include peptides known not to have sleep-aiding activities.

[0011] The active peptide dataset is divided into a training set, a validation set, and a test set according to a preset ratio.

[0012] A multimodal fusion network model is constructed, comprising an extraction layer, a fusion layer, and a prediction layer. The extraction layer extracts multidimensional features of peptides, including sequence features, three-dimensional structural features, and receptor binding features. The fusion layer fuses these multidimensional features to obtain fused features. The prediction layer takes the fused features as input and performs a prediction task.

[0013] The multimodal fusion network model is trained using a training set, and the performance of the multimodal fusion network model is monitored using a validation set to prevent overfitting. The multimodal fusion network model is validated using a test set to obtain the trained multimodal fusion network model.

[0014] Input peptide sequences with unknown activity, and use a trained multimodal fusion network model to predict and screen for peptides with sleep-aiding activity.

[0015] Furthermore, the peptides in the positive samples range in length from 3 to 50 amino acids, and each peptide has clearly defined activity data, such as the half-maximal inhibitory concentration (IC50). 50 , half effective concentration EC 50 Or results from animal experiments, etc.

[0016] Furthermore, the sources of the positive samples include: peptides labeled as sedative or anxiolytic in public databases (e.g., BIOPEP-UWM), peptides that can activate GABAA receptors reported in academic literature (e.g., PubMed), approved peptide sleep aids and their structural analogs, and active peptides identified and verified by the inventors' research group from tortoise plasmin and casein hydrolysate in previous studies.

[0017] Furthermore, the peptide lengths in the negative samples match the length distribution of the positive samples. The sources of the negative samples include: peptides randomly selected from general protein databases (e.g., UniProt), and other types of functional peptides (e.g., certain antimicrobial and antioxidant peptides) that have been experimentally proven not to have sleep-aiding functions. To avoid high sequence similarity between negative and positive samples—which would cause the model to "take shortcuts" rather than truly learn activity features—this invention uses the CD-HIT tool to perform redundancy removal on all samples (positive and negative samples) with a 40% sequence similarity threshold.

[0018] Furthermore, the active peptide dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio, ensuring that the ratio of positive to negative samples remains consistent in each subset. The training set is used for model training, the validation set is used to monitor model performance and prevent overfitting during model training, and the test set is used to finally evaluate the model's generalization ability.

[0019] Furthermore, the sequence features include amino acid composition features (AAC), dipeptide composition features (DPC), physicochemical property features, and sequence embedding vectors.

[0020] Furthermore, the amino acid composition characteristics are obtained by statistically analyzing the frequency of occurrence of 20 amino acids. Proteins in nature are composed of 20 standard amino acids. For a peptide segment of length L, the frequency of occurrence of each amino acid is calculated, resulting in a 20-dimensional amino acid composition feature vector. The formula for calculating the frequency of occurrence of the i-th amino acid is:

[0021] ;

[0022] In the formula, AAC(i) is the frequency of the i-th amino acid; N(i) is the number of times the i-th amino acid appears in the peptide; and L is the total length of the peptide. For example, in a peptide "WWRGH" containing 5 amino acids, tryptophan (W) appears twice, so the AAC value of tryptophan is 2 / 5 = 0.4.

[0023] Furthermore, the dipeptide composition characteristics are obtained by statistically analyzing the occurrence frequencies of 400 dipeptide combinations. Since knowing only the proportion of each amino acid is insufficient, the arrangement patterns between adjacent amino acids are equally important. By pairwise combining 20 amino acids, there are 20 × 20 = 400 dipeptide combinations. For a peptide segment of length L, the frequency of each dipeptide combination is statistically analyzed to obtain a 400-dimensional dipeptide composition feature vector, capturing local arrangement pattern information within the peptide segment. The formula for calculating the occurrence frequency of each dipeptide combination (i, j) is as follows:

[0024] ;

[0025] In the formula, DPC(i, j) is the frequency of occurrence of each dipeptide combination (i, j); N(i, j) is the number of times the combination of amino acid i followed by amino acid j appears in the peptide segment; L - 1 corresponds to the total number of adjacent amino acid pairs in a peptide segment of length L.

[0026] Furthermore, the physicochemical property characteristic vector includes the average hydrophobicity index, average isoelectric point, net charge, molecular weight, aliphatic index, and instability index.

[0027] The average hydrophobicity index is obtained by averaging the Kyte-Doolittle hydrophobicity values ​​of the amino acids in the peptide segment, and the calculation formula is as follows:

[0028] ;

[0029] In the formula, MH is the average Kyte-Doolittle hydrophobicity of the amino acids in the peptide; L is the total length of the peptide; h(k) is the Kyte-Doolittle hydrophobicity value of the kth amino acid in the sequence, which is a constant that can be obtained by looking up a table.

[0030] The average isoelectric point was calculated iteratively using the Henderson-Hasselbalch equation. The isoelectric point pI of a peptide is defined as the pH value at which the net charge of the peptide is zero, and the pH value is calculated using the following formula:

[0031] ;

[0032] In the formula, pH is the negative logarithm of the hydrogen ion concentration; pKa is the acid dissociation constant of the ionized group; [A - [HA] represents the concentration in the deprotonated form; [HA] represents the concentration in the protonated form.

[0033] The above equation was applied to all ionizable groups in the peptide (including N-terminus, C-terminus, and charged side chain amino acids), and the solution was obtained iteratively in the pH range of 0-14 using a dichotomy method.

[0034] The net charge is obtained by subtracting the number of negatively charged acidic amino acids (e.g., aspartic acid D, glutamic acid E) from the number of positively charged basic amino acids (e.g., lysine K, arginine R, histidine H) in the peptide under human physiological conditions (pH = 7.4).

[0035] The molecular weight is obtained by adding the molecular weights of all amino acid residues in the peptide and then subtracting the mass of water molecules lost in the condensation reaction. The calculation formula is as follows:

[0036] ;

[0037] In the formula, MW is the molecular weight of the peptide; m(k) is the molecular weight of the kth amino acid residue; 18 is the molecular weight of one water molecule (H2O) removed in each condensation reaction; L - 1 corresponds to a peptide of length L that needs to undergo L - 1 condensation reactions to form a peptide chain.

[0038] The formula for calculating the aliphatic index is as follows:

[0039] ;

[0040] In the formula, AI is the aliphatic index; A%, V%, I%, and L% are the molar percentages of alanine, valine, isoleucine, and leucine in the peptide, respectively; and the coefficients 2.9 and 3.9 reflect the relative volume of the aliphatic side chains of different amino acids.

[0041] The instability index was calculated based on the dipeptide instability weight table established by Guruprasad et al., and is used to predict the stability of peptides in vivo. The calculation formula is as follows:

[0042] ;

[0043] In the formula, II is the instability index; L is the peptide length; x i and x i+1 For the i-th and (i+1)-th adjacent amino acids in the sequence; DIWV(x i x i+1 The value (II) represents the instability weight of the dipeptide, which can be found in the standard dipeptide instability weight table. When II < 40, the peptide is stable in vivo; when II ≥ 40, the peptide is unstable.

[0044] Furthermore, the sequence embedding vector is obtained by inputting the peptide sequence into a pre-trained protein language model ESM-2.

[0045] Specifically, the operation of obtaining sequence embedding vectors using the pre-trained protein language model ESM-2 is as follows: The amino acid sequence of the peptide is input into the ESM-2 model. Each amino acid in the sequence is processed layer by layer by 33 Transformer layers, generating a 1280-dimensional hidden state vector in the last layer. This vector condenses the semantic information of the amino acid in the context of the entire sequence. A peptide of length L will generate L 1280-dimensional hidden state vectors. To obtain a fixed-length sequence-level representation, the average value of these L 1280-dimensional hidden state vectors is taken along the sequence dimension (average pooling), finally obtaining a 1280-dimensional sequence embedding vector, denoted as E. seq .

[0046] Furthermore, the three-dimensional structural features include secondary structure proportions and structural representation vectors.

[0047] Furthermore, the three-dimensional atomic coordinates of the peptide are first predicted using protein structure prediction tools such as ESMFold or AlphaFold2. Then, the proportion of secondary structures in the peptide is analyzed and statistically analyzed using the DSSP algorithm based on the three-dimensional atomic coordinates. Protein structure prediction tools, based on deep learning, can provide the position of each atom in three-dimensional space with high accuracy, accompanied by a residue-by-residue confidence score (pLDDT), ranging from 0 to 100. A higher score indicates a more reliable prediction for that position. Secondary structure refers to the local folding pattern of a peptide, mainly including three types: α-helix (a tight, helical structure), β-sheet (a sheet-like structure formed by multiple chains arranged side-by-side), and random coil (irregular, flexible regions). This invention uses the DSSP algorithm to analyze the predicted three-dimensional structure and statistically calculate the proportion of residues occupied by each of the above three secondary structures. The calculation formula is as follows:

[0048] ;

[0049] ;

[0050] ;

[0051] In the formula, P helix P sheet P coil These represent the proportions of residues in α-helix, β-sheet, and random coil states, respectively; L is the total length of the peptide, and N is the total length of the peptide. helix N sheet N coi l represents the number of amino acid residues in α-helix, β-sheet, and random coil states, respectively, and N represents the number of residues in these states. helix N sheet N coi The sum of l equals 1.

[0052] Furthermore, the three-dimensional atomic coordinates of the peptide are first predicted using the protein structure prediction tool ESMFold or AlphaFold2. Then, based on the three-dimensional atomic coordinates, a residue contact graph is constructed with Cα atoms of amino acid residues in the peptide as nodes and amino acid residue pairs with a spatial distance of less than 10 Å as edges. The graph is then input into a 4-layer E(3)-equivariant graph neural network (EGNN) for encoding to obtain the structure representation vector.

[0053] To capture the spatial relationships between amino acid residues in the three-dimensional structure of peptides more precisely, this invention models the three-dimensional structure of peptides as a residue contact "graph". This graph can be imagined as a social network: each amino acid residue is a "person" in the network, i.e., a node; if two amino acid residues are very close in space, i.e., the distance between the Cα atoms (the central carbon atoms in the amino acid backbone) of the amino acid residues is less than 10 Å, a line is drawn between the Cα atoms of the two amino acid residues, i.e., an edge, indicating that the two amino acid residues are spatial "neighbors".

[0054] Formally, the residue contact graph is denoted as G = (V, E), where V is the set of nodes and E is the set of edges. Each node in the node set V corresponds to a Cα atom of an amino acid residue. Each node carries a set of features: one-hot encoding of the amino acid type (a 20-dimensional vector, with the corresponding amino acid type position set to 1 and the rest to 0), main chain dihedral angles phi and psi (2-dimensional, describing local turns of the main chain), and solvent accessibility (1-dimensional, representing the degree to which the residue is exposed to the solvent), for a total of 23 dimensions. Each edge in the edge set E connects two amino acid residues whose spatial distance is less than 10 Å to their Cα atoms, and the feature of each edge is the Euclidean distance between these two Cα atoms.

[0055] After the residue contact map is constructed, it is encoded using an E(3)-isovariant graphical neural network. "E(3)-isovariant" means that regardless of rotation or translation of the peptide's three-dimensional structure, the features extracted by the model will not change—this is physically reasonable because the properties of a molecule should not depend on the observer's perspective. The core update rule of the E(3)-isovariant graphical neural network is as follows:

[0056] ;

[0057] In the formula, h i (l) This represents the node feature vector of the i-th node (corresponding to one amino acid residue) in the l-th layer; h i (l +1) This represents the node feature vector of the i-th node after being updated by the (l+1)-th layer E(3)-equivariant graph neural network; h j (l) This represents the feature vector of the j-th node adjacent to node i in the l-th layer. Indicates the update function; Represents a message function; x i -x j Represents the distance between node i and node j; e ij Represents the characteristics of an edge.

[0058] The above formula can be simply understood as follows: For node i, at the (l+1)th layer, its new feature h i (l+1) The calculation is as follows: First, for each of its spatial neighbors j, its current features h are comprehensively considered. i (l) Characteristics of neighbors h j (l) The square of the distance between the two ||x i -x j || 2 and the feature e of the edge ij Through a small neural network (This is called a message function) generates a "message"; then, it adds up the messages from all its neighbors (summation and aggregation), along with its current features, and then passes it through another small neural network. (This is called the update function) generates new features. Intuitively, each layer of the E(3)-equivariant graph neural network allows each residue to "listen" to information from its spatial neighbors. After multiple layers of transmission, the features of each node are integrated with structural information from the local area and even further away.

[0059] This invention uses a 4-layer E(3)-equivariant graph neural network to encode the residue contact graph G = (V, E). Finally, the output features of all nodes are averaged (average pooling) to obtain a 256-dimensional structural representation vector, denoted as E. struct .

[0060] Furthermore, the receptor binding features include binding free energy and binding characterization vector. The steps for extracting receptor binding features include: molecularly docking the peptide with GABAA receptor and 5-HT1A receptor to obtain the optimal binding conformation and binding free energy; voxelizing the binding interface of the optimal binding conformation; and inputting the data into a three-dimensional convolutional neural network to obtain a 128-dimensional binding characterization vector, denoted as E. bind .

[0061] Specifically, the receptor binding feature extraction operation is as follows:

[0062] (1) Selection of target receptors: Based on the neurobiological mechanisms of sleep regulation, this invention selects two core receptors as screening targets. The first is the α1 subunit of the GABAA receptor (the experimentally resolved three-dimensional structure is designated PDB ID: 6HUG). GABA is the most important inhibitory neurotransmitter in the brain. Activation of the GABAA receptor slows down neuronal firing, producing sedative and hypnotic effects. Most clinical sleep aids (such as benzodiazepines) exert their effects by acting on this receptor. The second is the 5-HT1A receptor (PDB ID: 7E2Y). 5-HT plays an important role in the regulation of the sleep-wake cycle. Activation of the 5-HT1A receptor can promote non-rapid eye movement (NREM) sleep.

[0063] (2) Molecular docking: The molecular docking simulation of peptides and receptors was performed using AutoDock Vina software. The principle of molecular docking is to allow the peptide molecule to continuously change position and orientation near the active site of the receptor protein in the computer, searching for the binding conformation with the lowest energy (i.e., the most stable). The specific steps include: first, preprocessing the receptor structure, including removing water of crystallization and the original ligands, completing hydrogen atoms, and calculating atomic charges; then, setting a 30 × 30 × 30 Å cuboid search space as the docking box with the active site of the receptor as the center; during docking, the exhaustiveness parameter was set to 32 (the larger the value, the more thorough the search), and 9 optimal binding conformations were output. The binding conformation was evaluated by the binding free energy ΔG (unit: kcal / mol). The larger the absolute value of ΔG, the tighter the binding and the stronger the affinity.

[0064] (3) Voxelization of the binding interface: From the optimal binding conformation obtained by molecular docking, the three-dimensional spatial information of the interface (i.e., the binding interface) where the peptide and receptor "embrace" each other is converted into numerical features that the model can process. This invention uses the "voxelization" method to convert the three-dimensional spatial information of the binding interface into numerical features that the model can process: a 20 × 20 × 20 three-dimensional grid (similar to three-dimensional pixels) is set in the binding interface region, and each small cell (voxel) is marked as "receptor atom", "peptide atom" or "empty" according to the type of atom it contains, thus transforming the complex three-dimensional binding conformation into a three-dimensional discrete matrix;

[0065] (4) 3D Convolutional Neural Network Processing: A 3D convolutional neural network (3D-CNN) is used to process the 3D discrete matrix obtained by voxelization. The principle of the 3D convolutional neural network is similar to that of the convolutional neural network in image processing, except that the convolution kernel of the 3D convolutional neural network is three-dimensional, which can capture local patterns in space and extract spatial features from low level to high level layer by layer. The process is as follows:

[0066] First layer: 3D convolution (1 input channel, 32 output channels, kernel size 3×3×3), followed by ReLU activation function and 3D max pooling (window size 2×2×2).

[0067] The second layer: 3D convolution (32 input channels, 64 output channels, kernel size 3×3×3), followed by ReLU activation and 3D max pooling;

[0068] The third layer consists of 3D convolution (64 input channels, 128 output channels, kernel size 3×3×3), followed by ReLU activation and global average pooling. Global average pooling compresses the entire 3D feature map into a 128-dimensional vector, which condenses the spatial interaction pattern information of the peptide-receptor binding interface.

[0069] Furthermore, a cross-attention mechanism is used to fuse multidimensional features (sequence features, 3D structural features, and receptor binding features) to obtain a fused feature. Specifically, the sequence features, 3D structural features, and receptor binding features are first concatenated within their respective modalities, and then mapped to a unified 256-dimensional model through fully connected layers to obtain enhanced sequence features, enhanced 3D structural features, and enhanced receptor binding features. Finally, the 256-dimensional enhanced features from the three modalities are fused to obtain the fused feature, denoted as F. fusion The specific steps are as follows:

[0070] First, at the sequence level, a cross-attention mechanism is used to concatenate amino acid composition features (20-dimensional vector), dipeptide composition features (400-dimensional vector), and physicochemical property features (6-dimensional vector) to form a 426-dimensional traditional sequence feature. This feature is then mapped to 256 dimensions through a fully connected layer and fused with the sequence embedding vector (1280-dimensional vector) to form a 256-dimensional enhanced sequence feature. At the three-dimensional structure level, the secondary structure ratio (3-dimensional vector) is dimensionally increased and fused with the structural representation vector (256-dimensional vector) to form a 256-dimensional enhanced three-dimensional structural feature. At the receptor binding level, the binding free energy ΔG (1-dimensional vector) is concatenated with the binding representation vector (128-dimensional vector) to form a 256-dimensional enhanced receptor binding feature.

[0071] The formulas for mapping feature vectors of three different dimensions to a 256-dimensional space using linear projection are as follows:

[0072] ;

[0073] ;

[0074] ;

[0075] In the formula, Q represents the Query matrix. It is obtained by linear mapping from the input sequence; K represents the Key matrix. V represents the Value matrix. n is the sequence length, d model W represents the hidden dimension of the model. q W k W v This is a learnable weight matrix. It borrows the "Query-Key-Value" paradigm from the cross-attention mechanism: sequence features serve as the "query," 3D structural features as the "key," and receptor binding features as the "value." Intuitively, the model "asks" which information in the structural features (key) is most relevant to the current sequence context based on the sequence features (query), and then extracts the corresponding information from the binding features (value).

[0076] The formula for calculating attention weights is:

[0077] ;

[0078] In the formula, QK T Calculate the dot product similarity between the query and the key. It is a scaling factor to prevent the gradient of the softmax function from vanishing due to excessively large dot product values. The softmax function converts similarity into probability weights between 0 and 1, and then uses these weights to perform a weighted summation of the value vectors.

[0079] To enable the model to focus on the relationships between features from multiple different "angles" simultaneously, this invention employs a multi-head attention mechanism, using eight attention heads to perform parallel computation. Each head performs attention computation in an independent 32-dimensional subspace (256 / 8 = 32).

[0080] ;

[0081] ;

[0082] In the formula, Q represents the Query matrix. It is obtained by linear mapping from the input sequence; K represents the Key matrix. V represents the Value matrix. n is the sequence length, d model The hidden dimension of the model; Let i be the Query projection matrix corresponding to the i-th head. ; Let i be the Key projection matrix corresponding to the i-th head. ; Let be the Value projection matrix corresponding to the i-th head. ;dk d represents the dimension of the Query / Key in each attention head. v This represents the dimension of the Value in each attention head.

[0083] Finally, the outputs of the three modalities after processing by the attention mechanism (each with 256 dimensions) are concatenated to obtain a fused feature vector F with 256 × 3 = 768 dimensions. fusion .

[0084] Furthermore, the prediction tasks include prediction of sleep-aiding activity, prediction of GABA pathway regulation, prediction of 5-HT pathway regulation, and prediction of drug-likeness.

[0085] Task 1: Predicting Sleep-Aid Activity: Predicting sleep-aid activity is the core prediction task. The network structure consists of three fully connected layers, with the dimensionality decreasing layer by layer (768→256→64→1). The hidden layers use the ReLU activation function to introduce non-linearity, and the output layer uses the Sigmoid function to compress the output value to between 0 and 1, representing the probability P of sleep-aid activity. sleep The closer the value is to 1, the more the model believes that the peptide has sleep-inducing activity;

[0086] Task 2: GABA Pathway Regulation Prediction: Predict the likelihood of peptides regulating the GABAA receptor. The network structure is the same as in Task 1, and the output is the probability P of GABA receptor regulation. GABA ;

[0087] Task 3: 5-HT Pathway Regulation Prediction: Predict the likelihood of peptides regulating the 5-HT1A receptor. The network structure is the same as in Task 1, outputting the 5-HT receptor regulation probability P. 5HT .

[0088] Task 4: Drugability Prediction: Predicting the practicality of peptides as drug or functional food ingredients. The network structure is 768→256→64→4, outputting four consecutive scores, corresponding to in vivo stability (not being rapidly degraded by proteases), water solubility (whether it can dissolve in physiological fluids), membrane permeability (whether it can cross the intestinal or blood-brain barrier), and metabolic stability (whether it can maintain a sufficient concentration in vivo).

[0089] Furthermore, when training the multimodal fusion network model, since the model performs four tasks simultaneously, this invention designs a comprehensive loss function to quantify the difference between the model's predicted probability and the true category (having sleep-aiding activity / not having sleep-aiding activity). The total loss function is the weighted sum of the losses from each task, and the calculation formula is as follows:

[0090] ;

[0091] In the formula, L totalλ1, λ2, λ3, and λ4 are the total loss function; λ1, λ2, λ3, and λ4 are weighting coefficients with values ​​of λ1 = 1.0, λ2 = 0.5, λ3 = 0.5, and λ4 = 0.3; L sleep L GABA L 5HT L drug The loss functions are used for predicting sleep-aiding activity, GABA pathway regulation, 5-HT pathway regulation, and drug-likeness, respectively. In this invention, sleep-aiding activity prediction is the primary task with the largest weight; GABA and 5-HT pathway prediction are secondary tasks with the next largest weight; and drug-likeness prediction has the smallest weight.

[0092] For the loss function L of the main task sleep Considering that the number of positive samples is usually much smaller than the number of negative samples, if a common cross-entropy loss function is used, the model will tend to predict all samples as negative to achieve a higher surface accuracy, but in reality, it will miss a large number of peptides with sleep-inducing activities. To solve this problem of positive and negative sample imbalance, this invention adopts the Focal Loss function:

[0093] ;

[0094] In the formula, FL(p) t ) represents the value of the Focal Loss function, used to measure the model's prediction error for a single sample; p t The model predicts the correct class of the sample; (1 - p) t ) γ p is the modulation factor; γ is the focusing parameter, γ = 2; the clever aspect of the modulation factor in this loss function is that for samples that are easily and correctly classified, p t Approaching 1, (1 - p) t ) γ When p approaches zero, the loss is significantly reduced; for samples that are difficult to classify, p t Smaller, (1 - p) t ) γ The loss remains constant even when the value is relatively large. This allows the model to automatically focus its training attention on samples that are difficult to classify. α t These are the class weights, used to further balance the influence of positive and negative samples.

[0095] For the loss function L of the GABA and 5-HT pathway prediction task GABA L 5HT This invention uses the standard binary cross-entropy loss (BCE) function. For the drug-likeness prediction task, the loss function L... drug Since the output is a continuous score, this invention uses the mean squared error loss (MSE) function.

[0096] Furthermore, when training the multimodal fusion network model, the AdamW optimizer is used for iterative optimization. AdamW is an improved version of Adam, which incorporates weight decay (weight decay=0.01) to prevent the model parameters from becoming too large and causing overfitting.

[0097] The initial learning rate is set to 1×10. -4 The learning rate is scheduled using a cosine annealing strategy. The learning rate smoothly decreases from its initial value to near zero over one cycle (50 epochs) in the shape of a cosine function, and then increases again. This strategy helps the model escape local optima and find better parameter solutions.

[0098] Each batch contains 32 peptides. The model is trained for 100 epochs (i.e., iterates through the entire training set 100 times) and employs an early stopping strategy: if the performance on the validation set does not improve for 10 consecutive epochs, training is terminated early to prevent overfitting.

[0099] In terms of data augmentation, the input sequence is randomly masked, that is, some amino acids in the sequence are randomly replaced with special labels with a 15% probability, forcing the model to learn more robust feature representations and reducing over-reliance on individual amino acids.

[0100] In terms of regularization, Dropout (random inactivation rate of 0.2, i.e., 20% of neurons in each layer are randomly turned off during training) and Layer Normalization (layer normalization, stabilizing the distribution of outputs of each layer) are used together to further suppress overfitting.

[0101] Furthermore, the method further includes the step of: performing interpretability analysis on the prediction results (peptides with higher prediction scores), wherein the interpretability analysis method includes:

[0102] (1) Visualization of attention weights

[0103] In a multi-head attention layer, there is an attention weight value between each pair of amino acid positions. The attention weight matrix is ​​extracted. , where element A ij This indicates the "level of attention" the model pays to position i over position j when making predictions. For each amino acid position i, its overall importance score is calculated:

[0104] ;

[0105] In the formula, Importance(i) is the overall importance score of the i-th amino acid position in the model prediction process, used to measure the contribution of this position to the model decision; L is the length of the amino acid sequence, i.e., the total number of amino acid residues in the sequence; A ij The elements in the attention weight matrix represent the attention weight (degree of attention) of the i-th amino acid position to the j-th amino acid position when the model makes predictions.

[0106] This refers to the average attention score of this location relative to all other locations. After normalizing the importance scores of all locations, a heatmap is plotted, with darker colors indicating a greater contribution to the model's predictions and potentially key amino acid sites influencing sleep-inducing activity.

[0107] (2) SHAP value analysis

[0108] SHAP (SHapley Additive exPlanations) is a feature attribution method derived from game theory. It treats each prediction of the model as a "cooperative game": each input feature is a "participant," and the model's prediction is the "total payout" created by all participants working together. The goal of SHAP value analysis is to fairly allocate the contribution of each participant.

[0109] For the i-th feature x in the input feature vector x i Its Shapley value is defined as:

[0110] ;

[0111] In the formula, N is the set of all features, S is any subset of N that does not contain the i-th feature, and f(S) represents the predicted output of the model when only the feature subset S is used. This represents the predicted output after adding the i-th feature to the subset S. The difference between the two is the marginal contribution of the i-th feature in this "pairing" in S. The formula takes a weighted average over all possible subsets S, with the weights determined by coefficients in combinatorics to ensure fair allocation.

[0112] Since accurately calculating the Shapley value requires traversing all possible combinations of feature subsets (an exponential complexity), the KernelSHAP approximation algorithm is used in practice to accelerate the calculation. Finally, the SHAP value analysis outputs a ranking of the contribution of each amino acid site and each physicochemical property feature to the prediction results. A positive contribution indicates that the feature helps improve activity prediction, while a negative contribution indicates that the feature reduces activity prediction.

[0113] (3) Virtual mutation analysis of key sites

[0114] After identifying key sites through attention weighting and SHAP value analysis, virtual mutation experiments were conducted to further validate the functional importance of these sites and explore optimization directions. Specifically, the original amino acids at each key site were sequentially replaced with 19 other amino acids. The predicted activity of the model was recalculated for each mutant, and the difference in activity before and after mutation was compared, as shown in the following formula:

[0115] ;

[0116] In the formula, ΔP represents the change in predicted activity before and after the mutation, used to measure the impact of a mutation at a certain amino acid site on the model's prediction results; P wildtype P represents the probability of predicted sleep-aiding activity for the original sequence (wild type). mutant This represents the predicted activity probability of the mutated sequence. If ΔP > 0.1, the mutation is likely to increase activity (beneficial mutation); if ΔP < -0.1, the mutation will significantly decrease activity (detrimental mutation), and the original site is crucial to activity.

[0117] A second aspect of the present invention provides a smart screening system for sleep-aiding active peptides for implementing the method, comprising:

[0118] The multidimensional feature extraction module is used to extract sequence features, three-dimensional structural features, and receptor binding features of peptides;

[0119] The multidimensional feature fusion module is used to fuse the sequence features, three-dimensional structural features, and receptor binding features of peptides.

[0120] The prediction module is used to train a multimodal fusion network model using fusion features, and then uses the trained multimodal fusion network model to predict peptides with unknown activity, thereby screening out peptides with sleep-aiding activity.

[0121] A third aspect of the present invention provides a computer device including a memory and a processor, the memory storing a computer program, and the processor implementing the method when processing the computer program.

[0122] Compared with the prior art, the advantages of the present invention are:

[0123] This invention utilizes multimodal characterization learning and fusion modeling of peptide sequence information, three-dimensional structural information, and receptor binding information to achieve high-throughput prediction, ranking and screening, and interpretability analysis of key sites for sleep-aiding active peptides. Compared to the traditional QSAR model, the multimodal fusion network model of this invention significantly outperforms other indices such as accuracy, precision, recall, F1 score, AUC-ROC, and AUC-PR, and possesses good generalization ability and accuracy in predicting active peptides. It can be widely applied to the discovery and optimization of sleep-aiding active peptides in the development of functional foods, health foods, and peptide drugs. Detailed Implementation

[0124] To enable those skilled in the art to clearly and completely understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art. Obviously, the embodiments described herein are only used to explain the present invention and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0125] This invention provides a method for intelligent screening of sleep-aiding active peptides based on multimodal fusion networks, comprising:

[0126] 1. Construct an active peptide dataset including positive and negative samples, and perform redundancy removal; wherein, the positive samples include peptides with sleep-aiding, sedative or anti-anxiety activities, and the negative samples include peptides known not to have sleep-aiding activities;

[0127] In the following specific embodiments, the peptides in the positive samples are between 3 and 50 amino acids in length, and each peptide has clear activity data to support it, such as the half-maximal inhibitory concentration (IC50). 50 , half effective concentration EC 50Or animal experimental results, etc.; the length distribution of peptides in the negative samples matches that of the positive samples. The sources of the positive samples include: peptides labeled as sedative or anxiolytic in public databases (e.g., BIOPEP-UWM), peptides that can activate GABAA receptors reported in academic literature (e.g., PubMed), approved peptide sleep aids and their structural analogs, and active peptides identified and verified by the inventors' research group in previous studies from tortoise shell protein and casein protein hydrolysates; the sources of the negative samples include: peptides randomly selected from general protein databases (e.g., UniProt), and other types of functional peptides that have been experimentally proven not to have sleep-aiding functions (e.g., certain antimicrobial peptides and antioxidant peptides). To avoid high sequence similarity between negative and positive samples, this invention uses the CD-HIT tool to perform redundancy removal on all samples with a 40% sequence similarity threshold.

[0128] 2. Divide the active peptide dataset into training, validation, and test sets according to a preset ratio;

[0129] In the following specific embodiments, the present invention divides the active peptide dataset into a training set, a validation set, and a test set in a ratio of 8:1:1, ensuring that the ratio of positive samples to negative samples remains consistent in each subset. The training set is used for model training, the validation set is used to monitor model performance and prevent overfitting during model training, and the test set is used to finally evaluate the generalization ability of the model.

[0130] 3. Construct a multimodal fusion network model; the multimodal fusion network model includes an extraction layer, a fusion layer, and a prediction layer; wherein, the extraction layer is used to extract multidimensional features of peptides, the multidimensional features including sequence features, three-dimensional structural features, and receptor binding features; the fusion layer is used to fuse the multidimensional features to obtain fused features; the prediction layer takes the fused features as input and performs a prediction task;

[0131] In the following specific embodiments, the sequence features include amino acid composition features (AAC), dipeptide composition features (DPC), physicochemical property features, and sequence embedding vectors;

[0132] The amino acid composition characteristics were obtained by statistically analyzing the frequency of occurrence of 20 amino acids. Proteins in nature are composed of 20 standard amino acids. For a peptide segment of length L, the frequency of occurrence of each amino acid was calculated, resulting in a 20-dimensional amino acid composition feature vector. The formula for calculating the frequency of the i-th amino acid is:

[0133] ;

[0134] In the formula, AAC(i) is the frequency of the i-th amino acid; N(i) is the number of times the i-th amino acid appears in the peptide; and L is the total length of the peptide. For example, in a peptide "WWRGH" containing 5 amino acids, tryptophan (W) appears twice, so the AAC value of tryptophan is 2 / 5 = 0.4.

[0135] The dipeptide composition features were obtained by statistically analyzing the frequency of 400 dipeptide combinations. Knowing only the proportion of each amino acid is insufficient; the arrangement pattern between adjacent amino acids is equally important. By pairwise combining 20 amino acids, there are 20 × 20 = 400 dipeptide combinations. For a peptide segment of length L, the frequency of each dipeptide combination is statistically analyzed to obtain a 400-dimensional dipeptide composition feature vector, capturing local arrangement pattern information within the peptide segment. The formula for calculating the frequency of each dipeptide combination (i, j) is as follows:

[0136] ;

[0137] In the formula, DPC(i, j) is the frequency of occurrence of each dipeptide combination (i, j); N(i, j) is the number of times the combination of amino acid i followed by amino acid j appears in the peptide segment; L - 1 corresponds to the total number of adjacent amino acid pairs in a peptide segment of length L.

[0138] The physicochemical properties include average hydrophobicity index, average isoelectric point, net charge, molecular weight, aliphatic index, and instability index.

[0139] The average hydrophobicity index measures the overall hydrophilic or lipophilic tendency of a peptide. Peptides with high hydrophobicity are more likely to cross cell membranes. The average Kyte-Doolittle hydrophobicity value of the amino acids in the peptide is calculated using the following formula:

[0140] ;

[0141] In the formula, MH is the average Kyte-Doolittle hydrophobicity of the amino acids in the peptide; L is the total length of the peptide; h(k) is the Kyte-Doolittle hydrophobicity value of the kth amino acid in the sequence, which is a constant that can be obtained by looking up a table.

[0142] The average isoelectric point is the pH value corresponding to zero net charge of the peptide, calculated iteratively using the Henderson-Hasselbalch equation. The formula for calculating the pH value is as follows:

[0143] ;

[0144] In the formula, pH is the negative logarithm of the hydrogen ion concentration; pKa is the acid dissociation constant of the ionized group; [A - [HA] represents the concentration in the deprotonated form; [HA] represents the concentration in the protonated form.

[0145] By applying the above equation to all ionizable groups in the peptide (including N-terminus, C-terminus, and charged side chain amino acids), the average isoelectric point is obtained by iteratively solving the problem using a dichotomy method within the pH range of 0-14.

[0146] The net charge is obtained by subtracting the number of negatively charged acidic amino acids (e.g., aspartic acid D, glutamic acid E) from the number of positively charged basic amino acids (e.g., lysine K, arginine R, histidine H) in the peptide under human physiological conditions (pH = 7.4).

[0147] The molecular weight is obtained by adding the molecular weights of all amino acid residues in the peptide and then subtracting the mass of water molecules lost in the condensation reaction. The calculation formula is as follows:

[0148] ;

[0149] In the formula, MW is the molecular weight of the peptide; m(k) is the molecular weight of the kth amino acid residue; 18 is the molecular weight of one water molecule (H2O) removed in each condensation reaction; L - 1 corresponds to a peptide of length L that needs to undergo L - 1 condensation reactions to form a peptide chain.

[0150] The formula for calculating the aliphatic index is as follows:

[0151] ;

[0152] In the formula, AI is the aliphatic index; A%, V%, I%, and L% are the molar percentages of alanine, valine, isoleucine, and leucine in the peptide, respectively; and the coefficients 2.9 and 3.9 reflect the relative volume of the aliphatic side chains of different amino acids.

[0153] The instability index was calculated based on the dipeptide instability weight table established by Guruprasad et al., and is used to predict the stability of peptides in vivo. The calculation formula is as follows:

[0154] ;

[0155] In the formula, II is the instability index; L is the peptide length; x i and x i+1 For the i-th and (i+1)-th adjacent amino acids in the sequence; DIWV(x i x i+1The value (II) represents the instability weight of the dipeptide, which can be found in the standard dipeptide instability weight table. When II < 40, the peptide is stable in vivo; when II ≥ 40, the peptide is unstable.

[0156] The sequence embedding vector is obtained by inputting the peptide sequence into a pre-trained protein language model, ESM-2. While the sequence features obtained above, such as amino acid composition features, dipeptide composition features, and physicochemical property features, are effective, they still struggle to capture the complex long-range dependencies within peptide sequences. Therefore, this invention utilizes the pre-trained protein language model ESM-2 to further obtain deeper sequence representations. The ESM-2 protein language model is a large-scale protein language model released by the Meta AI team, and its principle is similar to the BERT model in natural language processing. The ESM-2 model has been pre-trained on hundreds of millions of protein sequences, learning the "grammar of proteins"—that is, the arrangement rules and functional associations between amino acids.

[0157] Specifically, the operation of obtaining sequence embedding vectors using the pre-trained protein language model ESM-2 is as follows: The amino acid sequence of the peptide is input into the ESM-2 model. Each amino acid in the sequence is processed layer by layer by 33 Transformer layers, generating a 1280-dimensional hidden state vector in the last layer. This vector condenses the semantic information of the amino acid in the context of the entire sequence. A peptide of length L will generate L 1280-dimensional hidden state vectors. To obtain a fixed-length sequence-level representation, the average value of these L 1280-dimensional hidden state vectors is taken along the sequence dimension (average pooling), finally obtaining a 1280-dimensional sequence embedding vector, denoted as E. seq E seq This is the "identity card" of the peptide segment in the sequence modality.

[0158] In the following specific embodiments, the three-dimensional structural features include secondary structure proportions and structural representation vectors.

[0159] First, the three-dimensional atomic coordinates of the peptide are predicted using protein structure prediction tools such as ESMFold or AlphaFold2. Then, based on these coordinates, the DSSP algorithm is used to analyze and statistically determine the proportion of secondary structures in the peptide. Protein structure prediction tools, based on deep learning, can accurately determine the position of each atom in three-dimensional space, accompanied by a residue-by-residue confidence score (pLDDT), ranging from 0 to 100. A higher score indicates a more reliable prediction for that position. Secondary structure refers to the local folding pattern of a peptide, mainly including three types: α-helix (a tight, helical structure), β-sheet (a sheet-like structure formed by multiple chains arranged side-by-side), and random coil (irregular, flexible regions). This invention uses the DSSP algorithm to analyze the predicted three-dimensional structure and statistically determine the proportion of residues occupied by each of these three secondary structures. The calculation formula is as follows:

[0160] ;

[0161] ;

[0162] ;

[0163] In the formula, P helix P sheet P coil These represent the proportions of residues in α-helix, β-sheet, and random coil states, respectively; L is the total length of the peptide, and N is the total length of the peptide. helix N sheet N coi l represents the number of amino acid residues in α-helix, β-sheet, and random coil states, respectively, and N represents the number of residues in these states. helix N sheet N coi The sum of l equals 1.

[0164] First, the three-dimensional atomic coordinates of the peptide are predicted using the protein structure prediction tool ESMFold or AlphaFold2. Then, based on the three-dimensional atomic coordinates, a residue contact graph is constructed with Cα atoms of amino acid residues in the peptide as nodes and amino acid residue pairs with a spatial distance of less than 10 Å as edges. The graph is then input into a 4-layer E(3)-equivariant graph neural network (EGNN) for encoding to obtain the structure representation vector.

[0165] To capture the spatial relationships between amino acid residues in the three-dimensional structure of peptides more precisely, this invention models the three-dimensional structure of peptides as a residue contact "graph". This graph can be imagined as a social network: each amino acid residue is a "person" in the network, i.e., a node; if two amino acid residues are very close in space, i.e., the distance between the Cα atoms (the central carbon atoms in the amino acid backbone) of the amino acid residues is less than 10 Å, a line is drawn between the Cα atoms of the two amino acid residues, i.e., an edge, indicating that the two amino acid residues are spatial "neighbors".

[0166] Formally, the residue contact graph is denoted as G = (V, E), where V is the set of nodes and E is the set of edges. Each node in the node set V corresponds to a Cα atom of an amino acid residue. Each node carries a set of features: one-hot encoding of the amino acid type (a 20-dimensional vector, with the corresponding amino acid type position set to 1 and the rest to 0), main chain dihedral angles phi and psi (2-dimensional, describing local turns of the main chain), and solvent accessibility (1-dimensional, representing the degree to which the residue is exposed to the solvent), for a total of 23 dimensions. Each edge in the edge set E connects two amino acid residues whose spatial distance is less than 10 Å to their Cα atoms, and the feature of each edge is the Euclidean distance between these two Cα atoms.

[0167] After the residue contact map is constructed, it is encoded using an E(3)-isovariant graphical neural network. "E(3)-isovariant" means that regardless of rotation or translation of the peptide's three-dimensional structure, the features extracted by the model will not change—this is physically reasonable because the properties of a molecule should not depend on the observer's perspective. The core update rule of the E(3)-isovariant graphical neural network is as follows:

[0168] ;

[0169] In the formula, h i (l) This represents the node feature vector of the i-th node (corresponding to one amino acid residue) in the l-th layer; h i (l +1) This represents the node feature vector of the i-th node after being updated by the (l+1)-th layer E(3)-equivariant graph neural network; h j (l) This represents the feature vector of the j-th node adjacent to node i in the l-th layer. Indicates the update function; Represents a message function; x i -x j Represents the distance between node i and node j; e ij Represents the characteristics of an edge.

[0170] The above formula can be simply understood as follows: For node i, at the (l+1)th layer, its new feature h i (l+1) The calculation is as follows: First, for each of its spatial neighbors j, its current features h are comprehensively considered. i (l) Characteristics of neighbors h j (l) The square of the distance between the two ||x i -x j || 2 and the feature e of the edge ij Through a small neural network (This is called a message function) generates a "message"; then, it adds up the messages from all its neighbors (summation and aggregation), along with its current features, and then passes it through another small neural network. (This is called the update function) generates new features. Intuitively, each layer of the E(3)-equivariant graph neural network allows each residue to "listen" to information from its spatial neighbors. After multiple layers of transmission, the features of each node are integrated with structural information from the local area and even further away.

[0171] This invention uses a 4-layer E(3)-equivariant graph neural network to encode the residue contact graph G = (V, E). Finally, the output features of all nodes are averaged (average pooling) to obtain a 256-dimensional structural representation vector, denoted as E. struct .

[0172] In the following specific embodiments, the receptor binding features include binding free energy and binding characterization vector. The steps for extracting receptor binding features include: molecularly docking the peptide with GABAA receptor and 5-HT1A receptor to obtain the optimal binding conformation and binding free energy; voxelizing the binding interface of the optimal binding conformation; and inputting the data into a three-dimensional convolutional neural network to obtain a 128-dimensional binding characterization vector, denoted as E. bind .

[0173] The key to the function of sleep-aiding active peptides lies in whether the peptide can successfully bind to receptor proteins in the brain that regulate sleep. This is analogous to whether a key can be inserted and open a lock. Therefore, this invention simulates the binding process of peptides to target receptors and encodes the binding status as a binding characterization vector as a receptor binding feature.

[0174] Specifically, the receptor binding feature extraction operation is as follows:

[0175] (1) Selection of target receptors: Based on the neurobiological mechanisms of sleep regulation, this invention selects two core receptors as screening targets. The first is the α1 subunit of the GABAA receptor (the experimentally resolved three-dimensional structure is designated PDB ID: 6HUG). GABA is the most important inhibitory neurotransmitter in the brain. Activation of the GABAA receptor slows down neuronal firing, producing sedative and hypnotic effects. Most clinical sleep aids (such as benzodiazepines) exert their effects by acting on this receptor. The second is the 5-HT1A receptor (PDB ID: 7E2Y). 5-HT plays an important role in the regulation of the sleep-wake cycle. Activation of the 5-HT1A receptor can promote non-rapid eye movement (NREM) sleep.

[0176] (2) Molecular docking: The molecular docking simulation of peptides and receptors was performed using AutoDock Vina software. The principle of molecular docking is to allow the peptide molecule to continuously change position and orientation near the active site of the receptor protein in the computer, searching for the binding conformation with the lowest energy (i.e., the most stable). The specific steps include: first, preprocessing the receptor structure, including removing water of crystallization and the original ligands, completing hydrogen atoms, and calculating atomic charges; then, setting a 30 × 30 × 30 Å cuboid search space as the docking box with the active site of the receptor as the center; during docking, the exhaustiveness parameter is set to 32 (the larger the value, the more thorough the search), and 9 optimal binding conformations are output (selecting to output 9 optimal binding conformations is one of the recommended and default settings of AutoDock Vina, used to achieve a balance between computational efficiency and conformational diversity, ensuring that the main possible low-energy binding modes can be covered). The binding conformation is evaluated by the binding free energy ΔG (unit: kcal / mol). The larger the absolute value of ΔG, the tighter the binding and the stronger the affinity.

[0177] (3) Voxelization of the binding interface: From the optimal binding conformation obtained by molecular docking, the three-dimensional spatial information of the interface (i.e., the binding interface) where the peptide and receptor "embrace" each other is converted into numerical features that can be processed by the machine model. This invention uses the "voxelization" method to convert the three-dimensional spatial information of the binding interface into numerical features that can be processed by the model: a 20 × 20 × 20 three-dimensional grid (similar to three-dimensional pixels) is set in the binding interface region, and each small cell (voxel) is marked as "receptor atom", "peptide atom" or "empty" according to the type of atom it contains, thus transforming the complex three-dimensional binding conformation into a three-dimensional discrete matrix;

[0178] (4) 3D Convolutional Neural Network Processing: A 3D convolutional neural network (3D-CNN) is used to process the 3D discrete matrix obtained by voxelization. The principle of the 3D convolutional neural network is similar to that of the convolutional neural network in image processing, except that the convolution kernel of the 3D convolutional neural network is three-dimensional, which can capture local patterns in space and extract spatial features from low level to high level layer by layer. The process is as follows:

[0179] First layer: 3D convolution (1 input channel, 32 output channels, kernel size 3×3×3), followed by ReLU activation function and 3D max pooling (window size 2×2×2).

[0180] The second layer: 3D convolution (32 input channels, 64 output channels, kernel size 3×3×3), followed by ReLU activation and 3D max pooling;

[0181] The third layer consists of 3D convolution (64 input channels, 128 output channels, kernel size 3×3×3), followed by ReLU activation and global average pooling. Global average pooling compresses the entire 3D feature map into a 128-dimensional vector, which condenses the spatial interaction pattern information of the peptide-receptor binding interface.

[0182] In the following specific embodiment, a cross-attention mechanism is used to fuse multidimensional features (sequence features, three-dimensional structural features, and receptor binding features) to obtain fused features. Specifically, the sequence features, three-dimensional structural features, and receptor binding features are first concatenated within their respective modalities, and then mapped to a unified 256-dimensional matrix through fully connected layers to obtain enhanced sequence features, enhanced three-dimensional structural features, and enhanced receptor binding features. Finally, the 256-dimensional enhanced features from the three modalities are fused to obtain the fused features. The specific operation is as follows:

[0183] First, at the sequence level, a cross-attention mechanism is used to concatenate amino acid composition features (20-dimensional vector), dipeptide composition features (400-dimensional vector), and physicochemical property features (6-dimensional vector) to form a 426-dimensional traditional sequence feature. This feature is then mapped to 256 dimensions through a fully connected layer and fused with the sequence embedding vector (1280-dimensional vector) to form a 256-dimensional enhanced sequence feature. At the three-dimensional structure level, the secondary structure ratio (3-dimensional vector) is dimensionally increased and fused with the structural representation vector (256-dimensional vector) to form a 256-dimensional enhanced three-dimensional structural feature. At the receptor binding level, the binding free energy ΔG (1-dimensional vector) is concatenated with the binding representation vector (128-dimensional vector) to form a 256-dimensional enhanced receptor binding feature.

[0184] The formulas for mapping feature vectors of three different dimensions to a 256-dimensional space using linear projection are as follows:

[0185] ;

[0186] ;

[0187] ;

[0188] In the formula, Q represents the Query matrix. It is obtained by linear mapping from the input sequence; K represents the Key matrix. V represents the Value matrix. n is the sequence length, d model W represents the hidden dimension of the model. q W k W v This is a learnable weight matrix. It borrows the "Query-Key-Value" paradigm from the cross-attention mechanism: sequence features serve as the "query," 3D structural features as the "key," and receptor binding features as the "value." Intuitively, the model "asks" which information in the structural features (key) is most relevant to the current sequence context based on the sequence features (query), and then extracts the corresponding information from the binding features (value).

[0189] The formula for calculating attention weights is:

[0190] ;

[0191] In the formula, QK T Calculate the dot product similarity between the query and the key. It is a scaling factor to prevent the gradient of the softmax function from vanishing due to excessively large dot product values. The softmax function converts similarity into probability weights between 0 and 1, and then uses these weights to perform a weighted summation of the value vectors.

[0192] To enable the model to focus on the relationships between features from multiple different "angles" simultaneously, this invention employs a multi-head attention mechanism, using eight attention heads to perform parallel computation. Each head performs attention computation in an independent 32-dimensional subspace (256 / 8 = 32).

[0193] ;

[0194] ;

[0195] In the formula, Q represents the Query matrix. It is obtained by linear mapping from the input sequence; K represents the Key matrix. V represents the Value matrix. n is the sequence length, d model The hidden dimension of the model; Let i be the Query projection matrix corresponding to the i-th head. ; Let i be the Key projection matrix corresponding to the i-th head. ; Let be the Value projection matrix corresponding to the i-th head. ;d k d represents the dimension of the Query / Key in each attention head. v This represents the dimension of the Value in each attention head.

[0196] Finally, the outputs of the three modalities after processing by the attention mechanism (each with 256 dimensions) are concatenated to obtain a fused feature vector F with 256 × 3 = 768 dimensions. fusion .

[0197] In the following specific embodiments, the prediction tasks include sleep-aid activity prediction, GABA pathway regulation prediction, 5-HT pathway regulation prediction, and drug-likeness prediction. The multimodal fusion network model of this invention simultaneously performs four related but not identical prediction tasks. The advantage of this "multi-task learning" strategy is that different tasks share underlying feature representations and provide each other with additional supervision signals, thereby improving the prediction performance of each task.

[0198] Task 1: Predicting Sleep-Aid Activity: Predicting sleep-aid activity is the core prediction task. The network structure consists of three fully connected layers, with the dimensionality decreasing layer by layer (768→256→64→1). The hidden layers use the ReLU activation function to introduce non-linearity, and the output layer uses the Sigmoid function to compress the output value to between 0 and 1, representing the probability P of sleep-aid activity. sleep The closer the value is to 1, the more the model believes that the peptide has sleep-inducing activity;

[0199] Task 2: GABA Pathway Regulation Prediction: Predict the likelihood of peptides regulating the GABAA receptor. The network structure is the same as in Task 1, and the output is the probability P of GABA receptor regulation. GABA ;

[0200] Task 3: 5-HT Pathway Regulation Prediction: Predict the likelihood of peptides regulating the 5-HT1A receptor. The network structure is the same as in Task 1, outputting the 5-HT receptor regulation probability P. 5HT .

[0201] Task 4: Drugability Prediction: Predicting the practicality of peptides as drug or functional food ingredients. The network structure is 768→256→64→4, with four continuous scores output (the output layer does not use an activation function, but uses linear activation for four-dimensional continuous value regression), corresponding to in vivo stability (not rapidly degraded by proteases), water solubility (whether it can dissolve in physiological fluids), membrane permeability (whether it can cross the intestinal or blood-brain barrier), and metabolic stability (whether it can maintain a sufficient concentration in vivo).

[0202] 4. The multimodal fusion network model is trained using the training set, and the performance of the multimodal fusion network model is monitored using the validation set to prevent overfitting. The multimodal fusion network model is validated using the test set to obtain the trained multimodal fusion network model.

[0203] In the following specific embodiments, when training the multimodal fusion network model, since the model performs four tasks simultaneously, this invention designs a comprehensive loss function to quantify the difference between the model's predicted probability and the true category (having sleep-aiding activity / not having sleep-aiding activity). The total loss function is the weighted sum of the losses from each task, and the calculation formula is as follows:

[0204] ;

[0205] In the formula, L total λ1 represents the total loss function; λ2, λ3, and λ4 are weighting coefficients with values ​​of λ1 = 1.0, λ2 = 0.5, λ3 = 0.5, and λ4 = 0.3; L sleep L GABA L 5HT L drug The loss functions are used for predicting sleep-aiding activity, GABA pathway regulation, 5-HT pathway regulation, and drug-likeness, respectively. In this invention, sleep-aiding activity prediction is the primary task with the largest weight; GABA and 5-HT pathway prediction are secondary tasks with the next largest weight; and drug-likeness prediction has the smallest weight.

[0206] For the loss function L of the main task sleep Considering that the number of positive samples is usually much smaller than the number of negative samples, if a common cross-entropy loss function is used, the model will tend to predict all samples as negative to achieve a higher surface accuracy, but in reality, it will miss a large number of peptides with sleep-inducing activities. To solve this problem of positive and negative sample imbalance, this invention adopts the Focal Loss function:

[0207] ;

[0208] In the formula, FL(p) t) represents the value of the Focal Loss function, used to measure the model's prediction error for a single sample; p t The model predicts the correct class of the sample; (1 - p) t ) γ p is the modulation factor; γ is the focusing parameter, γ = 2; the clever aspect of the modulation factor in this loss function is that for samples that are easily and correctly classified, p t Approaching 1, (1 - p) t ) γ When p approaches zero, the loss is significantly reduced; for samples that are difficult to classify, p t Smaller, (1 - p) t ) γ The loss remains constant even when the value is relatively large. This allows the model to automatically focus its training attention on samples that are difficult to classify. α t These are the class weights, used to further balance the influence of positive and negative samples.

[0209] For the loss function L of the GABA and 5-HT pathway prediction task GABA L 5HT This invention uses the standard binary cross-entropy loss (BCE) function. For the drug-likeness prediction task, the loss function L... drug Since the output is a continuous score, this invention uses the mean squared error loss (MSE) function. If L sleep L GABA L 5HT L drug Using the same loss function for all tasks will lead to a mismatch between the task and the loss, which will directly reduce the effectiveness.

[0210] Furthermore, when training the multimodal fusion network model, the AdamW optimizer is used for iterative optimization. AdamW is an improved version of Adam, which incorporates weight decay (weight decay=0.01) to prevent the model parameters from becoming too large and causing overfitting.

[0211] The initial learning rate is set to 1×10. -4 The learning rate is scheduled using a cosine annealing strategy. The learning rate smoothly decreases from its initial value to near zero within one cycle (50 epochs) in the shape of a cosine function, and then increases again. This strategy helps the model escape local optima and find better parameter solutions.

[0212] Each batch contains 32 peptides. The model is trained for 100 epochs (i.e., iterates through the entire training set 100 times) and employs an early stopping strategy: if the performance on the validation set does not improve for 10 consecutive epochs, training is terminated early to prevent overfitting.

[0213] In terms of data augmentation, the input sequence is randomly masked, that is, some amino acids in the sequence are randomly replaced with special labels with a 15% probability, forcing the model to learn more robust feature representations and reducing over-reliance on individual amino acids.

[0214] In terms of regularization, Dropout (random inactivation rate of 0.2, i.e., 20% of neurons in each layer are randomly turned off during training) and Layer Normalization (layer normalization, stabilizing the distribution of outputs of each layer) are used together to further suppress overfitting.

[0215] 5. Input peptide sequences with unknown activity, use a trained multimodal fusion network model for prediction, perform interpretability analysis on peptides with high prediction scores, and screen out peptides with sleep-aiding activity.

[0216] In the following specific embodiments, the method for interpretability analysis includes:

[0217] (1) Visualization of attention weights

[0218] In a multi-head attention layer, there is an attention weight value between each pair of amino acid positions. The attention weight matrix is ​​extracted. , where element A ij This indicates the "level of attention" the model pays to position i over position j when making predictions. For each amino acid position i, its overall importance score is calculated:

[0219] ;

[0220] In the formula, Importance(i) is the overall importance score of the i-th amino acid position in the model prediction process, used to measure the contribution of this position to the model decision; L is the length of the amino acid sequence, i.e., the total number of amino acid residues in the sequence; A ij The elements in the attention weight matrix represent the attention weight (degree of attention) of the i-th amino acid position to the j-th amino acid position when the model makes predictions.

[0221] This refers to the average attention score of this location relative to all other locations. After normalizing the importance scores of all locations, a heatmap is plotted, with darker colors indicating a greater contribution to the model's predictions and potentially key amino acid sites influencing sleep-inducing activity.

[0222] (2) SHAP value analysis

[0223] SHAP (SHapley Additive exPlanations) is a feature attribution method derived from game theory. It treats each prediction of the model as a "cooperative game": each input feature is a "participant," and the model's prediction is the "total payout" created by all participants working together. The goal of SHAP value analysis is to fairly allocate the contribution of each participant.

[0224] For the i-th feature x in the input feature vector x i Its Shapley value is defined as:

[0225] ;

[0226] In the formula, N is the set of all features, S is any subset of N that does not contain the i-th feature, and f(S) represents the predicted output of the model when only the feature subset S is used. This represents the predicted output after adding the i-th feature to the subset S. The difference between the two is the marginal contribution of the i-th feature in this "pairing" in S. The formula takes a weighted average over all possible subsets S, with the weights determined by coefficients in combinatorics to ensure fair allocation.

[0227] Since accurately calculating the Shapley value requires traversing all possible combinations of feature subsets (an exponential complexity), the KernelSHAP approximation algorithm is used in practice to accelerate the calculation. Finally, the SHAP value analysis outputs a ranking of the contribution of each amino acid site and each physicochemical property feature to the prediction results. A positive contribution indicates that the feature helps improve activity prediction, while a negative contribution indicates that the feature reduces activity prediction.

[0228] (3) Virtual mutation analysis at key sites

[0229] After identifying key sites through attention weighting and SHAP value analysis, virtual mutation experiments were conducted to further validate the functional importance of these sites and explore optimization directions. Specifically, the original amino acids at each key site were sequentially replaced with 19 other amino acids. The predicted activity of the model was recalculated for each mutant, and the difference in activity before and after mutation was compared, as shown in the following formula:

[0230] ;

[0231] In the formula, ΔP represents the change in predicted activity before and after the mutation, used to measure the impact of a mutation at a certain amino acid site on the model's prediction results; P wildtype P represents the predicted sleep-inducing activity probability of the original sequence (wild type). mutantThis represents the predicted activity probability of the mutated sequence. If ΔP > 0.1, the mutation is likely to increase activity (beneficial mutation); if ΔP < -0.1, the mutation will significantly decrease activity (detrimental mutation), and the original site is crucial to activity.

[0232] Example 1

[0233] A Smart Screening Method for Sleep-Aid Active Peptides Based on Multimodal Fusion Networks

[0234] 1. Dataset Construction and Processing

[0235] In this embodiment, the positive samples of the dataset were derived from peptides labeled with sedative or anxiolytic functions in the BIOPEP-UWM database and peptides with GABAA receptor agonist activity reported in PubMed literature. Peptides with clear experimental validation data were retained after manual screening. Negative samples were randomly selected from the UniProt database, containing peptide sequences whose length distribution matched that of the positive samples, and other functional peptides confirmed in the literature to lack central nervous system regulatory activity were also included. Redundancy was removed from all samples using CD-HIT software, with a sequence similarity threshold set to 40%. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio, using a stratified sampling strategy to ensure a consistent ratio of positive to negative samples in each subset.

[0236] 2. Multidimensional feature extraction

[0237] This embodiment uses a tetrapeptide sequence “YWLH” as an example to illustrate the multidimensional feature extraction process.

[0238] Sequence feature extraction includes: statistically analyzing the frequency of occurrence of 20 amino acids to obtain amino acid composition features (20-dimensional vector); statistically analyzing the frequency of occurrence of 400 dipeptide combinations to obtain dipeptide composition features (400-dimensional vector); using the Biopython library to calculate the average hydrophobicity index (-0.475), isoelectric point (7.02), net charge (+0.12), molecular weight (601.68 Da), aliphatic index (32.5), and instability index (28.73, <40 indicates stability) to obtain physicochemical properties (6-dimensional vector); and inputting the peptide amino acid sequence into a pre-trained ESM-2 model to obtain a 1280-dimensional sequence embedding vector.

[0239] The three-dimensional structural feature extraction includes: using ESMFold to predict the three-dimensional atomic coordinates of the peptide, with a pLDDT confidence level of 72.6; then, based on the three-dimensional atomic coordinates, using the DSSP algorithm to analyze the proportion of secondary structures (3-dimensional vector) (α-helix proportion of 0%, β-sheet proportion of 25%, random coil proportion of 75%); based on the three-dimensional atomic coordinates, constructing a residue contact graph with Cα atoms as nodes and residue pairs with a spatial distance of less than 10 Å as edges, and inputting it into a 4-layer E(3)-isovariant graph neural network to obtain a 256-dimensional structural representation vector.

[0240] Receptor binding feature extraction includes: using AutoDock Vina to perform molecular docking of peptides with GABAA and 5-HT1A receptors to obtain the optimal binding conformation, and collecting the free binding energy (1-dimensional vector) of the optimal binding conformation; voxelizing the binding interface of the optimal binding conformation (free binding energy of -7.82 kcal / mol with GABAA receptor α1 subunit and -6.15 kcal / mol with 5-HT1A receptor); and inputting the data into a three-dimensional convolutional neural network to obtain a 128-dimensional binding characterization vector.

[0241] 3. Fusion of multidimensional features

[0242] In step 2, features from different sources are first normalized and linearly mapped to maintain consistency in numerical scale and dimensionality. These features are then concatenated along their respective feature dimensions to construct a fused feature vector. Finally, this fused feature vector serves as the unified input to the model for multi-task joint prediction, thereby achieving comprehensive modeling of peptide bioactivity and druggability.

[0243] 4. Model Training and Evaluation

[0244] The fused features are input into the multimodal fusion network model for training. The cross-attention module is configured with 8 attention heads and a projection dimension of 256. Training uses the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 The weight decay coefficient is 0.01, and the learning rate is scheduled using a cosine annealing strategy. The batch size is set to 32, the total number of training epochs is 100, and the early stopping patience value is set to 10. The training convergence of this embodiment is as follows:

[0245] The validation set loss reached its minimum value (0.1847) at the 67th epoch.

[0246] The early stop mechanism is triggered in the 77th epoch;

[0247] Final training set loss: 0.0923;

[0248] Final loss on the validation set: 0.1862;

[0249] Training time per epoch: approximately 2.8 minutes (batch size 32, GPU: NVIDIA A100 40GB);

[0250] The learning rate changes as follows:

[0251] Initial learning rate: 1×10 -4 ;

[0252] Learning rate at the 50th epoch: 3.2 × 10 -5 ;

[0253] Early stopping learning rate: 5.7 × 10 -5 (Second cycle of cosine annealing);

[0254] The evaluation was conducted on the test set, and metrics such as accuracy, precision, recall, F1 score, and AUC-ROC were recorded. The results were compared with baseline models using only sequence features, models with different feature combinations, and traditional QSAR models to verify the effectiveness of the multimodal fusion strategy.

[0255] The evaluation results of different feature combination models are as follows:

[0256]

[0257] The evaluation results of the multimodal fusion network model of this invention and the traditional QSAR model are as follows:

[0258]

[0259] As shown in the table above, the multimodal fusion network model of this invention outperforms the sequence feature-only model and the traditional QSAR model in terms of accuracy, precision, recall, F1 score, AUC-ROC, and AUC-PR.

[0260] The five-fold cross-validation results of the multimodal fusion network model of the present invention are as follows: the mean AUC-ROC is 0.943±0.018 and the mean F1 score is 0.869±0.024, which further proves that the multimodal fusion network model of the present invention has good generalization ability and accuracy in predicting active peptides.

[0261] 5. Interpretability Analysis

[0262] Interpretability analysis was performed on peptides with high prediction scores. Attention weights were used to visualize and identify amino acid positions of high model interest; SHAP values ​​were used to quantify the contribution of each input feature to the prediction results; virtual mutation scans were performed on key sites, and the original amino acids were sequentially replaced with 19 other amino acids before re-predicting activity. Activity change trends were analyzed to identify beneficial and harmful mutations, providing guidance for subsequent peptide optimization.

[0263] Example 2

[0264] High-throughput screening application of sleep-aiding active peptides

[0265] Using tortoise shell protein as the raw material, candidate peptide sequences were generated by mimicking the cleavage sites of pepsin and trypsin. After length filtering, all candidate peptides underwent feature extraction and activity prediction. Candidate peptides were ranked according to their prediction scores, and peptides with higher scores were selected as priority for validation. The results of interpretability analysis were used to assess whether their structural characteristics conformed to the general patterns of sleep-aiding active peptides, providing a basis for subsequent experimental validation.

[0266] The prediction results of the Top 5 candidate peptides in this embodiment are as follows:

[0267]

[0268] Example 3

[0269] This embodiment provides a screening system for implementing an intelligent screening method for sleep-aiding active peptides based on a multimodal fusion network. The system includes: a multidimensional feature extraction module for extracting sequence features, three-dimensional structural features, and receptor binding features of peptides; a multidimensional feature fusion module for fusing the sequence features, three-dimensional structural features, and receptor binding features of peptides; and a prediction module for training the multimodal fusion network model using the fused features, obtaining a trained multimodal fusion network model, and using the trained multimodal fusion network model to predict peptides with unknown activity, thereby screening for peptides with sleep-aiding activity.

[0270] Example 4

[0271] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor processes the computer program to implement a smart screening method for sleep-aiding active peptides based on a multimodal fusion network.

[0272] The above detailed embodiments describe the implementation of the present invention; however, the present invention is not limited to the specific details described in the above embodiments. Within the scope of the claims and technical concept of the present invention, various simple modifications and changes can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

Claims

1. A method for intelligent screening of sleep-aiding active peptides based on multimodal fusion networks, characterized in that, include: Construct an active peptide dataset including positive and negative samples; wherein the positive samples include peptides with sleep-aiding, sedative, or anti-anxiety activities, and the negative samples include peptides known not to have sleep-aiding activities; The active peptide dataset is divided into a training set, a validation set, and a test set according to a preset ratio. A multimodal fusion network model is constructed, comprising an extraction layer, a fusion layer, and a prediction layer. The extraction layer extracts multidimensional features of peptides, including sequence features, three-dimensional structural features, and receptor binding features. The fusion layer fuses these multidimensional features to obtain fused features. The prediction layer takes the fused features as input and performs a prediction task. The multimodal fusion network model is trained using a training set, and the performance of the multimodal fusion network model is monitored using a validation set to prevent overfitting. The multimodal fusion network model is validated using a test set to obtain the trained multimodal fusion network model. Input peptide sequences with unknown activity, and use a trained multimodal fusion network model to predict and screen for peptides with sleep-aiding activity.

2. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 1, characterized in that, The sequence features include amino acid composition features, dipeptide composition features, physicochemical property features, and sequence embedding vectors.

3. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 2, characterized in that, The amino acid composition characteristics were obtained by statistically analyzing the occurrence frequency of 20 amino acids; the dipeptide composition characteristics were obtained by statistically analyzing the occurrence frequency of 400 dipeptide combinations. The physicochemical properties include average hydrophobicity index, average isoelectric point, net charge, molecular weight, aliphatic index, and instability index. The sequence embedding vector is obtained by inputting the peptide sequence into a pre-trained protein language model.

4. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 1, characterized in that, The three-dimensional structural features include the proportion of secondary structures and structural representation vectors.

5. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 4, characterized in that, First, the three-dimensional atomic coordinates of the peptide are predicted using a protein structure prediction tool. Then, based on the three-dimensional atomic coordinates, the DSSP algorithm is used to analyze and statistically analyze the proportion of secondary structures of the peptide. Based on the three-dimensional atomic coordinates, a residue contact graph is constructed by using the Cα atom of the amino acid residue in the peptide as a node and the amino acid residue pairs with a spatial distance less than the distance threshold as edges. Then, a 4-layer E(3)-equivariant graph neural network is used to encode and obtain the structure representation vector.

6. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 1, characterized in that, The steps for extracting receptor binding features include: molecularly docking the peptide with GABAA and 5-HT1A receptors to obtain the optimal binding conformation and binding free energy; voxelizing the binding interface of the optimal binding conformation; and inputting the data into a three-dimensional convolutional neural network to obtain the binding characterization vector.

7. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 1, characterized in that, The prediction tasks include prediction of sleep-aiding activity, prediction of GABA pathway regulation, prediction of 5-HT pathway regulation, and prediction of drug-likeness.

8. The intelligent screening method for sleep-aiding active peptides based on multimodal fusion networks according to claim 7, characterized in that, When training a multimodal fusion network model, the loss function is calculated using the following formula: ； In the formula, L total λ1 represents the total loss function; λ2, λ3, and λ4 are weighting coefficients with values ​​of λ1 = 1.0, λ2 = 0.5, λ3 = 0.5, and λ4 = 0.3; L sleep L GABA L 5HT L drug The loss functions for predicting sleep-aiding activity, GABA pathway regulation, 5-HT pathway regulation, and drug-likeness are respectively; among them, the loss function for predicting sleep-aiding activity L... sleep Focal Loss, GABA, and 5-HT pathways are used to predict L GABA L 5HT Using the standard binary cross-entropy loss function, drugability prediction L drug The mean squared error loss function is used. Iterative optimization was performed using the AdamW optimizer; Cosine annealing is used for learning rate scheduling.

9. A smart screening system for sleep-aiding active peptides for implementing the method of any one of claims 1-8, characterized in that, include: The multidimensional feature extraction module is used to extract sequence features, three-dimensional structural features, and receptor binding features of peptides; The multidimensional feature fusion module is used to fuse the sequence features, three-dimensional structural features, and receptor binding features of peptides. The prediction module is used to train a multimodal fusion network model using fusion features, and then uses the trained multimodal fusion network model to predict peptides with unknown activity, thereby screening out peptides with sleep-aiding activity.

10. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor implements the method according to claims 1-8 when processing the computer program.

Images