Deep feature and automatic machine learning-based sgRNA activity prediction method and device

Abstract

The invention relates to the technical field of gene editing, and discloses an sgRNA activity prediction method and device based on deep features and automatic machine learning. The method comprises the following steps: acquiring sgRNA training data with editing efficiency labels and discretizing the sgRNA training data into classification labels; after the sgRNA and the PAM sequence are spliced, inputting the spliced sgRNA and PAM sequence into a pre-training depth sequence model to extract high-dimensional features; obtaining a low-dimensional feature vector through nonlinear dimensionality reduction; an automatic machine learning framework is combined with cross validation, and a prediction model is automatically trained and optimized based on low-dimensional features; and finally, carrying out activity prediction and sorting screening on candidate sgRNA by utilizing the model. According to the method, through fusion of deep semantic extraction, nonlinear structure maintenance and a full-automatic modeling process, the accuracy, robustness and development efficiency of sgRNA activity prediction are effectively improved, and an efficient pilot screening tool is provided for a gene editing experiment.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and gene editing technology, and particularly relates to a method and device for predicting sgRNA activity based on deep features and automated machine learning. Background Technology

[0002] In CRISPR gene editing technology, sgRNA activity directly determines the success rate and efficiency of gene editing. Screening for highly active sgRNAs one by one using traditional biological experimental methods is not only costly but also time-consuming. Therefore, developing accurate and efficient computational methods to predict sgRNA activity is of great significance for gene editing research and applications.

[0003] In the prior art, some computational methods have attempted to address this problem. For example, some methods (such as invention patents with publication numbers CN113257359A, CN120748488A, and CN111613274A) combine the sgRNA sequence with the prototype spacer neighbor motif (PAM) sequence as input and employ traditional machine learning models (such as linear models like Lasso regression) for feature selection and model construction. However, these methods have significant drawbacks: First, the linear models they employ struggle to fully extract the complex nonlinear patterns and deep semantic information within the sequence data, resulting in limited feature extraction capabilities and directly impacting prediction accuracy. Second, the linear dimensionality reduction or feature selection methods they employ are ineffective when processing high-dimensional features generated by complex models, potentially losing crucial nonlinear structural information, thus limiting the model's generalization ability and robustness. Finally, the model construction process heavily relies on extensive manual parameter tuning and model comparison, which is inefficient and makes it difficult to guarantee finding the optimal model combination. These shortcomings collectively result in the need to improve the prediction accuracy and robustness of existing methods. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for predicting sgRNA activity based on deep features and automated machine learning, in order to solve the technical problems of low prediction accuracy, poor model robustness and low modeling efficiency caused by insufficient feature extraction capabilities, inappropriate dimensionality reduction methods and cumbersome model building processes in sgRNA activity prediction methods.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for predicting sgRNA activity based on deep features and automated machine learning, comprising the following steps: S1. Training data acquisition: Obtain a training dataset containing multiple sgRNA sequences and their corresponding editing efficiencies; S2, Discretization of Editing Efficiency: Convert the continuous edit efficiency values ​​in the training dataset into categorical labels; S3. Sequence concatenation to construct input: For the sgRNA sequences in the training dataset, the prototype spacer neighbor motif (PAM) sequence is spliced ​​with the sgRNA sequence to form a complete input sequence; S4. Deep Feature Extraction: The complete input sequence is fed into a pre-trained deep sequence model to extract a high-dimensional semantic representation of the input sequence; S5. Nonlinear dimensionality reduction processing: A nonlinear dimensionality reduction algorithm is applied to the high-dimensional semantic representation to obtain a low-dimensional feature vector; S6, AutoML Model Training and Optimization: An automated machine learning framework, combined with a cross-validation strategy, is used to automatically train and optimize the model based on the low-dimensional feature vector and the categorical label to obtain an sgRNA activity prediction model. S7. Candidate sgRNA Activity Prediction and Screening: Candidate PAM sequences are spliced ​​with candidate sgRNA sequences and input into the trained sgRNA activity prediction model to predict the activity of candidate sequences. The sequences are then sorted according to their predicted activity levels, and the Top N sgRNAs are selected for experimental verification.

[0006] Optionally, in step S2, the category label is a binary label.

[0007] Optionally, in step S4, the pre-trained deep sequence model is a DNA model based on the Transformer architecture. Further, as an example in this embodiment of the invention, the pre-trained deep sequence model is the EVO2 model, but it is not limited thereto.

[0008] Optionally, in step S5, the nonlinear dimensionality reduction algorithm is the Uniform Manifold Approximation and Projection (UMAP) algorithm. Cross-validation is used to search across different dimensionality reduction dimensions to select the optimal low-dimensional feature representation.

[0009] Optionally, in step S6, the automated machine learning framework is AutoGluon, and the model is preferably based on at least one of macro-F1 score and balanced accuracy. The cross-validation strategy employs 5-fold hierarchical cross-validation.

[0010] The present invention also provides an sgRNA activity prediction device based on deep features and automated machine learning, a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. When the computer processor executes the computer program, it implements the above-described sgRNA activity prediction method based on deep features and automated machine learning.

[0011] The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-described method for predicting sgRNA activity based on deep features and automated machine learning.

[0012] The beneficial effects of this invention are: 1. Experimental verification proves the effectiveness of the method. The 13 highly active sgRNAs predicted by the model of this invention were experimentally verified, with a cleavage efficiency ranging from a minimum of 25% to a maximum of 70%, which is significantly better than random design, fully demonstrating the effectiveness and practical value of the method of this invention.

[0013] 2. Significantly improves prediction accuracy and robustness Deep sequence models mine sequence semantics, and nonlinear dimensionality reduction preserves key structural information, effectively overcoming the limitations of traditional linear models.

[0014] 3. Highly automated modeling process AutoML automatically completes model selection, hyperparameter search, and model fusion, significantly reducing manual intervention.

[0015] 4. Reduce noise impact and improve prediction stability Using classification labels to replace continuous noise editing improves efficiency and increases the model's tolerance to experimental errors.

[0016] 5. Significantly improves the efficiency of gene editing research and development. Researchers can screen out a large number of low-activity sgRNAs before experiments, thereby shortening the cycle and saving costs. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an sgRNA activity prediction method based on deep features and automated machine learning according to the present invention. Figure 2 This is a schematic diagram of an sgRNA activity prediction device based on deep features and automated machine learning according to the present invention. Figure 3 The efficiency of IAG gene editing is represented by discrete points, which are parallel samples. Figure 4 The efficiency of MIH gene editing is represented by discrete points, which are parallel samples. Figure 5 The efficiency of Dmrt gene editing is represented by discrete points, which are parallel samples. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0019] See Figure 1 This invention provides a method for predicting sgRNA activity based on deep features and automated machine learning, specifically including the following steps: S1. Training Data Acquisition The sgRNA sequences and their corresponding editing efficiency values ​​were collected from the experimental results to form a training dataset. Each record includes the sgRNA sequence, the corresponding PAM sequence, and the actual measured cutting efficiency.

[0020] S2, Discretization of Editing Efficiency Convert continuous editing efficiency values ​​into discrete tags based on a set threshold, for example: ≥A certain threshold is defined as high activity (1); This threshold is denoted as low activity (0).

[0021] In this embodiment, binary labels are used.

[0022] S3, Sequence splicing to construct input Each sgRNA sequence is spliced ​​with its corresponding PAM sequence so that the model simultaneously obtains the sequence information of both the sgRNA and PAM regions, forming a complete input sequence, in the format: sgRNA+PAM.

[0023] S4, Deep Feature Extraction The complete input sequence is fed into a pre-trained deep sequence model EVO2 to extract high-dimensional semantic representations of the input sequence.

[0024] The EVO2 model, trained on a large number of biological sequences, can automatically capture sequence structure, contextual information, and potential functional features. Specifically, the Evo2 model is a deep learning model pre-trained on large-scale genomic data, capable of understanding the syntax and semantics of DNA sequences. This model can map the input nucleotide sequence to a high-dimensional continuous vector space, generating a high-dimensional semantic representation (or embedding vector). This vector can capture complex, non-linear patterns and potential biological meanings within the sequence, which is difficult to achieve with traditional methods (such as one-hot encoding combined with linear models).

[0025] S5, Nonlinear Dimensionality Reduction Processing The UMAP (Uniform Manifold Approximation and Projection) algorithm is applied to the high-dimensional semantic vectors to perform nonlinear dimensionality reduction, resulting in a more compact low-dimensional feature representation that preserves the manifold structure, which can then be used for subsequent machine learning model training.

[0026] S6, AutoML Model Training and Optimization The AutoGluon automated machine learning framework is used to construct a classification model with low-dimensional feature vectors as input and categorical labels as output.

[0027] AutoGluon automatically performs: feature selection, model structure search, and hyperparameter optimization.

[0028] Model fusion: The optimal model is selected based on indicators such as macro-F1 score and balance accuracy as the final sgRNA activity prediction model.

[0029] S7. Candidate sgRNA Activity Prediction and Screening The candidate sgRNAs to be screened and their PAM sequences are input into the trained prediction model to obtain the activity score or classification result of each candidate sgRNA. The top N sgRNAs are selected for experimental validation based on their activity levels.

[0030] An embodiment of the sgRNA activity prediction method based on deep features and automated machine learning of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2 The diagram shown is a hardware structure diagram of any data processing-capable device, in which the sgRNA activity prediction method and apparatus based on deep features and automated machine learning of the present invention are located. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of that data processing device, which will not be elaborated further. The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0031] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0032] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements an sgRNA activity prediction method based on deep features and automated machine learning as described in the above embodiments.

[0033] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0034] Example 1: A method for predicting sgRNA activity based on deep features and automated machine learning This embodiment discloses a method for training an sgRNA activity prediction model. The method aims to construct a high-precision, robust prediction model from a small amount of sgRNA sequences and their activity data through an innovative process combining deep sequence models, nonlinear dimensionality reduction, and automated machine learning, and then use the model to predict sgRNA activity.

[0035] S1. Training Data Acquisition A training dataset containing multiple sgRNA sequences, their corresponding PAM sequences, and their editing efficiencies was obtained from the wet experiment results. The fields included sgRNA, PAM, and editing efficiency.

[0036] Basic cleaning was performed on each sgRNA and PAM sequence, including: capitalizing all letters; replacing non-A, C, G, and T characters with "N".

[0037] The loaded data is denoted as a DataFrame, where the editing efficiency is converted to floating-point values ​​between 0 and 1. The specific training set data is shown in Table 1.

[0038] Table 1: Training Set S2, Discretization of Editing Efficiency Discretize continuous editing efficiency values ​​into binary labels (the threshold is selected to maintain the same number of samples in each category): Values ​​less than the threshold of 0.21 are denoted as "low"; Values ​​greater than or equal to the threshold of 0.21 are denoted as "high".

[0039] S3, Sequence splicing to construct input The cleaned sequence of each sgRNA and its corresponding PAM sequence were sequentially spliced ​​together and used as input for the deep model.

[0040] The code is as follows: df["combined_seq"] = df["sgRNA_clean"] + df["PAM_clean"].

[0041] S4, Deep Feature Extraction The pre-trained deep sequence model EVO2 (EVO model: evo2_1b_base) is used to perform embedding operations on the concatenated sequence to obtain high-dimensional semantic vectors (by extracting the blocks.19.mlp.l3 layer of EVO2).

[0042] The code is as follows: def generate_evo2_embedding(sequence: str, model, device: str, layer:str = 'blocks.19.mlp.l3') -> np.ndarray: input_ids = torch.tensor(model.tokenizer.tokenize(sequence),dtype=torch.int).unsqueeze(0).to(device) with torch.no_grad(): _, embeddings = model(input_ids, return_embeddings=True,layer_names=[layer]) return embeddings[layer].mean(dim=1).squeeze(0).float().cpu().numpy().

[0043] S5, Nonlinear Dimensionality Reduction Processing UMAP nonlinear dimensionality reduction is performed on the high-dimensional embedding, and the dimensionality reduction dimension can be searched and optimized between 8 and 16. This embodiment explores different dimensions iteratively and uses cross-validation to select the optimal dimension.

[0044] The code is as follows: dim_sizes = range(8,16) results = [] for dim_size in dim_sizes: print(f"\n===== Exploring dim_size = {dim_size} =====") best_model, best_reducer, (bal_acc, f1) = train_and_evaluate_cv( X_scaled, y, method="UMAP", dim=dim_size, n_splits=5, time_limit=120, save_prefix=f"edit_eff_dim{dim_size}" ) results.append({ "dim_size": dim_size, "balanced_acc": bal_acc, "macro_f1": f1, "model": best_model, "reducer": best_reducer }).

[0045] S6, AutoML Model Training and Optimization AutoGluon Tabular is used to automate the modeling of dimensionality-reduced features. The dimensionality-reduced feature data is split into five parts: four parts are used to train the model, and one part serves as a validation set for testing. The training process includes: 1. Use the StratifiedKFold function to perform 5-fold stratified cross-validation; 2. Train the AutoGluon model for each fold and evaluate its performance on the validation set; 3. Performance indicators are measured using macro-F1 and balance accuracy. 4. Automatically select the model with the best fold and the corresponding UMAP dimensionality reducer as the final model.

[0046] The code is as follows: def train_and_evaluate_cv( X, y, method: str = None, # None / "PCA" / "UMAP" dim: int = None, n_splits: int = 5, time_limit: int = 120, random_state: int = 42, save_prefix: str = "model ): skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=random_state) bal_accs, f1s = [], [] fold_predictors, fold_reducers = [], [] for fold, (train_idx, val_idx) in enumerate(skf.split(X, y)): print(f"\n===== Fold {fold+1} / {n_splits} =====") X_train, X_val = X[train_idx], X[val_idx] y_train, y_val = y[train_idx], y[val_idx] # Dimensionality reduction if method is None: reducer = None X_train_df = pd.DataFrame(X_train) X_val_df = pd.DataFrame(X_val) else: X_train_df, reducer = reduce_dimension(X_train, method=method, n_components=dim) X_val_reduced = reducer.transform(X_val) col_names = [f"{method}_{i}" for i in range(dim)] X_val_df = pd.DataFrame(X_val_reduced, columns=col_names) X_train_df['Editing_class'] = y_train X_val_df['Editing_class'] = y_val predictor = TabularPredictor( label='Editing_class', # problem_type='multiclass', eval_metric='f1_macro' ).fit( X_train_df, presets='best_quality', time_limit=time_limit, verbosity=2 ) y_pred = predictor.predict(X_val_df).values bal_acc = balanced_accuracy_score(y_val, y_pred) f1 = f1_score(y_val, y_pred, average='macro') print(f"Fold {fold+1} | Balanced Acc={bal_acc:.4f} | Macro-F1={f1:.4f}") print(classification_report(y_val, y_pred)) print("Confusion matrix:\n", confusion_matrix(y_val, y_pred)) bal_accs.append(bal_acc) f1s.append(f1) fold_predictors.append(predictor) fold_reducers.append(reducer).

[0047] Through dimensional search and cross-validation, this embodiment finds the optimal dimension to be 10, corresponding to Macro-F1=0.59 and balanced accuracy=0.61.

[0048] Table 2: Hyperparameter selection for 5-fold cross-validation (dimension reduction by UMAP) S7. Candidate sgRNA Activity Prediction and Screening The pipeline involves inputting candidate sgRNAs and their PAMs into the trained model. 1. Sequence splicing; 2. EVO2 embedding; 3. Standardization; 4. UMAP dimensionality reduction; 5. AutoGluon model predicts categories and probabilities.

[0049] The code is as follows: def predict_test(test_file, scaler, reducer, predictor, method, dim, out_file="test_predictions.csv"): test_df = pd.read_csv(test_file) for col in ["ID", "sgRNA", "PAM"]: assert col in test_df.columns, f"Missing column {col}" test_df["sgRNA_clean"] = test_df["sgRNA"].apply(clean_dna_sequence) test_df["PAM_clean"] = test_df["PAM"].apply(clean_dna_sequence) test_df["combined_seq"] = test_df["sgRNA_clean"] + test_df["PAM_clean"] device = 'cuda:0' if torch.cuda.is_available() else 'cpu' evo2_model = Evo2('evo2_1b_base') embedding_file = "test_embeddings.npy" if os.path.exists(embedding_file): test_embeddings = np.load(embedding_file) else: test_embeddings = compute_embeddings(test_df["combined_seq"], evo2_model, device) np.save(embedding_file, test_embeddings) test_embeddings_scaled = scaler.transform(test_embeddings) if reducer is not None: test_reduced = reducer.transform(test_embeddings_scaled) col_names = [f"{method}_{i}" for i in range(dim)] test_reduced_df = pd.DataFrame(test_reduced, columns=col_names) else: test_reduced_df = pd.DataFrame(test_embeddings_scaled) test_preds = predictor.predict(test_reduced_df) test_probas = predictor.predict_proba(test_reduced_df) output = pd.concat([pd.DataFrame({"ID": test_df["ID"], "Prediction": test_preds}), test_probas], axis=1) output.to_csv(out_file, index=False) print(f" Test predictions saved to {out_file}").

[0050] Based on the predicted activity ranking, the Top N sgRNAs were selected for experimental validation (where N is the preset number). The prediction results can be output as a CSV file, containing the fields ID, prediction result (Prediction), P (low), and P (high).

[0051] Table 3: Prediction Results Example 2: Validation of sgRNA activity This embodiment aims to conduct prospective experimental verification of the final sgRNA activity prediction model constructed and determined in Example 1, in order to confirm its predictive ability and reliability in real application scenarios.

[0052] 1. Microinjection of sgRNA:SpCas9 complex into giant freshwater prawn embryos and detection of editing efficiency. Mature female shrimp that had just molted and reached sexual maturity, along with larger male shrimp with intact claws, were collected in the early morning and temporarily housed in a laboratory rearing tank at a female-to-male ratio of 2:1. The female shrimp were observed approximately every hour to check for spawning. After spawning, the fertilized egg mass was removed from the abdomen using forceps and enzymatically digested with 1 mg / ml trypsin to disperse the fertilized eggs into individual eggs. The eggs were then neatly arranged on a wedge-shaped agarose gel mold for microinjection. The injection system consisted of 1.5 pmol / μL Cas9 protein (amino acid sequence as shown in SEQ ID NO. 63) and 2.1 pmol / μL sgRNA, with approximately 0.5 nL injected per egg. The microinjection needle had an opening of approximately 1 μm. After injection, the fertilized eggs were transferred to filtered, sterilized seawater containing 4% PEG4000 and cultured on a horizontal shaking incubator at approximately 28°C.

[0053] 2. Sample collection and editing efficiency testing Embryos were collected 72 h after injection. The genome of the embryo samples was extracted using the Marine Animal Tissue Genomic DNA Extraction Kit (#DP324) (Tiangen). Three biological experimental replicates were set up for each group (sample groups are shown in Table 4). Primers were designed (primer sequence list is shown in Table 5). The target fragment was amplified by PCR using Phanta UniFi Super-Fidelity DNA Polymerase universal annealing ultra-high fidelity single enzyme (P506, #P516 / 526-01) from Novizan. The PCR products were sent to Youkang Biotechnology for Sanger sequencing. The sequencing results were statistically analyzed using the TIDE tool (Brinkman EK, Chen T, Amendola M, vanSteensel B. Easy quantitative assessment of genome editing by sequence trace decomposition. Nucleic Acids Res. 2014 Dec 16;42(22):e168.).

[0054] Table 4: Experimental Groups Table 5: Primer Sequence Listing 3. Editing efficiency results The editing efficiency statistics showed that, compared with the WT control group (developing embryos that were not injected at the same time), the average editing efficiencies of the five sgRNAs targeting the Dmrt11E gene were 35.30%, 25.03%, 32.57%, 36.43%, and 39.17%, respectively. Figure 5 The average editing efficiencies of the four sgRNAs targeting the MIH gene were 34.37%, 45.40%, 43.20%, and 25.47%, respectively. Figure 4 The average editing efficiencies of the four sgRNAs targeting the IAG gene were 32.43%, 35.63%, 38.23%, and 70.67%, respectively. Figure 3 ).

[0055] Table 6: Gene editing efficiency of each sgRNA target gene These editing efficiency results reveal the practical usability of the constructed Macrobrachium rosenbergii genome sgRNA prediction model. The predicted highly efficient sgRNAs all exhibit extremely high editing efficiency, making them suitable for research on functional gene editing in Macrobrachium rosenbergii and the creation of new gene-edited germplasm individuals. Specifically, the IAG-sgRNA4 sequence has been used to efficiently edit the sex-controlling gene IAG, and mutant individuals with IAG gene editing in Macrobrachium rosenbergii have already been obtained.

Claims

1. A method for predicting sgRNA activity based on deep features and automated machine learning, characterized in that, The method comprises the following steps: S1, training data acquisition: acquiring a training data set containing multiple sgRNA sequences and their corresponding editing efficiency; S2, editing efficiency discretization: converting continuous editing efficiency values in the training data set into categorical labels; S3, sequence splicing to construct input: splicing the protospacer adjacent motif sequence and the sgRNA sequence in the training data set to form a complete input sequence; S4, deep feature extraction: inputting the complete input sequence into a pre-trained deep sequence model to extract high-dimensional semantic representation of the complete input sequence; S5, nonlinear dimensionality reduction processing: applying a nonlinear dimensionality reduction algorithm to the high-dimensional semantic representation to obtain a low-dimensional feature vector; S6, AutoML model training and optimization: using an automated machine learning framework, combining a cross-validation strategy, and based on the low-dimensional feature vector and the categorical label, automatically training and optimizing the model to obtain an sgRNA activity prediction model; S7, candidate sgRNA activity prediction and screening: splicing the candidate protospacer adjacent motif and the candidate sgRNA sequence and inputting them into the sgRNA activity prediction model to obtain a predicted activity result, and sorting or screening the candidate sgRNA according to the predicted activity result. 2.The sgRNA activity prediction method of claim 1, wherein, In step S2, the categorical label is a binary classification label. 3.The sgRNA activity prediction method of claim 1, wherein, In step S4, the pre-trained deep sequence model is a DNA model based on a Transformer architecture. 4.The sgRNA activity prediction method of claim 3, wherein, The pre-trained deep sequence model is an EVO2 model. 5.The sgRNA activity prediction method of claim 1, wherein, In step S5, the nonlinear dimensionality reduction algorithm is a uniform manifold approximation and projection algorithm. 6.The sgRNA activity prediction method of claim 1, wherein, In step S6, the automated machine learning framework is AutoGluon. 7.The sgRNA activity prediction method of claim 1 or 6, wherein, When the model is optimized, at least one of the evaluation indicators is based on macro-F1 score and balanced accuracy. 8.The sgRNA activity prediction method of claim 1, wherein, In step S5, different dimensionality reduction dimensions are searched through cross-validation to select the optimal low-dimensional feature representation. 9.A device for sgRNA activity prediction based on deep features and automated machine learning, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein, The computer processor executes the computer program to implement the sgRNA activity prediction method of any one of claims 1-8.

10. A computer storage medium having stored thereon a computer program, characterized in that The computer program is processed to implement the sgRNA activity prediction method of any one of claims 1-8.

Images