A method for constructing a model for recognizing mussel adhesive proteins

Abstract

The application discloses a method for constructing a model for recognizing mussel adhesive proteins, and belongs to the technical field of bioinformatics, and specifically comprises the following steps: inputting an amino acid sequence of a target protein, obtaining standard electromagnetic response waveforms of each amino acid residue in a variety of non-aqueous solvent environments, and sorting to form an initial electromagnetic fingerprint; polypeptide fragment fingerprints are generated through sliding window cutting, shared frequency components of each fragment waveform are extracted, and characteristic sequences are formed by sorting according to intensity; local patterns and context dependencies of the sequences are extracted through an encoder network, and a context vector is output; a decoder network maps the vector into a fixed-dimension embedding representation; known protein sequences with or without adhesive functions are used as training samples, the above-mentioned processing is performed on each sample to obtain a set of embedding representations, a binary classification model is trained, and a recognition model is obtained. The application realizes function recognition by using electromagnetic response characteristics of proteins in non-aqueous environments.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a method for constructing a mussel adhesive protein recognition model. Background Technology

[0002] Mussel adhesive proteins, as biomacromolecules with excellent wet surface adhesion properties, have shown significant application potential in fields such as biomedical adhesives, biomimetic materials, and drug delivery systems. Accurate and efficient identification and characterization of proteins with these special functions are prerequisites for functional research, rational design, and engineering applications. Currently, with the rapid growth of protein sequence data, developing computational model-based identification methods has become a key technological direction in this field.

[0003] In existing technologies, protein function identification mainly relies on sequence homology analysis, conserved domain matching, or predictions based on machine learning models. Some methods attempt to incorporate physicochemical characteristics of proteins, such as hydrophobicity, charge distribution, or secondary structure predisposition, as input features to improve model performance. In recent years, research has also begun to focus on characterizing protein behavior under specific environments, such as obtaining solvent-accessible surface area or hydrogen bond network information through molecular dynamics simulations. These methods are generally based on modeling the classical physicochemical parameters or static structural features of proteins in aqueous solutions.

[0004] However, for mussel adhesive proteins, which rely on specific interfacial adhesion functions, existing identification methods mainly depend on amino acid sequence homology comparison or matching with known structural templates. These methods are usually based on physicochemical parameters or evolutionary conservation in aqueous environments, without considering the intrinsic electromagnetic response characteristics of proteins in non-aqueous interfaces or complex solvent environments. Since the functional realization of mussel adhesive proteins is closely related to their conformational dynamics and electron distribution in non-aqueous environments, identification based solely on traditional sequence or structural similarity is insufficient to fully capture their key properties at the real adhesion interface. This results in a limited identification range and a high functional false negative rate for potential adhesion proteins, especially novel protein sequences lacking significant homology or well-defined domains. Therefore, there is an urgent need to develop a novel identification method that can directly characterize and utilize the intrinsic electromagnetic response characteristics of proteins in non-aqueous environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a mussel adhesive protein recognition model, thereby solving the following technical problems:

[0006] Existing methods rely on sequence homology or structural template matching, primarily based on physicochemical parameters in aqueous solutions, failing to consider the inherent electromagnetic response characteristics of proteins at non-aqueous interfaces. Since the function of mussel adhesive proteins is directly related to their conformation and electron distribution in non-aqueous environments, traditional methods struggle to effectively identify novel potential adhesion proteins lacking homology, leading to limited identification scope and missed detections.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A method for constructing a mussel adhesive protein recognition model includes the following steps:

[0009] S1. Input the amino acid sequence of the target mussel adhesive protein, obtain the standard electromagnetic response waveform of each amino acid residue in the sequence under various non-aqueous solvent environments, arrange the waveforms in order to form the initial electromagnetic fingerprint.

[0010] S2. Use a sliding window to cut the initial electromagnetic fingerprint to generate multiple consecutive overlapping polypeptide fragment electromagnetic fingerprints. Extract the shared frequency components of all waveforms in each fragment, arrange the shared frequency components according to intensity, and form the electromagnetic feature sequence of the polypeptide fragment.

[0011] S3. Establish a neural network structure that includes an encoder network and a decoder network. The encoder network consists of convolutional layers and gated recurrent unit layers connected in series, and the decoder network consists of fully connected layers.

[0012] S4. Input the electromagnetic feature sequence of the polypeptide fragment into the encoder network, the convolutional layer extracts the local frequency patterns of the sequence, the gated recurrent unit layer captures the dependencies between the patterns, and outputs the context vector.

[0013] S5. The decoder network receives the context vector, and the fully connected layer maps the context vector into a dense vector of fixed dimensions. This dense vector is defined as the electromagnetic fingerprint-driven embedding representation of the polypeptide fragment.

[0014] S6. Collect known protein sequences with or without adhesion function as training samples. Perform steps S1 to S5 for each protein sequence to obtain the embedding representation set of all polypeptide fragments. Use the set to train a binary classification model, which is the mussel adhesive protein recognition model.

[0015] As a further aspect of the present invention: in step S1, the process of obtaining the standard electromagnetic response waveform of amino acid residues is as follows:

[0016] A solvation effect field calculation model is constructed based on the polarity, viscosity and dielectric constant parameters of each non-aqueous solvent. The three-dimensional structural conformation of amino acid residues is embedded in the model and the characteristic electron cloud distribution under the solvation effect field is calculated.

[0017] Based on the characteristic electron cloud distribution, the resonant absorption intensity at multiple discrete excitation frequency points is calculated to form the discrete spectrum of the amino acid residue in the corresponding solvent environment; the frequency points with intensity exceeding the background noise threshold in the discrete spectrum and their intensity values ​​are extracted and paired to form a standard electromagnetic response waveform;

[0018] The above process is performed for all amino acid residue types and predefined non-aqueous solvent environments, and the generated waveforms are stored in a structured manner to build a computational chemistry feature database. For each amino acid residue in the target sequence and the required solvent environment, the corresponding standard waveform is retrieved from the database.

[0019] As a further aspect of the present invention: in S2, the process of constituting the electromagnetic feature sequence of the polypeptide fragment is as follows:

[0020] Starting from the initial position of the electromagnetic fingerprint with a fixed window length and sliding step size, the system slides and cuts, and each time it captures the standard waveform corresponding to the amino acid position within the window to generate a polypeptide fragment electromagnetic fingerprint; for each generated polypeptide fragment electromagnetic fingerprint, each non-aqueous solvent environment is processed in sequence.

[0021] When processing any non-aqueous solvent environment, the waveforms of all amino acids in the fragment under this solvent are aligned and superimposed, and the envelope of the superimposed waveform is calculated. This envelope waveform is used as the comprehensive response waveform of the polypeptide fragment under this non-aqueous solvent environment.

[0022] After processing in all non-aqueous solvent environments, the comprehensive response waveforms of the same polypeptide fragment under each environment are compared. Frequency components that appear continuously in each waveform are identified and marked as shared frequency components. The average amplitude intensity of each shared frequency component in the comprehensive waveform under all non-aqueous solvent environments is calculated. The components are then sorted in descending order according to the average amplitude intensity and output as the electromagnetic feature sequence of the polypeptide fragment.

[0023] As a further aspect of the present invention: the process of identifying frequency components that continuously appear in each waveform and marking them as shared frequency components is as follows:

[0024] The overall response waveforms of the peptide fragments under all non-aqueous solvent environments were discretely sampled to obtain the discrete frequency sequence corresponding to each waveform. All discrete frequency sequences were then merged to form a total frequency set.

[0025] For each unique frequency value in the total frequency set, count the number of times the frequency value appears in different non-aqueous solvent environments. If the number of times the frequency value appears is equal to the total number of non-aqueous solvent environments, then mark the frequency value as a candidate sustained frequency. Calculate the amplitude value corresponding to each candidate sustained frequency in different non-aqueous solvent environments to obtain a set of cross-environment amplitudes for the frequency.

[0026] Calculate the standard deviation of the amplitude across the environment. If the standard deviation is lower than the preset amplitude fluctuation threshold, the candidate sustained frequency is finally determined as the shared frequency component. All the finally determined shared frequency components are summarized and marked as shared frequency components used to constitute the electromagnetic feature sequence.

[0027] As a further aspect of the present invention: in step S3, the process of establishing a neural network structure including an encoder network and a decoder network is as follows:

[0028] The encoder network is sequentially connected to the first one-dimensional convolutional layer, the first nonlinear activation layer, the first pooling layer, and the gated recurrent unit layer; the decoder network is a fully connected layer network.

[0029] The first one-dimensional convolutional layer is configured with multiple convolutional kernels of different widths to extract local features of different scales from the input sequence; the first nonlinear activation layer performs an element-wise nonlinear transformation on the convolutional output; the first pooling layer downsamples the transformed features to compress the sequence length; the gated recurrent unit layer receives the downsampled feature sequence and outputs a context vector.

[0030] The fully connected layer network receives a context vector and maps the dimension of the context vector to a preset fixed dimension through at least one fully connected layer with a non-linear activation function.

[0031] As a further aspect of the present invention: in S4, the process of the encoder network outputting the context vector is as follows:

[0032] Set up a set of trainable one-dimensional convolutional kernels with different receptive field widths, input the electromagnetic feature sequence of the peptide fragment into the convolutional layer of the encoder network, and each convolutional kernel independently performs sliding dot product calculation and generates an initial feature map.

[0033] For each initial feature map, a non-linear activation function is applied to transform it. Max pooling is then performed on the transformed feature map to compress the sequence dimension and output a sequence of feature vectors.

[0034] The feature vector sequence is input into the gated recurrent unit layer, and each feature vector is processed sequentially. The update gate and reset gate are calculated based on the current input and the state of the previous time step. Historical information is modulated and combined with the current input to generate candidate states. The candidate states and historical states are fused by updating the gate to obtain the hidden state at the current time step. The hidden state at the end of the sequence processing is output as the context vector.

[0035] As a further aspect of the present invention: in step S5, the process of mapping the context vector to the electromagnetic fingerprint driving embedding representation is as follows:

[0036] The fully connected layer of the decoder network receives the context vector from the encoder network. The fully connected layer consists of multiple sequentially connected linear transformation layers, where the first linear transformation layer transforms the dimension of the context vector into an intermediate dimension, and the second and subsequent linear transformation layers transform the intermediate dimension step by step.

[0037] Nonlinear activation function layers are introduced after the first and second linear transformation layers to perform nonlinear mapping on the output of the linear transformation. The last linear transformation layer converts the dimension of the feature vector after nonlinear mapping into a pre-defined fixed dimension. The dense vector with fixed dimension output by the last linear transformation layer is defined as the electromagnetic fingerprint-driven embedding representation of the polypeptide fragment.

[0038] As a further aspect of the present invention: in step S6, the process of training the binary classification model is as follows:

[0039] For each protein sequence in the training sample set, steps S1 to S5 are applied to generate electromagnetic fingerprint-driven embedding representations of all polypeptide fragments of the protein sequence. The standard deviation of the amplitude intensity of each embedding representation vector in all non-aqueous solvent environments corresponding to the polypeptide fragment is calculated. Vectors with standard deviations lower than a preset stability threshold are selected to form a stable embedding representation subset.

[0040] The magnitude of each vector in the stable embedding representation subset is calculated, and the magnitude is used as the significance weight of the vector. The significance weight is then used to calculate a weighted average of all vectors in the stable embedding representation subset to generate a feature vector representing the overall characteristics of the protein sequence.

[0041] The overall feature vectors of all training sample protein sequences and their corresponding binary labels of adhesion function are collected to form a training dataset. A support vector machine model is trained using the training dataset. The training process includes iteratively adjusting the normal vector and intercept parameters of the classification hyperplane in the feature space. The support vector machine model with fixed parameters after iterative adjustment is saved as the final mussel adhesive protein recognition model.

[0042] The beneficial effects of this invention are:

[0043] This invention constructs an initial electromagnetic fingerprint by acquiring standard electromagnetic response waveforms of amino acid residues in various non-aqueous solvent environments, directly characterizing the intrinsic physicochemical properties of proteins at near-realistic adhesion interfaces, overcoming the limitations of existing methods that rely on aqueous environment parameters. By using sliding window segmentation and multi-environment waveform fusion to extract shared frequency components, the common electromagnetic features of peptide fragments in heterogeneous solvent environments can be stably captured, enhancing the indicative power of these features for functionally relevant conformational dynamics. Furthermore, an encoder-decoder neural network structure is used to map the electromagnetic feature sequence into a fixed-dimensional electromagnetic fingerprint-driven embedding representation, which deeply integrates local sequence patterns and their long-range contextual dependencies. Finally, a weighted aggregation of all peptide fragment embedding representations generates a global feature vector for the protein sequence, which is then used to train a classification model. This allows the model to functionally discriminate protein sequences from the perspective of electromagnetic responses in non-aqueous environments. This method achieves a more fundamental functional relevance characterization of mussel adhesive proteins, improves the ability to identify proteins lacking sequence homology but exhibiting similar interfacial behaviors, and enhances the model's sensitivity and specificity in detecting potential adhesion functions. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0047] Please see Figure 1 As shown, this invention provides a method for constructing a mussel adhesive protein recognition model, comprising the following steps:

[0048] S1. Input the amino acid sequence of the target mussel adhesive protein, obtain the standard electromagnetic response waveform of each amino acid residue in the sequence under various non-aqueous solvent environments, arrange the waveforms in order to form the initial electromagnetic fingerprint.

[0049] S2. Use a sliding window to cut the initial electromagnetic fingerprint to generate multiple consecutive overlapping polypeptide fragment electromagnetic fingerprints. Extract the shared frequency components of all waveforms in each fragment, arrange the shared frequency components according to intensity, and form the electromagnetic feature sequence of the polypeptide fragment.

[0050] S3. Establish a neural network structure that includes an encoder network and a decoder network. The encoder network consists of convolutional layers and gated recurrent unit layers connected in series, and the decoder network consists of fully connected layers.

[0051] S4. Input the electromagnetic feature sequence of the polypeptide fragment into the encoder network, the convolutional layer extracts the local frequency patterns of the sequence, the gated recurrent unit layer captures the dependencies between the patterns, and outputs the context vector.

[0052] S5. The decoder network receives the context vector, and the fully connected layer maps the context vector into a dense vector of fixed dimensions. This dense vector is defined as the electromagnetic fingerprint-driven embedding representation of the polypeptide fragment.

[0053] S6. Collect known protein sequences with or without adhesion function as training samples. Perform steps S1 to S5 for each protein sequence to obtain the embedding representation set of all polypeptide fragments. Use the set to train a binary classification model, which is the mussel adhesive protein recognition model.

[0054] In a preferred embodiment of the present invention, the process of obtaining the standard electromagnetic response waveform of amino acid residues in step S1 is as follows:

[0055] First, a specific computational model for the solvation effect field needs to be constructed for each selected non-aqueous solvent. The solvent's polarity, viscosity, and dielectric constant are three core physicochemical parameters that collectively determine the microscopic force field generated by solvent molecules on solute molecules. Taking polarity as an example, its value is typically between 0 and 1; for instance, the polarity index of ethanol is approximately 0.654. ​​When constructing the model, these parameters are quantified as a series of potential energy functions acting on grid points, collectively forming a three-dimensional solvation effect field.

[0056] Next, the three-dimensional structural conformation of the target amino acid residue is embedded into this computational model. Here, the three-dimensional structural conformation does not refer to a crystal structure, but rather to a stable conformation of the residue with lower energy in the corresponding solvent environment, obtained through conformational search or molecular mechanics optimization. The embedding process involves aligning and superimposing the atomic coordinates of the residue with the grid of the solvation effect field. Subsequently, the characteristic electron cloud distribution of this conformation under the influence of the solvation effect field is calculated. This process is achieved by solving a parameterized electron density functional, the core of which is calculating the spatial probability density of the electron cloud on each atomic orbital of the amino acid residue in the presence of solvent molecules. The final characteristic electron cloud distribution is a three-dimensional data field that characterizes the electronic distribution properties of the residue in the solvent environment, which is the physical basis for subsequent calculations of its electromagnetic response.

[0057] Based on the obtained characteristic electron cloud distribution, the resonant absorption intensity at multiple discrete excitation frequency points is calculated. The range of excitation frequency points typically covers the spectral region from far-infrared to ultraviolet, for example, from wavenumber 100 cm⁻¹. -1 Up to 50000cm -1 and at fixed intervals such as 10cm -1 Discrete sampling is performed. The calculation of the resonant absorption intensity is based on time-dependent perturbation theory, simulating the perturbation effect of the external electromagnetic field on the characteristic electron cloud distribution, thereby obtaining the intensity of resonant absorption of the electron cloud under electromagnetic wave excitation at different frequencies. After calculating the intensity values ​​at all discrete frequency points, the discrete spectrum of the amino acid residue in the corresponding non-aqueous solvent environment is formed.

[0058] Subsequently, effective feature information needs to be extracted from this discrete spectrum. To this end, a background noise threshold is set, typically determined based on the statistical standard deviation of the baseline fluctuation across the entire discrete spectrum. All frequency points whose resonant absorption intensity exceeds this threshold are extracted, and each frequency point is paired with its corresponding intensity value. Each frequency-intensity pair constitutes a basic unit, and the collection of all these basic units constitutes the standard electromagnetic response waveform of the amino acid residue in this specific solvent environment.

[0059] To construct a comprehensive computational chemistry feature database, the aforementioned complete process must be executed for all twenty common amino acid residue types and all predefined combinations of non-aqueous solvent environments. This includes, but is not limited to, various polar, nonpolar, protic, and aprotic solvents such as alcohols, ethers, and halogenated hydrocarbons. Each calculated standard electromagnetic response waveform is assigned a unique index identifier containing information about the amino acid type, solvent type, and computational model version. All these waveform data are stored and archived in a structured format, such as hierarchical key-value pairs or relational tables, forming a computational chemistry feature database that can be efficiently queried.

[0060] In practical applications, after inputting the amino acid sequence of the target mussel adhesive protein, the system retrieves and calls the corresponding pre-calculated standard electromagnetic response waveform from the aforementioned database for each amino acid residue in the sequence and for each preset non-aqueous solvent environment, based on their type combination, for direct use in subsequent steps. This approach avoids repeating time-consuming quantum chemical calculations for each target sequence, greatly improving the efficiency of the entire identification model construction process.

[0061] In another preferred embodiment of the present invention, the process of forming the electromagnetic feature sequence of the polypeptide fragment in S2 is as follows:

[0062] First, set two key parameters: window length and sliding step size, both expressed in units of amino acid residues. For example, the window length can be set to 20, and the sliding step size to 5. This means that each time 20 consecutive amino acid residues are extracted as an analysis fragment, and the starting position between two adjacent fragments is offset by 5 residues, thus forming a continuous peptide fragment with most overlapping regions.

[0063] The initial electromagnetic fingerprint is a set of standard electromagnetic response waveforms arranged according to the amino acid sequence. Starting from the first amino acid position of this fingerprint, the analysis window is moved along the sequence direction in increments of 5. After each move, all standard electromagnetic response waveforms corresponding to the 20 amino acid positions covered by the current window are extracted. These waveforms are arranged in order within the window, collectively forming a "peptide fragment electromagnetic fingerprint". For a protein with a length of 100 amino acids, the above parameters will generate 17 such fragment fingerprints.

[0064] Next, for each generated polypeptide fragment electromagnetic fingerprint, it is necessary to process each preset non-aqueous solvent environment sequentially. Assume three preset non-aqueous solvent environments are solvent A, solvent B, and solvent C. When processing the current fragment in solvent A, it is necessary to extract the standard electromagnetic response waveforms corresponding to all 20 amino acid residues in the fragment fingerprint under solvent A conditions. These 20 waveforms need to be aligned and superimposed. Alignment refers to calibrating all waveforms on the frequency axis to ensure that the same frequency value is located at the same coordinate position. Superimposition involves arithmetically adding the amplitude values ​​of the calibrated waveforms at each frequency point to generate a preliminary superimposed waveform.

[0065] However, direct superposition may amplify noise or the specific responses of individual residues. Therefore, it is necessary to calculate the "envelope" of this superimposed waveform. In this invention, the envelope refers to a smooth curve formed by connecting all local maxima in the superimposed waveform. Specifically, the method is to first identify all frequency points on the superimposed waveform whose amplitude values ​​are greater than the preceding and following points, and mark these points as local maxima. Then, cubic spline interpolation is used to connect these discrete maxima to generate a smooth curve covering the entire frequency range. This curve is the "overall response waveform" of the polypeptide fragment in solvent A. It reflects the overall, smooth electromagnetic response trend of the fragment in solvent A, weakening the detailed fluctuations of individual amino acids. For the same sample fragment, the above process is repeated to generate its overall response waveforms in solvents B and C, respectively.

[0066] After obtaining the composite response waveforms of the same peptide fragment under all non-aqueous solvent conditions, the next step is to identify their common characteristics, i.e., "shared frequency components." First, each composite response waveform is discretely sampled. Since these waveforms are essentially continuous envelopes, they need to be converted into discrete frequency-amplitude pairs for computation. The sampling interval can be set to, for example, 5 wavenumbers, reading amplitude values ​​at equal intervals within a predefined frequency range, thus generating a discrete frequency sequence for each waveform. Each sequence contains hundreds of frequency points and their corresponding amplitude values.

[0067] Subsequently, the discrete frequency sequences of the current peptide fragment generated under all solvent conditions are merged to form a total frequency set. This set contains the frequency values ​​that have appeared in all solvents. For each unique frequency value in this total set, the number of times it appears in the discrete sequences of different solvents is counted. Here, "appearance" means that in the discrete sequence of that solvent, there exists a frequency point whose absolute value of the difference from this unique frequency value is less than a very small tolerance, such as 0.5 wavenumbers.

[0068] If a unique frequency value is detected in all discrete sequences of the three solvents (i.e., its occurrence count equals the total number of occurrences in the non-aqueous solvent environment, which is 3), then this frequency value is labeled as a "candidate persistence frequency." For example, suppose the frequency value is 1000.5 cm⁻¹. -1 If a matching point can be found in the sequences of solvents A, B, and C, then it becomes a candidate persistence frequency.

[0069] Once a candidate frequency is identified, its amplitude stability needs to be verified. For each candidate sustained frequency, its corresponding amplitude value in a discrete sequence for each solvent environment needs to be extracted. Thus, for a candidate frequency, a set of amplitude values ​​is obtained, the number of which equals the number of solvent types. For example, for three solvents, three amplitude values ​​are obtained [AmpA, AmpB, AmpC]. This set of data is called the "cross-environment amplitude" of that frequency.

[0070] Next, the standard deviation of this set of cross-environment amplitudes is calculated to measure the variability of the frequency component's response intensity in different solvents. An "amplitude variability threshold" is set, for example, 0.1 (assuming the amplitude values ​​have been normalized). If the calculated standard deviation is below this threshold of 0.1, the frequency component is considered to exhibit a stable response intensity in different solvent environments, thus ultimately being identified as a "shared frequency component." If the standard deviation is higher than or equal to the threshold, it is discarded.

[0071] After iterating through all candidate sustained frequencies and completing the stability screening described above, all remaining shared frequency components are summarized. Finally, for these shared frequency components, the average amplitude values ​​corresponding to them under all solvent conditions are calculated to obtain the "average amplitude intensity" of each component. These shared frequency components are then sorted in descending order of average amplitude intensity. The resulting ordered list of frequency components constitutes the "electromagnetic signature sequence" of the polypeptide fragment. This sequence focuses on the stable electromagnetic response characteristics exhibited by the protein fragment under various non-aqueous environments, providing core input for subsequent neural network analysis.

[0072] In another preferred embodiment of the present invention, the process of establishing a neural network structure including an encoder network and a decoder network in step S3 is as follows:

[0073] The encoder network is designed as a sequentially connected four-layer data processing architecture. The first layer is a first one-dimensional convolutional layer. The input to this layer is the electromagnetic feature sequence of a polypeptide fragment, which is a one-dimensional vector of length L, where the value at each position represents the average amplitude intensity of a shared frequency component ordered by intensity. The specific value of L might be, for example, 150. The first one-dimensional convolutional layer contains multiple trainable convolutional kernels with different widths. For example, three convolutional kernels with widths of 5, 7, and 9 can be used simultaneously. Here, "width" refers to the number of consecutive positions covered by the convolutional kernel in a single computation of the input sequence. Each convolutional kernel contains a set of trainable parameters, namely weights and biases. When processing the input sequence, a convolutional kernel with a width of 5 sequentially covers positions 1 to 5, positions 2 to 6, and so on, until the end of the sequence. At each covered position, the convolutional kernel multiplies its five internal weight parameters by the corresponding five sequence values, then sums all the products and adds the bias value to output a final computational result. After the convolutional kernel slides across the entire sequence, it generates a new numerical sequence, which is called the "initial feature map" corresponding to that kernel. Because three convolutional kernels of different widths are used, three initial feature maps of similar length but different content are generated in parallel. This design allows the network to simultaneously extract local feature patterns at different scales from the input sequence. Narrower kernels focus on fine-grained relationships between neighboring frequency components, while wider kernels capture overall trend changes over a wider frequency range.

[0074] Next, these three initial feature maps are fed into the first nonlinear activation layer. This layer independently performs a nonlinear function transformation on each value in each map. A typical and widely used function is the modified linear unit function (MRJ). The function operates on the rule that for any input value x, if x is greater than 0, output x; if x is less than or equal to 0, output 0. This operation is applied element-wise to the entire map. After this transformation, all values ​​less than or equal to 0 in the map are set to zero, while positive values ​​are retained. This introduces a crucial nonlinear element into the entire model, enabling it to learn and represent complex, nonlinear interactions between input features, rather than just linear combinations.

[0075] Subsequently, the feature maps after non-linear activation enter the first pooling layer. This embodiment uses max pooling to achieve downsampling. The pooling layer defines a fixed pooling window, for example, setting the window length to 2 and the sliding step size to 2. This window starts from the beginning position of each feature map, taking out two values ​​within the window each time, comparing their size, and outputting only the larger value to the next layer. Then the window moves forward two positions, repeating this comparison and selection process. For a map of length L, after this operation, its length is compressed to approximately L / 2. The core functions of max pooling are threefold: First, it reduces the number of parameters and computational cost required for subsequent calculations by reducing the sequence length; second, it selects the most representative features in local regions, enhancing the model's robustness to small changes in feature position; third, it helps to gradually build more abstract feature representations in subsequent layers. After pooling, the original three independent feature maps are compressed in length, and they can be concatenated along the feature channel dimension to form an integrated "feature vector sequence" that is shorter in length but has more feature channels.

[0076] This sequence of feature vectors is ultimately fed into the last layer of the encoder network, the gated recurrent unit (ROU) layer. A Gated Recurrent Unit (GRU) is a type of recurrent neural network specifically designed to process sequential data and model long-range dependencies between sequence elements. This layer sequentially processes the input feature vector sequence. Internally, it has a continuously updated "hidden state vector" as its memory unit. When processing the t-th feature vector in the sequence, the GRU performs a series of internal calculations. First, it combines the current input feature vector with the hidden state vector from the previous time step, and calculates the "update gate vector" and "reset gate vector" using two different trainable weight matrices and the Sigmoid activation function. The Sigmoid function maps the linearly transformed value to between 0 and 1, making the gate vector a soft switch that controls the rate of information flow.

[0077] The reset gating vector determines how much of the previous hidden state should be ignored when generating new candidate information. Specifically, the reset gating vector is element-wise multiplied with the previous hidden state vector. This filtered historical information is then concatenated with the current input feature vector and passed through another trainable weight matrix and a hyperbolic tangent activation function to generate a "candidate hidden state vector." The candidate state contains new information adjusted for the current input.

[0078] The updated gating vector determines how the final hidden state at the current time step should be composed of the historical state and the new candidate state. The specific calculation is as follows: the updated gating vector is multiplied element-wise by the candidate hidden state vector to obtain the contribution of new information; simultaneously, a vector is calculated by subtracting each element of the updated gating vector from the value 1, and this vector is multiplied element-wise by the hidden state vector from the previous time step to obtain the retained historical information; finally, these two parts are added together, and the result is the new "hidden state vector" at the current time step. This state conveys relevant historical memories and integrates new input content. After the gated recurrent unit layer processes the last element of the input feature vector sequence, the hidden state vector it holds at this point encodes all the context information of the entire input sequence from beginning to end. This final hidden state vector is defined as the "context vector," serving as a fixed-dimensional summary output of the encoder network for the input electromagnetic feature sequence.

[0079] The decoder network is constructed as a fully connected layer network whose sole task is to receive the context vector generated by the encoder and map it to a uniformly formatted low-dimensional representation. A fully connected layer means that each neuron in the network is connected to all outputs of the previous layer. A typical decoder network consists of two fully connected layers sequentially. The first fully connected layer receives the context vector, for example, a 128-dimensional vector, and linearly transforms it to an intermediate dimension, such as 64, using a weight matrix. This layer is typically followed by a ReLU non-linear activation function layer. The second fully connected layer then linearly transforms this 64-dimensional vector to a final, fixed target dimension, such as 32. This final output, a dense vector with a fixed dimension of 32, is explicitly defined as the "electromagnetic fingerprint-driven embedding representation" of the peptide fragment, which will serve as the core input for subsequent protein-level feature aggregation.

[0080] In another preferred embodiment of the present invention, the process of the encoder network outputting the context vector in S4 is as follows:

[0081] The entire process begins with the input of the electromagnetic feature sequence and convolutional feature extraction. First, a set of trainable one-dimensional convolutional kernels needs to be set up, each with a different receptive field width, for example, 3, 5, and 7. The receptive field width defines the local range of the input sequence that each convolutional kernel can observe at one time. The electromagnetic feature sequence of the peptide fragment, assumed to be a one-dimensional array of length 100, is input into the convolutional layer of the encoder network. Each value in the sequence represents the intensity of a shared frequency component. For a convolutional kernel with a width of 5, it starts from the beginning of the sequence, takes the first 5 values, multiplies them by the 5 trainable weights within the kernel, sums them, and adds a trainable bias parameter to obtain the first output value. Then, the convolutional kernel slides one step to the right, takes the 2nd to 6th values, and repeats the same multiplication and addition operation to obtain the second output value. This continues until the end of the sequence is reached. This convolutional kernel ultimately generates a new sequence of length 96, i.e., the "initial feature map". Because three convolutional kernels of different widths are used simultaneously, three initial feature maps of slightly different lengths are generated in parallel, which capture local intensity variation patterns at different scales in the input sequence.

[0082] The next step is a nonlinear transformation of the initial feature maps. For each initial feature map generated in the previous step, a nonlinear activation function is independently applied for element-wise transformation. A commonly used function is the modified linear unit (MRU). Specifically, each value in the map is iterated over; if the value is greater than 0, it remains unchanged; if the value is less than or equal to 0, it is set to 0. This operation is performed element-wise on each map. After this step, all negative or zero values ​​in the map are eliminated, positive features are preserved and may be enhanced, thus introducing a key component of the nonlinear decision boundary into the model.

[0083] Then, max pooling is performed to compress the data dimensionality. For each feature map after the nonlinear transformation, a max pooling operation is applied. A pooling window is defined in the operation, for example, with a window size of 2 and a stride of 2. This means that the window covers two consecutive values ​​in the map each time. At each window position, the two values ​​are compared, and only the larger value is passed to the next layer. The window then moves forward two positions, repeating this comparison and selection process. For a map of length 96, this operation reduces its length to 48. The purpose of max pooling is to downsample the features of local regions, preserving the most salient feature signals while reducing the spatial size of the data and the computational complexity of subsequent layers. After pooling, the three originally independent feature maps are compressed, and they can be stacked along a new dimension to form a three-dimensional tensor structure, but its sequence length dimension has been uniformly shortened. This structure can be reshaped and passed to the next layer as a more compact "sequence of feature vectors".

[0084] Finally, the feature vector sequence is processed by a gated recurrent unit (ROU) layer to generate a context vector. The feature vector sequence obtained in the previous step is input sequentially into the GU layer. This layer maintains an internal hidden state. When processing the t-th feature vector in the sequence, it performs the following calculations: First, the current input feature vector is concatenated with the hidden state vector from the previous time step, and a "reset gate vector" is calculated using a weight matrix and a sigmoid function. Each element of the reset gate vector has a value between 0 and 1. Next, the reset gate vector is multiplied element-wise with the hidden state from the previous time step; this step determines how much past information is forgotten. This result is then concatenated again with the current input feature vector, and a "candidate hidden state vector" is calculated using another weight matrix and a hyperbolic tangent activation function. This candidate hidden state vector represents a new state proposal generated based on the current input and filtered historical information.

[0085] Simultaneously, using the current input and the previous hidden state, an "update gate vector" is calculated through a separate weight matrix and a sigmoid function. The value of the update gate vector, also between 0 and 1, controls the fusion ratio of the old and new states. Finally, the current hidden state vector is calculated using the following formula: the new hidden state equals the update gate vector multiplied by the candidate hidden state vector, plus a difference vector (the difference between the update gate vector and the previous hidden state vector). This calculation is element-wise. When the update gate value is close to 1, the new state relies more on the candidate state; when it is close to 0, it retains more of the old state. The GRU layer processes all vectors in the feature vector sequence in this manner. When the last vector is processed, the corresponding hidden state vector integrates all the information from the entire input sequence. This final hidden state vector is output and defined as the context vector corresponding to the electromagnetic feature sequence of the peptide fragment.

[0086] In another preferred embodiment of the present invention, in step S5, the process of mapping the context vector to the electromagnetic fingerprint driving embedding representation is as follows:

[0087] First, the fully connected layers of the decoder network receive the context vector from the encoder network. Assume the encoder outputs a 128-dimensional real-valued vector. The decoder network itself is constructed as a module consisting of multiple sequentially connected fully connected layers. A fully connected layer, also known as a dense layer, is characterized by each neuron in that layer being connected to all output neurons of the previous layer.

[0088] The first layer of the decoder is the first linear transformation layer. This layer is defined by a trainable weight matrix and a bias vector. The weight matrix is ​​designed to be 128 rows by 64 columns. This means that when a 128-dimensional context vector is input, this layer performs the operation of multiplying the input vector by this 128x64 weight matrix and then adding a 64-dimensional bias vector. The mathematical result of this linear transformation is a new 64-dimensional vector. This step achieves the transformation from the original context dimension (128 dimensions) to an intermediate dimension (64 dimensions). The intermediate dimension is typically chosen to be smaller than the input dimension, aiming to perform lossy compression, retain the most essential information, and begin to build a more compact representation.

[0089] Immediately following this first linear transformation layer, a non-linear activation function layer is introduced. Typically, a modified linear unit (MLU) is used as the activation function. This function processes the 64-dimensional vector obtained in the previous step element-wise: for each value in the vector, if the value is greater than 0, the value itself is output; if the value is less than or equal to 0, 0 is output. This operation is not a simple linear scaling, but rather introduces non-linear mapping capabilities. It enables the network to learn and express complex, non-linear combinations of input features, which is crucial for the model's strong fitting ability; otherwise, stacking multiple linear transformations is equivalent to a single linear transformation.

[0090] The 64-dimensional vector, after non-linear activation, is then fed into a second linear transformation layer. This layer also consists of a trainable weight matrix and a bias vector. For example, the weight matrix could be designed to be 64 rows by 32 columns. This layer performs another linear transformation: multiplying the 64-dimensional input vector by the 64x32 weight matrix and then adding a 32-dimensional bias vector, resulting in a 32-dimensional output vector.

[0091] Similarly, after the second linear transformation layer, another nonlinear activation function layer (e.g., another ReLU layer) is introduced to perform an element-wise nonlinear transformation on the 32-dimensional vector.

[0092] At this point, the network has completed two cycles of "linear transformation followed by non-linear activation." In some designs, more such layer pairs may be added to progressively extract more abstract features. However, an output layer is ultimately needed to map the dimensions to a predefined fixed target. Therefore, a final linear transformation layer is added at the end of the network. The dimensions of the weight matrix for this layer need to be set based on its input dimensions and the final target dimension. Assuming the second non-linear activation layer outputs a 32-dimensional vector, and the predefined fixed target dimension is 16 dimensions, then the weight matrix of the final linear transformation layer will be 32 rows by 16 columns, with a 16-dimensional bias vector. This layer performs the final linear projection, transforming the 32-dimensional vector into a 16-dimensional vector. After this layer, non-linear activation functions are typically not used again, or only linear activation is used to output a continuous, unrestricted, dense vector.

[0093] The final output, a real vector with a fixed dimension of 16, is the product of the entire decoding process. It is formally defined as the "electromagnetic fingerprint-driven embedding representation" of the peptide fragment. This name emphasizes its origin: it is generated by a feature sequence characterizing the electromagnetic response fingerprint of the peptide fragment in various non-aqueous solvents; meanwhile, the term "embedding representation" in machine learning specifically refers to mapping high-dimensional or symbolic data to a point in a low-dimensional continuous vector space. This 16-dimensional vector, as a standardized mathematical object, fully represents the comprehensive electromagnetic properties of the peptide fragment and is ready for use in subsequent protein-level polymerase chain analysis.

[0094] In another preferred embodiment of the present invention, the process of training the binary classification model in step S6 is as follows:

[0095] For each protein sequence in the training sample set, steps S1 to S5 are applied independently and completely. Assuming a protein sequence generates N polypeptide fragments through the sliding window processing in S2, then S5 will correspondingly generate N electromagnetic fingerprint-driven embedding representations, each a 16-dimensional vector. These N vectors constitute the "embedding representation set" for the protein sequence.

[0096] Next, a crucial stability screening step is implemented. It's important to recognize that the embedding representation of each peptide fragment originates from its electromagnetic response in various non-aqueous solvent environments. To ensure the reliability and consistency of the information used to construct the overall protein signature, the "generation quality" of each embedding representation needs to be evaluated. The evaluation method involves retrospectively analyzing all non-aqueous solvent environments (e.g., three solvents) involved in generating the peptide fragment for this embedding representation in step S2, and extracting the amplitude intensities of the shared frequency components used to construct the final electromagnetic signature sequence from the composite response waveform of the fragment in each solvent. The standard deviation of this set of cross-solvent amplitude intensities is then calculated. The standard deviation is a statistical measure of the dispersion of a set of data; a smaller value indicates that the data points are closer to the mean, meaning a more consistent response across different solvent environments.

[0097] To this end, a "stability threshold" is preset, for example, 0.08. Then, the N 16-dimensional vectors in the protein sequence embedding representation set are traversed. For each vector, its corresponding cross-solvent amplitude intensity standard deviation is calculated. If the standard deviation of a vector is higher than or equal to 0.08, the electromagnetic response of the polypeptide fragment is considered to fluctuate greatly in different solvents, and its embedding representation reliability is low, so it is excluded. If the standard deviation is lower than 0.08, the vector is retained. After this screening, from the original N vectors, only M vectors may be retained, forming a "stable embedding representation subset". This step aims to filter out noisy fragment representations caused by abnormal responses in specific solvents, thereby improving the signal-to-noise ratio of the overall features constructed subsequently.

[0098] Subsequently, feature vectors representing the entire protein sequence are constructed based on this stable subset. First, the magnitude (i.e., Euclidean norm) of each of the M 16-dimensional vectors in the subset is calculated. For a 16-dimensional vector, its magnitude is the square root of the sum of the squares of all its dimensions. This magnitude is interpreted as the "salience weight" of the polypeptide fragment represented by the vector within the context of the entire protein sequence. A larger magnitude means that the vector has higher "intensity" or "salience" in the representation space, potentially corresponding to a more prominent and important functional segment of the protein.

[0099] Next, a weighted average is calculated. Instead of a simple arithmetic average of the M vectors, they are weighted using their significance weights. Specifically, each 16-dimensional vector is multiplied by its corresponding magnitude (scalar weight), resulting in M ​​weighted vectors. These M weighted vectors are then summed dimensionally to obtain a cumulative vector. Finally, this cumulative vector is divided by the sum of all weights (i.e., the M magnitudes). The result is a new 16-dimensional vector, which is the "feature vector representing the overall characteristics of the protein sequence." This feature vector not only summarizes information from all stable fragments but also gives fragments that contribute more weight to the overall characteristics a higher proportion.

[0100] The above process is repeated for each protein sequence in the training sample set, thereby generating a unique 16-dimensional overall feature vector for each sample protein. At the same time, each sample protein has a "binary label of adhesion function" that has been experimentally verified beforehand, for example, "1" is marked for those with adhesion function and "0" is marked for those without adhesion function.

[0101] The overall feature vectors of all training samples and their corresponding binary labels are combined to form the "training dataset" for training supervised machine learning models.

[0102] Finally, a Support Vector Machine (SVM) model is trained using this training dataset. SVM is a classic binary classification algorithm whose core idea is to find an optimal linear hyperplane in a high-dimensional feature space that maximizes the margin between the two classes. In this embodiment, the feature space is 16-dimensional. The training process is an iterative optimization process aimed at adjusting two key parameters of this hyperplane: the "normal vector" and the "intercept." The normal vector determines the orientation of the hyperplane, and the intercept determines its position. By solving a convex optimization problem, the normal vector and intercept that minimize the classification error and maximize the class margin are found.

[0103] After the training process is completed and the optimal parameters are determined, these parameters are fixed and saved along with the computational logic of the support vector machine model. This saved model with fixed parameters is the final "mussel adhesive protein recognition model". For any new protein sequence with unknown function, simply apply steps S1 to S5 and generate its 16-dimensional overall feature vector through the stability screening and weighted averaging process described above. Then, input it into this recognition model, and the model will output a predicted category label or a probability score belonging to a certain category, thereby achieving automatic recognition of the protein's adhesion function.

[0104] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for constructing a mussel adhesive protein recognition model, characterized in that, Includes the following steps: S1. Input the amino acid sequence of the target mussel adhesive protein, obtain the standard electromagnetic response waveform of each amino acid residue in the sequence under various non-aqueous solvent environments, arrange the waveforms in order to form the initial electromagnetic fingerprint. S2. Use a sliding window to cut the initial electromagnetic fingerprint to generate multiple consecutive overlapping polypeptide fragment electromagnetic fingerprints. Extract the shared frequency components of all waveforms in each fragment, arrange the shared frequency components according to intensity, and form the electromagnetic feature sequence of the polypeptide fragment. S3. Establish a neural network structure that includes an encoder network and a decoder network. The encoder network consists of convolutional layers and gated recurrent unit layers connected in series, and the decoder network consists of fully connected layers. S4. Input the electromagnetic feature sequence of the polypeptide fragment into the encoder network, the convolutional layer extracts the local frequency patterns of the sequence, the gated recurrent unit layer captures the dependencies between the patterns, and outputs the context vector. S5. The decoder network receives the context vector, and the fully connected layer maps the context vector into a dense vector of fixed dimensions. This dense vector is defined as the electromagnetic fingerprint-driven embedding representation of the polypeptide fragment. S6. Collect known protein sequences with or without adhesion function as training samples. Perform steps S1 to S5 for each protein sequence to obtain the embedding representation set of all polypeptide fragments. Use the set to train a binary classification model, which is the mussel adhesive protein recognition model. In step S1, the process of obtaining the standard electromagnetic response waveform of amino acid residues is as follows: A solvation effect field calculation model is constructed based on the polarity, viscosity and dielectric constant parameters of each non-aqueous solvent. The three-dimensional structural conformation of amino acid residues is embedded in the model and the characteristic electron cloud distribution under the solvation effect field is calculated. Based on the characteristic electron cloud distribution, the resonant absorption intensity at multiple discrete excitation frequency points is calculated to form the discrete spectrum of the amino acid residue in the corresponding solvent environment; the frequency points with intensity exceeding the background noise threshold in the discrete spectrum and their intensity values ​​are extracted and paired to form a standard electromagnetic response waveform; The above process is performed for all amino acid residue types and predefined non-aqueous solvent environments, and the generated waveforms are stored in a structured manner to build a computational chemistry feature database. For each amino acid residue in the target sequence and the required solvent environment, the corresponding standard waveform is retrieved from the database.

2. The method for constructing a mussel adhesive protein recognition model according to claim 1, characterized in that, In S2, the process of forming the electromagnetic feature sequence of the polypeptide fragment is as follows: Starting from the initial position of the electromagnetic fingerprint with a fixed window length and sliding step size, the system slides and cuts, and each time it captures the standard waveform corresponding to the amino acid position within the window to generate a polypeptide fragment electromagnetic fingerprint; for each generated polypeptide fragment electromagnetic fingerprint, each non-aqueous solvent environment is processed in sequence. When processing any non-aqueous solvent environment, the waveforms of all amino acids in the fragment under this solvent are aligned and superimposed, and the envelope of the superimposed waveform is calculated. This envelope waveform is used as the comprehensive response waveform of the polypeptide fragment under this non-aqueous solvent environment. After processing in all non-aqueous solvent environments, the comprehensive response waveforms of the same polypeptide fragment under each environment are compared. Frequency components that appear continuously in each waveform are identified and marked as shared frequency components. The average amplitude intensity of each shared frequency component in the comprehensive waveform under all non-aqueous solvent environments is calculated. The components are then sorted in descending order according to the average amplitude intensity and output as the electromagnetic feature sequence of the polypeptide fragment.

3. The method for constructing a mussel adhesive protein recognition model according to claim 2, characterized in that, The process of identifying frequency components that appear continuously in every waveform and marking them as shared frequency components is as follows: The overall response waveforms of the peptide fragments under all non-aqueous solvent environments were discretely sampled to obtain the discrete frequency sequence corresponding to each waveform. All discrete frequency sequences were then merged to form a total frequency set. For each unique frequency value in the total frequency set, count the number of times the frequency value appears in different non-aqueous solvent environments. If the number of times the frequency value appears is equal to the total number of non-aqueous solvent environments, then mark the frequency value as a candidate sustained frequency. Calculate the amplitude value corresponding to each candidate sustained frequency in different non-aqueous solvent environments to obtain a set of cross-environment amplitudes for the frequency. Calculate the standard deviation of the amplitude across the environment. If the standard deviation is lower than the preset amplitude fluctuation threshold, the candidate sustained frequency is finally determined as the shared frequency component. All the finally determined shared frequency components are summarized and marked as shared frequency components used to constitute the electromagnetic feature sequence.

4. The method for constructing a mussel adhesive protein recognition model according to claim 1, characterized in that, In step S3, the process of establishing a neural network structure that includes an encoder network and a decoder network is as follows: The encoder network is sequentially connected to the first one-dimensional convolutional layer, the first nonlinear activation layer, the first pooling layer, and the gated recurrent unit layer; the decoder network is a fully connected layer network. The first one-dimensional convolutional layer is configured with multiple convolutional kernels of different widths to extract local features of different scales from the input sequence; the first nonlinear activation layer performs an element-wise nonlinear transformation on the convolutional output; the first pooling layer downsamples the transformed features to compress the sequence length; the gated recurrent unit layer receives the downsampled feature sequence and outputs a context vector. The fully connected layer network receives a context vector and maps the dimension of the context vector to a preset fixed dimension through at least one fully connected layer with a non-linear activation function.

5. The method for constructing a mussel adhesive protein recognition model according to claim 1, characterized in that, In step S4, the process of the encoder network outputting the context vector is as follows: Set up a set of trainable one-dimensional convolutional kernels with different receptive field widths, input the electromagnetic feature sequence of the peptide fragment into the convolutional layer of the encoder network, and each convolutional kernel independently performs sliding dot product calculation and generates an initial feature map. For each initial feature map, a non-linear activation function is applied to transform it. Max pooling is then performed on the transformed feature map to compress the sequence dimension and output a sequence of feature vectors. The feature vector sequence is input into the gated recurrent unit layer, and each feature vector is processed sequentially. The update gate and reset gate are calculated based on the current input and the state of the previous time step. Historical information is modulated and combined with the current input to generate candidate states. The candidate states and historical states are fused by updating the gate to obtain the hidden state at the current time step. The hidden state at the end of the sequence processing is output as the context vector.

6. The method for constructing a mussel adhesive protein recognition model according to claim 1, characterized in that, In step S5, the process of mapping the context vector to the electromagnetic fingerprint-driven embedding representation is as follows: The fully connected layer of the decoder network receives the context vector from the encoder network. The fully connected layer consists of multiple sequentially connected linear transformation layers, where the first linear transformation layer transforms the dimension of the context vector into an intermediate dimension, and the second and subsequent linear transformation layers transform the intermediate dimension step by step. Nonlinear activation function layers are introduced after the first and second linear transformation layers to perform nonlinear mapping on the output of the linear transformation. The last linear transformation layer converts the dimension of the feature vector after nonlinear mapping into a pre-defined fixed dimension. The dense vector with fixed dimension output by the last linear transformation layer is defined as the electromagnetic fingerprint-driven embedding representation of the polypeptide fragment.

7. The method for constructing a mussel adhesive protein recognition model according to claim 1, characterized in that, In S6, the process of training the binary classification model is as follows: For each protein sequence in the training sample set, steps S1 to S5 are applied to generate electromagnetic fingerprint-driven embedding representations of all polypeptide fragments of the protein sequence. The standard deviation of the amplitude intensity of each embedding representation vector in all non-aqueous solvent environments corresponding to the polypeptide fragment is calculated. Vectors with standard deviations lower than a preset stability threshold are selected to form a stable embedding representation subset. The magnitude of each vector in the stable embedding representation subset is calculated, and the magnitude is used as the significance weight of the vector. The significance weight is then used to calculate a weighted average of all vectors in the stable embedding representation subset to generate a feature vector representing the overall characteristics of the protein sequence. The overall feature vectors of all training sample protein sequences and their corresponding binary labels of adhesion function are collected to form a training dataset. A support vector machine model is trained using the training dataset. The training process includes iteratively adjusting the normal vector and intercept parameters of the classification hyperplane in the feature space. The support vector machine model with fixed parameters after iterative adjustment is saved as the final mussel adhesive protein recognition model.

Images