An anti-freeze protein and screening method and application thereof based on protein language model

Abstract

The application belongs to the technical field of biotechnology, and discloses an antifreeze protein and a screening method and application thereof based on a protein language model, wherein training samples are constructed by integrating an antifreeze protein data set, sequence features of an ESM-AA-35M protein language model, K-mer component features and four types of physicochemical features of amino acids, and an antifreeze protein binary classification prediction model is constructed based on an XGBoost algorithm; and a high-throughput screening system is further established, so that high-activity antifreeze peptides can be quickly screened from a large-scale random peptide library. The AUC of the prediction model of the application reaches 0.978 in independent test set, and the screening reliability is strong; and the obtained artificial antifreeze protein has excellent thermal hysteresis activity, ice crystal recrystallization inhibition activity and single ice crystal growth inhibition activity, is low in toxicity and good in stability, and can be widely applied to the fields of organ transplantation low-temperature storage, biological sample preservation, low-temperature medical operation and the like.

Description

Technical Field

[0001] This invention belongs to the field of biotechnology, and in particular relates to an antifreeze protein and its screening method and application based on a protein language model. Background Technology

[0002] Cryogenic freezing technology is widely used in food storage, preservation of ex vivo organ transplants, and cryogenic ablation surgery for tumors. Food preservation and organ preservation are essentially forms of biological sample preservation. Low temperatures can effectively slow down cell metabolism and extend the preservation time of cells and tissues. However, the main reason limiting the large-scale application of cryogenic technology is the high water content of cells. The phase transition of water at low temperatures damages cells. Simultaneously, as ice crystals grow, solutes gradually accumulate inside the cells, causing a sharp increase in osmotic pressure and cell death. Organs and tissues lose their original functions, leading to decreased utilization and wasted medical resources. The preservation time of most organs and tissues is limited to only a few hours, and a large number of transplanted hearts and lungs are discarded due to loss of function. Therefore, there is an urgent need for safe and effective preservation methods for the long-term preservation of biological samples.

[0003] Existing commercial cryoprotectants are mainly composed of small organic molecules such as dimethyl sulfoxide (DMSO) and ethylene glycol. They possess strong cell permeability and hydrogen bond formation capabilities, driving the slow outflow of intracellular water during extracellular freezing, reducing intracellular free water content, decreasing the amount of ice formed, and minimizing damage to cell membranes from sharp ice crystals. However, they have significant drawbacks: high concentrations can damage internal cell structures, and in severe cases, cause gene damage and epigenetic effects. This limits the application of small organic molecules to large-volume tissues and organs.

[0004] Antifreeze proteins (AFPs), also known as thermal hysteresis proteins, ice structural proteins, or ice-binding proteins, protect cells from damage in extremely cold environments by inhibiting ice crystal formation. They primarily rely on thermal hysteresis and ice crystal adsorption inhibition mechanisms to suppress ice crystal growth. They are effective at low concentrations, exhibit good biocompatibility, and are core materials for next-generation cryoprotectants. However, natural antifreeze proteins face three major obstacles to industrialization: 1) extremely low abundance of AFPs in natural organisms; 2) low sequence similarity and significant structural differences among different species; and 3) long screening cycles, high costs, and extremely low throughput of traditional experimental methods such as DSC and cryogenic microscopy (IRI).

[0005] Language models were first applied to machine translation. By learning from large amounts of text data, they extract patterns and store them as parameters. In practice, these parameters are calculated and converted into text output by a decoder. Protein sequences and natural language share a natural similarity; a single amino acid can be considered a single term, and a protein sequence is a sentence composed of multiple terms. Protein language models are the application of language models in the protein field. By learning from protein family sequence information, they store information such as the physicochemical properties and interactions between residues, evolutionary conserved sites, and mutational relationships as parameters within the model. Current technology lacks a complete antifreeze peptide design system, making it difficult to rapidly obtain antifreeze peptides with high ice-inhibiting activity. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes an antifreeze protein and its screening method and application based on a protein language model. Addressing the shortcomings of existing small-molecule cryopreservatives (high toxicity), low screening efficiency of natural antifreeze proteins, and incomplete feature capture by traditional protein prediction models, this invention constructs a high-precision prediction model for high-throughput screening of novel artificial antifreeze short peptides. Simultaneously, it provides two low-toxicity, high-ice-inhibiting-activity 10AA antifreeze peptides, expanding the selection of cryogenic biological preservation materials.

[0007] To achieve the above objectives, this invention provides a method for screening antifreeze proteins based on a protein language model, comprising the following steps: S1. Integrate antifreeze protein sequences as positive samples, remove redundant sequences to obtain positive samples, extract non-antifreeze protein sequences as negative samples in a 1:1 ratio of positive samples to negative samples, construct a comprehensive dataset, and split the positive samples and negative samples into training datasets and independent test datasets. S2. Using a protein language model, multiple candidate antifreeze proteins are converted into feature vectors and K-mer component features and amino acid physicochemical scoring features are simultaneously fused to construct the model input feature vector. S3. Based on K-mer component characteristics and amino acid physicochemical scoring characteristics, a screening interval was set to perform initial screening of multiple candidate antifreeze proteins and obtain a preliminary screening candidate peptide library. S4. The candidate peptide libraries screened in the initial screening were subjected to antifreeze activity prediction, cytotoxicity filtering, solubility filtering, and stability filtering in sequence, and finally the antifreeze proteins were obtained.

[0008] Preferably, in step S1, the antifreeze protein sequences include the AFP481, AFP920, and AFP604 datasets; positive and negative samples are split into training datasets and independent test datasets in an 8:2 ratio.

[0009] Preferably, in step S2, the protein language model is the ESM-AA-35M protein language model; the process of extracting the sequence feature vector of the protein language model is as follows: input a protein sequence of length L, which is converted into a 1×480-dimensional sequence feature vector after average pooling; K-mer component features and amino acid physicochemical scoring features include hydrophobicity coefficient score GOLD730101, solubility free energy score CHAM820102, isoelectric point score ZIMJ680104, and stability contribution score ZHOH040102.

[0010] Preferably, in step S3, the K-mer component characteristics and amino acid physicochemical scoring characteristics are the sum of the scores of the K-mer component characteristics and amino acid physicochemical scoring characteristics corresponding to the amino acids in each candidate antifreeze protein; the screening interval is set with the mean of the K-mer component characteristics and amino acid physicochemical scoring characteristics as the center and 0.5 standard deviations as the radius.

[0011] Preferably, in step S4, the antifreeze activity prediction is performed by using the XGBoost algorithm to test the antifreeze probability of the initially screened candidate peptide library. Peptides with an antifreeze probability higher than the screening range are considered antifreeze proteins, while those lower than the screening range are considered non-antifreeze proteins. Cytotoxicity filtering is performed using the ToxinPred tool to predict and retain non-toxic peptides. Solubility filtering is performed using a combination of the Camsol and Peptide tools to predict and retain peptides with a Camsol value greater than 1 and a Peptide prediction solubility greater than 70%. Stability filtering is performed using the ProtParam tool to predict and retain peptides with an instability index lower than 40.

[0012] An antifreeze protein is also provided, the amino acid sequence of which is shown in SEQ ID NO.1 or SEQ ID NO.2.

[0013] Preferably, the antifreeze protein has single ice crystal growth inhibitory activity.

[0014] The application of the antifreeze protein in cryopreservation of biological samples, temporary storage of organ transplants, or cryomedical procedures is also provided.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: (1) Based on the protein language model, the present invention constructs a high-precision antifreeze protein prediction model, and uses the model to design and screen in a dataset containing 1 billion 10AA random peptides, successfully obtaining two antifreeze peptides with superior antifreeze activity, Peptide11 (SEQ ID NO.1) and Peptide6 (SEQ ID NO.2). This screening method solves the problems of long screening cycle and high cost, and can be further applied to databases with larger data volume, and quickly complete large-scale sequence evaluation, significantly reducing R&D costs. At the same time, the qualified rate of antifreeze activity of the candidate peptides obtained by screening reaches 100%, the false positive rate is low, and the screening reliability is strong.

[0016] (2) This invention integrates the AFP481, AFP920, and AFP604 antifreeze protein datasets and removes redundancy to construct training samples. It integrates three types of features to construct the input vector: extracts the deep evolution and residue interaction information of protein sequences through the ESM-AA-35M protein language model, and superimposes K-mer component features and four types of amino acid physicochemical property features to comprehensively cover the sequence attributes and functional basis of antifreeze proteins. This reduces the problem of insufficient capture of sequence functional information by traditional single feature models. The obtained prediction model can achieve an AUC of 0.978 and a PR-AUC of 0.981 on the independent test set. The accuracy, recall, and F1 score are all 0.924, which shows excellent prediction accuracy and has the ability to efficiently and accurately identify the antifreeze potential of sequences.

[0017] (3) The screening method of the present invention provides a general technical platform for the research and development of novel antifreeze biomaterials; the obtained antifreeze short peptides can be used as core functional components in multiple fields such as food low-temperature storage, low-temperature preservation of ex vivo organs and tissues, and low-temperature ablation surgery of tumors, providing new material support for the upgrading of low-temperature biological preservation technology and having industrial application value.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 The figure shows the test results of the antifreeze protein prediction model on the independent test set. In the figure, A is the independent test performance index, B is the ROC curve, and C is the PR curve. Figure 2 The mass spectrum of the synthesized antifreeze peptides is shown in the figure. A represents peptide Peptide11 and B represents peptide Peptide6. Figure 3The results of the single ice crystal growth inhibition activity assay are shown in the figure. A is the single ice crystal growth image at different time points, with a scale bar of 100 micrometers. B is the single ice crystal area growth at different time points. PBS represents the negative control group of phosphate buffer. Peptide6 represents the peptide Peptide6 group, Peptide11 represents the peptide Peptide11 group, ns represents P>0.05, and ** represents P<0.01. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention. Experimental methods in the following embodiments that do not specify specific conditions are generally determined according to national standards. Experimental instruments, equipment, and reagents in the following embodiments that do not specify their sources are all commercially available materials.

[0023] Unless otherwise defined or stated, all technical and scientific terms used in this invention have the same meaning as those skilled in the art. Furthermore, any methods and materials similar to or equivalent to those described herein can be applied to the methods of this invention. It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0024] Example 1 Construction of an antifreeze protein prediction model.

[0025] The dataset constructed in this embodiment includes antifreeze protein sequences and non-antifreeze protein sequences, which are used as positive and negative data, respectively, and integrates three published and widely used datasets.

[0026] The first dataset is AFP481. This dataset is widely used and serves as a benchmark dataset for multiple antifreeze protein prediction models. The Pfam database was searched using antifreeze protein keywords, and based on 221 seed sequences of antifreeze proteins, high-similarity sequences were searched in a non-redundant sequence database using the PSI-BLAST tool, resulting in 481 antifreeze protein sequences. The corresponding negative data consisted of 9493 non-antifreeze protein sequences in the Pfam database.

[0027] The second dataset is AFP920. This dataset was collected by searching the UniProtKB database. As of June 20, 2023, a total of 6,509 antifreeze proteins were collected. Redundancy was removed using the CD-HIT tool based on a 40% sequence similarity, resulting in 920 antifreeze protein sequences.

[0028] The third dataset is AFP604. This dataset was obtained by searching the Unipro and Interpro databases using keywords and manually verifying 7567 antifreeze proteins. Redundancy was removed using CD-HIT, and sequences similar to AFP481 were removed using BLAST sequence similarity comparison, retaining 604 antifreeze protein sequences.

[0029] 1. Dataset Integration and Construction: The AFP481, AFP920, and AFP604 antifreeze protein sequence datasets were integrated, and redundant sequences were removed using CD-HIT, resulting in 1512 antifreeze protein sequences. Negative non-antifreeze protein sequences were randomly selected from the 9493 non-antifreeze protein sequences in the baseline dataset AFP481. Due to the large difference in the number of positive and negative samples, the 1512 positive samples were kept constant, and sequences were randomly selected from the negative sample database at a 1:1 ratio to obtain the comprehensive dataset.

[0030] Positive and negative samples in the comprehensive dataset are split into training dataset and independent test dataset in an 8:2 ratio.

[0031] 2. Feature Extraction: Converting protein sequences into feature vectors containing sequence information is a key step in achieving binary classification prediction. Three sets of descriptors are selected to construct scoring vectors, including protein language models, K-mer composition and physicochemical property descriptors.

[0032] Specifically, the ESM-AA-35M protein language model is selected to transform protein sequences into feature vectors, hereinafter referred to as ESM-AA. A protein sequence of length L is input, and mean pooling is performed to transform it into a 1×480-dimensional vector. Simultaneously, K-mer component features and four types of amino acid physicochemical scoring features from the AAindex database (hydrophobicity coefficient score GOLD730101, solubility free energy score CHAM820102, isoelectric point score ZIMJ680104, and stability contribution score ZHOH040102) are concatenated to obtain the complete input features.

[0033] 3. XGBoost Model Training: Based on the above characteristics, the extreme gradient boosting (XGBoost) algorithm was selected for model training and independent testing. During XGBoost model training, the input dimension was 480, the threshold was 0.5, and predicted probabilities above 0.5 were considered antifreeze proteins, while those below 0.5 were considered non-antifreeze proteins.

[0034] Performance Evaluation: The antifreeze protein prediction model was evaluated on an independent test set, and the results are as follows: Figure 1 As shown, the AUC was 0.978, the accuracy was 0.924, the F1 score was 0.924, the recall was 0.924, and the precision was 0.924. The PR-AUC on the independent test set was 0.981. This indicates that the model exhibits good predictive performance and can accurately predict antifreeze proteins.

[0035] In addition, the performance of this model was evaluated and compared with other AFP prediction models, including AFP-CKSAAP and AFP-LSE, on the same independent test set. The results are shown in Table 1. On the same test dataset, the AUC and accuracy of the comparable model AFP-CKSAAP were only 0.920 and 0.836, respectively; while the AUC and accuracy of AFP-LSE were only 0.830 and 0.821, respectively. This model significantly outperforms the other models.

[0036] Table 1 Comparison of Predictive Performance

[0037] Example 2 High-throughput screening design for random peptides.

[0038] 1. Random Peptide Library Generation and Initial Physicochemical Screening: 1 billion 10AA peptides were randomly generated to construct a random candidate peptide library. The physicochemical properties of 387 experimentally validated 10AA peptides from the Uniport database were calculated using indicators from the AAindex database: hydrophobicity coefficient (GOLD730101), solubility free energy (CHAM820102), isoelectric point (ZIMJ680104), and stability contribution (ZHOH040102). Each indicator was summed over all amino acid scores for each peptide to obtain its score. A reasonableness interval for each indicator was defined with a radius of 0.5 standard deviations, centered on the mean of the overall scores, for filtering. A total of 22,422,000 peptides were obtained to construct the peptide library.

[0039] 2. Multi-stage filtration.

[0040] (1) Prediction and screening of antifreeze activity of candidate peptides: The XGBoost model obtained in Example 1 was used to predict the antifreeze probability of each candidate peptide (the model is input, and the model predicts the probability value that the peptide is an antifreeze peptide. If the probability value is greater than 0.5, it is an antifreeze protein, and if the probability value is less than or equal to, it is not). In order to reduce the probability of false positives, the predicted probability values ​​of each peptide were sorted from high to low.

[0041] (2) Toxicity filtering: For the constructed peptide library, the online toxicity prediction tool ToxinPred (https: / / webs.iiitd.edu.in / raghava / toxinpred / ) was used to predict the cytotoxicity of each peptide and only non-toxic peptides were retained.

[0042] (3) Solubility and stability filtering: The solubility and stability of non-toxic peptides were predicted using the online solubility prediction tools Camsol (https: / / www-cohsoftware.ch.cam.ac.uk / / index.php) and Peptide (https: / / peptide.bio / ), as well as the online stability prediction tool ProtParam (https: / / web.expasy.org / protparam / ). A Camsol value greater than 1 and a Peptide value greater than 70% were considered to indicate that the peptide had high solubility, while an Instability Index lower than 40 was considered to indicate that the peptide was stable.

[0043] Retain highly soluble and stable peptides, that is, peptides that simultaneously meet the following conditions: Campol value greater than 1, Peptide value greater than 70%, and Instability Index less than 40.

[0044] Finally, two optimal candidate antifreeze peptides were screened and synthesized by Sangon Biotech (Shanghai) Co., Ltd.: Peptide11: DKLATWTPGM (SEQ ID NO.1) and Peptide6: AAIDALKSTM (SEQ ID NO.2).

[0045] Example 3 Purity testing: The quality of the synthesized antifreeze peptides was verified using a linear quadrupole ion trap (LTQ) liquid chromatography-mass spectrometry (LC-MS / MS). The results were obtained from chromatograms (e.g., [image of chromatogram]). Figure 2As shown in the figure, the main peak elution retention time of sample Peptide11 is 11.328 min, and only trace impurity peaks are present at 9.553 min, 9.945 min, 10.275 min, 10.541 min, 12.034 min, 12.416 min, and 12.901 min, with no obvious impurity peaks. Sample Peptide6 shows a sharp and symmetrical strong absorption main peak at 8.567 min, with only one small peak of extremely low impurity content at 7.260 min, and no obvious impurity absorption signal in the other elution intervals. Calculations using the peak area normalization method show that the retention time of Peptide11 is 11.328 min, and the peak area of ​​the target main peak accounts for ≥98% of the total peak area. The retention time of Peptide6 is 8.567 min, and the peak area of ​​the target main peak accounts for ≥98% of the total integral area. This proves that the purity of the prepared Peptide11 and Peptide6 meets the requirements for activity testing and practical applications.

[0046] Example 4 In vitro wet assay activity verification.

[0047] Determination of single ice crystal growth inhibition activity.

[0048] Take 2.5 μL of peptide 6 and peptide 11 samples and phosphate buffer (PBS group) and place them on the stage; lower the temperature to -40℃ at 50℃ / min and hold for 5 min; warm up to -10℃ at 10℃ / min and hold for 5 min; warm up to -5℃ at 5℃ / min and hold for 5 min; warm up to 1.3℃ at 0.5℃ / min and then lower the temperature to 0.3℃, and measure the size of single ice crystals within 30 s.

[0049] The results are as follows Figure 3 As shown, at the three observation times of 10s, 20s, and 30s, the area of ​​a single ice crystal in the peptide-treated group was smaller than that in the PBS control group. At 30s of isothermal treatment, the average area of ​​a single ice crystal in each group was as follows: PBS group: 111432.63 μm. 2 Peptide 6 group has a value of 53943.76 μm. 2 Peptide 11 group has a particle size of 26384.03 μm. 2 The increase in single ice crystal area from 10 min to 30 min is as follows: PBS group: 108152.94 μm 2 Peptide group 6 has a value of 50591.52 μm. 2 Peptide 11 group has a size of 23163.62 μm. 2 The results showed that both peptides possessed single ice crystal growth inhibitory activity.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for screening antifreeze proteins based on a protein language model, characterized in that, Includes the following steps: S1. Integrate antifreeze protein sequences as positive samples, remove redundant sequences to obtain positive samples, extract non-antifreeze protein sequences as negative samples in a 1:1 ratio of positive samples to negative samples, construct a comprehensive dataset, and split the positive samples and negative samples into training datasets and independent test datasets. S2. Multiple candidate antifreeze proteins are converted into feature vectors using a protein language model, and K-mer component features and amino acid physicochemical scoring features are simultaneously fused to construct the input feature vector. S3. Based on K-mer component characteristics and amino acid physicochemical scoring characteristics, a screening interval was set to perform initial screening of multiple candidate antifreeze proteins and obtain a preliminary screening candidate peptide library. S4. The candidate peptide library screened in the initial screening is subjected to antifreeze activity prediction, cytotoxicity filtering, solubility filtering, and stability filtering in sequence to finally screen and obtain antifreeze proteins.

2. The method for screening antifreeze proteins according to claim 1, characterized in that, In step S1, the antifreeze protein sequences include the AFP481, AFP920, and AFP604 datasets; the positive and negative samples are split into the training dataset and the independent test dataset in an 8:2 ratio.

3. The method for screening antifreeze proteins according to claim 1, characterized in that, In step S2, the protein language model is the ESM-AA-35M protein language model; the process of extracting the sequence feature vector of the protein language model is as follows: input a protein sequence of length L, and after average pooling, convert it into a 1×480-dimensional sequence feature vector; the K-mer component features and amino acid physicochemical scoring features include hydrophobicity coefficient score, solubility free energy score, isoelectric point score, and stability contribution score.

4. The method for screening antifreeze proteins according to claim 1, characterized in that, In step S3, the K-mer component characteristics and amino acid physicochemical scoring characteristics are the sum of the scores of the K-mer component characteristics and amino acid physicochemical scoring characteristics corresponding to the amino acids in each candidate antifreeze protein; the screening interval is set with the mean of the K-mer component characteristics and amino acid physicochemical scoring characteristics as the center and 0.5 standard deviations as the radius.

5. The method for screening antifreeze proteins according to claim 1, characterized in that, In step S4, the antifreeze activity prediction is performed by using the XGBoost algorithm to test the antifreeze probability of the initial screening candidate peptide library. Peptides with an antifreeze probability higher than the screening range are considered antifreeze proteins, while those lower than the screening range are considered non-antifreeze proteins. The cytotoxicity filtering is performed using the ToxinPred tool to predict and retain non-toxic peptides. The solubility filtering is performed using a combination of Camsol and Peptide tools to predict and retain peptides with a Camsol value greater than 1 and a Peptide predicted solubility greater than 70%. The stability filtering is performed using the ProtParam tool to predict and retain peptides with an instability index lower than 40.

6. An antifreeze protein, characterized in that, The amino acid sequence of the antifreeze protein is shown in SEQ ID NO.1 or SEQ ID NO.

2.

7. The antifreeze protein according to claim 6, characterized in that, The antifreeze protein has single ice crystal growth inhibition activity.

8. The application of the antifreeze protein as described in claim 6 in cryopreservation of biological samples, temporary storage for organ transplantation, or cryomedical surgery.

Images