Method, device and electronic device for processing protein mutation effects

Abstract

The present invention relates to the interdisciplinary field of artificial intelligence and biotechnology, and provides a method, apparatus, and electronic device for processing protein mutation effects. The method comprises: obtaining the amino acid sequence of the protein before and after mutation; extracting features from the amino acid sequences of the protein before and after mutation using a pre-trained protein language model to obtain features of the protein before and after mutation; determining protein mutation characteristics based on the features of the protein before and after mutation; and inputting the protein mutation characteristics into a pre-trained first natural language model to generate a text description corresponding to the protein mutation. This method improves the interpretability of protein mutation effects by combining the protein language model with natural language processing technology.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of artificial intelligence and biotechnology, and in particular to a method, device and electronic equipment for processing protein mutation effects. Background Art

[0002] In modern life sciences, proteins, as fundamental molecules that carry out biological functions, are crucial for understanding biological processes, disease mechanisms, and the development of new drugs. Amino acid mutations within protein sequences are an important tool for studying protein functional diversity and evolution. With the development of bioinformatics and computational biology, the use of computational models to predict and design protein mutations has become an active research area.

[0003] Studying amino acid mutations in protein sequences is of paramount importance in life sciences. Protein language models have demonstrated powerful capabilities in a wide range of biological applications. However, due to their architectural design and lack of supervision, they are insufficient in explaining protein mutations. Summary of the Invention

[0004] The present invention provides a method, device and electronic device for processing protein mutation effects, which are used to address the shortcomings of the existing technology in insufficient explanation of protein mutations. By combining protein language models and natural language processing technology, the interpretability of protein mutation effects is improved.

[0005] The present invention provides a method for processing protein mutation effects, comprising: obtaining an amino acid sequence of a protein before mutation and an amino acid sequence of the protein after mutation; performing feature extraction on the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation using a pre-trained protein language model to obtain features of the protein before mutation and features of the protein after mutation; determining protein mutation features based on the features of the protein before mutation and the features of the protein after mutation; and inputting the protein mutation features into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0006] According to the method for processing protein mutation effects provided by the present invention, the protein mutation characteristics are determined based on the characteristics of the protein before mutation and the characteristics of the protein after mutation, including: determining the characteristic difference between the characteristics of the protein before mutation and the characteristics of the protein after mutation; encoding the characteristic difference based on the encoder branch of the protein mutation network of the pre-trained encoder-decoder architecture to obtain the protein mutation characteristics.

[0007] According to the method for processing protein mutation effects provided by the present invention, the encoder of the protein mutation network uses a cross-attention mechanism to map the feature difference into several fixed-length mutation features.

[0008] According to the method for processing protein mutation effects provided by the present invention, the decoder of the protein mutation network adopts a cross-attention mechanism symmetrical to the encoder to map the mutation features to the difference of the features before encoding.

[0009] According to the method for processing protein mutation effects provided by the present invention, the decoder branch of the protein mutation network also includes a mutation position prediction head and a mutation amino acid prediction head, the mutation position prediction head is used to predict the position of the mutation, and the mutation amino acid prediction head is used to predict the mutated amino acid. The inputs of the mutation position prediction head and the mutation amino acid prediction head are the characteristics of the post-mutation protein calculated by dividing the difference between the pre-encoding characteristics obtained by the decoder mapping and the characteristics of the pre-mutation protein.

[0010] According to the method for processing protein mutation effects provided by the present invention, the protein mutation network is jointly trained with the first natural language model and a second natural language model for mapping text descriptions into mutation features. Mutation features are obtained by inputting virtual words into the second natural language model. The joint training includes a pre-training stage. In the pre-training stage, sample protein amino acid sequences and sample documents are accepted as input, and training is performed through a conditionally controlled next word prediction task and a conditionally controlled mask prediction task, respectively, to achieve preliminary alignment of text modality and protein sequence modality.

[0011] According to the method for processing protein mutation effects provided by the present invention, the joint training also includes a fine-tuning stage, which adopts a two-round dialogue thinking chain form. The first round of dialogue requires the model to predict its function based on the protein before mutation. The second round of dialogue requires the model to integrate the results of the previous round of dialogue and predict the mutation effect based on the mutation characteristics, or predict the location and content of the mutation based on the mutation effect description.

[0012] According to the method for processing protein mutation effects provided by the present invention, the sample protein amino acid sequence and the sample literature are derived from a pre-established sample data set, and the steps for processing the data in the sample data set include: obtaining protein mutation records from a pre-established protein knowledge database, the mutation records including descriptions of mutation effects written by experts; searching based on reference information associated with the mutation records to obtain abstracts of publications associated with the mutation records; processing the abstracts of the retrieved publications based on a large language model to obtain enriched descriptions of the mutation effects; balancing the sample data set by constructing reverse samples for benign and malignant mutations; and removing low-quality descriptions from the descriptions of the sample data set using a large language model.

[0013] According to the method for processing protein mutation effects provided by the present invention, the method also includes: obtaining text instructions corresponding to the mutation; inputting the text instructions into a pre-trained second natural language model to generate mutation characteristics of the protein; using the pre-trained protein language model, extracting features of the amino acid sequence of the protein before mutation to obtain features of the protein before mutation; obtaining features of the protein after mutation based on the mutation features of the protein generated by the second natural language model and the features of the protein before mutation; and restoring the amino acid sequence of the protein after mutation based on the features of the protein after mutation.

[0014] The present invention also provides a device for processing protein mutation effects, comprising: an acquisition module, configured to acquire the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation; an extraction module, configured to perform feature extraction on the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation through a pre-trained protein language model to obtain features of the protein before mutation and features of the protein after mutation; a determination module, configured to determine protein mutation features based on the features of the protein before mutation and the features of the protein after mutation; and a generation module, configured to input the protein mutation features into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0015] 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 computer program, the method for processing the protein mutation effect as described above is implemented.

[0016] 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 any of the above-described methods for processing protein mutation effects.

[0017] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned methods for processing protein mutation effects.

[0018] The protein mutation effect processing method, device and electronic device provided by the present invention improve the interpretability of protein mutation effects by combining protein language models and natural language processing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to 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 any creative work.

[0020] Figure 1 It is a schematic flow chart of the method for processing protein mutation effects provided by the present invention.

[0021] Figure 2 It is a schematic diagram of the data processing flow of the protein mutation network provided by the present invention.

[0022] Figure 3 Schematic diagram of the joint training process of the protein mutation network provided by the present invention.

[0023] Figure 4 It is a schematic structural diagram of the device for processing protein mutation effects provided by the present invention.

[0024] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0026] Unless otherwise defined, the technical or scientific terms used in the present invention shall have the usual meaning understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one", "an" or "the" do not indicate a quantity limitation, but rather indicate the presence of at least one. Words such as "include" or "comprise" mean that the elements or objects preceding the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0027] The following is a brief explanation of the terms involved in the present invention.

[0028] The following combination Figure 1-Figure 5 The present invention describes a method, device and electronic device for processing the effects of protein mutations.

[0029] Figure 1 Schematic diagram of the process of treating protein mutation effect provided by the present invention. Figure 1 As shown, the method includes the following:

[0030] Step 101: Obtain the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation.

[0031] In this embodiment, the protein before mutation may include a wild-type protein, which is a protein existing in nature.

[0032] Step 102: Using a pre-trained protein language model, extract features from the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation to obtain features of the protein before mutation and features of the protein after mutation.

[0033] In this example, protein language models (PLMs) are a type of computational model specifically designed for processing and analyzing protein sequences. By learning from large amounts of protein sequence data, these models can capture the complex relationships and patterns between amino acids, enabling them to be used for various tasks such as protein structure prediction, functional annotation, and mutation effect prediction. Representative PLMs include ESM, ProteinBERT, and ProtTrans. PLMs can be pre-trained on an evolutionary scale database or other sample databases.

[0034] Step 103: Determine the protein mutation characteristics based on the characteristics of the protein before mutation and the characteristics of the protein after mutation.

[0035] In this embodiment, the protein mutation characteristics can be determined based on the characteristics of the protein before mutation and the characteristics of the protein after mutation to perform differential analysis to obtain the protein mutation characteristics. The protein mutation characteristics can also be obtained by further feature extraction or transformation of the differential analysis results by establishing a model or other means.

[0036] Step 104: Input the protein mutation feature into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0037] In this embodiment, the first natural language model is used to characterize the correspondence between protein mutation signatures and textual descriptions corresponding to the protein mutations. The first natural language model can be trained using sample protein mutation signatures and textual descriptions corresponding to the sample protein mutations. Alternatively, the model can be initialized using an existing large language model (LLM), such as BioMedGPT-LM or BERT. BioMedGPT-LM is a scientific language model built through continuous pre-training on a large-scale biomedical corpus. BERT is a pre-trained language representation model that understands natural language through deep bidirectional representation.

[0038] Existing models for interpreting protein mutation effects are often designed for static protein sequences, and their interpretation can only predict evolutionary fitness values. However, the present invention uses protein mutation signatures to model the changes introduced by mutations. Combined with a pre-trained first natural language model, the present invention can generate detailed textual descriptions of protein mutations, revealing their potential impacts on protein function, stability, and interactions. This significantly enriches the information available for mutation analysis and enables human experts to more intuitively understand mutation effects.

[0039] In some optional implementations, protein mutation signatures are determined based on pre-mutation and post-mutation protein signatures, including: determining the difference between pre-mutation and post-mutation protein signatures; and encoding the difference to obtain the protein mutation signature using an encoder branch of a protein mutation network based on a pre-trained encoder-decoder architecture. The difference between pre-mutation and post-mutation protein signatures contains rich mutation information and is suitable for extracting mutation signatures. Protein mutation networks using encoder-decoder architectures can explicitly model protein mutations, provide a deeper understanding of the effects of mutations on protein structure and function, and help provide a more accurate interpretation of mutation effects.

[0040] In some optional implementations, the encoder of the protein mutation network uses a cross-attention mechanism to map feature differences into several fixed-length mutation features. Figure 2 , Figure 2It is a schematic diagram of the data processing flow of the protein mutation network provided by the present invention. The encoder and decoder of the protein mutation network promote the bidirectional conversion between feature differences and protein mutation features. In the encoder branch of the protein mutation network, given the feature difference, the encoder is expected to extract information that retains the mutation features in a unified feature space. However, the length of protein sequences varies greatly, ranging from dozens to thousands of amino acids. Using a cross-attention module, the sequence representation can be converted into a fixed number of potential features. This module part uses K trainable features as queries and generates outputs with sequence representations as keys and values. In addition, two parallel modules can be used for encoding wild-type proteins and mutations, further enriching the interpretation of protein mutation effects.

[0041] In some optional implementations, the decoder of the protein mutation network uses a cross-attention mechanism symmetric to the encoder, mapping mutation features to differences in pre-encoded features. For example, the feature representation of the wild-type protein can be used as the query, with the mutation features as the key and value. The output can then be processed through a two-layer feedforward network (FFN) or other neural network architecture to reconstruct the differences in protein features before and after the mutation.

[0042] In some optional implementations, the decoder branch of the protein mutation network may further include a mutation position prediction head and a mutation amino acid prediction head. The mutation position prediction head is used to predict the position of the mutation, and the mutation amino acid prediction head is used to predict the mutated amino acid. The inputs of the mutation position prediction head and the mutation amino acid prediction head are the features of the mutated protein calculated by the difference between the pre-encoding features obtained by the decoder mapping and the features of the pre-mutation protein. The mutation position prediction head can be a fully connected layer or other neural network structure that can predict whether an amino acid should be replaced. The mutation amino acid prediction head can be initialized from a protein language model to predict the probability distribution of the mutated amino acid. To facilitate the subsequent use of text instructions for protein engineering, K trainable virtual words are introduced. These virtual words are additionally input into the word embedding so that the large language model can summarize the text semantics. The output representation of the virtual words is processed by the decoder to generate mutations.

[0043] In addition, the prediction of mutation location and type can be integrated into the final output layer of the model, or by designing features that can indicate the possibility of mutation, these features can be directly input into the model for training instead of through a separate prediction head. It is also possible to use known biological rules and patterns, such as protein domain knowledge and known mutation hotspots, to predict possible mutation locations, or to compare the target protein sequence with proteins with known mutations based on sequence comparison to predict the location and potential impact of the mutation.

[0044] In some optional implementations, the protein mutation network is jointly trained with a first natural language model and a second natural language model used to map textual descriptions to mutation signatures. Mutation signatures are obtained by inputting soft tokens into the second natural language model. Joint training includes a pretraining phase, in which the network receives sample protein amino acid sequences and sample literature as inputs and is trained using the conditional next token prediction task and the conditional masked language modeling task, respectively, to achieve preliminary alignment between the textual and protein sequence modalities. The soft tokens are trainable parameters embedded in the model to help capture and represent mutation signatures. The first and second natural language models constitute the forward and reverse language models. The joint training of the protein mutation network with the two natural language models aims to integrate protein knowledge and textual descriptions from scientific literature to improve the understanding and prediction of protein mutations.

[0045] refer to Figure 3 , Figure 3 This is a schematic diagram of the joint training process for the protein mutation network provided by the present invention. The goal is to integrate protein knowledge from sample literature using a bidirectional language modeling objective. For the encoding workflow, the output representation of the wild-type encoder can be used as the input to the first natural language model, and a next-word prediction objective is calculated to generate descriptive text. For the decoding workflow, a conditional masked language modeling objective can be used on the protein sequence. As an example, 15% of the amino acids can be masked, and the PLM is required to recover the masked amino acids based on the remaining amino acid sequence and mutation signature. Overall, the sum of these two language modeling objectives can be optimized.

[0046] In some optional implementations, joint training also includes a fine-tuning phase, which uses a two-round conversational chain of thoughts. The first round requires the model to predict the function of the protein before mutation. The second round requires the model to integrate the results of the previous round and predict the effect of the mutation based on the mutation signature, or to predict the location and content of the mutation based on the description of the mutation effect. This two-round chain of thoughts (CoT) is an effective strategy for improving model performance on specific tasks. The fine-tuning phase can further refine the model's capabilities, particularly in understanding and predicting the effects of protein mutations.

[0047] like Figure 3As shown, protein evolution knowledge can be used to guide the model training process and fine-tuning can be used to promote mutation interpretation and design. Since the effect of mutations usually involves enhancing or weakening protein function, a chain of thoughts (CoT) strategy can be adopted to seamlessly connect protein function and mutation effect in a two-round dialogue. In the first round, the natural language model can be prompted to describe the function of the wild-type protein using an encoding workflow. In the second round, two tasks can be introduced, namely describing the effect of the mutation and predicting the mutation based on text instructions. Both tasks use the predicted function and potential wild-type representation in the first round of dialogue as additional input. Specifically, the overall goal of fine-tuning is the sum of the following three parts: (1) next token prediction based on protein function, (2) next token prediction based on mutation effect, and (3) cross-entropy loss between the predicted position and amino acid and the true mutation.

[0048] In some optional implementations, the sample protein amino acid sequences and sample literature are derived from a pre-established sample dataset. The data processing steps in the sample dataset include: obtaining protein mutation records from a pre-established protein knowledge database, including expert-written descriptions of the mutation effects; searching for abstracts of publications associated with the mutation records based on reference information associated with the mutation records; processing the retrieved abstracts of the publications using a large-scale language model to obtain enriched descriptions of the mutation effects; balancing the sample dataset by constructing reverse samples for benign and malignant mutations; and removing low-quality descriptions from the sample dataset using the large-scale language model. The protein knowledge database can be established based on an existing database, such as the Universal Protein Knowledge Database (UniProtKB), a widely used protein database containing 107,000 single-point mutations. Expert-written descriptions of the mutation effects can be collected from the database, and abstracts of corresponding publications can be retrieved based on available reference information. Publication abstracts can be retrieved from literature search platforms such as PubMed and Web of Science. Given that descriptions in UniProtKB are often short and homogeneous, large language models such as GPT-3.5-turbo can be used to enrich text annotations by retrieving relevant descriptions from the original PubMed abstracts. GPT-3.5-turbo is a state-of-the-art natural language processing (NLP) model developed by the artificial intelligence research lab OpenAI. It is a successor to the GPT-3 (third-generation generative pre-trained Transformer) model family, designed to provide more efficient and faster text generation capabilities.

[0049] Furthermore, the dataset can be balanced by constructing reverse samples for benign and malignant mutations. Specifically, for each protein mutation, one can attempt to create a reversed data sample by swapping the wild-type and mutant versions and prompting GPT-3.5-turbo to write a description that is the opposite of the original mutation's effect. GPT-3.5-turbo can also be prompted to filter out low-quality descriptions, such as those that only mention the original species. This step helps ensure that the dataset contains high-quality and informative annotations. Finally, the dataset can be split into training, validation, and test sets.

[0050] In some optional implementations, the method further includes: obtaining a text instruction corresponding to the mutation; inputting the text instruction into a pre-trained second natural language model to generate a mutation feature of the protein; performing feature extraction on the amino acid sequence of the protein before mutation using the pre-trained protein language model to obtain features of the protein before mutation; obtaining features of the protein after mutation based on the mutation features of the protein generated by the second natural language model and the features of the protein before mutation; and restoring the amino acid sequence of the protein after mutation based on the features of the protein after mutation.

[0051] Obtaining a feature of the post-mutation protein based on the protein mutation feature and the feature of the pre-mutation protein may include: decoding the protein mutation feature using a decoder branch of a protein mutation network based on a pre-trained encoder-decoder architecture to obtain a feature difference between the pre-mutation protein feature and the post-mutation protein feature; and obtaining a feature of the post-mutation protein based on the feature difference and the pre-mutation protein feature.

[0052] Existing techniques for mutation design using protein language models can only suggest evolutionarily plausible mutations, which may not align with real-world human preferences for directed evolution. For example, enhancing the catalytic activity of a bacterial enzyme might be detrimental to its survival due to the increased energetic cost of its survival, but beneficial for industrial applications. In such cases, the utility of protein language models in assisting protein engineering is significantly diminished.

[0053] In this implementation, natural language is used to explicitly model protein mutations, which effectively improves the customizability of mutation design and enhances the user experience for researchers.

[0054] The device for processing protein mutation effects provided by the present invention is described below. The device for processing protein mutation effects described below and the method for processing protein mutation effects described above can be referred to in correspondence with each other.

[0055] Figure 4 A schematic diagram of the structure of a device for processing protein mutation effects provided in an embodiment of the present application is shown in FIG. Figure 4As shown, it specifically includes: an acquisition module 401, configured to obtain the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation; an extraction module 402, configured to extract features of the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation through a pre-trained protein language model to obtain features of the protein before mutation and features of the protein after mutation; a determination module 403, configured to determine protein mutation features based on the features of the protein before mutation and the features of the protein after mutation; a generation module 404, configured to input the protein mutation features into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0056] In some optional implementations, the determination module 404 is further configured to: determine the feature difference between the feature of the protein before mutation and the feature of the protein after mutation; and encode the feature difference based on the encoder branch of the protein mutation network of the pre-trained encoder-decoder architecture to obtain the protein mutation feature.

[0057] In some optional implementations, the encoder of the protein mutation network uses a cross-attention mechanism to map feature differences into several fixed-length mutation features.

[0058] In some optional implementations, the decoder of the protein mutation network obtains mutation features by adding virtual words, and uses a cross-attention mechanism symmetrical to the encoder to map the mutation features to the difference of the features before encoding.

[0059] In some optional implementations, the decoder branch of the protein mutation network also includes a mutation position prediction head and a mutation amino acid prediction head. The mutation position prediction head is used to predict the position of the mutation, and the mutation amino acid prediction head is used to predict the mutated amino acid. The input of the mutation position prediction head and the mutation amino acid prediction head is the feature of the mutated protein calculated by the difference between the pre-encoding feature obtained according to the decoder mapping and the feature of the protein before mutation.

[0060] In some optional implementations, the protein mutation network is jointly trained with a first natural language model and a second natural language model for mapping text descriptions to mutation features. The joint training includes a pre-training stage. In the pre-training stage, sample protein amino acid sequences and sample documents are accepted as input, and training is performed through a conditionally controlled next word prediction task and a conditionally controlled mask prediction task, respectively, to achieve preliminary alignment of the text modality and the protein sequence modality.

[0061] In some optional implementations, joint training also includes a fine-tuning phase, which adopts the form of a two-round dialogue chain. The first round of dialogue requires the model to predict the function of the protein based on its pre-mutation protein. The second round of dialogue requires the model to integrate the results of the previous round of dialogue and predict the mutation effect based on the mutation characteristics, or predict the location and content of the mutation based on the mutation effect description.

[0062] In some optional implementations, the sample protein amino acid sequence and the sample literature are derived from a pre-established sample data set, and the steps for processing the data in the sample data set include: obtaining protein mutation records from a pre-established protein knowledge database, the mutation records including descriptions of the mutation effects written by experts; searching based on reference information associated with the mutation records to obtain abstracts of publications associated with the mutation records; processing the abstracts of the retrieved publications based on a large language model to obtain enriched descriptions of the mutation effects; balancing the sample data set by constructing reverse samples for benign and malignant mutations; and removing low-quality descriptions from the descriptions of the sample data set through a large language model.

[0063] In some optional implementations, the device also includes a mutation design module, which is configured to: obtain text instructions corresponding to the mutation; input the text instructions into a pre-trained second natural language model to generate mutation characteristics of the protein; extract features of the amino acid sequence of the protein before mutation through the pre-trained protein language model to obtain features of the protein before mutation; obtain features of the mutated protein based on the mutation features of the protein generated by the second natural language model and the features of the protein before mutation; and restore the amino acid sequence of the mutated protein based on the features of the mutated protein.

[0064] The protein mutation effect processing device provided by the present invention improves the interpretability of protein mutation effects by combining protein language models and natural language processing technology.

[0065] Figure 5 An example of a physical structure diagram of an electronic device is shown below. Figure 5As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communications bus 540. The processor 510 may call logic instructions in the memory 530 to execute a method for processing protein mutation effects, which includes: obtaining an amino acid sequence of the protein before mutation and an amino acid sequence of the protein after mutation; extracting features of the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation using a pre-trained protein language model to obtain features of the protein before mutation and features of the protein after mutation; determining protein mutation features based on the features of the protein before mutation and the features of the protein after mutation; and inputting the protein mutation features into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0066] Furthermore, the logic instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0067] 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 protein mutation effect processing method provided by the above-mentioned methods, which includes: obtaining the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation; using a pre-trained protein language model, extracting features of the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation to obtain features of the protein before mutation and features of the protein after mutation; determining the protein mutation features based on the features of the protein before mutation and the features of the protein after mutation; inputting the protein mutation features into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0068] 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, implements a method for processing protein mutation effects provided by the above-mentioned methods, the method comprising: obtaining the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation; performing feature extraction on the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation through a pre-trained protein language model to obtain features of the protein before mutation and features of the protein after mutation; determining protein mutation features based on the features of the protein before mutation and the features of the protein after mutation; and inputting the protein mutation features into a pre-trained first natural language model to generate a text description corresponding to the protein mutation.

[0069] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0070] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0071] 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 various embodiments of the present invention.

Claims

1. A method for treating protein mutation effects, characterized in that: include: Obtaining the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation; Using a pre-trained protein language model, feature extraction is performed on the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation to obtain features of the protein before mutation and features of the protein after mutation; Determining protein mutation characteristics based on the characteristics of the protein before mutation and the characteristics of the protein after mutation includes: determining a characteristic difference between a characteristic of the protein before mutation and a characteristic of the protein after mutation; Encoding the feature difference based on an encoder branch of a protein mutation network of a pre-trained encoder-decoder architecture to obtain the protein mutation feature; Inputting the protein mutation feature into a pre-trained first natural language model to generate a text description corresponding to the protein mutation; The protein mutation network is jointly trained with the first natural language model and a second natural language model for mapping text descriptions to mutation features. Mutation features are obtained by inputting virtual words into the second natural language model. The joint training includes a pre-training phase. In the pre-training phase, sample protein amino acid sequences and sample documents are accepted as input and trained through a conditionally controlled next word prediction task and a conditionally controlled mask prediction task, respectively, to achieve preliminary alignment of text modality and protein sequence modality. The joint training also includes a fine-tuning phase, which adopts a two-round dialogue thinking chain. The first round of dialogue requires the model to predict the function of the protein before mutation. The second round of dialogue requires the model to integrate the results of the previous round of dialogue and predict the mutation effect based on the mutation characteristics, or predict the location and content of the mutation based on the mutation effect description. The sample protein amino acid sequence and the sample literature are derived from a pre-established sample data set, and the steps for processing the data in the sample data set include: Obtaining a mutation record of a protein from a pre-established protein knowledge database, wherein the mutation record includes an expert-written description of the effect of the mutation; Searching based on reference information associated with the mutation record to obtain abstracts of publications associated with the mutation record; Processing the abstracts of retrieved publications based on a large language model to obtain enriched descriptions of mutation effects; The sample dataset was balanced by constructing reverse samples for benign and malignant mutations; Use a large language model to remove low-quality descriptions from the descriptions of the sample dataset.

2. The method for treating protein mutation effects according to claim 1, characterized in that: The encoder of the protein mutation network uses a cross-attention mechanism to map the feature difference into several fixed-length mutation features.

3. The method for treating protein mutation effects according to claim 2, characterized in that: The decoder of the protein mutation network uses a cross-attention mechanism symmetrical to the encoder to map the mutation features to the difference of the features before encoding.

4. The method for treating protein mutation effects according to claim 3, characterized in that: The decoder branch of the protein mutation network also includes a mutation position prediction head and a mutation amino acid prediction head. The mutation position prediction head is used to predict the position of the mutation, and the mutation amino acid prediction head is used to predict the mutated amino acid. The inputs of the mutation position prediction head and the mutation amino acid prediction head are the characteristics of the mutated protein calculated by the difference between the pre-encoding characteristics obtained by the decoder mapping and the characteristics of the pre-mutation protein.

5. The method for treating protein mutation effects according to claim 1, characterized in that: The method further comprises: Get the text instructions corresponding to the mutation; Inputting the text instruction into a pre-trained second natural language model to generate a mutation signature of the protein; Using a pre-trained protein language model, the features of the protein before mutation are extracted from the amino acid sequence of the protein before mutation. Obtaining features of the protein after mutation based on the mutation features of the protein generated by the second natural language model and features of the protein before mutation; The amino acid sequence of the mutated protein is restored based on the characteristics of the mutated protein.

6. A device for processing protein mutation effects, characterized in that: include: an acquisition module, configured to acquire the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation; an extraction module configured to perform feature extraction on the amino acid sequence of the protein before mutation and the amino acid sequence of the protein after mutation using a pre-trained protein language model to obtain features of the protein before mutation and features of the protein after mutation; A determination module is configured to determine a protein mutation feature based on the feature of the protein before mutation and the feature of the protein after mutation, comprising: determining a characteristic difference between a characteristic of the protein before mutation and a characteristic of the protein after mutation; Encoding the feature difference based on an encoder branch of a protein mutation network of a pre-trained encoder-decoder architecture to obtain the protein mutation feature; a generation module configured to input the protein mutation feature into a pre-trained first natural language model to generate a text description corresponding to the protein mutation; The protein mutation network is jointly trained with the first natural language model and a second natural language model for mapping text descriptions to mutation features. Mutation features are obtained by inputting virtual words into the second natural language model. The joint training includes a pre-training phase. In the pre-training phase, sample protein amino acid sequences and sample documents are accepted as input and trained through a conditionally controlled next word prediction task and a conditionally controlled mask prediction task, respectively, to achieve preliminary alignment of text modality and protein sequence modality. The joint training also includes a fine-tuning phase, which adopts a two-round dialogue thinking chain. The first round of dialogue requires the model to predict the function of the protein before mutation. The second round of dialogue requires the model to integrate the results of the previous round of dialogue and predict the mutation effect based on the mutation characteristics, or predict the location and content of the mutation based on the mutation effect description. The sample protein amino acid sequence and the sample literature are derived from a pre-established sample data set, and the steps for processing the data in the sample data set include: Obtaining a mutation record of a protein from a pre-established protein knowledge database, wherein the mutation record includes an expert-written description of the effect of the mutation; Searching based on reference information associated with the mutation record to obtain abstracts of publications associated with the mutation record; Processing the abstracts of retrieved publications based on a large language model to obtain enriched descriptions of mutation effects; The sample dataset was balanced by constructing reverse samples for benign and malignant mutations; Use a large language model to remove low-quality descriptions from the descriptions of the sample dataset.

Images