Multi-modal instruction guided protein design method and device

Through the multimodal instruction-guided protein design method, the deep learning model combined with cross-modal transformation and transformer architecture is used to solve the problem of limited space in traditional protein design, realize more efficient protein sequence design and visual interpretation, and explore a new structure .

CN118969059APending Publication Date: 2024-11-15ZHEJIANG UNIV

Patent Information

Application Number
CN202411091775.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Traditional protein design methods have limited design space, making it difficult to explore new protein structures, and deep learning models have shortcomings in capturing long context information and parallelism.

Method used

Using multimodal instructions to guide protein design method, by extracting paired text instructions and protein sequences from the SwissProtCLAP dataset, cross-modal conversion is carried out using SciBERT and Gaussian distribution models to build a protein design model that integrates multimodal instructions, combined with The transformer decoder performs protein sequence design and visualizes it.

Benefits of technology

It improves the accuracy and efficiency of protein design, explores a broader new protein structure, provides visual explanations, and makes up for the limitations of traditional designs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969059A_ABST
    Figure CN118969059A_ABST
Patent Text Reader

Abstract

The invention discloses a protein design method and device based on multi-modal instruction guidance, and the method comprises the steps: randomly extracting paired text instructions and protein sequences from a SwisProtCLAP data set as single samples, and constructing a sample set; the method comprises the following steps of: obtaining a generation condition in a character form by utilizing a character instruction to serve as a modal instruction, and performing cross-modal conversion on the generation condition in the character form into a generation condition in a protein form to serve as another modal instruction, so as to obtain a multi-modal instruction; constructing a protein design model which comprises a multi-modal fusion device fusing the multi-modal instructions and a decoder processing the multi-modal instructions and performing protein design, and performing protein sequence-based supervised training on the protein design model by using the sample set; and carrying out protein sequence design by utilizing the trained protein design model, and visualizing the designed protein sequence, so as to solve the technical problem of limited protein design space in practical application.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of synthetic biotechnology, and specifically relates to a protein design method and device guided by multimodal instructions. Background Art

[0002] The task of using deep learning technology to assist protein design has attracted more and more attention from practitioners. Protein design is an important emerging branch and frontier discipline of synthetic biology. Deep learning technology has played an important role in assisting protein design, providing greater support for synthetic biology research. Proteins are macromolecules in organisms with complex primary, secondary and tertiary structures, and play a variety of biological functions. In biology, protein sequences are often used as descriptions of proteins, and protein sequences are linear chains composed of 20 natural amino acids, with lengths ranging from tens to thousands of amino acids. Protein design is based on biological principles and computer simulation technology, aiming to design protein molecules with new structures and new activities, and promote the development of biotechnology and medicine. However, the design space of traditional protein design is limited, and it can often only be designed based on known protein sequences. The generated results have a high degree of overlap with existing proteins, making it difficult to explore a wider range of new protein structures. In order to make up for the limitations of traditional protein design, using the current hot deep learning technology to assist protein design is a work of great practical significance.

[0003] Deep learning-assisted protein design is a generation task that takes the input instruction data as the generation condition and generates the corresponding target data under its guidance. There are various types of generation tasks, including variational generation, autoregressive generation, adversarial generation, streaming generation, etc. Among them, autoregressive generation, as an important generation method, has many advantages such as good generation effect, strong flexibility, and strong interpretability. It is widely used in the generation of text, image, protein, time series and other data. The recurrent neural network RNN ​​has been continuously improved. It introduces time dependencies in the network and can effectively capture the time information in the sequence data. LSTM introduces a gating mechanism to solve the gradient vanishing and gradient exploding problems of RNN, making it more flexible and efficient when generating sequence data. However, the above model still has the problem of memory decay, cannot capture very long context information, and has poor parallelism.

[0004] The introduction of transformers laid the foundation for the arrival of the era of large models. Many classic models have subsequently adopted the transformer architecture and greatly increased the scale of model parameters and training data, becoming truly large models, such as GPT and Dall-E. In general, researchers continue to promote the development of generative models from multiple perspectives such as model structure, model scale, and data type. Based on transformers, models are no longer confined to a single text modality, but pay more attention to the understanding and generation of multimodal data. The protein design method based on multimodal instruction guidance focuses on biological protein data and text data, selects a suitable deep network model structure, and generates corresponding protein sequences according to text instructions, making up for the limitations of traditional protein design and promoting the long-term development of synthetic biology. Summary of the invention

[0005] In view of the above, an object of the present invention is to provide a protein design method and apparatus guided by multimodal instructions to solve the technical problem of limited protein design space in practical applications.

[0006] To achieve the above-mentioned object of the invention, an embodiment of the present invention provides a protein design method guided by multimodal instructions, comprising:

[0007] Randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design;

[0008] First, a textual instruction is used to obtain a textual generation condition as a modal instruction, and then the textual generation condition is cross-modally converted into a protein-form generation condition as another modal instruction, thereby obtaining a multi-modal instruction;

[0009] Constructing a protein design model, which includes a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and using a sample set to perform supervised training on the protein design model based on protein sequences;

[0010] Use the trained protein design model to design protein sequences and visualize the designed protein sequences.

[0011] Preferably, in a single sample, the text instruction describes at least one property of a protein by natural language, and the protein sequence correspondingly possesses at least one described property and is composed of tens to thousands of amino acids.

[0012] Preferably, a pre-trained SciBERT encoder is used to convert the text instruction into a generation condition in text form as a modal instruction.

[0013] Preferably, a pre-trained Gaussian distribution model is used for cross-modal conversion, and the generation conditions in text form are converted into generation conditions in protein form as another modal instruction.

[0014] Preferably, the process of fusing the multimodal instructions by the multimodal fuser includes:

[0015] After concatenating or averaging the multimodal instructions, the attention module and the fully connected layer are connected to reorganize to obtain fused instructions that are beneficial to protein design.

[0016] Preferably, the decoder includes at least two transformer decoders and a fully connected layer, and processes and performs protein design according to the fusion instructions output by the multimodal fuser to obtain the designed protein sequence.

[0017] Preferably, the designed protein sequence is visualized, comprising:

[0018] The designed protein sequence is folded using the AlphaFold 2 model to obtain a three-dimensional structure. The protein sequence is interpretably operated through the PyMOL visualization tool, and a visual three-dimensional structure image of the protein sequence is output.

[0019] Preferably, it also includes: verifying the accuracy and effectiveness of the protein design model in generating the protein sequence, specifically: first calculating the similarity between the input text instruction and the designed protein sequence; then randomly selecting 4 non-corresponding text instructions, and calculating their similarity with the designed protein sequence. If the input text instruction has the highest similarity with the designed protein sequence, it means that the generation process is effectively guided by the text instruction, and the generation result is considered accurate, otherwise it is inaccurate.

[0020] To achieve the above-mentioned purpose of the invention, the embodiment further provides a protein design device guided by multimodal instructions, comprising:

[0021] Sample construction, which is used to randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design;

[0022] An instruction processing module, which is used to first use the text instruction to obtain a text-form generation condition as a modal instruction, and then cross-modally convert the text-form generation condition into a protein-form generation condition as another modal instruction, thereby obtaining a multi-modal instruction;

[0023] A model building module, which is used to build a protein design model, comprising a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and using a sample set to perform supervised training on the protein design model based on protein sequences;

[0024] The protein design module is used to design protein sequences using the trained protein design model and to visualize the designed protein sequences.

[0025] To achieve the above-mentioned purpose of the invention, an embodiment further provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the protein design method guided by the above-mentioned multimodal instructions.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1) Apply deep learning to the design of protein sequences, improve the accuracy and effectiveness of protein production, and provide visual explanations for those working in the field of protein synthesis.

[0028] 2) Use text instructions to obtain instructions in two modes, text form and protein form, and perform multiple multimodal instruction fusions to extract more effective instruction features to promote the generation effect.

[0029] 3) To a certain extent, it can effectively improve the efficiency of protein synthesis, make up for the limited design space in the traditional protein design field, and explore a broader range of new protein structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 is a flow chart of a protein design method guided by multimodal instructions provided in an embodiment;

[0032] Figure 2 It is a flowchart of a protein design method guided by multimodal instructions provided in an embodiment;

[0033] Figure 3 It is the structure and flow chart of the SciBERT encoder provided in the embodiment;

[0034] Figure 4 is a schematic diagram of the structure of a Gaussian distribution model provided in an embodiment;

[0035] Figure 5 It is a structure and flow chart of a multimodal fusion device provided in an embodiment;

[0036] Figure 6 is the structure and flow chart of the decoder provided by the embodiment;

[0037] Figure 7 It is a schematic diagram of the structure of a protein design device guided by multimodal instructions provided in an embodiment. DETAILED DESCRIPTION

[0038] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0039] With regard to the task of assisting workers in the field of protein synthesis in protein design, the goal of the present invention is to generate corresponding protein sequences based on input text instructions, ultimately providing workers in the field of protein synthesis with a reference design result and providing three-dimensional structural visualization of the generated results.

[0040] like Figure 1 and Figure 2 As shown, the protein design method guided by multimodal instructions provided in the embodiment comprises the following steps:

[0041] S1. Randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design.

[0042] In the embodiment, a certain number of paired text instructions and protein sequences are randomly selected from the SwissProtCLAP data set as samples, and a sample set based on multimodal instructions to guide protein design is constructed. A single sample contains paired text instructions and protein sequences, wherein the text instructions describe one or several properties of the protein and consist of one or several natural languages; and the protein sequence correspondingly has the described properties and consists of dozens to thousands of amino acids. The effective use of text instructions is the key to a good model generation effect.

[0043] Overall, the data randomly extracted from the SwissProtCLAP dataset is well organized, so only the simplest data sorting and cleaning is needed without changing the data characteristics and distribution. Then, some preprocessing operations such as data sorting and cleaning are performed without affecting the overall data characteristics and distribution. Finally, the training set and validation set are divided into 7:3. Before the division, the data needs to be randomly disturbed to prevent the same data from being stacked.

[0044] S2. First, use the text instruction to obtain the text-based generation condition as a modal instruction, and then convert the text-based generation condition into a protein-based generation condition across modalities as another modal instruction, thereby obtaining a multimodal instruction.

[0045] When processing text instructions, we strive to maximize the multimodal information of text instructions and obtain instructions of multiple modes to guide protein design.

[0046] Specifically, the pre-trained SciBERT encoder is first used to convert the text instruction into a text-based generation condition as a modal instruction. Figure 3 As shown, the SciBERT encoder contains 12 transformer encoders. The text instructions are input into the SciBERT encoder and encoded multiple times to obtain the generation conditions in text form.

[0047] Then, the pre-trained Gaussian distribution model is used for cross-modal conversion, and the text-based generation conditions are converted into protein-based generation conditions as another modal instruction, thereby obtaining instructions of two modalities and maximizing the mining of multiple modal information of text instructions. Figure 4 As shown, in the Gaussian distribution model, the text-based generation conditions are input through two fully connected layers to calculate the mean and standard deviation, and then a random vector is introduced to perform element-wise product calculation with the standard deviation and then added to the mean to obtain the protein-based generation conditions.

[0048] S3. Construct a protein design model, which includes a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and use the sample set to perform supervised training on the protein design model based on protein sequences.

[0049] During the model building stage, different multimodal instruction fusion methods and complex multimodal instruction decoding methods will be used to build the model, striving to achieve the best generation effect.

[0050] In the embodiment, the constructed protein design model includes a multimodal fuser and a decoder, wherein the multimodal fuser is used to fuse multimodal instructions, such as Figure 5 The process shown includes: the instructions in text form and protein form have the same dimension, the two modal instructions are spliced ​​or the multimodal instructions are averaged, where the splicing dimension will be doubled, and the dimension remains unchanged after averaging, the attention module is connected, weights are assigned to different channel features, the modal instructions are reorganized based on the weights, and then passed through the fully connected layer to obtain a fusion instruction that is more conducive to protein design.

[0051] The decoder is used to process the fusion instructions output by the multimodal fusion device and perform protein design to obtain the designed protein sequence. Figure 6As shown in the figure, the decoder contains 6 transformer decoders and 1 fully connected layer, showing powerful protein design capabilities.

[0052] In the embodiment, a loss function of the training model is also constructed. The loss function can be constructed based on the difference between the designed protein sequence output by the decoder and the real protein sequence contained in the sample. The loss function is an autoregressive loss, which is a special cross entropy loss.

[0053] In the training phase, after constructing the entire model structure and loss function, back propagation can be performed to update the network parameters. When the model converges and is optimized to a limit, the model iteration can be stopped. The upper limit of the specific number of iterations is set to 80,000 times. Each time a batch of data is input, the loss is calculated and back propagated, and the model parameters are updated until the training is completed. During the iteration process, the model with the best verification effect will be saved. If the effect of the verification set is in a continuous decline for 10 consecutive times, the current training will be exited to obtain the final protein design model.

[0054] S4. Use the trained protein design model to design protein sequences and visualize the designed protein sequences.

[0055] In the embodiment, for the data to be predicted, the data is sorted and cleaned according to the method of obtaining the training data, and then input into the model to output the generated protein sequence. In the verification stage, the similarity between the input text instruction and the protein sequence is first calculated; then 4 non-corresponding text instructions are randomly selected to calculate their similarity with the protein sequence. If the similarity between the input text instruction and the protein sequence is the highest, it means that the generation process is effectively guided by the text instruction, and the generation result is considered accurate; otherwise, it is inaccurate. Finally, AlphaFold 2 is used to fold the three-dimensional structure of the generated result, and the design result is explained by the PyMOL visualization tool. The visual three-dimensional structure image of the protein sequence is output to verify whether the judgment at this time has a certain degree of correctness.

[0056] like Figure 7As shown, the embodiment also provides a protein design device 70 guided by multimodal instructions, including a sample construction module 71, an instruction processing module 72, a model construction module 73, and a protein design module 74, wherein the sample construction module 71 is used to randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP data set as single samples, and construct a sample set based on multimodal instruction-guided protein design; the instruction processing module 72 is used to first use the text instructions to obtain text-based generation conditions as one modal instruction, and then cross-modally convert the text-based generation conditions into protein-based generation conditions as another modal instruction, thereby obtaining multimodal instructions; the model construction module 73 is used to construct a protein design model, which includes a multimodal fusion device for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and uses the sample set to perform supervised training on the protein design model based on protein sequences; the protein design module 74 is used to use the trained protein design model to perform protein sequence design, and visualize the designed protein sequence.

[0057] It should be noted that, when the multimodal instruction-guided protein design device provided in the above embodiment performs multimodal instruction-guided protein design, the division of the above functional modules should be used as an example for explanation, and the above functional allocation can be completed by different functional modules as needed, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the functions described above. In addition, the multimodal instruction-guided protein design device provided in the above embodiment and the multimodal instruction-guided protein design construction method embodiment belong to the same concept, and the specific implementation process is detailed in the multimodal instruction-guided protein design method embodiment, which will not be repeated here.

[0058] Based on the same inventive concept, an embodiment further provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the method for protein design guided by the multimodal instruction is implemented, specifically including the following steps:

[0059] S1. Randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design;

[0060] S2, first using the text instruction to obtain a text-form generation condition as a modal instruction, and then converting the text-form generation condition into a protein-form generation condition across modalities as another modal instruction, thereby obtaining a multimodal instruction;

[0061] S3, constructing a protein design model, which includes a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and using the sample set to perform supervised training on the protein design model based on protein sequences;

[0062] S4. Use the trained protein design model to design protein sequences and visualize the designed protein sequences.

[0063] The computing device provided in the embodiment, in addition to the processor and memory, also includes hardware required for other services such as internal bus, network interface, memory, etc. at the hardware level. The memory is a non-volatile memory, and the processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the protein design method guided by the multimodal instructions described in S1-S4 above. Of course, in addition to the software implementation, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0064] Based on the same inventive concept, an embodiment further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the protein design method guided by the multimodal instruction is implemented, which specifically includes the following steps:

[0065] S1. Randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design;

[0066] S2, first using the text instruction to obtain a text-form generation condition as a modal instruction, and then converting the text-form generation condition into a protein-form generation condition across modalities as another modal instruction, thereby obtaining a multimodal instruction;

[0067] S3, constructing a protein design model, which includes a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and using the sample set to perform supervised training on the protein design model based on protein sequences;

[0068] S4. Use the trained protein design model to design protein sequences and visualize the designed protein sequences.

[0069] In the embodiment, computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data.

[0070] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A protein design method guided by multimodal instructions, It is characterized in that The following steps are involved: Randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design; First, a textual instruction is used to obtain a textual generation condition as a modal instruction, and then the textual generation condition is cross-modally converted into a protein-form generation condition as another modal instruction, thereby obtaining a multi-modal instruction; Constructing a protein design model, which includes a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and using a sample set to perform supervised training on the protein design model based on protein sequences; Use the trained protein design model to design protein sequences and visualize the designed protein sequences.

2. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that In a single sample, the textual instructions describe at least one property of a protein through natural language. The protein sequence correspondingly possesses at least one of the described properties and is composed of dozens to thousands of amino acids.

3. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that The pre-trained SciBERT encoder is used to convert text instructions into generation conditions in text form as a modal instruction.

4. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that The pre-trained Gaussian distribution model is used for cross-modal conversion, converting the text-based generation conditions into protein-based generation conditions as another modal instruction.

5. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that The process of fusing multimodal instructions by the multimodal fuser includes: After concatenating or averaging the multimodal instructions, the attention module and the fully connected layer are connected to reorganize to obtain fused instructions that are beneficial to protein design.

6. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that The decoder includes at least two transformer decoders and a fully connected layer, which processes and designs proteins according to the fusion instructions output by the multimodal fuser to obtain the designed protein sequence.

7. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that Visualize designed protein sequences, including: The designed protein sequence is folded using the AlphaFold 2 model to obtain a three-dimensional structure. The protein sequence is interpretably operated through the PyMOL visualization tool, and a visual three-dimensional structure image of the protein sequence is output.

8. The protein design method guided by multimodal instructions according to claim 1, It is characterized in that Also includes: Verify the accuracy and effectiveness of the protein design model in generating protein sequences, specifically: first calculate the similarity between the input text instructions and the designed protein sequence; Then randomly select 4 non-corresponding text instructions and calculate their similarity with the designed protein sequence. If the input text instruction has the highest similarity with the designed protein sequence, it means that the generation process is effectively guided by the text instruction, and the generation result is considered accurate, otherwise it is inaccurate.

9. A protein design device guided by multimodal instructions, It is characterized in that include: Sample construction, which is used to randomly extract pairs of text instructions and protein sequences from the SwissProtCLAP dataset as single samples, and construct a sample set based on multimodal instructions to guide protein design; An instruction processing module, which is used to first use the text instruction to obtain a text-form generation condition as a modal instruction, and then cross-modally convert the text-form generation condition into a protein-form generation condition as another modal instruction, thereby obtaining a multi-modal instruction; A model building module, which is used to build a protein design model, comprising a multimodal fuser for fusing multimodal instructions and a decoder for processing multimodal instructions and performing protein design, and using a sample set to perform supervised training on the protein design model based on protein sequences; The protein design module is used to design protein sequences using the trained protein design model and to visualize the designed protein sequences.

10. A computing device comprising a memory and one or more processors, wherein the memory stores executable code, It is characterized in that When the one or more processors execute the executable code, they are used to implement the protein design method guided by the multimodal instructions according to any one of claims 1 to 8.

Citation Information

Patent Citations

  • Protein model training method, protein data acquisition method and device

    CN117275570A

  • Method and system for generating amino acid sequence of protein with specific function and attribute by using deep learning technology

    CN118335190A

  • AU2005211654A1

Cited By

  • Multi-modal protein sequence generation method based on deep learning

    CN119943134A