Artificial intelligence reverse design method and system of multifunctional antifreeze peptide
By constructing a mapping model between two-dimensional infrared spectroscopy and peptide backbone, and combining it with deep learning technology, multifunctional antifreeze peptides are designed in reverse. This solves the problems of low design efficiency and unreasonable generation results in existing technologies, and realizes efficient and automated multifunctional peptide generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to efficiently design multifunctional peptides that possess both antifreeze and antibacterial capabilities. Furthermore, the generation process lacks physical constraints, has unreasonable structures, and the generation results lack interpretability and functional screening.
By constructing a mapping model between two-dimensional infrared spectra and peptide backbones, and combining it with deep learning technology, multifunctional antifreeze peptides are designed in reverse, including protein/peptide three-dimensional structure prediction, spectral calculation, backbone generation, and sequence design, realizing the automated generation and screening from functional requirements to structure and sequence.
This improves the design efficiency and success rate of multifunctional antifreeze peptides, and the generated results are physically reasonable and interpretable, realizing efficient and automated design from functional requirements to structure and sequence.
Smart Images
Figure CN122157782A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of peptide generation technology, and more specifically to an artificial intelligence reverse design method and system for multifunctional antifreeze peptides. Background Technology
[0002] With the rapid development of protein engineering, synthetic biology, and artificial intelligence in the life sciences, our understanding of the relationship between peptide molecular structure, conformation, and biological function has become increasingly important. Antifreeze peptides and antimicrobial peptides, as two classes of highly valuable functional peptides, have demonstrated significant importance in areas such as cell cryopreservation, food preservation, medical materials, and anti-infective therapy. However, the structural characteristics of these peptides are highly dependent on intricate hydrogen bond networks, secondary structure conformations, hydrophobic-hydrophilic stripe distributions, and surface charge patterns, making their design process complex and requiring substantial experimental and computational resources. How to efficiently obtain multifunctional peptides with both antifreeze and antimicrobial capabilities, starting from functional requirements, is a key scientific and engineering challenge that urgently needs to be addressed internationally. Among existing technologies, deep learning-based molecular and protein generation methods have received widespread attention, including various diffusion models, generative adversarial networks, graph neural networks, and the application of Transformers in the field of biomacromolecules. Researchers are continuously attempting to improve and optimize generative models to enable them to generate molecular or protein structures with specific functions or physical properties. For example, targeted generation of sequences or structures can be achieved by designing conditional diffusion models, introducing controllable attribute vectors, or using graph-based generation methods. Some works utilize protein three-dimensional structural information for sequence dedesign to obtain peptide molecules with target biological activities. This type of research has made some progress in the field of molecular generation, laying the foundation for the intelligent design of biofunctional molecules. However, existing methods still have significant limitations. Most current peptide generation technologies start directly from one-dimensional text or functional tags, attempting to generate three-dimensional protein structures or amino acid sequences. This involves traversing a high-dimensional mapping process from requirement to conformation, easily leading to instability and uncontrollability caused by this "dimensional leap." Due to the lack of physical constraints, peptides generated by these methods often suffer from structural inconsistencies, discontinuous twist angles, and broken hydrogen bond networks, resulting in a lack of physical feasibility. Furthermore, the function of antifreeze peptides depends on highly specific geometric conformations, such as ice crystal surfaces and hydrophobic stripe arrangements, and these key features are difficult to effectively capture through direct sequence generation methods. Similarly, antimicrobial peptides rely on a balance between positive charge distribution and hydrophobicity, and their structural requirements are not entirely consistent with those of antifreeze peptides. This makes it difficult for traditional models to simultaneously satisfy both types of functions, hindering the joint design of multi-objective and multi-functional peptides. On the other hand, although two-dimensional infrared spectroscopy (2DIR) can reflect the hydrogen bond network, conformational distribution, and secondary structure changes of peptide molecules, existing molecular generation techniques almost never utilize spectroscopy as an intermediate physical characterization tool. Current methods generally directly generate structures or sequences instead of using physical signals such as spectra as an intermediate layer for regulation in the "spectrum-structure-function" relationship. This results in a lack of understanding of the essential conformation of molecules in the models, making it difficult to generate structures with interpretability and physical plausibility. Furthermore, there are still shortcomings in functional screening, especially in the field of antifreeze peptides, where a long-standing lack of reliable computational methods makes large-scale peptide screening difficult.The computational prediction of ice-binding capacity only became feasible with the emergence of next-generation ice crystal interface prediction models. However, no existing technology has yet integrated spectral generation, structure generation, sequence generation, and functional screening into a single closed-loop system, and there is no complete AI method for multifunctional peptide design. In summary, existing peptide design methods generally face problems such as a lack of physical constraints in structure generation, inability of sequence design to capture complex conformational patterns, difficulty in achieving multi-objective joint optimization of models, lack of interpretability of generated structures, and a lack of effective functional screening mechanisms. Summary of the Invention
[0003] The technical problem to be solved by this invention is how to improve the efficiency and success rate of reverse designing multifunctional antifreeze peptides starting from functional requirements.
[0004] The present invention solves the above-mentioned technical problems through the following technical means:
[0005] This invention provides an artificial intelligence reverse design method for multifunctional antifreeze peptides, comprising the following steps: S1. Obtain the three-dimensional structure of natural multifunctional antifreeze peptide sequences collected from public databases using a protein / peptide three-dimensional structure prediction model. S2. Based on the three-dimensional structure, calculate its two-dimensional infrared spectrum; label each two-dimensional infrared spectrum with functional text information, and construct a training set for mapping text information to two-dimensional infrared spectra; S3. Using the text-conditional image generation model as the base model, and employing an efficient parameter fine-tuning method, the model is trained based on the training set obtained in step S2 to obtain a two-dimensional infrared spectrum generation model. S4. Based on the multifunctional antifreeze peptide backbone in the three-dimensional structure, construct a training set for mapping two-dimensional infrared spectra to the multifunctional antifreeze peptide backbone. S5. Construct a two-dimensional infrared spectrum to backbone prediction model, and complete the model training based on the training set obtained in step S4 to obtain the peptide backbone generation model. S6. Input function text information, and obtain the predicted peptide backbone through the two-dimensional infrared spectroscopy generation model and peptide backbone generation model; S7. Refine the predicted peptide backbone and call the backbone conditional sequence design module to generate the corresponding peptide sequence. S8. Evaluate the generated peptide sequences and screen out peptide sequences that meet the comprehensive evaluation criteria as candidate peptide sequences for subsequent molecular simulation or experimental verification.
[0006] Furthermore, the protein / peptide three-dimensional structure prediction models described in step S1 include: the AlphaFold model and the RoseTTAFold model.
[0007] Further, step S2 includes the following steps: S21. Construct a molecular dynamics system for two-dimensional infrared spectroscopy calculations and output a set of structure frames that can be used for two-dimensional infrared spectroscopy calculations; S22. The AIM method is used to quantify the interactions of the multifunctional peptide system corresponding to each structural frame, and the set of interaction parameters that can be used for spectral calculations is output. S23. The NISE method is used to perform two-dimensional infrared spectral calculations. The structural frame and interaction parameters are input together and converted into two-dimensional infrared spectral data with time-resolved characteristics. S24. Add text information describing the function of each two-dimensional infrared spectrum image data to generate a training set.
[0008] Furthermore, step S21 specifically involves the following operations: Molecular dynamics simulations of the three-dimensional structure were performed in the GROMACS environment. First, the multifunctional peptide was placed in a pre-sized cubic water box for solvation, and then Na was added using the genion tool. + / Cl - Ions neutralize the overall charge of the system; based on this, environmental parameters such as temperature, pressure, and periodic boundary conditions are set, unreasonable contacts are eliminated and the initial configuration is optimized through an energy minimization process, and then the system evolution is driven by a time integration algorithm, outputting molecular dynamics trajectories and related dynamic data containing information on the change of atomic coordinates over time; the trajectories are used to construct the set of structural frames required for two-dimensional infrared spectroscopy calculations, that is, the conformations of multifunctional peptides at different time points are extracted from the trajectories at preset time intervals as structural inputs for spectral calculations.
[0009] Furthermore, step S22 specifically involves the following operations: By identifying the electron density-related topological features corresponding to the structural frames and extracting the key information of bond critical points, the criteria for identifying and quantifying the intensity of hydrogen bonds, covalent interactions, and non-bonded interactions are realized, forming interaction parameter results that correspond one-to-one with the structural frame index. The set of interaction parameters is used to characterize the differences in microscopic interactions under different time points of conformation and serves as the coupling / interaction input in two-dimensional infrared spectroscopy calculations, enabling subsequent spectral reconstruction to reflect the interaction changes caused by the evolution of multifunctional peptides over time.
[0010] Furthermore, step S23 specifically involves the following operations: The NISE method is used to perform two-dimensional infrared spectral calculations. The structural frame and interaction parameters are jointly input and converted into two-dimensional infrared spectral data with time-resolved characteristics. NISE calls the corresponding interaction parameters for each structural frame, generates the spectral contribution of the frame, and performs summarization / reconstruction in the time dimension according to preset rules, thereby obtaining a two-dimensional infrared spectral image characterized by time or statistical average.
[0011] Further, step S5 includes the following steps: S51. Construct a prediction model from two-dimensional infrared spectra to the backbone, the prediction model including an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is used to extract multi-scale features from the two-dimensional infrared spectral image, and the decoding subnetwork is used to convert the multi-scale features into a two-dimensional structural constraint representation for backbone reconstruction, thereby obtaining the three-dimensional coordinate output of the peptide backbone; the prediction model is implemented using a deep convolutional network. S52. Using a two-dimensional infrared spectral image as input and the actual backbone coordinates of the corresponding peptide as supervision labels, train the prediction model from the two-dimensional infrared spectrum to the backbone. During training, a joint loss function is used, including a backbone coordinate error term and a backbone geometry rationality constraint term. The backbone geometry rationality constraint term is used to constrain the bond length, bond angle, and conformational continuity between adjacent residues to avoid unreasonable distortions or atomic conflicts in the output backbone. The network parameters are updated using a gradient descent optimization method, and the training is iteratively continued until the loss function on the verification dataset converges or reaches the preset stopping condition.
[0012] Further, step S7 includes the following steps: S71. The design tool SCUBA-D is used to perform diffusion sampling and noise reduction reconstruction on the input peptide backbone. While ensuring that the overall folding topology remains unchanged, the local twist angle, backbone geometry and spatial conflicts are corrected, and the refined backbone structure PDB file is output. S72. Based on the refined backbone structure PDB file, the corresponding peptide sequence is generated using the design tool ProteinMPNN.
[0013] Further, step S8, which involves evaluating the generated peptide sequence, includes: (1) Perform legality checks and length screening on peptide sequences; (2) Perform three-dimensional structure prediction on the screened peptide sequences and evaluate the reliability of their structures; (3) Evaluate the antifreeze properties of the peptide structure; (4) Control screening of peptide sequences for novelty and homology; (5) Predict and screen the antibacterial activity of peptide sequences.
[0014] This invention also provides an artificial intelligence reverse design system for multifunctional antifreeze peptides. The system executes the above-described method during operation and includes the following modules: The natural sequence acquisition module is used to obtain the three-dimensional structure of natural multifunctional antifreeze peptide sequences collected from public databases through protein / peptide three-dimensional structure prediction models. The text-to-spectrum mapping module is used to calculate the two-dimensional infrared spectrum based on the three-dimensional structure; it annotates the functional text information for each two-dimensional infrared spectrum and constructs a training set for text-to-two-dimensional infrared spectrum mapping. The spectrum generation module uses a text-conditional image generation model as the base model and employs an efficient parameter fine-tuning method to complete model training based on the training set obtained from the step text-to-spectrum mapping module, thereby obtaining a two-dimensional infrared spectrum generation model. The spectral-to-backbone mapping module is used to construct a training set for mapping two-dimensional infrared spectra to multifunctional antifreeze peptide backbones based on the three-dimensional structure. The backbone generation module is used to construct a prediction model from two-dimensional infrared spectra to backbones. The model is trained based on the training set obtained by the spectrum-to-backbone mapping module to obtain a peptide backbone generation model. The integrated reasoning module is used to input functional text information, and obtain the predicted peptide backbone through a two-dimensional infrared spectroscopy generation model and a peptide backbone generation model. The sequence generation module is used to refine the predicted peptide backbone and call the backbone conditional sequence design module to generate the corresponding peptide sequence. The sequence screening module is used to evaluate the generated peptide sequences and screen out peptide sequences that meet the comprehensive evaluation criteria as candidate peptide sequences for subsequent molecular simulation or experimental verification.
[0015] The advantages of this invention are: (1) This invention establishes a quantitative relationship between the function of antifreeze peptides and their spectral patterns and spatial conformations, providing a reverse design approach for multifunctional antifreeze peptides based on functional requirements. Compared with existing technologies that directly generate and screen from sequences or three-dimensional structures, this invention improves the design efficiency of multifunctional antifreeze peptides.
[0016] (2) This invention enables the model to generate two-dimensional infrared spectra with corresponding physical characteristics based on user requirements such as “antifreeze”, “high heat hysteresis”, “ice binding”, and “antibacterial”. Compared with existing methods that only perform conditional generation in SMILES or structural space, this invention avoids the “dimensional jump” problem of directly jumping from one-dimensional requirements to three-dimensional structures by performing conditional sampling at the spectral level. Functional constraints are first completed in the intermediate physical space of the spectrum, which greatly improves the controllability of the design space search and the relevance of the generated results.
[0017] (3) This invention realizes the physical integration of “sequence-structure-spectrum”, and obtains two-dimensional spectral data reflecting the hydrogen bond network and secondary structure characteristics based on the real molecular structure, which improves the physical rationality and interpretability of subsequent generation and prediction.
[0018] (4) This invention constructs an integrated design pipeline of "spectral generation → skeleton prediction → skeleton refinement → sequence design" by connecting the spectral generation model, the spectral-structure mapping model and the design tools SCUBA-D and ProteinMPNN. Compared with the existing method of manually defining the skeleton and manually selecting sequences, this invention achieves automation and high throughput of skeleton refinement and sequence dedesign, and significantly reduces the dependence on human experience. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the artificial intelligence reverse design method for multifunctional antifreeze peptides according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the framework of the two-dimensional infrared spectrum generation model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the peptide backbone generation model according to an embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.
[0021] Example 1 This embodiment provides an artificial intelligence-based reverse design method for multifunctional antifreeze peptides. It constructs a reverse design system capable of describing the conformational characteristics of peptide molecules using two-dimensional infrared spectroscopy, enabling the correlation of spectrum, sequence, structure, and function within a unified framework. Based on the user's proposed multifunctional requirements, a structural backbone corresponding to the target spectrum is generated. Then, a sequence is generated from the backbone, and usable peptide molecules are obtained through stability and multifunctional activity screening. This method achieves spectrum-driven backbone prediction, backbone-driven sequence generation, and sequence-driven functional screening, making the entire process automated and ensuring structural rationality and sequence novelty. The specific implementation process is as follows: Figure 1 As shown, it includes the following steps: S1. Obtain the three-dimensional structure of natural multifunctional antifreeze peptide sequences collected from public databases using protein / peptide three-dimensional structure prediction models; the protein / peptide three-dimensional structure prediction models include: AlphaFold model and RoseTTAFold model.
[0022] S2. Based on the three-dimensional structure, calculate its two-dimensional infrared spectrum; label each two-dimensional infrared spectrum with functional text information, and construct a training set for mapping text information to two-dimensional infrared spectra; the specific implementation includes the following steps: S21. Construct a molecular dynamics system for two-dimensional infrared spectroscopy calculations, and output a set of structure frames that can be used for two-dimensional infrared spectroscopy calculations; specifically including: Molecular dynamics simulations of the three-dimensional structure were performed in the GROMACS environment. First, the multifunctional peptide was placed in a pre-sized cubic water box for solvation, and then Na was added using the genion tool. + / Cl - Ions neutralize the overall charge of the system; based on this, environmental parameters such as temperature, pressure, and periodic boundary conditions are set, unreasonable contacts are eliminated and the initial configuration is optimized through an energy minimization process, and then the system evolution is driven by a time integration algorithm, outputting molecular dynamics trajectories and related dynamic data containing information on the change of atomic coordinates over time; the trajectories are used to construct the set of structural frames required for two-dimensional infrared spectroscopy calculations, that is, the conformations of multifunctional peptides at different time points are extracted from the trajectories at preset time intervals as structural inputs for spectral calculations.
[0023] S22. The AIM method is used to quantify the interactions of the multifunctional peptide system corresponding to each structural frame, outputting a set of interaction parameters that can be used for spectral calculations. This embodiment introduces the AIM (Atoms in Molecules) analysis method based on Bader's electron density theory, aiming to deeply explore the microscopic forces within the peptide. By topologically mapping the electron density and its gradient of the system, the bond critical point (BCP) is accurately identified, thereby achieving a quantitative assessment of the strength of hydrogen bonds, covalent interactions, and non-bonded forces. This process provides a solid theoretical basis for determining the stability of peptide structures and predicting their physicochemical functions. Specifically, it includes: By identifying the electron density-related topological features corresponding to the structural frames and extracting the key information of bond critical points, the criteria for identifying and quantifying the intensity of hydrogen bonds, covalent interactions, and non-bonded interactions are realized, forming interaction parameter results that correspond one-to-one with the structural frame index. The set of interaction parameters is used to characterize the differences in microscopic interactions under different time points of conformation and serves as the coupling / interaction input in two-dimensional infrared spectroscopy calculations, enabling subsequent spectral reconstruction to reflect the interaction changes caused by the evolution of multifunctional peptides over time.
[0024] S23. Two-dimensional infrared spectroscopy calculations are performed using the NISE method. The structural frame and interaction parameters are jointly input and converted into two-dimensional infrared spectral data with time-resolved characteristics. To analyze the real-time evolution of peptides in complex environments, this embodiment utilizes the NISE (Nonlinear Infrared Spectroscopy Estimator) algorithm to perform nonlinear infrared spectral simulations on the simulated trajectory. This method successfully constructs a 2DIR spectral image with time resolution by decoupling the energy transfer paths between molecular vibrational modes. This not only intuitively reflects the structural dynamics of the peptide but also provides a key dimension for establishing the intrinsic relationship between "spectrum and conformation," specifically including: The NISE method is used to perform two-dimensional infrared spectral calculations. The structural frame and interaction parameters are jointly input and converted into two-dimensional infrared spectral data with time-resolved characteristics. NISE calls the corresponding interaction parameters for each structural frame, generates the spectral contribution of the frame, and performs summarization / reconstruction in the time dimension according to preset rules, thereby obtaining a two-dimensional infrared spectral image characterized by time or statistical average.
[0025] S24. Add text information describing the function of each two-dimensional infrared spectrum image data, such as antifreeze, heat retention, ice binding, antibacterial, etc., to generate a training set.
[0026] S3. A text-conditional image generation model is used as the basic model, and the model framework is as follows: Figure 2 As shown, in this embodiment, the Stable Diffusion XL base is used as the pre-trained model, and the LoRA method for efficient parameter fine-tuning is employed. Model training is completed based on the training set obtained in step S2. During model training, the training resolution is set to 512×512, the training batch size is 1, the gradient accumulation step count is 4, and the maximum learning step count is 6000 steps; the optimizer learning rate is set to 1×10⁻⁶. -4 The learning rate scheduling strategy used was cosine annealing, the mixed precision was set to fp16, and the random seed was 42. A checkpoint was saved every 500 steps, and the training process log was recorded to the wandb platform. Multi-GPU training was initiated using the accelerate function, with the GPU device specified using the environment variable CUDA_VISIBLE_DEVICES. After training, a two-dimensional spectral generation model based on SDXL was obtained.
[0027] S4. Based on the multifunctional antifreeze peptide backbone in the three-dimensional structure, construct a training set for mapping two-dimensional infrared spectra to the multifunctional antifreeze peptide backbone. S5. Construct a two-dimensional infrared spectrum to backbone prediction model, and complete model training based on the training set obtained in step S4 to obtain a peptide backbone generation model; the specific implementation includes the following steps: S51. Construct a prediction model from two-dimensional infrared spectra to the backbone, wherein the prediction model includes an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is used to extract multi-scale features from the two-dimensional infrared spectral image, and the decoding subnetwork is used to convert the multi-scale features into a two-dimensional structural constraint representation for backbone reconstruction, thereby obtaining the three-dimensional coordinate output of the peptide backbone; specifically, as shown... Figure 3 As shown, the prediction model described in this embodiment adopts a deep convolutional encoder-decoder structure. Its encoding part is a DCNN with atrous convolutions (DCNN), and its decoding part is a structure combining upsampling and convolution, specifically including: (1) The encoding end uses several layers of 3×3 dilated convolution to extract low-level features, and uses a multi-scale atrous convolution module (dilation rate of 6, 12, etc.) with global image pooling to fuse features from different receptive fields to obtain multi-scale spectral features; (2) After feature fusion, the two-dimensional correlation pattern of the spectrum is further extracted by the 3×3 convolution and spatial convolution modules to obtain high-dimensional characterization related to peptide secondary structure and hydrogen bond network. (3) The decoding end adopts a cascaded structure, which fuses the high-level features and low-level features of the encoding end through 1×1 convolution and feature concatenation, and then gradually restores the spatial resolution through two upsampling and 3×3 convolution to output a continuous two-dimensional feature map. (4) The decoded features are mapped to a two-dimensional tensor representing the contact map of residues or the probability of secondary structures through the terminal 1×1 convolutional layer, and then the tensor is transformed into three-dimensional coordinates of the Cα skeleton through the gradient-based folding module.
[0028] S52. During model training, a two-dimensional infrared spectral image is used as input, and the actual backbone coordinates of the corresponding peptide are used as supervision labels to train the prediction model from the two-dimensional infrared spectrum to the backbone. A joint loss function is used during training, including a backbone coordinate error term and a backbone geometry rationality constraint term. The backbone geometry rationality constraint term is used to constrain the bond length, bond angle and conformational continuity between adjacent residues to avoid unreasonable distortion or atomic conflicts in the output backbone. Gradient descent optimization method is used to update the network parameters, and iterative training is performed until the loss function on the verification dataset converges or reaches the preset stopping condition.
[0029] S6. In the integrated reasoning stage, input functional text information, and obtain the predicted peptide backbone through the two-dimensional infrared spectroscopy generation model and the peptide backbone generation model. S7. Refine the predicted peptide backbone and use the backbone conditional sequence design module to generate the corresponding peptide sequence; the specific implementation includes the following steps: S71. The design tool SCUBA-D is used to perform diffusion sampling and noise reduction reconstruction on the input peptide backbone. While ensuring that the overall folding topology remains unchanged, the local twist angle, backbone geometry and spatial conflicts are corrected, and the refined backbone structure PDB file is output. Specifically, in this embodiment, the officially released checkpoint_clean.pt weight file is preferably used, including: (1) Install and configure the SCUBA-D runtime environment in the Linux environment, create a Conda environment named "SCUBA-D", install Python and the dependencies listed in requirements.txt, and execute postInstall.sh to complete the initialization configuration of model weights and scripts; (2) Based on the skeleton output by the spectrum → structure model, write the corresponding skeleton structure PDB file path and design control parameters into the SCUBA-D control JSON file. It is preferred to use the structure_refine.json template provided by the official website. In the template, specify the coordinate file to be optimized, the number of residues, the number of diffusion steps and other parameters so that SCUBA-D can refine the structure of the given skeleton in the “refine_prior” mode. S72. Based on the refined backbone structure PDB file, the corresponding peptide sequence is generated using the design tool ProteinMPNN. ProteinMPNN is a deep learning-based protein sequence design model that can generate amino acid sequences that are compatible with and foldably stable for a given protein backbone structure. This embodiment uses its CA-only mode to perform sequence dedesign on the SCUBA-D refined backbone, including: (1) Install the dependencies required for ProteinMPNN in an independent Conda environment, including Python, PyTorch and NumPy, and download the CA-only weight files such as ca_model_weights / v_48_020.pt released by the official website to the directory specified by path_to_model_weights; (2) Use the refined backbone PDB file output by SCUBA-D as the input of ProteinMPNN, and write the chain number to be designed into pdb_path_chains or the corresponding jsonl configuration file. In this embodiment, it is preferred to design the single-chain peptide chain with all residues and not fix any position. (3) Set the --ca_only flag to enable the CA-only model, --model_name is preferably v_48_020, --num_seq_per_target is set to greater than 1 (e.g. 16 or 32) so as to generate multiple candidate sequences for each skeleton, and --sampling_temp is set to a temperature in the range of 0.1–0.3 to balance sequence diversity and probability distribution; (4) Set parameters such as --omit_AAs or --bias_AA_jsonl as needed to inhibit or bias specific amino acids in order to facilitate the formation of Thr-rich or positively charged surface features common to antifreeze peptides. (5) After running ProteinMPNN, an output file containing candidate sequences and corresponding scores is obtained. The output records the score, global_score and other indicators of each sequence. In this embodiment, candidate sequences with lower global_score and sequence length of no more than 50 amino acids are selected to form the initial sequence library for subsequent multifunctional screening steps.
[0030] S8. Evaluate the generated peptide sequences and screen out peptide sequences that meet the comprehensive evaluation criteria as candidate peptide sequences for subsequent molecular simulation or experimental verification.
[0031] The evaluation of the generated peptide sequence includes: (1) Perform legality checks and length screening on peptide sequences; the legality check includes checking the amino acid character set and removing sequences containing non-standard amino acid characters or illegal symbols. The length screening includes retaining candidate peptides whose length does not exceed a preset length threshold. The length threshold is set according to the needs of subsequent evaluation and experimental verification and is not limited to a fixed value; (2) The reference three-dimensional structure of the candidate peptide is predicted using AlphaFold3 and a pLDDT score is obtained. The reference three-dimensional structure is then compared with the refined three-dimensional structure. After the structure is superimposed, the root mean square deviation is calculated to characterize the structural difference between the two. When the root mean square deviation is within a preset reasonable range, the three-dimensional structure generated by this method is considered to be in good agreement with the reference predicted structure, thus the structural result of the candidate peptide is considered to have usability and reliability. (3) Evaluate the antifreeze performance of peptide structures; input the three-dimensional structure into the antifreeze performance prediction module, such as the ICEPIC tool, to evaluate its ice-binding potential and heat hysteresis-related activities, and select candidate peptides with antifreeze potential according to the preset screening rules. (4) Controlled screening of peptide sequences for novelty and homology; specifically including: Global sequence alignment involves aligning candidate sequences one by one with sequences in the database using a global sequence alignment method to obtain the alignment results. The alignment method can be implemented using well-known sequence alignment algorithms, and there are no restrictions on specific tools or scoring matrices.
[0032] Similarity index calculation: Based on the alignment results, a sequence identity index is calculated. The sequence identity index is the ratio of the number of identical amino acids after alignment to the total alignment length. Optionally, a sequence similarity index can be calculated to characterize the degree of similarity of conservative substitutions.
[0033] Novelty threshold screening retains only candidate peptides whose sequence identity index relative to any known antifreeze peptide in the database is not higher than a preset threshold, to ensure that the candidate peptides have sufficient novelty at the sequence level.
[0034] (5) Predict and screen the antibacterial activity of peptide sequences. Specifically, input the peptide sequences into the antibacterial activity prediction module, such as the EvoGradient tool, and output the antibacterial activity prediction results for each sequence. The output can be a binary classification result or a continuous score. According to the preset screening rules, candidate sequences predicted to have antibacterial potential or whose scores meet the preset standards are retained to form a candidate set that has both antifreeze potential and antibacterial potential.
[0035] Finally, candidate peptides that simultaneously meet the following conditions are summarized as multifunctional antifreeze peptide candidate molecules output from the reverse design of this embodiment. The conditions include: (1) The length of the candidate peptide meets the preset length threshold; (2) The results of the antifreeze performance prediction screening meet the preset rules, including the ice binding potential meeting the requirements; the ice-related category belonging to the antifreeze-related category; the thermal hysteresis-related activity index reaching the threshold and the structure prediction reliability index reaching the threshold; (3) The sequence novelty screening results meet the preset rules, that is, the sequence identity index relative to the known antifreeze peptide database is not higher than the threshold; (4) The antibacterial activity prediction screening results meet the preset rules, that is, the prediction is that it has antibacterial potential or the score reaches the threshold. The candidate sequences mentioned above can be further used for molecular simulation or experimental verification. The entire screening process can be automatically executed by an AI-aided design system, featuring scalability and reproducibility.
[0036] Example 2 It should be further explained that, based on the same inventive concept, this embodiment also provides an artificial intelligence reverse design system for multifunctional antifreeze peptides. When the system runs, it executes the method described in Embodiment 1, including the following modules: The natural sequence acquisition module is used to obtain the three-dimensional structure of natural multifunctional antifreeze peptide sequences collected from public databases through protein / peptide three-dimensional structure prediction models. The text-to-spectrum mapping module is used to calculate the two-dimensional infrared spectrum based on the three-dimensional structure; it annotates the functional text information for each two-dimensional infrared spectrum and constructs a training set for text-to-two-dimensional infrared spectrum mapping. The spectrum generation module uses a text-conditional image generation model as the base model and employs an efficient parameter fine-tuning method to complete model training based on the training set obtained from the step text-to-spectrum mapping module, thereby obtaining a two-dimensional infrared spectrum generation model. The spectral-to-backbone mapping module is used to construct a training set for mapping two-dimensional infrared spectra to multifunctional antifreeze peptide backbones based on the three-dimensional structure. The backbone generation module is used to construct a prediction model from two-dimensional infrared spectra to backbones. The model is trained based on the training set obtained by the spectrum-to-backbone mapping module to obtain a peptide backbone generation model. The integrated reasoning module is used to input functional text information, and obtain the predicted peptide backbone through a two-dimensional infrared spectroscopy generation model and a peptide backbone generation model. The sequence generation module is used to refine the predicted peptide backbone and call the backbone conditional sequence design module to generate the corresponding peptide sequence. The sequence screening module is used to evaluate the generated peptide sequences and screen out peptide sequences that meet the comprehensive evaluation criteria as candidate peptide sequences for subsequent molecular simulation or experimental verification.
[0037] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An artificial intelligence reverse design method for multifunctional antifreeze peptides, characterized in that, Includes the following steps: S1. Obtain the three-dimensional structure of natural multifunctional antifreeze peptide sequences collected from public databases using a protein / peptide three-dimensional structure prediction model. S2. Based on the three-dimensional structure, calculate its two-dimensional infrared spectrum; label each two-dimensional infrared spectrum with functional text information, and construct a training set for mapping text information to two-dimensional infrared spectra; S3. Using the text-conditional image generation model as the base model, and employing an efficient parameter fine-tuning method, the model is trained based on the training set obtained in step S2 to obtain a two-dimensional infrared spectrum generation model. S4. Based on the multifunctional antifreeze peptide backbone in the three-dimensional structure, construct a training set for mapping two-dimensional infrared spectra to the multifunctional antifreeze peptide backbone. S5. Construct a two-dimensional infrared spectrum to backbone prediction model, and complete the model training based on the training set obtained in step S4 to obtain the peptide backbone generation model. S6. Input function text information, and obtain the predicted peptide backbone through the two-dimensional infrared spectroscopy generation model and peptide backbone generation model; S7. Refine the predicted peptide backbone and call the backbone conditional sequence design module to generate the corresponding peptide sequence. S8. Evaluate the generated peptide sequences and screen out peptide sequences that meet the comprehensive evaluation criteria as candidate peptide sequences for subsequent molecular simulation or experimental verification.
2. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 1, characterized in that, The protein / peptide three-dimensional structure prediction models mentioned in step S1 include: AlphaFold model and RoseTTAFold model.
3. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 1, characterized in that, Step S2 includes the following steps: S21. Construct a molecular dynamics system for two-dimensional infrared spectroscopy calculations and output a set of structure frames that can be used for two-dimensional infrared spectroscopy calculations; S22. The AIM method is used to quantify the interactions of the multifunctional peptide system corresponding to each structural frame, and the set of interaction parameters that can be used for spectral calculations is output. S23. The NISE method is used to perform two-dimensional infrared spectral calculations. The structural frame and interaction parameters are input together and converted into two-dimensional infrared spectral data with time-resolved characteristics. S24. Add text information describing the function of each two-dimensional infrared spectrum image data to generate a training set.
4. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 3, characterized in that, The specific operation of step S21 is as follows: Molecular dynamics simulations of the three-dimensional structure were performed in the GROMACS environment. First, the multifunctional peptide was placed in a pre-sized cubic water box for solvation, and then Na was added using the genion tool. + / Cl - Ions neutralize the overall charge of the system; Based on this, environmental parameters such as temperature, pressure, and periodic boundary conditions are set. Unreasonable contacts are eliminated and the initial configuration is optimized through an energy minimization process. Then, the system evolution is driven by a time integration algorithm, and molecular dynamics trajectories and related dynamic data containing information on the change of atomic coordinates over time are output. The trajectories are used to construct the set of structural frames required for two-dimensional infrared spectroscopy calculations. That is, the conformations of multifunctional peptides at different time points are extracted from the trajectories at preset time intervals as structural inputs for spectral calculations.
5. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 4, characterized in that, The specific operation of step S22 is as follows: By identifying the electron density-related topological features corresponding to the structural frames and extracting the key information of bond critical points, the criteria for identifying and quantifying the intensity of hydrogen bonds, covalent interactions, and non-bonded interactions are realized, forming interaction parameter results that correspond one-to-one with the structural frame index. The set of interaction parameters is used to characterize the differences in microscopic interactions under different time points of conformation and serves as the coupling / interaction input in two-dimensional infrared spectroscopy calculations, enabling subsequent spectral reconstruction to reflect the interaction changes caused by the evolution of multifunctional peptides over time.
6. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 5, characterized in that, The specific operation of step S23 is as follows: The NISE method is used to perform two-dimensional infrared spectral calculations. The structural frame and interaction parameters are jointly input and converted into two-dimensional infrared spectral data with time-resolved characteristics. NISE calls the corresponding interaction parameters for each structural frame, generates the spectral contribution of the frame, and performs summarization / reconstruction in the time dimension according to preset rules, thereby obtaining a two-dimensional infrared spectral image characterized by time or statistical average.
7. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 1, characterized in that, Step S5 includes the following steps: S51. Construct a prediction model from two-dimensional infrared spectra to the backbone, the prediction model including an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is used to extract multi-scale features from the two-dimensional infrared spectral image, and the decoding subnetwork is used to convert the multi-scale features into a two-dimensional structural constraint representation for backbone reconstruction, thereby obtaining the three-dimensional coordinate output of the peptide backbone; the prediction model is implemented using a deep convolutional network. S52. Using a two-dimensional infrared spectral image as input and the actual backbone coordinates of the corresponding peptide as supervision labels, train the prediction model from the two-dimensional infrared spectrum to the backbone. During training, a joint loss function is used, including a backbone coordinate error term and a backbone geometry rationality constraint term. The backbone geometry rationality constraint term is used to constrain the bond length, bond angle, and conformational continuity between adjacent residues to avoid unreasonable distortions or atomic conflicts in the output backbone. The network parameters are updated using a gradient descent optimization method, and the training is iteratively continued until the loss function on the verification dataset converges or reaches the preset stopping condition.
8. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 7, characterized in that, Step S7 includes the following steps: S71. The design tool SCUBA-D is used to perform diffusion sampling and noise reduction reconstruction on the input peptide backbone. While ensuring that the overall folding topology remains unchanged, the local twist angle, backbone geometry and spatial conflicts are corrected, and the refined backbone structure PDB file is output. S72. Based on the refined backbone structure PDB file, the corresponding peptide sequence is generated using the design tool ProteinMPNN.
9. The artificial intelligence reverse design method for multifunctional antifreeze peptides according to claim 1, characterized in that, Step S8, which involves evaluating the generated peptide sequence, includes: (1) Perform legality checks and length screening on peptide sequences; (2) Perform three-dimensional structure prediction on the screened peptide sequences and evaluate the reliability of their structures; (3) Evaluate the antifreeze properties of the peptide structure; (4) Control screening of peptide sequences for novelty and homology; (5) Predict and screen the antibacterial activity of peptide sequences.
10. An AI-based reverse design system for multifunctional antifreeze peptides, characterized in that, Includes the following modules: The natural sequence acquisition module is used to obtain the three-dimensional structure of natural multifunctional antifreeze peptide sequences collected from public databases through protein / peptide three-dimensional structure prediction models. The text-to-spectrum mapping module is used to calculate the two-dimensional infrared spectrum based on the three-dimensional structure; it annotates the functional text information for each two-dimensional infrared spectrum and constructs a training set for text-to-two-dimensional infrared spectrum mapping. The spectrum generation module uses a text-conditional image generation model as the base model and employs an efficient parameter fine-tuning method to complete model training based on the training set obtained from the step text-to-spectrum mapping module, thereby obtaining a two-dimensional infrared spectrum generation model. The spectral-to-backbone mapping module is used to construct a training set for mapping two-dimensional infrared spectra to multifunctional antifreeze peptide backbones based on the three-dimensional structure. The backbone generation module is used to construct a prediction model from two-dimensional infrared spectra to backbones. The model is trained based on the training set obtained by the spectrum-to-backbone mapping module to obtain a peptide backbone generation model. The integrated reasoning module is used to input functional text information, and obtain the predicted peptide backbone through a two-dimensional infrared spectroscopy generation model and a peptide backbone generation model. The sequence generation module is used to refine the predicted peptide backbone and call the backbone conditional sequence design module to generate the corresponding peptide sequence. The sequence screening module is used to evaluate the generated peptide sequences and screen out peptide sequences that meet the comprehensive evaluation criteria as candidate peptide sequences for subsequent molecular simulation or experimental verification.