SiRNA inhibition rate prediction method based on antisense strand multi-dimensional embedding representation

Abstract

A siRNA inhibition rate prediction method based on antisense chain multidimensional embedding representation, in the offline stage, a deep neural network is constructed and a multidimensional feature is constructed for the antisense chain sequence of siRNA for training, in the online stage, given the target gene name, all candidate siRNA sequences are obtained, and the trained deep neural network is used to predict the real-time input of the candidate siRNA sequence and sort the silence efficiency. The present application uses one-hot encoding and pre-training RNA-FM model embedding representation features as input for the antisense chain sequence of siRNA, constructs a siRNA inhibition rate prediction model in the form of a deep neural network, improves the effectiveness and accuracy of the prediction, and fuses it with the gene expression sequence search algorithm to construct a complete process from the target gene name input to the screening of efficient siRNA sequences.

Description

Technical Field

[0001] This invention relates to a technology in the field of small interfering RNA (siRNA), specifically a method for predicting siRNA inhibition rate based on multi-dimensional embedding characterization of antisense strands. Background Technology

[0002] Existing methods for determining rules in biological experiments or statistically analyzing validated datasets have significant limitations. The former is time-consuming and labor-intensive for high-throughput data, while the latter has poor generalization, with rules not being universally applicable to different research subjects. With the rise of deep learning, deep learning-based siRNA design has seen significant development. However, existing neural network-based siRNA sequence silencing efficiency prediction techniques only utilize a two-dimensional base matrix for feature representation, neglecting the correlation between different types of bases, as well as the potential impact of base thermodynamic properties, base combination paradigms at different positions, and other biological characteristics on silencing efficiency. This results in limited accuracy in predicting silencing efficiency. Furthermore, existing prediction techniques must be based on the already acquired siRNA sequence to be detected; accurate prediction is impossible if only the initial target gene sequence, or even just the target gene name, is given. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a method for predicting siRNA inhibition rate based on multi-dimensional embedding representation of the antisense strand. For the antisense strand sequence of siRNA, the method utilizes one-hot encoding and the embedding representation features of a pre-trained RNA-FM model as inputs to construct an siRNA inhibition rate prediction model using a deep neural network, thereby improving the effectiveness and accuracy of the prediction. Furthermore, this model is integrated with a gene expression sequence search algorithm to construct a complete process from inputting the target gene name to screening efficient siRNA sequences.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a method for predicting siRNA inhibition rate based on multi-dimensional embedding representation of antisense strands. In the offline stage, a deep neural network is constructed and trained using multi-dimensional features built with antisense strand sequences targeting siRNA. In the online stage, given the target gene name, all candidate siRNA sequences are obtained, and the trained deep neural network is used to predict and rank the silencing efficiency of the real-time input candidate siRNA sequences.

[0006] The aforementioned multidimensional features are obtained as follows: after collecting nucleotide information from several heterologous datasets, one-hot encoding is performed on the antisense strand AS of siRNA, and the antisense strand is embedded and characterized using a pre-trained RNA-FM model; candidate secondary structures are calculated for the AS sequence; the frequency of different bases at different positions is statistically analyzed for the AS sequence to obtain the position-specific scoring matrix (PSSM), single nucleotide frequency, dinucleotide frequency, trinucleotide frequency, GC base content and the ratio of GC continuous base fragment length, the sum of Gibbs free energies of the entire AS core fragment, and the difference in Gibbs free energies between the 5' and 3' ends.

[0007] The nucleotide information refers to: collecting Reynolds, Vickers, Haborth, Takayuki, and UI-Tei datasets, unifying the sequence length of the antisense strand to 21 nucleotides, and unifying its inhibition rate index to the range of 0-100%; preferably dividing it into training and test sets in a 9:1 ratio, and performing 10-fold cross-validation.

[0008] The position-specific scoring matrix is ​​obtained by statistically analyzing the frequency of different bases at different positions in the entire training set. During training, the PSSM score is calculated for each new siRNA input as a measure of sequence similarity with the training set.

[0009] The sum of the Gibbs free energies and the difference between the Gibbs free energies at the 5' and 3' ends are obtained by looking up the table to obtain the Gibbs free energy between each two adjacent nucleotides, and then the sum of the Gibbs free energies of the entire AS core fragment and the difference between the Gibbs free energies at the 5' and 3' ends are calculated.

[0010] The one-hot encoding described above targets the antisense strand AS of siRNA and performs one-hot encoding on different bases (A / G / C / U).

[0011] The candidate secondary structures are obtained using, but are not limited to, RNAFold calculations.

[0012] The pre-trained RNA-FM model described herein is implemented using, but is not limited to, the method described by Jiayang Chen et al. in "Interpretable RNA Foundation Model from Unannotated Data for Highly Accurate RNA Structure and Function Predictions" (arXiv:2204.00300[q-bio.QM]).

[0013] The deep neural network comprises an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The input layer extracts various multi-dimensional features from a given AS antisense strand sequence. The convolutional layers perform convolutional transformations on the one-hot encoded sequences in the input features. The pooling layers apply average pooling and max pooling to the convolution results, fusing and simplifying complex node information. The pooling results from different convolutional operations are then concatenated into a one-dimensional feature vector. The fully connected layer concatenates the pooled sequence feature vector with all other features (Gbibs free energy features, nucleotide frequency features, etc.), and sequentially passes the information to the output layer through several fully connected mappings, non-linear activations, and random deactivation operations. Finally, a sigmoid non-linear activation function is used to constrain the final output to the range of 0 to 1.

[0014] The convolutional layer consists of n one-dimensional convolutions with similar structures but different kernel sizes and dilation coefficients. Each convolution operation is followed by a non-linear activation function: Output = Relu(Conv(input)), where input is a one-hot encoding matrix, Conv is a one-dimensional convolution, and Relu is an activation function.

[0015] The information transmission is as follows: Output = Dropout(Relu(Linear(input))).

[0016] The training described uses the absolute mean error as the loss function and Pearson correlation coefficient (PCC), Spearman correlation coefficient (SPCC), and area under the receiver acceptance curve (AUC) as evaluation metrics. Multi-dimensional features are processed through different components in the deep neural network and finally fused into a single node's inhibition rate score at the output layer. The score is then compared with the true label on the entire test set to calculate the correlation coefficient and AUC.

[0017] Given a target gene name, all candidate siRNA sequences are obtained. The corresponding target gene sequence is searched in the GENOMES database. Then, a sliding window algorithm with a step size of 1 is used to sample and obtain the sense strand (SS) of the double-stranded structure of all candidate siRNA sequences. Finally, the antisense strand (AS) in the double-stranded structure is obtained according to the criterion of reverse base complementary pairing.

[0018] Technical effect

[0019] This invention, based on one-hot encoding, adds a pre-trained RNA-FM model to embed the antisense strand sequence, thus taking into account the intrinsic sequence knowledge learned by large-scale pre-trained models. It also extracts features from more dimensions, including secondary structure and Gibbs free energy, as input, and automatically learns potential paradigms through a deep neural network to select the most effective feature combinations. Tenfold cross-validation was performed on multiple heterogeneous datasets such as Reynold and Takasaki, demonstrating that the method in this embodiment achieves superior performance metrics compared to traditional approaches.

[0020] This method was compared with s-Biopredsi, DSIR, i-score methods and graph neural network (GNN) methods on the same dataset to predict siRNA silencing efficiency. The two correlation coefficients PCC and SPCC between the predicted and true values ​​were compared, as well as the AUC test under binary classification with a 70% threshold. This method improved by 10% to 25%, and improved by 5% to 10% compared with the convolutional neural network (CNN) method. Attached Figure Description

[0021] Figure 1 This is a structural diagram of a deep neural network-based siRNA inhibition rate prediction model.

[0022] Figure 2 A flowchart illustrating the design process from a given target gene name to the selection of highly efficient siRNAs;

[0023] Figure 3 Test set bin diagrams for the T and RVHU datasets. Detailed Implementation

[0024] like Figure 1 As shown in the figure, this embodiment relates to a method for predicting siRNA inhibition rate based on multi-dimensional embedding characterization of the antisense strand, which includes the following steps:

[0025] Step 1) Using the Reynolds, Vickers, Haborth, Takayuki, and UI-Tei datasets, all with a uniform length of 21 nucleotides, and with normalized inhibition rate labels ranging from 0% to 100%, the datasets were sorted in descending order based on inhibition rate and then divided into training and testing sets in a 9:1 ratio. One-hot encoding and embedded characterization of the local mRNA targeted by the siRNA were then performed, along with the calculation of its PSSM score, secondary structure, Gibbs free energy, and other features.

[0026] Step 2) Construct a deep neural network, specifically including: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer extracts and preprocesses multi-dimensional features from the antisense chain sequence. The convolutional layer uses one-dimensional convolutions with different kernel sizes (5-20). The pooling layer fuses the convolution results into a one-dimensional vector of length 30. The fully connected layer transforms the feature dimension of the nodes to 25 through three consecutive linear layers, a ReLU non-linear activation layer, and a random deactivation layer (Dropout). Finally, the output layer fuses the information from 25 nodes and outputs the inhibition rate score through the Sigmoid activation function.

[0027] Step 3) The constructed deep neural network is trained using Adam as the optimizer: the initial learning rate is 0.01, the weight decay is 5e-4, the learning rate adjustment method used is CosineAnnealingWarmRestarts, the batch size of the training set is 128, and the training process adopts an early stopping strategy using Pearson correlation coefficient (PCC) as the monitoring indicator to terminate the training early to avoid overfitting. The batch size of the test set is the total number of test samples. The loss function is the mean absolute error (MAELoss).

[0028] Step 4) The Pearson correlation coefficient (PCC), Spearman correlation coefficient (SPCC), and area under the receiver acceptance curve (AUC) are used as evaluation metrics for testing: On the test set, the correlation between the predicted scores and the true labels of all test samples is calculated. The PCC index represents the direct correlation coefficient between two sets of data, while the SPCC index represents the correlation coefficient between the sorted indices of the two sets of data. In the siRNA inhibition rate prediction task, the siRNAs in the target mRNA candidate pool are fixed; the task is to select the most inhibitory target. Therefore, the SPCC index is the most important among all evaluation metrics. The closer the PCC and SPCC values ​​are to 1, the stronger the positive correlation. The AUC index is used for binary classification tasks to evaluate the accuracy of classifying predicted objects into negative and positive samples. Both require binarizing the true labels before calculation. Based on relevant experience in this field, a 70% inhibition rate is set as the threshold for binary classification. Again, the closer the AUC value is to 1, the higher the classification accuracy and the better the classifier's performance.

[0029] Through specific practical experiments, using Ubuntu version 20.04, Linux version 5.15.0, CUDA version 11.5, and a single 3090 graphics card, the following parameters were set: Adam as the optimizer (initial learning rate 0.01, weight decay 5e-4), CosineAnnealingWarmRestarts as the learning rate adjuster, 128 as the training batch size, the total number of test set samples equal to the test batch size, PCC as the monitoring metric for early stopping, and MAELoss as the loss function.

[0030] On the Takayuki dataset (T), the results of this embodiment and the other four models on tenfold cross-validation are as follows:

[0031]

[0032]

[0033] On the Reynolds, Vickers, Haborth, and UI-Tei datasets (RVHU), the results of this embodiment and the other four models on tenfold cross-validation are as follows:

[0034]

[0035]

[0036] like Figure 3 As shown, the prediction results of the test set under the above two dataset partitions (T and RVHU) are represented by box plots, which are more intuitive and easier to understand.

[0037] This embodiment performed tenfold cross-validation on two datasets: the Takayuki dataset T, which targets only one mRNA, and the RVHU dataset, which targets multiple heterologous mRNAs and consists of Reynolds, Vickers, Haborth, and UI-Tei datasets. The average values ​​of each metric were calculated under both test set partitioning conditions. The results show that the present invention improves upon traditional feature selection-based and CNN-based algorithms in all three metrics, demonstrating that the present invention has good accuracy and generalization.

[0038] like Figure 2 As shown, after the name of the target gene is given, all candidate siRNA sequences are searched in the GENOMES database using the pyGB algorithm. Then, the trained deep neural network model is used to predict and rank the silencing efficiency of each real-time input siRNA sequence.

[0039] Compared with existing technologies, this method, in the process of converting base sequences into feature inputs that can be recognized by deep learning methods, uses a pre-trained RNA-FM model for embedded representation in addition to the commonly used one-hot encoding. The representation includes, but is not limited to, the correlation between different types of bases and different positions, and considers the potential impact of base thermodynamic properties, base combination paradigms at different positions, and other biological characteristics on silencing efficiency. In addition, this embodiment uses a relatively deep fully connected layer, supplemented by ReLU nonlinear activation and Dropout random deactivation, to fully integrate multi-dimensional features and extract effective information, while avoiding overfitting problems caused by limited data.

[0040] In addition to performance improvements, this invention establishes a complete process compared to existing technologies, from obtaining all candidate siRNA sequences from a given target gene name to further predicting high-efficiency targets, which greatly facilitates research and design in this field.

[0041] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for predicting siRNA inhibition rate based on multi-dimensional embedding characterization of antisense strands, characterized in that, In the offline stage, a deep neural network was constructed and trained using antisense strand sequences targeting siRNA to build multi-dimensional features. In the online stage, given the target gene name, all candidate siRNA sequences were obtained, and the trained deep neural network was used to predict and rank the silencing efficiency of the real-time input candidate siRNA sequences. The multi-dimensional features mentioned above are obtained by collecting nucleotide information from several heterogeneous datasets, performing one-hot encoding on the antisense strand AS of siRNA, and embedding the antisense strand using a pre-trained RNA-FM model. Candidate secondary structures were calculated for the AS sequence; the frequency of different bases at different positions in the AS sequence was statistically analyzed to obtain the position-specific scoring matrix, single nucleotide frequency, dinucleotide frequency, trinucleotide frequency, GC base content and GC continuous base fragment length ratio, the sum of Gibbs free energies of the entire AS core fragment, and the difference in Gibbs free energies between the 5' and 3' ends.

2. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1, characterized in that, The nucleotide information refers to: collecting Reynolds, Vickers, Haborth, Takayuki, and UI-Tei datasets, unifying the sequence length of the antisense strand to 21 nucleotides, unifying its inhibition rate index to the range of 0 to 100%, dividing it into training and test sets at a ratio of 9:1, and performing 10-fold cross-validation.

3. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1, characterized in that, The position-specific scoring matrix is ​​obtained by statistically analyzing the frequency of different bases at different positions in the entire training set. During training, the PSSM score is calculated for each new siRNA input as a measure of sequence similarity with the training set.

4. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1 or 2, characterized in that, The sum of the Gibbs free energies and the difference between the Gibbs free energies at the 5' and 3' ends are obtained by looking up the table to obtain the Gibbs free energy between each two adjacent nucleotides, and then the sum of the Gibbs free energies of the entire AS core fragment and the difference between the Gibbs free energies at the 5' and 3' ends are calculated.

5. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1 or 2, characterized in that, The one-hot encoding described above targets the antisense strand AS of siRNA and encodes different bases using one-hot encoding.

6. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1, characterized in that, The deep neural network comprises an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Specifically: the input layer extracts the multi-dimensional features mentioned above from a given AS antisense chain sequence; the convolutional layer performs convolutional transformations on the one-hot encoded sequence in the input features; the pooling layer applies average pooling and max pooling to the convolution results, fusing and simplifying complex node information; then, the pooling results from different convolutional operations are concatenated into a one-dimensional feature vector; the fully connected layer concatenates the pooled sequence feature vector with all other features, sequentially passing the information to the output layer through several fully connected mappings, non-linear activations, and random deactivation operations; finally, the Sigmoid non-linear activation function is used to constrain the final output to the range of 0 to 1.

7. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 6, characterized in that, The convolutional layer consists of n one-dimensional convolutions with similar structures but different kernel sizes and dilation coefficients. Each convolution operation is followed by a non-linear activation function, i.e., Output = Relu(Conv(input)), where: input is a one-hot encoding matrix, Conv is a one-dimensional convolution, and Relu is an activation function. The information is passed to the output layer, i.e., Output = Dropout(Relu(Linear(input))).

8. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1, characterized in that, The training described uses the absolute mean error as the loss function and Pearson correlation coefficient (PCC), Spearman correlation coefficient (SPCC), and area under the receiver acceptance curve (AUC) as evaluation metrics. Multi-dimensional features are processed through different components in the deep neural network and finally fused into a single node's inhibition rate score at the output layer. The score is then compared with the true label on the entire test set to calculate the correlation coefficient and AUC.

9. The siRNA inhibition rate prediction method based on multi-dimensional embedding characterization of antisense strands according to claim 1, characterized in that, Given a target gene name, all candidate siRNA sequences are obtained. The corresponding target gene sequence is searched in the GENOMES database. Then, a sliding window algorithm with a step size of 1 is used to sample and obtain the sense strand (SS) of the double-stranded structure of all candidate siRNA sequences. Finally, the antisense strand (AS) in the double-stranded structure is obtained according to the criterion of reverse base complementary pairing.

Images