Artificial intelligence model and system for assessing expression of recombinant human-derived proteins

Abstract

This invention discloses an artificial intelligence model and system for evaluating the expression of recombinant human proteins. The AI ​​model can predict the expression data of a given sequence (natural or artificial). It exhibits high prediction accuracy and a low MSE value, which can be beneficial for the actual production of proteins.

Description

Technical Field

[0001] This invention belongs to the field of biotechnology, specifically relating to an artificial intelligence model and system for evaluating the expression of recombinant human proteins. Background Technology

[0002] Despite advancements in protein structure analysis and drug discovery, high-quality protein expression remains a bottleneck. Research institutions and pharmaceutical companies preparing human proteins (especially membrane proteins, such as G protein-coupled receptors, GPCRs, one of the most important known drug targets) generally employ suspension cell culture, relying on manual trial-and-error methods for expression and purification. Due to numerous influencing factors and their interactions (insertion, mutation, and truncation in target DNA construction; substrate ratios, temperature, and humidity in cloning, plasmid construction, transfection, cell growth, and protein expression), this trial-and-error process is extremely time-consuming, inefficient, and lacks reproducibility. The analysis of each protein typically takes 3–5 years, making human protein-based drug development both time-consuming and labor-intensive.

[0003] The hundreds of GPCRs structures resolved to date mostly exhibit diverse ligand binding and conformations, fundamentally altering our understanding of GPCR structure, receptor activation mechanisms, and the molecular determinants of ligand efficacy. Three complementary techniques have been developed for GPCR structure resolution to achieve better expression and purity: (i) fusing small soluble proteins (such as T4 lysozyme (T4L) and thermostable cytochrome b562 (BRIL)) with intracellular loop 3 (ICL3) or the N-terminus of the receptor; (ii) isolating camel or mouse antibodies with specific conformations to lock the receptor in a specific state; and (iii) achieving thermostability of the receptor through systematic mutagenesis. In these strategies, the stability of the resulting GPCR-ligand complex is crucial for expression. The stability of receptors with T4L or BRIL fusions is ligand-dependent; typically, only ligands with very high affinity (10 pM to 10 nM) can achieve stable, high-expression of the target protein. The same applies to antibodies. For example, camel heavy chain antibodies (nanobodies) provide sufficient stability to the receptor, allowing for the determination of the structure of the β2-adrenergic receptor (β2AR) that binds to the natural ligand adrenaline. For highly unstable human proteins like GPCRs, in addition to stabilizing the receptor's specific conformation by adding ligands to aid expression, specific amino acid mutations can enhance its thermal stability. These mutations can increase overall stability by adding internal hydrogen bonds, hydrophobic interactions, or introducing new disulfide bonds, thereby improving expression success rates. Engineering disulfide bonds in the receptor structure was one of the early strategies used to stabilize the rhodopsin structure.

[0004] When the structures of identical or similar receptors are determined using thermostable receptors or receptor-T4L fusions, these structures are nearly identical within the experimental error range, with major differences typically confined to the circulating region due to different crystal contacts. However, conformational thermostabilization strategies offer significant advantages when multiple receptor structures are needed to understand ligand efficacy or for hit-to-guide optimization in drug development. For example, the thermostable β1 adrenergic receptor (β1AR) has been co-crystallized with 12 different ligands (inverse agonists, weak partial agonists, partial agonists, and full agonists), allowing for detailed understanding of the molecular determinants of ligand efficacy and the initial process upon agonist binding, thus increasing the probability that the receptor becomes active and can bind to G proteins. In the thermostable adenosine A2A receptor, structures with multiple lead compounds binding to the receptor have facilitated the development of preclinical candidate compounds for Parkinson's disease. Thermostable GPCRs are ideal for both structure determination and drug development, but the thermostabilization process is time-consuming and costly.

[0005] Point mutations have proven particularly important for improving the thermal stability of protein structures and making them suitable for structure-based drug design applications. For example, G protein-coupled receptors involve co-crystallization of the receptor with drugs that typically have low affinity. For instance, point mutations have been used to improve the thermal stability of the β1-adrenergic receptor (ADRB1) and the A2A adenosine receptor (A2AAR), resulting in more than a dozen structures for each receptor, including co-crystallization with ligands in the micromolar affinity range, in both agonist and antagonist bound states (DOI: https: / / doi.org / 10.1073 / pnas.1100185108; DOI: https: / / doi.org / 10.1021 / jm400140q; DOI: https: / / doi.org / 10.1038 / nature07101; DOI: https: / / doi.org / 10.1038 / nature09746). In the thermally stable A2AAR case, the structure and biophysical properties of the initial hits were characterized, leading to structure-based discovery and optimization of preclinical candidate compounds for Parkinson's disease (DOI:https: / / doi.org / 10.1021 / jm201455y). Furthermore, the construction of thermostable GPCRs can simplify the characterization of ligand binding biochemical properties in sensor-based high-throughput screening (HTS) (DOI: https: / / doi.org / 10.1016 / j.molmed.2015.09.002) and the measurement of ligand binding kinetics via surface plasmon resonance (SPR) (DOI: https: / / doi.org / 10.1021 / jm400140q; DOI: https: / / doi.org / 10.1016 / B978-0-12-381274-2.00005-4; DOI: https: / / doi.org / 10.1016 / j.ab.2008.12.011).

[0006] However, identifying stable mutations experimentally using alanine scanning (DOI:https: / / doi.org / 10.1007 / 978-1-4939-2914-6_1) or directed evolution methods (DOI:https: / / doi.org / 10.1073 / pnas.1317903111; DOI:https: / / doi.org / 10.1073 / pnas.1202107109) is a very resource-intensive process, and to date only a few GPCRs have been successfully stabilized (DOI:https: / / doi.org / 10.1016 / j.tips.2015.09.005). Furthermore, the stable mutations obtained by these methods exhibit very limited transferability between different GPCRs (DOI:https: / / doi.org / 10.3389 / fphar.2015.00082; DOI:https: / / doi.org / 10.3109 / 09687680903208239), requiring extensive stabilization activities for each individual receptor. Summary of the Invention

[0007] To address the lack of existing methods for predicting the expression level and purity of target proteins based on their sequences and in different cloning optimization systems, this invention provides an artificial intelligence model and system for evaluating the expression of recombinant human proteins.

[0008] To address the aforementioned technical problems, the present invention provides a technical solution as follows: a method for constructing an artificial intelligence model for evaluating the expression of recombinant human proteins, the method comprising:

[0009] Protein data is input into a Random Forest (RF) classifier for training. The data includes sequence data and expression data. The sequence data includes the original protein sequence category, insertion site, insertion category, mutation site, mutated amino acid category, truncation site, and truncation length. The expression data includes the yield in a certain amount of biomass, the expression percentage of recombinant human protein, the peak position in the purification curve, the highest point of the Gaussian fit of the purification curve, and the confidence level of the Gaussian fit of the purification curve. The training involves predicting the expression data of the corresponding protein based on the protein sequence data, thus constructing the artificial intelligence model for evaluating the expression of recombinant human protein.

[0010] In some preferred embodiments of the present invention, one or more of the original protein sequence category, insertion sequence site, insertion sequence category, mutation site, mutated amino acid category, truncation site, and truncation sequence length are normalized based on the corresponding original sequence data of the recombinant human protein.

[0011] In some preferred embodiments of the present invention, the expression data includes one or more of the following: yield in a certain amount of biomass, percentage of recombinant human protein expression, peak position in the purification curve, highest point of Gaussian fitting of the purification curve, and confidence level of Gaussian fitting of the purification curve, which are normalized based on the corresponding original expression data of recombinant human protein.

[0012] In this invention, the category to which the original protein sequence belongs is the category to which a protein belongs within its family. Preferably, the original protein sequence is standardized as follows: it is classified by homology sequence similarity clustering, dividing the protein sequence into specific categories based on sequence similarity. For example, a multiple sequence similarity clustering algorithm is used to classify GPCRs proteins into 138 categories, each denoted by a numerical label 1-138. The implementation parameters of the clustering algorithm are as described in Example 1.

[0013] In this invention, the insertion site is a specified amino acid position, typically located at the N-terminus, C-terminus, or middle of the protein sequence. Preferably, the insertion site is standardized as follows: based on its position in the protein sequence, it is classified into three tags [0, 1, 2], where the insertion site is located at the N-terminus, C-terminus, or in between (referred to as the middle). The classification of the insertion site is based on its absolute position region in the protein sequence, specifically defined as follows: N-terminal tag (0): the insertion site is located before the first amino acid of the protein sequence; C-terminal tag (1): the insertion site is located after the last amino acid of the protein sequence; Middle tag (2): the insertion site is located at all positions between the first and last amino acids.

[0014] In this invention, the inserted sequence is a specific functional fragment, such as a sequence tag or a specific structural domain, which can classify the inserted sequence in a certain way to obtain the category to which the inserted sequence belongs. Preferably, the category to which the inserted sequence belongs is standardized as follows: according to the specific sequence type of the inserted sequence, each sequence is assigned a tag, and a corresponding digital tag is generated. The tag is not directly related to biological classification, but only indicates the classification result.

[0015] In this invention, the mutation site is a specified amino acid position. Preferably, the mutation site is standardized as follows: each mutation site in the sequence is extracted and integrated into a standardized input data. Then, the maximum number of mutation sites m in a single sample in the training dataset is determined, and the label vector of the mutation site count for each sample is padded to length m, with the remaining positions padded with 0. For example, for the mutation site ['L42A','F88A','G103A','Y202F'], the corresponding standardized input data is [42,88,103,202]. Based on the maximum number of mutations in its sample being 6, the standardized data is adjusted to [42,88,103,202,0,0].

[0016] In this invention, the mutant amino acid category refers to amino acids with specific functions or structural characteristics. Preferably, the mutant amino acid category is standardized as follows: the mutated amino acid is matched with 22 standard amino acids, and a label (1 to 22) is assigned to each amino acid, where each label represents a specific amino acid type. If the amino acid belongs to a non-standard amino acid of another category (e.g., a heterologous protein), then label 23 is assigned. All mutation categories in a sequence are integrated into a single standardized input data. Then, the maximum number of mutation sites m in a single sample in the training dataset is determined, and the label vector of the number of mutation sites for each sample is padded to length m, with the remaining positions padded with 0. For example, for the mutation sites ['L42A', 'F88A', 'G103A', 'Y202F'], the corresponding standardized input data is [1, 1, 1, 5,]. Based on the fact that the maximum number of mutations in its sample is 6, the standardized data is adjusted to [1, 1, 1, 5, 0, 0,].

[0017] In this invention, the truncated site is the site or region where the cleaved amino acid is located. Preferably, the mutated site is standardized as follows: the start site of each truncated sequence is recorded, and all truncated sites in a single sequence are integrated into a single standardized input data. Then, the maximum number of truncated sites n in a single sample in the training dataset is determined, and the label vector of the number of truncated sites for each sample is padded to length n, with the remaining positions padded with 0. For example, for the truncated sequence [[1,52],[251,249],[348,368]], the corresponding truncated site standardized input data is [1,251,348]. For example, when the maximum number of truncated sites in the sample is 5, the standardized data is adjusted to [1,251,348,0,0].

[0018] In this invention, the truncated sequence length is 2-900 amino acid residues. Preferably, the truncated sequence length is standardized as follows: the length of each truncated sequence is recorded, and all truncated lengths in a single sequence are integrated into a standardized input data. Then, the maximum number of truncated sites n in a single sample in the training dataset is determined, and the label vector of the number of truncated sites for each sample is filled to length n, with the remaining positions padded with 0. For example, for the truncated sequence [[1,52],[251,249],[348,368]], the corresponding standardized length is [52,3,21]. When the maximum number of truncated sites in the sample is 5, the standardized data is adjusted to [52,3,21,0,0].

[0019] In this invention, the yield in a certain quantitative biomass represents the total amount of recombinant human protein contained in a certain total volume (e.g., 40 ml) of expression and purification product. The yield in the certain quantitative biomass directly reflects the richness of recombinant human protein expression. Its representation can be, for example, by converting the peak absorbance measured by the instrument into the total amount of the corresponding product in that total volume of solution; the specific form can be any form that reflects the yield. Preferably, the yield in the certain quantitative biomass is standardized as follows: the peak absorbance measured by the instrument is converted into the total amount of the corresponding product in a certain quantitative solution, in units such as μg.

[0020] In this invention, apart from the location and sequence length, the numbers representing the labels in the aforementioned input data are only used to distinguish the labels and have no actual meaning. Within the same type of label, the numerical codes can be interchanged.

[0021] Specifically, when the recombinant human protein is a membrane protein, its expression percentage is the percentage of cells on the cell surface where the recombinant human protein expression can be detected on the cell surface in the total number of cells tested. For membrane proteins, this reflects whether the recombinant human protein can be correctly localized to the cell membrane, which is crucial for its functional activity. Preferably, the expression percentage of the recombinant human protein on the cell surface as a membrane protein is standardized by dividing the number of cells detecting recombinant human protein expression on the cell surface by the total number of cells tested to obtain a relative percentage value. When the recombinant human protein is a water-soluble protein, its expression percentage can be obtained by detecting a specific band formed by a specific molecular weight of the target recombinant human protein. Preferably, the expression percentage of the recombinant human protein as a water-soluble protein is standardized by normalizing the band intensity and calculating the relative content of the target recombinant human protein in the sample. That is, the proportion of the target recombinant human protein band signal to the total signal is obtained, thus reflecting the percentage of protein expression.

[0022] In this invention, the peak position in the purification curve represents the difference between the time of separation of the recombinant human protein in high-performance liquid chromatography (HPLC) during the purification process and the mean of the separation data for all proteins of this type. Because recombinant human proteins belonging to the same protein class share similar physical properties, closer separation times indicate a higher probability that they are recombinant human proteins. Preferably, the peak position in the purification curve is standardized as follows: the mean of the times corresponding to the highest peak positions obtained from the separation of proteins during all purification processes is calculated, and then the absolute time difference (which can be positive or negative) between the current recombinant human protein sample's highest peak position and this mean is calculated, in units such as minutes.

[0023] In this invention, the highest point of the Gaussian fit of the purification curve represents the abundance of recombinant human protein during separation in high-performance liquid chromatography (HPLC) during the purification process. A higher value at the highest point indicates a greater final yield. Preferably, the highest point of the Gaussian fit of the purification curve is standardized as follows: the highest peak value of the Gaussian fit relative to the actual experimental curve is calculated, representing the absorbance value, for example, in mAU.

[0024] In this invention, the reliability of the Gaussian fitting of the purification curve represents the purity of the recombinant human protein isolated during the purification process. A higher reliability indicates higher purity of the final recombinant human protein. Preferably, the reliability of the Gaussian fitting of the purification curve is standardized as follows: the proportion of the overlap area between the fitted Gaussian curve and the actual curve within a range of 3 sigma is calculated and converted into a relative percentage value.

[0025] In some preferred embodiments of the present invention, the parameters of the random forest classifier are set as follows: the random forest classifier at each layer adopts the following parameter settings: the maximum tree depth is 17 and the maximum number of trees is 200.

[0026] In some preferred embodiments of the present invention, the data of the recombinant human protein is randomly divided into a training set and a test set during the training process. The ratio of the data in the training set to the test set is preferably 6:4-9:1, and more preferably 8:2.

[0027] In some preferred embodiments of the present invention, the metrics used to evaluate the artificial intelligence model during training are mean-square error (MSE) and / or AUC (area under the ROC curve). MSE reflects the difference between the predicted and actual values ​​of the artificial intelligence model; the smaller the difference, the closer the predicted value is to the actual value, and the more accurate the prediction. For example, the model's performance can be determined by evaluating the MSE of the model's predictions against the actual data for each expression of data.

[0028] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0029] In this invention, the artificial intelligence model can be used to predict the expression level and purity of proteins in at least the following cells: prokaryotic cells E.coli BC21, insect cells Sf9, Hi5, and Ssf9-3, and mammalian cells 293F, 293T, and GnTi.

[0030] To address the aforementioned technical problems, the present invention provides a technical solution: a model for evaluating protein expression levels and purity, wherein the artificial intelligence model is constructed using the construction method described in the present invention.

[0031] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0032] To address the aforementioned technical problems, the present invention provides a technical solution: a system for evaluating the expression of recombinant human proteins, the system comprising:

[0033] A data receiving module is used to receive or input protein sequence data, as described herein;

[0034] The judgment and output module is used to, after the reception or input is completed, output the evaluation result of the expression data of the recombinant human protein through an artificial intelligence model as described herein, wherein the expression data is as described herein.

[0035] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0036] To address the aforementioned technical problems, the present invention provides a technical solution: a method for evaluating the expression of recombinant human proteins, the method comprising inputting the sequence data of the recombinant human protein into an artificial intelligence model as described in the present invention, and obtaining an evaluation result of the expression data of the recombinant human protein.

[0037] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0038] To solve the above-mentioned technical problems, the present invention provides a technical solution as follows: an evaluation system for the expression of recombinant human protein, the evaluation system comprising: the yield in a certain amount of biomass, the percentage of recombinant human protein expressed on the cell surface, the position of the highest peak in the purification curve, the highest point of the Gaussian fitting of the purification curve, and the reliability of the Gaussian fitting of the purification curve.

[0039] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0040] To address the aforementioned technical problems, the present invention provides a technical solution: the application of sequence data of recombinant human protein expression as described herein in evaluating expression data of the recombinant human protein as described herein.

[0041] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0042] To address the aforementioned technical problems, the present invention provides a technical solution: a computer-aided method for evaluating protein expression levels and purity, the evaluation method comprising the following steps:

[0043] Step 1: Receive or input the sequence data of the protein to be tested, as described herein;

[0044] Step 2: Input the sequence data received or input in Step 1 into the artificial intelligence model as described in this invention, and output the evaluation results of the expression data of the protein to be tested, as described herein.

[0045] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0046] To solve the above-mentioned technical problems, the present invention provides a technical solution as follows: a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it can realize the functions of the system as described in the present invention, or realize the steps of the method for evaluating protein expression level and purity as described in the present invention.

[0047] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0048] To solve the above-mentioned technical problems, the present invention provides a technical solution as follows: an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is used to execute the computer program to realize the function of the system as described in the present invention, or to implement the steps of the method for evaluating protein expression level and purity as described in the present invention.

[0049] In some preferred embodiments of the present invention, the recombinant human protein is a GPCR.

[0050] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of the present invention.

[0051] The reagents and raw materials used in this invention are all commercially available.

[0052] The positive and progressive effects of this invention are as follows:

[0053] The artificial intelligence model provided by this invention can be used to predict the expression data of a given sequence (natural or artificial). For example, it can predict expression data in different expression systems (such as insect cells, prokaryotic cells, and mammalian cells). Existing related models only focus on local problems (such as codon optimization, structure prediction, etc.). The artificial intelligence model integrates relevant data from multiple expression systems to establish a mapping relationship between sequence characteristics and expression purification performance, and has high prediction accuracy and low MSE value, which can help facilitate the actual production of proteins. Attached Figure Description

[0054] Figure 1 This is to output a comparison between the actual and predicted values ​​of the data.

[0055] Figure 2 As part of the workflow of the intelligent protein workstation, the local intelligent system interface can predict protein expression and purification based on the user's experimental conditions.

[0056] Figure 3 The cloud service platform continuously collects new experimental data (including original protein sequences, insertion sites and types, mutation sites and mutated amino acids, and truncation sites) and iteratively trains the model to optimize it. Detailed Implementation

[0057] The present invention is further illustrated below by way of embodiments, but the invention is not limited to the scope of the embodiments described herein. Experimental methods in the following embodiments that do not specify specific conditions were performed according to conventional methods and conditions, or as selected according to the product instructions.

[0058] Example 1

[0059] This embodiment uses random forest to accurately map the correspondence between parameters involved in plasmid construction and protein expression and purification data. Specifically, the input includes the original protein sequence, insertion site and type, mutation site and mutated amino acid, and truncation site. The output includes the yield of the target recombinant human protein in a certain quantitative biomass (i.e., cell lysate of the expression system) (e.g., the protein content of the target protein, i.e., CXCR3, in 40 mL of cell lysate of the expression system expressing recombinant CXCR3), the expression percentage of the target recombinant human protein on the cell surface, the position of the peak in the purification curve, the highest point of the Gaussian fit of the purification curve, and the confidence level of the Gaussian fit of the purification curve.

[0060] Specifically as follows:

[0061] To describe the expression effect of the target recombinant human protein in this experiment, the following parameters were selected: yield in a certain amount of biomass, percentage of target recombinant human protein expression on the cell surface, expression level of target recombinant human protein on the cell surface, peak position in the purification curve, highest point of Gaussian fitting of the purification curve, and confidence level of Gaussian fitting of the purification curve. For example, after C-terminal His tag insertion and L140W site mutation of the CXCR3 receptor protein, and after sequence truncation (Δ1-52, Δ249-251, Δ348-368), the yield of target recombinant human protein in 40 mL of biomass was 0.072683 μg, the surface expression percentage was 5.17%, and the purification peak position shifted to 12.468 min (shift +0.46843 min, average for similar proteins 12.0 min). The highest point of Gaussian fitting of the purification curve was 0.061951 mA, and the confidence level of Gaussian fitting of the purification curve was 59.4606%. These parameter values ​​will change accordingly depending on the modification conditions (e.g., C-terminal truncation). The yield in a given amount of biomass represents the total amount of the target recombinant human protein contained in a given amount of purified expression product, directly reflecting the richness of the target recombinant human protein expression. The percentage of target recombinant human protein expression on the cell surface measures the proportion of cells in a sample of 2000 tested cells where the target recombinant human protein expression can be detected on the cell surface. For membrane protein studies, this reflects whether the protein can be correctly localized to the cell membrane, which is crucial for its functional activity. The cells used for protein expression include: prokaryotic E.coli BC21 cells, insect cells Sf9, Hi5, and Ssf9-3, and mammalian cells 293F, 293T, and GnTi. In actual model training and prediction, these expression cells were not explicitly distinguished. The experimental data are sourced from the PBC Consortium database (https: / / www.pbcconsortium.org / ). The peak position in the purification curve represents the difference between the time of separation of the target recombinant human protein in HPLC and the mean of the separation data for all proteins of this type during the purification process. Since the target recombinant human proteins belong to the same protein class and share similar physical properties, closer separation times indicate a higher probability that it is indeed the target recombinant human protein. Similarly, the highest point of the Gaussian fit of the purification curve reflects the peak value of the target recombinant human protein during HPLC separation during the purification process; a higher value indicates a higher final yield. Finally, the reliability of the Gaussian fit of the purification curve reflects the purity of the target recombinant human protein separation during the purification process; a higher reliability indicates a higher purity of the final expressed protein. Therefore, a comprehensive evaluation of these parameters can, to a certain extent, reflect the expression level and purity of the target recombinant human protein.Higher protein expression percentages and levels, better purification (represented by the peak position in the purification curve and the peak value of the Gaussian fit), and higher yield (measured by the yield per 40 ml biomass) indicate good performance at each stage of the experimental procedure, thus increasing the likelihood of obtaining excellent expression of the target recombinant human protein. Predicting the final expression effect of the target recombinant human protein using these quantifiable indicators not only allows for the prediction of the protein's expression and purification data but also helps researchers select appropriate cloning optimization systems, providing strong support for experimental optimization.

[0062] The model and system used in this experiment were developed in Python. Taking GPCR protein as an example, the feasibility of the method of this invention was verified. First, experimental data on protein expression and purification were collected from the aforementioned GPCR protein database. This data consisted of real data accumulated during the experiment. The data included the original protein sequence, insertion sites and types, mutation sites and mutated amino acids, truncation sites, yield in a specific amount of biomass, percentage of expression of the target recombinant human protein on the cell surface, the peak position in the purification curve, the highest point of the Gaussian fit to the purification curve, and the reliability of the Gaussian fit to the purification curve.

[0063] To train the model, the feature data was standardized and split to ensure a balance in the impact of different features during model training. The original protein sequences were transformed into GPCR classification categories using a similarity clustering algorithm. The 138 standardized protein family classification labels were generated through the following multi-sequence similarity clustering steps: 1) Obtain the full-length sequences of all GPCRs from the public protein sequence database UniProt; 2) Calculate the sequence similarity score using a substitution matrix generated by the sequence alignment tool BLAST; 3) Set a gap penalty rule to balance the impact of insertions / deletions on evolutionary distance, with a linear gap initiation penalty of -0.3 and an extension penalty of -0.1; 4) Set a filtering threshold of sequence identity ≥75% to ensure high conservation of functional domains within the same cluster; 5) Based on the filtered similarity matrix, perform hierarchical clustering using the bottom-up aggregation UPGMA algorithm; 6) Stop aggregation when the average evolutionary distance between clusters is ≥0.45, i.e., the corresponding sequence similarity is ≤55%. 7) Assign a unique numerical label 1-138 to each final cluster. The label is not directly related to the biological classification and only indicates the sequence similarity classification result.

[0064] The insertion sequence sites and types in the feature data are split into two categories: insertion sequence sites and insertion sequence categories. The insertion sequence categories are transformed into the categories of the insertion sequences contained in the database through similarity clustering. The insertion sequence sites are transformed into a certain category based on the C-terminus, N-terminus, and middle position of the sequence.

[0065] The mutation site and mutated amino acid are divided into two categories: mutation site and mutated amino acid category. The mutation site is the sequence in which it is located, and the mutated amino acid category is the type of amino acid after mutation. The type is classified according to the total amino acid category. The truncated site is converted into the truncated sequence start sequence and truncated length. Therefore, the feature data input to the model includes: the category of the original protein sequence, the insertion sequence site, the insertion sequence category, the mutation site, the mutated amino acid category, the truncated site, and the truncated sequence length. The output data includes the yield (μg) of a certain amount of biomass (in ml, for example, 40 ml), the expression percentage (%) of the target recombinant human protein on the cell surface, the peak position (minute) of the purification curve, the highest point of the Gaussian fitting of the purification curve (UV280 detection signal / mAU), and the confidence of the Gaussian fitting of the purification curve. The output data are all continuous variables. After being normalized to values ​​between 0 and 1, they can be directly converted into the output features of the model. The standardized input data of some exemplary proteins (protein numbers are (1)-(10)) are shown in Table 1, and the original input data of some exemplary proteins are shown in Table 2. Examples of inserted sequences are shown in Table 3, and amino acid codes are shown in Table 4.

[0066] Table 1. Raw input data for some exemplary proteins.

[0067]

[0068]

[0069] Table 2. Sample input data after protein standardization.

[0070]

[0071]

[0072] Note: The original protein sequence IDs are normalized based on the categories of the original sequences in Table 1 above.

[0073] Table 3. Exemplary Inserted Label Sequences

[0074]

[0075]

[0076] Table 4 Amino Acid Encoding Table

[0077] Serial Number Single letter abbreviation Three-letter abbreviation Chinese name 1 A Ala alanine 2 C Cys Cysteine 3 D Asp Aspartic acid 4 E Glu glutamic acid 5 F Phe Phenylalanine 6 G Gly glycine 7 H His Histidine 8 I Ile Isoleucine 9 K Lys Lysine 10 L Leu Leucine 11 M Met Methionine 12 N Asn Asparagine 13 O Pyl Pyrrolidone 14 P Pro proline 15 Q Gln glutamine 16 R Arg Arginine 17 S Ser Serine 18 T Thr threonine 19 U Sec selenocysteine 20 V Val Valine 21 W Trp Tryptophan 22 Y Tyr Tyrosine 23 X X Unknown amino acids

[0078] Among the better locations:

[0079] When the target recombinant human protein is a membrane protein, its expression percentage is the percentage of its expression on the cell surface. This percentage represents the proportion of cells on the cell surface where the target recombinant human protein expression can be detected in the total number of cells tested. The expression percentage of the target recombinant human protein on the cell surface is standardized by dividing the number of cells expressing the target recombinant human protein on the cell surface by the total number of cells tested to obtain the relative percentage value. When the target recombinant human protein is a water-soluble protein, its expression percentage can be obtained by detecting a specific band formed by a specific molecular weight of the target recombinant human protein. Preferably, the expression percentage of the recombinant human protein as a water-soluble protein is standardized by normalizing the band intensity and calculating the relative content of the target recombinant human protein in the sample. This yields the proportion of the target recombinant human protein band signal to the total signal, thus reflecting the percentage of protein expression.

[0080] The peak position in the purification curve represents the difference between the time of separation of the target recombinant human protein in high-performance liquid chromatography (HPLC) during the purification process and the mean of the separation data for all proteins of this type. Because the target recombinant human proteins belong to the same protein class and share similar physical properties, the closer the separation time, the greater the likelihood that it is the target recombinant human protein. The peak position in the purification curve is standardized as follows: the mean of the times corresponding to the highest peak positions obtained from the separation of proteins during all purification processes is calculated, and then the absolute time difference (which can be positive or negative) between the current target recombinant human protein sample's highest peak position and this mean is calculated, in minutes.

[0081] The highest point of the Gaussian fit of the purification curve represents the abundance of the target recombinant human protein during high-performance liquid chromatography (HPLC) separation. A higher value at the highest point indicates a greater final yield. Preferably, the highest point of the Gaussian fit of the purification curve is standardized as follows: the highest peak value of the Gaussian fit relative to the actual experimental curve is calculated, representing the absorbance value in mA.

[0082] The reliability of the Gaussian fitting of the purification curve represents the purity of the target recombinant human protein isolated during the purification process. A higher reliability indicates higher purity of the final target recombinant human protein. Preferably, the reliability of the Gaussian fitting of the purification curve is standardized as follows: the proportion of the overlap area between the fitted Gaussian curve and the actual curve within a range of 3 sigma is calculated and converted into a relative percentage value.

[0083] The training program for the random forest classifier was developed using a Python compiler. Each layer of the random forest classifier used the following parameter settings: maximum tree depth 17, maximum number of trees 200. A total of 1200 data points were obtained, including both training and validation sets.

[0084] Specific data are shown in Tables 1-3. Due to the large amount of data, only exemplary data are listed. Of all the data, 80% was used for training and 20% was used for testing (external validation). The mean squared error (MSE) was used as the evaluation metric to measure the difference between the model's predicted values ​​and the actual values.

[0085] Validation results on the test set showed that the model's prediction accuracy for different output parameters ranged from 60% to 95%. The MSEs corresponding to protein expression percentage, peak position of the purification curve, highest point of Gaussian fitting of the purification curve, yield of 40 ml of biomass, and confidence level of Gaussian fitting of the purification curve were 0.00427506, 0.02539316, 0.02057277, 0.0238477, and 0.02046304, respectively. The original output prediction data for exemplary corresponding proteins are shown in Table 5, and the normalized output prediction data for exemplary corresponding proteins are shown in Table 6.

[0086] Table 5 shows the output data before normalization.

[0087]

[0088] Table 6 shows the normalized prediction data.

[0089]

[0090] like Figure 1 As shown, most data points are located near the expected red line, verifying the effectiveness of this method in predicting protein expression information based on protein sequence information. This method can reduce the number of blind screenings and improve screening efficiency.

[0091] Conclusions (qualitative and / or quantitative):

[0092] This invention develops and validates an artificial intelligence model based on the random forest algorithm for predicting and optimizing the experimental process of protein purification and expression. By processing and analyzing characteristic parameters such as the original protein sequence, insertion site and type, mutation site and mutated amino acid type, and truncation site, this model can effectively predict the expression and purification results of proteins in experiments. For the protein purification and expression process, the model's output parameters use the following data: yield in a certain amount of biomass, percentage of protein expression, peak position of the purification curve, highest point of the Gaussian fit of the purification curve, and confidence level of the Gaussian fit of the purification curve. Results on the test set show that the model's prediction accuracy ranges from 60% to 95% for different output parameters. This invention's method, by combining the random forest model with some features from actual experimental data, can improve the prediction accuracy of protein expression and purification results, providing support for experimental screening and laying the foundation for the research and development of other similar proteins.

[0093] Example 2

[0094] This example provides an interactive prediction system based on random forests. Users can access the system through its interactive interface (such as...). Figure 2 As shown, input the key parameters in plasmid construction, and correct and evaluate the model prediction results at the end of the experiment (e.g.) Figure 3 (As shown). This system consists of two modules:

[0095] 1) Model parameter input and prediction module

[0096] 2) Prediction result feedback and model update module

[0097] Similar to Example 1, the system collects and inputs relevant information about the designed protein sequence (including GPCRs or other types of proteins), including the original protein sequence category, insertion site, insertion sequence category, mutation site, mutated amino acid category, truncation site, and truncation length. The system then converts the input information into the model's corresponding input data format (e.g., standardized data format). Afterward, based on the existing model, the system outputs the predicted features, including: yield in a given amount of biomass, expression percentage of the target recombinant human protein, peak position in the purification curve, highest point of the Gaussian fit in the purification curve, and confidence level of the Gaussian fit in the purification curve. Users can choose whether to conduct subsequent experiments or readjust parameters based on the given prediction results.

[0098] After completing an experiment, users can feed back new data generated during the experiment to the system, inputting actual experimental data to replace the original predictions in the system, and then submitting the updated data. The system will use this new data to retrain the model, improving prediction accuracy by optimizing the model's parameters. During this process, the system uses evaluation metrics (such as mean squared error, MSE, in this case) to determine whether the new model outperforms the current model. If the evaluation results show that the new model performs better, it will be applied to the system for predictions in subsequent experiments.

Claims

1. A method for constructing an artificial intelligence model for evaluating the expression of recombinant human proteins, the method comprising: The protein data is input into a random forest classifier for training. The data includes sequence data and expression data. The sequence data includes the original protein sequence category, insertion site, insertion category, mutation site, mutated amino acid category, truncation site, and truncation length. The expression data includes the yield in a certain amount of biomass, the expression percentage of the target recombinant human protein, the peak position in the purification curve, the highest point of the Gaussian fit of the purification curve, and the confidence level of the Gaussian fit of the purification curve. The training involves predicting the expression data of the corresponding protein based on the protein sequence data, thereby constructing the artificial intelligence model for evaluating the expression of recombinant human protein.

2. The construction method as described in claim 1, characterized in that, The construction method includes one or more of the following: The original protein sequence belongs to the category of a protein within its family; preferably, the original protein sequence is standardized as follows: it is classified by homology sequence similarity clustering, dividing the protein sequence into specific categories based on sequence similarity to obtain standardized protein family classification labels. The insertion site is a specified amino acid position, typically located at the N-terminus, C-terminus, or middle of the protein sequence; preferably, the insertion site is standardized as follows: based on its position in the protein sequence, it is classified into several corresponding tags; and The inserted sequence belongs to a specific functional segment category, such as a sequence label or a specific structural domain; preferably, the inserted sequence category is standardized as follows: each sequence is assigned a label based on its specific sequence type; and The mutation site is a specified amino acid position; preferably, the mutation site is standardized as follows: each mutation site in the sequence is extracted and integrated into a standardized input dataset; and The mutated amino acid category is an amino acid with specific functional or structural characteristics; preferably, the mutated amino acid category is standardized as follows: the mutated amino acid is matched with 22 standard amino acids, corresponding tags are assigned, and all mutation categories in a sequence are integrated into a standardized input data; and The truncated site is the site or region where the cleaved amino acid is located; preferably, the mutated site is standardized as follows: the start site of each truncated sequence is recorded, and all truncated sites in a single sequence are integrated into a single standardized input data set; and The truncated sequence length is 2 to 900 amino acid residues; preferably, the truncated sequence length is standardized as follows: the length of each truncated sequence is recorded, and all truncated lengths in a single sequence are integrated into a standardized input data.

3. The construction method as described in claim 1, characterized in that, The construction method includes one or more of the following: The yield in a certain quantitative biomass represents the amount of expression and purification product in a certain total quantity, and the yield in the certain quantitative biomass directly reflects the richness of recombinant human protein expression; preferably, the yield in the certain quantitative biomass is standardized as follows: the peak absorbance measured by the instrument is converted into the mass value of the corresponding product of a certain quantitative solution; and When the recombinant human protein is a membrane protein, its expression percentage is the percentage of cells on the cell surface where the recombinant human protein expression can be detected on the cell surface in the total number of cells tested. Preferably, the expression percentage of the recombinant human protein on the cell surface is standardized by dividing the number of cells expressing the recombinant human protein on the cell surface by the total number of cells tested to obtain a relative percentage value. When the recombinant human protein is a water-soluble protein, its expression percentage can be obtained by detecting a specific band formed by a specific molecular weight of the target recombinant human protein. Preferably, the expression percentage of the recombinant human protein as a water-soluble protein is standardized by normalizing the band intensity and calculating the relative content of the target recombinant human protein in the sample. The peak position in the purification curve represents the difference between the time position of recombinant human protein separation in high-performance liquid chromatography (HPLC) during the purification process and the mean value of all such protein separation data. Preferably, the peak position in the purification curve is standardized as follows: the mean value of the time corresponding to the highest peak position obtained from the separation of proteins during all purification processes is calculated, and then the absolute time difference between the individual peak position of the recombinant human protein and this mean value is calculated. The highest point of the Gaussian fit of the purification curve represents the abundance of recombinant human protein during high-performance liquid chromatography (HPLC) separation in the purification process; preferably, the highest point of the Gaussian fit of the purification curve is standardized as follows: the highest peak value of the Gaussian fit relative to the actual experimental curve is calculated, representing the absorbance value; and The reliability of the Gaussian fitting of the purification curve is the purity of the recombinant human protein isolated during the purification process; preferably, the reliability of the Gaussian fitting of the purification curve is standardized as follows: the proportion of the overlap area between the fitted Gaussian curve and the actual curve within a range of 3 sigma to the total area is calculated and converted into a relative percentage value.

4. The construction method according to any one of claims 1-3, characterized in that, The construction method includes one or more of the following: The parameters of the random forest classifier are set as follows: Each layer of the random forest classifier uses the following parameter settings: maximum tree depth 17, maximum number of trees 200; and... In the training process, the data of the recombinant human protein are randomly divided into a training set and a test set. The preferred ratio of the training set to the test set is 6:4-9:1, more preferably 8:

2. The metrics used to evaluate the model during training are MSE and / or AUC, with MSE being preferred; and The recombinant human protein is a GPCR.

5. A model for evaluating the expression of recombinant human proteins, characterized in that, The artificial intelligence model is constructed by the construction method described in any one of claims 1-4; Preferably, the recombinant human protein is a GPCR.

6. A system for evaluating the expression of recombinant human proteins, characterized in that, The system includes: A data receiving module is used to receive or input protein sequence data, said sequence data as described in claim 1 or 2; The judgment and output module is used to, after the reception or input is completed, output the evaluation result of the expression data of the recombinant human protein according to the model as described in claim 5, wherein the expression data is as described in claim 1 or 3; Preferably, the recombinant human protein is a GPCR.

7. A method for evaluating the expression of recombinant human proteins, characterized in that, The method includes inputting the sequence data of the recombinant human protein into the model as described in claim 5 to obtain the evaluation results of the expression data of the recombinant human protein; Preferably, the recombinant human protein is a GPCR.

8. An evaluation system for recombinant human protein expression, characterized in that, The evaluation system includes: the yield of a certain amount of biomass, the percentage of recombinant human protein expressed on the cell surface, the peak position in the purification curve, the highest point of the Gaussian fit of the purification curve, and the reliability of the Gaussian fit of the purification curve. Preferably, the recombinant human protein is a GPCR.

9. The use of sequence data defined by the construction method of recombinant human protein expression as described in claim 1 or 2 in evaluating expression data defined by the construction method of recombinant human protein expression as described in claim 1 or 3; Preferably, the recombinant human protein is a GPCR.

10. A computer-aided method for evaluating the expression of recombinant human proteins, characterized in that, The evaluation method includes the following steps: Step 1: Receive or input the sequence data of the protein to be tested, wherein the sequence data is as described in claim 1 or 2; Step 2: Input the sequence data received or input in Step 1 into the model as described in claim 5, and output the evaluation results of the expression data of the protein to be tested, wherein the expression data is as described in claim 1 or 3; Preferably, the recombinant human protein is a GPCR.

11. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement the functions of the system as described in claim 6, or implement the steps of the method for evaluating recombinant human protein expression as described in claim 7. Preferably, the recombinant human protein is a GPCR.

12. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor is used to execute the computer program to implement the functions of the system as described in claim 6, or Implement the steps of the method for evaluating recombinant human protein expression as described in claim 7; Preferably, the recombinant human protein is a GPCR.

Images