A synthetic protein design method based on protein feature engineering and diffusion model

Abstract

The application discloses a synthetic protein design method based on protein feature engineering and diffusion model. First, a P-BLAST tool is used to search for homologous sequences of a query sequence in a non-redundant NR database by taking a target protein sequence as the query sequence; feature engineering is performed on the screened homologous sequences by using ProtParam-7 and ProtFactor-5; data represented by the ProtParam-7 and ProtFactor-5 features are taken as original protein input diffusion model; multidimensional screening is performed on the generated candidate proteins; wet experiments are adopted to evaluate the functional activity and structural stability of the candidate proteins, and proteins meeting the conditions are taken as the final generated results. The method combines protein sequence feature extraction and diffusion model generation, and through multidimensional screening and experimental verification, it is ensured that the generated protein is close to the target protein in structure and function.

Description

Technical Field

[0001] This invention relates to the field of computational biology, and in particular to a synthetic protein design method based on protein feature engineering and diffusion models. Background Technology

[0002] Proteins, as core molecules in life processes, have wide applications in medicine, industrial production, and environmental governance. Traditional protein design methods mainly rely on experimental screening and structural biology techniques. While these methods can achieve certain results, they often suffer from strong data dependence, long design cycles, and difficulty in exploring potential sequence spaces. Crucially, traditional methods typically require a large amount of known protein sequence, structural, and functional data to train models or guide experiments. When data is scarce, model performance is easily limited, thus affecting design effectiveness. Secondly, traditional protein optimization methods often require multiple rounds of experimental validation, a cumbersome and time-consuming process that further increases research costs. Furthermore, due to the complexity of natural proteins, empirical or random mutation-based methods cannot fully cover the potential effective sequence space, limiting the diversity and innovation of designs.

[0003] In recent years, the rapid development of deep learning technology has brought new opportunities for protein design. By training deep neural networks, researchers can model the relationship between protein sequences and their structure and function using large-scale data, thereby rapidly obtaining functional proteins with fewer experimental screenings. Among them, diffusion models, as a powerful generative model, have achieved significant results in fields such as image generation and text generation. Diffusion models can gradually recover high-quality target data from random noise during the data generation process, showing unique advantages in modeling complex data distributions. Introducing diffusion models into the field of protein design allows for the exploration of potential functional proteins in a broader sequence space, breaking through the limitations of traditional methods. Meanwhile, ProtParam-7 (p7) and ProtFactor-5 (f5), as classic protein feature representation methods, can effectively capture key information such as the physicochemical properties, hydrophobicity, and conformational features of amino acids. Using ProtParam-7 (p7) and ProtFactor-5 (f5) to characterize protein sequences can provide high-quality feature representations for diffusion models, thereby improving the reliability and functionality of generated proteins. However, existing technologies lack solutions for combining factors such as ProtParam-7 (p7) and ProtFactor-5 (f5), which represent rich physicochemical properties, with diffusion models for application in protein synthesis. Summary of the Invention

[0004] The purpose of this invention is to provide a synthetic protein design method based on protein feature engineering and diffusion models. This method combines protein sequence feature extraction with diffusion model generation, and through multidimensional screening and experimental verification, ensures that the generated protein is structurally and functionally close to the target protein, thereby improving the accuracy and applicability of the design.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A synthetic protein design method based on protein feature engineering and diffusion models, the method comprising:

[0007] Step 1: First, use the P-BLAST tool with the target protein sequence as the query sequence to search for homologous sequences of the query sequence in the non-redundant NR database. By filtering out sequences with abnormal lengths, a dataset containing the filtered homologous sequences is constructed. Here, P-BLAST is an iterative search algorithm based on the position-specific scoring matrix PSSM; the non-redundant NR database is a protein sequence database that has been filtered and deduplicated.

[0008] Step 2: Use ProtParam-7 and ProtFactor-5 feature engineering on the screened homologous sequences to convert the protein sequences into data represented by ProtParam-7 and ProtFactor-5 features. ProtParam-7 and ProtFactor-5 feature engineering are protein feature representation methods that can capture the physicochemical properties, hydrophobicity, and conformational features of amino acids.

[0009] Step 3: Input the data represented by ProtParam-7 and ProtFactor-5 features into the diffusion model as the original protein, and use the diffusion model to generate candidate proteins with diversity and structural rationality.

[0010] Step 4: Perform multidimensional screening on the generated candidate proteins. The screening criteria include the root mean square deviation of the backbone distance, solubility, and physicochemical properties to ensure that the functional characteristics of the candidate proteins meet the actual requirements.

[0011] Step 5: In the experimental verification stage, wet experiments are used to evaluate the functional activity and structural stability of candidate proteins, including protein expression, purification and functional detection. Proteins that meet the criteria are taken as the final results.

[0012] As can be seen from the technical solution provided by the present invention, the above method combines protein sequence feature extraction and diffusion model generation, and through multidimensional screening and experimental verification, ensures that the generated protein is close to the target protein in structure and function, improves the accuracy and applicability of the design, and provides an efficient and comprehensive optimization strategy for protein engineering, drug development, enzyme optimization and other fields. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the synthetic protein design method based on protein feature engineering and diffusion model provided in an embodiment of the present invention. Detailed Implementation

[0015] 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, and do not constitute a limitation 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 protection scope of the present invention.

[0016] like Figure 1 The diagram shows a flowchart of a synthetic protein design method based on protein feature engineering and diffusion models provided in an embodiment of the present invention. The method includes:

[0017] Step 1: First, using the P-BLAST tool with the target protein sequence as the query sequence, search for homologous sequences in the Non-Redundant (NR) database. The Non-Redundant (NR) database is a filtered and deduplicated protein sequence database, primarily providing a non-repetitive, high-quality collection of protein sequences for accurate and efficient protein analysis and comparison. Then, by filtering out sequences with abnormal lengths, a dataset containing the filtered homologous sequences is constructed.

[0018] In this step, P-BLAST (Position-Specific Iterated BLAST) is an iterative search algorithm based on the position-specific scoring matrix PSSM, which has a strong ability to identify distantly related homologous sequences.

[0019] P-BLAST first initializes the position-specific scoring matrix (PSSM) based on the amino acid composition and conserved region characteristics of the target protein sequence. When searching in a non-redundant NR database, the PSSM is continuously updated in each iteration to improve the sensitivity and accuracy of the search. Through this iterative optimization mechanism, P-BLAST can identify homologous sequences with highly similar conserved regions and potential functional properties, thereby capturing more comprehensive and diverse data information related to the target protein.

[0020] To ensure dataset quality, sequences with abnormal lengths were removed from the obtained candidate homologous sequences. Specifically, entries shorter or longer than the target protein sequence were eliminated to avoid introducing noisy data due to sequencing errors, pseudogenes, or incomplete fragments. This screening process helps retain sequence samples with greater biological significance and structural integrity. The resulting homologous sequence dataset covers diverse mutation patterns and conserved region features, providing high-quality, information-rich input data for subsequent diffusion model training.

[0021] Step 2: Use ProtParam-7 and ProtFactor-5 feature engineering on the screened homologous sequences to convert the protein sequences into data represented by ProtParam-7 and ProtFactor-5 features. ProtParam-7 and ProtFactor-5 feature engineering are protein feature representation methods that can capture the physicochemical properties, hydrophobicity, and conformational features of amino acids.

[0022] In this step, protein sequence characterization is specifically performed using ProtParam-7 and ProtFactor-5 factors. The data represented by ProtParam-7 (i.e., p7) features include:

[0023] Stereo parameters are used to characterize the complexity, branching degree, and symmetry of amino acid side chains;

[0024] Polarizability is used to characterize the molecular polarization ability of amino acids under electrostatic conditions.

[0025] Volume parameter used to characterize the standardized van der Waals volume of amino acids;

[0026] Hydrophobicity parameters are used to characterize the hydrophilic or hydrophobic properties of amino acids.

[0027] The isoelectric point parameter is used to characterize the charge state of amino acids under different pH conditions.

[0028] Helical tendency parameter, used to characterize the probability of amino acids forming α-helices;

[0029] β-sheet propensity parameter, used to characterize the probability of an amino acid forming a β-sheet;

[0030] The data represented by ProtFactor-5 (i.e., f5) features includes:

[0031] The polarity index is used to characterize the polarity, hydrophilicity, hydrophobicity, and number of hydrogen bond donors of amino acids.

[0032] Secondary structure index is used to characterize the structural tendency of amino acids in α-helix, β-sheet and random coil;

[0033] Molecular volume index is used to characterize the molecular weight, volume, and side chain size of amino acids.

[0034] Codon diversity index, used to characterize the genetic codon diversity and relative abundance of amino acids;

[0035] The charge index is used to characterize the isoelectric point and charge distribution characteristics of amino acids.

[0036] Among them, the data represented by ProtParam-7 and ProtFactor-5 features are used to characterize the physicochemical properties and conformational characteristics of amino acids, and to transform protein sequences into low-dimensional numerical feature matrices.

[0037] In the specific implementation, the p7 parameter is obtained by collecting statistical characteristics of amino acid side chain volume, polarizability, hydrophobicity, isoelectric point, and secondary structure, and then performing dimensionality reduction. The f5 parameter is obtained by performing factor analysis and rotation processing on 544 amino acid properties. Both p7 and f5 parameters are standardized, and a fixed 7-dimensional or 5-dimensional vector is assigned to each natural amino acid. During sequence encoding, each amino acid residue in the protein sequence is replaced with the corresponding vector, thereby constructing a numerical matrix of length ×12. Table 1 shows the standardized values ​​of p7 and f5.

[0038] Table 1. Values ​​of ProtParam-7 (p7) and ProtFactor-5 (f5) factors

[0039]

[0040] Step 3: Input the data represented by ProtParam-7 and ProtFactor-5 features into the diffusion model as the original protein, and use the diffusion model to generate candidate proteins with diversity and structural rationality.

[0041] In this step, the diffusion model, also known as the generative model, is used to generate samples similar to the training data. Its core working principle is to progressively add Gaussian noise to the training data to disrupt its structure, and then learn to reverse the noise process to recover the data, thereby generating new samples. This method not only generates samples similar to the original data but also introduces diversity to a certain extent, providing new possibilities for protein design. The diffusion model employs a generative strategy combining forward diffusion and reverse diffusion to progressively optimize protein feature representations and generate structurally sound protein sequences, where:

[0042] During forward diffusion, the protein feature matrix is ​​represented by ProtParam-7 (p7 features) and ProtFactor-5 (f5 features). :

[0043] (1)

[0044] in, Indicates the first indivual Quantity; Indicates the first indivual Quantity; and These are the corresponding scalar coefficients; For bias terms; subscript This indicates the sample number, i.e., the sequence position, used to distinguish different residues or different samples;

[0045] The diffusion model gradually evolves into a protein feature matrix. Adding Gaussian noise to the structure gradually causes it to lose its original structural features. This process is represented by the following formula:

[0046] (2)

[0047] in, This represents the feature matrix after adding noise at time step t; It is the preset noise figure, a small positive number used to control the intensity of noise; It is Gaussian noise and follows a standard normal distribution. As time step t increases, the feature matrix... It will gradually become chaotic and eventually lose its original characteristic information;

[0048] The forward diffusion process can be viewed as a Markov chain, with the following state transition probability distribution:

[0049] (3)

[0050] This means that in each step, the feature matrix Based on the feature matrix of the previous step and Gaussian noise The goal of the entire forward diffusion process is to generate the initial feature matrix. Transformed into a fully noisy feature matrix , where t is the total number of steps in the diffusion process;

[0051] In the reverse diffusion process, the diffusion model evolves from the final noise matrix. Begin by gradually denoising to recover the original feature matrix. At each step t, through the neural network To predict and reduce noise, the goal is to ultimately recover a feature matrix that approximates the one generated by the temple coefficients. Each step of the reverse diffusion process is formulated as follows:

[0052] (4)

[0053] in, The noise is predicted by a neural network and used to estimate the actual noise during the forward diffusion process; to control the randomness of the generation process, a scalar or diagonal matrix is ​​introduced into the model. The formula for adjusting the variance of the forecast and the mean is:

[0054] (5)

[0055] in It is the output of the noise prediction network, used to estimate the noise during the forward process; This represents noise intensity, which directly affects training stability and sampling quality. The cumulative retention factor is the one that is closer to 1, the more of the original signal is retained after t steps.

[0056] The reverse diffusion process is also a Markov chain, and its state transition probability distribution is as follows:

[0057] (6)

[0058] in The introduced scalar or diagonal matrix is ​​the variance of the predictions used to control the randomness of the generation, optimizing the mean prediction. The reverse generation step starts from a pure noise matrix. Start by sampling sequentially. Until the original feature matrix is ​​recovered .

[0059] Furthermore, during the training of the diffusion model, the goal is to train the denoiser in the reverse diffusion process, that is, to train the model how to remove noise. This process is achieved by minimizing the noise prediction error, specifically:

[0060] The model optimizes the output of the noise prediction network to make it as close as possible to the actual noise during the forward diffusion process, with the loss function... The mean squared error (MSE) is used, and the formula is:

[0061] (7)

[0062] in, It is the output of the noise prediction network. These are network parameters; It is the original feature matrix; Represents Gaussian noise;

[0063] By minimizing the loss function The model can learn how to accurately predict noise and gradually recover the original feature matrix. To achieve this, a generator architecture based on a multilayer perceptron (MLP) is adopted. The MLP consists of multiple fully connected layers, and the ReLU activation function is used after the hidden layers to introduce non-linearity. The model's input includes a noisy feature matrix. and time-step embedding vectors generated by sinusoidal coding Time step embedding vector Its purpose is to provide the model with information about the current time step, helping the model better understand the progress of the denoising process. The time step embedding vector... The formula is:

[0064] (8)

[0065] in, It is a frequency parameter, usually taken as . In this way, the time step embedding vector It is embedded into a high-dimensional vector, providing the model with rich contextual information;

[0066] The specific structure of a multilayer perceptron (MLP) consists of three parts: the input layer receives the noisy feature matrix. and time step embedding vector The hidden layer contains multiple fully connected layers, each followed by a ReLU activation function to introduce nonlinearity; the final output layer outputs the predicted noise. ;

[0067] By transforming the input features layer by layer, the Multilayer Perceptron (MLP) can progressively predict noise and remove it through inverse inference, ultimately recovering the original feature matrix. In practical implementation, MLP performs well in global feature processing and can effectively recover data structures. In the denoising task of diffusion models, MLP can capture complex relationships in the feature matrix and generate new feature matrices by learning the denoising process.

[0068] The final diffusion model will generate a new feature matrix. The matrix Compared with the original wooden temple coefficient characteristic matrix Similar, but with some diversity, in order to incorporate the newly generated feature matrix The protein sequence is restored, and the similarity between the newly generated features and the original features is calculated using cosine similarity. Based on this similarity, the new generated sequence is then reconstructed. The cosine similarity calculation formula is as follows:

[0069] (9)

[0070] When the similarity reaches a set threshold, the generated features are considered to be close to the original protein features, and then based on this feature matrix... Restore the protein sequence.

[0071] The generated feature matrix can be further used for protein structure prediction and function optimization, providing new ideas and methods for protein design.

[0072] Step 4: Perform multidimensional screening on the generated candidate proteins. The screening criteria include the root mean square deviation of the backbone distance (RMSD), solubility, and physicochemical properties (isoelectric point and hydrophobicity) to ensure that the functional characteristics of the candidate proteins meet the actual requirements.

[0073] In this step, the AlphaFold3 tool (an advanced protein structure prediction platform based on deep learning) is first used to calculate the backbone distance between each generated candidate protein and the target protein sequence. Specifically, the root mean square deviation (RMSD) is calculated to measure the difference between the two structures, assessing the average distance between protein atoms, typically concentrated on the aligned backbone atoms. The unit of RMSD is... , The formula for calculating the root mean square deviation (RMSD) is:

[0074] (10)

[0075] Where N represents the number of atoms involved in the calculation (usually framework atoms, such as Cα, N, C, and O atoms); δ i δ represents the positional deviation of the i-th atom between the two structures, i.e., the Euclidean distance between the coordinates of the two structures after alignment; i 2 The square of the positional deviation is taken as the final square root, which is the RMSD. The smaller the RMSD value, the closer the two structures are; the larger the value, the greater the difference.

[0076] During the virtual screening process, the solubility of the generated candidate proteins is compared. Specifically, the NovoPro tool is used. NovoPro is a tool specifically designed for protein solubility prediction. Based on the protein's sequence information, this tool can predict its solubility and give a solubility score between 0 and 1. A solubility value close to 1 indicates that the protein has good solubility, while a solubility value close to 0 indicates that the protein has poor solubility.

[0077] Based on solubility prediction, a solubility threshold of no less than 0.8 is set to screen out protein sequences with good solubility, thereby improving the feasibility of subsequent experiments and applications.

[0078] In terms of physicochemical property screening, to more comprehensively evaluate the characteristics of the generated proteins, two different amino acid characterization methods, namely the ProtFactor-5 parameter and the ProtParam-7 parameter, were used to compare the generated candidate proteins with the target protein. By using these two different amino acid characterization methods, the generated candidate proteins were cross-validated from different perspectives to ensure that their physicochemical properties were highly similar to those of the target protein and to accurately identify differences in key features. Ultimately, this multi-dimensional screening strategy can improve the screening accuracy and ensure that the selected proteins meet the experimental requirements in terms of structure, solubility, and functional properties, thereby improving the reliability of subsequent applications.

[0079] Step 5: In the experimental verification stage, wet experiments are used to evaluate the functional activity and structural stability of candidate proteins, including protein expression, purification and functional detection. Proteins that meet the criteria are taken as the final results.

[0080] In this step, during the experimental validation phase, the gene sequence of the candidate protein is first cloned into the pET28a vector, an E. coli expression vector with a His tag. The His tag is a common multihistidine tag (usually 6×His) that specifically binds to nickel ions (Ni²⁺), thus facilitating protein purification using Ni-NTA affinity chromatography. The vector was then transformed into BL21(DE3) bacteria for protein expression. BL21(DE3) is a commonly used Escherichia coli expression host strain that carries the T7 RNA polymerase gene, enabling efficient expression of the target protein under IPTG induction. Under 220 rpm shaking culture, when the OD600 value reached 0.6 (OD600 refers to the optical density of bacterial culture at a wavelength of 600 nm, which usually reflects the bacterial growth status; when OD600 ≈ 0.6, it indicates that the bacteria are in the logarithmic growth phase, suitable for protein induction expression), isopropyl-β-D-thiogalactoside IPTG at a final concentration of 2 mmol / L was added to induce protein expression.

[0081] After culture, the bacterial pellet was collected by centrifugation, and the cells were lysed by sonication to release the target protein. The protein was then purified using Ni-NTA affinity chromatography (Ni-NTA is a nickel-nitrogen triacetic acid packing material that specifically binds to the His tag), and eluted with PBS + 500 mM imidazole solution. PBS is a commonly used phosphate buffer with a high concentration of imidazole (a nitrogen-containing five-membered heterocyclic organic compound with coordination ability similar to histidine side chains), which competitively binds to Ni²⁺, effectively eluting the His-tagged protein. The purified protein was dialyzed to 1×PBS to remove impurities, and its purity was assessed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) to ensure a purity of over 95% for subsequent functional assays.

[0082] To evaluate the functional activity of candidate proteins, affinity assays were performed using Bio-Layer Interferometry (BLI) on a ForteBio Octet K2 instrument. The ForteBio Octet K2 is a BLI-based biomolecular interaction analyzer capable of real-time monitoring of protein-protein or protein-antibody binding kinetics. During the experiment, the purified protein was first immobilized on a His sensor (4 μg / mL, immobilization for 120 s), and then equilibrated in PBST (phosphate-buffered saline + Tween 20) buffer for 120 s to reduce background noise. The sensor was then sequentially immersed in solutions of human IgG4 monoclonal antibody (mAb1) at different concentrations (50 nM, 25 nM, 12.5 nM, 6.25 nM) for binding experiments. IgG4 (Immunoglobulin G4) is a subclass of IgG that plays a unique role in human immune tolerance and chronic inflammation, and is commonly used in the development of therapeutic antibodies. Each concentration was incubated for 120 s and then dissociated in PBST for 180 s to determine the protein binding stability. After the experiment, the sensor was regenerated using a pH 1.7 glycine solution to remove residual conjugates and proceed to the next round of measurements. Finally, the binding ability of the candidate protein to the target antibody was evaluated by calculating the dissociation constant (KD) value, where a lower KD value indicates stronger binding affinity and higher stability.

[0083] Based on experimental data, candidate proteins with similar functional activity to the target protein and good structural stability were screened. Combined with the results of BLI affinity assay, proteins with dissociation constant KD values ​​within a reasonable range were selected as the final generated results.

[0084] In practice, to ensure the structural stability of the protein, circular dichroism (CD) spectroscopy was used to assess the secondary structure composition of the protein, including the proportions of α-helices, β-sheets, and random coils, to determine whether the protein maintained the correct conformation. Simultaneously, differential scanning fluorometry (DSF) was used to record the folding state of the protein at different temperatures and analyze its thermal stability. Finally, through protein expression, purification, BLI binding experiments, and structural stability analysis, candidate proteins were comprehensively evaluated to ensure that the selected proteins met experimental requirements in terms of functional activity and structural stability, laying the foundation for subsequent applications.

[0085] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0086] The method described in this embodiment of the invention is illustrated below with specific examples. The Protein A sequence used in the examples is derived from the immunoglobulin-binding B region of Protein A in Staphylococcus aureus. SPA is a protein isolated from the cell wall of Staphylococcus aureus and possesses various immunomodulatory properties. For example, it can bind IgG in humans and several animals, including guinea pigs, pigs, mice, and monkeys. SPA has a wide range of biochemical, biotechnological, and medical applications, particularly in diagnostic techniques such as immunoprecipitation, enzyme-linked immunosorbent assay (ELISA), and Western blotting. It is also widely used in pharmaceutical applications to bind immune complexes and serum immunoglobulins. Furthermore, SPA interacts directly with B cells, playing a role in preventing infectious diseases and influencing the development of autoimmune diseases. This example uses the B domain sequence of SPA (PDB identifier: 1BDC:a).

[0087] Example 1: Protein generation by mimicking the Protein A sequence

[0088] 1. Generate protein sequences

[0089] This example first uses the P-BLAST tool to search for homologous sequences of SPA proteins in the NR database, and filters out sequences with abnormal lengths to construct a dataset of 56 sequences. Next, the original SPA protein sequences and homologous sequences are transformed into feature vectors using ProtParam-7 (p7) and ProtFactor-5 (f5) feature engineering to characterize their physicochemical properties and conformational features. These feature vectors are then input into a diffusion model to generate a large number of diverse and structurally sound candidate proteins. A total of 9 sequences were generated, as shown in Table 2 below.

[0090] Table 2. SPA protein and sequences of 9 synthetic proteins

[0091]

[0092] To distinguish the generated protein from the original protein, the location and number of site changes were compared to show the differences. The site change parameters of the generated protein sequence are shown in Table 3 below:

[0093] Table 3. Parameters for generating protein sequence site changes

[0094]

[0095] Although the number and location of site changes are the same for FC3 and FC4, the types of amino acids resulting from these changes are different. Even minor alterations can lead to serious consequences such as protein structural instability, water insolubility, or poor drug efficacy. The parameters for site changes in the resulting protein sequence are shown in Table 4 below.

[0096] Table 4. Parameters for changing protein sequence sites

[0097]

[0098] 2. Comparison of protein sequence solubility and Peptide Rank.

[0099] After obtaining the protein sequence output by the model, its solubility compared to the synthesized protein sequence was calculated using the NovoPro solubility prediction tool. This process first eliminated sequences with low solubility to prevent them from being completely insoluble in water. Simultaneously, the online tool PeptideRank was used for comparison. Its main function is to predict whether a peptide can penetrate the cell membrane based on the characteristics of the peptide sequence and rank the peptides according to their probability of cell penetration. This is primarily to identify peptides with high penetration ability to improve drug delivery efficiency. The specific protein sequence parameters are shown in Table 5 below.

[0100] Table 5. Parameters for Generating Protein Sequences

[0101]

[0102] According to the parameter table, the solubility of each synthesized protein sequence is checked to ensure it is greater than 0.80, guaranteeing that the synthesized protein sequence is soluble in water. At the same time, sequences with excessively high Peptide Rank need to be screened to prevent affecting drug performance. The remaining synthesized protein sequences enter the rescreening stage.

[0103] 3. Generate RMSD (backbone distance) and feature similarity of protein sequences.

[0104] After initial screening, sequences that did not meet the solubility and PeptideRank thresholds were eliminated. Then, the secondary structure of the generated protein sequence was calculated using the AlphaFold2 tool. Next, the backbone distance between the generated protein sequence and the target sequence was calculated using Pymol software, and the feature distance was calculated based on ESM2. The backbone distance and feature distance between the generated protein and the original protein are shown in Table 6 below:

[0105] Table 6. Skeletal distance and feature distance between the generated protein and the original protein.

[0106]

[0107] Subsequently, to comprehensively assess the similarity between the generated protein and the original protein, a feature similarity method was employed to compare their physicochemical properties and structural characteristics. During the comparison, different strategies, one using 7 features and the other using 5 features, were employed to evaluate the quality of the generated protein and its similarity to the target protein from different dimensions. The feature similarity between the generated protein and the original protein is shown in Table 7 below:

[0108] Table 7. Feature Similarity Table between Generated Protein and Original Protein

[0109]

[0110] Based on the above multi-dimensional screening results, only four sequences—FC-1, FC-3, FC-6, and FC-7—were ultimately retained for wet experimental verification. These four sequences exhibited significant site variations, their physicochemical properties closely matched the original sequences, their RMSD values ​​were similar, and their solubility exceeded the threshold, thus meeting all the required standards.

[0111] Following this, we performed wet-layer experiments to validate the generated proteins. In this stage, we used biolayer interferometry (BLI) to assess the affinity of different candidate proteins at different analyte concentrations. The experiments involved immobilizing candidate proteins (4 μg / mL) and performing binding experiments with different concentrations of analytes (TY101-500, TY101-250, TY101-125, TY101-67.5) to calculate the dissociation constant KD value, thus measuring the binding ability of the candidate proteins.

[0112] Experimental data showed that FC1 and FC7 exhibited extremely low KD values ​​at all analyte concentrations. FC-1 showed a binding affinity close to <1.0E-12 M, while FC-7 reached levels ranging from 1.40E-11 M to 1.32E-10 M, indicating extremely tight binding to the target molecule and significantly superior affinity compared to other candidate proteins. In contrast, proA-FC-3 and proA-FC-6 showed significantly higher KD values ​​at lower concentrations (TY101-125 and TY101-67.5), indicating weaker binding affinity. Considering the BLI affinity test results at different analyte concentrations, proA-FC-1 and proA-FC-7 demonstrated stronger binding stability and lower dissociation constants (KD values), indicating a greater advantage in binding target molecules. Therefore, these two proteins can be considered as preferred candidate proteins for further in-depth research or application development to improve the specificity and stability of protein binding. The dissociation constants of the generated protein and the original protein are shown in Table 8 below.

[0113] Table 8. Dissociation constants between generated and original proteins

[0114]

[0115] After further verifying the purification effect and binding stability of the candidate proteins, Ni-NTA affinity chromatography was used to purify proA-FC-1, proA-FC-7, proA-FC-6, and proA-FC-3, and their binding ability and recovery efficiency were analyzed. Ni-NTA purification experiments can evaluate the binding ability, elution characteristics, and purity of proteins. By detecting the UV 280 nm absorption peak, elution volume, and loss of non-specifically bound proteins, the optimal candidate proteins were further screened.

[0116] Ni-NTA affinity chromatography revealed significant UV 280 nm elution peaks for FC-1 and FC-7 at 100.14 ml and 98.17 ml, respectively. These high peak values ​​indicate efficient binding of these two proteins to the Ni-NTA affinity column and their release at high concentrations during elution. Furthermore, the low UV absorbance of FC-1 and FC-7 during the flow-through phase (0-20 ml) suggests fewer unbound proteins, indicating better protein binding to the Ni-NTA medium and reduced loss. In contrast, the relatively low elution peaks for FC-6 and FC-3 at 98.52 ml and 95.66 ml suggest slightly weaker protein binding, with some protein potentially being lost during flow-through or washing.

[0117] Further analysis of the washing phase (20-80 ml) revealed lower UV 280 nm signals for FC-1 and FC-7, indicating fewer non-specifically bound proteins and higher purity. FC-6 and FC-3, however, still exhibited low UV absorption during the washing phase, suggesting the possible presence of some unbound proteins. Furthermore, during the elution phase, the UV 280 nm absorption intensity of FC-1 and FC-7 increased rapidly with rising Conc B (green curve, imidazole concentration), demonstrating efficient elution and good purification even at high imidazole concentrations.

[0118] Therefore, FC-1 and FC-7 showed superior binding ability and recovery rate compared to FC-3 and FC-6 in Ni-NTA purification experiments. Combined with BLI affinity assay results, this further confirms proA-FC-1 and proA-FC-7 as preferred candidate proteins, possessing stronger binding stability, lower dissociation constants, and better purification recovery rates. Therefore, these two proteins can be considered as key areas for further in-depth research or application development to improve the specificity and stability of protein binding and enhance the reliability of experimental data.

[0119] In summary, the method described in the embodiments of the present invention has the following advantages:

[0120] 1. Efficient Data-Driven Screening Method: In this method, the P-BLAST tool is used to screen homologous protein sequences from the NR database. By removing sequences with abnormal lengths, the obtained homologous sequence dataset is ensured to have a high correlation in structural and functional characteristics. Compared with traditional manual screening methods, P-BLAST, combined with iterative optimization, can quickly and accurately identify sequences similar to the target protein, thus providing a reliable data foundation for subsequent generation. This efficient data screening greatly improves the speed of protein design and can handle large-scale datasets.

[0121] 2. ProtParam-7 (p7) and ProtFactor-5 (f5) have significant advantages in sequence characterization: they compress the physicochemical and structural statistics of amino acids into low-dimensional vectors of 7 and 5 dimensions, respectively. This preserves key physicochemical information (such as volume, polarizability, hydrophobicity, isoelectric point, and secondary structure tendency) while significantly reducing input dimensionality and redundancy, thereby improving the training efficiency and robustness of the generative model. p7, through data-driven network dimensionality reduction, retains the ability to reconstruct the original attributes, thus making it more effective in capturing signals related to secondary structure and local conformation. f5, on the other hand, extracts biologically interpretable dominant factors through factor analysis, which helps to introduce interpretability constraints into the modeling process. By combining p7 and f5, the physicochemical properties and conformational features of 20 natural amino acids are compressed into 12 factors, which not only preserves the physicochemical information of each amino acid but also effectively reduces dimensionality, avoids redundancy, and improves the learning efficiency of the generative model. Furthermore, this feature representation method can more comprehensively characterize the structural and functional properties of proteins, enabling diffusion models to learn the deep-seated patterns of proteins more accurately.

[0122] 3. Innovative Application of Diffusion Models: In this method, diffusion models are used to generate candidate protein sequences. By progressively adding noise to the original data and restoring the data structure through a reverse denoising process, diffusion models can generate protein sequences that combine diversity and structural rationality. Compared to traditional random mutagenesis or Markov chain Monte Carlo (MCMC) methods, this approach allows for better control over the structural stability and functional properties of proteins during the generation process. By efficiently modeling the protein sequence space, diffusion models enable the generated proteins to not only be similar to the target protein in physicochemical properties but also possess greater diversity and innovation, thereby enhancing the breadth and depth of protein design.

[0123] 4. Multidimensional Screening Ensures Protein Quality: The generated candidate proteins undergo rigorous evaluation through a multidimensional screening process. Screening criteria include the root mean square deviation (RMSD) of the backbone distance, solubility, and physicochemical properties (such as isoelectric point and hydrophobicity). These criteria ensure the rationality and stability of the candidate proteins in terms of structure and function. The backbone distance (RMSD), as a key indicator of the similarity of protein three-dimensional structure, effectively guarantees the consistency between the structure of the generated protein and the target protein. Furthermore, screening for physicochemical properties such as solubility further optimizes the stability and biological function of the protein. Multidimensional screening provides a more comprehensive evaluation system, ensuring the quality of the generated protein from multiple perspectives and avoiding the limitations that may arise from a single evaluation criterion.

[0124] 5. Experimental Validation Ensures Biological Activity and Function: Unlike traditional computer simulations and theoretical analyses, this method experimentally validates the functional activity and structural stability of candidate proteins, ensuring the effectiveness and operability of the final generated proteins in practical applications. The wet experimental process includes protein expression, purification, and functional detection, using various experimental techniques to verify the protein's biological function and stability. This experimental validation not only enhances the credibility of protein design but also provides reliable experimental data support for subsequent practical applications (such as drug development and enzyme engineering). The inclusion of experimental steps allows this method to not only possess theoretical advantages but also be closely integrated with practical applications.

[0125] 6. Improved success rate and efficiency of protein design: This method combines traditional protein design steps with modern computational methods. Through efficient data screening, feature engineering, generative models, and multidimensional screening, the entire protein design process becomes more efficient and accurate. In traditional methods, protein design often relies on extensive experimental validation, which is inefficient. This method, however, uses diffusion models and multidimensional screening to preliminarily screen for high-quality candidate proteins before experiments, significantly shortening the design and optimization cycle and improving the success rate. This approach not only saves time and resources but also allows for the identification of potential candidate proteins in a broader protein space.

[0126] 7. Enhanced diversity and innovation in protein design: Compared to traditional template-based design methods, generating protein sequences through diffusion models can produce more innovative proteins. These proteins not only retain the core functional characteristics of the target protein but also offer more options for protein engineering by generating different variants, thereby promoting the discovery and application of novel functional proteins. For example, in the fields of drug development and enzyme engineering, generating proteins with new structures and functions may bring more research opportunities and technological breakthroughs.

[0127] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for designing synthetic proteins based on protein feature engineering and diffusion models, characterized in that, The method includes: Step 1: First, use the P-BLAST tool with the target protein sequence as the query sequence to search for homologous sequences of the query sequence in the non-redundant NR database. By filtering out sequences with abnormal lengths, a dataset containing the filtered homologous sequences is constructed. Here, P-BLAST is an iterative search algorithm based on the position-specific scoring matrix PSSM; the non-redundant NR database is a protein sequence database that has been filtered and deduplicated. Step 2: Use ProtParam-7 and ProtFactor-5 feature engineering on the screened homologous sequences to convert the protein sequences into data represented by ProtParam-7 and ProtFactor-5 features. ProtParam-7 and ProtFactor-5 feature engineering are protein feature representation methods that can capture the physicochemical properties, hydrophobicity, and conformational features of amino acids. Step 3: Input the data represented by ProtParam-7 and ProtFactor-5 features into the diffusion model as the original protein, and use the diffusion model to generate candidate proteins with diversity and structural rationality. In step 3, the diffusion model employs a generation strategy combining forward and backward diffusion to progressively optimize protein feature representation and generate structurally sound protein sequences, wherein: During forward diffusion, the protein feature matrix is ​​represented by ProtParam-7 (p7 features) and ProtFactor-5 (f5 features). : （1） in, Indicates the first indivual Quantity; Indicates the first indivual Quantity; and These are the corresponding scalar coefficients; For bias terms; subscript This indicates the sample number, i.e., the sequence position, used to distinguish different residues or different samples; The diffusion model gradually evolves into a protein feature matrix. Adding Gaussian noise to the structure gradually causes it to lose its original structural features. This process is represented by the following formula: （2） in, This represents the feature matrix after adding noise at time step t; It is a preset noise figure used to control the intensity of noise; It is Gaussian noise and follows a standard normal distribution. As time step t increases, the feature matrix... It will gradually become chaotic and eventually lose its original characteristic information; The forward diffusion process can be viewed as a Markov chain, with the following state transition probability distribution: （3） This means that in each step, the feature matrix Based on the feature matrix of the previous step and Gaussian noise The goal of the entire forward diffusion process is to generate the initial feature matrix. Transformed into a fully noisy feature matrix , where t is the total number of steps in the diffusion process; In the reverse diffusion process, the diffusion model evolves from the final noise matrix. Begin by gradually denoising to recover the original feature matrix. At each step t, through the neural network To predict and reduce noise, the goal is to ultimately recover a feature matrix that approximates the one generated by the temple coefficients. Each step of the reverse diffusion process is formulated as follows: （4） in, The noise is predicted by a neural network and used to estimate the actual noise during the forward diffusion process; to control the randomness of the generation process, a scalar or diagonal matrix is ​​introduced into the model. The formula for adjusting the variance of the forecast and the mean is: （5） in It is the output of the noise prediction network, used to estimate the noise during the forward process; This represents noise intensity, which directly affects training stability and sampling quality. The cumulative retention factor is the one that is closer to 1, the more of the original signal is retained after t steps. The reverse diffusion process is also a Markov chain, and its state transition probability distribution is as follows: （6） in The introduced scalar or diagonal matrix is ​​the variance of the predictions used to control the randomness of the generation, optimizing the mean prediction. The reverse generation step starts from a pure noise matrix. Start by sampling sequentially. Until the original feature matrix is ​​recovered ; During the training of the diffusion model, the goal is to train the denoiser in the reverse diffusion process, that is, to train the model how to remove noise. This process is achieved by minimizing the noise prediction error, specifically: The model optimizes the output of the noise prediction network to make it as close as possible to the actual noise during the forward diffusion process, with the loss function... The mean squared error (MSE) is used, and the formula is: （7） in, It is the output of the noise prediction network. These are network parameters; It is the original feature matrix; Represents Gaussian noise; By minimizing the loss function The model can learn how to accurately predict noise and gradually recover the original feature matrix. To achieve this, a generator architecture based on a multilayer perceptron (MLP) is adopted. The MLP consists of multiple fully connected layers, and a ReLU activation function is used after the hidden layers to introduce non-linearity. The model's input includes a noisy feature matrix. and time-step embedding vectors generated by sinusoidal coding Time step embedding vector Its function is to provide the model with information about the current time step, the time step embedding vector. The formula is: （8） in, It is a frequency parameter; in this way, the time step embedding vector It is embedded into a high-dimensional vector, providing the model with rich contextual information; The specific structure of a multilayer perceptron (MLP) consists of three parts: the input layer receives the noisy feature matrix. and time step embedding vector The hidden layer contains multiple fully connected layers, each followed by a ReLU activation function to introduce nonlinearity; the final output layer outputs the predicted noise. ; By transforming the input features layer by layer, the Multilayer Perceptron (MLP) can progressively predict noise and remove it through inverse inference, ultimately recovering the original feature matrix. ; The final diffusion model will generate a new feature matrix. The matrix Compared with the original wooden temple coefficient feature matrix Similarly, in order to generate the new feature matrix The protein sequence is restored, and the similarity between the newly generated features and the original features is calculated using cosine similarity. Based on this similarity, the new generated sequence is then reconstructed. The cosine similarity calculation formula is as follows: （9） When the similarity reaches a set threshold, the generated features are considered to be close to the original protein features, and then based on this feature matrix... Restore the protein sequence; Step 4: Perform multidimensional screening on the generated candidate proteins. The screening criteria include the root mean square deviation of the backbone distance, solubility, and physicochemical properties to ensure that the functional characteristics of the candidate proteins meet the actual requirements. Step 5: In the experimental verification stage, wet experiments are used to evaluate the functional activity and structural stability of candidate proteins, including protein expression, purification and functional detection. Proteins that meet the criteria are taken as the final results.

2. The synthetic protein design method based on protein feature engineering and diffusion models according to claim 1, characterized in that, In step 1, P-BLAST first initializes the position-specific scoring matrix PSSM based on the amino acid composition and conserved region features of the target protein sequence. When searching in a non-redundant NR database, the position-specific scoring matrix PSSM is continuously updated in each iteration to improve the sensitivity and accuracy of the search. Among the obtained candidate homologous sequences, sequences with abnormal lengths are screened out, specifically entries that are shorter or longer than the target protein sequence, in order to avoid introducing noisy data due to sequencing errors, pseudogenes, or incomplete fragments.

3. The synthetic protein design method based on protein feature engineering and diffusion models according to claim 1, characterized in that, In step 2, protein sequence characterization is specifically performed using ProtParam-7 and ProtFactor-5 factors. The data represented by ProtParam-7 features include: Stereo parameters are used to characterize the complexity, branching degree, and symmetry of amino acid side chains; Polarizability is used to characterize the molecular polarization ability of amino acids under electrostatic conditions. Volume parameter used to characterize the standardized van der Waals volume of amino acids; Hydrophobicity parameters are used to characterize the hydrophilic or hydrophobic properties of amino acids. The isoelectric point parameter is used to characterize the charge state of amino acids under different pH conditions. Helical tendency parameter, used to characterize the probability of amino acids forming α-helices; β-sheet propensity parameter, used to characterize the probability of an amino acid forming a β-sheet; The data represented by ProtFactor-5 features include: The polarity index is used to characterize the polarity, hydrophilicity, hydrophobicity, and number of hydrogen bond donors of amino acids. Secondary structure index is used to characterize the structural tendency of amino acids in α-helix, β-sheet and random coil; Molecular volume index is used to characterize the molecular weight, volume, and side chain size of amino acids. Codon diversity index, used to characterize the genetic codon diversity and relative abundance of amino acids; The charge index is used to characterize the isoelectric point and charge distribution characteristics of amino acids. Among them, the data represented by ProtParam-7 and ProtFactor-5 features are used to characterize the physicochemical properties and conformational characteristics of amino acids, and to transform protein sequences into low-dimensional numerical feature matrices.

4. The synthetic protein design method based on protein feature engineering and diffusion models according to claim 1, characterized in that, In step 4, the AlphaFold3 tool is used to calculate the backbone distance between each generated candidate protein and the target protein sequence. AlphaFold3 is an advanced protein structure prediction platform based on deep learning. Specifically, the root mean square deviation (RMSD) is calculated to measure the difference between the two structures, assessing the average distance between protein atoms; the unit of RMSD is... , The formula for calculating the root mean square deviation (RMSD) is: （10） where N represents the number of atoms participating in the calculation; δ i represents the position deviation of the i-th atom between the two structures, that is, the Euclidean distance of the coordinates of the two after alignment; δ i 2 is the square of the position deviation; finally, taking the square root is the RMSD, the smaller the RMSD value, the closer the two structures; the larger the RMSD value, the greater the difference; During the virtual screening process, the solubility of the generated candidate proteins is compared. Specifically, the NovoPro tool is used. NovoPro is a tool specifically designed for protein solubility prediction. Based on the protein's sequence information, this tool can predict its solubility and give a solubility score between 0 and 1. A solubility value close to 1 indicates that the protein has good solubility, while a solubility value close to 0 indicates that the protein has poor solubility. Based on solubility prediction, a solubility threshold of no less than 0.8 is set to screen out protein sequences with good solubility.

5. The synthetic protein design method based on protein feature engineering and diffusion models according to claim 4, characterized in that, In step 4, for screening physicochemical properties, two different amino acid characterization methods, namely the ProtFactor-5 parameter and the ProtParam-7 parameter, are used to compare the generated candidate proteins with the target proteins. By using these two different amino acid characterization methods, the generated candidate proteins are cross-validated from different perspectives to ensure that their physicochemical properties are highly similar to those of the target protein and to accurately identify differences in key features.

6. The synthetic protein design method based on protein feature engineering and diffusion models according to claim 1, characterized in that, In step 5, during the experimental verification phase, the gene sequence of the candidate protein was first cloned into the pET28a vector, a His-tagged E. coli expression vector, and transformed into BL21(DE3) bacteria for protein expression. BL21(DE3) is an E. coli expression host strain. Under 220 rpm shaking culture, when the OD600 value reached 0.6, isopropyl-β-D-thiogalactoside IPTG was added at a final concentration of 2 mmol / L to induce protein expression. OD600 refers to the optical density of the bacterial culture medium measured at a wavelength of 600 nm, used to reflect the bacterial growth status. After culture, bacterial pellets were collected by centrifugation, and cells were lysed by sonication to release the target protein. Protein purification was then performed using Ni-NTA affinity chromatography, where Ni-NTA is a nickel-nitrogen triacetic acid (NiNTA) packing material. The target protein was separated using Ni²⁺ binding to the His tag and eluted with PBS + 500mM imidazole. PBS is a commonly used phosphate buffer containing a high concentration of imidazole, which competitively binds to Ni²⁺, effectively eluting the His-tagged protein. The purified protein was dialyzed into 1×PBS to remove impurities, and the purity of the protein was assessed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) to ensure that its purity reached more than 95% for subsequent functional testing. To evaluate the functional activity of candidate proteins, affinity assays were performed using biolayer interferometry (BLI) on a ForteBio Octet K2 instrument. The ForteBio Octet K2 is a BLI-based biomolecular interaction analyzer capable of real-time monitoring of protein-protein or protein-antibody binding kinetics. First, the purified protein was immobilized on a His sensor and then equilibrated for 120 seconds in PBST (phosphate-buffered saline + Tween 20 buffer) to reduce background noise. Next, the sensor was sequentially immersed in different concentrations of human IgG4 monoclonal antibody solutions (IgG4 is the G4 subclass of immunoglobulin) for binding experiments. Each concentration was incubated for 120 seconds, followed by dissociation in PBST for 180 seconds to determine protein binding stability. After the experiments, the sensor was regenerated using a pH 1.7 glycine solution to remove residual conjugates and for the next round of assays. Finally, the binding ability of the candidate protein to the target antibody was evaluated by calculating the dissociation constant (KD) value, where a lower KD value indicates stronger binding affinity and higher stability. Based on experimental data, candidate proteins with similar functional activity to the target protein and good structural stability were screened. Combined with the BLI affinity assay results, proteins with dissociation constant KD values ​​within the set range were selected as the final generated results.

Images