Recombinant collagen sequence analysis method based on AI deep learning

Abstract

The invention belongs to the technical field of recombinant collagen sequence analysis, and discloses a recombinant collagen sequence analysis method based on AI deep learning. Comprising the following steps: extracting a natural III-type collagen sequence; applying the natural III-type collagen sequence to the trained model to generate a thermodynamic diagram; generating an optimized natural III type collagen sequence based on the thermodynamic diagram; performing fine tuning on the AlphaFold2 model by using the fine tuning data set, and obtaining an integrated structure model by introducing spiral torsion constraint and hydroxyproline simulation parameters; loading the optimized natural III-type collagen sequence, and predicting the stability of the triple helix structure of the sequence through the structure model; and extracting a natural III-type collagen sequence which passes the stability detection, and outputting a report file by using a skin permeability prediction model to realize recombinant collagen sequence analysis based on AI deep learning.

Description

Technical Field

[0001] This invention relates to the field of recombinant collagen sequence analysis technology, and more specifically, to a recombinant collagen sequence analysis method based on AI deep learning. Background Technology

[0002] Recombinant type III collagen is an important bioactive protein with a complete triple helix structure. It possesses multiple functions, including hemostasis, promoting tissue growth, repairing damage, moisturizing and locking in moisture, and reducing pigmentation, and is widely used in the fields of beauty, medicine, and tissue engineering. However, existing recombinant type III collagen faces significant technical challenges in preparation and application: First, its sequence is susceptible to protease degradation, leading to the disintegration of the triple helix structure and rapid loss of bioactivity; second, the triple helix lacks stability, being affected by temperature and pH (e.g., it easily disintegrates at skin pH 5-8), and its thermal stability (Tm) value is often below 40℃, failing to maintain its "coiled" shape; third, its transdermal absorption efficiency is low, and its molecular rigidity and hydrophilicity are poor, resulting in a skin barrier permeability of <1 μg / cm² / h, limiting the long-term effectiveness of cosmetic products.

[0003] Traditional methods rely on manual sequence alignment or simple mutations, which cannot accurately predict cleavage sites and structural dynamics. They also ignore the protection of collagen-specific patterns (such as Gly-XY repeats and GLLG-like motifs), often damaging active motifs and resulting in reduced efficacy (e.g., a 50% decrease in hemostasis rate). Consequently, the resulting structure is not ideal. Existing AI tools such as AlphaFold can predict the structure, but if the Hyp site in the sequence is not correctly labeled, the model may predict a loose double helix instead of a compact triple helix. In transdermal absorption scenarios, this means that collagen unwinds at skin temperature (approximately 32-37°C) and cannot effectively penetrate the stratum corneum.

[0004] Based on the above problems, there is an urgent need to propose a recombinant collagen sequence analysis method based on AI deep learning. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a recombinant collagen sequence analysis method based on AI deep learning, comprising: S1: Extracted natural type III collagen sequence; S2: Construct a collagen-specific enzyme cleavage database and pre-train a dynamic prediction model for enzyme cleavage sites. Input the natural type III collagen sequence into the trained model to generate a heatmap. S3: Obtain the optimized natural type III collagen sequence based on heatmap; S4: Fine-tuning the AlphaFold2 model using the fine-tuning dataset, and simultaneously obtaining the integrated structural model by introducing helical torsion constraints and hydroxyproline simulation parameters; S5: Load the optimized natural type III collagen sequence and predict the stability of the triple helix structure of the sequence using a structural model; S6: Extract the natural type III collagen sequence that has passed the stability test, and output a report file based on skin permeability prediction.

[0006] Preferably, the method for constructing a collagen-specific enzyme digestion database includes: Collect the enzymes required for the target and extract known cleavage site data for each enzyme; the known cleavage site data includes substrate sequence, cleavage site location, and experimental validation data; The search scope was limited to human or other mammalian sources. Type III collagen sequences and cleavage site annotations were extracted, and collagen-specific patterns in the type III collagen sequences were marked. The substrate sequence of each enzyme is compared with the sequence of the collagen-specific pattern. If the similarity is greater than the expected value, it is retained; otherwise, the corresponding substrate sequence is deleted. For the preserved sequence, a window sequence of 8 amino acids before and after the cleavage site is labeled, and a pH-dependent cleavage rate is introduced to obtain positive samples; negative samples are generated by randomly perturbing the positive samples. For positive and negative samples, after similarity filtering, the final dataset is obtained, which is in the form of a table. The columns in the table are in the form of {enzyme ID, sequence, cleavage site location, context label, pH-dependent cleavage rate, positive or negative label}.

[0007] Preferably, inputting the natural type III collagen sequence into the trained model includes: The dynamic prediction model for protease cleavage sites includes a two-layer learning structure: the first layer is a bidirectional long short-term memory network layer, and the second layer is a multi-head attention layer. The output of the first layer is the input of the second layer, and a final output layer is added. After standardizing the collagen-specific enzyme cleavage database, it was used to pre-train a dynamic prediction model for protease cleavage sites. The natural type III collagen sequence is converted into a numerical embedding vector and input into the trained model. A sliding window is used to scan segment by segment to obtain the probability of the cut site at each position in the sequence.

[0008] Preferably, the method for generating a heatmap includes: Define a first risk threshold and a second risk threshold. If the probability of the cut site is greater than the first risk threshold, it is marked as a high-risk area. If the probability of the cut site is greater than or equal to the first risk threshold and is greater than the second risk threshold, it is marked as a medium-risk area. If the probability of the cut site is less than or equal to the second risk threshold and contains an active motif, it is marked as a protected area. For natural type III collagen sequences, risk regions are marked position by position: high-risk regions are marked as non-functional regions, medium-risk regions are marked as potential functional regions, and protected regions are marked as active motif regions. After image processing, a pixelated probability map is output.

[0009] Preferably, the step of predicting the stability of the sequence triple helix structure through a structural model includes: Set the AlphaFold2 model as the starting point for fine-tuning, construct a fine-tuning dataset and use it to fine-tune the AlphaFold2 model. Fine-tuning is complete when the model's accuracy in testing reaches the standard. For the fine-tuned model, a helical torsion constraint is forcibly applied, and hydroxyproline simulation parameters are introduced to obtain the integrated structural model; The optimized natural type III collagen sequence is loaded and input into the structural model. The inter-chain hydrogen bond density and melting temperature are then output. If the inter-chain hydrogen bond density and melting temperature do not meet the expected standards, the model is rolled back and re-optimized until the expected standards are met.

[0010] Preferably, the method for fine-tuning the dataset and using it to fine-tune the AlphaFold2 model includes: Collagen crystal structures were collected from the PDB database, and amino acid sequences, 3D coordinates, and annotations were extracted as the basis of the dataset. Samples containing collagen-specific patterns were loaded from a collagen-specific enzyme digestion database. If the similarity between the samples and the sequences in the dataset reached the expected level, the information from the samples was used to supplement the dataset, forming a fine-tuned dataset. Freeze the bottom layer of the AlphaFold2 model, only unfreeze the top layer, set the fine-tuning parameters, and use the fine-tuning dataset for iterative training.

[0011] Preferably, the method of forcibly applying helical torsional constraints and introducing hydroxyproline simulation parameters includes: Based on the collagen-specific mode, torque parameters, target torsion angle, and deviation threshold are set. The custom script of the AlphaFold2 model is modified to select the inter-chain dihedral angle. If the torsion deviates and the energy cost increases, the coordinates are forcibly adjusted to curl towards the target torsion angle. The total energy is periodically detected, and if it exceeds the expected threshold, the torque parameters are automatically adjusted until the total energy is minimized and then stopped. Based on a collagen-specific enzyme digestion database, hydroxyproline simulation parameters were defined, and potential hydroxyproline sites were labeled in the optimized natural type III collagen sequence. Then, the force field was modified to add hydrogen bond bonuses to the labeled sites.

[0012] Preferably, the optimized natural type III collagen sequence includes: Annotate each position in the natural type III collagen sequence, including amino acid sequence, cleavage site probability, labeled risk regions, and additional annotations; Based on the annotation results and heatmaps, high-risk areas were deleted or replaced; medium-risk areas were replaced; and protected areas retained their original sequences and added protective modifications.

[0013] Preferably, the method for predicting and outputting a report file based on skin permeability includes: The net charge, hydrophilicity index, and molecular rigidity of natural type III collagen sequences were extracted, and pH-dependent cleavage rates were integrated from a collagen-specific enzyme cleavage database. The skin permeability prediction incorporates a dual-objective optimization mechanism: maximizing the retention of the natural transdermal domain and minimizing structural perturbations. The output transdermal optimized sequence is then analyzed to determine whether the transdermal coefficient and the proportion of circular dichroism α-helices meet the output criteria. If not, the optimization is rolled back; otherwise, a report file is output, including the transdermal optimized sequence, transdermal coefficient, and proportion of circular dichroism α-helices.

[0014] An AI deep learning-based application for recombinant collagen sequence analysis includes: Memory, used to store computer programs; The processor is used to execute computer programs to implement the aforementioned AI deep learning-based recombinant collagen sequence analysis method.

[0015] The technical effects and advantages of the AI ​​deep learning-based recombinant collagen sequence analysis method of this invention are as follows: By constructing a collagen-specific enzyme cleavage database and integrating multi-source bioinformatics data, we can efficiently process heterogeneous sequence features, overcoming the limitations of traditional methods in single enzyme cleavage site analysis and significantly improving data utilization and prediction accuracy. Through a deep learning model combined with attention mechanisms and transfer learning, we constructed a dynamic prediction model for protein cleavage sites with strong generalization capabilities. This model can adapt to the optimization needs of different collagen variants and degradation environments, solving the problem of strong dependence on specific sequences in existing methods.

[0016] Risk region labeling and heatmap visualization techniques were introduced to assess the degradation uncertainty of the predicted results, generating a risk distribution map. This provided a reliable structural protection basis for sequence optimization decisions and significantly reduced the risk of functional loss. A dual-objective evolutionary algorithm was used to refine the initially optimized sequences, comprehensively considering stability indicators and transdermal function, achieving a balance between degradation resistance and bioactivity, and greatly improving the scientific rigor and practicality of the optimized sequences.

[0017] By employing techniques such as feature extraction and constraint simulation, the computational complexity of sequence processing and structure prediction is reduced, thereby improving the efficiency and scalability of model training. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of the AI ​​deep learning-based recombinant collagen sequence analysis method in this invention. Figure 2 This is a structural diagram of an electronic device for recombinant collagen sequence analysis based on AI deep learning, as described in this invention. Detailed Implementation

[0019] 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.

[0020] Example 1 Please see Figure 1 This embodiment provides a method for recombinant collagen sequence analysis based on AI deep learning, which may specifically include the following steps: Extract the sequence of natural type III collagen; Natural type III collagen sequences, such as the human COL3A1 sequence, are 1466 amino acids long and contain Gly-XY repeats and GLLG-like motifs such as "GGLLGADFAGD".

[0021] Current recombinant type III collagen sequence analysis suffers from data limitations and inaccurate predictions. Specifically, traditional methods rely on general databases (such as MEROPS) and lack specific data for type III collagen. This results in an inability to accurately capture the restriction enzyme sites and contextual features (such as pH-dependent cleavage rates) of collagen-specific patterns (such as Gly-XY repeats and GLLG-like motifs), leading to excessively high degradation prediction bias, insufficient stability, and low transdermal absorption. Therefore, the following design was proposed: A collagen-specific enzyme cleavage database was constructed, and a dynamic prediction model for enzyme cleavage sites was pre-trained. Natural type III collagen sequences were input into the trained model to generate heatmaps. Methods for constructing collagen-specific enzyme digestion databases include: Collect the enzymes required for the target and extract known cleavage site data for each enzyme; the known cleavage site data includes substrate sequence, cleavage site location, and experimental validation data; Specifically, the enzymes are identified and their types are listed, including 10 collagen degradation-related enzymes such as matrix metalloproteinase-1, matrix metalloproteinase-2, matrix metalloproteinase-3, matrix metalloproteinase-7, matrix metalloproteinase-8, matrix metalloproteinase-9, and matrix metalloproteinase-13, as well as trypsin, chymotrypsin, elastase, collagenase, cathepsin (such as cathepsin K), neutral protease, acidic protease, serine protease, cysteine ​​protease, and metalloproteinase.

[0022] Use a browser or script to query the MEROPS database, confirm the available data entries for each enzyme, and obtain a preliminary enzyme list file (e.g., enzymes_list.txt) containing enzyme names and preliminary entry counts.

[0023] Using the MEROPS database API or a batch download tool, download known cleavage site data for each enzyme, including substrate sequence, cleavage site location, and experimental validation data (experimental validation data refers to reliable data on cleavage sites verified through experiments; it is the cornerstone of building a high-quality database and training reliable AI models, such as catalytic efficiency constants, pH, temperature, etc.; all data required in this design can be provided). Prioritize records with experimental evidence, ensuring coverage of at least 100 per enzyme. Obtain the original CSV file (e.g., merops_data.csv), containing approximately 5000 records.

[0024] The search scope was limited to human or other mammalian sources. Type III collagen sequences and cleavage site annotations were extracted, and collagen-specific patterns in the type III collagen sequences were marked. Search and download type III collagen sequences from the UniProt database, specifying human or other mammalian origin using the keyword "COL3A1", targeting over 5000 sequences (including variants and fragments). Obtain the sequence file (e.g., collagen_sequences.fasta) containing collagen-specific sequences, outputting the data in CSV format.

[0025] Collagen-specific patterns refer to unique, recurring amino acid sequences within the collagen protein sequence. These patterns form the basis of the collagen triple helix structure and influence its sensitivity to enzymes. In collagen (especially type III), the most typical collagen-specific pattern is the "Gly-XY repeat" (where Gly is glycine, X is usually proline or lysine, and Y is usually hydroxyproline or other amino acids). This pattern repeats every three amino acids, forming a triple helix "coiled" structure. It is rare in other proteins and is primarily found in the collagen family. This makes collagen susceptible to degradation by specific enzymes, such as matrix metalloproteinases, especially near the GLLG-like motif (Gly-Leu-Leu-Gly).

[0026] The substrate sequence of each enzyme is compared with the collagen-specific sequence. If the similarity is greater than the expected value, it is retained; otherwise, the corresponding substrate sequence is deleted. Sequence alignment tools such as BLAST or Clustal Omega can be used, and expected values ​​can be set based on expert experience. If the substrate sequence is a short fragment, only matching regions are compared. The goal is to retain at least 5000 collagen-specific records after filtering.

[0027] Example: For a MEROPS substrate sequence “GGLLGADFAGD” (possibly from a cleavage site of an enzyme), the tool checks if it has a similarity >80% to any part of P02461 (e.g., if there is a similar fragment “GGLQGADFAGD” in P02461 with 85% similarity, it is retained). Based on the alignment results, records with a similarity >80% are retained; those that do not match are discarded (e.g., a substrate sequence from a non-collagen source with only 50% similarity is excluded).

[0028] Context tags are added to the preserved sequences. Context tags are characteristic expressions of the sequences, including but not limited to the window sequence of 8 amino acids before and after the cleavage site and the pH-dependent cleavage rate. The pH-dependent cleavage rate is extracted from the literature or estimated using tools at the cleavage rate of the site at pH 5-8 [it can be estimated by weighted average of amino acid pKa values ​​(e.g., cleavage rate = basal rate multiplied by ionization adjustment factor)]. After that, the data at this time is recorded as positive samples, and negative samples are generated by randomly perturbing the positive samples. Assuming approximately 7,000 positive samples (with verified cleavage sites) are retained and approximately 3,000 negative samples (without cleavage sites) are generated, the total number of samples is ensured to reach 10,000, and the enzyme types are balanced (at least 500 samples for each enzyme).

[0029] For positive and negative samples, remove sequences with high similarity, update the data, and obtain the final dataset. The structure is a table, and the columns are in the form of {enzyme ID, sequence, cleavage site location, context label, pH-dependent cleavage rate, positive / negative label}. Similarity filtering removes sequences with excessively high similarity. Clustering tools can be used to remove sequences with similarity greater than 95%, preserving diversity. Then, the collagen specificity of 1000 samples is manually or automatically validated (e.g., checking for Gly-XY patterns). A variant generation tool is used to generate 2-3 variants for each positive sample (maintaining collagen-specific patterns, such as random point mutations) to augment the data, ensuring the final augmented dataset has a stable total of 10,000 samples.

[0030] The dataset is exported in HDF5 format and stored in an SQLite database, supporting queries for enzyme IDs. All data is limited to human or other mammalian sources.

[0031] By constructing a database to provide collagen-specific and balanced training data, the problems of data noise and poor generalization are addressed. Filtering substrate sequences with similarity greater than the expected value and adding contextual labels significantly improves data utilization and captures the influence of collagen-specific patterns. Positive and negative sample generation and mutation augmentation ensure a balanced and diverse dataset, reducing overfitting, providing data support for triple-helix stability and transdermal absorption, avoiding damage to active motifs, and overall enhancing the bioactivity of recombinant collagen (such as improved hemostasis).

[0032] Current recombinant type III collagen sequence analysis cannot process sequence context (such as pH-dependent cleavage rate and collagen-specific Gly-XY repeat patterns) in real time, leading to significant prediction biases for protease degradation sites and easy disintegration of the triple helix structure. While existing AI tools can predict structure, they lack position-by-position dynamic scanning and intuitive visualization (such as heatmaps) of cleavage site probabilities, failing to highlight high-risk, intermediate-risk, and protected areas. This results in optimization processes (such as cleavage site removal) damaging active motifs (such as GLLG-like motifs), affecting bioactivity (such as low transdermal absorption efficiency). Furthermore, the model cannot effectively integrate multidimensional features (such as pH-dependent cleavage rate), limiting its application in skin environments (pH 5-8). Therefore, the following design is proposed: The protease cleavage site dynamic prediction model includes a two-layer learning structure. The first layer is a bidirectional long short-term memory network layer, with the input being a numerical embedding vector. The second layer is a multi-head attention layer, with the output being the probability of the cleavage site at each position in the input sequence. The output of the first layer is used as the input of the second layer, and a final output layer is added. After standardizing the collagen-specific enzyme cleavage database, a dynamic prediction model for protease cleavage sites was pre-trained. Standardization involves converting all sequences into padding vectors of the same length (e.g., a maximum length of 100 amino acids, with short sequences padded with zeros), converting labels to binary (e.g., positive samples = 1, negative samples = 0), and integrating multidimensional features of the database (e.g., pH-dependent cut rates) as additional input channels. The standardized database is then divided into a training set (80%, 8000 samples) and a validation set (20%, 2000 samples).

[0033] The bidirectional long short-term memory network layer is configured with 256 hidden units, 2 layers, and bidirectional=True, with numerical embedding vectors as input. The multi-head attention layer is configured with 8 heads, with the output of the first layer as input, and a weighted representation is calculated. The final output layer is a fully connected linear layer with an output dimension of 1, representing the probability of the cut point. Random or pre-trained embeddings are used to initialize the model's weights. The loss function is set to binary cross-entropy (calculating the difference between the predicted probability and the true label for positive and negative labels). The optimizer is Adam (learning rate 0.001). The number of training epochs is 20, and the batch size per epoch is 32.

[0034] The batch of data is input into the model. The first layer, a bidirectional long short-term memory network, processes the sequence context and generates hidden states. The second layer, a multi-head attention layer, weights the hidden states (calculates the soft-maximization weights of the query vector, key vector, and value vector, with the weights calculated based on the dot product of the query vector and key vector after normalization), and outputs the predicted probabilities.

[0035] After each round, the accuracy is calculated using the validation set [accuracy is calculated as (number of correctly predicted positions / total number of positions) × 100%]. If the accuracy is less than 85%, the learning rate is adjusted (e.g., reduced to 0.0005) and training continues for an additional number of rounds (up to 5 rounds). If the training accuracy is >95% but the validation accuracy is <80%, regularization such as L2 penalty is added. Finally, if the validation accuracy is greater than or equal to 85%, the model parameters are saved (e.g., trained_model.pt).

[0036] The natural type III collagen sequence is converted into a numerical embedding vector, and a trained model is applied. A sliding window (window size 4-10 amino acids, step size 1) is used to scan segment by segment to generate the hidden state at each position. Then, the hidden state is weighted through an attention mechanism to highlight key regions. The weighted output is processed using a sigmoid activation function to obtain the probability of the cut site at each position in the sequence.

[0037] Define a first risk threshold and a second risk threshold. If the probability of the cut site is greater than the first risk threshold, it is marked as a high-risk area. If the probability of the cut site is greater than or equal to the first risk threshold and is greater than the second risk threshold, it is marked as a medium-risk area. If the probability of the cut site is less than or equal to the second risk threshold and contains an active motif, it is marked as a protected area. For natural type III collagen sequences, risk regions are marked position by position: high-risk regions are marked as non-functional regions, medium-risk regions are marked as potential functional regions, and protected regions are marked as active motif regions. After image processing (such as using a convolutional neural network with a kernel size of 3x3), a pixelated probability map is output (color-coded: red for high-risk areas, yellow for medium-risk areas, and green for protected areas) to visualize the distribution of risk areas.

[0038] The first and second risk thresholds can be preset using expert experience. The specific image processing method is as follows: First, the probability of the cut points in the sequence (e.g., an array of length N, where each element represents a probability of 0-1) is converted into a two-dimensional matrix. Next, a color encoding function is used to map the probability values ​​to colors (red for high-risk areas > first risk threshold, e.g., >0.8; yellow for medium-risk areas 0.5-0.8; green for protected areas ≤0.5 and containing active motifs). If enhancement is needed (e.g., smoothing noise or highlighting region boundaries), a one-dimensional or two-dimensional convolutional neural network (kernel size 3x3, stride 1, padding mode 'same') is applied to convolve the matrix (e.g., an average filtering kernel calculates the local average probability). This step outputs a smoothed matrix. Finally, a plotting function is used to render a pixelated probability map (each pixel corresponds to the color of the sequence position), and axis labels are added (e.g., the x-axis represents the sequence position, and the y-axis represents the probability or risk type). The image is then saved as an image file.

[0039] Active motifs are short, biologically active amino acid sequences within a collagen protein. They are responsible for collagen's physiological functions (such as hemostasis, tissue repair, and moisturizing) and play a crucial role in the triple helix structure. Active motifs are regions that need to be protected to prevent optimization processes (such as removing cleavage sites) from disrupting the protein's biological activity (e.g., transdermal absorption and damage repair efficacy).

[0040] Suppose we are processing a short sequence fragment of recombinant type III collagen, with the input sequence fragment being "Gly-Pro-Hyp-Gly-Leu-Leu-Gly-Ala-Asp-Phe-Ala-Gly-Asp" (13 amino acids, containing the Gly-XY repeat and the GLLG active motif "Gly-Leu-Leu-Gly"). The probabilities of the cleavage sites are expressed as vectors [0.3, 0.4, 0.6, 0.9, 0.85, 0.7, 0.92, 0.55, 0.2, 0.45, 0.65, 0.1, 0.75] (corresponding to each position).

[0041] Define the first risk threshold as 0.8 and the second risk threshold as 0.5.

[0042] digit-by-digit annotation: Location 4 (probability 0.9 > 0.8): Marked as a high-risk area (non-functional region, possibly susceptible to matrix metalloproteinase cleavage).

[0043] Location 5 (0.85>0.8): High-risk area.

[0044] Location 6 (0.7:0.8≥0.7>0.5): Medium-risk area (potential functional area).

[0045] Location 7 (0.92>0.8): High-risk area.

[0046] Location 3 (0.6:0.8≥0.6>0.5): Medium risk zone.

[0047] Positions 1-2 and 8-13: Check the portion with a probability ≤ 0.5 (e.g., position 1 = 0.3 ≤ 0.5).

[0048] The results were labeled as follows: Positions 1-2 (low risk, unlabeled); Position 3 (medium risk area); Positions 4-5, 7 (high risk area); Position 6 (medium risk area); Positions 8-13 (protected area, if minor active motifs are detected; otherwise low risk).

[0049] Convert the probability vector into a 2D matrix (1 row and 13 columns).

[0050] Apply color mapping: red (high-risk areas, such as locations 4-5 and 7); yellow (medium-risk areas, such as locations 3 and 6); green (protected areas, such as parts of locations 8-13).

[0051] The matrix is ​​smoothed using a convolutional neural network (kernel size 3x3), and the average neighbor probability is averaged (e.g., 0.7 at position 6 becomes 0.75 after smoothing). An image is generated (x-axis = sequence positions 1-13, y-axis = probability 0-1), with red blocks highlighting high-risk clusters, yellow indicating intermediate-risk transitions, and green protecting active motif regions.

[0052] By employing a pre-trained bi-layer learning structure (a bidirectional long short-term memory network layer and a multi-head attention layer) combined with a standardized collagen-specific enzyme digestion database, this system enables positional cleavage site probability calculations for natural type III collagen sequences, improving accuracy. A sliding window scanning method captures contextual dependencies (such as the influence of collagen-specific patterns). It visually displays the distribution of risk regions, facilitating the identification and protection of active motifs (such as GLLG-like motifs), reducing the risk of loss of function during optimization, and enhancing overall biological activity (such as improved hemostasis). It is applicable to various sequence variants.

[0053] Existing recombinant type III collagen structure prediction methods suffer from bias and insufficient stability assessment. Specifically: the traditional AlphaFold2 model has limited collagen structure data in its training data, leading to inaccurate predictions of triple-helix coil morphology (such as the collagen-specific Gly-XY repeat pattern), neglecting the influence of hydroxyproline sites, and failing to accurately simulate the melting risk at pH 5-8. The lack of mandatory constraints and parameter introduction results in triple-helix disintegration of optimized sequences (such as after deletion of high-risk regions), thermal stability (Tm < 40℃), inability to assess interchain hydrogen bond density and melting temperature, and impacts transdermal absorption and bioactivity (such as reduced hemostatic efficacy). Therefore, the following design is proposed: Set the AlphaFold2 model (version 2 or higher) loaded from the official source as the starting point for fine-tuning, construct the fine-tuning dataset, and use transfer learning to fine-tune the AlphaFold2 model. Fine-tuning is complete when the model's accuracy reaches the standard. The accuracy standard can be set as pLDDT>80 or RMSD<1.5A, calculated based on the validation set of the collagen-specific enzyme digestion database. For the fine-tuned model, a helical torsion constraint is forcibly applied, and hydroxyproline simulation parameters are introduced to obtain the integrated structural model. The integrated structural model is exported as a PDB format, which can be imported into PyMOL for visualization and used for downstream transdermal coefficient calculation. The optimized natural type III collagen sequence is loaded, and the interchain hydrogen bond density is obtained through structural modeling. The melting temperature is estimated. If the interchain hydrogen bond density and melting temperature Tm do not meet the expected standards (density > 1.5 / residue, Tm > 42℃), the high-risk region deletion strategy is rolled back and adjusted first, with a maximum of 3 iterations, until the expected standards are met. Then, the interchain hydrogen bond density and melting temperature Tm are used as quantitative indicators and output. The unwinding temperature Tm was estimated using molecular dynamics simulations (running a 10-nanosecond trajectory using GROMACS software) with a target value greater than 42 degrees Celsius.

[0054] Methods for constructing fine-tuning datasets and using transfer learning to fine-tune the AlphaFold2 model include: Collagen crystal structures were collected from the PDB database, and amino acid sequences, 3D coordinates, and annotations were extracted as the basis of the dataset. Samples containing collagen-specific patterns were loaded from the collagen-specific enzyme digestion database. If the similarity between the sample and the sequence in the dataset reached the expected level (a preset similarity threshold, determined based on experimental measurements or historical data analysis), the information from the sample was used to supplement the dataset and form a fine-tuned dataset. Specifically, a PDB file is a standard format file that contains the coordinates of a protein's 3D structure (the x, y, z positions of each atom in space), like a 3D blueprint, used to visualize and simulate the triple-helix 3D "curled" structure; Access the PDB database using a browser or script by entering the keyword "collagen type III". Search results will display relevant PDB IDs. For each PDB file, the amino acid sequence is read from the file. The x, y, and z coordinates of each atom are read from the "ATOM" line of the PDB file. Annotations provide additional explanatory information, such as "hydroxyproline site," which indicates the position in the sequence where proline (Pro) is modified to hydroxyproline (Hyp), which is crucial for enhancing hydrogen bonds.

[0055] The supplementary information includes: adding the location of the cleavage site (e.g., in the core sequence Gly-Pro-Hyp-Gly-Leu-Leu-Gly, add positions 5-7 as matrix metalloproteinase cleavage sites); adding labeling information (e.g., pH-dependent cleavage rate = 0.7, indicating easy degradation at pH 5-8); adding multidimensional features (e.g., hydrophobicity = -0.5), and marking high stability (e.g., labeling the Gly-XY repeat region of positive samples as "stable triple helix").

[0056] Freeze the bottom layers of the AlphaFold2 model (such as the first 50% of layers, to retain general folding knowledge), unfreeze only the top layers (such as the remaining last 50% of layers), set fine-tuning parameters, and use the fine-tuning dataset for iterative training.

[0057] Example of setting fine-tuning parameters: Define the loss function as mean squared error; the optimizer as Adam (learning rate 0.0001); the number of training epochs as 10; and the batch size as 8.

[0058] Load the training batch of the fine-tuning dataset, predict the triple helix structure through forward propagation; calculate the loss; update the weights of the unfrozen layer through backpropagation. If the loss does not decrease, adjust the learning rate or increase the number of rounds (up to 5 rounds). Stop fine-tuning when the test accuracy reaches the standard.

[0059] Methods that forcibly apply helical torsional constraints and introduce hydroxyproline simulation parameters include: Based on the collagen-specific model, torsion parameters (a numerical parameter for interchain dihedral angles that defines the energy cost required when a molecule deviates from the target torsion angle), the target torsion angle, and the deviation threshold are set. The custom script for the AlphaFold2 model is modified (e.g., the energy function in run_alphafold.py). The interchain dihedral angle (such as the Cα-Cβ bond in Gly-Pro-Hyp) is selected. If the torsion deviation (referring to the difference between the calculated actual dihedral angle value and the preset ideal target value based on known structural knowledge during protein structure prediction or optimization) is considered, the deviation threshold is set. If the energy cost increases due to differences between iterations (e.g., E=0.5kcal / mol when deviating by 10 degrees), the coordinates are forced to curl towards θ0; the total energy is periodically detected (total energy refers to the sum of the energy costs of all iterations, such as checking the total energy every 1000 steps); if it exceeds the expected threshold (e.g., total energy > 5kcal / mol), the torque parameter k is automatically adjusted (increased to 0.2) and the run is restarted until the energy is minimized (the curled structure is stable) and then it stops; if the simulated trajectory continues to diverge (e.g., energy > 10kcal / mol), the integrity of the input sequence is checked and alternative parameters (e.g., reducing the learning rate) are used.

[0060] Interchain dihedral angles are angular parameters used in protein structure simulations, referring to the bond angles (e.g., phi / psi angles, where phi is the C-N bond rotation and psi is the Cα-C bond rotation) between multiple protein chains (such as the triple helix of collagen), measured in degrees. By selecting constraint positions, the stability of the triple helix structure is ensured.

[0061] The energy function is given by the energy cost E = (1 / 2) × k × (θ - θ0). 2 Where k is the torque parameter (representing strength, in kcal / mol / degree^2, e.g., k=0.1), θ is the current torsion angle (degrees), and θ0 is the target torsion angle. If the database shows that the stable torsion angle of the GLLG region is 100-110 degrees, then set θ0=105 degrees, the deviation threshold=5 degrees (the cost increases exponentially when it exceeds the threshold), and the total energy=original energy+costE; the inter-chain dihedral angle refers to the phi / psi angle).

[0062] Setting torque parameters, target torsion angle, and deviation threshold based on collagen-specific patterns involves adjusting these parameters according to the type of collagen-specific pattern (e.g., Gly-XY repetition). Based on a collagen-specific enzyme digestion database, simulated parameters for hydroxyproline were defined (this refers to extracting data from the database to set simulated values. For example, if the hydroxyproline equilibrium constant is greater than 10 from positive samples in the database, then parameters such as a free energy change of -1.5 kcal / mol are defined to indicate enhanced stability from hydroxyproline). Potential hydroxyproline sites were labeled in the optimized natural type III collagen sequence. Then, the force field was modified (e.g., modifying the model's force field file to add extra terms for hydroxyproline sites, making the model prefer stable conformations). Hydrogen bond bonuses were added to the labeled sites [this refers to adding a positive energy bonus to potential hydroxyproline sites to encourage the model to form more hydrogen bonds, improve stability (e.g., increase Tm), and ensure that the triple helix does not disintegrate after optimization. Adding a hydrogen bond bonus, such as reducing the total energy by 0.5 kcal / mol / bond, simulates the effect of hydroxyproline enhancing the hydrogen bond network].

[0063] By fine-tuning through transfer learning and introducing constraints / parameters, an integrated structural model is generated, improving the prediction accuracy of the collagen triple helix and effectively capturing the influence of collagen-specific patterns (such as the coil stability of Gly-XY repeats). By using interchain hydrogen bond density (if the expected standard is set to be greater than 1.5 / residue) and melting temperature Tm (if the expected standard is set to be greater than 42℃) as quantitative indicators, a rollback mechanism (such as re-optimization when standards are not met) is supported, ensuring that the optimized sequence maintains bioactivity (such as improved transdermal absorption) under pH 5-8 conditions. Molecular dynamics simulations (10 nanosecond trajectories) combined with parameter adjustments reduce computational bias, providing a reliable basis for downstream transdermal optimization and achieving a structure-function balance.

[0064] Based on the structural disruption and activity loss issues encountered in existing recombinant type III collagen sequence optimization, the following design was proposed: The optimized natural type III collagen sequence includes: Label each position of the natural type III collagen sequence, including amino acid sequence, cleavage site probability, labeled risk region (high-risk, medium-risk, and protected), and additional annotations (such as whether it contains GLLG-like motifs with active motifs). Based on the annotation results and heatmap, high-risk regions are deleted (removing short fragments <10 amino acids, if they are flexible linkers) or replaced (e.g., replacing Lys with Ala, using a BLOSUM62 matrix to ensure conservation); intermediate-risk regions are replaced (e.g., replacing Arg with Lys) to avoid disrupting collagen-specific patterns (e.g., Gly-XY repeats); protected regions retain the original sequence and add protective modifications (e.g., inserting glycosylation sites Asn-X-Ser / Thr to shield against potential degradation); the integrity of the optimized sequence is checked (integrity criteria: length change <5%, retention of at least 95% of collagen-specific patterns), and the loaded sequence is compared with the original annotation results: checking whether high-risk regions have been deleted / replaced, whether intermediate-risk regions have been optimized, and whether protected regions (e.g., GLLG-like motifs) have been completely preserved. If inconsistencies are found (e.g., the green protected area shown in the heatmap is missing), the sequence is rolled back and re-optimized. The optimized sequence is exported in FASTA format, and a log file is generated to record change details (e.g., a list of replaced amino acids).

[0065] When deleting or replacing high-risk areas, replacement should be preferred over deletion. If the impact of replacement on collagen-specific patterns is greater than 5%, then roll back to minimum deletion. By using positional annotation and heatmap guidance, high-risk areas are deleted / replaced, medium-risk areas are conservatively adjusted, and protected areas are modified (such as glycosylation). This ensures that the optimized sequence retains the biological activity of active motifs (such as GLLG-like motifs) while avoiding disruption of collagen-specific patterns and enhancing anti-degradation capabilities (such as extended half-life).

[0066] Existing optimization methods for recombinant type III collagen sequences suffer from a conflict between transdermal absorption and structural stability. Specifically, neglecting the net charge, hydrophilicity index, and molecular rigidity of the sequence makes it difficult for the molecule to cross the skin barrier (permeability <1 μg / cm² / h), especially prone to degradation in pH 5-8 environments, limiting cosmetic applications (e.g., reduced moisturizing and water-locking effects). Existing optimization methods (such as inserting transdermal domains) often disrupt triple helix stability (e.g., α-helix percentage <92%), causing sequence collapse and failing to retain the bioactivity of active motifs (such as GLLG-like motifs). Furthermore, the lack of dual-target detection and rollback mechanisms leads to low iteration efficiency. Therefore, the following design is proposed: The net charge (calculated by subtracting the total amount of aspartic acid and glutamic acid from the total amount of arginine and lysine), hydrophilicity index (GRAVY score less than -0.4), and molecular rigidity (rigidity vector based on principal component analysis) of the optimized natural type III collagen sequence were extracted and integrated with pH-dependent cleavage rates from the collagen-specific enzyme cleavage database. A dual-objective optimization was added to the pre-constructed skin permeability prediction. This optimization included maximizing the retention of native transdermal domains (e.g., GLLG motif similarity greater than 90%) and minimizing structural perturbations (e.g., RMSD less than 0.15 nm). The output transdermal optimized sequence was then checked to ensure that the transdermal coefficient and the proportion of circular dichroism α-helices met the output conditions, i.e., whether the values ​​were greater than expected (e.g., transdermal coefficient > 1.5 × 10⁻⁶). -6 (cm / s, percentage >92%). If this condition is not met, roll back for optimization. Rollback will prioritize adjusting the replacement strategy for the medium-risk area, with a maximum of 3 rollbacks. If the condition is still not met, manual intervention will be triggered. Otherwise, a report file will be output, including the transdermal optimized sequence, transdermal coefficient, and circular dichroism α-spiral percentage. The report file will be exported as a PDF or CSV file, including visualization charts (such as α-spiral percentage curves).

[0067] Circular dichroism α-helix percentage refers to the percentage of α-helices in the secondary structure of a protein measured by circular dichroism spectroscopy (a percentage >92% indicates structural integrity). The α-helix percentage is commonly used to quantify the proportion of coiled structures and to verify the triple-helix stability of optimized sequences (e.g., a high percentage indicates resistance to disintegration and support for transdermal absorption).

[0068] The transdermal permeability coefficient is an indicator that quantifies the rate at which molecules penetrate the skin (unit: cm / s, >1.5 × 10⁻⁶). -6 (Indicates high penetration). Used to evaluate the skin absorption efficiency of the sequence (e.g., a high coefficient ensures that collagen penetrates rapidly without degradation at pH 5-8).

[0069] The report documents are used to guide industrial production (such as selecting cosmetic formulations based on transdermal penetration coefficient).

[0070] Specifically, the pre-built skin permeability prediction model uses the Potts-Guy equation to predict: transdermal coefficient Kp = 10^(0.71×logP - 0.0061×MW - 2.74), where logP is the partition coefficient (calculated based on the hydrophilicity index; hydrophilic sequences have low logP); MW is the molecular weight (optimized <500Da is beneficial for absorption). If active motifs (such as GLLG-like motifs) are included, a correction factor (transdermal coefficient Kp × 1.2, to simulate enhanced penetration by biological activity) is added. Then, it is input into a regression model (such as random forest or neural network) to predict and refine the transdermal coefficient Kp. If Kp < the target value, a generative adversarial network is used to generate sequence variants (the generator produces a hydrophilic enhanced version, and the discriminator evaluates the authenticity), and Kp is recalculated.

[0071] By calculating net charge, hydrophilicity index, and molecular rigidity, and integrating pH-dependent shear rates, a transdermal optimized sequence that meets production requirements is generated, improving the balance between absorption efficiency and stability. Dual-objective optimization (maximizing the retention of natural transdermal domains and minimizing structural perturbations) and a rollback mechanism ensure automatic adjustments when production requirements are not met, reducing ineffective optimizations and shortening the development cycle.

[0072] The output and process of each of the above steps can be visualized.

[0073] Example 2 Please see Figure 2 The present invention also provides an electronic device. Figure 2 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of the invention. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the AI ​​deep learning-based recombinant collagen sequence analysis method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be a computer.

[0074] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this invention, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0075] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0076] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the AI ​​deep learning-based recombinant collagen sequence analysis method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0077] Furthermore, the present invention also provides a computer-readable storage medium for storing a computer program; wherein, when executed by a processor, the computer program implements the aforementioned disclosed method for recombinant collagen sequence analysis based on AI deep learning. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0078] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for recombinant collagen sequence analysis based on AI deep learning, characterized in that, include: S1: Extracted natural type III collagen sequence; S2: Construct a collagen-specific enzyme cleavage database and pre-train a dynamic prediction model for enzyme cleavage sites. Input the natural type III collagen sequence into the trained model to generate a heatmap. S3: Obtain the optimized natural type III collagen sequence based on heatmap; S4: Fine-tuning the AlphaFold2 model using the fine-tuning dataset, and simultaneously obtaining the integrated structural model by introducing helical torsion constraints and hydroxyproline simulation parameters; S5: Load the optimized natural type III collagen sequence and predict the stability of the triple helix structure of the sequence using a structural model; S6: Extract the natural type III collagen sequence that has passed the stability test, and output a report file based on skin permeability prediction.

2. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 1, characterized in that, The method for constructing a collagen-specific enzyme digestion database includes: Collect the enzymes required for the target and extract known cleavage site data for each enzyme; the known cleavage site data includes substrate sequence, cleavage site location, and experimental validation data; The search scope was limited to human or other mammalian sources. Type III collagen sequences and cleavage site annotations were extracted, and collagen-specific patterns in the type III collagen sequences were marked. The substrate sequence of each enzyme is compared with the sequence of the collagen-specific pattern. If the similarity is greater than the expected value, it is retained; otherwise, the corresponding substrate sequence is deleted. For the preserved sequence, a window sequence of 8 amino acids before and after the cleavage site is labeled, and a pH-dependent cleavage rate is introduced to obtain positive samples; negative samples are generated by randomly perturbing the positive samples. For positive and negative samples, after similarity filtering, the final dataset is obtained, which is in the form of a table. The columns in the table are in the form of {enzyme ID, sequence, cleavage site location, context label, pH-dependent cleavage rate, positive or negative label}.

3. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 2, characterized in that, The step of inputting the natural type III collagen sequence into the trained model includes: The dynamic prediction model for protease cleavage sites includes a two-layer learning structure: the first layer is a bidirectional long short-term memory network layer, and the second layer is a multi-head attention layer. The output of the first layer is the input of the second layer, and a final output layer is added. After standardizing the collagen-specific enzyme cleavage database, it was used to pre-train a dynamic prediction model for protease cleavage sites. The natural type III collagen sequence is converted into a numerical embedding vector and input into the trained model. A sliding window is used to scan segment by segment to obtain the probability of the cut site at each position in the sequence.

4. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 3, characterized in that, The method for generating the heatmap includes: Define a first risk threshold and a second risk threshold. If the probability of the cut site is greater than the first risk threshold, it is marked as a high-risk area. If the probability of the cut site is greater than or equal to the first risk threshold and is greater than the second risk threshold, it is marked as a medium-risk area. If the probability of the cut site is less than or equal to the second risk threshold and contains an active motif, it is marked as a protected area. For natural type III collagen sequences, risk regions are marked position by position: high-risk regions are marked as non-functional regions, medium-risk regions are marked as potential functional regions, and protected regions are marked as active motif regions. After image processing, a pixelated probability map is output.

5. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 4, characterized in that, The method of predicting the stability of a sequence triple helix structure using a structural model includes: Set the AlphaFold2 model as the starting point for fine-tuning, construct a fine-tuning dataset and use it to fine-tune the AlphaFold2 model. Fine-tuning is complete when the model's accuracy in testing reaches the standard. For the fine-tuned model, a helical torsion constraint is forcibly applied, and hydroxyproline simulation parameters are introduced to obtain the integrated structural model; The optimized natural type III collagen sequence is loaded and input into the structural model. The inter-chain hydrogen bond density and melting temperature are then output. If the inter-chain hydrogen bond density and melting temperature do not meet the expected standards, the model is rolled back and re-optimized until the expected standards are met.

6. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 5, characterized in that, The methods for fine-tuning the dataset and using it to fine-tune the AlphaFold2 model include: Collagen crystal structures were collected from the PDB database, and amino acid sequences, 3D coordinates, and annotations were extracted as the basis of the dataset. Samples containing collagen-specific patterns were loaded from a collagen-specific enzyme digestion database. If the similarity between the samples and the sequences in the dataset reached the expected level, the information from the samples was used to supplement the dataset, forming a fine-tuned dataset. Freeze the bottom layer of the AlphaFold2 model, only unfreeze the top layer, set the fine-tuning parameters, and use the fine-tuning dataset for iterative training.

7. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 6, characterized in that, The method of forcibly applying helical torsional constraints and introducing hydroxyproline simulation parameters includes: Based on the collagen-specific mode, torque parameters, target torsion angle, and deviation threshold are set. The custom script of the AlphaFold2 model is modified to select the inter-chain dihedral angle. If the torsion deviates and the energy cost increases, the coordinates are forcibly adjusted to curl towards the target torsion angle. The total energy is periodically detected, and if it exceeds the expected threshold, the torque parameters are automatically adjusted until the total energy is minimized and then stopped. Based on a collagen-specific enzyme digestion database, hydroxyproline simulation parameters were defined, and potential hydroxyproline sites were labeled in the optimized natural type III collagen sequence. Then, the force field was modified to add hydrogen bond bonuses to the labeled sites.

8. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 7, characterized in that, The optimized natural type III collagen sequence includes: Annotate each position in the natural type III collagen sequence, including amino acid sequence, cleavage site probability, labeled risk regions, and additional annotations; Based on the annotation results and heatmaps, high-risk areas were deleted or replaced; medium-risk areas were replaced; and protected areas retained their original sequences and added protective modifications.

9. The method for recombinant collagen sequence analysis based on AI deep learning according to claim 8, characterized in that, The method for generating a report file based on skin permeability prediction includes: The net charge, hydrophilicity index, and molecular rigidity of natural type III collagen sequences were extracted, and pH-dependent cleavage rates were integrated from a collagen-specific enzyme cleavage database. The skin permeability prediction incorporates a dual-objective optimization mechanism: maximizing the retention of the natural transdermal domain and minimizing structural perturbations. The output transdermal optimized sequence is then analyzed to determine whether the transdermal coefficient and the proportion of circular dichroism α-helices meet the output criteria. If not, the optimization is rolled back; otherwise, a report file is output, including the transdermal optimized sequence, transdermal coefficient, and proportion of circular dichroism α-helices.

10. An application for recombinant collagen sequence analysis based on AI deep learning, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program to implement the AI ​​deep learning-based recombinant collagen sequence analysis method as described in any one of claims 1 to 9.

Images