Functional specific therapeutic peptide design and multi-stage screening method based on protein large language model and multi-attribute constraints

Abstract

The application discloses a functional specificity therapeutic peptide design and multi-stage screening method based on a protein large language model and multi-attribute constraints in the cross field of biological medicine and artificial intelligence. The method comprises the following steps: data construction and input; function prompt and continuous attribute coding; double condition protein large language model fine tuning; candidate therapeutic peptide generation; and multi-stage calculation screening. The method solves the problems of poor controllability, difficulty in multi-attribute synergistic regulation and low reliability of screening results in the intelligent design of existing therapeutic peptides, significantly improves the therapeutic peptide discovery efficiency and candidate molecule quality, has the characteristics of modularity and scalability, and can be widely applied to the intelligent design of various functional polypeptides and protein drugs.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of biomedicine and artificial intelligence, specifically to a method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints. Background Technology

[0002] Therapeutic peptides are a class of functional biomolecules composed of amino acid sequences that can participate in specific biological regulation or target action. With advantages such as high target specificity, good biocompatibility, and designability, they have shown great application potential in the treatment of complex diseases such as drug-resistant bacterial infections, tumors, and autoimmune diseases. Traditional therapeutic peptide discovery relies on screening natural bioactive peptides, empirical sequence modification, and local optimization, which suffers from bottlenecks such as long development cycles, high costs, and low sequence space exploration efficiency, making it difficult to meet the demands of modern drug development for rapid discovery of candidate molecules.

[0003] In recent years, generative artificial intelligence (AI) technology has provided a new paradigm for therapeutic peptide design, enabling the learning of sequence composition patterns and potential property preferences from large-scale protein or peptide sequences, achieving batch generation and high-throughput initial screening of candidate sequences. However, existing AI-based therapeutic peptide generation methods still face three major technical challenges: 1. Insufficient functional controllability: Most methods are limited to single activity or weak condition control, making it difficult to achieve stable and controllable generation for a specified therapeutic function category, resulting in poor functional consistency of the generated sequence.

[0004] 2. Weak ability to coordinate multiple attributes: It cannot effectively handle the coupling and conflict between therapeutic function and key physicochemical properties such as solubility and stability, resulting in unreasonable distribution of candidate sequence attributes and poor drug-likeness.

[0005] 3. Incomplete screening system: Existing computational screening methods have a single evaluation dimension, mainly relying on sequence property prediction or single docking score, lacking a comprehensive evaluation of target space matching, interface interaction and complex dynamic stability, resulting in low reliability of screening results and low success rate of subsequent experimental verification.

[0006] Therefore, there is an urgent need to develop a smart design technology for therapeutic peptides that can simultaneously achieve functionally specific and controllable generation, multi-attribute synergistic optimization, and multi-dimensional precise screening, in order to improve the efficiency of therapeutic peptide discovery and the quality of candidate molecules. Summary of the Invention

[0007] The purpose of this invention is to provide a method for designing and screening functionally specific therapeutic peptides based on a large protein language model and multi-attribute constraints. This method addresses the technical problems of poor functional controllability, difficulty in multi-attribute synergistic regulation, and imperfect screening systems in existing intelligent design of therapeutic peptides, and realizes an integrated process from large-scale sequence generation to the selection of highly reliable candidate molecules.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints includes the following steps: S1. Data Construction and Input: Collect active peptide sequences with clear therapeutic function labels from the database, calculate key physicochemical properties, construct a training dataset containing functional labels, sequences and properties, and select representative protein targets with corresponding functions; S2. Functional hints and continuous attribute encoding: The therapeutic function categories are converted into discrete functional hint tokens and added to the model vocabulary. Solubility and stability attributes are encoded as continuous attribute embeddings. Soft modulation of attribute pair sequence generation is achieved through additive injection. S3. Fine-tuning of a dual-conditional protein big language model: Using a pre-trained protein big language model as the backbone, functional cue embeddings and attribute-modulated sequence embeddings are fused together, and fine-tuning is performed using an autoregressive conditional language modeling task to obtain a multi-conditional constrained therapeutic peptide generation model. S4. Candidate therapeutic peptide generation: Input the target therapeutic function category, the model generates candidate peptide sequences in batches, and obtains a non-redundant candidate peptide library after redundancy removal, and predicts the complex structure of candidate peptides with corresponding targets. S5. Multi-stage computational screening: Candidate complexes are screened step by step through rigid docking, flexible docking and molecular dynamics simulation, and conformational stability and interfacial interactions are comprehensively evaluated to obtain high-quality candidate therapeutic peptides.

[0009] Furthermore, S1 includes the following steps: S11. Collect active therapeutic peptide sequences with clearly labeled therapeutic functions from the APD3 database to form a set of original therapeutic peptide sequences. ,in, This represents the nth therapeutic peptide sequence. Indicates the corresponding treatment function category; S12. The original therapeutic peptide sequences are cleaned and standardized to remove sequences with format errors, abnormal amino acid characters, missing or duplicate functional tags, and to obtain a set of standardized therapeutic peptide sequences D'. S13. For each therapeutic peptide sequence, calculate the solubility-related attribute value and the stability-related attribute value; n S14. Integrate the therapeutic peptide sequences, functional category labels, solubility attribute values, and stability attribute values ​​according to a unified data format to construct a high-quality bioactive peptide dataset, TheraPep19. Each therapeutic peptide sample is represented in a unified FASTA format. ,in, The treatment function category corresponding to the nth sample; S15. According to the functional categories, the TheraPep19 dataset is divided into 80% training set and 20% test set using a stratified random sampling strategy; S16. For each therapeutic function category, select a representative protein target with pharmacological or biochemical basis, obtain its three-dimensional structure data from the RCSB database, and construct a protein target dataset ProteinSet.

[0010] Furthermore, the solubility-related attribute values ​​in S13 are represented by the GRAVY average hydrophobicity index, denoted as... :

[0011] in, For sequence length, For the i-th amino acid, h( ) represents the hydrophobicity value of amino acids; Stability-related attribute values ​​are represented by the instability index calculated using the ProtParam tool, denoted as . :

[0012] in, This represents the nth candidate peptide sequence.

[0013] Furthermore, S2 includes the following steps: S21. Encode each training sample as an autoregressive input sequence with a start symbol, function cue, therapeutic peptide sequence, padding symbol, and end symbol:

[0014] Where START is the start symbol, END is the end symbol, and PAD is the padding symbol. The token is the function category hint for the nth sample. The token is the amino acid at the i-th position in the n-th therapeutic peptide sequence; S22. Add all function category hints to the model vocabulary and assign trainable embedding vectors to the newly added hint tokens. (Based on the original unknown tokens) <unk>Embedded mean and variance initialization, add new hint embedding:

[0015] in, and Let represent the mean and standard deviation of the original unknown label embedding vector, respectively; S23. Solubility properties and stability properties Projecting each vector onto the model embedding dimension d yields the soluble embedding vector. and stability embedding vector ,in To share the linear projection matrix; S24. Merge and normalize the two attribute vectors to obtain a unified attribute representation. : ; S25. Will Broadcast along the sequence dimension, injecting the polypeptide sequence body token embedding, excluding start, end, and padding characters.

[0016] in, Embed the original i-th amino acid token. For attribute-enhanced embedding, This is a set of token indexes for the polypeptide body.

[0017] Furthermore, S3 includes the following steps: S31. Map the function hint token corresponding to the function label to a function embedding vector: ,in, This represents the feature token corresponding to the nth sample. This indicates the embedded representation of its corresponding function; S32. Concatenate the embedded functionality with the attribute-modulated sequence token representation to construct the model input sequence: ,in, This represents the embedding of the i-th amino acid token after fusing continuous attribute information; S33. Load the pre-trained protein large language model progen2-small-mix7 as the backbone of the autoregressive Transformer decoder, and reconstruct the word embedding layer and the language model output layer according to the number of tokens added to the new function prompts; S34. The input sequence is fed into a multi-layer Transformer decoder for feature extraction. The output of the k-th layer Transformer is represented as: ,in, This represents the hidden state of the (k-1)th layer; S35. The model learns the conditional probability distribution of therapeutic peptide sequences under functional category and attribute constraints:

[0018] in, This represents the amino acid token at position t in the nth therapeutic peptide sequence. This represents all generated tokens before the current position t in the sequence. This represents the attribute condition of the nth therapeutic peptide; S36. For the current position t, the model calculates the predicted probability of the next token based on the hidden state:

[0019] in, Indicates the current location is hidden. and These represent the output layer weight matrix and bias term, respectively. S37. The model is fine-tuned using an autoregressive conditional language modeling task, enabling it to predict the amino acid token at the next position in a sequence based on input functional category cues, preceding tokens, and continuous attribute embeddings. The optimization objective is the conditional negative log-likelihood loss function.

[0020] Where N is the batch size. Let n be the length of the nth sequence. For the t-th token in the n-th sequence, The token is the function category hint for the nth sample. This represents the attribute condition of the nth therapeutic peptide; S38. The Adam optimizer and cosine annealing learning rate scheduling strategy are used to update the model parameters. In the early stage of training, the number of warm-up steps is set to improve the stability of fine-tuning, and the model parameters are updated according to the gradient of the loss function. ,in, Represents the set of model parameters. Indicates the learning rate. This represents the gradient of the loss function with respect to the model parameters.

[0021] S39. Complete the model training according to the preset number of training rounds, and save the model checkpoints, word segmenter and optimizer status at preset intervals.

[0022] Furthermore, S4 includes the following steps: S41. Input the target treatment function category and construct the initial prompt sequence: ,in, Indicates the target treatment function category; START is the start symbol. S42. The model performs forward inference based on the generated token sequence, and is evaluated using a temperature coefficient. The probability distribution of the next token after scaling:

[0023] in, This represents the output logits at the t-th generation position. Indicates the temperature coefficient; S43. On probability distribution Perform top-k sampling, retain only the top k candidate tokens with the highest probabilities, and set the probabilities of the remaining tokens to approximately zero. Then, sample from the retained candidate tokens to generate the amino acid token for the current position. S44. Append the sampled token to the end of the generated sequence, and continue to generate the next step using the cached past key values. Iterate until the model generates the END token or reaches the preset maximum sequence length. S45. Perform post-processing on the generated sequences to remove special markers from the sequences and use CD-HIT for redundancy removal to obtain a non-redundant candidate peptide library. S46. Pair each non-redundant candidate peptide with the protein target corresponding to its functional category to form a candidate peptide-target protein input pair, and use the complex structure prediction tool Boltz-2 to predict the three-dimensional conformation of the peptide-protein complex.

[0024] Furthermore, S5 includes the following steps: S51. Perform ZDOCK rigid docking on each candidate peptide and its corresponding protein target to generate 20 initial complex conformations, and retain the top-1 score conformation; S52. Perform Rosetta flexible docking on each candidate peptide obtained by rigid docking to generate 5 flexible optimized conformations, and select the top-5 conformations based on the results; S53. Energy minimization, NVT equilibrium, NPT equilibrium, and 40 ns molecular dynamics simulations were performed sequentially on the screened candidate complexes. Calculations were made on multiple indicators including RMSD, RMSF, radius of gyration Rg, solvent accessible surface area SASA, number of hydrogen bonds, and interfacial contact frequency within 2.5 Å. The results were comprehensively evaluated and further screened to obtain high-quality candidate peptides.

[0025] By adopting the above technical solution, the present invention has the following advantages: 1. This invention provides a method for designing and screening function-specific therapeutic peptides based on a protein big language model and multi-attribute constraints. It proposes for the first time a dual-condition mechanism that integrates hard cues for functional categories and soft modulation of continuous attributes, enabling the directional generation of candidate peptides under specified therapeutic functions. At the same time, it guides the sequences to converge toward reasonable solubility and stability distributions, significantly improving functional consistency and attribute rationality.

[0026] 2. This invention provides a method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints. It adopts an additive attribute injection approach, which enhances the attributes of the peptide main token without changing the main structure of the pre-trained model. While preserving the original sequence representation ability, it significantly improves the fine-tuning efficiency and task adaptability. Functional and attribute constraints have been introduced in the generation stage, which significantly reduces the number of low-quality sequences entering the high-cost docking and simulation stages and reduces the consumption of computing resources.

[0027] 3. This invention provides a method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints. It constructs a multi-stage screening process from sequence redundancy removal to molecular dynamics verification, and comprehensively evaluates from multiple dimensions such as sequence rationality, target space matching, interfacial interactions, and dynamic stability, thereby improving the reliability of candidate molecule selection results. The functional hint set, physicochemical property constraint indicators, and drugability evaluation can be expanded according to needs, making it highly scalable and applicable to the design of different types of functional peptides and protein molecules. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process of the functional-specific therapeutic peptide design and multi-stage screening method based on protein large language model and multi-attribute constraints of the present invention. Figure 2 This is a diagram of an ablation experiment based on solubility properties. Figure 3 Figure for ablation experiment to assess stability properties; Figure 4 This is a comparison chart of the ZDOCK scores of the Target-AMP model and the control model PepMLM under different therapeutic functional categories. Figure 5 This is a comparison chart of the ZDOCK scores of the Target-AMP model and the control model DiffPepBuilder under different therapeutic function categories. Figure 6 A comparison chart showing the consistency of physicochemical properties between real active therapeutic peptides and candidate peptides; Figure 7 A comparison plot of sequence diversity of candidate therapeutic peptides generated for different models; Figure 8 This is a graph used to verify the molecular dynamics stability of the candidate therapeutic peptide-target protein complex of the present invention. Figure 9 This diagram illustrates the molecular structure stability verification of the candidate therapeutic peptide-target protein complex of this invention. Detailed Implementation

[0029] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0030] This invention proposes a method for the design and multi-stage screening of therapeutic peptides based on a protein big language model and multi-attribute constraints. This method uses a pre-trained protein big language model as its foundation, employing a dual-condition mechanism combining functional prompts and continuous attribute modulation to incorporate constraints such as therapeutic function category, solubility, and stability into the candidate peptide generation process. This enables the controllable generation of candidate therapeutic peptides targeting specific therapeutic functions and key physicochemical properties. Furthermore, this invention integrates multi-stage screening steps, including target-related structure prediction, rigid docking, flexible optimization, and molecular dynamics stability verification, to evaluate and progressively optimize the generated candidate peptides using multiple indicators. This forms an integrated technical process from therapeutic peptide sequence design to multi-indicator calculation and screening of candidate molecules.

[0031] Therapeutic peptide design and multi-stage screening methods based on protein large language models and multi-attribute constraints are detailed below. Figure 1 As shown, it includes the following steps: S1. Data Construction and Input: Collect active peptide sequences with clear therapeutic function labels from the database, calculate key physicochemical properties, construct a training dataset containing functional labels, sequences and properties, and select representative protein targets with corresponding functions; S1 includes the following specific steps: S11. Collect active therapeutic peptide sequences with clearly labeled therapeutic functions from the APD3 database to form a set of original therapeutic peptide sequences. ,in This represents the nth therapeutic peptide sequence. This indicates the corresponding therapeutic function category, which includes categories such as antifungal, antitumor, and antiviral, as detailed below. Figure 1 Step 101 in the process.

[0032] S12. The original therapeutic peptide sequences are cleaned and standardized to remove sequences with format errors, abnormal amino acid characters, missing or duplicate functional tags, resulting in a standardized set of therapeutic peptide sequences. , specifically Figure 1 Step 102 in the process.

[0033] S13. Calculate key physicochemical properties for each therapeutic peptide sequence, including solubility-related property values ​​and stability-related property values.

[0034] Among them, the solubility-related attribute values ​​are represented by the GRAVY average hydrophobicity index, denoted as . :

[0035] in, For sequence length, For the i-th amino acid, h( ) represents the hydrophobicity value of amino acids.

[0036] Stability-related attribute values ​​are represented by the instability index calculated using the ProtParam tool, denoted as . :

[0037] in, This represents the nth candidate peptide sequence, specifically as follows: Figure 1 Step 103 in the process.

[0038] S14. Include the therapeutic peptide sequence, functional category label, and solubility property value. and stability attribute value A high-quality bioactive peptide dataset, TheraPep19, covering 19 therapeutic functions, was constructed by integrating data according to a unified data format. Each therapeutic peptide sample is represented in FASTA format. .in, This indicates that the nth sample corresponds to the treatment function category. This represents the nth therapeutic peptide sequence. This represents the solubility attribute value of the nth sample. This represents the stability attribute value of the nth sample, specifically as follows: Figure 1 Step 104 in the process.

[0039] S15. The TheraPep19 dataset is divided according to functional categories using a stratified random sampling strategy. 80% of the samples from each category form the training set TheraPep19-Train, and the remaining 20% ​​form the test set TheraPep19-Test. Specifically, as follows... Figure 1 Step 105 in the process.

[0040] S16. For each therapeutic function category, select a representative protein target with pharmacological or biochemical basis, and obtain its three-dimensional structure data from the RCSB database to construct a protein target dataset, ProteinSet, as follows: Figure 1 Step 106 in the process.

[0041] Reading therapeutic peptide pretreated samples As input 1, specifically as follows: Figure 1 Step 107 in the middle; Read the protein target dataset ProteinSet as input 2, specifically as follows: Figure 1 Step 108 in the process.

[0042] S2. Functional hints and continuous attribute encoding: The therapeutic function categories are converted into discrete functional hint tokens and added to the model vocabulary. Solubility and stability attributes are encoded as continuous attribute embeddings. Soft modulation of attribute pair sequence generation is achieved through additive injection. S2 includes the following specific steps: S21: Encode each training sample as an autoregressive input sequence with a start symbol, function cue, therapeutic peptide sequence, padding symbol, and end symbol:

[0043] Where START is the start symbol, END is the end symbol, and PAD is the padding symbol. The token is the function category hint for the nth sample. This refers to the amino acid token at the i-th position in the n-th therapeutic peptide sequence; specifically, as shown below. Figure 1 Step 201 in the process.

[0044] S22: Add all functional category cues to the model vocabulary and assign trainable embedding vectors to the new cue tokens. To ensure that the new embeddings are consistent with the original protein large language model representation space, they are based on the original unknown tags. <unk>Embedded mean and variance initialization, add new hint embedding:

[0045] in, and represents the mean and standard deviation of the original unknown label embedding vector, respectively.

[0046] Expanded glossary: , =19, as detailed below Figure 1 Step 202 in the process.

[0047] S23: Solubility properties and stability properties Projecting each vector onto the model embedding dimension d yields the soluble embedding vector. and stability embedding vector ,in To share the linear projection matrix. Specifically, as follows: Figure 1 Step 203 in the process.

[0048] S24: Merge and normalize the two attribute vectors to obtain a unified attribute representation, Eattr, as follows: Figure 1 Step 204 in the process.

[0049]

[0050] S25: Will Broadcast along the sequence dimension, injecting polypeptide sequence body token embeddings (excluding special tokens such as start, end, and padding tokens):

[0051] in, Embed the original i-th amino acid token. For attribute-enhanced embedding, This is a set of token indexes for the polypeptide body, specifically as follows: Figure 1 Step 205 in the process.

[0052] S3. Fine-tuning of a dual-conditional protein large language model: Using a pre-trained protein large language model as the backbone, functional cue embeddings and attribute-modulated sequence embeddings are fused together, and fine-tuning is performed using an autoregressive conditional language modeling task to obtain a multi-conditional constrained therapeutic peptide generation model.

[0053] S3 includes the following specific steps: S31: Map the function hint token corresponding to the function tag to a function embedding vector: .in, This represents the feature token corresponding to the nth sample. This indicates the corresponding embedded function representation, specifically as follows: Figure 1 Step 301 in the process.

[0054] S32: Concatenate the embedded functionality with the attribute-modulated sequence token representation to construct the model input sequence: .in, This represents the embedding of the i-th amino acid token after fusing continuous attribute information, specifically as follows: Figure 1 Step 302 in the process.

[0055] S33: Load the pre-trained protein-based large language model progen2-small-mix7 as the backbone of the autoregressive Transformer decoder, and reconstruct the word embedding layer and language model output layer according to the number of newly added feature suggestion tokens, so that the model vocabulary size is consistent with the set of newly added suggestions, as detailed below. Figure 1 Step 303 in the process.

[0056] S34: The input sequence is fed into a multi-layer Transformer decoder for feature extraction. The output of the k-th layer Transformer is represented as follows: .in, The hidden state of the (k-1)th layer represents the updated hidden state after processing by the kth Transformer decoder layer.

[0057] S35: The model learns the conditional probability distribution of therapeutic peptide sequences under functional category and attribute constraints:

[0058] in, This represents the amino acid token at position t in the nth therapeutic peptide sequence. This represents all generated tokens before the current position t in the sequence. This represents the attribute condition of the nth therapeutic peptide.

[0059] S36: For the current position t, the model calculates the predicted probability of the next token based on the hidden state:

[0060] in, Indicates the current location is hidden. and These represent the output layer weight matrix and bias term, respectively.

[0061] S37: The model is fine-tuned using an autoregressive conditional language modeling task, specifically as follows: Figure 1 Step 307 in the model enables the model to predict the amino acid token at the next position in the sequence based on the input functional category hints, preceding tokens, and continuous attribute embeddings. The optimization objective is the conditional negative log-likelihood loss function.

[0062] Where N is the batch size. Let n be the length of the nth sequence. For the t-th token in the n-th sequence, The token is the function category hint for the nth sample. This represents the attribute condition of the nth therapeutic peptide.

[0063] S38: The model parameters are updated using the Adam optimizer and a cosine annealing learning rate scheduling strategy. A warm-up step count is set during the initial training phase to improve fine-tuning stability. Model parameters are updated based on the loss function gradient. .in, Represents the set of model parameters. Indicates the learning rate. This represents the gradient of the loss function with respect to the model parameters.

[0064] S39: Complete model training according to the preset number of training rounds, and save the model checkpoints, word segmenter, and optimizer states at preset intervals, as detailed below. Figure 1 Step 310 in the process.

[0065] S4. Candidate therapeutic peptide generation: Input the target therapeutic function category, the model generates candidate peptide sequences in batches, and obtains a non-redundant candidate peptide library after redundancy removal, and predicts the complex structure of candidate peptides with corresponding targets. S4 includes the following specific steps: S41: Input the target treatment function category and construct the initial prompt sequence: .in, This indicates the target therapeutic function category, with START as the start symbol. During the generation phase, only the target function prompt is required. The model generates candidate therapeutic peptide sequences based on the functional category features and attribute distributions learned in the fine-tuning phase, as detailed below. Figure 1 Step 401 in the process.

[0066] S42: At each generation step, the model performs forward inference based on the generated token sequence and outputs logits from the amino acid vocabulary, denoted as... Temperature coefficient The probability distribution of the next token after scaling:

[0067] in, This represents the output logits at the t-th generation position. This represents the temperature coefficient.

[0068] S43: On probability distribution Perform top-k sampling, retaining only the top k candidate tokens with the highest probabilities, and setting the probabilities of the remaining tokens to approximately zero. Then, sample from the retained candidate tokens to generate the amino acid token for the current position: Where k represents the number of candidate tokens retained. This represents the amino acid token obtained by sampling at the t-th generation position, specifically as follows: Figure 1 Step 403 in the process.

[0069] S44: Append the sampled token to the end of the generated sequence, and continue the next generation using the cached past key values. This process continues until the model generates an END marker or reaches the preset maximum sequence length, as detailed below. Figure 1 Step 404 in the process.

[0070] S45: Post-process the generated sequence to remove function hints, START markers, END markers, and padding characters, obtaining a standard peptide sequence. If a new START or END marker appears in the generated sequence, use the position of its first occurrence as the cutoff point, and use CD-HIT to remove redundancy from the generated candidate sequence, obtaining a non-redundant candidate therapeutic peptide library, as detailed below. Figure 1 Steps 405 and 406 in the process.

[0071] S46: Pair each non-redundant candidate peptide with its corresponding protein target for its functional category to form candidate peptide-target protein input pairs. Use the complex structure prediction tool Boltz-2 to predict the three-dimensional conformation of the peptide-protein complex and extract the peptide structure for docking and subsequent simulations, specifically as follows: Figure 1 Step 407 in the process.

[0072] S5. Multi-stage computational screening: Candidate complexes are screened step by step through rigid docking, flexible docking and molecular dynamics simulation, and conformational stability and interfacial interactions are comprehensively evaluated to obtain high-quality candidate therapeutic peptides.

[0073] S5 includes the following steps: S51: Perform ZDOCK rigid docking on each candidate peptide and its corresponding protein target to generate 20 initial complex conformations. The top-scoring conformations are retained for the next stage, as detailed below. Figure 1 Step 501 in the process.

[0074] S52: Each candidate peptide obtained from rigid docking is subjected to Rosetta flexible docking, generating 5 optimized flexible conformations. The top-5 conformations are then selected for further adaptation of interfacial side chains and local conformations. Specifically, as follows... Figure 1 Step 502 in the process. The Rosetta interface score statistics of the top-5 conformational candidate peptide-target protein complexes are shown in Table 1 below: Table 1. Statistical table of Rosetta interface scores for the top 5 candidate peptide-target protein complexes in each therapeutic function category of the present invention.

[0075] S53: The candidate complexes obtained from the screening were subjected to energy minimization, NVT equilibrium, NPT equilibrium, and 40 ns molecular dynamics simulations in sequence. Based on the simulated trajectories, multiple indices were calculated, including RMSD, RMSF, radius of gyration Rg, solvent-accessible surface area (SASA), number of hydrogen bonds, and interfacial contact frequency within 2.5 Å, to evaluate the global conformational stability and interfacial interaction persistence in solution. Further screening yielded a few high-quality candidate peptides, specifically as follows: Figure 1 Step 503 in the process.

[0076] This method takes active therapeutic peptide sequences with clearly defined therapeutic functions, functional category labels, solubility and stability attributes, and representative protein targets as inputs. Through data construction and input, it completes sequence cleaning, physicochemical property calculation, functional category labeling, and target protein selection. Subsequently, through functional hints and continuous attribute encoding, the therapeutic function category is converted into discrete functional hint information, and physicochemical properties such as solubility and stability are converted into continuous attribute conditions. Based on this, a pre-trained protein big language model is used for dual-condition fine-tuning, enabling the model to learn the correspondence between functional categories, sequence patterns, and physicochemical properties. This allows for the batch generation of candidate therapeutic peptide sequences with specified functional categories by inputting only the target functional hints during the generation phase. Finally, through sequence redundancy removal, target-related structure prediction, and binding stability analysis, candidate peptides are progressively screened and validated to obtain candidate therapeutic peptides with high functional matching, reasonable attributes, and target binding potential. In the specific implementation, the start symbol START in the model input sequence is set to 1, and the end symbol END is set to 2. The initial learning rate during the model fine-tuning phase is set to 1×10⁻⁶. -4 The warm-up steps were set to 200, and the training rounds to 15. During the candidate therapeutic peptide generation phase, the sampling script was iteratively called for 19 functional tags, with batch size set to 100, max length to 64, temperature to 1.0, and top-k to 10. After generation, CD-HIT was used for sequence deduplication, and the sequence similarity threshold was set to 0.6.

[0077] Substitutability of models and tools: 1. The pre-trained protein large language model described in this invention is not limited to progen2-small-mix7. In specific implementation, it can be replaced with other models capable of generating protein or peptide sequences, depending on the scale of training data, computing resources, and generation task requirements.

[0078] 2. The tools and parameters in the multi-stage computational screening process of this invention can be replaced or adjusted according to actual needs. For example, other structure prediction tools can be used in the structure prediction stage, and molecular modeling and docking software with the same functions can be used in the docking stage.

[0079] 3. The candidate therapeutic peptides obtained by this invention are a set of preferred candidate molecules obtained by calculation and screening, which can be used as candidates for subsequent wet experiments.

[0080] The following series of experiments validate the functional-specific therapeutic peptide design and multi-stage screening method based on a large protein language model and multi-attribute constraints of the present invention: Figure 2 The effect of solubility property modulation on the distribution of the generated therapeutic peptide GRAVY was verified to illustrate the technical effect of solubility property modulation in this invention. Figure 2 The complete Target-AMP model was compared with a control model that removed solubility modulation. The horizontal axis represents different therapeutic function categories, and the vertical axis represents the normalized GRAVY score. Box plots were used to represent the GRAVY distribution of candidate peptides generated in each group, and the mean point was used to represent the average level. Figure 2 It can be seen that after introducing continuous solubility property modulation, the solubility distribution of candidate peptides generated by the model can shift to a higher range, thus proving that continuous property modulation can effectively constrain the physicochemical properties of the generated sequence.

[0081] Figure 3 This method is used to verify the effect of stability property modulation on the stability composition of generated therapeutic peptides, and to illustrate the technical effect of continuous stability property modulation in this invention. The horizontal axis represents the proportion of stable and unstable peptides, and the vertical axis represents different therapeutic function categories. Under each therapeutic function category, the stability composition of the complete Target-AMP model and the control model with stability modulation removed (w / o Stability) are compared. Figure 3 It can be seen that continuous stability property modulation can increase the proportion of stable peptides in the generated candidate peptides, reduce the possibility of low-stability candidate peptides entering the subsequent high-cost structure screening stage, and thus improve the efficiency of candidate molecule selection.

[0082] Figure 4 The diagram shows a comparison of ZDOCK scores between the Target-AMP model of this invention and the control model PepMLM under different therapeutic functional categories. Figure 5 This figure shows a comparison of ZDOCK scores between the Target-AMP model and the control model DiffPepBuilder under different therapeutic function categories, illustrating the target binding potential of the candidate therapeutic peptides generated by this invention. The horizontal axis represents different therapeutic function categories, and the vertical axis represents the ZDOCK score. Box plots of Target-AMP and the corresponding control model are drawn for each therapeutic function category, with the mean point representing the average ZDOCK score of each group of candidate peptides. The boxes represent the quartile range of the score distribution, the median line represents the median, and the upper and lower extension lines represent the score dispersion range. Figure 4 , Figure 5 It can be seen that, under the same protein target conditions, the candidate therapeutic peptides generated by this invention have a higher rigidity docking score distribution, indicating that they are more likely to form a peptide-target protein binding conformation with spatial matching relationship.

[0083] Figure 6 A comparison chart of the consistency of physicochemical properties between the actual active therapeutic peptides and candidate peptides is shown to illustrate the consistency of the physicochemical properties of the candidate therapeutic peptides generated in this invention. The horizontal axis represents eight physicochemical property indicators, including hydrophobicity, hydrophobic moment, net charge, sequence length, Boman index, isoelectric point, aliphatic index, and α-helix ratio; the vertical axis represents the normalized property values. For each property indicator, bar charts are plotted for the actual active therapeutic peptide sample, the candidate peptide generated by Target-AMP, and the candidate peptide generated by the corresponding comparative model, with error bars representing the standard deviation. Figure 6 It can be seen that the candidate therapeutic peptides generated by the present invention are closer to real active therapeutic peptide samples in terms of multidimensional physicochemical properties, thus indicating that the present invention can maintain a more reasonable distribution of biochemical properties under functional constraints.

[0084] Figure 7 This diagram illustrates the sequence diversity of candidate therapeutic peptides generated by different models. It serves to demonstrate the sequence-level diversity of candidate therapeutic peptides generated by the Target-AMP model of this invention. Each sub-plot corresponds to the sequence similarity evaluation results under a specific therapeutic function category, including two indicators: sequence identity and sequence similarity. The horizontal axis represents percentages, the bar charts represent the average sequence identity or average sequence similarity of candidate peptide libraries corresponding to different data sources or generation models, and the error bars represent the mean standard error. The comparison objects include real training peptide libraries, Target-AMP generated candidate peptide libraries, and candidate peptide libraries generated by comparison models such as PepMLM, DiffPepBuilder, and BindCraft. The overall sequence identity and sequence similarity of the candidate peptide libraries are relatively low, indicating that the model does not simply generate highly homologous or repetitive candidate sequences, but rather explores a broader sequence space while maintaining functional and attribute constraints. This demonstrates that this invention, when generating candidate therapeutic peptides with target functional categories, can balance functional and attribute controllability with sequence diversity, reducing the redundancy of the candidate peptide library.

[0085] Figure 8 The diagram illustrates the molecular dynamics stability verification of the candidate therapeutic peptide-target protein complex of the present invention. The diagram includes curves showing the change in the number of hydrogen bonds versus the RMSD (Mean Reactive Molecular Strain). The hydrogen bond number curve illustrates the persistence of the interfacial interaction between the candidate peptide and the target protein during the simulation, while the RMSD curve illustrates the conformational stability of the candidate peptide and target protein backbone structures during the simulation. If the number of hydrogen bonds remains relatively stable during the simulation and the RMSD curve does not exhibit a sustained large drift, it indicates that the candidate therapeutic peptide-target protein complex obtained through the multi-stage screening of the present invention can maintain relatively stable interfacial interactions and overall conformational state. Figure 8 This demonstrates that the structure prediction, rigid docking, flexible optimization, and molecular dynamics verification process of this invention can effectively screen candidate therapeutic peptide complexes with good dynamic stability.

[0086] Figure 9 This figure shows the molecular structural stability verification of the candidate therapeutic peptide-target protein complex of the present invention. The figure includes the time-varying results of three indicators: radius of gyration (Rg), root mean square fluctuation (RMSF), and solvent accessible surface area (SASA). Rg represents the overall compactness of the candidate peptide, target protein, and their complex during molecular dynamics simulation; RMSF represents the local flexibility and fluctuation amplitude of protein residues and peptide chain residues during the simulation; and SASA characterizes the surface area changes of the candidate peptide, target protein, and complex exposed to the solvent environment. If the Rg and SASA curves do not show a sustained abnormal increase or collapse during the simulation, and the high fluctuation region of RMSF is mainly concentrated at the ends or non-critical interface regions, it indicates that the candidate therapeutic peptide-target protein complex obtained through the multi-stage screening of the present invention has good overall conformational stability, local structural stability, and solvent environment adaptability.

[0087] Finally, it should be noted that although the present invention has been described with reference to specific embodiments, those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, any changes or modifications to the above embodiments within the essential spirit of the present invention will fall within the scope of the claims of the present invention.< / unk> < / unk>

Claims

1. A method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints, characterized in that: Includes the following steps: S1. Data Construction and Input: Collect active peptide sequences with clear therapeutic function labels from the database, calculate key physicochemical properties, construct a training dataset containing functional labels, sequences and properties, and select representative protein targets with corresponding functions; S2. Functional hints and continuous attribute encoding: The therapeutic function categories are converted into discrete functional hint tokens and added to the model vocabulary. Solubility and stability attributes are encoded as continuous attribute embeddings. Soft modulation of attribute pair sequence generation is achieved through additive injection. S3. Fine-tuning of a dual-conditional protein big language model: Using a pre-trained protein big language model as the backbone, functional cue embeddings and attribute-modulated sequence embeddings are fused together, and fine-tuning is performed using an autoregressive conditional language modeling task to obtain a multi-conditional constrained therapeutic peptide generation model. S4. Candidate therapeutic peptide generation: Input the target therapeutic function category, the model generates candidate peptide sequences in batches, and obtains a non-redundant candidate peptide library after redundancy removal, and predicts the complex structure of candidate peptides with corresponding targets. S5. Multi-stage computational screening: Candidate complexes are screened step by step through rigid docking, flexible docking and molecular dynamics simulation, and conformational stability and interfacial interactions are comprehensively evaluated to obtain high-quality candidate therapeutic peptides.

2. The method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints according to claim 1, characterized in that, S1 includes the following steps: S11. Collect active therapeutic peptide sequences with clearly labeled therapeutic functions from the APD3 database to form a set of original therapeutic peptide sequences. ,in, This represents the nth therapeutic peptide sequence. Indicates the corresponding treatment function category; S12. The original therapeutic peptide sequences are cleaned and standardized to remove sequences with format errors, abnormal amino acid characters, missing or duplicate functional tags, and to obtain a set of standardized therapeutic peptide sequences D'. S13. For each therapeutic peptide sequence, calculate the solubility-related attribute value and the stability-related attribute value; n S14. Integrate the therapeutic peptide sequences, functional category labels, solubility attribute values, and stability attribute values ​​according to a unified data format to construct a high-quality bioactive peptide dataset, TheraPep19. Each therapeutic peptide sample is represented in a unified FASTA format. ,in, The treatment function category corresponding to the nth sample; S15. According to the functional categories, the TheraPep19 dataset is divided into 80% training set and 20% test set using a stratified random sampling strategy; S16. For each therapeutic function category, select a representative protein target with pharmacological or biochemical basis, obtain its three-dimensional structure data from the RCSB database, and construct a protein target dataset ProteinSet.

3. The method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints according to claim 2, is characterized in that... The solubility-related attribute values ​​in S13 are represented by the GRAVY average hydrophobicity index, denoted as... : in, For sequence length, For the i-th amino acid, h( ) represents the hydrophobicity value of amino acids; Stability-related attribute values ​​are represented by the instability index calculated using the ProtParam tool, denoted as . : in, This represents the nth candidate peptide sequence.

4. The method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints according to claim 3, is characterized in that... S2 includes the following steps: S21. Encode each training sample as an autoregressive input sequence with a start symbol, function cue, therapeutic peptide sequence, padding symbol, and end symbol: Where START is the start symbol, END is the end symbol, and PAD is the padding symbol. The token is the function category hint for the nth sample. The token is the amino acid at the i-th position in the n-th therapeutic peptide sequence; S22. Add all function category hints to the model vocabulary and assign trainable embedding vectors to the newly added hint tokens. (Based on the original unknown tokens) <unk> Embedded mean and variance initialization, add new hint embedding:< / unk> in, and Let represent the mean and standard deviation of the original unknown label embedding vector, respectively; S23. Solubility properties and stability properties Projecting each vector onto the model embedding dimension d yields the soluble embedding vector. and stability embedding vector ,in To share the linear projection matrix; S24. Merge and normalize the two attribute vectors to obtain a unified attribute representation. : ； S25. Will Broadcast along the sequence dimension, injecting the polypeptide sequence body token embedding, excluding start, end, and padding characters. in, Embed the original i-th amino acid token. For attribute-enhanced embedding, This is a set of token indexes for the polypeptide body.

5. The method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints according to claim 4, characterized in that, S3 includes the following steps: S31. Map the function hint token corresponding to the function label to a function embedding vector: ,in, This represents the feature token corresponding to the nth sample. This indicates the embedded representation of its corresponding function; S32. Concatenate the embedded functionality with the attribute-modulated sequence token representation to construct the model input sequence: ,in, This represents the embedding of the i-th amino acid token after fusing continuous attribute information; S33. Load the pre-trained protein large language model progen2-small-mix7 as the backbone of the autoregressive Transformer decoder, and reconstruct the word embedding layer and the language model output layer according to the number of tokens added to the new function prompts; S34. The input sequence is fed into a multi-layer Transformer decoder for feature extraction. The output of the k-th layer Transformer is represented as: ,in, This represents the hidden state of the (k-1)th layer; S35. The model learns the conditional probability distribution of therapeutic peptide sequences under functional category and attribute constraints: in, This represents the amino acid token at position t in the nth therapeutic peptide sequence. This represents all generated tokens before the current position t in the sequence. This represents the attribute condition of the nth therapeutic peptide; S36. For the current position t, the model calculates the predicted probability of the next token based on the hidden state: in, Indicates the current location is hidden. and These represent the output layer weight matrix and bias term, respectively. S37. The model is fine-tuned using an autoregressive conditional language modeling task, enabling it to predict the amino acid token at the next position in a sequence based on input functional category cues, preceding tokens, and continuous attribute embeddings. The optimization objective is the conditional negative log-likelihood loss function. Where N is the batch size. Let n be the length of the nth sequence. For the t-th token in the n-th sequence, The token is the function category hint for the nth sample. This represents the attribute condition of the nth therapeutic peptide; S38. The Adam optimizer and cosine annealing learning rate scheduling strategy are used to update the model parameters. In the early stage of training, the number of warm-up steps is set to improve the stability of fine-tuning, and the model parameters are updated according to the gradient of the loss function. ,in, Represents the set of model parameters. Indicates the learning rate. This represents the gradient of the loss function with respect to the model parameters. S39. Complete the model training according to the preset number of training rounds, and save the model checkpoints, word segmenter and optimizer status at preset intervals.

6. The method for designing and screening functionally specific therapeutic peptides based on a large protein language model and multi-attribute constraints according to claim 5, is characterized in that... S4 includes the following steps: S41. Input the target treatment function category and construct the initial prompt sequence: ,in, Indicates the target treatment function category; START is the start symbol. S42. The model performs forward inference based on the generated token sequence, and is evaluated using a temperature coefficient. The probability distribution of the next token after scaling: in, This represents the output logits at the t-th generation position. Indicates the temperature coefficient; S43. On probability distribution Perform top-k sampling, retain only the top k candidate tokens with the highest probabilities, and set the probabilities of the remaining tokens to approximately zero. Then, sample from the retained candidate tokens to generate the amino acid token for the current position. S44. Append the sampled token to the end of the generated sequence, and continue to the next generation using the cached past key values, iterating until the model generates the END token or reaches the preset maximum sequence length; S45. Perform post-processing on the generated sequences to remove special markers from the sequences and use CD-HIT for redundancy removal to obtain a non-redundant candidate peptide library. S46. Pair each non-redundant candidate peptide with the protein target corresponding to its functional category to form a candidate peptide-target protein input pair, and use the complex structure prediction tool Boltz-2 to predict the three-dimensional conformation of the peptide-protein complex.

7. The method for designing and screening function-specific therapeutic peptides based on a large protein language model and multi-attribute constraints according to claim 6, characterized in that, S5 includes the following steps: S51. Perform ZDOCK rigid docking on each candidate peptide and its corresponding protein target to generate 20 initial complex conformations, and retain the top-1 score conformation; S52. Perform Rosetta flexible docking on each candidate peptide obtained by rigid docking to generate 5 flexible optimized conformations, and select the top-5 conformations based on the results; S53. Energy minimization, NVT equilibrium, NPT equilibrium, and 40 ns molecular dynamics simulations were performed sequentially on the screened candidate complexes. Calculations were made on multiple indicators including RMSD, RMSF, radius of gyration Rg, solvent accessible surface area SASA, number of hydrogen bonds, and interfacial contact frequency within 2.5 Å. The results were comprehensively evaluated and further screened to obtain high-quality candidate peptides.

Images