Deep mutation scanning method based on protein-protein interaction in bacterial cells and binding affinity prediction method

Abstract

The invention discloses a deep mutation scanning method based on protein-protein interaction in bacterial cells and a binding affinity prediction method. In the invention, split dCas9 proteins (SEQ ID NO.1 and SEQ ID NO.2) are respectively subjected to fusion expression with a receptor and a ligand with binding affinity to be detected, and are combined with a toxic sgRNA guide sequence (SEQ ID NO.3), and when PPI occurs, a toxic dCas9: sgRNA complex is formed to inhibit bacterial growth, so that a PPI affinity size signal is converted into a bacterial cell generation speed signal, and the signal can be quantified into a copy number value through next-generation sequencing. The invention further designs a deep learning binding affinity prediction model, high-affinity sequences can be screened from theoretical mutation space, the workload and cost of experimental verification are remarkably reduced, and a reliable basis is provided for subsequent protein function optimization and design.

Description

Technical Field

[0001] The present invention belongs to the field of protein sequence-function relationship detection and prediction, and particularly relates to a deep mutation scanning method and a binding affinity prediction method based on protein-protein interaction in bacterial cells. Background Art

[0002] Proteins carry out most biochemical functions within cells and do so by interacting with each other. The specificity of these protein-protein interactions (PPIs), i.e. the information about which two proteins interact, is encoded in the DNA sequence and is therefore central to sequence-function relationships. PPIs are required for a variety of processes such as cell division, signaling, metabolism, transcription, and translation, and are central to the pathogenic mechanisms of most diseases. Therefore, the study of PPIs is of great interest in medical and pharmacological research.

[0003] With the demand for artificial intelligence protein design, many innovatively designed proteins need to systematically study the relationship between their core sites and phenotypes, such as the effect of specific sites on protein-protein interactions. Deep mutational scanning (DMS) detection can simultaneously evaluate the effects of hundreds of thousands or even millions of sequence variants on the function of a given molecule, and is a very suitable tool for large-scale research on protein-protein interactions. DMS can systematically analyze the sequence-function relationship of protein mutants by integrating large-scale mutation library construction, high-throughput screening and NGS sequencing. In deep mutation scanning detection, a variant library is created by mutagenesis or purposeful mutation of the sequence of interest, and then the library is selectively screened. Since scanning screening is usually performed in a pooled form, high-activity variants and low-activity variants under different selective conditions will be enriched or depleted accordingly due to the design of experimental conditions. Then, the second-generation high-throughput sequencing can be used to quantify the enrichment and depletion of the mutation library, so as to obtain the copy number of the corresponding mutant, reflecting the corresponding mutant phenotypic data, such as the affinity of protein-protein binding.

[0004] In the biological systems of bacteriophage, bacteria, yeast and mammalian cells, deep mutation scanning (DMS) technology has its own characteristics and is suitable for different research needs. Deep mutation scanning (DMS) based on bacterial systems is particularly suitable for applications in enzyme activity optimization, metabolic pathway modification and biosynthesis. The bacterial system introduces mutant genes into cells and achieves expression through plasmid transformation. Since bacteria have a short growth cycle, the experimental scale can be expanded to tens of thousands or even hundreds of thousands of mutants. Phenotypic analysis usually reacts to the corresponding mutants through growth competition experiments, and can also be combined with second-generation sequencing to quantify the growth rate of all mutants to obtain the phenotypic data of the mutants. However, due to the lack of the post-translational modification mechanism of eukaryotes, the application of bacterial deep mutation scanning (DMS) in studying the function of proteins that rely on post-translational modifications is limited.

[0005] With the surge in demand for artificial intelligence protein design, the demand for protein mutation data has also increased accordingly. Correspondingly, deep mutation scanning (DMS) technology has also posed challenges, and a deep mutation scanning (DMS) technology platform that can meet the needs of artificial intelligence protein design is needed, such as high throughput, high stability, ease of use, data format and high data quality. Summary of the invention

[0006] In response to the research needs of protein-protein binding in the field of artificial intelligence protein design, the purpose of the present invention is to provide a deep mutation scanning method based on protein-protein interactions in bacterial cells, as well as a method for predicting protein binding affinity by artificial intelligence modeling based on data.

[0007] The first aspect of the present invention is to provide a deep mutation scanning method based on protein-protein interactions in bacterial cells, comprising the following steps: S1 Construction of mutation library based on split dCas9 protein The split dCas9 protein includes C-dCas9 (SEQ ID NO.1) and N-dCas9 (SEQ ID NO.2); the split N-dCas9 and C-dCas9 are respectively fused and expressed with the target proteins (ligands and receptors) of the binding affinity to be tested through molecular biology technology, and only when specific interaction occurs between the target proteins, N-dCas9 and C-dCas9 are reassembled to form a complete functional dCas9 enzyme.

[0008] Preferably, the ligand amino acid sequence is designed at the C-terminus of N-dCas9, and a ligand plasmid is constructed; the receptor amino acid sequence is designed at the N-terminus of C-dCas9, and a receptor plasmid is constructed. The vectors of the ligand plasmid or the receptor plasmid are both Escherichia coli.

[0009] The mutation library is derived from DNA synthesis technology, and can generate target protein gene sequences containing multiple mutations through oligonucleotide array design. These sequences are cloned into affinity binding plasmids using seamless cloning technology. The mutation library here can be a random library of specified sites or a specified library, both of which can be obtained by primer synthesis or gene synthesis.

[0010] Preferably, the library designed based on the mutation site of the ligand amino acid sequence is amplified using primers to introduce the corresponding mutation, and the corresponding ligand mutation library is constructed using a seamless cloning method. In the case of specific experimental requirements, the ligand sequence can also be fixed, and then the library can be built based on the receptor sequence, that is, mutations are designed based on the mutation site of the receptor amino acid sequence to construct a receptor mutation library.

[0011] It should be noted that the mutation library here can be a random library of a specified site or a specified library, both of which can be synthesized by primers. However, if it is a specified library, the corresponding barcode needs to be synthesized at the same time for subsequent sequencing identification.

[0012] S2 Phenotypic analysis based on bacterial expression system The preferred bacterial expression system of the present invention is Escherichia coli; the phenotype of protein-protein interaction is binding affinity.

[0013] The present invention uses E. coli strains with different backgrounds, such as DH5alpha, MG1655 and BL21(DE3).

[0014] The constructed ligand mutant library is transfected into a host bacterium containing a receptor plasmid, or the constructed receptor mutant library is transfected into a host bacterium containing a ligand plasmid, a bacterial expression system is constructed for affinity binding reaction, and corresponding samples are collected for the next step of high-throughput sequencing.

[0015] Preferably, the above-mentioned bacterial system also includes an sgRNA plasmid, wherein the sgRNA is an sgRNA sequence that is toxic or kills host bacteria, and its expression combined with the dCas9 enzyme is toxic to cells, thereby inhibiting bacterial growth.

[0016] Furthermore, the nucleotide sequence of the sgRNA is shown in SEQ ID NO. 3. The present invention also tests sgRNAs that can directly inhibit essential genes of E. coli, but the combination of the toxic sgRNA of the present invention with dCas9 can make the affinity and cell growth more linear.

[0017] The present invention designs sgRNA so that it exhibits toxicity to bacterial growth when combined with complete dCas9. There are two options for bacterial toxic sgRNA. One is to use the toxic sgRNA guide sequence of the present invention to inhibit bacterial growth in the form of non-specific toxicity. The other is a sgRNA guide sequence designed to target essential bacterial genes (such as replication and transcription-related genes). The toxic effect of the sgRNA will be activated only after the split dCas9 protein is reassembled into a complete enzyme for the two differently designed sgRNA guide sequences. The growth state of bacteria is directly related to the interaction strength between the target proteins, which is manifested in that the stronger the affinity, the higher the reassembly efficiency of the split dCas9, the more active the expression of the toxic sgRNA, and the more obvious the inhibition of bacterial growth. In this way, the difference in bacterial growth is converted into a quantitative phenotype of the interaction between the target proteins, and the present invention provides a highly sensitive and linear protein interaction detection method.

[0018] Specifically, taking the ligand mutation library as an example, electro-competent cells of Escherichia coli including a receptor plasmid and an sgRNA plasmid are prepared, and the ligand mutation library (ligand mutation plasmid library) is transfected into the prepared electro-competent cells of Escherichia coli to obtain a recombinant strain; the recombinant strain is divided into two groups, and cultured in resistant LB solid culture medium with / without IPTG, respectively, the cells are eluted, and plasmids are extracted respectively to obtain the corresponding IPTG-free group and IPTG group after affinity reaction.

[0019] S3 high-throughput sequencing The samples that have completed the affinity reaction are subjected to high-throughput sequencing to obtain the corresponding data. Due to the relatively short target sequence and high throughput characteristics of the present invention, the Illumina sequencing platform is selected for the second-generation high-throughput sequencing.

[0020] Since the purpose of this step is to detect high mutations in a specific region, a two-step PCR method is used to prepare sequencing samples. Among them, the downstream primer sequence of PCR amplification contains the i7 index sequence. The sequencing samples are loaded together with other samples to increase the diversity of the samples according to the conventional sequencing steps, and sequenced using the Illumina Novasek 6000 platform.

[0021] S4 Data Processing According to the sequencing results, the corresponding copy number is calculated, the data is balanced, and the log2 multiple ratio is calculated to quantify the binding affinity of the amino acid sequence corresponding to each mutated DNA sequence, which is the sequence-binding affinity data. Data processing mainly includes balancing the data and then calculating the ratio of the copy number of different mutants under affinity and no affinity conditions.

[0022] As in the exemplary specific embodiment of the present invention, after obtaining the sequencing results, the corresponding copy numbers are first calculated for the 15552 different designed DNA sequences. Then the sequencing data is balanced. In the present invention, the copy number of the negative control is used to balance the data of the IPTG-free group and the IPTG-containing group. Based on the obtained copy number, the ratio of log2 multiples is calculated to quantify the binding affinity of each mutant DNA sequence to the corresponding amino acid sequence.

[0023] In a second aspect, the present invention provides a prediction method based on protein-protein binding affinity data in bacterial cells, collecting the binding affinity data set of the high-throughput ligand mutation library obtained by the above-mentioned DMS and the receptor protein to construct an artificial intelligence prediction model, comprising the following steps: (1) Artificial intelligence model construction: Data processing: First, the binding affinity dataset is preprocessed through the ESM2 model to generate fixed-length embeddings as model input data, and each sequence corresponds to an affinity value as the target label. 10% of the data in the dataset is set aside as a validation set to validate the model. The remaining data in the dataset is divided into training and test sets in an 80:20 ratio using the train_test_split function of scikit-learn, and randomization is set to ensure the repeatability of the results. The input embedding data and target values ​​are converted to PyTorch tensors to adapt to the deep learning framework.

[0024] (2) Model structure and training: The present invention designs a hybrid model combining convolutional neural network (CNN) and fully connected neural network (FFNN) to capture local features of protein sequences and realize nonlinear regression. The model structure includes four one-dimensional convolutional layers, the number of convolution kernels in each layer increases layer by layer, and the feature dimension reduction is performed through the maximum pooling layer. The parameters of the convolutional layer include the number of convolution kernels is 64, the convolution kernel size is 5, and the pooling layer size is 4. The output of the convolutional layer is flattened and input to the fully connected layer. The fully connected layer contains three layers of neurons, the number of hidden units in the first layer is 64, and the last layer outputs a continuous scalar as the predicted value. A Dropout layer is added to the fully connected layer to avoid overfitting, and the Dropout rate is set to 0.0. The Adam optimizer is used in the optimization process, the learning rate is set to 0.001, and the learning rate is adjusted according to the verification loss in combination with the dynamic learning rate scheduler. The mean square error (MSE) is used as the loss function during the training process. The model is trained for 20 cycles with a batch size of 128.

[0025] (3) Model evaluation and optimization: During the model training process, a random search was performed on the hyperparameter space, and a total of 20 iterations were performed to select the best hyperparameter combination. Finally, the model obtained the optimal R-squared value (R 2 ) is 0.9017, and the corresponding optimal parameters are: number of convolution kernels 64, convolution kernel size 5, pooling layer size 4, hidden layer units 64, learning rate 0.001, batch size 128, training cycle 20, Dropout rate 0.0. The model performance is evaluated by the R square value of the test set, R 2 The value is 0.9136. Finally, save the model for prediction.

[0026] (4) The amino acid sequence of the protein to be tested is input into the model, and the predicted value is the binding affinity prediction value. Specifically, an amino acid sequence is selected from the theoretical mutant protein space, and then the constructed model is used to predict the binding affinity of the amino acid sequence.

[0027] In a third aspect, the present invention provides a high-affinity leucine zipper protein ligand.

[0028] In a specific embodiment of the present invention, the interaction between leucine zipper proteins is used as a technical verification. Leucine zipper CZ (SEQ ID NO.4) is used as a receptor, leucine zipper NZ (SEQ ID NO.5) is used as a ligand, and a mutation library is constructed based on the ligand amino acid sequence.

[0029] The 8 amino acid residues at position 5 and positions 7-13 of SEQ ID NO.5 were selected for mutation, and a mutation library was constructed by PCR amplification. The primer sequence of the PCR amplification is shown in SEQ ID NO.8.

[0030] The above-mentioned deep mutation scanning method was used to construct a mutation library, perform phenotypic analysis and high-throughput sequencing, and build an artificial intelligence model to predict binding affinity. The theoretical number of different proteins with mutations at 8 amino acid sites is 2.5 × 10 10 , we selected the amino acid sequences of 1 million proteins from the theoretical protein space, and then used the constructed model for prediction. We first analyzed the relationship between the affinity of the predicted sequence and the amino acid sequence. For low-affinity and high-affinity sequences, the constructed model has obvious amino acid preferences. For example, for the low-affinity group, it prefers amino acid Q or amino acid P. However, for the high-affinity group, there is a clear amino acid preference for specific sites, such as amino acid E for site 8, amino acid N for site 10, and amino acid I for site 13.

[0031] For the predicted sequences, two high affinity sequences and two low affinity sequences were taken and cloned into pARB118 for in vivo verification. The high affinity sequences were SEQ ID NO.17 and SEQ ID NO.18, and the corresponding predicted affinity values ​​were -4.35 and -3.94. The low affinity sequences were SEQ ID NO.19 and SEQ ID NO.20, and the corresponding predicted affinity values ​​were 0.07 and 0.06.

[0032] The present invention combines the split dCas9 protein and the toxic sgRNA guide sequence, thereby converting the affinity signal of the protein-protein interaction into a signal of the toxicity of the dCas9:sgRNA complex to the cell, which is ultimately reflected as the growth rate of the bacterial cell. The stronger the affinity, the higher the concentration of the composite dCas9 produced, and thus the greater the toxicity of dCas9:sgRNA to the cell, and the lower the copy number of the corresponding cell. In this way, the affinity signal of the protein-protein interaction is converted into a cell copy number signal, and the cell copy number can be collected in the form of a pool experiment using second-generation sequencing for high-throughput data collection. This design achieves linear data presentation of the affinity signal. Based on the above design, the present invention has the following advantages over the prior art: 1. Solve the defects of the prior art. The present invention effectively overcomes the various limitations of the prior art. Compared with the phage system, the present invention solves the problem that the non-cellular in vivo environment cannot simulate complex biological functions through the design of split dCas9 protein in bacteria, while improving the efficiency of specific protein expression. In addition, compared with the yeast system, the present invention reduces the operational complexity and experimental cost by optimizing the experimental process, avoiding the efficiency bottleneck caused by the slow growth and low conversion efficiency of yeast. In comparison with the mammalian cell system, the present invention uses bacteria as a platform, with lower cost and high-throughput experimental capabilities, to meet the needs of large-scale screening and phenotypic data, while taking into account research accuracy.

[0033] 2. Combination of high throughput and high sensitivity. The present invention utilizes innovative split dCas9 protein and toxic sgRNA design to achieve high-sensitivity detection of protein-protein interactions. When the target protein interacts, the split dCas9 reassembles to form an enzyme with full functionality, and combines with the toxic sgRNA to form a dCas9:sgRNA toxic complex that is toxic to the host cell, so that the affinity change can be intuitively reflected by the cell growth rate. This design greatly simplifies the process of collecting phenotypic data. At the same time, the present invention uses seamless cloning technology and barcode labeling to construct a mutation library, combined with high-throughput second-generation sequencing, which greatly improves the processing capacity of the mutation library and supports the screening of millions of mutants, thereby meeting the research needs of large-scale sequence-phenotype relationships.

[0034] 3. The present invention closely combines deep learning technology with experimental data, generates protein sequence embedding through the ESM2 model, and designs a hybrid model of convolutional neural network (CNN) and fully connected neural network (FFNN) to perform nonlinear neural network modeling on mutant sequences and phenotypic data. The R-square value of the optimized model reaches 0.9136, showing extremely high affinity prediction accuracy. Through artificial intelligence modeling, the present invention can screen high-affinity sequences from the theoretical mutation space, significantly reduce the workload and cost of experimental verification, and provide a reliable basis for subsequent protein function optimization and design.

[0035] 4. The present invention significantly reduces the experimental cost of deep mutation scanning. The design based on the bacterial system avoids the high reagent consumption and equipment dependence of the yeast and mammalian systems, and the construction and screening of the mutation library can be completed under standard laboratory conditions. The optimized experimental steps and the application of multiple strain backgrounds simplify the operation process and improve the efficiency of mutation library construction and screening. While ensuring high-quality data, the experimental cycle is shortened, which is particularly suitable for rapid screening of large-scale mutants.

[0036] 5. Adapt to the needs of artificial intelligence protein design. The present invention is oriented to the needs of artificial intelligence-driven protein design, provides high-throughput, high-quality phenotypic data, and meets the urgent needs of current protein engineering for large-scale mutation data. Combined with the affinity prediction model, the present invention realizes the ability to quickly screen functional mutants from large-scale mutants, providing a new solution for optimizing protein function. Through accurate modeling and prediction of sequence and affinity, the present invention can significantly accelerate the progress of protein engineering, drug development and related basic research, and provide powerful data and technical support for artificial intelligence protein design. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of converting protein-protein affinity signals into cytotoxic signals based on split dCas9 and toxic sgRNA in bacterial cells of the present invention.

[0038] Figure 2 The figure is a schematic diagram of the process of high-throughput mutation scanning based on protein-protein interactions in bacterial cells of the present invention.

[0039] Figure 3 The scatter plot is a comparison of the copy number of all mutant ligand IPTG groups and No_IPTG groups. The horizontal axis represents the copy number without the addition of IPTG (No_IPTG group), and the vertical axis represents the copy number under the addition of IPTG (IPTG group). Each point represents the copy number distribution of a specific mutant under two conditions.

[0040] Figure 4The Log2 fold change (Log2FC) distribution diagram of all ligand mutation sequences. The horizontal axis represents the Log2 fold change value (Log2FC), and the vertical axis represents the frequency. Each bar represents the frequency of the mutation sequence within the corresponding Log2FC range.

[0041] Figure 5 The training architecture of the artificial intelligence model of the present invention is shown, which combines a one-dimensional convolutional neural network (Conv1D) and a fully connected neural network to extract features from input data and perform regression prediction.

[0042] Figure 6 Schematic diagram of the amino acid distribution of the predicted low-affinity group and high-affinity group of ligands. The horizontal axis represents the different amino acid mutation sites, the vertical axis represents the frequency, and the size of the amino acid letter reflects the frequency of occurrence of the amino acid at a specific mutation site.

[0043] Figure 7 Growth curve ratios of two high-affinity sequences (SEQ ID NO.17 and SEQ ID NO.18) and two low-affinity sequences (SEQ ID NO.19 and SEQ ID NO.20) of the predicted leucine zipper ligands verified in bacterial cells. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with the embodiments, but they should not be construed as limiting the scope of protection of the present invention.

[0045] Figure 1 The figure is a schematic diagram of converting protein-protein affinity signals into cytotoxic signals based on split dCas9 and toxic sgRNA in bacterial cells, showing the mechanism of the present invention for converting protein-protein interaction signals into cytotoxic signals based on the split dCas9 system and toxic sgRNA. In bacterial cells, plasmid A and plasmid B encode split N-terminal dCas9 (N-dCas9) and C-terminal dCas9 (C-dCas9), respectively. When the target protein-protein interaction occurs, N-dCas9 and C-dCas9 are successfully assembled into functional dCas9 by affinity. Subsequently, the functional dCas9 binds to the toxic sgRNA from another plasmid to form a dCas9:sgRNA complex. The complex non-specifically targets DNA sequences in the bacterial genome, resulting in inhibition of bacterial growth.

[0046] Figure 2It is a schematic diagram of the process of high-throughput mutation scanning based on protein-protein interactions in bacterial cells. The schematic diagram shows the specific process of high-throughput mutation scanning based on protein-protein interactions in bacterial cells of the present invention. First, a plasmid A mutation library is constructed to encode a ligand mutant protein library. The plasmid A mutation library and plasmid B drive the reassembly of split dCas9 through protein interactions in bacteria in bacterial cells, bind to toxic sgRNA, and generate toxic signals. Subsequently, the composition of mutants in the mutation library and their correlation with toxic signals are analyzed by second-generation sequencing technology to generate a large number of functional-related data sets. These data are further used to train artificial intelligence models to build models that can predict high-affinity and specific functional protein mutants. Ultimately, the model can be applied to the screening and design of protein mutants with specific functions or high affinity.

[0047] The present invention uses the interaction between leucine zipper proteins as a technical verification. Leucine zipper CZ (SEQ ID NO.4) is used as a receptor, leucine zipper NZ (SEQ ID NO.5) is used as a ligand, and mutations are performed based on the ligand sequence to construct a mutation library. The implementation case is based on the following 7 steps: Step 1: Design and construct a mutation library based on NZ. Based on the plasmid A (receptor plasmid) and plasmid B (ligand plasmid) of the present invention, a receptor plasmid pARB117 (for kanamycin resistance) containing CZ-dCas9(C) and a ligand plasmid pARB118 (for chloramphenicol resistance) containing dCas9(N)-NZ were constructed respectively. The amino acid sequence of CZ-dCas9(C) is SEQ ID NO.6, and the amino acid sequence of dCas9(N)-NZ is SEQ ID NO.7. Based on SEQ ID NO.2 sequence (ALKK E L QANKKEL AQLKWELQALKKELAQ), select the 5th and 7-13th (underlined parts), a total of 8 mutation sites for mutation. Since the theoretical library size of 8 mutation sites is 2.5 × 10 10 , which is currently impossible to construct. The present invention constructs a library of 15552 in size based on the 8 selected sites, uses primers (SEQ ID NO.8) to introduce corresponding mutations, and uses seamless cloning methods to construct corresponding mutant plasmid libraries.

[0048] The specific cloning is as follows: primers SEQ ID NO.8 and SEQ ID NO.9 are used to amplify clone fragment 1, and primers SEQ ID NO.10 and SEQ ID NO.11 are used to amplify clone fragment 2. After PCR amplification to obtain fragment 1 and fragment 2, the fragments are gel-recovered to obtain purified fragments 1 and fragment 2. Then fragments 1 and fragment 2 are cloned using the homologous recombination method (Gibson Assembly). Since the mutant library is constructed, clones with a mutation library of more than 100 times are required. The final cloned product is electrotransfected into DH5alpha Escherichia coli, a total of 5 transfections are performed, and all transfection products are plated on 10 chloramphenicol selective LB solid culture square plates (10 cm × 10 cm). The culture plate is placed at 37°C for overnight culture, and then all clones are eluted as much as possible, collected and subjected to plasmid extraction to obtain pARB118-NZ-N8-library. A total of 2.6 × 10 6 clones, which is 167 times the size of the designed mutation library.

[0049] Step 2: Transfect the mutant library and perform affinity reaction of the mutant library. The purpose of this step is to transfect the constructed ligand mutant library into Escherichia coli containing the receptor plasmid, perform affinity binding reaction, and collect the corresponding samples for the next step of next-generation sequencing. Prepare electroporation competent cells of MG1655 Escherichia coli containing pARB117 and sgRNA plasmid (containing SEQ ID NO.3 as sgRNA guide sequence). The pARB118-NZ-N8-library plasmid library prepared in the previous step was transfected into the prepared competent cells. A total of 10 transfections were performed, and after mixing, 5 ml was equally plated onto 10 LB solid culture square plates (10 cm × 10 cm) containing chloramphenicol and kanamycin with or without IPTG. A total of 5.1 × 10 6 The number of clones is 328 times the size of the designed mutation library and 1.96 times the size of the plasmid cloned from the library, which meets the requirements of the pooled experiment. The culture plate was placed in a 37°C incubator for 8 hours, and then all the cells were washed off. The cells without IPTG induction were kept as one group, and the cells with IPTG induction were kept as another group. The plasmids of the two groups of cells were extracted separately to obtain the corresponding IPTG-free group and IPTG group after affinity reaction.

[0050] Step 3: Perform second-generation sequencing. The purpose of this part is to perform second-generation sequencing on the samples that have completed the affinity reaction to obtain the corresponding data. Since the purpose of this experiment is to detect the high mutation in a specific region, a two-step PCR method is used to prepare the sequencing sample. First, primers P1-F (SEQ ID NO.12) and P1-R (SEQ ID NO.13) are used to perform PCR with the plasmid prepared in the previous step as a template. After obtaining the first-step PCR product, gel recovery is performed, and then the gel recovery product is used as a template for the second-step PCR. The second-step PCR uses primers P2-F (SEQ ID NO.14) and P2-R1 (SEQ ID NO.15), and P2-R2 (SEQ ID NO.16) as primers. P2-R1 corresponds to the sample without IPTG group, and P2-R2 corresponds to the sample with IPTG group. P2-R1 contains the i7index sequence TAACGATT, and P2-R2 contains the i7 index sequence ATGTAGAC. The sequencing samples were loaded together with other samples according to the conventional sequencing steps to increase the diversity of the samples, and sequenced using the Illumina Novasek 6000 platform.

[0051] Step 4: Data analysis. After obtaining the sequencing results, we first calculated the corresponding copy numbers for the 15,552 different DNA sequences we designed. A total of 15,549 sequences were found in the sequencing, and some sequences could not be found because the copy numbers were too low. Then we balanced the sequencing data. We used the copy number of the negative control without protein-protein interaction to balance the data to make the data of the IPTG-free group and the IPTG-containing group more balanced. From the copy number of the sequencing results, the results were in line with expectations ( Figure 3 ), that is, the copy number of some sequences in the IPTG group is lower than its corresponding copy number in the non-IPTG group. Based on the obtained copy number, the log2 multiple ratio is calculated to quantify the affinity of the amino acid sequence corresponding to the DNA sequence of each mutant ligand to interact with the receptor amino acid. The affinity distribution of all mutant ligands is shown in Figure 4 .

[0052] Figure 3 The scatter plot is a comparison of the copy number of all mutant ligand IPTG groups and the No_IPTG group, comparing the copy number of all ligand mutant IPTG groups and the No IPTG group. The horizontal axis represents the copy number without the addition of IPTG (No_IPTG group), and the vertical axis represents the copy number under the addition of IPTG (IPTG group). Each point represents the copy number distribution of a specific mutant under two conditions. Most of the data points are distributed along the diagonal, but some data points deviate from the diagonal, reflecting that the copy number of some mutants has changed significantly under the addition of IPTG.

[0053] Figure 4 The figure shows the Log2 fold change (Log2FC) distribution histogram of all ligand mutant sequences. The horizontal axis represents the Log2 fold change value (Log2FC), and the vertical axis represents the frequency. Each bar represents the frequency of the mutant sequence in the corresponding Log2FC range. The figure shows that the Log2FC values ​​of most mutants are concentrated between -1 and 0, indicating that the expression levels of these mutants between the control group and the experimental group change little, and only a few mutants have Log2FC values ​​lower than -3 or higher than 0, indicating that these mutants have significant expression differences between the two groups.

[0054] Step 5: AI model building: Data processing: First, the protein sequence data is preprocessed by the ESM2 model to generate fixed-length embeddings as model input data, and each sequence corresponds to an affinity value as the target label. The dataset contains 13,982 samples (1567 samples are taken out for model validation). The dataset is divided into training and test sets in a ratio of 80:20 using the train_test_split function of scikit-learn, and randomization is set to ensure the repeatability of the results. The input embedding data and target values ​​are converted to PyTorch tensors to adapt to the deep learning framework.

[0055] Model structure and training: This paper designs a hybrid model that combines convolutional neural network (CNN) and fully connected neural network (FFNN) to capture sequence-level local features and implement nonlinear regression. For the specific AI model training architecture, see Figure 5 First, the input layer receives one-dimensional data containing 640 nodes, which is adjusted to the format of (1, 640) by the reshape layer to adapt to the convolution operation. The model passes through four convolutional layers (Conv1D) in sequence, each layer uses 32, 64, 128 and 256 convolutional filters, respectively, and the output length is gradually reduced to 320, 160, 80 and 40 with pooling operations. The convolution output is then flattened into a one-dimensional tensor containing 10,240 nodes by the flattening layer for input into the fully connected layer. The fully connected layer includes three layers. The first layer maps 10,240 nodes to 128 nodes and applies activation functions and Dropout operations to prevent overfitting; the second layer further maps 128 nodes to 64 nodes and applies activation functions. Finally, the output layer maps 64 nodes to a single node for generating a regression output of the target value. The artificial intelligence training architecture of the present invention extracts local features of amino acids through multi-layer convolution, and integrates features through the fully connected layer to realize training models and predictions.

[0056] Model evaluation and optimization: During the model training process, a random search was performed on the hyperparameter space for 20 iterations to select the best hyperparameter combination. Finally, the model obtained the optimal R-squared value (R 2 ) is 0.9017, and the corresponding optimal parameters are: number of convolution kernels 64, convolution kernel size 5, pooling layer size 4, hidden layer units 64, learning rate 0.001, batch size 128, training cycle 20, Dropout rate 0.0. The model performance is evaluated by the R square value of the test set, R 2 The value is 0.9136. Finally, save the model for prediction.

[0057] Step 6: Prediction of high affinity leucine zippers and their verification. The theoretical number of different proteins with mutations at the 8 amino acid sites is 2.5 × 10 10 , we select 1 million protein sequences from the theoretical protein space and then use the constructed model to make predictions. We first analyze the relationship between the affinity of the predicted sequence and the amino acid sequence, such as Figure 6 As shown, the horizontal axis represents different amino acid mutation sites, the vertical axis represents frequency, and the size of the amino acid letter reflects the frequency of occurrence of the amino acid at a specific mutation site. For low-affinity and high-affinity sequences, the constructed model has obvious amino acid preferences. For example, for the low-affinity group, it prefers amino acid Q or amino acid P. However, for the high-affinity group, there is an obvious amino acid preference for specific sites, such as site 8 prefers amino acid E, site 10 prefers amino acid N, and site 13 obviously prefers amino acid I.

[0058] For the predicted sequences, two high-affinity sequences and two low-affinity sequences were taken respectively, cloned into pARB118, and verified in bacteria in vivo. The high-affinity sequences are SEQ ID NO.17 and SEQ ID NO.18, and the corresponding predicted affinity values ​​are -4.35 and -3.94. The low-affinity sequences are SEQ ID NO.19 and SEQ ID NO.20, and the corresponding predicted affinity values ​​are 0.07 and 0.06. The DNA sequences of the above four sequences were individually cloned into the ligand plasmid, and then transfected into Escherichia coli DH5alpha containing the receptor plasmid and the toxic sgRNA plasmid. Then a single clone was picked and placed in a 96-well plate containing 200 ul LB culture medium and culture medium containing IPTG in each well for culture. The culture was carried out for a total of 18 hours, and the OD600 value was detected once every 5 minutes in each cycle. After the experiment, the ratio of the OD600 values ​​of the non-IPTG group and the IPTG group was calculated. The results are shown in the table. Figure 7 . from Figure 7It can be seen that the growth curve ratio of SEQ ID 17 and SEQ ID 18 increases with the increase of cycle number, indicating that after IPTG induces affinity reaction, cell growth is inhibited, indicating that SEQ ID 17 and SEQ ID 18 sequences can bind to the receptor with high affinity. The growth curve ratio of SEQ ID 19 and SEQ ID 20 remains stable, indicating that their corresponding sequences have no affinity for receptor binding. This growth curve ratio is highly consistent with the predicted affinity data, indicating that the model constructed based on deep mutation scanning data in the case of the present invention can successfully predict the affinity of different ligand amino acid sequences.

[0059] It should be noted that the method of the present invention for constructing a mutation library based on split dCas9 for deep mutation scanning can also be used for the detection and analysis of other protein-ligand functional relationships besides leucine zipper proteins. The sgRNA of the present invention can also be other reporter genes that can reflect protein-ligand binding affinity. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A deep mutation scanning method based on protein-protein interactions in bacterial cells, characterized in that: The steps include: S1 constructs a mutation library based on split dCas9 protein: the split dCas9 protein includes C-dCas9 and N-dCas9, the amino acid sequence of C-dCas9 is shown in SEQ ID NO.1, and the amino acid sequence of N-dCas9 is shown in SEQ ID NO.2; Fusing C-dCas9 and N-dCas9 to the receptor and ligand of the binding affinity to be tested, respectively; mutating the amino acid sequence of the receptor to obtain a receptor mutation library, or mutating the amino acid sequence of the ligand to obtain a ligand mutation library; S2 Phenotypic analysis based on bacterial expression system: The constructed ligand mutation library is transfected into a host bacterium containing a receptor plasmid and an sgRNA plasmid to construct a bacterial expression system, or the constructed receptor mutation library is transfected into a host bacterium containing a ligand plasmid and an sgRNA plasmid to construct a bacterial expression system, inducing the expression of ligands and receptors in bacterial cells to perform affinity binding reactions, and collecting corresponding samples for the next step of high-throughput sequencing; the sgRNA expression combined with dCas9 protein can inhibit bacterial growth; S3 high-throughput sequencing: Prepare samples for sequencing and perform high-throughput sequencing on samples that have completed affinity reactions; S4 Data processing: According to the sequencing results, the corresponding copy number is calculated, and the log2 multiple ratio is calculated to quantify the binding affinity of the amino acid sequence corresponding to each mutated DNA sequence, which is the sequence-binding affinity data.

2. The deep mutation scanning method based on protein-protein interaction in bacterial cells according to claim 1, characterized in that: In step S1, the ligand or ligand mutant library is designed at the C-terminus of N-dCas9; the receptor or receptor mutant library is designed at the N-terminus of C-dCas9.

3. The deep mutation scanning method based on protein-protein interaction in bacterial cells according to claim 1, characterized in that: The receptor mutation library and the ligand mutation library further comprise a barcode sequence.

4. The deep mutation scanning method based on protein-protein interaction in bacterial cells according to claim 1, characterized in that: The nucleotide sequence of the sgRNA is shown in SEQ ID NO.

3.

5. The deep mutation scanning method based on protein-protein interaction in bacterial cells according to claim 1, characterized in that: The amino acid sequence of the receptor is shown in SEQ ID NO.4, and the amino acid sequence of the C-dCas9 after being connected to the C-terminus of the receptor is shown in SEQ ID NO.6; The amino acid sequence of the ligand is shown in SEQ ID NO.5, and the amino acid sequence after the N-dCas9 is connected to the N-terminus of the ligand is shown in SEQ ID NO.

7.

6. The deep mutation scanning method based on protein-protein interaction in bacterial cells according to claim 5, characterized in that: The 5th and 8 amino acid residues at positions 7-13 of the amino acid sequence of the ligand were selected for mutation, and a mutation library was constructed by PCR amplification. The nucleotide sequence of the PCR amplification primer is shown in SEQ ID NO.

8.

7. The deep mutation scanning method based on bacterial protein-protein interaction according to claim 1, characterized in that: In the bacterial expression system, the plasmid vector is Escherichia coli, and the host bacteria is Escherichia coli.

8. A method for predicting binding affinity based on protein-protein interactions, characterized in that: The steps include: (1) using the deep mutation scanning method described in any one of claims 1 to 7 to obtain a high-throughput protein binding affinity dataset; preprocessing the binding affinity dataset through the ESM2 model to generate fixed-length embedded data as model input data, with each sequence corresponding to an affinity value as a target label; The input embedding data and target values ​​are converted into PyTorch tensors to adapt to the deep learning framework; the dataset is divided into training set, test set and validation set; (2) Model structure and training: A hybrid model of convolutional neural network and fully connected neural network is used to capture local features of the sequence and implement nonlinear regression. The Adam optimizer is used in the model optimization process in combination with a dynamic learning rate, and the learning rate is adjusted according to the verification loss. The mean square error is used as the loss function during training; (4) Model evaluation and optimization: Perform random search and iteration on the hyperparameter space to select the best hyperparameter combination; evaluate the performance of the model through the R-squared value of the test set, and save the model for prediction; (5) Input the amino acid sequence of the protein to be tested into the model and obtain the predicted value, which is the binding affinity prediction value.

9. The method for predicting binding affinity based on protein-protein interaction according to claim 8, characterized in that: The convolutional neural network includes four one-dimensional convolutional layers, the number of convolution kernels in each layer increases layer by layer, and feature dimension reduction is performed through the maximum pooling layer; the fully connected layer includes three layers of neurons, the number of hidden units in the first layer is 64, and the last layer of neurons outputs a continuous scalar as a predicted value.

10. A high-affinity leucine zipper protein ligand, characterized in that: The high-affinity leucine zipper protein ligand is obtained according to the prevention method according to claim 8 or 9, and the amino acid sequence of the leucine zipper protein ligand is shown in SEQ ID NO.17 or SEQ ID NO.18.

Images