Guided editing efficiency prediction method and system based on deep learning and transfer learning
By employing deep learning and transfer learning methods, K-mer features of Spacer, PBS, and RTT sequences are extracted. The editing efficiency of pegRNA is predicted using a Transformer encoder and a fully connected feedforward network. This solves the bias and generalization problems of existing models, and achieves efficient and accurate prediction and automatic design of pegRNA editing efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ACADEMY OF MILITARY MEDICAL SCIENCES
- Filing Date
- 2022-12-23
- Publication Date
- 2026-04-28
AI Technical Summary
Existing guided editing technology models suffer from significant bias, low accuracy, poor generalization, and lack of interpretability, making it difficult to effectively predict the editing efficiency of pegRNA and thus limiting their application in clinical gene therapy.
We employ deep learning and transfer learning to extract K-mer features from Spacer, PBS, and RTT sequences. Feature extraction is performed using an attention network and a Transformer encoder, combined with a multi-layer fully connected feedforward network for nonlinear transformation to predict pegRNA editing efficiency. Furthermore, we enhance the model's generalization ability through transfer learning.
It achieves unbiased and accurate prediction of arbitrary editing efficiency and can automatically design optimal pegRNAs, improving the accuracy and generalization ability of editing efficiency prediction and simplifying the pegRNA design process.
Smart Images
Figure CN115862743B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of gene editing technology, and in particular to a guided editing efficiency prediction method and system based on deep learning and transfer learning. Background Technology
[0002] Prime editing, a novel and precise gene-editing technology proposed in 2019, utilizes pre-designed prime editing guide RNAs (pegRNAs) to arbitrarily perform base substitutions, insertions, and deletions. Therefore, prime editing is a versatile and precise gene-editing technique. Since point mutations, insertion mutations, and deletion mutations cover most pathogenic genetic variations in humans, prime editing holds great potential in clinical gene therapy research, potentially correcting up to 89% of known human disease-related gene mutations. However, prime editing research is still in its early stages. The factors influencing its editing efficiency are unclear. Editing efficiency depends on pegRNAs and sgRNAs, making effective prediction through computational models difficult. Furthermore, the design of pegRNAs and sgRNAs still relies on experimental trials, which is extremely time-consuming and labor-intensive. Finding the optimal pegRNAs and sgRNAs from countless options to maximize prime editing efficiency is challenging, significantly limiting the application of prime editing technology.
[0003] A study published in Nature Biotechnology in 2020 constructed a library of guided editing and trained three different machine learning models, DeepPE, PE_type, and PE_position, based on the library to predict the efficiency of different types of editing. However, the machine learning models in this paper have several problems: First, the models are highly biased. The DeepPE model is a biased model trained and evaluated on a dataset of single-base edits at the +5 position (G to C), and it only has some effect on predicting single-base edits at the +5 position. The PE_type model is a biased model trained and evaluated on datasets of different edit types, and it only has some effect on predicting different types of edits. The PE_position model is a biased model trained and evaluated on datasets of different edit positions, and it only has some effect on predicting edits at different positions. Second, the models are not accurate, especially the PE_type and PE_position models, which perform very poorly. The Pearson correlation coefficients between the predicted and actual edit efficiencies are 0.48 and 0.56, respectively, which severely limits the application of the models in predicting edit efficiencies of different types and positions. Third, the models have poor generalization ability. All models are biased and have low accuracy, so they do not provide a general, highly generalizable model that can accurately predict the efficiency of any edit simultaneously. Fourth, the models are black boxes and lack interpretability. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to provide a guided editing efficiency prediction method and system based on deep learning and transfer learning. The trained artificial intelligence model can predict the efficiency of arbitrary editing unbiasedly and accurately, and provides some interpretability for the results, thereby effectively predicting the editing efficiency of different pegRNAs. Furthermore, based on this artificial intelligence prediction system, an optimal pegRNA system that can be automatically designed has been developed, thereby quickly and accurately obtaining the optimal design from countless possible structures to achieve the highest editing efficiency.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: The present invention discloses a guided editing efficiency prediction method based on deep learning and transfer learning, the method comprising:
[0006] K-mer sequences were extracted from the Spacer sequence, PBS sequence, and RTT sequence, respectively, where K = 1, 2, and 3;
[0007] The hidden vector of each K-mer sequence is summed with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence.
[0008] The sequence description vectors corresponding to the Spacer sequence, the PBS sequence, and the RTT sequence are concatenated to form the description vector of the pegRNA.
[0009] The description vector of pegRNA is nonlinearly transformed using a multilayer fully connected feedforward network and then input into a guided editing efficiency prediction system. The guided editing efficiency prediction system outputs the predicted editing efficiency of the pegRNA description vector.
[0010] Preferably, before the step of weighted summation of the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence, the method further includes: using an attention network to map the latent vector of each K-mer sequence to a scalar score.
[0011] The SoftMax function is used to compute the latent vector of each K-mer sequence based on the weight score of the attention network.
[0012] Preferably, the step of extracting the K-mer sequences of the Spacer sequence, PBS sequence, and RTT sequence respectively, wherein K = 1, 2, 3, further includes...
[0013] Based on the M-nt length sequences of the Spacer sequence, the M-nt length sequences of the PBS sequence, and the M-nt length sequences of the RTT sequence, a K-mer sequence of length M-K+1 is obtained.
[0014] Each K-mer sequence is transformed into a fixed-dimensional d using word embeddings. model Numerical vectors;
[0015] Positional encoding using sine and cosine functions with different frequencies was incorporated into the word embedding.
[0016] Preferably, before the step of weighted summation of the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence, the method further includes:
[0017] Extract the Spacer sequence using the Transformer encoder;
[0018] The PBS sequence was extracted and encoded using a Transformer decoder;
[0019] RTT sequences are extracted and encoded using a Transformer decoder;
[0020] Normalize each K-mer sequence hidden by the Spacer sequence, each K-mer sequence hidden by the PBS sequence, and each K-mer sequence hidden by the RTT sequence using the Transformer.
[0021] Preferably, the training process of the guided editing efficiency prediction system includes:
[0022] Multiple pegRNA samples were obtained from the HT-training, Type-training, and Position-trainingsets using random resampling with replacement.
[0023] Multiple pegRNA samples were merged together to construct a merged training set, which was then used to train and guide the editing efficiency prediction system.
[0024] Preferably, the training process of the guided editing efficiency prediction system includes: initializing the guided editing efficiency prediction system using the optimal guided editing efficiency prediction system that has been pre-trained on the HT-trainingset training set of Library1, and modifying the upper fully connected feedforward network into multiple hidden layers.
[0025] Preferably, the retained HT-test, Type-test, and Position-test datasets are used as three independent test datasets. The final guided editing efficiency prediction system is evaluated based on the Pearson correlation coefficient r, Spearman correlation coefficient R, mean absolute error MAE, and root mean square error RMSE between the predicted editing efficiency and the actual measured efficiency in the experiment.
[0026] The second objective of this invention can be achieved by adopting the following technical solution: a guided editing efficiency prediction system based on deep learning and transfer learning, the system comprising:
[0027] The input unit is used to extract K-mer sequences from Spacer, PBS, and RTT sequences, respectively, where K = 1, 2, 3;
[0028] The first data processing unit is used to sum the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence.
[0029] The second data processing unit is used to concatenate the sequence description vectors corresponding to the Spacer sequence, the PBS sequence, and the RTT sequence to form the description vector of the pegRNA.
[0030] The prediction unit is used to perform a nonlinear transformation on the description vector of pegRNA using a multilayer fully connected feedforward network and then input it into the guide editing efficiency prediction system. The guide editing efficiency prediction system outputs the predicted editing efficiency of the pegRNA description vector.
[0031] The third objective of this invention can be achieved by adopting the following technical solution:
[0032] A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described guided editing efficiency prediction method based on deep learning and transfer learning.
[0033] The fourth objective of this invention can be achieved by adopting the following technical solution:
[0034] A storage medium that stores a program, which, when executed by a processor, implements the aforementioned guided editing efficiency prediction method based on deep learning and transfer learning.
[0035] This invention extracts K-mer sequences from Spacer, PBS, and RTT sequences, where K = 1, 2, and 3 respectively. The latent vector of each K-mer sequence is weighted and summed with its attention weight to obtain the sequence description vectors corresponding to the Spacer, PBS, and RTT sequences. These vectors are then concatenated to form the pegRNA description vector. A multilayer fully connected feedforward network is used to perform a nonlinear transformation on the pegRNA description vector, which is then input into a guided editing efficiency prediction system. This system outputs the predicted editing efficiency of the pegRNA description vector. The artificial intelligence model trained in this application can predict the efficiency of arbitrary editing unbiasedly and accurately, and provides some interpretability to the results. This effectively predicts the editing efficiency of different pegRNAs. Furthermore, based on this artificial intelligence prediction system, an optimal pegRNA system can be automatically designed, quickly and accurately obtaining the optimal design from countless possible structures to achieve the highest editing efficiency. Attached Figure Description
[0036] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings:
[0037] Figure 1This is a flowchart of the guided editing efficiency prediction method based on deep learning and transfer learning of the present invention;
[0038] Figure 2 This is a schematic diagram of the guided editing efficiency prediction method based on deep learning and transfer learning of the present invention. Detailed Implementation
[0039] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.
[0040] Existing technologies suffer from low model accuracy, poor generalization, bias, and are essentially black boxes lacking interpretability. This invention addresses these issues by extracting K-mer sequences from Spacer, PBS, and RTT sequences (K = 1, 2, 3). The latent vector of each K-mer sequence is weighted and summed with its attention weight to obtain the sequence description vectors for the Spacer, PBS, and RTT sequences. These vectors are then concatenated to form the pegRNA description vector. A multilayer fully connected feedforward network is used to perform a nonlinear transformation on the pegRNA description vector before inputting it into a guided editing efficiency prediction system. This system then outputs the predicted editing efficiency of the pegRNA description vector. The artificial intelligence model trained in this application can predict the efficiency of arbitrary editing unbiasedly and accurately, and has a certain interpretability of the results, thus effectively predicting the editing efficiency of different pegRNAs. Furthermore, based on this artificial intelligence prediction system, an optimal pegRNA system that can be automatically designed has been developed, thereby quickly and accurately obtaining the optimal design from countless possible structures to achieve the highest editing efficiency.
[0041] Example 1
[0042] This application provides a guided editing efficiency prediction method based on deep learning and transfer learning, including: step 100, extracting K-mer sequences of Spacer sequence, PBS sequence and RTT sequence respectively, where K = 1, 2, 3;
[0043] Step 200: The hidden vector of each K-mer sequence is weighted and summed with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence.
[0044] Step 300: Concatenate the sequence description vectors corresponding to the Spacer sequence, the PBS sequence, and the RTT sequence to form the description vector of the pegRNA.
[0045] Step 400: The description vector of pegRNA is nonlinearly transformed using a multilayer fully connected feedforward network and then input into the guided editing efficiency prediction system. The guided editing efficiency prediction system outputs the predicted editing efficiency of the pegRNA description vector.
[0046] In this implementation, the core of the pegRNA consists of three subsequences: a spacer, a primer binding site (PBS), and a reverse transcription template (RTT). The RTT embeds the intended arbitrary editing, including base substitution, insertion, and deletion. To fully extract the effective features of the target DNA and pegRNA pair, we designed and developed a guided editing efficiency prediction system based on deep learning and transfer learning, tailored to the pegRNA structure. This system is based on an artificial intelligence network model using Transformer and Attention. Its input data consists of the target DNA and pegRNA sequences, including the spacer (47 bp of upstream and downstream target sequences), a variable-length PBS sequence, and the RTT sequence—a total of three subsequences as input features. The system automatically extracts the inverse complementary sequences of the PBS and RTT, and then further extracts the K-mer (K = 1, 2, 3) sequences from these three sequences.
[0047] Preferably, before step 200, which involves weighted summation of the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence, the method further includes: step 200a, which uses an attention network to map the latent vector of each K-mer sequence to a scalar score.
[0048] Step 200b: Apply the SoftMax function to calculate the weight score of the latent vector for each K-mer sequence based on the attention network.
[0049] Preferably, in step 100, the K-mer sequences of the Spacer sequence, PBS sequence, and RTT sequence are extracted respectively, wherein after K=1, 2, 3, the method further includes...
[0050] Step 101: Based on the M-nt length sequence of the Spacer sequence, the M-nt length sequence of the PBS sequence, and the M-nt length sequence of the RTT sequence, obtain a K-mer sequence of length M-K+1.
[0051] Step 102: Use word embeddings to transform each K-mer sequence into a fixed-dimensional d. model Numerical vectors;
[0052] Step 103: Positional encoding using sine and cosine functions of different frequencies is added to the word embedding.
[0053] For a sequence of length M-nt, a K-mer sequence of length M-K+1 can be obtained. Then, the system uses word embeddings to convert each K-mer sequence into a fixed-dimensional d. model Numerical vectors.
[0054] To take advantage of the spatial order of K-mer sequences, we incorporated positional encoding using sine and cosine functions with different frequencies into the word embeddings:
[0055]
[0056]
[0057] To learn more general high-level features, a Transformer encoder is used to extract the feature sequence of the target, a Transformer decoder is used to extract and encode the PBS sequence, and another Transformer decoder is used to extract and encode the RTT sequence. The Transformer encoder and decoder are suitable for variable sequence lengths, can well describe the spatial continuity of sequence bases, and possess good computational parallelism. Each Transformer encoder and decoder contains two Transformer encoding and decoding layers, and each Transformer encoding and decoding layer consists of a multi-head attention mechanism and a feedforward network. The multi-head attention mechanism in the Transformer is calculated as follows:
[0058] MultiHeadQ,K,V=Concathead1,…,head h W O
[0059] The calculation formula for feedforward networks is as follows:
[0060] FFNx=max0,xW1+b1W2+b2
[0061] Each K-mer in the obtained hidden vector sequence is normalized through a Transformer layer, where the layer normalization formula is:
[0062]
[0063] The system then uses an attention network, which is a feedforward linear layer that maps the hidden vector of each K-mer to a scalar score, and then applies the SoftMax function to compute the attention-based weight score for each K-mer.
[0064] The formula for calculating attention weights is:
[0065]
[0066] s i =MLP(h i )
[0067] Attention weights explain the importance of K-mer in sequences very well.
[0068] The sequence description vector is obtained by weighted summation of the latent vector of each K-mer sequence with its corresponding attention weight, as shown in the formula: The sequence description vector is a feature extraction and generalization of the entire subsequence.
[0069] The description vectors of all sequences corresponding to the target sequence, PBS, and RTT are concatenated to form a large description vector for each target sequence and pegRNA pair. This vector is used to encode and describe the information of the entire pegRNA and serve as the input feature for the next step.
[0070] Finally, a multi-layer fully connected feedforward network is used to perform a nonlinear transformation on the pegRNA description vector. The system uses the modified linear unit function (ReLU) as the nonlinear activation function and uses Dropout to prevent overfitting, outputting the predicted pegRNA editing efficiency.
[0071] Because this system requires learning a large number of parameters, and because there is no sufficiently large guided edit dataset covering various types and locations, it is difficult to train the optimal model parameters. Therefore, we use transfer learning to pre-train and fine-tune the model to achieve optimal training. This model can be used to predict pegRNA editing efficiency for various edit types and locations, even data from different sources. Since the Library1 dataset published by Kim et al. has sufficient data, we first train the system on the HT-trainingset of the Library1 dataset. We select the optimal hyperparameter set from numerous hyperparameter combinations based on 5x cross-validation. The number of epochs is 200, the batch size is 512, the word embedding dimension is 64, the dimension of the feedforward network model is 2048, the number of Transformer encoder sub-encoder layers is 1, the number of Transformer decoder sub-decoder layers is 1, the number of heads of the multi-head-attention model is 64, the dropout value is 0.1, and the multi-layer fully connected feedforward network consists of 3 hidden layers with 512, 64, and 512 hidden units respectively. We used mean squared error (L2 norm squared) as the loss function, with a learning rate of 0.001 and a weight decay (L2 penalty) of 0.1 for the Adam optimization algorithm. Using the optimal set of the above hyperparameters, we trained a system from scratch on the HT-trainingset with random initialization. Evaluation on the independent HT-testset yielded a Pearson correlation coefficient of 0.793, a Spearman correlation coefficient of 0.820, a mean absolute error of 3.95%, and a root mean square error of 5.59%. However, the system trained on the Library1 HT-trainingset only achieved good performance on the independent Library1 HT-testset, performing poorly on other datasets such as the Type-test and Position-test datasets, which include target sequences and various editing types (replacement, insertion, deletion, and combined editing) and editing positions.
[0072] To improve the model's generalization ability, we further fine-tuned it using transfer learning techniques. We obtained 5000 pegRNA samples with replacement from the HT-training, Type-training, and Position-trainingsets respectively, and then merged them to construct a combined training set. Instead of random initialization, we initialized it using the pre-trained optimal system previously trained on the HT-trainingset in Library1, and modified the upper fully connected feedforward network to have 6 hidden layers, containing 1,024, 2,048, 2,048, 1,024, 1,024, and 256 hidden units respectively. We optimized and adjusted the combined training set and saved the fine-tuned model as the final model for further use.
[0073] We used the preserved HT-test, Type-test, and Position-test datasets as three independent test datasets, and evaluated the final model based on the Pearson correlation coefficient (r), Spearman correlation coefficient (r), mean absolute error (MAE), and root mean square error (RMSE) between the predicted editing efficiency and the experimental actual measurement efficiency.
[0074] The model achieved the following results on the HT-test set: r = 0.769, R = 0.798, MAE = 4.28%, RMSE = 6.01%; on the Type-test set: r = 0.612, R = 0.624, MAE = 5.32%, RMSE = 6.91%; and on the Position-test set: r = 0.628, R = 0.562, MAE = 4.14%, RMSE = 5.17%. In two other cell types not used for training (HCT116 and MDA-MB-231), the results across different cell types and biological and technical replicates were as follows: HCT116, r = 0.569–0.611, R = 0.677–0.728, MAE = 8.48%–8.99%, RMSE = 11.89%–12.47%; MDA-MB-231, r = 0.636–0.663, R = 0.730–0.766, MAE = 6.36%–6.64%, RMSE = 9.05%–9.56%. These results demonstrate a significant positive correlation between predicted editing efficiency and actual experimental editing efficiency on different test datasets, with relatively small errors, indicating that the model possesses good learning and generalization abilities in predicting editing efficiency across different cell types and locations. To further evaluate the model's predictive power, we compared it with three state-of-the-art computational models: DeepPE, PE_type, and PE_position, which achieved state-of-the-art performance on the HT-test, Type-test, and Position-test datasets, respectively. We found that our predictive model consistently outperformed these three state-of-the-art models across all three test sets. For further comparisons, we developed and trained nine other different machine learning methods using the same training set and the same input features as DeepPE, PE_type, and PE_position. We then evaluated them on the same three test sets. We found that our predictive model also outperformed these nine different machine learning models across all three test sets. In conclusion, comparisons with other methods demonstrate that our predictive model is versatile and accurate in terms of pegRNA prediction efficiency.
[0075] Example 2:
[0076] This embodiment provides a guided editing efficiency prediction system based on deep learning and transfer learning, the system comprising:
[0077] The input unit is used to extract K-mer sequences from Spacer, PBS, and RTT sequences, respectively, where K = 1, 2, 3;
[0078] The first data processing unit is used to sum the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence.
[0079] The second data processing unit is used to concatenate the sequence description vectors corresponding to the Spacer sequence, the PBS sequence, and the RTT sequence to form the description vector of the pegRNA.
[0080] The prediction unit is used to perform a nonlinear transformation on the description vector of pegRNA using a multilayer fully connected feedforward network and then input it into the guide editing efficiency prediction system. The guide editing efficiency prediction system outputs the predicted editing efficiency of the pegRNA description vector.
[0081] This invention extracts K-mer sequences from Spacer, PBS, and RTT sequences, where K = 1, 2, and 3 respectively. The latent vector of each K-mer sequence is weighted and summed with its attention weight to obtain the sequence description vectors corresponding to the Spacer, PBS, and RTT sequences. These vectors are then concatenated to form the pegRNA description vector. A multilayer fully connected feedforward network is used to perform a nonlinear transformation on the pegRNA description vector, which is then input into a guided editing efficiency prediction system. This system outputs the predicted editing efficiency of the pegRNA description vector. The artificial intelligence model trained in this application can predict the efficiency of arbitrary editing unbiasedly and accurately, and provides some interpretability to the results. This effectively predicts the editing efficiency of different pegRNAs. Furthermore, based on this artificial intelligence prediction system, an optimal pegRNA system can be automatically designed, quickly and accurately obtaining the optimal design from countless possible structures to achieve the highest editing efficiency.
[0082] The specific implementation of each module in this embodiment can be found in Embodiment 1 above, and will not be repeated here. It should be noted that the device provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.
[0083] It is understood that the terms "first," "second," etc., used in the apparatus of the above embodiments can be used to describe various units, but these units are not limited by these terms. These terms are only used to distinguish the first module from another module. For example, without departing from the scope of the invention, the first transmitting module can be referred to as the second transmitting module, and similarly, the second transmitting module can be referred to as the first transmitting module. Both the first transmitting module and the second transmitting module are transmitting modules, but they are not the same transmitting module.
[0084] Example 3:
[0085] This embodiment provides a computer device, including a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the above-described guided editing efficiency prediction method based on deep learning and transfer learning.
[0086] Example 4:
[0087] This embodiment provides a storage medium, which is a computer-readable storage medium that stores a computer program. When the program is executed by a processor, the processor executes the computer program stored in the memory to implement the guided editing efficiency prediction method based on deep learning and transfer learning in Embodiment 1 above.
[0088] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.
[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A guided editing efficiency prediction method based on deep learning and transfer learning, characterized in that, The method includes: K-mer sequences were extracted from the Spacer sequence, PBS sequence, and RTT sequence, respectively, where K=1,2,3; The hidden vector of each K-mer sequence is summed with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence. The sequence description vectors corresponding to the Spacer sequence, the PBS sequence, and the RTT sequence are concatenated to form the description vector of the pegRNA. The description vector of pegRNA is nonlinearly transformed using a multilayer fully connected feedforward network and then input into a guided editing efficiency prediction system. The guided editing efficiency prediction system outputs the predicted editing efficiency of the pegRNA description vector. Before the step of weighted summation of the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence, the following steps are included: An attention network is used to map the latent vector of each K-mer sequence to a scalar score; The SoftMax function is used to compute the latent vector for each K-mer sequence based on the weight scores of the attention network; After extracting the K-mer sequences from the Spacer, PBS, and RTT sequences respectively, the process further includes: Based on the M-nt length sequences of the Spacer sequence, the M-nt length sequences of the PBS sequence, and the M-nt length sequences of the RTT sequence, a K-mer sequence of length M-K+1 is obtained. Each K-mer sequence is transformed into a fixed dimension using word embeddings. Numerical vectors; Positional encoding using sine and cosine functions of different frequencies was incorporated into the word embedding; The training process of the guided editing efficiency prediction system includes: Multiple pegRNA samples were obtained from the HT-training, Type-training, and Position-training sets using random resampling with replacement; Multiple pegRNA samples were merged together to construct a merged training set, which was then used to train a system to predict editing efficiency.
2. The guided editing efficiency prediction method based on deep learning and transfer learning according to claim 1, characterized in that, Before the step of weighted summation of the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence, the following steps are included: Extract the Spacer sequence using the Transformer encoder; The PBS sequence was extracted and encoded using a Transformer decoder; RTT sequences are extracted and encoded using a Transformer decoder; Normalize each K-mer sequence hidden by the Spacer sequence, each K-mer sequence hidden by the PBS sequence, and each K-mer sequence hidden by the RTT sequence using the Transformer.
3. The guided editing efficiency prediction method based on deep learning and transfer learning according to claim 1, characterized in that, The training process of the guided editing efficiency prediction system includes: The guided editing efficiency prediction system was initialized using the optimal guided editing efficiency prediction system that was previously pre-trained on the HT-training set in Library 1, and the upper fully connected feedforward network was modified to have multiple hidden layers.
4. The guided editing efficiency prediction method based on deep learning and transfer learning according to claim 3, characterized in that, The retained HT-test, Type-test, and Position-test datasets were used as three independent test datasets. The final guided editing efficiency prediction system was evaluated based on the Pearson correlation coefficient r, Spearman correlation coefficient R, mean absolute error MAE, and root mean square error RMSE between the predicted editing efficiency and the experimental actual measured efficiency.
5. A guided editing efficiency prediction system based on deep learning and transfer learning, used to implement the guided editing efficiency prediction method based on deep learning and transfer learning as described in any one of claims 1 to 4, characterized in that, The system includes: The input unit is used to extract K-mer sequences from Spacer, PBS, and RTT sequences, respectively. The first data processing unit is used to sum the latent vector of each K-mer sequence with the attention weight of each K-mer sequence to obtain the sequence description vector corresponding to the Spacer sequence, the sequence description vector corresponding to the PBS sequence, and the sequence description vector corresponding to the RTT sequence. The second data processing unit is used to concatenate the sequence description vectors corresponding to the Spacer sequence, the PBS sequence, and the RTT sequence to form the description vector of the pegRNA. The prediction unit is used to perform a nonlinear transformation on the description vector of pegRNA using a multilayer fully connected feedforward network and then input it into the guide editing efficiency prediction system. The guide editing efficiency prediction system outputs the predicted editing efficiency of the pegRNA description vector.
6. A computer device, characterized in that, The computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, the computer device performs the method as described in any one of claims 1 to 4.
7. A storage medium, characterized in that, A stored program, which, when executed by a processor, performs the method according to any one of claims 1 to 4.