Training method and device of interaction protein generation model

Abstract

The invention provides a training method and device for an interaction protein generation model, and the method comprises the steps: obtaining a sample protein pair which comprises a sample protein and a sample interaction protein of the sample protein; extracting and fusing sequence characteristics and spatial structure characteristics of each amino acid residue in the sample protein pair to obtain residue characteristics of the sample protein pair; generating a sample predicted protein of the sample protein based on an initial generation model and the residue features; and constructing generation loss based on the sample prediction protein, and performing parameter iteration on the initial generation model based on the generation loss to obtain an interaction protein generation model. According to the method provided by the invention, the first-order sequence information and the three-dimensional structure information of the protein are considered at the same time during training, so that the accuracy of generating the interaction protein is greatly improved, and efficient and universal generation of the interaction protein with affinity is realized.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a training method and device for generating a model of interacting proteins. Background Art

[0002] Antibodies and antigens are interactive proteins. Traditional methods of antibody production are mainly through animal immunization, phage display technology, or hybridoma technology to achieve the production of antibody proteins. Among them, the animal immunization method is to inject antigens into animals to induce the animal immune system to produce antibodies, and then extract antibodies from the animal serum; phage display technology uses phage display systems to screen antibody fragments that bind to specific antigens in vitro; hybridoma technology is to fuse immune B cells and myeloma cells to produce hybridoma cells that can proliferate indefinitely and secrete monoclonal antibodies.

[0003] However, traditional methods for generating interacting proteins all have common defects, namely, they have long cycles, require a lot of resources, and are only applicable to the generation of individual interacting proteins. Summary of the invention

[0004] The present invention provides a training method and device for generating a model of interacting proteins, which are used to solve the defects of the generating methods of interacting proteins in the prior art, such as long cycle, high resource cost and low versatility.

[0005] The present invention provides a training method for generating a model of interacting proteins, comprising: obtaining a sample protein pair, wherein the sample protein pair comprises a sample protein and a sample interacting protein of the sample protein; Extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; generating a sample predicted protein of the sample protein based on the initial generation model and the residue features; Predicting protein construction generation loss based on the sample, performing parameter iteration on the initial generation model based on the generation loss, and obtaining an interaction protein generation model.

[0006] According to a training method for an interaction protein generation model provided by the present invention, the sample predicted protein of the sample protein is generated based on the initial generation model and the residue features, comprising: Based on the residue features of the sample protein and the above residue features in the sample interacting protein, instructing the initial generation model to gradually generate sample amino acid residues at each site; Based on the sample amino acid residues at each site, obtaining the sample predicted protein; The above residue features include residue features of the amino acid residues preceding the current generated site in the sample interacting proteins.

[0007] According to a training method for an interacting protein generation model provided by the present invention, the initial generation model is constructed based on a large language model; and the large language model is constructed based on a decoder.

[0008] According to a training method for generating a model of interacting proteins provided by the present invention, the sample protein pairs include general sample protein pairs and fine-tuning sample protein pairs; The step of predicting protein construction generation loss based on the sample, and iterating parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model comprises: Constructing a universal generation loss based on sample predicted proteins of the universal sample protein pair, and iterating parameters of the initial generation model based on the universal generation loss to obtain a universal protein generation model; A fine-tuning generation loss is constructed based on the sample predicted protein of the fine-tuning sample protein pair, and based on the fine-tuning generation loss, parameters of the general protein generation model are fine-tuned to obtain the interacting protein generation model.

[0009] The present invention also provides a method for producing an interacting protein, comprising: Obtaining a reference protein sequence; extracting and fusing reference sequence features of each amino acid residue in the reference protein sequence; Based on the reference sequence features and the interacting protein generation model, generating the interacting protein corresponding to the reference protein sequence; The interacting protein generation model is obtained based on the training method for generating an interacting protein model as described in any one of the above items.

[0010] According to a method for generating an interacting protein provided by the present invention, the interacting protein corresponding to the reference protein sequence is generated based on the reference sequence characteristics and the interacting protein generation model, comprising: Based on the reference sequence features and the sequence features of the generated sites, instructing the interacting protein generation model to gradually generate the amino acid residues of the next site; Based on the generated amino acid residues at each site, the interacting proteins corresponding to the reference protein sequence are obtained.

[0011] The present invention also provides a training device for generating a model of interacting proteins, comprising: An acquisition unit is configured to acquire a sample protein pair, wherein the sample protein pair comprises a sample protein and a sample interacting protein of the sample protein; A feature extraction unit extracts and fuses the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; A prediction unit, which generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; The training unit predicts protein construction generation loss based on the sample, and iterates parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model.

[0012] The present invention also provides an interactive protein generating device, comprising: Obtain sequence units, obtain reference protein sequences; A sequence feature extraction unit, which extracts and fuses reference sequence features of each amino acid residue in the reference protein sequence; A generating unit, which generates the interacting protein corresponding to the reference protein sequence based on the reference sequence characteristics and the interacting protein generation model; The interacting protein generation model is obtained based on the training method for generating an interacting protein model as described in any one of the above items.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the training method for generating a model of interacting proteins as described in any one of the above, or the method for generating interacting proteins as described in any one of the above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a training method for generating a model of interacting proteins as described in any one of the above, or a method for generating interacting proteins as described in any one of the above.

[0015] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the training method for generating a model of interacting proteins as described in any one of the above, or the method for generating interacting proteins as described in any one of the above.

[0016] The training method and device of the interaction protein generation model provided by the present invention obtain the residue features of the sample protein pair by extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair, and generate a sample prediction protein of the sample protein based on the initial generation model and the residue features; a generation loss is constructed based on the sample prediction protein, and the initial generation model is parameterized based on the generation loss to obtain the interaction protein generation model. The first-order sequence information and three-dimensional structure information of the protein are simultaneously considered during training, which greatly improves the accuracy of generating the interaction protein and realizes efficient, universal and affinity-based generation of interaction proteins. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a schematic diagram of the flow chart of the training method for generating a model of interacting proteins provided by the present invention; Figure 2 It is a schematic diagram of a process for obtaining residue features of a sample protein pair provided by the present invention; Figure 3 This is one of the schematic diagrams of the process for generating the interacting protein provided by the present invention; Figure 4 This is the second schematic diagram of the process for generating the interacting protein provided by the present invention; Figure 5 It is a schematic diagram of the structure of a training device for generating a model of interacting proteins provided by the present invention; Figure 6 is a schematic diagram of the structure of the interactive protein generating device provided by the present invention; Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] It should be noted that traditional antibody production methods mainly include the following: (1) Animal immunization method: by preparing the target antigen into a form suitable for injection, it is usually necessary to mix it with an adjuvant to enhance the immune response. Then, the antigen is injected into the animal (such as mice, rabbits, etc.), usually multiple injections every few weeks, and the immunized animals (such as mice, rabbits) produce polyclonal antibodies or monoclonal antibodies. A few weeks after the immunization injection, the blood of the animal is collected and polyclonal antibodies are extracted from the serum. However, this method is time-consuming, usually taking several months, and requires complex screening and purification steps. In addition, it is not suitable for the production of interacting proteins other than antibodies.

[0021] (2) Phage display technology: First, a phage display library containing a large number of different antibody fragments is constructed, and each phage displays an antibody fragment. Then, the target antigen is fixed on a solid phase carrier and mixed with the phage library, so that the phage displaying the specific antibody fragment binds to the antigen. Next, the unbound phage is eluted, and the bound phage is retained and amplified. Repeat multiple rounds of screening to improve the specificity and affinity of the antibody fragment. Finally, the screened phage is sequenced to identify the antibody fragment sequence, which is then expressed and purified in vitro. Using the phage display system, antibody fragments that bind to specific antigens are screened in vitro. Although this method is more efficient than animal immunization, it still requires a large number of experimental operations and screening steps, still consumes a lot of time and resources, and is not very versatile.

[0022] (3) Hybridoma technology: First, the antigen is injected into the mouse to stimulate its B cells to produce antibodies. Then, the B cells in the mouse spleen are mixed with myeloma cells and cell fusion is performed using polyethylene glycol (PEG) or electrofusion technology. Next, the fused cells are cultured in a selective culture medium (such as HAT culture medium) to screen out hybridoma cells that have successfully fused and can proliferate indefinitely. The screened hybridoma cells are cloned and cultured to ensure that monoclonal antibodies are obtained. Finally, the hybridoma cells are amplified, and the antibodies in their culture supernatant are collected and purified. That is, hybridoma technology produces monoclonal antibodies by fusing B cells and myeloma cells to produce hybridoma cells with unlimited proliferation ability. This method also requires a lot of time and resources and is not very versatile.

[0023] In view of the above problems, the present invention provides a training method for an interacting protein generation model to achieve efficient, more versatile and more affinity interacting protein generation. Figure 1 is a flow chart of the training method for generating a model of interacting proteins provided by the present invention, such as Figure 1 As shown, the method includes: Step 110, obtaining a sample protein pair, wherein the sample protein pair comprises a sample protein and a sample interacting protein of the sample protein; Here, the sample protein pair refers to two interacting proteins, which may include sample proteins and sample interacting proteins that interact with sample proteins. For example, antigens and antibodies, antibodies can specifically recognize and bind to antigens, thereby triggering an immune response; for example, transport proteins and receptor proteins regulate the exchange of substances inside and outside cells and signal transduction through interaction. It should be noted that interacting proteins may include enzyme complexes, signal transduction proteins, structural protein complexes, interacting proteins in the immune system, and other types. In order to improve the versatility and training effect of the interacting protein generation model, the sample protein pairs here may include various types of interacting protein pairs.

[0024] Specifically, various types of interacting protein pairs can be obtained as sample protein pairs. Then, the structural information of the sample protein pairs can be obtained by searching in the PDB (Protein Data Bank). The sequence information here may include the types of amino acids in the protein and their arrangement order on the protein chain. The structural information here may be the spatial position information of each atom in the protein. It can be understood that since the sequence information is always greater than the structural information, the corresponding structural information of the remaining protein sequences that do not have corresponding structural information can be obtained from the alphafold database. Therefore, the sequence information and structural information of the sample protein pairs can be used as sample data sets for the training of the interacting protein generation model. For example, the representation of a sequence of a protein: A = (a1, a2..., a n ), A represents a protein sequence, a1, a2..., a n represents the sequence information and structural information of the amino acid residues at each site of a protein, and n represents the number of amino acids. Similarly, protein B is represented by B = (b1,b2...,b n ). Then for a sample protein pair, the complete input x=(A,B). For all sample protein pairs in a data set, D = {x 1 ,x 2 ,…,x |D|}, where D represents the total number of sample data in the data set.

[0025] Step 120, extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; Specifically, the sample protein pair can be input into the encoding model, and the feature extraction of the sample protein pair can be performed through the encoding model. It should be noted that different encoders are usually used to extract the corresponding features for extracting the sequence features and structural features of each amino acid residue in the protein. For example, the sequence features of the sample protein pair can be extracted by ESM2 (Evolutionary SequenceModel 2). The structural features of the sample protein pair can be extracted by a pre-trained encoder, such as GVP transformer (Geometric Vector Perceptron Transformer). Then, the sequence features and structural features of the sample protein pair are fused, such as splicing, to obtain the residue features of the sample protein pair.

[0026] Figure 2 is a schematic diagram of a process for obtaining residue features of a sample protein pair provided by the present invention, such as Figure 2 As shown in the figure, the sample protein pair includes Sequence A (protein sequence A) and Sequence B (protein sequence B). Sequence A and Sequence B are input into the sequence encoder for sequence encoding (Sequence embedding) to obtain the sequence features of the sample protein pair. At the same time, Sequence A and Sequence B can be input into the GVP transformer, and structure encoding (structure embedding) can be performed through the GVP transformer to obtain structural features. Finally, the sequence features of Sequence A and its structural features can be fused, and the sequence features of Sequence B and its structural features can be feature fused to obtain the fusion features of Sequence A and Sequence B (feature fusion), that is, the residue features.

[0027] It should be noted that when extracting features, the extracted features can be divided into feature scalar information and feature vector information, and the extracted feature scalar information and feature vector information are represented in the form of a tuple. Among them, the feature scalar information contains the invariant features of the sequence features and the spatial structure. The feature vector information is related to the spatial structure and does not depend on the spatial position of the biological structure and the transformation of the coordinate system.

[0028] Step 130, generating a sample predicted protein of the sample protein based on the initial generation model and the residue features; Here, the initial generative model can be constructed based on the decoder. Specifically, the residue features can be input into the initial generative model, and the initial generative model predicts the interacting protein corresponding to the sample protein through the residue features of the sample protein and the residue features of the sample interacting protein, that is, generates a sample predicted protein. It can be understood that if the training effect of the initial generative model is good, the sample predicted protein here can be a protein with the same protein sequence and structure as the sample interacting protein.

[0029] It should be noted that the residue features here include the sequence features and structural features of proteins. The structural features can include the spatial position information of each amino acid residue in the protein sequence, and can especially reflect the spatial position information of the side chain groups in the amino acid residues compared to the main chain atoms (such as Ca atoms). The position information of the side chain groups in the amino acid residues can reflect the surface characteristics of the protein, form the binding interface for the interaction between proteins, and can affect the specificity and strength of the interaction, and participate in the recognition and binding of proteins.

[0030] Therefore, compared with the prior art that predicts protein generation only by sequence features, the sequence design based on illusion often results in proteins that cannot be folded, and it is difficult to design larger proteins. The residue features in the embodiments of the present invention include the side chain groups of amino acid residues compared to the spatial structure of the main chain atoms, which can be combined with the sequence features of the protein to more accurately predict the sample predicted protein of the sample protein. That is, the spatial structure features can enable the initial generation model to learn the subtle spatial features of the interaction region in the sample protein pair, and control the initial generation model to generate interacting proteins with higher affinity.

[0031] In addition, some interacting proteins, such as antibodies, are very sensitive to local three-dimensional spatial structures. Therefore, during the training process of the interacting protein generation model, strong local structural features are input into the model to make the trained interacting protein generation model more versatile.

[0032] It should also be noted that the method provided in the embodiment of the present invention predicts the interacting proteins of the sample proteins in the sample protein pair through the residue features of the two protein sequences of the sample protein pair, so that the model not only learns the residue information between the two interacting proteins, but also learns the association between the residues of the protein chain to be generated itself, thereby making the sequence and structure of the predicted residue protein generated by the trained model more accurate, and the affinity of the generated interacting proteins higher.

[0033] Step 140, predicting protein construction generation loss based on the sample, performing parameter iteration on the initial generation model based on the generation loss, and obtaining an interaction protein generation model.

[0034] Specifically, the generation loss can be constructed by predicting the sample protein. Alternatively, the sample interaction protein can be used as a label, and the generation loss can be calculated by the sample interaction protein and the sample prediction protein. Then, the parameters of the initial generation model are iterated based on the generation loss to obtain the final interaction protein generation model. For example, the parameters of the initial generation model can be adjusted by minimizing the negative log-likelihood as the generation loss. The generation loss here can be calculated by the following formula, as shown below:

[0035] In the formula, represents the generation loss; D represents the total number of samples in the sample protein pair dataset; represents the kth sample protein pair; n represents the number of amino acid residues in a sample protein; i represents the amino acid residue at the i-th position in the protein sequence; represents the model parameters of the initial generative model; represents the amino acid residue at the i-th position in the protein predicted by the k-th sample; represents the amino acid residue before the i-th site in the k-th sample interacting protein; represents the initial generative model (parameters ) generates the probability that the k-th sample predicts the amino acid residue at the i-th site in the protein under the condition that the k-th sample interacts with the amino acid residue before the i-th site in the protein.

[0036] It should be noted that compared with the existing technology of generating antibody proteins based on animal immunization, phage display technology, and hybridoma technology, the generation of interacting proteins based on the trained interacting protein generation model greatly reduces the time cost and resource consumption.

[0037] The method provided by the embodiment of the present invention obtains the residue features of the sample protein pair by extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair, and generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; constructs a generation loss based on the sample predicted protein, and iterates the parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model. During training, the first-order sequence information and three-dimensional structure information of the protein are simultaneously considered, which greatly improves the accuracy of generating interaction proteins and realizes efficient, universal and affinity-based generation of interaction proteins.

[0038] Based on any of the above embodiments, step 130 includes: Based on the residue features of the sample protein and the above residue features in the sample interacting protein, instructing the initial generation model to gradually generate sample amino acid residues at each site; Based on the sample amino acid residues at each site, obtaining the sample predicted protein; The above residue features include residue features of the amino acid residues preceding the current generated site in the sample interacting proteins.

[0039] Here, the above residue feature refers to the residue feature of the amino acid residue before the current generation site in the interacting protein. It can be understood that the sample prediction protein can be understood as a sentence, and each amino acid residue in the sample prediction protein can be understood as a token corresponding to each word in the sentence. Since there is often a certain connection between tokens to generate a complete sentence, in the training of the language model, the input generally includes all the tokens in the sentence. Based on the mechanism of calculating the attention between tokens by the neural network, the meaning of all tokens in the sentence is learned, and the next word can be predicted based on the context of a sentence. Therefore, the above residue amino acid here can be used to reflect the residue features of the amino acid before the current generation site generated by the prediction.

[0040] Specifically, the residue features of the sample protein and the above residue features in the sample interacting protein can be input into the initial generation model to instruct the initial generation model to gradually generate sample amino acid residues at each site. Then, according to the sample amino acid residues of each site generated in sequence, the protein sequence and structure are obtained, and the obtained protein sequence and structure are used as sample predicted proteins. In particular, the initial generation model uses an autoregressive decoding method from left to right, and the generated protein sequence will be generated from left to right, that is, the sample amino acid residues of each site from left to right are obtained. The generated sample predicted proteins here can be expressed by the following formula, as shown below: ; In the formula, It represents the probability of generating protein B (sample predicted protein) under the condition of protein A (sample protein); Indicates the sample prediction protein The amino acid residue at each position; Indicates the first interacting protein in the sample The residue characteristics of the amino acid residue before the position; Represents the residue characteristics of the entire sample protein.

[0041] The method provided by the embodiment of the present invention provides relevant information between two interacting protein chains through the residue characteristics of the entire sample protein, as well as the contextual information in a single protein chain reflected by the residue characteristics before the current generation site in the sample interacting protein, and performs step-by-step prediction of the amino acid residues at each site in another sample predicted protein corresponding to the sample protein, which greatly improves the accuracy of the sample predicted protein and can be applied to the generation of any type of interacting proteins, thereby improving the versatility of the interacting protein generation model and the accuracy of the generated interacting proteins.

[0042] Based on any of the above embodiments, the initial generation model is constructed based on a large language model; and the large language model is constructed based on a decoder.

[0043] It should be noted that the initial generative model is constructed based on the model structure of the large language model. With the help of the powerful language understanding ability of the large language model, the initial generative model is trained to have a strong understanding ability between "protein languages". It can gradually predict the amino acid residues of the next site based on the connection between interacting proteins and the connection between each site of the protein sequence itself, and obtain a complete sample prediction protein. The large language model here is constructed based on the decoder of Transformer.

[0044] Based on any of the above embodiments, the sample protein pairs include universal sample protein pairs and fine-tuning sample protein pairs; Step 140 includes: Constructing a universal generation loss based on sample predicted proteins of the universal sample protein pair, and iterating parameters of the initial generation model based on the universal generation loss to obtain a universal protein generation model; A fine-tuning generation loss is constructed based on the sample predicted protein of the fine-tuning sample protein pair, and based on the fine-tuning generation loss, parameters of the general protein generation model are fine-tuned to obtain the interacting protein generation model.

[0045] Here, the sample protein pairs can be divided into general sample protein pairs and fine-tuning sample protein pairs. The general sample protein pairs can include various types of interacting protein pairs, and the fine-tuning sample protein pairs can be a certain type of interacting protein pairs, such as antigens and antibodies. It is understandable that if it is necessary to train to generate a certain type of interacting protein, for example, it is necessary to train to obtain an interacting protein generation model for generating antibodies, then antigens and antibodies can be used as fine-tuning sample protein pairs.

[0046] Specifically, in the stage of obtaining the sample data set, the fine-tuning sample protein pairs in the sample protein pairs can be selected, and the remaining sample protein pairs can be used as universal sample protein pairs. In the model training stage, the model training stage can be divided into a universal training stage and a fine-tuning stage. In the universal training stage, the residue features based on the universal sample protein pairs can be input into the initial generation model, and the sample predicted proteins of the universal sample protein pairs can be gradually generated through the initial generation model. Then, a universal generation loss is constructed based on the sample predicted proteins of the universal sample protein pairs. For example, the negative log-likelihood can be minimized as the universal generation loss to iterate the parameters of the initial generation model to obtain a universal protein generation model.

[0047] In order to further improve the affinity of the protein generated by the model, the general protein generation model can be fine-tuned based on the fine-tuning sample protein pairs. Similarly, the residue features of the fine-tuning sample protein pairs can be input into the general protein generation model, and the sample predicted proteins of the fine-tuning sample protein pairs can be gradually generated through the general protein generation model. Then, based on the sample predicted proteins of the fine-tuning sample protein pairs, a fine-tuning generation loss is constructed. For example, the negative log-likelihood can be minimized as the fine-tuning generation loss to iterate the parameters of the general generation model and obtain the mutual protein generation model that is ultimately used to generate a specific type of protein.

[0048] It should be noted that after the fine-tuning training is completed, the sample predicted proteins generated in the fine-tuning training phase can be evaluated. The rationality of the sequence can be judged by calculating the scores of amino acids at all sites in the sample predicted proteins. The evaluation score can be achieved by the following formula, as shown below:

[0049] In the formula, represents the overall evaluation score of the sample predicted protein B; n represents the number of amino acid residues contained in the sample predicted protein B, Indicates The amino acid residue at each position; The model is good at fine-tuning protein A and in Under the condition of the amino acid residue before the position, the sample prediction protein B is predicted The probability of an amino acid residue at a position.

[0050] In addition, the designed sample prediction protein sequence needs to be tested by wet experiment means (such as ITC) to detect whether the antibodies interact with each other, and whether the antibodies have high affinity. Based on the results of the progen enzyme generation in the experimental data, the method of finding characteristics from the family is more conducive to generating antibodies with stronger affinity.

[0051] The method provided by the embodiment of the present invention obtains a universal protein generation model by training data with large-scale universal sample protein pairs. On the basis of the universal protein generation model, fine-tuning training is performed by fine-tuning sample protein pairs to obtain an interaction protein generation model such as an antibody generation model, thereby obtaining a high-affinity and universal interaction protein generation model.

[0052] Based on any of the above embodiments, Figure 3 is one of the flow diagrams of the method for generating the interacting protein provided by the present invention, such as Figure 3 As shown, the method includes: Step 310, obtaining a reference protein sequence; Step 320, extracting and fusing reference sequence features of each amino acid residue in the reference protein sequence; Step 330, generating the interacting proteins corresponding to the reference protein sequence based on the reference sequence features and the interacting protein generation model; The interacting protein generation model is obtained based on the training method for the interacting protein generation model described in any of the above embodiments.

[0053] Specifically, first, a reference protein sequence input by a user can be obtained, such as a certain antigen sequence. Next, the sequence feature of each amino acid residue in the reference protein sequence can be encoded by an encoder. Then, the reference sequence feature of the reference protein sequence is input into the interaction protein generation model, indicating that the interaction protein generation model generates an interaction protein corresponding to the reference protein sequence. It is understandable that the interaction protein generation model provided by the embodiment of the present invention can also be applied to optimize existing antibodies. For example, a prompt word can be constructed based on an existing antibody sequence, and the prompt word can be input into the interaction protein generation model, and the interaction protein generation model is used to generate an optimized antibody that meets the prompt word requirements.

[0054] It should be noted that in the actual application of generating interacting proteins, there may be a situation where the three-dimensional spatial structure information of the reference protein is located. In addition, the sequence information of the protein determines the three-dimensional structure of the protein to a certain extent, and the interacting protein can be generated by only inputting the sequence information of the reference protein to indicate the interacting protein sequence.

[0055] In addition, the interaction protein generation model is trained based on the residue features of sample protein pairs, and the relationship between the interacting proteins and the residues in a single protein chain in the sequence and structure dimensions is learned, so that the generated interacting proteins have a stronger affinity with the reference protein sequence. In addition, it can efficiently and widely realize the interaction proteins corresponding to different reference protein sequences.

[0056] In one embodiment, Figure 4 FIG. 2 is a flow chart of the method for generating the interacting protein provided by the present invention. Figure 4 As shown, the method includes: first, protein A can be sequence-encoded to extract sequence features. Then, the sequence features of protein A are input into a transformer decoder constructed by an N-layer neural network, and the transformer decoder is used to predict the next amino acid residue in protein B one by one. ,but The constructed sequence can be the interacting protein B corresponding to protein A.

[0057] The method provided by the embodiment of the present invention obtains the residue features of the sample protein pair by extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair, and generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; a generation loss is constructed based on the sample predicted protein, and the parameters of the initial generation model are iterated based on the generation loss to obtain an interaction protein generation model. During training, the first-order sequence information and three-dimensional structure information of the protein need to be considered simultaneously, which greatly improves the accuracy of generating interaction proteins and realizes efficient, universal and affinity-based generation of interaction proteins.

[0058] Based on any of the above embodiments, the interacting protein generation model is constructed based on a large language model; the large language model is constructed based on a decoder; Step 330 includes: Based on the reference sequence features and the sequence features of the generated sites, instructing the interacting protein generation model to gradually generate the amino acid residues of the next site; Based on the generated amino acid residues at each site, the interacting proteins corresponding to the reference protein sequence are obtained.

[0059] Specifically, the reference sequence features are input into the interaction protein generation model, and the interaction protein generation model is used to decode one by one to predict the amino acid residue of the next site. In the process of step-by-step decoding, first, the first amino acid residue of the interaction protein sequence is predicted based on the reference sequence features of the complete reference protein sequence. Then, when predicting the generation of the second amino acid residue, the second amino acid residue can be predicted based on the reference sequence features of the complete reference protein sequence and the sequence features of the amino acid residues at the generated sites, that is, the sequence features of the first amino acid residue. And so on, until the amino acid residue of the last site is generated, and finally, the amino acid residues generated at each site are combined to obtain the interaction protein corresponding to the reference protein sequence.

[0060] Based on any of the above embodiments, Figure 5 Schematic diagram of the structure of the training device for generating the interaction protein model provided by the present invention. Figure 5 As shown, the device comprises: An acquiring unit 510 acquires a sample protein pair, wherein the sample protein pair includes a sample protein and a sample interacting protein of the sample protein; A feature extraction unit 520 extracts and fuses the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; A prediction unit 530 generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; The training unit 540 predicts protein construction generation loss based on the sample, and iterates parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model.

[0061] The device provided by the embodiment of the present invention obtains the residue features of the sample protein pair by extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair, and generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; constructs a generation loss based on the sample predicted protein, and iterates the parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model. During training, the first-order sequence information and three-dimensional structure information of the protein are simultaneously considered, which greatly improves the accuracy of generating interaction proteins and realizes efficient, universal and affinity-based generation of interaction proteins.

[0062] Based on any of the above embodiments, the prediction unit is specifically used for: Based on the residue features of the sample protein and the above residue features in the sample interacting protein, instructing the initial generation model to gradually generate sample amino acid residues at each site; Based on the sample amino acid residues at each site, obtaining the sample predicted protein; The above residue features include residue features of the amino acid residues preceding the current generated site in the sample interacting proteins.

[0063] Based on any of the above embodiments, the initial generation model is constructed based on a large language model; and the large language model is constructed based on a decoder.

[0064] Based on any of the above embodiments, the sample protein pairs include universal sample protein pairs and fine-tuning sample protein pairs; The training unit is specifically used for: Constructing a universal generation loss based on sample predicted proteins of the universal sample protein pair, and iterating parameters of the initial generation model based on the universal generation loss to obtain a universal protein generation model; A fine-tuning generation loss is constructed based on the sample predicted protein of the fine-tuning sample protein pair, and based on the fine-tuning generation loss, parameters of the general protein generation model are fine-tuned to obtain the interacting protein generation model.

[0065] Based on any of the above embodiments, Figure 6 Schematic diagram of the structure of the interactive protein generating device provided by the present invention, such as Figure 6 As shown, the device comprises: A sequence acquisition unit 610 acquires a reference protein sequence; A sequence feature extraction unit 620 extracts and fuses reference sequence features of each amino acid residue in the reference protein sequence; A generating unit 630 generates an interacting protein corresponding to the reference protein sequence based on the reference sequence feature and the interacting protein generation model; The interacting protein generation model is obtained based on the training method for the interacting protein generation model described in any one of the above embodiments.

[0066] The device provided by the embodiment of the present invention obtains the residue features of the sample protein pair by extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair, and generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; a generation loss is constructed based on the sample predicted protein, and the parameters of the initial generation model are iterated based on the generation loss to obtain an interaction protein generation model. During training, the first-order sequence information and three-dimensional structure information of the protein need to be considered simultaneously, which greatly improves the accuracy of generating interaction proteins and realizes efficient, universal and affinity-based generation of interaction proteins.

[0067] Based on any of the above embodiments, the generating unit is specifically used for: Based on the reference sequence features and the sequence features of the generated sites, instructing the interacting protein generation model to gradually generate the amino acid residues of the next site; Based on the generated amino acid residues at each site, the interacting proteins corresponding to the reference protein sequence are obtained.

[0068] Figure 7 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the training method of the interactive protein generation model, the method comprising: obtaining a sample protein pair, the sample protein pair comprising a sample protein and a sample interactive protein of the sample protein; extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; generating a sample predicted protein of the sample protein based on the initial generation model and the residue features; constructing a generation loss based on the sample predicted protein, and performing parameter iteration on the initial generation model based on the generation loss to obtain the interactive protein generation model.

[0069] A method for generating an interacting protein can also be performed, which includes: obtaining a reference protein sequence; extracting and fusing reference sequence features of each amino acid residue in the reference protein sequence; generating an interacting protein corresponding to the reference protein sequence based on the reference sequence features and an interacting protein generation model; the interacting protein generation model is obtained based on the training method for the interacting protein generation model as described above.

[0070] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0071] On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the training method of the interaction protein generation model provided by the above-mentioned methods, and the method includes: obtaining a sample protein pair, wherein the sample protein pair includes a sample protein and a sample interaction protein of the sample protein; extracting and fusing the sequence characteristics and spatial structure characteristics of each amino acid residue in the sample protein pair to obtain the residue characteristics of the sample protein pair; generating a sample predicted protein of the sample protein based on an initial generation model and the residue characteristics; constructing a generation loss based on the sample predicted protein, and performing parameter iteration on the initial generation model based on the generation loss to obtain an interaction protein generation model.

[0072] A method for generating an interacting protein can also be performed, which includes: obtaining a reference protein sequence; extracting and fusing reference sequence features of each amino acid residue in the reference protein sequence; generating an interacting protein corresponding to the reference protein sequence based on the reference sequence features and an interacting protein generation model; the interacting protein generation model is obtained based on the training method for the interacting protein generation model as described above.

[0073] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the training method for the interaction protein generation model provided by the above-mentioned methods, the method comprising: obtaining a sample protein pair, the sample protein pair comprising a sample protein, and a sample interaction protein of the sample protein; extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; generating a sample predicted protein of the sample protein based on an initial generation model and the residue features; constructing a generation loss based on the sample predicted protein, and performing parameter iteration on the initial generation model based on the generation loss to obtain an interaction protein generation model.

[0074] A method for generating an interacting protein can also be performed, which includes: obtaining a reference protein sequence; extracting and fusing reference sequence features of each amino acid residue in the reference protein sequence; generating an interacting protein corresponding to the reference protein sequence based on the reference sequence features and an interacting protein generation model; the interacting protein generation model is obtained based on the training method for the interacting protein generation model as described above.

[0075] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0076] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for generating a model of interacting proteins, characterized in that: include: obtaining a sample protein pair, wherein the sample protein pair comprises a sample protein and a sample interacting protein of the sample protein; Extracting and fusing the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; generating a sample predicted protein of the sample protein based on the initial generation model and the residue features; Predicting protein construction generation loss based on the sample, performing parameter iteration on the initial generation model based on the generation loss, and obtaining an interaction protein generation model.

2. The method for training an interacting protein generation model according to claim 1, characterized in that: The step of generating a sample predicted protein of the sample protein based on the initial generation model and the residue features comprises: Based on the residue features of the sample protein and the above residue features in the sample interacting protein, instructing the initial generation model to gradually generate sample amino acid residues at each site; Based on the sample amino acid residues at each site, obtaining the sample predicted protein; The above residue features include residue features of the amino acid residues preceding the current generated site in the sample interacting proteins.

3. The training method for generating a model of interacting proteins according to claim 2, characterized in that: The initial generation model is constructed based on a large language model; and the large language model is constructed based on a decoder.

4. The method for training a model for generating interacting proteins according to any one of claims 1 to 3, characterized in that: The sample protein pairs include universal sample protein pairs and fine-tuning sample protein pairs; The step of predicting protein construction generation loss based on the sample, and iterating parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model comprises: Constructing a universal generation loss based on sample predicted proteins of the universal sample protein pair, and iterating parameters of the initial generation model based on the universal generation loss to obtain a universal protein generation model; A fine-tuning generation loss is constructed based on the sample predicted protein of the fine-tuning sample protein pair, and based on the fine-tuning generation loss, parameters of the general protein generation model are fine-tuned to obtain the interacting protein generation model.

5. A method for producing an interacting protein, characterized in that: include: Obtaining a reference protein sequence; extracting and fusing reference sequence features of each amino acid residue in the reference protein sequence; Based on the reference sequence features and the interacting protein generation model, generating the interacting protein corresponding to the reference protein sequence; The interacting protein generation model is obtained based on the training method for generating an interacting protein model according to any one of claims 1 to 4.

6. The method for producing an interacting protein according to claim 5, characterized in that: The step of generating the interacting protein corresponding to the reference protein sequence based on the reference sequence features and the interacting protein generation model comprises: Based on the reference sequence features and the sequence features of the generated sites, instructing the interacting protein generation model to gradually generate the amino acid residues of the next site; Based on the generated amino acid residues at each site, the interacting proteins corresponding to the reference protein sequence are obtained.

7. A training device for generating a model of interacting proteins, characterized in that: include: An acquisition unit is configured to acquire a sample protein pair, wherein the sample protein pair comprises a sample protein and a sample interacting protein of the sample protein; A feature extraction unit extracts and fuses the sequence features and spatial structure features of each amino acid residue in the sample protein pair to obtain the residue features of the sample protein pair; A prediction unit, which generates a sample predicted protein of the sample protein based on the initial generation model and the residue features; The training unit predicts protein construction generation loss based on the sample, and iterates parameters of the initial generation model based on the generation loss to obtain an interaction protein generation model.

8. An interactive protein production device, characterized in that: include: Obtain sequence units, obtain reference protein sequences; A sequence feature extraction unit, which extracts and fuses reference sequence features of each amino acid residue in the reference protein sequence; A generating unit, which generates the interacting protein corresponding to the reference protein sequence based on the reference sequence characteristics and the interacting protein generation model; The interacting protein generation model is obtained based on the training method for generating an interacting protein model according to any one of claims 1 to 4.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for training an interacting protein generation model according to any one of claims 1 to 4 is implemented, or the method for generating interacting proteins according to any one of claims 5 to 6 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for training an interacting protein generation model according to any one of claims 1 to 4 or the method for generating interacting proteins according to any one of claims 5 to 6 is implemented.

Images