Methods, models, apparatuses, devices, and media for predicting messenger RNA stability
By combining multimodal feature information prediction methods with deep learning models, the problem of low accuracy in mRNA stability prediction has been solved, improving prediction accuracy and biological interpretability, and promoting gene therapy and vaccine development.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING YUEKANGKECHUANG PHARM TECH CO LTD
- Filing Date
- 2026-02-13
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies have low accuracy in predicting mRNA stability. Traditional methods are labor-intensive and computationally intensive, making it difficult to effectively optimize mRNA design.
A multimodal feature information prediction method is adopted, including first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. These feature information are extracted and fused through a deep learning model to predict the stability of mRNA.
It significantly improves the accuracy and robustness of mRNA stability prediction, reduces prediction costs, provides biological interpretability, and enhances the effectiveness of gene therapy and vaccine development.
Smart Images
Figure CN121725892B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics, specifically to a method, model, device, equipment, and medium for predicting the stability of messenger RNA. Background Technology
[0002] With the rapid development of gene therapy and vaccine research, messenger RNA (mRNA) has shown great potential as an emerging biotherapy platform. The stability of mRNA directly affects its intracellular function and therapeutic efficacy. In recent years, advancements in high-throughput sequencing technology have enabled researchers to delve deeper into the regulatory mechanisms of RNA stability and utilize this information to optimize mRNA design. Therefore, developing accurate and effective methods for predicting mRNA sequence stability is particularly important. Summary of the Invention
[0003] This invention provides a method, model, device, equipment, and medium for predicting messenger RNA stability, in order to solve the problem of low accuracy in mRNA stability prediction.
[0004] In a first aspect, the present invention provides a method for predicting the stability of messenger RNA, the method comprising:
[0005] Obtain the target sequence information of the target messenger RNA whose stability needs to be predicted;
[0006] Using the feature extraction module in the messenger RNA stability prediction model, target feature information is obtained based on the target sequence information. The target feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count.
[0007] Using the prediction head in the messenger RNA stability prediction model, the stability of the target messenger RNA is predicted based on the target feature information.
[0008] Secondly, the present invention provides a predictive model for messenger RNA stability, comprising:
[0009] The input module is used to input the target sequence information of the target messenger RNA whose stability needs to be predicted;
[0010] The feature extraction module is used to obtain target feature information based on the target sequence information. The target feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count.
[0011] A prediction head is used to predict the stability of the target messenger RNA based on the target feature information.
[0012] Thirdly, the present invention provides a device for predicting messenger RNA stability, the device comprising:
[0013] The input unit is used to obtain the target sequence information of the target messenger RNA whose stability is to be predicted.
[0014] The feature extraction unit is used to obtain target feature information based on the target sequence information using the feature extraction module in the messenger RNA stability prediction model. The target feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count.
[0015] The prediction unit is used to predict the stability of the target messenger RNA based on the target feature information using the prediction head in the messenger RNA stability prediction model.
[0016] Fourthly, the present invention provides an electronic device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the messenger RNA stability prediction method of the first aspect or any corresponding embodiment described above.
[0017] Fifthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method for predicting messenger RNA stability according to the first aspect or any corresponding embodiment thereof.
[0018] In a sixth aspect, the present invention provides a computer program product, including computer instructions for causing a computer to execute the method for predicting messenger RNA stability according to the first aspect or any corresponding embodiment thereof.
[0019] In this embodiment of the invention, multimodal feature information is extracted from the target sequence information of the messenger RNA whose stability is to be predicted. This multimodal feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. By using multimodal feature information to predict the stability of messenger RNA, the determinants of mRNA stability can be captured from different levels, which can significantly improve the accuracy of mRNA stability prediction. Attached Figure Description
[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0021] Figure 1 This is a schematic flowchart of a first method for predicting the stability of messenger RNA according to an embodiment of the present invention;
[0022] Figure 2 This is a schematic diagram of the data preprocessing process according to an embodiment of the present invention;
[0023] Figure 3 This is a schematic diagram of the prediction process for messenger RNA stability according to an embodiment of the present invention;
[0024] Figure 4 This is a schematic diagram of the first sequence feature extraction process according to an embodiment of the present invention;
[0025] Figure 5 This is a schematic diagram of the Kozak sequence feature extraction process according to an embodiment of the present invention;
[0026] Figure 6 This is a schematic diagram of the Motif attention feature extraction process according to an embodiment of the present invention;
[0027] Figure 7 This is a schematic diagram of the manual feature vector generation process according to an embodiment of the present invention;
[0028] Figure 8 This is a schematic diagram of a second flowchart of a method for predicting the stability of messenger RNA according to an embodiment of the present invention;
[0029] Figure 9 This is a schematic diagram of the test results of the messenger RNA stability prediction model according to an embodiment of the present invention;
[0030] Figure 10 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0032] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.
[0033] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0034] RNA stability refers to the rate of RNA molecule degradation within cells, affecting the dynamic balance of gene expression. RNA concentration is determined by the dynamic balance between its synthesis (transcription) and degradation. Traditionally, RNA stability has been measured using time-series experiments or chemical inhibition experiments, but these methods are often labor-intensive and have limitations. In recent years, researchers have proposed various algorithms to estimate the relative half-life of RNA. While this method is relatively simple, it often relies on complex sequencing result processing and involves enormous computational demands.
[0035] Studies have shown that mRNA stability is closely related to its sequence characteristics. For example, RNA splicing-related features are positively correlated with RNA stability, while features related to miRNA binding and DNA methylation are negatively correlated. Furthermore, the GC content in the DNA sequence also significantly affects RNA stability; stable RNA is often rich in A+T, while unstable RNA is rich in G+C.
[0036] Structural equation modeling (SEM) allows researchers to analyze the impact of transcriptional unit characteristics on RNA stability and reveal the complex relationships between different characteristics. The application of SEM enables researchers to explore the potential determinants of RNA stability while controlling for transcriptional effects. This approach provides a new perspective for predicting RNA stability.
[0037] With the rapid development of artificial intelligence (AI) technology, related technologies are beginning to explore its application in mRNA design and optimization. Through machine learning models, researchers can analyze large amounts of RNA sequence data to predict RNA stability and translation efficiency. This approach not only improves the accuracy of predictions but also accelerates the mRNA design process, providing new tools for personalized medicine and gene therapy.
[0038] As research into RNA stability deepens, mRNA design methods incorporating AI technology will revolutionize gene therapy and vaccine development. By systematically optimizing mRNA sequences, researchers can improve mRNA stability and bioactivity, thereby enhancing its clinical efficacy. Furthermore, future research can explore the interactions between RNA stability and other biological processes, further advancing the biomedical field.
[0039] According to an embodiment of the present invention, a method for predicting the stability of messenger RNA is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.
[0040] This embodiment provides a method for predicting messenger RNA stability, which can be used in various electronic devices, such as desktop computers and mobile terminal devices. Figure 1 This is a flowchart of a method for predicting messenger RNA stability according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps:
[0041] Step S101: Obtain the target sequence information of the target messenger RNA whose stability is to be predicted.
[0042] Specifically, the target sequence information can be the original mRNA sequence string.
[0043] This embodiment can perform the following steps on the obtained raw mRNA sequence string of the target messenger RNA: Figure 2 The data preprocessing shown includes:
[0044] Encoding: The original mRNA sequence string is encoded by encoding each base in the original mRNA sequence as an integer identifier (ID) to obtain the encoded RNA sequence identifier;
[0045] Manual feature calculation: First, replace T with U in the original mRNA sequence string. Then, calculate the sequence length, GC content (the proportion of G (guanine) bases and C (cytosine) bases in the sequence) and AUG count as manual features (features extracted manually). Finally, normalize the calculated manual features.
[0046] Step S102: Using the feature extraction module in the messenger RNA stability prediction model, target feature information is obtained based on the target sequence information. The target feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, cis-acting element (Motif) attention feature information, and handcrafted feature information; the messenger RNA stability prediction model is a deep learning model.
[0047] The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module; the first sequence feature information is general sequence feature information, and the neural network module can be a convolutional neural network module (CNN).
[0048] The handcrafted feature information includes at least one of GC content, sequence length, and AUG count. The handcrafted feature information can be calculated based on the original mRNA sequence string;
[0049] For example, the feature extraction module of the messenger RNA stability prediction model includes four modules, each used to extract the four target feature information mentioned above. Among them, the neural network module is used to extract the first sequence feature information.
[0050] Step S103: Using the prediction head in the messenger RNA stability prediction model, predict the stability of the target messenger RNA based on the target feature information.
[0051] For example, such as Figure 3 As shown, the prediction head can be a multilayer perceptron (MLP), which includes one or two linear layers and a nonlinear activation function (e.g., ReLU). The final output layer is a single neuron without an activation function, used to map the target feature information to the final mRNA stability value (half-life).
[0052] In addition, such as Figure 3 As shown, before predicting the stability of the target messenger RNA based on the target feature information using the prediction head, multiple target feature information can be deeply fused to form a comprehensive feature vector. Specifically, multiple target feature information can be concatenated along the feature dimension to form the comprehensive feature vector.
[0053] The messenger RNA stability prediction method provided in this embodiment extracts multimodal feature information from the target sequence information of the messenger RNA whose stability is to be predicted. This multimodal feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. By predicting the stability of messenger RNA through multimodal feature information, the determinants of mRNA stability can be captured from different levels, which can significantly improve the accuracy of mRNA stability prediction.
[0054] Specifically, this embodiment allows the model to comprehensively understand mRNA stability from multiple dimensions, such as local sequence patterns, translation initiation efficiency, cis-acting elements, and macroscopic biological statistics, thereby improving the accuracy and robustness of predictions.
[0055] Moreover, using deep learning models to predict mRNA stability can reduce prediction costs and improve prediction efficiency.
[0056] The following examples illustrate the extraction process and structure of the extraction modules for the four types of target feature information mentioned above.
[0057] Regarding the first sequence feature information, as mentioned above, the original mRNA sequence string can be encoded to obtain the encoded RNA sequence identifier, and then the neural network module can be used to extract its features to obtain the first sequence feature information.
[0058] In some optional implementations, the neural network module for extracting the first sequence feature information can be a sequence encoder based on a convolutional neural network (SeqEncoderCNN), such as... Figure 4 As shown, it can specifically include:
[0059] The base embedding layer is used to convert the encoded target sequence information (i.e., the encoded RNA sequence identifier) into base embeddings; specifically, each base ID is mapped to a dense vector of fixed dimension, and the output base embedding sequence has the shape (Batch_Size, Max_Len, Embedding_Dim)).
[0060] One-dimensional convolutional layers are used to extract local sequence patterns from the base embeddings; specifically, multiple one-dimensional convolutional layers can be stacked, each followed by an activation function (ReLU) and a batch normalization layer. The kernel size and the number of filters can be adjusted according to the actual data to capture local sequence patterns at different scales.
[0061] The global average pooling layer is used to aggregate the output of the one-dimensional convolutional layer into a fixed-length first sequence feature information along the sequence dimension. In this embodiment, the global average pooling layer performs average pooling on the output of the last convolutional layer along the sequence dimension, aggregating the variable-length feature map into a fixed-length vector, which serves as the first sequence feature vector.
[0062] The neural network module provided in this embodiment is used to extract general and local base pattern features in mRNA sequences.
[0063] Regarding Kozak sequence feature information, the module that extracts this feature information can be a differentiable Kozak module (a type of neural network module). This differentiable Kozak module generates Kozak sequence feature information by learning position weights, identifying AUG sites (which can be called AUG promoter sites or promoter sites), and combining them with a multilayer perceptron (MLP) to predict Kozak intensity.
[0064] Some optional implementations, such as Figure 5 As shown, step S102, which is the feature extraction module in the messenger RNA stability prediction model, obtains target feature information based on the target sequence information, including:
[0065] Step S1021: Obtain a set of differentiable position weight parameters obtained from model training. The position weight parameters correspond to each position within the Kozak sequence context window. In other words, this set of position weight parameters represents the initial weights of each position within the Kozak sequence context window.
[0066] Step S1022: Apply a non-negative activation function (e.g., the Softplus function) to the position weight parameters and normalize them (e.g., normalize them using the softmax function) to ensure that the position weight parameters are non-negative and their sum is 1.
[0067] Step S1023: Identify all AUG sites in the target sequence information;
[0068] As mentioned above, the target sequence information can be the original RNA sequence string;
[0069] Step S1024: For the identified AUG sites, extract the base sequence within a context window of a preset length;
[0070] Additionally, when the window extends beyond the sequence boundary, a padding character (such as N) can be used for padding.
[0071] Step S1025: The extracted base sequence is converted into an embedding vector sequence through a base embedding layer;
[0072] Step S1026: Multiply the embedding vector sequence element-wise with the normalized position weight parameters and sum them to obtain the weighted embedding features of the AUG site.
[0073] Step S1027: Input the weighted embedded features into a multilayer perceptron (MLP) to predict the Kozak intensity score of the AUG site;
[0074] Step S1028: Average the Kozak intensity scores of the AUG sites to generate the Kozak sequence feature information. Specifically, average the Kozak intensity scores of all AUG sites in the target sequence information.
[0075] The Kozak sequence feature information obtained here is a feature vector. The Kozak sequence feature vector can be mapped to a preset output dimension using an additional multilayer perceptron (MLP).
[0076] In this embodiment, a differentiable Kozak module is introduced to quantify the importance of the Kozak sequence context by learning positional weights, and this is combined with MLP to predict the Kozak intensity at AUG sites. This design allows the model to adaptively learn the biological patterns of the Kozak sequence, rather than relying on fixed pattern matching, thus enhancing the model's flexibility. Furthermore, the differentiable positional weight learning mechanism can adaptively learn the importance of each position within the Kozak sequence context window for translation initiation efficiency.
[0077] In the above embodiments, the Kozak sequence feature information is obtained through learnable positional weight parameters. In some alternative embodiments, the Kozak sequence feature information can also be obtained through either method one or method two:
[0078] Method 1: Scoring method based on position weight matrix (PWM)
[0079] This method utilizes the known Kozak consensus sequence in bioinformatics to construct a predefined position weight matrix (PWM). Specific steps include:
[0080] Identify all AUG sites in the target sequence information;
[0081] Extract sequence fragments of a predetermined length (e.g., -6 to +4 positions) around each AUG site;
[0082] The sequence fragments are scored using a predefined PWM matrix (which records the probability scores of A, U, G, and C at various positions around the translation start site) to obtain a matching score for each AUG site;
[0083] The matching scores are statistically aggregated (e.g., the maximum value, average value, or weighted sum is taken) and used as the feature information of the Kozak sequence.
[0084] Method 2: Direct mapping method based on one-hot coding
[0085] This method does not rely on preset weights, but directly preserves sequence information for the model to learn. Specific steps include:
[0086] Identify the main AUG sites in the target sequence (such as the first AUG or the AUG with the highest predicted score).
[0087] Extract bases from specific positions upstream and downstream of the AUG site (e.g., from upstream -3 to downstream +4).
[0088] The truncated bases are subjected to one-hot encoding, and the resulting vector is directly flattened or concatenated as the feature information of the Kozak sequence.
[0089] Regarding Motif attention feature information, the module that extracts this feature information can be a Motif attention module (a type of neural network module). This module generates Motif attention features through k-mer hash embedding, multi-head self-attention, and weighted pooling mechanisms.
[0090] Some optional implementations, such as Figure 6 As shown, step S102, which is the feature extraction module in the messenger RNA stability prediction model, obtains target feature information based on the target sequence information, including:
[0091] Step S102a: Decompose the target sequence information into multiple k-mers.
[0092] As mentioned above, the target sequence information can be the original RNA sequence string; in addition, the adjacent k-mers obtained from the decomposition may or may not overlap.
[0093] Step S102b: Generate the k-mer embedding vectors corresponding to the plurality of k-mers respectively;
[0094] Specifically, each k-mer can be hashed first to map it to an integer hash identifier (ID) within a predefined range, and then the hash ID is input into the k-mer embedding layer to generate the embedding vector for each k-mer;
[0095] Step S102c: Input the embedding vectors of multiple k-clusters into a multi-head self-attention mechanism to capture the contextual dependencies between different k-clusters and generate context-aware representations of each k-cluster.
[0096] This embodiment captures the complex dependencies between k-mers from different "angles" and generates context-aware representations.
[0097] Step S102d: Input the context-aware representation into the score predictor and output the attention score of the k-cluster; the score predictor can be a small MLP;
[0098] Step S102e: Normalize the attention score to generate normalized attention weights;
[0099] For example, the softmax function can be used for normalization;
[0100] Step S102f: Use the normalized attention weights to perform weighted pooling on the k-cluster embedding vector to generate the Motif attention feature information.
[0101] In some optional implementations, the k-cluster embedding layer may use a hash bucket mechanism to map the multiple k-clusters to a fixed-size embedding space through hash modulo.
[0102] In this embodiment, a k-mer hash embedding combined with a multi-head self-attention mechanism is used to dynamically identify and weight cis-acting elements (Motifs) that have a significant impact on stability.
[0103] In this embodiment, a Motif attention module based on hash k-mer embedding and multi-head self-attention mechanism is adopted. It can not only identify important k-mer patterns in the sequence, but also understand the interaction between k-mers and their contextual importance in the sequence through the self-attention mechanism, and achieve interpretable feature extraction through attention weights.
[0104] In the above embodiments, the Motif attention feature information is obtained based on a multi-head self-attention mechanism. In some alternative embodiments, the Motif attention feature information can also be obtained through one or two methods:
[0105] Method 1: K-mer Frequency-Based Approach
[0106] This method focuses on the frequency of motif occurrences rather than contextual dependencies. The specific steps include:
[0107] Set the value of k (e.g., k=3, 4, 5);
[0108] Scan the target sequence information and count the number of times or frequency of various possible k-mers in the sequence (Term Frequency, TF).
[0109] Calculate the TF-IDF value by combining the inverse document frequency (IDF);
[0110] The statistical values of all k-clusters are combined into a feature vector, which is used as the Motif feature information.
[0111] Method 2: A method based on one-dimensional convolutional neural networks (1D-CNN)
[0112] This method utilizes convolutional kernels to automatically detect motif patterns. The specific steps include:
[0113] Base insertion is performed on the target sequence information;
[0114] The embedded sequence is convolved using multiple one-dimensional convolution kernels of different sizes (corresponding to motifs of different lengths);
[0115] Max pooling or average pooling is performed on the convolutional output to extract significant features;
[0116] The pooling result is used as a feature vector that can represent Motif information, which is the Motif attention feature information.
[0117] Regarding the handcrafted feature information, the module for extracting the feature information may include one or more linear layers and activation functions, used to map the handcrafted feature information to a preset output dimension to obtain the corresponding handcrafted feature vector.
[0118] In this embodiment, hand-crafted feature vectors (hand_feats) are first calculated based on the original mRNA sequence string. Then, these hand-crafted feature vectors are normalized. Finally, a multilayer perceptron (MLP) composed of one or more fully connected layers (linear layers) and nonlinear activation functions (such as ReLU or GELU) maps the low-dimensional hand-crafted features to a higher-dimensional feature space (e.g., ...). Figure 7 As shown in the figure, it serves as the final handcrafted feature vector.
[0119] In this embodiment, prior biological knowledge (such as GC content, length, and AUG count) is incorporated into the model and works in conjunction with features extracted by deep learning.
[0120] In summary, in this embodiment, the modules for extracting feature information of each target are independent and configurable, which facilitates the expansion and optimization of the model.
[0121] In other optional embodiments, the method for predicting messenger RNA stability also includes:
[0122] Step S201: Output explanatory information related to the stability prediction of the target messenger RNA, the explanatory information including at least one of the following:
[0123] The AUG site of the target sequence information and the Kozak intensity score corresponding to the AUG site;
[0124] The target sequence information includes the list of k-mers and the attention weights corresponding to the k-mers;
[0125] The positional weights of the Kozak sequence. Specifically, these can be normalized positional weights.
[0126] For details on how and how the explanatory information was obtained, please refer to the above text; it will not be repeated here.
[0127] To address the lack of biological interpretability in messenger RNA stability prediction schemes in related technologies, this embodiment provides a method for predicting messenger RNA stability that outputs interpretive information related to stability prediction. This includes positional weight parameters in the Kozak sequence feature extraction process, Kozak intensity scores at each AUG site, and k-mer attention weights in the Motif attention feature extraction process. This provides direct biological explanations, improving the interpretability of using messenger RNA stability prediction models and aiding in understanding the basis of model predictions. Specifically, outputting k-mer attention weights indicates which k-mers contribute most to mRNA stability prediction, helping to identify potential regulatory motifs. Outputting positional weight parameters in the Kozak sequence feature extraction process and Kozak intensity scores at each AUG site directly reveals the positions in the Kozak sequence that contribute most to translation initiation efficiency.
[0128] In this embodiment, the messenger RNA stability prediction model used to predict messenger RNA stability needs to be constructed and trained in advance to accurately predict messenger RNA stability. The training process of this messenger RNA stability prediction model is illustrated below.
[0129] The messenger RNA stability prediction model comprises four parallel feature extraction modules, a feature fusion layer, and a prediction head. The four parallel feature extraction modules extract four different target features: first sequence features, Kozak sequence features, Motif attention features, and handcrafted features. The feature fusion layer deeply fuses the four extracted target features, and the prediction head outputs an mRNA stability value based on the fused feature information. For the process of predicting messenger RNA stability using this model, please refer to [link to relevant documentation]. Figure 8 .
[0130] The training process includes:
[0131] I. Collecting training data:
[0132] Data related to mRNA stability were collected, including the ratio of PRO-Seq sequencing data to RNA-Seq sequencing data, to express the half-life of the mRNA.
[0133] II. Data Preprocessing:
[0134] 1. Encoding: The original mRNA sequence string is encoded, specifically by encoding each base in the original mRNA sequence as an integer identifier (ID), resulting in the encoded RNA sequence identifier;
[0135] 2. Manual feature calculation: First, replace T with U in the original mRNA sequence string. Then, calculate the sequence length, GC content (the proportion of G (guanine) bases and C (cytosine) bases in the sequence) and AUG count as manual features. Finally, normalize the calculated manual features.
[0136] III. Dataset Partitioning:
[0137] The preprocessed dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for learning model parameters, the validation set was used for hyperparameter tuning and early stopping, and the test set was used to evaluate the model's final generalization performance.
[0138] IV. Model Training:
[0139] 1. Hardware environment:
[0140] Model training was performed on a computing server equipped with a high-performance graphics processing unit (GPU, V100), large memory capacity, and a multi-core central processing unit (CPU). The deep learning framework used was PyTorch to support dynamic computation graphs and efficient GPU parallel computing.
[0141] 2. Loss function:
[0142] Since mRNA stability prediction is a regression task, this embodiment uses Mean Squared Error (MSE) as the loss function. MSE measures the average squared difference between the model's predicted values and the true values, aiming to minimize the prediction error.
[0143] 3. Optimizer:
[0144] This embodiment uses the Adam optimizer to update the model parameters. The Adam optimizer combines the advantages of Adagrad and RMSprop, and can adaptively adjust the learning rate of each parameter, generally exhibiting good convergence performance. The initial learning rate is set to 0.001, and the default Adam parameters are used.
[0145] 4. Training process:
[0146] Batch Processing: Training data is divided into batches of a fixed size (e.g., batch_size=64), and each batch of data is fed into the model sequentially for training.
[0147] Epochs: The model is trained iteratively for multiple epochs on the entire training set.
[0148] Forward propagation: Data from each batch is forward propagated through the model to calculate the predicted mRNA stability value.
[0149] Loss calculation: Calculate the MSE loss based on the predicted and actual values.
[0150] Backpropagation and parameter update: The gradient of the loss with respect to the model parameters is calculated using the backpropagation algorithm, and the model parameters are updated based on the gradient using the Adam optimizer.
[0151] Validation and Early Stopping: After each epoch, the model is evaluated on a validation set. If the loss on the validation set does not improve within a preset number of consecutive epochs, training is stopped to prevent overfitting. Simultaneously, the parameters of the best-performing model on the validation set are saved.
[0152] Learning Rate Scheduler: Employs a learning rate decay strategy. For example, when the validation loss does not decrease within a certain period, the learning rate is decayed by a factor to help the model escape local optima and converge more finely.
[0153] The following examples illustrate performance evaluation methods after model training, and how to utilize the model's intrinsic design for biological interpretability analysis.
[0154] 1. Model Evaluation:
[0155] Optimal model loading: Load the model parameters that perform best on the validation set.
[0156] Test set evaluation: The model is evaluated using a completely independent test set to obtain the model's generalization performance on unknown data.
[0157] Evaluation indicators:
[0158] Root Mean Squared Error (RMSE): The square root of MSE, in units consistent with the original data, represents the average deviation between the predicted and actual values. A smaller RMSE value indicates higher prediction accuracy.
[0159] Mean Absolute Error (MAE): The average of the absolute differences between predicted and actual values. It is insensitive to outliers. A smaller MAE value indicates higher prediction accuracy.
[0160] Results Analysis: Experimental results show that the messenger RNA stability prediction model proposed in this invention achieves significantly better prediction performance than existing baseline methods on the test set, with significantly reduced RMSE and MAE values, demonstrating its effectiveness and accuracy in the mRNA stability prediction task. Test results are as follows: Figure 9 As shown.
[0161] 2. Interpretability Analysis:
[0162] This invention incorporates an interpretable design that can provide valuable insights for biological research.
[0163] Kozak position weight analysis:
[0164] Extraction method: The learned normalized Kozak position weights are directly extracted from the trained differentiable Kozak module.
[0165] Analysis: These weights visually demonstrate the relative importance of each nucleotide position within the Kozak sequence context window to translation initiation efficiency. For example, key positions such as upstream and downstream of AUG show high weights, consistent with known Kozak sequence consensus, and may also reveal new important positions.
[0166] Biological significance: This study validates the biological validity of the model and may lead to the discovery of new mechanisms regulating translation initiation. The positional weights can be clearly visualized.
[0167] k-mer attention weight analysis:
[0168] Extraction method: Extract the attention weights of each k-mer from the trained Motif attention module.
[0169] Analysis content: These weights indicate which k-mer patterns in the sequence contribute most to the prediction of mRNA stability. All k-mers can be sorted according to their attention weights to identify the k-mers with the highest weights.
[0170] Biological significance: High-weighted k-mers likely represent important cis-acting elements. Analyzing these high-weighted motifs can provide biologists with new experimental directions to delve into the regulatory mechanisms of mRNA stability.
[0171] The detailed description and experimental results of the above embodiments fully demonstrate that the method described in this invention not only verifies the high accuracy of the model in predicting mRNA stability, but more importantly, through interpretable outputs such as Kozak position weights and k-mer attention weights, it provides in-depth biological insights into understanding the molecular mechanisms of mRNA stability regulation, showcasing the great application potential of this invention in biomedical research and drug development.
[0172] This embodiment also provides a predictive model for messenger RNA stability, including:
[0173] The input module is used to input the target sequence information of the target messenger RNA whose stability needs to be predicted;
[0174] The feature extraction module is used to obtain target feature information based on the target sequence information. The target feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count.
[0175] A prediction head is used to predict the stability of the target messenger RNA based on the target feature information.
[0176] For a detailed description of the structure and function of the predictive model for messenger RNA stability, please refer to the above method implementation examples; it will not be repeated here.
[0177] This embodiment also provides a device for predicting messenger RNA stability, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0178] This embodiment provides a device for predicting messenger RNA stability, comprising:
[0179] The input unit is used to obtain the target sequence information of the target messenger RNA whose stability is to be predicted.
[0180] The feature extraction unit is used to obtain target feature information based on the target sequence information using the feature extraction module in the messenger RNA stability prediction model. The target feature information includes at least two of the following: first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count.
[0181] The prediction unit is used to predict the stability of the target messenger RNA based on the target feature information using the prediction head in the messenger RNA stability prediction model.
[0182] In one optional embodiment, when the target feature information includes Kozak sequence feature information, the feature extraction unit includes a first Kozak sequence feature information extraction subunit, which is used to:
[0183] A non-negative activation function is applied to a set of differentiable position weight parameters obtained during training, and then normalized. The position weight parameters correspond to each position within the Kozak sequence context window.
[0184] Identify the AUG site in the target sequence information;
[0185] For each identified AUG site, the base sequence within a context window of a preset length is extracted.
[0186] The extracted base sequence is converted into an embedding vector sequence through a base embedding layer;
[0187] The weighted embedding features of the AUG site are obtained by multiplying the embedding vector sequence element-wise with the normalized position weight parameters and summing the results.
[0188] The weighted embedding features are input into a multilayer perceptron to predict the Kozak intensity score of the AUG site;
[0189] The Kozak intensity scores of the AUG sites are averaged to generate the Kozak sequence feature information.
[0190] In one optional embodiment, when the target feature information includes Motif attention feature information, the feature extraction unit includes a first Motif attention feature information extraction subunit, used for:
[0191] The target sequence information is decomposed into multiple k-mers;
[0192] Generate the k-mer embedding vectors corresponding to the multiple k-mers respectively;
[0193] Multiple k-clusters are embedded into a vector and input into a multi-head self-attention mechanism to capture the contextual dependencies between different k-clusters and generate context-aware representations of each k-cluster.
[0194] The context-aware representation is input into the score predictor, which outputs the attention score of the k-cluster.
[0195] The attention scores are normalized to generate normalized attention weights;
[0196] The normalized attention weights are used to perform weighted sum pooling on the k-cluster embedding vector to generate the Motif attention feature information.
[0197] In one optional implementation, the k-cluster embedding layer uses a hash bucket mechanism to map the plurality of k-clusters to a fixed-size embedding space through hash modulo.
[0198] In one optional embodiment, the neural network module for extracting the first sequence feature information includes:
[0199] A base embedding layer is used to convert the encoded target sequence information into base embeddings;
[0200] A one-dimensional convolutional layer is used to extract local sequence patterns from the base embeddings;
[0201] A global average pooling layer is used to aggregate the output of the one-dimensional convolutional layer into a fixed-length first sequence feature information along the sequence dimension.
[0202] In one optional embodiment, the messenger RNA stability prediction device further includes:
[0203] An explanatory information output unit is configured to output explanatory information related to the stability prediction of the target messenger RNA, wherein the explanatory information includes at least one of the following:
[0204] The AUG site of the target sequence information and the Kozak intensity score corresponding to the AUG site;
[0205] The target sequence information includes the list of k-clusters and the attention weights corresponding to the k-clusters;
[0206] Position weight parameters of the Kozak sequence.
[0207] In another optional embodiment, when the target feature information includes Kozak sequence feature information, the feature extraction unit includes a second Kozak sequence feature information extraction subunit, which is specifically used for:
[0208] Identify the AUG site in the target sequence information;
[0209] Extract the context sequence within a preset range of the AUG site;
[0210] Based on a predefined position weight matrix, the matching score between the context sequence and the Kozak consensus sequence is calculated;
[0211] The Kozak sequence feature information is generated based on the matching score.
[0212] In another optional embodiment, when the target feature information includes Motif attention feature information, the feature extraction unit includes a second Motif attention feature information extraction subunit, which is specifically used for:
[0213] The frequency of occurrence or TF-IDF value of k-clusters of different lengths in the target sequence information is statistically analyzed, and a k-cluster statistical vector is generated as the Motif attention feature information;
[0214] Alternatively, a one-dimensional convolutional neural network can be used to perform convolution and pooling operations on the embedded representation of the target sequence information to extract local sequence pattern features as the Motif attention feature information.
[0215] The messenger RNA stability prediction device provided in this embodiment of the invention can execute the messenger RNA stability prediction method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.
[0216] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.
[0217] The following is a detailed reference. Figure 10 This diagram illustrates a suitable structural schematic for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from memory 1008 into random access memory (RAM) 1003. The RAM 1003 also stores various programs and data required for the operation of the electronic device. The processor 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.
[0218] Typically, the following devices can be connected to the I / O interface 1005: input devices 1006 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 1007 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; memory devices 1008 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 10 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.
[0219] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 1009, or installed from a memory 1008, or installed from a ROM 1002. When the computer program is executed by the processor 1001, it performs the functions defined in the method for predicting messenger RNA stability according to embodiments of the present invention.
[0220] Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.
[0221] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the method for predicting messenger RNA stability shown in the above embodiments is implemented.
[0222] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.
[0223] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.
Claims
1. A method for predicting the stability of messenger RNA, characterized in that, The method includes: Obtain the target sequence information of the target messenger RNA whose stability needs to be predicted; Using the feature extraction module in the messenger RNA stability prediction model, target feature information is obtained based on the target sequence information. The target feature information includes first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count. The Kozak sequence feature information is determined through the following steps: applying a non-negative activation function to a set of differentiable position weight parameters obtained during training, and then normalizing the parameters. The numbers correspond to various positions within the Kozak sequence context window; AUG sites in the target sequence information are identified; for each identified AUG site, the base sequence within a context window of a preset length is extracted; the extracted base sequence is converted into an embedding vector sequence through a base embedding layer; the embedding vector sequence is multiplied element-wise with the normalized position weight parameters and summed to obtain the weighted embedding feature of the AUG site; the weighted embedding feature is input into a multilayer perceptron to predict the Kozak intensity score of the AUG site; the Kozak intensity scores of the AUG sites are averaged to generate the Kozak sequence feature information. The stability of the target messenger RNA is predicted using the prediction head in the messenger RNA stability prediction model, based on the target feature information; the prediction head is a multilayer perceptron.
2. The method according to claim 1, characterized in that, The feature extraction module in the messenger RNA stability prediction model obtains target feature information based on the target sequence information, including: The target sequence information is decomposed into multiple k-mers; Generate the k-mer embedding vectors corresponding to the multiple k-mers respectively; Multiple k-clusters are embedded into a vector and input into a multi-head self-attention mechanism to capture the contextual dependencies between different k-clusters and generate context-aware representations of each k-cluster. The context-aware representation is input into the score predictor, which outputs the attention score of the k-cluster. The attention scores are normalized to generate normalized attention weights; The normalized attention weights are used to perform weighted sum pooling on the k-cluster embedding vector to generate the Motif attention feature information.
3. The method according to claim 2, characterized in that, The k-cluster embedding layer uses a hash bucket mechanism to map the multiple k-clusters to a fixed-size embedding space through hash modulo.
4. The method according to claim 1, characterized in that, The neural network module for extracting the feature information of the first sequence includes: A base embedding layer is used to convert the encoded target sequence information into base embeddings; A one-dimensional convolutional layer is used to extract local sequence patterns from the base embeddings; A global average pooling layer is used to aggregate the output of the one-dimensional convolutional layer into a fixed-length first sequence feature information along the sequence dimension.
5. The method according to claim 1, characterized in that, Also includes: Output explanatory information related to the stability prediction of the target messenger RNA, the explanatory information including at least one of the following: The AUG site of the target sequence information and the Kozak intensity score corresponding to the AUG site; The target sequence information includes the list of k-clusters and the attention weights corresponding to the k-clusters; Position weight parameters of the Kozak sequence.
6. The method according to claim 1, characterized in that, The feature extraction module in the messenger RNA stability prediction model obtains target feature information based on the target sequence information, including: Identify the AUG site in the target sequence information; Extract the context sequence within a preset range of the AUG site; Based on a predefined position weight matrix, the matching score between the context sequence and the Kozak consensus sequence is calculated; The Kozak sequence feature information is generated based on the matching score.
7. The method according to claim 1, characterized in that, The feature extraction module in the messenger RNA stability prediction model obtains target feature information based on the target sequence information, including: The frequency of occurrence or TF-IDF value of k-clusters of different lengths in the target sequence information is statistically analyzed, and a k-cluster statistical vector is generated as the Motif attention feature information; Alternatively, a one-dimensional convolutional neural network can be used to perform convolution and pooling operations on the embedded representation of the target sequence information to extract local sequence pattern features as the Motif attention feature information.
8. A device for predicting the stability of messenger RNA, characterized in that, The device includes: The input unit is used to obtain the target sequence information of the target messenger RNA whose stability is to be predicted. The feature extraction unit is used to obtain target feature information based on the target sequence information using the feature extraction module in the messenger RNA stability prediction model. The target feature information includes first sequence feature information, Kozak sequence feature information, Motif attention feature information, and handcrafted feature information. The first sequence feature information is obtained by encoding the target sequence information and then extracting it using a neural network module. The handcrafted feature information includes at least one of GC content, sequence length, and AUG count. A prediction unit is used to predict the stability of the target messenger RNA based on the target feature information using the prediction head in the messenger RNA stability prediction model; the prediction head is a multilayer perceptron. The feature extraction unit includes a first Kozak sequence feature information extraction subunit, which is used to: apply a non-negative activation function to a set of differentiable position weight parameters obtained through training and normalize them, wherein the position weight parameters correspond to each position within the Kozak sequence context window; identify AUG sites in the target sequence information; extract base sequences within a preset length context window for each identified AUG site; convert the extracted base sequences into embedding vector sequences through a base embedding layer; multiply the embedding vector sequences element-wise with the normalized position weight parameters and sum them to obtain the weighted embedding features of the AUG sites; input the weighted embedding features into a multilayer perceptron to predict the Kozak intensity score of the AUG sites; and average the Kozak intensity scores of the AUG sites to generate the Kozak sequence feature information.
9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the method for predicting messenger RNA stability according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the method for predicting the stability of messenger RNA as described in any one of claims 1 to 7.