Peptide sequence generation model and method utilizing intermediate sequence MSA and diffusion masking mechanism

Abstract

This invention relates to a peptide sequence generation model and method utilizing intermediate sequence semantic masking (MSA) and diffusion masking mechanisms, specifically a peptide sequence AI model training method incorporating evolutionary information. It innovatively introduces intermediate sequences to address the problem that peptide sequences are too short to directly extract evolutionary information. First, an MSA is constructed for the peptide dataset. For sequences where an effective MSA cannot be constructed, their sequences are compared with a protein database to select intermediate sequences with high homology and appropriate length. Then, MSA data of these intermediate sequences is constructed to guide the AI ​​model in capturing evolutionary features. Model training employs a diffusion language masking mechanism and an MSA Transformer. All regions are randomly masked in the peptide MSA, and peptide mapping regions are masked in the intermediate MSA, enabling embedding modeling and generation of peptides. This method can capture sparse peptide evolutionary information, achieving semantic modeling and efficient generation of peptides, and is applicable to peptide sequence optimization and synthesis in fields such as drug design and protein engineering.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to the field of biological computing technology, and more specifically, to a method for generating polypeptide sequences. Background Technology

[0002] Peptide drugs possess advantages such as high activity and safety, strong specificity, good determinism, and good drug-likeness, making them widely applicable and promising in clinical practice. The goal of peptide drug design is to controllably and massively generate peptide sequences that satisfy specific properties. Peptide drug design is typically carried out manually by experts and then tested; however, this approach suffers from problems such as heavy reliance on expert knowledge and long design cycles.

[0003] To develop more peptide drugs, the applicant's research team has developed an antimicrobial peptide generation model, AMPGen. By integrating a pre-trained diffusion model, axial attention mechanism, and cascade screening system, they successfully designed novel antimicrobial peptide (AMP) sequences with 81.58% antimicrobial activity. Of the 38 synthetic peptides generated by this model, 31 showed significant inhibitory effects against Escherichia coli and Staphylococcus aureus (MIC ≤ 25 μM), providing an AI-driven solution to address the antibiotic resistance crisis.

[0004] CN115512763B discloses a method for generating peptide sequences, a method for training a peptide generation model, and an apparatus. The specific implementation involves: during peptide drug design of a specified protein, obtaining a reference peptide sequence corresponding to the specified protein, determining the spatial structural features of the protein binding site of the specified protein, and inputting the reference peptide sequence and spatial structural features into a pre-trained peptide generation model to obtain the target peptide sequence. Therefore, a method for generating peptide sequences by combining the spatial structural features of the protein binding site of a specified protein is proposed.

[0005] CN117809749B discloses a method, apparatus, memory, and electronic device for generating functional polypeptide sequences, belonging to the field of protein design technology. The generation method includes: acquiring training data; encoding the target functional polypeptide sequence in the training data into a two-dimensional feature matrix; training a functional polypeptide sequence feature generation model using the two-dimensional feature matrix to obtain a trained functional polypeptide sequence feature generation model; using the trained functional polypeptide sequence feature generation model, generating two-dimensional feature matrices of various new functional polypeptide sequences by adjusting the noise sampling method; and decoding the two-dimensional feature matrices of each new functional polypeptide sequence into the corresponding new functional polypeptide sequence. This invention fully utilizes the feature extraction capability of a pre-trained protein large language model and the generation capability of a generative model, achieving the goal of de novo synthesis of diverse functional polypeptide or protein sequences without templates and with unknown target structures by effectively extracting sequence features of specific categories of functional polypeptides.

[0006] CN114724643B discloses a method and related apparatus for screening peptide compounds, applicable to the field of artificial intelligence. The method involves acquiring a peptide generation model; then generating at least one peptide sequence based on the model. The peptide generation model includes an encoding layer and a decoding layer. The encoding layer generates a latent space based on training data, and the latent space provides sampling points. The peptide sequence is obtained by decoding the sampling points using the decoding layer. The peptide sequence is then input into an activity prediction model to obtain activity prediction information. This achieves a rapid and accurate peptide compound activity screening process. Since the input to the activity prediction model uses peptide sequences sampled from the feature dimensions of the peptide generation model, the effectiveness of the sampling is guaranteed, and no manual intervention is required, further improving the accuracy of peptide compound screening.

[0007] Existing methods for generating peptide sequences suffer from drawbacks such as high screening noise, redundancy, and low efficiency. Recently, multiple sequence alignment (MSA) has been introduced into the field of artificial intelligence-generated peptide sequences. MSA is used in bioinformatics to compare and align multiple biological sequences, such as protein or nucleic acid sequences, to identify common evolutionary features, conserved regions, and variation patterns.

[0008] In practice, MSA is more effective for substances with longer chain lengths, such as proteins or nucleic acids. However, for peptides composed of fewer amino acids, due to insufficient depth (i.e. fewer sequence numbers), MSA is less effective because the parameter sharing mechanism cannot function properly.

[0009] How to use MSA to generate sequences of peptides with a small number of amino acids is an urgent research question. Summary of the Invention

[0010] In view of the problems existing in the prior art, the present invention provides a method for sequence generation of polypeptides with insufficient depth (i.e., a small number of sequences).

[0011] Specifically, in order to solve the above-mentioned technical problems, the present invention provides the following technical solution.

[0012] Technical Solution 1: In a first aspect, the present invention provides a method for generating peptide sequences using an intermediate sequence MSA, comprising the following steps:

[0013] a. Construct the peptide sequence MSA,

[0014] b. Constructing intermediate sequence mappings for peptide sequences.

[0015] c. Construct the mediator sequence MSA.

[0016] d. Training masking diffusion modeling based on MSA.

[0017] e. Use the trained model to perform the generation task.

[0018] Technical Solution 2: Preferably, according to the method of generating peptide sequences using intermediate sequence MSA according to the above technical solution 1, wherein, in step a, peptide sequences of length 2-50aa are selected from the peptide dataset, and after quality control and deduplication of these peptide sequences, a sequence search tool is used to search for homologous sequences for each sequence in the database. These homologous sequences are peptide sequences, and peptide sequence MSA is constructed using these peptide sequences.

[0019] Technical Solution 3: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 2, wherein, in step a, the peptide dataset is UniProtKB.

[0020] Technical Solution 4: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 2 or 3, wherein, in step a, the sequence search tool is HHblits.

[0021] Technical Solution 5: Preferably, a method for generating peptide sequences using an intermediate sequence MSA according to any one of the above technical solutions 2-4, wherein, in step a, the database is UniRef30.

[0022] Technical Solution 6: Preferably, a method for generating peptide sequences using an intermediate sequence MSA according to any one of the above technical solutions 1-5, wherein, in step a, the peptide sequence MSA is in A3M format.

[0023] Technical Solution 7: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to any one of the above technical solutions 1-6, wherein, in step b, the BLAST tool is used to search for similar sequences in the UniRef100 database for sequences that cannot be constructed into peptide sequence MSA in step a and sequences with MSA depth less than or equal to 64, and candidate intermediate sequences are screened out.

[0024] In one specific implementation, it is preferred to construct intermediate MSAs only for peptide sequences with fewer than 20 amino acids based on reducing computational load, since most sequences with fewer than 20 amino acids are difficult to construct MSAs directly.

[0025] Technical Solution 8: Preferably, according to the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 7, wherein in step b, the screening conditions are identity greater than or equal to 90%, length greater than the peptide itself and not exceeding 20 times its length, preferably not exceeding 17 times the peptide length.

[0026] Technical solution 9: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 7 or 8, wherein, in step b, candidate intermediate sequences can be further screened to retain several intermediate sequences corresponding to each peptide.

[0027] Technical solution 10: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 9, wherein, in step b, the further screening condition is to retain the intermediate sequence with the largest bitscore value.

[0028] More specifically, after obtaining multiple candidate mediator sequences, the candidate mediator sequence with the longest align_len is selected. If there are multiple mediator sequences, the sequence with the highest bitscore is selected as the final mediator sequence. The start and end positions of the peptide on the mediator sequence are recorded. If the peptide completely matches a certain region of the mediator sequence, that region is used as the mask region.

[0029] Technical Solution 11: Preferably, in the method for generating peptide sequences using intermediate sequence MSA according to the above-described technical solution 9, in step b, the further screening condition is the intermediate sequence with the largest constructed MSA depth. Specifically, in one embodiment, the further screening condition is that the constructed MSA depth is greater than or equal to 64, and those less than 64 are deleted.

[0030] Technical Solution 12: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to any one of the above technical solutions 1-11, wherein in step c, a sequence search tool is used to search for homologous sequences of each intermediate sequence in the database, and these homologous sequences are intermediate sequences, and intermediate sequence MSA is constructed using these intermediate sequences.

[0031] Technical solution 13: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 12, wherein, in step c, the sequence search tool is HHblits.

[0032] Technical solution 14: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solutions 12 or 13, wherein, in step c, the database is UniRef30.

[0033] Technical Solution 15: Preferably, the method for generating peptide sequences using an intermediate sequence MSA according to any one of the above technical solutions 12-14, wherein, in step c, the intermediate sequence MSA is in A3M format.

[0034] Technical solution 16: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to any one of the above technical solutions 12-15, wherein in step c, redundant sequences are further removed from the intermediate sequence MSA, and MSA files with too low alignment depth are filtered out.

[0035] Technical Solution 17: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to any one of the above technical solutions 1-16, wherein, in step d, an OADM diffusion-like language masking strategy is introduced, a separate encoder model framework is used, and a 12-layer MSA Transformer is used as the backbone encoding structure.

[0036] Technical Solution 18: Preferably, according to the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 17, wherein, in step d, a directional random masking strategy is adopted during training to mask the alignment position column corresponding to the peptide on the intermediate sequence, that is, to mask all characters of the corresponding column in the entire MSA, so that the model learns the ability to reconstruct the original information of the peptide and realizes evolutionary feature-driven sequence generation.

[0037] Technical solution 19: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solutions 17 or 18, wherein, in step d, the goal of model training is to recover the correct amino acid sequence.

[0038] Technical solution 20: Preferably, a method for generating peptide sequences using an intermediate sequence MSA according to any one of the above technical solutions 1-19, wherein, in step e, the generation task is implemented in a de novo design mode.

[0039] Technical solution 21: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to the above technical solution 19 or 20, wherein in step e, the generation task is performed in a conditional generation mode.

[0040] Technical solution 22: Preferably, the method for generating peptide sequences using intermediate sequence MSA according to any one of the above technical solutions 19-21, wherein in step e, the generation task is achieved by peptide sequence evolution mode.

[0041] Technical Solution 23: In a second aspect, the present invention also provides an AI model for generating peptide sequences using intermediate sequence MSA, including a peptide dataset, peptide sequence MSA, intermediate sequence mapping of peptide sequences, intermediate sequence MSA, a separate encoder model framework, and a 12-layer MSA Transformer as the backbone coding structure.

[0042] Technical solution 24: Preferably, according to the AI ​​model of the above technical solution 23, the peptide dataset includes UniProtKB.

[0043] Technical solution 25: Preferably, according to the AI ​​model of the above technical solution 23, the peptide sequence MSA is constructed by the method of generating peptide sequences using intermediate sequence MSA according to the first aspect of the present invention.

[0044] Technical solution 26: Preferably, according to the AI ​​model of the above technical solutions 23 or 24, the intermediate sequence mapping of the peptide sequence is constructed by the method of generating peptide sequences using intermediate sequence MSA according to the first aspect of the present invention.

[0045] Technical Solution 27: Preferably, according to the AI ​​model of any one of the above technical solutions 23-25, the intermediate sequence MSA is constructed by the method of generating peptide sequences using the intermediate sequence MSA as described in any one of the first aspect technical solutions 12-16 of the present invention.

[0046] Technical Solution 28: In a third aspect, the present invention also provides a training method for the AI ​​model according to any one of the technical solutions 23-27 of the second aspect of the present invention. The method uses a peptide sequence MSA constructed using the method for generating peptide sequences using intermediate sequence MSA as described in any one of the technical solutions 1-6 of the first aspect of the present invention, an intermediate sequence mapping of the peptide sequence constructed using the method for generating peptide sequences using intermediate sequence MSA as described in any one of the technical solutions 7-11 of the first aspect of the present invention, and an intermediate sequence MSA constructed using the method for generating peptide sequences using intermediate sequence MSA as described in any one of the technical solutions 12-16 of the first aspect of the present invention for model training. The method introduces a diffusion-like language masking strategy of OADM for training. During training, a directional random masking strategy is used to mask the alignment position column corresponding to the peptide on the intermediate sequence, that is, to mask all characters in the corresponding column in the entire MSA, enabling the model to learn the ability to reconstruct the original information of the peptide and achieve evolutionary feature-driven sequence generation.

[0047] Through the above technical solution, this invention utilizes the evolutionary information of the intermediary sequence as a "mediator" to achieve the connection between the peptide and the complete MSA structure. This allows peptide sequences that were originally too short to construct an MSA to be embedded into a structured multi-sequence semantic space for effective modeling. This method combines the advantages of generative diffusion models in sequence modeling with the ability of MSATransformer to capture evolutionary context, possessing good versatility and scalability. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the process of obtaining evolutionary information from peptide sequences through intermediate sequences and constructing MSA in this invention.

[0049] Figure 2 This is a schematic diagram of the AI ​​model for targeted masking diffusion training based on intermediate sequence MSA according to the present invention. Detailed Implementation

[0050] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0051] This invention aims to provide a method for generating peptide sequences using an intermediate sequence MSA, the method comprising the following steps:

[0052] a. Construct the peptide sequence MSA,

[0053] b. Constructing intermediate sequence mappings for peptide sequences.

[0054] c. Construct the mediator sequence MSA.

[0055] d. Training masking diffusion modeling based on MSA.

[0056] e. Use the trained model to perform the generation task.

[0057] Step a is implemented as follows: The peptide dataset includes sequences of length 2-50 amino acids from UniprotKB. First, the sequences are quality controlled and deduplicated. Then, a sequence search tool such as HHblits is used to search for homologous sequences for each sequence in the UniRef30 database, constructing an A3M format multiple sequence alignment file. For sequences where MSA cannot be constructed (especially peptide sequences shorter than 20 amino acids) or where the MSA depth is too small (below 64), additional intermediate sequence mapping is performed.

[0058] UniProtKB is a collection of protein databases that have been validated by experts.

[0059] In biology, "aa" refers to amino acids, which are the basic building blocks of proteins. 5aa means five amino acids, and so on.

[0060] The main purpose of sequence quality control is to remove low-quality sequences, ensure that sequencing quality meets analytical requirements, and monitor the reliability of the experimental process through quality assessment. Specifically, removing low-quality sequences includes those containing N bases, sequences with sequencing quality values ​​(Q values) below a set threshold (e.g., Q < 20), and fragments that need to be discarded due to excessively high sequencing error rates. Assessing sequencing quality involves statistically analyzing parameters such as base quality distribution and error rate distribution to determine if there are systematic biases in the sequencing process (e.g., instrument performance issues or improper sample processing). Optimizing subsequent analyses ensures that the data can be used for complex analyses (e.g., assembly, diversity analysis) and avoids result bias caused by low-quality data. Commonly used tools, such as fastx_toolkit, are used to statistically analyze base quality distribution, set quality thresholds (e.g., Q ≥ 20) to filter sequences, and use software such as QIIME or Trimmomatic to remove low-quality tail bases, retaining high-quality core sequences.

[0061] HHblits is a tool for long-range protein sequence alignment. It is an iterative alignment method based on Hidden Markov Models (HMMs) used to identify similarities between structurally and functionally related protein sequences.

[0062] UniRef is an important database in the fields of biology and biochemistry. It integrates information from multiple protein databases and is designed to help researchers identify and classify protein sequences.

[0063] MSA depth refers to the number of aligned sequences in an MSA.

[0064] Step b is implemented as follows: For sequences that cannot construct a valid MSA or have an MSA depth too small (below 64), a BLAST tool is used to search for similar sequences in the UniRef100 database. The screening criteria are: identity greater than or equal to 90%, preferably 100%, and length greater than the peptide itself but not exceeding 10 times its length, as candidate mediator sequences. Depending on computational resources, further screening can be performed to retain several mediator sequences corresponding to each peptide, or only the mediator sequence with the highest bitscore or the largest constructed MSA depth can be retained. The peptide sequence, the mediator sequence, and the mapping interval of the peptide on the mediator sequence are obtained.

[0065] BLAST stands for Basic Local Alignment Search Tool, a commonly used software tool in bioinformatics. It can compare input nucleic acid or protein sequences with known sequences in a database to obtain information such as sequence similarity, thereby determining the origin or evolutionary relationship of the sequences.

[0066] The bitscore value is a key metric in BLAST alignment analysis, used to quantify sequence similarity. A higher value indicates a lower probability of sequence similarity and is often used to assess sequence uniqueness or homology.

[0067] Step c is implemented in the same way as the peptide sequence MSA construction method in step a. A sequence search tool such as HHblits is used to search for homologous sequences for each intermediate sequence in the UniRef30 database, construct an A3M format multiple sequence alignment file, and further remove redundant sequences and filter MSA files with too low alignment depth.

[0068] After step ac, approximately 190,000 peptide sequence MSAs and approximately 60,000 mediator sequence MSAs were finally obtained for model training.

[0069] Step d is implemented by introducing a diffusion-like language masking strategy from the Order-Agnostic Diffusion Model (OADM) and using a separate encoder model framework with a 12-layer MSATransformer as the backbone encoding structure. During training, a directed random masking strategy is employed to mask the alignment column corresponding to the peptide in the intermediate sequence. That is, all characters in the corresponding column are masked throughout the MSA, allowing the model to focus on learning the ability to reconstruct the original peptide information, thereby achieving evolutionary feature-driven sequence generation.

[0070] Order-agnostic diffusion models (OADMs) are a subclass of diffusion models that do not require ordering constraints during training or generation. Under certain assumptions, they can generalize autoregressive models and absorb discrete diffusion. OADMs do not require causal masking of the model representation, meaning they can be trained more efficiently and support parallel generation at test time. They can adapt to any given generation budget and require fewer steps to achieve the same performance as discrete diffusion models.

[0071] Standalone encoder model framework: This refers to the encoder part of the Transformer architecture, primarily used to process input sequences and extract features. Its core structure includes a multi-head self-attention mechanism and a feedforward neural network, achieving efficient data encoding by capturing long-range dependencies and extracting complex features. The encoder is its core component, consisting of multiple identical layers stacked together, each layer containing two sub-layers.

[0072] After model pre-training, step e, namely peptide sequence generation, can be performed. The model can implement three modes: de novo design, conditional design, and peptide sequence evolution. De novo design involves generating sequences from a given sequence length using a complete mask (*) of that length. Conditional design mode inputs a MSA (which can be a regular MSA or an intermediate MSA) and the desired peptide sequence length into the model. This input MSA is called the conditional MSA. If the conditional MSA depth is less than 64, the model will first pad it to 64, and then use this MSA as a basis to generate the required sequence length. In sequence evolution, an MSA or intermediate MSA representing the target sequence is input, along with the range of evolutionary mutations, and the evolved sequence is output.

[0073] The present invention will be further illustrated below with a representative embodiment.

[0074] The method for implementing step a is a conventional method in this field and will not be described in detail here. The following mainly describes step be.

[0075] 1: Construction of mediator mapping and multiple sequence alignment for peptide sequences

[0076] In this embodiment, a 12-amino acid peptide sequence RWRCWRMRGRRC is used as an example. Similar sequences are searched in the UniRef100 database using a BLAST tool. The screening criteria are: Identity ≥90%, Hit sequence length ≥12 and ≤200 (less than 20 times the peptide sequence length, preferably less than 17 times). After obtaining multiple candidate mediator sequences, the candidate mediator sequence with the longest align_len is selected. If there are multiple candidates, the sequence with the highest bitscore is selected as the final mediator sequence. The start and end positions of the peptide on the mediator sequence are recorded. For example, if the peptide completely matches positions 45 to 56 of the mediator sequence, this position is used as the mask region.

[0077] Then, the HHblits tool is used to retrieve a set of homologous sequences of the intermediate sequence from the UniRef30 database, generating an MSA file in A3M format. If the depth of the aligned sequences is greater than 64, the MSA file is retained for subsequent modeling.

[0078] 2: Training for directional masking diffusion modeling based on MSA

[0079] Using peptide sequence MSA and intermediate sequence MSA files as the training set, an AI training framework was employed for modeling. The model backbone is the MSA Transformer, which takes A3M format MSA as input and outputs an embedding vector sequence. The masking diffusion training mechanism is constructed as follows:

[0080] In MSA, the columns that map the corresponding peptides to the intermediate sequence are masked, that is, these columns in the entire MSA are set to the * symbol.

[0081] Masking is performed using a random ratio from 0 to 100, where the masking process is independent of the sequence order.

[0082] The goal of model training is to recover the correct amino acid sequence.

[0083] The Loss function is computed only within the mask region to focus on peptide generation effects.

[0084] This strategy enables the indirect establishment of peptide evolutionary context through intermediate sequences in the absence of evolutionary background information in the whole sequence, effectively making up for the blind spots of traditional MSA Transformer in peptide processing.

[0085] 3: Model Usage

[0086] After the model is trained, it can perform generation tasks. Take the conditional design pattern as an example.

[0087] For example, the known sequence FTCDVLSAEGSFRGVSVKLNHSACATHCLFLKKRGGYCNNKAICVCRN (https: / / pubmed.ncbi.nlm.nih.gov / 26522790 / ) is an antimicrobial peptide with antimicrobial function, and its length is 48 amino acids. First, the sequence search tool HHblits is used to search for its homologous sequences in the UniRef30 database, and a multiple sequence alignment file is constructed.

[0088] Then, this multiple sequence alignment file is input into the model, along with the required sequence length, which is a random 6-15 amino acids, to reduce the synthesis cost of the designed novel antimicrobial peptides and enhance their industrial value. The model is run 9 times, generating nine sequences. The nine sequences generated and output by the model are as follows:

[0089] 1. TCDLLSVETKKL

[0090] 2. INCNKHCIK

[0091] 3. WCWKRWRKRGRCW

[0092] 4. KSKKKKKK

[0093] 5. MITNLKHKTC

[0094] 6. KFAPNHAF

[0095] 7. AGS VN WND GLC AVONC

[0096] 8. AVGLFHKHACCAARC

[0097] 9. HALNLGRRAGKCDS

[0098] The generated sequences are novel and have not been reported in any antimicrobial peptide databases. They can be used to overcome the limitations of sequencing databases and explore and discover novel antimicrobial peptides.

[0099] The generated sequences were then scored and sorted using computational methods, such as the discriminator and scorer modules of AMPGen. The nine sequences were input into the AMPGen discriminator, which outputs a binary classification result to identify the sequence set belonging to the antimicrobial peptide sequence. The sequence set belonging to the antimicrobial peptide sequence was then input into the AMPGen scorer, which outputs a score for the antimicrobial performance against *E. coli* (see Table 1). The lower the score, the better the predicted antimicrobial performance. The selection criteria were set as follows: a discriminator result of 1, a scorer threshold of less than 2, and the sequence containing two or more spaced amino acid Cs to facilitate disulfide ring formation and increase stability.

[0100]

[0101] Sequence 3 was the best candidate sequence that met the requirements, and it was chemically synthesized and verified by wet experiments.

[0102] Sequence 3 was chemically synthesized, and the minimum inhibitory concentration (MIC) (μg / ml) was determined for different bacteria. The results are shown in Table 2 below:

[0103]

[0104] The above describes a method for designing and obtaining a novel functional antimicrobial peptide sequence that does not exist in the database using this model.

[0105] The above embodiments are only intended to explain the present invention in more detail and to make it easier for those skilled in the art to understand the present invention, and are not intended to limit the present invention. It is understood that various improvements and changes can be made to the present invention without departing from the spirit and principles of the present invention, and such improvements and changes are all within the protection scope of the present invention.

Claims

1. A method for generating peptide sequences using intermediate sequence MSA, comprising the following steps: a. Constructing a peptide sequence MSA, b. Constructing an intermediate sequence map of the peptide sequence, c. Constructing an intermediate sequence MSA, d. Masked diffusion modeling training based on the constructed peptide sequence MSA and the constructed intermediate sequence MSA as needed, and e. Implementing the generation task with the trained model.

2. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 1, wherein, In step a, polypeptide sequences of 2-50 aa length are selected from a peptide dataset, and after quality control and deduplication, sequence search tools are used to search for homologous sequences for each sequence in the database, which are peptide sequences, and these peptide sequences are used to construct a peptide sequence MSA.

3. The method for generating a peptide sequence using an intermediate sequence MSA according to claim 2, wherein, In step a, the peptide dataset is UniProtKB.

4. The method for generating a peptide sequence using an intermediate sequence MSA according to claim 2, wherein, In step a, the sequence search tool is HHblits.

5. The method for generating a peptide sequence using an intermediate sequence MSA according to claim 2, wherein, In step a, the database is UniRef30.

6. The method for generating a peptide sequence using an intermediate sequence MSA according to claim 1, wherein, In step a, the peptide sequence MSA is in A3M format.

7. The method of generating a peptide sequence using an intermediate sequence MSA according to any one of claims 1 to 6, wherein, In step b, BLAST tools are used to search for similar sequences in the UniRef100 database for sequences that cannot construct a peptide sequence MSA in step a and sequences with MSA depth less than or equal to 64, and candidate intermediate sequences are screened out.

8. The method for generating a peptide sequence using an intermediate sequence MSA according to claim 7, wherein, In step b, the screening condition is that the identity is greater than or equal to 90%, the length is greater than the peptide itself and does not exceed 20 times the length of the peptide.

9. The method for generating a peptide sequence using an intermediate sequence MSA according to claim 7, wherein, In step b, the candidate intermediate sequences can be further screened to retain a number of intermediate sequences corresponding to each peptide.

10. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 9, wherein, In step b, the further screening condition is to retain the intermediate sequence with the largest bitscore value.

11. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 9, wherein, In step b, the further screening condition is to retain the intermediate sequence with the largest MSA depth.

12. The method of generating a peptide sequence using an intermediate sequence MSA according to any one of claims 1-6, wherein, In step c, sequence search tools are used to search for homologous sequences for each intermediate sequence in the database, which are intermediate sequences, and these intermediate sequences are used to construct an intermediate sequence MSA.

13. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 12, wherein, In step c, the sequence search tool is HHblits.

14. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 12, wherein, In step c, the database is UniRef30.

15. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 12, wherein, In step c, the intermediate sequence MSA is in A3M format.

16. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 12, wherein, In step c, the intermediate sequence MSA is further removed from redundant sequences, and MSA files with too low alignment depth are filtered out.

17. The method of generating a peptide sequence using an intermediate sequence MSA according to any one of claims 1-6, wherein, In step d, the diffusion pattern masking strategy of OADM is introduced, and a separate encoder model framework is used, with a 12-layer MSA Transformer as the main encoding structure.

18. The method of generating a peptide sequence using a mediating sequence (MSA) according to claim 17, wherein, In step d, a directional random masking strategy is used during training, and the corresponding column of the peptide in the intermediate sequence is masked, i.e., all characters in the corresponding column in the entire MSA are masked, so that the model learns the reconstruction ability of the original information of the peptide and realizes the sequence generation driven by evolutionary characteristics.

19. The method of generating a peptide sequence using a mediating sequence (MSA) according to claim 17, wherein, In step d, the goal of model training is to restore the correct amino acid sequence.

20. The method of generating a peptide sequence using an intermediate sequence MSA according to any one of claims 1-6, wherein, In step e, the generation task is implemented in a de novo design mode.

21. The method for generating a peptide sequence using a mediating sequence (MSA) according to claim 19, wherein, In step e, the generation task is implemented in a conditional generation mode.

22. The method of generating a peptide sequence using a mediating sequence (MSA) according to claim 19, wherein, In step e, the generation task is implemented in a peptide sequence evolution mode.

23. A generation method of an AI model for generating a peptide sequence using a mediating sequence (MSA), comprising the following steps: 1) constructing a peptide sequence MSA using a peptide dataset, 2) constructing a mediating sequence mapping and a mediating sequence MSA of the peptide sequence as needed, and 3) using the peptide sequence MSA and the mediating sequence MSA for masked diffusion AI model training; wherein, A separate encoder model framework and a 12-layer MSA Transformer as the main encoding structure are used to process input sequences and extract features for training the model. 24.The method of claim 23, wherein, The peptide dataset comprises UniProtKB. 25.The method of claim 23, wherein, The peptide sequence MSA is constructed according to the method for generating a peptide sequence using an intermediate sequence MSA according to any one of claims 1-6. 26.The method of Claim 23 or 24, wherein, The intermediate sequence mapping of the peptide sequence is constructed according to the method for generating a peptide sequence using an intermediate sequence MSA according to any one of claims 7-11.

27. The method of Claim 23-25, wherein, The intermediate sequence MSA is constructed according to the method for generating a peptide sequence using an intermediate sequence MSA according to any one of claims 12-16.

28. A training method of an AI model for generating a peptide sequence using an intermediate sequence MSA, the peptide sequence MSA constructed according to the method for generating a peptide sequence using an intermediate sequence MSA according to any one of claims 1-16 and the intermediate sequence MSA constructed as required are used for model training; the training is performed by introducing the diffusion-like language mask strategy of OADM, and the alignment position column corresponding to the peptide on the intermediate sequence is masked during the training, that is, all characters in the corresponding column in the entire MSA are masked, so that the model learns the reconstruction ability of the original information of the peptide and realizes the sequence generation driven by the evolutionary characteristics.

Images