Method for predicting cytosine deaminase function based on machine learning model and application

Abstract

The invention discloses a method for predicting cytosine deaminase functions based on a machine learning model, the model and application, and belongs to the technical field of molecular biology. According to the method, a high-quality data set containing experimental data of 1 to 100 cytosine deaminase is constructed, and an AlphaCD model is trained in combination with multi-dimensional features, so that high-precision and high-throughput function prediction of the uncharacterized cytosine deaminase is realized. According to the method, the catalytic characteristics of 21 and 345 CDs in a Uniprot database are successfully predicted, and the prediction accuracy is verified through experiments. In addition, the off-target effect of candidate CDs is optimized by using an alanine scanning technology, and an efficient and low-off-target cytosine base editor is developed. The method not only improves the efficiency of protein function characterization, but also provides powerful support for the biotechnical fields of gene editing and the like, and shows a wide application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of molecular biology, and specifically relates to a method and application for predicting the function of cytosine deaminase based on a machine learning model. Background Art

[0002] The rapid development of genomics has made it possible to explore a large number of new proteins, bringing unprecedented opportunities and new challenges for characterizing protein functions in organisms. Although traditional biochemical experimental methods occupy an important position in protein function research, their limitations of being time-consuming, laborious, and difficult to screen efficiently on a large scale are becoming increasingly prominent. The primary structure of a protein (i.e., amino acid sequence) is the basis for predicting protein function. Protein sequence analysis can identify functional motifs and conserved domains, while sequence alignment can reveal homologous proteins with potential functional information. The development of structure prediction tools such as AlphaFold2 (AF2) has made it possible to classify a large number of proteins with potential similar functions using clustering methods based on structural (rather than sequence) homology. However, these methods are still insufficient when dealing with complex, multifunctional proteins and cannot fully evaluate their functional properties. Therefore, we urgently need high-precision, quantitative methods to effectively characterize protein functions to overcome the limitations of current methods.

[0003] As a powerful modeling framework, machine learning (ML) can integrate feature information from multiple dimensions to learn and predict protein functions. Through cross-dimensional feature learning, machine learning may revolutionize the identification of known molecular relationships, three-dimensional (3D) structures, or other specific variable features that determine protein function, and is expected to replace traditional univariate or monotonic clustering methods to improve the accuracy and efficiency of predictions. Currently, machine learning models based on short peptide sequence datasets have been used to discover new antibiotics and predict antibacterial activity, hemolytic activity, and antibiotic resistance. However, current machine learning models based on short peptide sequence datasets still face challenges in accuracy and efficiency in predicting the functions of complex structural proteins.

[0004] Standard datasets play a key role in advancing artificial intelligence (AI) capabilities, especially machine learning and deep learning models, as they provide benchmarks for the development, training, and evaluation of predictive models. Despite the critical importance of these datasets, the preparation of large-scale, high-quality experimental datasets faces many challenges, especially in predicting protein function. Currently, the lack of comprehensive and validated datasets restricts the application of machine learning methods, resulting in limited accuracy and stability of prediction results, which is insufficient for research goals that require accurate and quantitative characterization of unknown proteins. Therefore, bridging this gap by developing robust experimental datasets is crucial to enhance the precision of machine learning-driven methods and improve our understanding of the breadth, reliability, and depth (i.e., level of detail) of protein function. Summary of the invention

[0005] In view of the problem that the current protein function prediction methods are limited in prediction accuracy and stability due to incomplete data sets, and cannot achieve high-precision and high-throughput functional characterization of a large number of unknown proteins, the present invention aims to provide a method, model and application for predicting cytosine deaminase function based on machine learning.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: The present invention provides a method for predicting cytosine deaminase function based on a machine learning model, comprising: S1, based on the Uniprot database, screened and synthesized cytosine deaminase and constructed a multidimensional feature dataset, which included catalytic properties, amino acid sequence properties, three-dimensional structural properties, and related biophysical and chemical properties; S2, cluster analysis is performed based on the multidimensional feature data set constructed in S1 to extract features that contribute significantly to the prediction of cytosine deaminase function; S3, using machine learning algorithms, combines the features extracted by S2 with the data set of multi-dimensional features obtained by S1 to build a machine learning model; S4, using the machine learning model constructed in S3, performs functional prediction on uncharacterized cytosine deaminases, and screens out cytosine deaminases for constructing efficient and high-fidelity cytosine base editors based on the prediction results.

[0007] The cluster analysis includes sequence-based cluster analysis and structure-based cluster analysis; wherein the sequence-based cluster analysis uses CD-HIT and MMseqs tools, and the structure-based cluster analysis uses Foldseek and TM-Cluster tools.

[0008] The extracted features that contribute significantly to the prediction of CDase function include protein length, sequence- and structure-based clustering information, subfamily and physicochemical properties.

[0009] The machine learning algorithm is any one of a random forest, a support vector machine, a neural network, a gradient boosting tree, and a deep learning algorithm.

[0010] Furthermore, the machine learning algorithm is random forest.

[0011] The functional prediction includes catalytic efficiency, off-target catalytic activity, catalytic site window and catalytic motif preference.

[0012] The present invention provides a cytosine deaminase function prediction model, wherein the cytosine deaminase function prediction model is obtained by the above-mentioned method for predicting cytosine deaminase function based on a machine learning model.

[0013] The present invention provides an application of the above-mentioned cytosine deaminase function prediction model in screening proteins or their functional conservative variants for constructing cytosine base editors, wherein the proteins or their functional conservative variants serve as cytosine deaminase.

[0014] Furthermore, the cytosine base editor is used for C-to-T gene editing.

[0015] The protein or its functional conservative variant is selected from: CD0029, CD0276, CD0718, CD0376, CD0183, CD0760, CD0628, CD0311, CD0991, CD0503, CD0093, CD0791, CD0336, CD0317, CD0288, CD0702, CD0191, CD1101, CD0796, CD0608, CD0392, CD0759, CD0289, CD0752, CD0409, CD1008, CD0798, CD1049, CD0596, CD0506, CD0290, CD0537, CD0027, CD 0723, CD0529, CD1104, CD0342, CD0222, CD0142, CD1107, CD0509, CD0558, CD0655, CD0691, CD0082, CD0193, CD0393, CD0974, CD0746, CD0382, CD0981, CD0685, CD0813, CD1019, CD0711, CD0751, CD1022, CD0595, CD0256, CD0749, CD0562, CD0452, CD0354, CD0479, CD0357, CD0459, CD0481, CD0757, CD0155 , CD0225, CD0132, CD0178, CD0389, CD0223, CD0294, CD0934, CD0473, CD0128, CD0293, CD1058, CD0549, CD0347, CD0065, CD0805, CD0535, CD0237, CD08 92. CD1097, CD0390, CD0896, CD0546, CD1189, CD0461, CD0450, CD0381, CD0196, CD0491, CD1054, CD0545, CD0877, CD0425, CD1094, CD1105, CD1114, CD0 419, CD1128, CD1109, CD0420, CD0322, CD1193, CD0362, CD0804, CD0571, CD1112, CD0350, CD0973, CD0766, CD0527, CD0582, CD0933, CD0371, CD0488, C D0334, CD1031, CD1181, CD0335, CD1111, CD0489, CD0669, CD0076, CD0982, CD0455, CD0531, CD0446, CD0893, CD0528, CD0525, CD0245, CD0230, CD0255,CD1131、CD0602、CD0735、CD0738、CD0176、CD0635、CD0241、CD0331、CD0395、CD0150、CD0319、CD0659、CD0465、CD0557、CD1135、CD0072、CD1174、CD0295、CD0895、CD0668、CD0677、CD0063、CD1199、CD0098、CD0249、CD0433、CD0404、CD1172、CD0797、CD1006、CD0663、CD0876、CD1092、CD1059、CD0875、CD0819、CD0976、CD0830、CD0464、CD0402、CD0753、CD1195、CD0166、CD1082、CD0476、CD0271、CD0449、CD0325、CD0138、CD0352、CD0470、CD1002、CD0917、CD0431、CD0316、CD0188、CD0364、CD0576、CD0179、CD0739、CD0706、CD0514、CD1034、CD1093、CD0227、CD0365、CD0587、CD0720、CD0605、CD0520、CD1184、CD0505、CD0384、CD0838、CD0388、CD0707、CD0953、CD1043、CD0740、CD0284、CD0667、CD0229、CD0927、CD1084、CD0951、CD0560、CD0742、CD1042、CD0656、CD0463、CD0010、CD0594、CD0729、CD0915、CD0761、CD0883、CD1134、CD1005、CD0474、CD0888、CD0002、CD0521、CD0764、CD0682、CD0172、CD0658、CD1183、CD1028、CD1190、CD0101、CD0156、CD1137、CD0422、CD1186、CD0715、CD1017、CD0019、CD0168、CD0928、CD0043、CD0344、CD1188、CD0724、CD1064、CD0209、CD1194、CD1113、CD0462、CD0105、CD0621、CD0552、CD0307、CD0997、CD0840、A0A0K0MIC1、A0A455BCL3、A0A0K0MJ52、A0A671DNI5、A0A2R2Z4E4、A0A6P6C7T7, P31941, G8GPV4, P51908, A0A2R2X2I5, A0A1U7Q2F3, A0A4W2FT84, A0A803TEP0, G1LKJ4, A0A7J8FW00, A0A2K6MSA4, K7BTL7, A0A822EZF5, F7F6M6, Q8NFB8, A0A6P5KZH5, A0A4W2HPD5, Q4VB33, A0A834EE11, G1RYZ5, A0A7X8SJJ0, A7VGG7, A0A4Q9H5D8, A0A5J4P2S6, or A0A1J5GBA4.

[0016] The protein sequences are shown in SEQ ID NO.25 to SEQ ID NO.328.

[0017] The protein or its functional conservative variant is a functional conservative variant of A0A2R2Z4E4, selected from: H56A, N45A, M1A, I82A, R54A, Y114A, P86A, R112A, F6A, L124A or E100A.

[0018] The functional conservative variant sequences of A0A2R2Z4E4 are shown in SEQ ID NO.329 to SEQ ID NO.339.

[0019] The cytosine deaminase function prediction model is used in screening for the preparation of reagents or kits for C-to-T gene editing, wherein the reagents or kits contain at least one of the above-mentioned applied proteins or their functional conservative variants.

[0020] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for predicting the function of cytosine deaminase based on a machine learning model. For the first time, machine learning is applied to the prediction of the function of cytosine deaminase. By constructing a multidimensional feature data set and using an advanced machine learning algorithm for model training, the accurate prediction of the function of cytosine deaminase is achieved, breaking the limitations of traditional experimental methods, improving the prediction efficiency and accuracy, and contributing to the development of more types of cytosine base editors, thereby expanding the scope of application of gene editing. By screening and synthesizing cytosine deaminase from the Uniprot database and constructing an expression vector of a cytosine base editor, a multidimensional feature data set containing catalytic properties, amino acid sequences, three-dimensional structures, and related biophysical and chemical properties can be obtained, providing rich information for subsequent machine learning models, which helps to improve the accuracy and efficiency of predictions; using a machine learning algorithm, combined with the extracted features and multidimensional feature data sets, a machine learning model is constructed that can automatically learn the complex relationship between the function of cytosine deaminase and its features, thereby achieving accurate prediction of the function of uncharacterized cytosine deaminase; the screened cytosine deaminase can be used to construct an efficient and high-fidelity cytosine base editor, providing powerful tool support for the field of gene editing.

[0021] Furthermore, by constructing a data set containing multi-dimensional features and combining it with a machine learning algorithm, the AlphaCD model showed high accuracy in predicting catalytic efficiency, off-target catalytic activity, etc. (e.g., catalytic efficiency prediction R²=0.92, off-target effect prediction R²=0.84).

[0022] Furthermore, through sequence- and structure-based clustering analysis, features that significantly contribute to the prediction of cytosine deaminase function can be extracted and used as input to the machine learning model, which helps to improve the accuracy of the prediction.

[0023] The cytosine deaminase function prediction model provided by the present invention has high efficiency, accuracy, generalization ability, robustness, and broad application prospects and scientific research value, and is an important technological innovation in the field of biotechnology. The prediction model is based on a comprehensive and multi-dimensional cytosine deaminase feature data set, which provides a rich and accurate information basis for the model; the constructed model can automatically learn the complex relationship between the cytosine deaminase function and its characteristics, thereby achieving accurate prediction of the uncharacterized cytosine deaminase function; the model can greatly shorten the time for cytosine deaminase function prediction, improve prediction efficiency, and provide fast and reliable decision support for scientific research and practical applications; the model can screen out cytosine deaminases with specific catalytic properties, provide strong support for the construction of efficient and high-fidelity cytosine base editors, and promote the development and application of gene editing technology.

[0024] The application provided by the present invention utilizes the constructed cytosine deaminase function prediction model to efficiently and accurately screen out cytosine deaminases with specific catalytic properties from a large number of candidate proteins. Compared with traditional experimental screening, this screening method greatly shortens the screening cycle and improves the screening efficiency; a variety of cytosine deaminase candidate proteins including CD0029, CD0276, CD0718, etc., and functional conservative variants of A0A2R2Z4E4 are obtained, providing rich choices for constructing cytosine base editors; the cytosine deaminase or its functional conservative variant obtained by screening has specific catalytic properties, can more accurately identify and edit target gene sites, thereby improving the fidelity of the cytosine base editor; the screened cytosine deaminase performs well in catalytic efficiency, off-target catalytic activity, etc., can significantly improve the editing efficiency of the cytosine base editor, and provide strong support for gene editing experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The diversity characterization of CD catalytic properties used for training the machine learning-based functional prediction model in the present invention, wherein a is the workflow for high-throughput screening of CD catalytic properties of 102 sgRNA-target sequence libraries expressed in HEK293T cells; b is the phylogenetic distribution of 1,100 CDs across 11 APOBEC-like protein subfamilies; c is a heat map of the average catalytic efficiency of 1,100 CDs at 102 sgRNA-target sites, where the X-axis represents the 102 sgRNA-target sites, and the Y-axis represents the distribution of 1,100 CDs in the APOBEC protein subfamilies, with red Indicates high efficiency, blue indicates low efficiency; d is the average catalytic efficiency distribution of 1,100 CDs in 102 sgRNA-target sites; e is a scatter plot between the catalytic efficiency of 298 CDs (efficiency>10%) in 102 sgRNA-target sites in HEK293T cells and the off-target activity of four R-loop sites; f is the catalytic window distribution of 322 CDs (catalytic efficiency>10%) in 102 sgRNA-target sites; g is the motif preference distribution of 322 CDs (catalytic efficiency>10%) in 102 sgRNA-target sites; Figure 2 The catalytic activities of the 322 deaminases of the present invention are shown in Figure 2. The C-to-T catalytic efficiencies of the 322 deaminases (efficiency > 10%) were evaluated by deep sequencing of 102 sgRNA-target sequence libraries. The center line represents the median. The upper and lower edges of the box represent the first and third quartiles of the editing efficiency of the 102 sites, respectively, and the tails extend to the minimum and maximum values. Figure 3The catalytic specificity of the 298 deaminases of the present invention is the ratio of the catalytic efficiency to the off-target effect of the 298 deaminases with an efficiency greater than 10%; Figure 4 The catalytic windows of 322 deaminases in the present invention were evaluated by deep sequencing of 102 sgRNA-target sites. The data represent the average of four independent experiments, and the gradient from white to blue represents the catalytic efficiency from low to high, respectively.

[0026] Figure 5 is the preferential catalytic motif of 322 deaminases of the present invention, and the vertical axis represents the catalytic efficiency distribution of each deaminase under four sequence environments (AC, CC, GC and TC) in 102 sgRNA-target sites; Figure 6 The clustering effect of APOBEC family deaminases based on sequence and structural similarity in the present invention is used to compare the similarity of deaminases within and between clusters; wherein a is CD-HIT, b is MMSeq, c is Foldseek, and d is TM-Cluster. The structural similarity is calculated by comparing deaminases within the same cluster (red) or between different clusters (blue). The horizontal line represents the average value of each group, and the numerical value represents the mean ± standard deviation (SD) of the structural similarity between different deaminases. The P value is calculated by a two-tailed unpaired t-test. eh is the distribution of catalytic activity within the cluster; e is CD-HIT, f is MMSeqs, g is Foldseek, and h is TM-Cluster. The X-axis represents the cluster number; and the Y-axis represents the catalytic efficiency (%).

[0027] Figure 7 Identification of features that potentially affect catalytic efficiency, off-target effect, catalytic window and preferential catalytic motifs of the present invention, wherein a is catalytic efficiency, b is off-target effect, c is catalytic window, and d is correlation analysis between preferential catalytic motifs; Figure 8The model constructed based on machine learning (ML) in this invention is used to characterize 21,345 CDs. a is the ML workflow for predicting protein properties (including catalytic efficiency, off-target effects, catalytic windows, and preferential catalytic motifs); b is the SHAP summary diagram (left) showing the contribution of each feature to the model training process, with the SHAP value shown on the X-axis; blue represents low and red represents high, and the feature importance bar chart (right) shows the average absolute SHAP value of each feature, indicating their overall impact on model prediction; c is a comparison of predicted and observed catalytic efficiency, each point represents a sample; the size of the point corresponds to the catalytic efficiency; d is a comparison of predicted and observed off-target effects, each point represents a samples, the size of the dots indicates off-target activity; e is the ROC curve analysis of the catalytic window prediction model (the gray line indicates different window categories; the blue line indicates the average of all categories); f is the ROC curve analysis of the motif prediction model (the gray line indicates different motif categories; the blue line indicates the average of all categories); g is a 3D scatter plot of predicted catalytic efficiency, catalytic window, and preferential catalytic motifs, and the color indicates off-target catalytic activity (blue indicates low, red indicates high); h is the Pearson correlation analysis of predicted and observed catalytic efficiencies, R value = 0.86; significance is determined by P value < 0.05; Fig. 9 The AlphaCD of the present invention is used to construct and verify the CD function. a is the construction and operation process of AlphaCD. b is the SHAP summary diagram (left) showing the contribution of each feature to the off-target model training process. The SHAP value is shown on the X-axis; blue indicates low and red indicates high. The feature importance bar graph (right) shows the average absolute SHAP value of each feature, indicating their overall impact on the off-target model prediction. c and d are feature importance bar graphs of feature contributions during model training. c is the catalytic window, d is the catalytic motif, and e is the accuracy of AlphaFold2's prediction of CD structures in the prediction database; f is a scatter plot comparing the predicted catalytic efficiency of 21,345 CDs at 102 sgRNA-target sites with the off-target activity (efficiency > 10%) of four R-loop sites in HEK293T cells; g is the catalytic window distribution of 21,345 CDs (catalytic efficiency > 10%) in 102 sgRNA-target sites; h is the motif preference distribution of 21,345 CDs (catalytic efficiency > 10%) in 102 sgRNA-target sites; i is the correlation between predicted and observed off-target effects, and the R value and P value are calculated by Pearson's correlation test; j is the correlation between predicted and observed catalytic windows, using Cramér's V method evaluation; k is the correlation between the predicted and observed preferred catalytic motifs, and accuracy is calculated as the ratio of the number of accurate predictions to the total number of predictions.

[0028] Fig.10 The present invention is an efficient cytosine deaminase with reduced off-target effects through engineering design, wherein a is an experimental process for reducing the off-target effects of the candidate efficient deaminase A0A2R2Z4E4 by alanine scanning mutagenesis; b is the catalytic efficiency and sgRNA-independent off-target effects of the A0A2R2Z4E4 variants at different positions, and the catalytic efficiency is evaluated by 102 sgRNA-target sequence libraries; the sgRNA-independent off-target effect is evaluated by R-loop experiment, and each point represents the off-target effect of 102 sgRNA-target sequence libraries. get site or the average off-target effect on the four R-loop off-target sites; c is the catalytic efficiency and off-target effect of the A0A2R2Z4E4 variant, each point represents a pairwise comparison of the average catalytic efficiency (Y-axis) at 102 sgRNA-target sites and the average off-target effect (X-axis) at the four R-loop off-target sites for each A0A2R2Z4E4 variant relative to rAPOBEC1; d is the ratio of the catalytic efficiency to the off-target effect of the A0A2R2Z4E4 variant in c.

[0029] Fig.11 The catalytic efficiency, catalytic specificity and preferential catalytic motif of the A0A2R2Z4E4 variant of the present invention, wherein a is the catalytic efficiency of the A0A2R2Z4E4 variant, and the C-to-T catalytic efficiency of the A0A2R2Z4E4 variant was evaluated by deep sequencing of 102 sgRNA-target sites. The center line represents the median; the upper and lower edges of the box represent the first and third quartiles of the editing efficiency of the 102 sgRNA-target sites, respectively. The tail extends to the minimum and maximum values; b is the catalytic specificity of the A0A2R2Z4E4 variant. The ratio statistics of the catalytic efficiency and off-target effect of the A0A2R2Z4E4 variant; c is the preferential catalytic motif of the A0A2R2Z4E4 variant. The vertical axis represents the distribution of the catalytic efficiency of the deaminase in four sequence environments (AC, CC, GC and TC) in the 102 sgRNA-target sites.

[0030] Fig.12 The catalytic window of the A0A2R2Z4E4 variant of the present invention was evaluated by deep sequencing of 102 sgRNA-target sites. The data represent the average of four independent experiments, and the gradient from white to blue represents the catalytic efficiency from low to high, respectively. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0032] Example 1 This embodiment provides a method for predicting cytosine deaminase function based on a machine learning model, which specifically includes the following steps: 1. Data preparation To construct the training dataset, we obtained 1483 CDs containing at least one APOBEC-like N-terminal (Pfam ID: PF08210) or APOBEC-like C-terminal (Pfam ID: PF05240) domain from the Pfam database; 375 CDs lacking experimental data and 8 non-APOBEC family variants were excluded, and finally 1100 experimentally verified CDs were obtained for model training. To construct the test dataset, the Uniprot database (www.uniprot.org) was queried using "APOBEC" as a keyword, and 144,632 sequences were downloaded. The obtained sequences covered 11 protein subfamilies: APOBEC1, APOBEC2, APOBEC3A, APOBEC3B, APOBEC3C, APOBEC3DE, APOBEC3F, APOBEC3G, APOBEC3H, APOBEC4, and AID. Multiple sequence alignments (MSA) were generated for each family using the --auto option of MAFFT (v7.490). The hmmscan function in HMMER (3.1b2) was used to annotate the training and prediction data sets, and 30,457 sequences were successfully annotated with family information, of which 23,436 protein PDB files were downloaded from the AlphaFold protein structure database; after filtering out sequences lacking methionine start codons or stop codons, 21,345 sequences were finally retained.

[0033] 2. Clustering method for CD sequence data mining For sequence-based clustering, CD-HIT (v4.8.1) was used with a sequence similarity threshold of 60% and a coverage threshold of 60%; MMseqs (v15.6) was used with a sequence similarity threshold of 40% and a coverage threshold of 40%.

[0034] For structure-based clustering, Foldseek (v8.ef4e960) and TM-Cluster (v1) were used with 90% coverage and a threshold of 0.7, and the TM scores within and between clusters were calculated by TM-align (20190822).

[0035] 3. Feature preparation To generate the training set, sequence-based clustering was performed using CD-HIT (v4.8.1) in different subfamilies, with a sequence similarity threshold of 70% and 40%. Clustering based on three-dimensional structure was performed using TM-Cluster with a threshold of 0.7; protein features such as protein length, molecular weight, isoelectric point, instability index, and hydrophobicity (GRAVY) were calculated using the Bio.SeqUtils.ProtParam function in Biopython (v1.83); to generate the prediction set, unknown APOBEC family deaminases were assigned to CD-HIT and TM-Cluster groups based on the highest sequence and structural similarity, respectively, using Blast (v2.15.0) and Foldseek (v8.ef4e960); physicochemical properties were tested as described above.

[0036] 4. Correlation Analysis Spearman (ρ) correlation was used to analyze potential relationships between continuous variables (catalytic efficiency, off-target effects, protein length, molecular weight, isoelectric point, instability index, and GRAVY). Eta square (η²) correlation was used to test the correlation between continuous variables and categorical variables (clusters, families). Cramér's V (φc) correlation coefficient was used to evaluate the relationship between catalytic window, motifs, and categorical variables.

[0037] 5. Random Forest The features selected for model training included protein length, protein family, CD-HIT clustering, TM-Cluster clustering, isoelectric point, instability index and GRAVY; the RandomForestRegressor function in scikit-learn (v1.4.2) was used to build regression models for catalytic efficiency and off-target catalytic activity, and the RandomForestClassifier was used to build classification models for catalytic window and motif; the regression model data was sampled and binned using pandas (v2.1.4), and the classification model data was oversampled using the synthetic minority oversampling technique (SMOTE) to improve the accuracy of the model; RandomizedSearchCV was used for hyperparameter tuning and initial model evaluation, and a ten-fold cross-validation strategy was adopted. The optimal hyperparameters for each model were determined using GridSearchCV; model evaluation indicators included precision, accuracy, recall, F1 score, receiver operating characteristic (ROC) curve and area under the curve (AUC); SHAP value was calculated using SHAP (v0.45.1) to evaluate the contribution of each feature to the prediction of the regression model. Feature importance is used to evaluate the contribution of each feature to the prediction of the classification model.

[0038] 6. Plasmid Construction The CD sequences downloaded from UniProt were codon-optimized for human codons using GenSmart™ Codon Optimization (https: / / www.genscript.com.cn / gensmart-free-gene-codon-optimization.html); the optimized sequences were synthesized by the Beijing Genomics Institute and subsequently cloned into the pCMV-nSpCas9-CBE vector backbone via EcoR I and Hind III restriction sites; the oligonucleotides used to construct sgRNA were synthesized by GENEWIZ and cloned into the U6-sgRNA vector backbone after BpiI digestion.

[0039] 7. Cell Culture and Transfection The wild-type HEK293T cell line and the HEK293T cell line stably expressing the 102 sgRNA-target sequence library used in the present invention were constructed and maintained by the applicant's laboratory; the culture medium of all HEK293T cells was DMEM supplemented with 10% fetal bovine serum (Biological Industries, 04-001-1ACS) and 1% penicillin-streptomycin (Beyotime, C0224-100mL); the cells were incubated at 37°C, 5% CO 2 Incubator. 2For the sgRNA-target library experiment, 1×10 6 Cells were seeded in a 6-well plate, and 6 μg of the plasmid expressing the base editor was transfected into the cells using 12 μL of PEI. 24 hours after transfection, the culture medium was replaced with a medium containing 3 μg / mL puromycin (Invivogen, ant-pr-1), cultured for 96 hours, and positive cells were screened. The surviving cells were collected and used for genomic DNA extraction. For the orthogonal R-loop experiment, 2×10 5 Wild-type HEK293T cells were seeded in a 24-well plate, and 0.75 μg of dSaCas9-sgRNA plasmid and 0.75 μg of base editor plasmid were transfected into the cells using 3 μL of PEI. 24 hours after transfection, the cells were cultured in a medium containing 3 μg / mL puromycin for 96 hours to screen positive cells. The surviving cells were cells that were successfully co-transfected and expressed dSaCas9-sgRNA and base editors, and were used for genomic DNA extraction.

[0040] 8. Orthogonal R-loop Experiment In the present invention, orthogonal R-loop experiments were used to detect the sgRNA-independent off-target effects of base editors carried by different deaminases in DNA. Specifically, we selected four commonly used dSaCas9-sgRNA target sites (Sasite 3, Sa site 4, Sa site 5, and Sa site 6) as potential sgRNA-independent DNA off-target sites; dSaCas9-sgRNA plasmids and base editor plasmids of each deaminase were co-transfected into HEK293T cells using PEI; plasmid-positive transfected cells were enriched in a culture medium containing puromycin, and genomic DNA was extracted from cells that survived for 4 days; four potential off-target sites were PCR amplified using specific primers with barcodes (see the table below for a list of all primers used in the present invention), and off-target efficiency was detected by high-throughput sequencing.

[0041] Table 1: Primers for four potential off-target sites

[0042] 9. PCR amplification and deep sequencing To sequence HEK293T cells with 102 sgRNA-target libraries treated with different base editors (CBEs), genomic DNA (gDNA) was extracted using the Tiangen Tissue Cell Genomic DNA Extraction Kit (Tiangen, DP304-03) according to the kit instructions; for orthogonal R-loop experiments, HEK293T cells selected with puromycin were collected and gDNA was extracted using 30 μL lysis buffer (Vazyme, P073) according to the kit instructions; gDNA was used as a template for PCR amplification using TakaraEx Taq polymerase (Takara, RR902A); in the second round of PCR, specific barcode sequences were added to the primers to distinguish samples; PCR products were purified using the Universal DNA Purification Kit (Tiangen, DP219) according to the manufacturer's instructions; the purified DNA was sequenced by 150 bp paired-end high-throughput sequencing.

[0043] Table 2: 102 sgRNA-target primers

[0044] 10. Bioinformatics Analysis Next-generation sequencing data were segmented using fastq-multx (v1.4.2), and gene editing results were calculated using CRISPResso2 (v2.2.7). Phylogenetic trees were constructed using the ggtree (v3.10.1) script in R (v4.3.1). Images in the article were drawn using R (v4.3.1), Python (v3.11.7), and GraphPad Prism (v.8.2.1). All values ​​are expressed as mean ± standard deviation (SD). Two-tailed unpaired Student's t test was used to compare differences between groups of sequence information, and P < 0.05 was considered statistically significant.

[0045] The protein or its functional conservative variant screened by the present invention is selected from: CD0029, CD0276, CD0718, CD0376, CD0183, CD0760, CD0628, CD0311, CD0991, CD0503, CD0093, CD0791, CD0336, CD0317, CD0288, CD0702, CD0191, CD1101, CD0796, CD0608, CD0392, CD0759, CD0289, CD0752, CD0409, CD1008, CD0798, CD1049, CD0596, CD0506, CD0290, CD0537, CD 0027, CD0723, CD0529, CD1104, CD0342, CD0222, CD0142, CD1107, CD0509, CD0558, CD0655, CD0691, CD0082, CD0193, CD0393, CD0974, CD0746, CD0382, CD0981, CD0685, CD0813, CD1019, CD0711, CD0751, CD1022, CD0595, CD0256, CD0749, CD0562, CD0452, CD0354, CD0479, CD0357, CD0459, CD0481, CD0757 , CD0155, CD0225, CD0132, CD0178, CD0389, CD0223, CD0294, CD0934, CD0473, CD0128, CD0293, CD1058, CD0549, CD0347, CD0065, CD0805, CD0535, CD02 37. CD0892, CD1097, CD0390, CD0896, CD0546, CD1189, CD0461, CD0450, CD0381, CD0196, CD0491, CD1054, CD0545, CD0877, CD0425, CD1094, CD1105, CD1 114. CD0419, CD1128, CD1109, CD0420, CD0322, CD1193, CD0362, CD0804, CD0571, CD1112, CD0350, CD0973, CD0766, CD0527, CD0582, CD0933, CD0371, C D0488, CD0334, CD1031, CD1181, CD0335, CD1111, CD0489, CD0669, CD0076, CD0982, CD0455, CD0531, CD0446, CD0893, CD0528, CD0525, CD0245, CD0230,CD0255、CD1131、CD0602、CD0735、CD0738、CD0176、CD0635、CD0241、CD0331、CD0395、CD0150、CD0319、CD0659、CD0465、CD0557、CD1135、CD0072、CD1174、CD0295、CD0895、CD0668、CD0677、CD0063、CD1199、CD0098、CD0249、CD0433、CD0404、CD1172、CD0797、CD1006、CD0663、CD0876、CD1092、CD1059、CD0875、CD0819、CD0976、CD0830、CD0464、CD0402、CD0753、CD1195、CD0166、CD1082、CD0476、CD0271、CD0449、CD0325、CD0138、CD0352、CD0470、CD1002、CD0917、CD0431、CD0316、CD0188、CD0364、CD0576、CD0179、CD0739、CD0706、CD0514、CD1034、CD1093、CD0227、CD0365、CD0587、CD0720、CD0605、CD0520、CD1184、CD0505、CD0384、CD0838、CD0388、CD0707、CD0953、CD1043、CD0740、CD0284、CD0667、CD0229、CD0927、CD1084、CD0951、CD0560、CD0742、CD1042、CD0656、CD0463、CD0010、CD0594、CD0729、CD0915、CD0761、CD0883、CD1134、CD1005、CD0474、CD0888、CD0002、CD0521、CD0764、CD0682、CD0172、CD0658、CD1183、CD1028、CD1190、CD0101、CD0156、CD1137、CD0422、CD1186、CD0715、CD1017、CD0019、CD0168、CD0928、CD0043、CD0344、CD1188、CD0724、CD1064、CD0209、CD1194、CD1113、CD0462、CD0105、CD0621、CD0552、CD0307、CD0997、CD0840、A0A0K0MIC1、A0A455BCL3、A0A0K0MJ52、A0A671DNI5、A0A2R2Z4E4、A0A6P6C7T7, P31941, G8GPV4, P51908, A0A2R2X2I5, A0A1U7Q2F3, A0A4W2FT84, A0A803TEP0, G1LKJ4, A0A7J8FW00, A0A2K6MSA4, K7BTL7, A0A822EZF5, F7F6M6, Q8NFB8, A0A6P5KZH5, A0A4W2HPD5, Q4VB33, A0A834EE11, G1RYZ5, A0A7X8SJJ0, A7VGG7, A0A4Q9H5D8, A0A5J4P2S6, or A0A1J5GBA4.

[0046] Example 2 Build and characterize a high-quality, annotated CD diversity dataset for training machine learning models.

[0047] See attached Figures 1 to 5 In order to construct a high-throughput machine learning model that can efficiently predict the catalytic properties of uncharacterized APOBEC family-like deaminases, 1,100 CDs were synthesized and experimentally tested to establish a training data set for catalytic properties. First, 1,483 homologous CD protein sequences with APOBEC-like N-terminal or C-terminal catalytic domains were screened from the Uniprot database for synthesis; 281 nucleoside sequences synthesized by the applicant in the early stage and 777 nucleoside sequences synthesized later were used to construct a putative catalytic base editor (CBE) ( Figure 1 a). In addition, we selected 42 well-characterized wild-type (WT) or mutant deaminases as controls, ultimately constituting a combined set of 1,100 sequences. These sequences cover 11 APOBEC-like protein subfamilies, including AID (16.91%), APOBEC1 (11.82%), APOBEC2 (22.55%), APOBEC3A (7.55%), APOBEC3B (14.82%), APOBEC3C (1.36%), APOBEC3DE (7.45%), APOBEC3F (2.09%), APOBEC3G (7.55%), APOBEC3H (7.64%), and APOBEC4 (0.27%) ( Figure 1 b).

[0048] After optimizing the sequences of these 1,100 CDs using human codons, each sequence was cloned into the pCMV-nSpCas9-CBE expression vector; all 1,100 constructs were transfected into HEK293T cells stably expressing the 102 sgRNA-target sequence libraries; 5 days after transfection, genomic DNA was extracted from the cells, and the sgRNA-target sequences were amplified by PCR for Illumina high-throughput sequencing to analyze the catalytic properties of each deaminase, such as catalytic efficiency, catalytic site window, and catalytic motif preference ( Figure 1 a); In addition, R-loop experiments were performed on 298 editors with catalytic efficiency exceeding 10% to study their sgRNA-independent off-target catalytic activity.

[0049] Sequencing analysis of 102 gRNA-target sequences revealed that 322 of the 1,100 CBEs had an average C-to-T catalytic efficiency of more than 10% in the 102 targets. Notably, 224 editors showed higher catalytic efficiency than the widely used high-fidelity deaminase YE1 CBE, 85 exceeded the rAPOBEC1 CBE, and 24 exceeded the hAPOBEC3A (hA3A) CBE, one of the most efficient editors currently available. It is also noteworthy that the deaminases with the highest catalytic efficiency were mainly from the APOBEC3A and APOBEC1 families, while almost all deaminases from the APOBEC2 family exhibited negligible catalytic efficiency ( Figure 1 cd, Figure 2 ). R-loop experimental analysis of sgRNA-independent off-target effects at four potential off-target sites for 298 editors with efficiencies exceeding 10% showed that off-target catalytic activity was significantly positively correlated with on-target catalytic efficiency (p value = 0.88) ( Figure 1 e). Analysis of the ratio of catalytic efficiency to off-target catalytic activity showed that the target specificity of the 298 deaminases varied greatly, with some deaminases (such as 9.6 for CD0419 and 5.9 for CD0509) showing higher fidelity than YE1 (3.3), while others introduced significantly higher off-target catalytic activity (such as 50.0% for CD0730 and 48.3% for CD0288) than WT hA3A (35.8%), which is known to have high off-target activity ( Figure 1 e and Figure 3 ).

[0050] Analysis of the catalytic windows of deaminase editors with catalytic efficiencies greater than 10% revealed diverse catalytic windows, including 1-4 / 5 / 6 / 7 / 8 / 9 / 10 / 11 / 14, 2-4 / 5 / 6 / 7 / 8 / 9 / 10 / 11 / 12, 3-5 / 6 / 7 / 8 / 9 / 10 / 11 / 12 / 13, 4-5 / 6 / 7 / 8 / 9 / 10 / 11 / 12, 5-5 and 9-11, compared to the 3-8 catalytic window of rAPOBEC1 CBE. In addition, several deaminases exhibited unusually narrow (e.g., 5-5 and 4-5) or broad (e.g., 1-11 / 14, 2-12 and 3-13) catalytic windows, which have not been reported before ( Figure 1 f, Figure 4 ). Analysis of the motif preferences of deaminase editors with catalytic efficiencies exceeding 10% showed that 58.43% of the editors preferentially introduced base edits at TC motifs, while 28.61% of the editors preferentially modified GC sites, which is highly consistent with the previously reported catalytic properties of the APOBEC family. In contrast, a lower proportion of deaminases preferentially edited CC and AC motifs (7.53% and 2.41%, respectively) ( Figure 1 g, Figure 5 The extensive and diverse deaminase catalytic property data included in this CBE panel will provide sufficient data support for training AI-based models to predict the functions of other newly discovered naturally occurring deaminases.

[0051] Example 3 In order to gain a deeper understanding of the relationship between protein sequence, three-dimensional structure, and other physicochemical properties and CD catalytic properties, various cluster analysis methods were used, and the correlation between various features and CD functional properties was examined. The following are the detailed analysis results: See attached Figure 6-7 To explore the relationship between protein sequence or three-dimensional structure and CD activity, 1,100 CDs were clustered using four different methods, including the sequence-based CD-HIT method and the MMseqs method, which generated 133 and 102 clusters, respectively, and the structure-based Foldseek method and the TM-Cluster method, which generated 85 and 203 clusters, respectively. To validate the subgroups obtained by these four methods, we compared the CD sequence and structural similarities within the same cluster, as well as the CD similarities between different clusters. The analysis showed that CDs within the same cluster showed almost uniformly high similarities (TM scores were 0.85, 0.86, 0.81, and 0.89, respectively), while the CD sequence / structure similarities between different clusters were significantly lower (0.64, 0.61, 0.64, and 0.58, respectively) ( Figure 6ad), which confirms that these clusters do reflect sequence and structural similarities. However, the clusters generated by TM-Cluster also contain CDs with widely different catalytic activities, ranging from >50% to <10%, and even include inactive candidates. In addition, the results of the other three methods also show that there is no consistent correlation between sequence or structure and catalytic efficiency ( Figure 6 This observation led us to conclude that structural or sequence similarity alone cannot reliably distinguish differences in protein function.

[0052] To identify other factors that may affect protein function, other features were examined, including protein length, subfamily, and physicochemical properties (molecular weight, isoelectric point, instability index, and grand average hydropathicity (GRAVY) score). These features were then tested for Spearman (ρ), squared eta (η²), and Cramer's V coefficient (φc) correlations with functional CD attributes, including catalytic efficiency, off-target catalytic activity, catalytic window, and preferred catalytic motifs (AC, CC, GC, and TC). Consistent with the hypothesis, η² correlation analysis showed that structure-based clusters (such as TM-Cluster) had the highest correlation with catalytic efficiency (η² = 0.52), while CD-HIT and MMseqs showed correlation coefficient η² values ​​of 0.41 and 0.45 between protein sequence and catalytic efficiency, respectively. In contrast, peptide chain length and molecular weight had the lowest correlation with catalytic efficiency, with ρ coefficients of -0.07 and -0.04, respectively. In addition, the isoelectric point (0.11) and GRAVY score (0.15) are also positively correlated with the catalytic efficiency to a certain extent, and the ρ value of the instability index and catalytic efficiency is 0.28, which is close to the correlation between catalytic efficiency and subfamily (0.30) ( Figure 7 a). Although the above positive correlations were detected, none of them were strong enough to distinguish catalytic efficiency or other functional properties.

[0053] The study found that the correlation between protein structure and catalytic efficiency was high, especially in TM-Cluster classification (φc = 0.52), while the correlation between structure clusters and catalytic efficiency was weak in Foldseek method (φc = 0.39). The correlation strength between clusters obtained by CD-HIT and MMseqs methods and catalytic efficiency was similar (φc 0.43 and 0.45, respectively) ( Figure 7a). These results, together with simpler input and better usability, motivated us to include cluster information in TM-Cluster and CD-HIT analyses as potentially informative input features. In addition, sequence and structure had similar effects on the other three functional properties (off-target catalytic activity, catalytic window, and preferred catalytic motif), with φc values ​​ranging from 0.13 to 0.52, while physicochemical properties had no significant effect on function (all ρ and η² < 0.1) ( Figure 7 bd). These combined results suggest that multidimensional information is needed to improve the accuracy of machine learning-based CD function prediction.

[0054] Example 4 After successfully building and validating the AlphaCD model, the model was further applied to predict the functional properties of a large number of APOBEC family deaminases (CDs) to demonstrate its potential and value in practical applications. The AlphaCD model was used to predict the functional properties of 21,345 CDs, revealing key information such as the catalytic efficiency, off-target effects, catalytic site windows, and preferred catalytic motifs of these enzymes.

[0055] See attached Figures 8 and 9 As an ensemble learning algorithm, random forest (RF) can enhance the diversity and robustness of a single evaluation tool and improve the ability of function prediction based on structure alone. Therefore, we used this method to combine CD-HIT clustering and structure-based TM-Cluster groups (as mentioned above, CD-HIT and TM-Cluster together have a higher correlation with catalytic efficiency), isoelectric point, protein subfamily, instability index, GRAVY score and protein length (due to its universality) to construct two regression models (catalytic efficiency and off-target effect) and two classification models (catalytic site window and preferred catalytic motif) ( Figure 8 a and Fig. 9a). In the catalytic efficiency model, SHAP (SHapley Additive exPlanations) values ​​and feature importance were calculated as criteria for evaluating the contribution of each feature. The results showed that sequence clustering based on CD-HIT contributed the most to the accuracy of catalytic efficiency prediction (average |SHAP| = 0.045), followed by TM-Cluster structure group (0.036). This jointly supports the effectiveness of combining protein sequence and structure classification in learning factors that affect protein function. Isoelectric point also significantly improved the prediction accuracy in the catalytic efficiency model (0.044). These seven features also made similar contributions in the other three prediction models. Although TM-Cluster structure classification made significant contributions in the three models, it contained significantly less information in predicting motif preference, while GRAVY made a larger contribution ( Figure 8 b and Fig. 9 bd), which is related to the association between motif preference and protein hydrophilicity, and the results show that combining protein sequence, structure and other physicochemical properties can significantly improve the accuracy of function prediction. In particular, CD-HIT clustering, TM-Cluster structure group and isoelectric point show important contributions in the model.

[0056] In order to better understand the predictive ability of the model, after sample binning and parameter tuning optimization, the RF regression results between the predictions and observed catalytic properties of the catalytic efficiency and off-target effect models showed that the catalytic efficiency regression model had an R2 for predicting CDs with catalytic efficiencies greater than 0.5. 2 The value was 0.92, the mean absolute error (MAE) was low (0.02), and the root mean square error (RMSE) was 0.05 (Figure 8c). In addition, the off-target effect regression prediction model had a high consistency between the predicted and observed activities and the catalytic efficiency model (R 2 = 0.84)( Figure 8 d). These results indicate that the RF regression model based on multimodal data can reliably estimate the catalytic efficiency and off-target catalytic activity of CD.

[0057] ROC analysis of the catalytic window RF classification model showed reasonable accuracy (0.73) and a high total area under the curve (AUC) (0.95) after SMOTE undersampling and parameter tuning optimization, but low precision, recall, and F1 score (0.63, 0.64, and 0.61, respectively) ( Figure 8 e), which may be due to the small sample size of some window types. In contrast, the motif preference model showed higher accuracy, precision, recall, and F1 score (all 0.78), with an AUC of 0.93 ( Figure 8f). These results indicate that the two models have sufficient accuracy in characterizing CDs with specific catalytic windows or preferential catalytic motif features.

[0058] To further validate the predictive power of each model in estimating the functional properties of uncharacterized deaminases, 21,345 APOBEC family deaminases were searched and screened from the UniProt database using the same homology-based search strategy with the keyword "APOBEC". AF2 structure prediction provided high-confidence 3D structures (pLDDT value = 93.29 ± 5.39) for all 21,345 CDs ( Fig. 9 e). Sequences annotated to 11 APOBEC subfamilies were extracted, and sequences without methionine start or stop codons were filtered out. Four respective RF models were then used to generate predictions for catalytic efficiency, off-target effects, catalytic window, and preferred catalytic motifs. Of the 21,345 CDs, 18,138 were predicted.

[0059] Example 5 In the field of gene editing, high-efficiency and low-off-target deaminases are the key to developing high-performance base editors (CBEs). In this invention, the A0A2R2Z4E4 deaminase with high catalytic efficiency but relatively high off-target effect was optimized by alanine scanning strategy to reduce the off-target effect while maintaining or improving its catalytic efficiency.

[0060] See attached Figures 10-12 , Evolution of candidate CDs to reduce off-target effects: A0A2R2Z4E4 (186 amino acids) has a high catalytic efficiency (59.9%) and shows great potential for application in efficient CBEs, but its relatively high off-target effect limits its further development. To address this issue, an alanine scanning strategy was used to reduce the off-target effects of A0A2R2Z4E4. Specifically, we constructed 185 CBE variants with a single amino acid mutation (converting it to alanine (A)) at each position in the A0A2R2Z4E4 protein sequence ( Fig.10 a). Then, by expressing each variant in HEK293T cells containing a library of 102 sgRNA-target sequences, the A0A2R2Z4E4 variant CBE with higher catalytic efficiency and lower off-target effects than the wild type (WT) was screened ( Fig.10 a). R-loop assay was used to evaluate sgRNA-independent off-target effects. The results showed that 98 variants improved the on-target editing efficiency, 91 variants reduced the off-target effects, and 13 mutations both improved the on-target editing efficiency and reduced the off-target effects ( Fig.10 bc and Fig.11a). It is worth noting that the off-target effect of A0A2R2Z4E4E100A is 43.9% of that of wild-type A0A2R2Z4E4, but its on-target editing efficiency is increased by 1.09 times ( Fig.10 cd and Fig.11 b). In addition, the off-target effect of A0A2R2Z4E4M1A is 89.3% of that of the most widely used high-fidelity deaminase YE1, but its catalytic efficiency is 1.5 times that of YE1. Statistics of the catalytic efficiency and off-target effect ratios of the A0A2R2Z4E4 variant CBEs showed that the editing specificity of 18 variants was higher than that of the widely used YE1 CBE ( Fig.10 cd and Fig.11 b). In addition, the analysis of motif preference and catalytic window showed that almost all variants with higher catalytic efficiency than WT A0A2R2Z4E4 retained the motif preference and catalytic window characteristics of WT ( Fig.11 c, Fig.12 Collectively, these results suggest that alanine scanning can be used to improve the fidelity of CBEs while maintaining or increasing WT catalytic efficiency, thereby increasing their potential for use in editing-challenging cell types and sites.

[0061] By using experimental data from 1,100 CDs, a dataset for training supervised machine learning models was generated. Supervised learning systems use annotated data for model training, combined with appropriate data preprocessing steps, model selection, and parameter tuning techniques to ultimately achieve high-precision predictions. In addition, supervised learning models require a large number of high-quality annotated datasets for extensive training, and an iterative cycle of training and testing to generate high-precision machine learning models. Currently, tens of thousands of sequences of natural APOBEC-like CDs are available. To investigate the power of machine learning in large-scale enzyme biochemical characterization, we fused APOBEC CDs with nCas9, constructed an intelligent system, and systematically analyzed the biochemical features of 1,100 APOBEC CDs. Based on the high-quality training dataset generated from these CDs, we developed a highly accurate supervised learning model, AlphaCD, to predict the biochemical features of 21,345 uncharacterized APOBEC CDs. In addition, AlphaCD was able to efficiently validate test data, which we used to further optimize the model.

[0062] Although the one-dimensional peptide sequence of a protein determines its structure, and structural features determine the biological function of the protein, we found that clustering methods based solely on one-dimensional sequence similarity or three-dimensional structural similarity cannot adequately cluster proteins with the same function. This finding suggests that physical properties of candidate proteins and other features in annotation information may help improve their functional prediction. To explore this possibility, we constructed AlphaCD, a multimodal machine learning model that integrates one-dimensional sequence, protein structure, protein family annotation, and physicochemical properties to accurately predict the functional features of a large set of APOBEC-like family CDs. Experimental validation showed that AlphaCD has high prediction accuracy in mining efficient and high-fidelity deaminases with specific catalytic properties, indicating that AlphaCD is superior to traditional methods based on sequence alignment or structural homology, and is strongly supported by resampling the prediction dataset.

[0063] When characterizing the catalytic properties of APOBEC family proteins from public databases, we noticed that some predicted CD functions were slightly lower than the observed values. This discrepancy may be attributed to the imbalanced distribution of samples in our self-built CD database, where more than half of the CDs had catalytic efficiencies below 0.1, while the highest prediction accuracy (e.g., ROC AUC) of AlphaCD reached 0.92. We expect that this imbalance will decrease with the continuous update of new functional data from other protein databases. Although this study aims to provide a proof-of-concept for applying AlphaCD to screen efficient CDs from large-scale datasets, future work will focus on rigorously validating CDs with high catalytic efficiency, low off-target activity, and narrow catalytic windows or other specific functions.

[0064] Biological enzymes have a wide range of applications in agriculture and industry, but experimental characterization of their biochemical properties is often costly. In addition to building a high-quality annotated training dataset and a multimodal machine learning model to screen APOBEC deaminases based on predicted functions, our study provides a framework to guide the discovery and functional characterization of other enzymes that are valuable for industrial or research applications. Our results also show that protein function is influenced not only by protein sequence and structure, but also by a variety of other factors. It is critical to consider these collective features in de novo protein design. In addition, the efficient and high-fidelity candidate CDs obtained through our model and screening process can serve as potentially valuable tools for the development of new high-performance base editors (CBEs) for research and personalized precision medicine applications.

[0065] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting cytosine deaminase function based on a machine learning model, characterized in that: include: S1, Screening synthetic cytosine deaminase based on Uniprot database and constructing multidimensional feature dataset, Multidimensional features include catalytic properties, amino acid sequence properties, three-dimensional structural properties, and related biophysical and chemical properties; S2, cluster analysis is performed based on the multidimensional feature data set constructed in S1 to extract features that contribute significantly to the prediction of cytosine deaminase function; S3, using machine learning algorithms, combines the features extracted by S2 with the data set of multi-dimensional features obtained by S1 to build a machine learning model; S4, using the machine learning model constructed in S3, performs functional prediction on uncharacterized cytosine deaminases, and screens out cytosine deaminases for constructing efficient and high-fidelity cytosine base editors based on the prediction results.

2. The method for predicting cytosine deaminase function based on a machine learning model according to claim 1, characterized in that: The cluster analysis includes sequence-based cluster analysis and structure-based cluster analysis; wherein the sequence-based cluster analysis uses CD-HIT and MMseqs tools, and the structure-based cluster analysis uses Foldseek and TM-Cluster tools.

3. The method for predicting cytosine deaminase function based on a machine learning model according to claim 1, characterized in that: The extracted features that contribute significantly to the prediction of CDase function include protein length, sequence- and structure-based clustering information, subfamily and physicochemical properties.

4. The method for predicting cytosine deaminase function based on a machine learning model according to claim 1, characterized in that: The machine learning algorithm is any one of a random forest, a support vector machine, a neural network, a gradient boosting tree and a deep learning algorithm.

5. The method for predicting cytosine deaminase function based on machine learning according to claim 1, characterized in that: The functional prediction includes catalytic efficiency, off-target catalytic activity, catalytic site window and catalytic motif preference.

6. A cytosine deaminase function prediction model, characterized in that: The model for predicting the function of cytosine deaminase is obtained by the method for predicting the function of cytosine deaminase based on a machine learning model as described in any one of claims 1 to 5.

7. The use of the cytosine deaminase function prediction model according to claim 6 in screening proteins or their functional conservative variants for constructing cytosine base editors, characterized in that: The protein or its functional conservative variant is used as a cytosine deaminase to construct an efficient and high-fidelity cytosine base editor.

8. The use according to claim 7, characterized in that: The protein or its functional conservative variant is selected from: CD0029, CD0276, CD0718, CD0376, CD0183, CD0760, CD0628, CD0311, CD0991, CD0503, CD0093, CD0791, CD0336, CD0317, CD0288, CD0702, CD0191, CD1101, CD0796, CD0608, CD0392, CD0759, CD0289, CD0752, CD0409, CD1008, CD0798, CD1049, CD0596, CD0506, CD0290, CD0537, CD0027, CD 0723, CD0529, CD1104, CD0342, CD0222, CD0142, CD1107, CD0509, CD0558, CD0655, CD0691, CD0082, CD0193, CD0393, CD0974, CD0746, CD0382, CD0981, CD0685, CD0813, CD1019, CD0711, CD0751, CD1022, CD0595, CD0256, CD0749, CD0562, CD0452, CD0354, CD0479, CD0357, CD0459, CD0481, CD0757, CD0155 , CD0225, CD0132, CD0178, CD0389, CD0223, CD0294, CD0934, CD0473, CD0128, CD0293, CD1058, CD0549, CD0347, CD0065, CD0805, CD0535, CD0237, CD08 92. CD1097, CD0390, CD0896, CD0546, CD1189, CD0461, CD0450, CD0381, CD0196, CD0491, CD1054, CD0545, CD0877, CD0425, CD1094, CD1105, CD1114, CD0 419, CD1128, CD1109, CD0420, CD0322, CD1193, CD0362, CD0804, CD0571, CD1112, CD0350, CD0973, CD0766, CD0527, CD0582, CD0933, CD0371, CD0488, C D0334, CD1031, CD1181, CD0335, CD1111, CD0489, CD0669, CD0076, CD0982, CD0455, CD0531, CD0446, CD0893, CD0528, CD0525, CD0245, CD0230, CD0255,CD1131、CD0602、CD0735、CD0738、CD0176、CD0635、CD0241、CD0331、CD0395、CD0150、CD0319、CD0659、CD0465、CD0557、CD1135、CD0072、CD1174、CD0295、CD0895、CD0668、CD0677、CD0063、CD1199、CD0098、CD0249、CD0433、CD0404、CD1172、CD0797、CD1006、CD0663、CD0876、CD1092、CD1059、CD0875、CD0819、CD0976、CD0830、CD0464、CD0402、CD0753、CD1195、CD0166、CD1082、CD0476、CD0271、CD0449、CD0325、CD0138、CD0352、CD0470、CD1002、CD0917、CD0431、CD0316、CD0188、CD0364、CD0576、CD0179、CD0739、CD0706、CD0514、CD1034、CD1093、CD0227、CD0365、CD0587、CD0720、CD0605、CD0520、CD1184、CD0505、CD0384、CD0838、CD0388、CD0707、CD0953、CD1043、CD0740、CD0284、CD0667、CD0229、CD0927、CD1084、CD0951、CD0560、CD0742、CD1042、CD0656、CD0463、CD0010、CD0594、CD0729、CD0915、CD0761、CD0883、CD1134、CD1005、CD0474、CD0888、CD0002、CD0521、CD0764、CD0682、CD0172、CD0658、CD1183、CD1028、CD1190、CD0101、CD0156、CD1137、CD0422、CD1186、CD0715、CD1017、CD0019、CD0168、CD0928、CD0043、CD0344、CD1188、CD0724、CD1064、CD0209、CD1194、CD1113、CD0462、CD0105、CD0621、CD0552、CD0307、CD0997、CD0840、A0A0K0MIC1、A0A455BCL3、A0A0K0MJ52、A0A671DNI5、A0A2R2Z4E4、A0A6P6C7T7, P31941, G8GPV4, P51908, A0A2R2X2I5, A0A1U7Q2F3, A0A4W2FT84, A0A803TEP0, G1LKJ4, A0A7J8FW00, A0A2K6MSA4, K7BTL7, A0A822EZF5, F7F6M6, Q8NFB8, A0A6P5KZH5, A0A4W2HPD5, Q4VB33, A0A834EE11, G1RYZ5, A0A7X8SJJ0, A7VGG7, A0A4Q9H5D8, A0A5J4P2S6, or A0A1J5GBA4.

9. The use according to claim 7, characterized in that: The screened protein function conservative variant is a function conservative variant of A0A2R2Z4E4, selected from: H56A, N45A, M1A, I82A, R54A, Y114A, P86A, R112A, F6A, L124A or E100A.

10. Use of the cytosine deaminase function prediction model according to claim 6 in screening reagents or kits for preparing C-to-T gene editing.

Images