An RNA sequence representation enhancement method based on sequence synchronization optimization

Abstract

The application discloses an RNA sequence representation enhancement method based on sequence synchronization optimization, which comprises five processes of one-hot coding, random clipping, structure perception learning, sequence representation and hierarchical contrast loss calculation. The method can reliably represent RNA sequences and well retains feature information of the RNA sequences.

Description

Technical Field

[0001] This invention belongs to the field of biotechnology, specifically relating to a method for enhancing RNA sequence representation. Background Technology

[0002] Currently, RNA sequence analysis methods are mainly divided into two categories: traditional methods and deep learning methods. Traditional RNA sequence analysis methods rely primarily on expertise in biochemistry, including functional annotation and physicochemical properties. Specifically, these methods involve the reactivity and stability of bases in the RNA sequence. RNA sequences are characterized by constructing relevant physicochemical property matrices. Deep learning models, such as E2Efold, first perform one-hot encoding of the RNA sequence and then embedding it as a representation. The LTPonstraint model, on the other hand, uses 1, 2, 3, and 4 to encode the RNA sequence and then embeds it, obtaining a reliable representation through iterative training. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention provides an RNA sequence representation enhancement method based on sequence synchronization optimization, comprising five processes: one-hot encoding, random pruning, structure-aware learning, sequence representation, and hierarchical contrastive loss calculation. This method can reliably represent RNA sequences and effectively preserve their characteristic information.

[0004] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0005] Step 1: One-hot encoding;

[0006] For a linear RNA sequence x of length L, X is obtained after one-hot encoding:

[0007] X∈R L·4

[0008] Step 2: Randomly cut;

[0009] The encoded RNA sequence was randomly pruned to obtain S m and Sn, where S m Sn satisfies the following equation:

[0010]

[0011] Set the hyperparameter λ to restrict the minimum length of the overlapping sequence:

[0012] |Sm∩Sn≥λ|

[0013] Step 3: Structure-aware learning;

[0014] Two overlapping subsequences are mapped in high dimension through a linear layer to obtain the latent vector V of the RNA sequence; W and b represent the parameters to be optimized and the vector bias parameter, respectively.

[0015] V = WX + b

[0016] Step 4: Sequence characterization;

[0017] Each element in the potential vector V of the RNA sequence will be subjected to a probability mask operation (mask∈{0,1}).

[0018] Step 5: Calculate the loss by comparing different layers;

[0019] The stratified comparison loss calculation includes sequence consistency and fragment consistency; sequence consistency is used to compare the information loss of bases at different positions in an RNA sequence, while fragment consistency is used to compare the information loss of bases at the same position in different RNA sequences.

[0020] Preferably, step 5 specifically comprises:

[0021] Formula for calculating the similarity between bases:

[0022] Sim(γ1,γ2)=γ1γ2

[0023] Where γ1 and γ2 represent the high-dimensional vectors corresponding to the bases in the two overlapping subsequences, respectively;

[0024] Similarity measurement formula:

[0025]

[0026] Based on the above similarity measures, the sequence consistency information loss is calculated:

[0027]

[0028] Where m and n are both RNA sequence indices, and u and v are both base indices in the RNA sequence;

[0029] Calculate the fragment consistency information loss:

[0030]

[0031] Finally, the loss function combines the two types of losses mentioned above, and the total loss function is as follows:

[0032]

[0033] In the above formula, S represents the sample set, Ps represents the location set of the sample, and |S| and |PS| represent the number of elements in their respective sets.

[0034] The beneficial effects of this invention are as follows:

[0035] The linear RNA sequence characterization model constructed by the method of this invention can effectively characterize RNA sequences. Validation results show that the characterization vectors retain the characteristic information of the RNA sequence very well. Attached Figure Description

[0036] Figure 1 This is a model architecture diagram of the present invention.

[0037] Figure 2 This presents the results of a family classification experiment using the SVM classification model and the KNN clustering model to classify RNA sequence representation vectors. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] Traditional RNA sequence characterization methods have inherent limitations in capturing the complex linear relationships within RNA sequences, leading to the omission of potentially crucial information and being time-consuming, labor-intensive, and costly. Existing deep learning methods primarily focus on downstream secondary structure prediction tasks after RNA sequence characterization, exhibiting certain shortcomings in RNA sequence representation. This invention mainly utilizes the sequence information within RNA sequences, as well as the similarities and differences between RNA sequences, to achieve reliable RNA sequence characterization.

[0040] To obtain more reliable RNA sequence characterization, this invention mainly consists of five steps: one-hot encoding, random pruning, structure-aware learning, sequence characterization, and hierarchical contrastive loss calculation.

[0041] ●One-hot encoding: For a linear RNA sequence x of length L, one-hot encoding yields X:

[0042] X∈R L·4

[0043] ●Random trimming: The encoded RNA sequence is randomly trimmed to obtain S m and Sn, where S m Sn satisfies the following equation:

[0044]

[0045] In addition, the hyperparameter λ is set to limit the minimum length of the overlapping sequence:

[0046] |Sm∩Sn≥λ|

[0047] ●Structure-aware learning: Two overlapping subsequences are mapped in high dimension through a linear layer to obtain the latent vector V of the RNA sequence. W and b represent the parameters to be optimized and the vector bias parameter, respectively.

[0048] V = WX + b

[0049] ● Sequence Representation: To mitigate overfitting, noise masks are integrated into the processed RNA sequence data before the training phase. Each element in the latent vector V of the RNA sequence is passed through a probability mask mask∈

[0050] {0, 1} operation.

[0051] ● Stratified Contrast Loss Calculation: Stratified contrast loss calculation includes sequence similarity and fragment similarity. Sequence similarity is used to compare the information loss of bases at different positions in an RNA sequence, while fragment similarity is used to compare the information loss of bases at the same positions in different RNA sequences. The formula for calculating base similarity is as follows:

[0052] Sim(γ1,γ2)=γ1γ2

[0053] Where γ1 and γ2 represent the high-dimensional vectors corresponding to the bases in the two overlapping subsequences. The general similarity measurement formula is given below:

[0054]

[0055] Based on the above similarity measures, the sequence consistency information loss is calculated:

[0056]

[0057] Where m is the RNA sequence index, and u and v are the base indices in the RNA sequence, respectively.

[0058] Calculate the fragment consistency information loss:

[0059]

[0060] Ultimately, the loss function should take into account both types of loss mentioned above; therefore, the total loss function is as follows:

[0061]

[0062] In the above formula, S represents the sample set, and Ps represents the location set of the sample. |S| and |PS| represent the number of elements in their respective sets.

[0063] Example:

[0064] The model architecture of this invention is as follows: Figure 1As shown: First, the RNA sequence is one-hot encoded, then randomly pruned. The resulting overlapping subsequences are then passed through a structure-aware layer and a latent representation layer, followed by hierarchical contrastive loss calculation. After model training, the random pruning, the masking module in the latent representation layer, and the contrastive loss module are removed. Inputting the RNA sequence yields the corresponding representation vector.

[0065] In the experiment of the deep learning-based RNA sequence representation model, the obtained representation vectors were used to perform RNA family classification experiments using an SVM classification model. The model's performance was evaluated based on the classification accuracy, and the AUPRC value was used as an evaluation of the model's reliability. The specific steps are as follows:

[0066] 1) Sequences with a length of 600 or less from the RNAStralign dataset were selected for the experiment. These sequences came from eight RNA families and totaled 20,923 sequences.

[0067] 2) RNA sequences were divided into three groups according to length: 1-128, 129-512, and 1-600, and experiments were conducted separately for each group. For each group, the representation model was first trained using the full dataset. After training, the sequence was represented using the full dataset to obtain the representation vector.

[0068] 3) The representation vectors are randomly divided into a training set and a test set in a ratio of 2:8;

[0069] 4) The training set is used to train the SVM classification model. After the model is trained, the test set is used to evaluate the classification model.

[0070] 5) Calculate the classification accuracy by comparing the prediction results and labels on the test set; at the same time, calculate the AUPRC value to reflect the reliability of the model.

[0071] AUPRC stands for the area under the precision-recall curve. It is a metric used to evaluate the reliability of a model in classification problems, and is particularly suitable for reflecting the reliability of classification results on imbalanced datasets. Since the dataset used to verify this invention is an imbalanced dataset, this metric is more effective in reflecting the reliability of the invention.

[0072] To further verify the practicality of this invention, the deep learning-based RNA secondary structure prediction model LTPConstraint was used. The representation module was added to LTPConstraint and compared with the original model. Performance was measured by the F1 score of the downstream RNA secondary structure prediction. The results further demonstrate that this invention can improve the accuracy of RNA secondary structure prediction. The specific steps are as follows:

[0073] 1) Select sequence data from three RNA families: 5S rRNA, SRP, and PDB. These data were obtained from the bpRNA-1m dataset.

[0074] 2) Use the CD-HIT-EST tool to perform redundancy removal, removing sequence data with a redundancy rate exceeding 80%;

[0075] 3) Select the Rfam dataset from bpRNA-1m and perform step 2) in the same way;

[0076] 4) The processed 5S rRNA, SRP, and PDB family data were randomly divided into training and testing sets at a ratio of 8:2.

[0077] 5) Use the dataset obtained in step 3) for pre-training;

[0078] 6) Use the training set obtained in step 4) to train the model and obtain the prediction model;

[0079] 7) Use the test set obtained in step 4) and the prediction model obtained in step 6) to make predictions and calculate the F1 score;

[0080] 8) Add the RNA sequence characterization module of this invention to the original LTPonstraint model, and repeat steps 5), 6), and 7);

[0081] 9) The reliability of the invention is verified by comparing the F1 values ​​of the two tests.

[0082] The results in Table 1 show the results of family classification experiments using the SVM classification model on the RNA sequence representation vectors obtained through this invention. The results indicate that the accuracy is close to 1.0, and the AUPRC values ​​are all above 0.9, demonstrating the reliability of this invention.

[0083] Table 1

[0084]

[0085] Reference Figure 2 The results are from experiments using SVM and KNN clustering models to classify RNA sequence representation vectors into families. The results show that the KNN clustering model can still achieve very accurate family classification of representation vectors, further demonstrating the reliability of this invention.

[0086] Table 2 compares the experimental results of the deep learning-based RNA secondary structure prediction model LTPConstraint with and without this invention. The results show that using this invention can improve the F1 score of the LTPConstraint model.

[0087] Table 2

[0088]

[0089] The above results all demonstrate that the RNA sequence representation enhancement technology based on sequence synchronization optimization constructed by the method of the present invention can reliably characterize RNA sequences.

Claims

1. A method for enhancing RNA sequence representation based on sequence synchronization optimization, characterized in that, Includes the following steps: Step 1: One-hot encoding; For a linear RNA sequence x of length L, X is obtained after one-hot encoding: Step 2: Randomly cut; The encoded RNA sequence was randomly pruned to obtain and ,in and Satisfy the following formula: Set hyperparameters Limit the minimum length of overlapping sequences: Step 3: Structure-aware learning; Two overlapping subsequences are mapped in high dimension through a linear layer to obtain the latent vector V of the RNA sequence; W and b represent the parameters to be optimized and the vector bias parameter, respectively. +b Step 4: Sequence characterization; Each element in the latent vector V of the RNA sequence will be passed through a probability mask. operate; Step 5: Calculate the loss by comparing different layers; Hierarchical comparison loss calculation includes sequence consistency and fragment consistency; sequence consistency is used to compare the information loss of bases at different positions in an RNA sequence, while fragment consistency is used to compare the information loss of bases at the same position in different RNA sequences. Step 5 specifically involves: Formula for calculating the similarity between bases: in and These represent the high-dimensional vectors corresponding to the bases in the two overlapping subsequences; Similarity measurement formula: Based on the above similarity measures, the sequence consistency information loss is calculated: Where m and n are both RNA sequence indices, and u and v are both base indices in the RNA sequence; Calculate the fragment consistency information loss: Finally, the loss function combines the two types of losses mentioned above, and the total loss function is as follows: In the above formula, S represents the sample set. This represents the location set of the sample; and These represent the number of elements in their respective sets.

Images