A siRNA screening method and device based on deep learning

Abstract

The application relates to the technical field of deep learning, and discloses a siRNA screening method and device based on deep learning, which comprises the following steps: constructing a siRNA inhibition rate prediction model; obtaining a preliminary siRNA screening result corresponding to a target mRNA sequence to be predicted based on the siRNA inhibition rate prediction model; performing molecular dynamics simulation on the target mRNA sequence to be predicted to generate a molecular dynamics trajectory corresponding to the target mRNA sequence to be predicted; constructing a neural relation reasoning model to generate an interaction relation graph corresponding to the target mRNA sequence to be predicted; and performing secondary screening on the preliminary siRNA screening result based on the interaction relation graph to obtain a final siRNA screening result. The application effectively improves the accuracy of a deep learning model in a real screening task and avoids problems such as weak generalization ability of the deep learning model in screening siRNA due to the bias of a training set.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and particularly relates to a siRNA screening method and device based on deep learning. BACKGROUND

[0002] RNA interference (RNAi), also known as post-transcriptional gene silencing, is a mechanism that can resist parasitic and pathogenic nucleic acids and regulate the expression of specific genes, and can manually regulate gene expression and detect gene function. The regulatory factors of RNAi include small interfering RNA (siRNA) of 21 to 23 nucleotides and microRNA (miRNA) of about 22 nucleotides. Among them, the antisense strand (AS) in the siRNA duplex binds to the target mRNA through Watson-Crick base pairing. If the AS completely hybridizes with the target mRNA, it will cause the target mRNA to be cut and prevent translation; if only the seed region of the AS binds to the target mRNA, it will induce the target mRNA to be degraded and inhibit translation.

[0003] The target mRNA is generally composed of hundreds to thousands of nucleotides, and a large number of candidate siRNAs can be generated using the sliding window method. However, the generated siRNAs have significant differences in inhibition rates, and their compositions also differ. The siRNA inhibition rate is mainly affected by the sequence pattern, binding affinity, and secondary structure around the binding region. It is very difficult to predict the siRNA inhibition rate, so researchers have been exploring models for accurately predicting the siRNA inhibition rate.

[0004] The existing siRNA inhibition rate prediction models mainly include expert experience-based methods, machine learning-based methods and deep learning-based methods. The expert experience-based method formulates a set of siRNA design principles based on the design experience and statistical analysis of experts, which is usually suitable for the case of less siRNA data. For example, Reynolds et al. summarized some principles for rationally designing siRNA by using statistical methods for 180 siRNAs designed for two genes: the GC content is between 0.30 and 0.52; there are at least 3 A or U at the 15th-19th positions of the sense strand; avoid structures that can form hairpin loops; the 19th position of the sense strand is A, the 3rd position is A, the 10th position is U, the 19th position is not G or C, and the 13th position is not G. The machine learning-based and deep learning-based methods are data-driven, and features are extracted from the siRNA data set by machine learning and deep learning algorithms to predict the inhibition rate of siRNA. The machine learning-based method needs to manually extract features, which inevitably causes errors due to feature extraction; and the deep learning-based method does not need human intervention in feature extraction. In recent years, graph convolutional neural networks, graph neural networks and Transformer models have been used for siRNA inhibition rate prediction.

[0005] Although the deep learning method has made good progress in siRNA inhibition rate prediction, the deep learning method is extremely dependent on the quality of the data set. When there is a large difference between the distribution of siRNA data in the real screening task and the distribution of siRNA data in the training set, the deep learning model is difficult to accurately screen the results.

[0006] And in the actual application scenario, there is often a large difference between the distribution of siRNA data in the real screening task and the distribution of siRNA data in the training set. SUMMARY

[0007] Therefore, the present application provides a siRNA screening method and device based on deep learning to solve the problem that the existing siRNA prediction method based on deep learning relies too much on the model training data set.

[0008] In a first aspect, the present application provides a siRNA screening method based on deep learning, which comprises:

[0009] constructing an siRNA inhibition rate prediction model;

[0010] obtaining a preliminary siRNA screening result corresponding to the target mRNA sequence to be predicted based on the siRNA inhibition rate prediction model;

[0011] performing molecular dynamics simulation on the mRNA sequence of the target to be predicted, to generate a molecular dynamics trajectory corresponding to the mRNA sequence of the target to be predicted;

[0012] constructing a neural relation inference model to generate an interaction relation graph corresponding to the mRNA sequence of the target to be predicted;

[0013] based on the interaction relation graph, re-screening the preliminary siRNA screening result to obtain a final siRNA screening result.

[0014] The application greatly improves the accuracy of siRNA screening by constructing an interaction relation graph between target mRNA residues, improves the accuracy of deep learning models in real screening tasks, and effectively avoids problems such as weak generalization ability of deep learning models when screening siRNA due to bias of training sets.

[0015] In an optional embodiment, constructing an siRNA inhibition rate prediction model comprises:

[0016] obtaining an siRNA data set, the siRNA data set comprising: an siRNA sequence, a target mRNA sequence corresponding to the siRNA sequence, and an inhibition rate of the target mRNA sequence corresponding to the siRNA sequence pair;

[0017] obtaining a thermodynamic property feature and an expert experience feature corresponding to the siRNA sequence;

[0018] splicing the siRNA sequence in the siRNA data set and a local target mRNA sequence corresponding to the siRNA sequence, and adding a corresponding position code to the spliced sequence to obtain a coded spliced sequence; wherein the local target mRNA sequence is a sequence matched with the siRNA sequence plus a preset number of residue sequences before and after the siRNA sequence is matched;

[0019] inputting the coded spliced sequence into a pre-constructed feature network structure for feature extraction to obtain a sequence feature;

[0020] training a preset fully connected neural network based on the inhibition rate of the target mRNA sequence corresponding to the siRNA sequence pair, the sequence feature, the thermodynamic property feature, and the expert experience feature to obtain the siRNA inhibition rate prediction model after training.

[0021] In this embodiment, local sequence splicing and position coding are used to enhance feature representation, effectively improving the accurate prediction of the model on the siRNA inhibition rate. At the same time, the combination of thermodynamic property features and expert experience features enables the model to more comprehensively capture the interaction relationship between siRNA and target mRNA, effectively improving the reliability and effectiveness of the model prediction.

[0022] In an optional implementation, the preliminary siRNA screening result corresponding to the target mRNA sequence to be predicted is obtained based on an siRNA inhibition rate prediction model, and the preliminary siRNA screening result comprises:

[0023] An antisense sequence of the siRNA sequence of the target mRNA sequence to be predicted is constructed;

[0024] The antisense sequence of the siRNA sequence is input into the siRNA inhibition rate prediction model to obtain an siRNA predicted inhibition rate corresponding to the siRNA sequence;

[0025] The siRNA predicted inhibition rates are sequentially sorted from high to low;

[0026] The siRNA sequences of a preset proportion are sequentially selected as the preliminary siRNA screening result from the highest siRNA predicted inhibition rate.

[0027] In the embodiment, the siRNA antisense sequence corresponding to the target mRNA sequence is constructed, and the inhibition rate is accurately predicted by the siRNA inhibition rate prediction model, so that the siRNA with the best inhibition effect can be effectively screened.

[0028] In an optional implementation, the antisense sequence of the siRNA sequence of the target mRNA sequence to be predicted comprises:

[0029] The 5' end of the target mRNA sequence is taken as a starting point, and translation is sequentially performed to the 3' end until the end of the antisense sequence of the siRNA sequence is flush with the 3' end of the target mRNA; and the length of the antisense sequence of the siRNA sequence ranges from 19 nt to 21 nt.

[0030] In the embodiment, the siRNA antisense sequence is matched by translation from the 5' end to the 3' end of the local target mRNA sequence and by matching siRNA antisense sequences of different lengths, so that the method can comprehensively evaluate multiple binding sites on the mRNA, and the design flexibility and prediction accuracy of the siRNA are improved.

[0031] In an optional implementation, the neural relationship inference model is constructed based on a variational autoencoder model, the neural relationship inference model is constructed to generate an interaction relationship graph corresponding to the target mRNA sequence to be predicted, and the method comprises:

[0032] Node data corresponding to a molecular dynamics trajectory is obtained, the node is a residue in the target mRNA sequence to be predicted, and the node data comprises spatial coordinates and velocity of the node at multiple time steps;

[0033] The node data is input into an encoder structure in the neural relationship inference model to extract residue features and generate hidden variables based on the residue features;

[0034] The decoder structure in the neural relational inference model is input with the latent variable, a molecular dynamics trajectory is reconstructed, a reconstruction error is calculated according to the reconstructed molecular dynamics trajectory and the original molecular dynamics trajectory, the encoder parameters, the latent variable and the decoder parameters are updated by back propagation according to the reconstruction error, and the training is completed after multiple iterations.

[0035] The neural relational inference model after training is obtained, and the latent variable generated by the neural relational inference model after training is the interaction relationship graph corresponding to the target mRNA sequence to be predicted.

[0036] In the embodiment, the variational autoencoder structure can effectively extract the features of the residues in the molecular dynamics trajectory, capture complex molecular interaction information, and further improve the accuracy of siRNA screening.

[0037] In an optional embodiment, based on the interaction relationship graph, the preliminary siRNA screening result is screened again to obtain a final siRNA screening result, including:

[0038] Based on the adjacency matrix included in the interaction relationship graph, the degree of each node is determined, wherein the degree is the number of nodes connected to other nodes;

[0039] A relationship curve graph of nodes and degrees is generated, wherein the abscissa of the relationship curve graph is the residue number, and the ordinate is the degree;

[0040] It is judged whether the residue number of the target mRNA sequence corresponding to the siRNA sequence in the preliminary siRNA screening result is within the residue number range corresponding to the trough region in the relationship curve graph;

[0041] If it is within the residue number range corresponding to the trough region in the relationship curve graph, the corresponding siRNA sequence is retained;

[0042] If it is not within the residue number range corresponding to the trough region in the relationship curve graph, the corresponding siRNA sequence is deleted from the preliminary siRNA screening result.

[0043] In the embodiment, the neural relational inference model and the siRNA inhibition rate prediction model are combined, which can further improve the accuracy of siRNA screening based on the interaction relationship graph.

[0044] In a second aspect, the present application provides a siRNA screening device based on deep learning, which comprises:

[0045] The construction module is used for constructing an siRNA inhibition rate prediction model.

[0046] The siRNA inhibition rate prediction module is configured to obtain a preliminary siRNA screening result corresponding to the target mRNA sequence to be predicted based on an siRNA inhibition rate prediction model.

[0047] The molecular dynamics trajectory generation module is configured to perform molecular dynamics simulation on the target mRNA sequence to be predicted to generate a molecular dynamics trajectory corresponding to the target mRNA sequence to be predicted.

[0048] The relationship graph generation module is configured to construct a neural relationship reasoning model to generate an interaction relationship graph corresponding to the target mRNA sequence to be predicted.

[0049] The screening module is configured to perform re-screening on the preliminary siRNA screening result based on the interaction relationship graph to obtain a final siRNA screening result.

[0050] In a third aspect, the present application provides a computer device, comprising a memory and a processor, the memory and the processor are communicatively connected with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the siRNA screening method based on deep learning of the first aspect or any of the corresponding embodiments thereof.

[0051] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are used to make a computer execute the siRNA screening method based on deep learning of the first aspect or any of the corresponding embodiments thereof.

[0052] In a fifth aspect, the present application provides a computer program product, which comprises computer instructions, and the computer instructions are used to make a computer execute the siRNA screening method based on deep learning of the first aspect or any of the corresponding embodiments thereof.

[0053] It should be noted that the siRNA screening device based on deep learning, the computer device, the computer readable storage medium and the program product provided by the present application correspond to the siRNA screening method based on deep learning described above. Therefore, the beneficial effects of the siRNA screening device based on deep learning, the computer device, the computer readable storage medium and the program product will be described in the corresponding beneficial effects of the siRNA screening method based on deep learning above, and will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the description of the specific embodiments or the prior art. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0055] Figure 1 is a flowchart of a deep learning-based siRNA screening method according to an embodiment of the present application;

[0056] Figure 2 is a flowchart of another deep learning-based siRNA screening method according to an embodiment of the present application;

[0057] Figure 3 is a training flowchart of a Transformer deep learning model according to an embodiment of the present application;

[0058] Figure 4 is a local target mRNA selection diagram according to an embodiment of the present application;

[0059] Figure 5 is a training flowchart of a neural relation inference model according to an embodiment of the present application;

[0060] Figure 6 is a distribution diagram of target mRNA residue degree according to an embodiment of the present application;

[0061] Figure 7 is a structural block diagram of a deep learning-based siRNA screening device according to an embodiment of the present application;

[0062] Figure 8 is a hardware structure diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0064] According to an embodiment of the present application, a siRNA screening method based on deep learning is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0065] In this embodiment, a siRNA screening method based on deep learning is provided, which can be executed by a server, a terminal, a mobile terminal, and the like, Figure 1 is a flowchart of the siRNA screening method according to an embodiment of the present application, as shown in Figure 1 The flowchart includes the following steps:

[0066] Step S101, constructing a siRNA inhibition rate prediction model. In this embodiment, the siRNA inhibition rate prediction model includes but is not limited to a convolutional neural model, a graph neural network model, and a Transformer model.

[0067] Step S102, obtaining a preliminary siRNA screening result corresponding to the target mRNA sequence to be predicted based on the siRNA inhibition rate prediction model.

[0068] Step S103, performing molecular dynamics simulation on the target mRNA sequence to be predicted to generate a molecular dynamics trajectory corresponding to the target mRNA sequence to be predicted. Specifically, molecular dynamics simulation software can be used to perform molecular dynamics simulation on the target mRNA sequence to obtain the dynamics trajectory of the target mRNA, wherein the molecular dynamics simulation should use a full-atom dynamics simulation method, and the molecular dynamics simulation software includes but is not limited to Gromacs, Amber, etc. In this embodiment, Gromacs molecular dynamics software is used to perform molecular dynamics simulation on the target mRNA sequence, wherein the simulation force field is CHARMM27 force field, the water model is TITP3, the target mRNA molecule is placed in a cubic water box and solvated, the system charge is neutralized by adding chloride and sodium ions. The energy of the system is initially minimized using the steepest descent minimization method, and the minimum is 50000 steps, and then the position of the heavy atom is limited to balance the system. The system can be first heated and balanced, and the Langevin temperature control method is used to balance for 100 ps, then the pressure is increased, and the isotropic Berendsen pressure control method is used for pressure balancing, and the simulation time is 100 ps in total, and finally the molecular dynamics simulation is completed at a temperature of 300 K for 50 ns.

[0069] Step S104, constructing a neural relation inference model. The neural relation inference model in the embodiment is constructed using a variational autoencoder model. The encoder of the variational autoencoder model extracts features according to the input target mRNA molecular dynamics trajectory, and generates hidden variables. Then, the decoder reconstructs the dynamics trajectory of the target mRNA molecule according to the hidden variables, and calculates the reconstruction error according to the reconstructed trajectory and the original trajectory. The reconstruction error is back-propagated to update the encoder parameters, hidden variables and decoder parameters. After multiple iterations, the training is completed. The model after training generates hidden variables corresponding to the interaction relationship graph of the target mRNA sequence to be predicted.

[0070] Step S105, based on the interaction relationship graph, the preliminary siRNA screening result is screened again to obtain the final siRNA screening result.

[0071] The interaction relationship graph contains the interaction between each residue. The siRNA sequence screened preliminarily can be screened again according to the residue characteristics in the interaction relationship graph, so as to determine the final screening result.

[0072] The present application provides a virtual screening method for improving the siRNA inhibition rate of a deep learning model based on a neural relation inference model, as shown in Figure 2 The preliminary siRNA screening result is screened according to the siRNA inhibition rate prediction model; the molecular dynamics trajectory of the target mRNA is obtained by using molecular dynamics simulation software to simulate the molecular dynamics of the target mRNA; the molecular dynamics trajectory is reconstructed by using the neural relation inference model, and the interaction relationship graph of the target mRNA is obtained; finally, the siRNA is screened based on the preliminary siRNA screening result by using the interaction relationship graph of the target mRNA obtained based on the neural relation inference model. The present application greatly improves the accuracy of siRNA screening by constructing the interaction relationship graph between the residues of the target mRNA, improves the accuracy of the deep learning model in the real screening task, and effectively avoids the problem that the deep learning model has weak generalization ability when screening siRNA due to the bias of the training set.

[0073] In some optional embodiments, the siRNA inhibition rate prediction model is established by the following steps:

[0074] The siRNA data set is obtained, which includes: siRNA sequence, target mRNA sequence corresponding to the siRNA sequence, and inhibition rate of the target mRNA sequence corresponding to the siRNA sequence.

[0075] The thermodynamic property characteristics and expert experience characteristics corresponding to the siRNA sequence are obtained.

[0076] The siRNA sequence in the siRNA dataset is spliced with the local target mRNA sequence corresponding to the siRNA sequence, and the corresponding position code is added to the spliced sequence to obtain a coding spliced sequence; wherein the local target mRNA sequence is a sequence matched with the siRNA sequence plus a preset number of residue sequences before and after the siRNA sequence is matched.

[0077] The coding spliced sequence is input into a pre-constructed feature network structure for feature extraction to obtain sequence features.

[0078] A preset fully connected neural network is trained based on the inhibition rate of the siRNA sequence on the corresponding target mRNA sequence, sequence features, siRNA thermodynamic property features, and expert experience features to obtain a trained siRNA inhibition rate prediction model.

[0079] Before constructing the siRNA inhibition rate prediction model, an siRNA dataset needs to be constructed first, which should include siRNA sequences, mRNA target points corresponding to the siRNA, and inhibition rates of the siRNA corresponding target points. In this embodiment, the siRNA data measured by Huesken, Reynolds, and others can be used as the dataset. After processing, the collected siRNA inhibition rate data has a total of 2816 entries. The siRNA dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1 using a random division method.

[0080] Further, a siRNA inhibition rate deep learning prediction model is constructed, which is trained according to the siRNA dataset. The trained model can be used to predict the inhibition rate of a new siRNA. In this embodiment, a siRNA inhibition rate deep learning prediction model is constructed based on a Transformer model, as shown in Figure 3 The input of the model includes two parts, the sequence of the siRNA and the sequence of the local target mRNA. The sequence of the local target mRNA is a sequence matched with the siRNA plus a sequence of 20 residues before and after the matching, as shown in Figure 4In addition, not all local target mRNA sequences can be matched with siRNA to cut 20 residues in front and back, and for those that cannot be cut, fill in the characters "0". Before inputting the Transformer encoder, the siRNA sequence is spliced with the local target mRNA sequence, and then the position encoding is added to the spliced sequence. In this embodiment, the Transformer encoder has two layers, and after extracting the features by the Transformer encoder, a linear layer is used to change the dimension of the feature vector, and then the feature vector is unfolded by the unfolding layer. In addition, the model also integrates the features of the thermodynamic properties of siRNA and the features of expert experience, i.e. the expert experience thermodynamic feature, and inputs the two features and the inhibition rate of the siRNA sequence to the corresponding target mRNA sequence, sequence features into a fully connected neural network for further training. The siRNA inhibition rate prediction model after training can be used to predict the inhibition rate of siRNA and output.

[0081] In this embodiment, the siRNA sequence and the local target mRNA sequence are spliced and position encoded to enhance the feature representation, which effectively improves the accurate prediction of the siRNA inhibition rate by the model. At the same time, by combining the thermodynamic property features and the expert experience features, the model can more comprehensively capture the interaction between siRNA and target mRNA, effectively improving the reliability and effectiveness of the model prediction.

[0082] In some optional embodiments, based on the siRNA inhibition rate prediction model, a preliminary siRNA screening result corresponding to a to-be-predicted target mRNA sequence is obtained, including:

[0083] An antisense sequence of the siRNA sequence of the to-be-predicted target mRNA sequence is constructed.

[0084] The antisense sequence of the siRNA sequence is input into the siRNA inhibition rate prediction model to obtain the siRNA prediction inhibition rate corresponding to the siRNA sequence.

[0085] The siRNA prediction inhibition rates are sequentially sorted from high to low.

[0086] Starting from the highest siRNA prediction inhibition rate, a preset proportion of siRNA sequences are sequentially selected as the preliminary siRNA screening result.

[0087] In this embodiment, the antisense sequence of the siRNA sequence of the target mRNA is constructed according to the principle of base mutual matching. When constructing, the 5' end of the target mRNA is taken as the starting point, and translation is performed to the 3' end in turn until the end of the antisense sequence of the siRNA sequence is flush with the 3' end of the target mRNA, and the length of the antisense sequence of the siRNA sequence should be 19 nt-21 nt. Then the constructed antisense sequence of the siRNA sequence is input into the constructed siRNA inhibition rate deep learning prediction model to predict the inhibition rate. Finally, the siRNA can be sorted in descending order of the predicted siRNA inhibition rate, for example, the top 10% of siRNA is selected. In this embodiment, 702 siRNAs are constructed, and 70 siRNAs are obtained after screening by the siRNA inhibition rate deep learning model.

[0088] In this embodiment, by constructing the siRNA antisense sequence corresponding to the target mRNA sequence and accurately predicting the inhibition rate by the siRNA inhibition rate prediction model, siRNAs with the best inhibition effect can be effectively screened out.

[0089] In some optional embodiments, the antisense sequence of the siRNA sequence of the target mRNA sequence to be predicted is constructed, including:

[0090] The 5' end of the target mRNA sequence is taken as the starting point, and translation is performed to the 3' end in turn until the end of the antisense sequence of the siRNA sequence is flush with the 3' end of the target mRNA; the length of the antisense sequence of the siRNA sequence ranges from 19 nt to 21 nt.

[0091] In this embodiment, by translating from the 5' end of the local target mRNA sequence to the 3' end and matching siRNA antisense sequences of different lengths, this method can comprehensively evaluate multiple binding sites on the mRNA, improving the design flexibility and prediction accuracy of siRNA.

[0092] In some optional embodiments, the neural relation inference model is constructed based on a variational autoencoder model, and the molecular dynamics trajectory is reconstructed based on the neural relation inference model to generate an interaction relationship graph corresponding to the target mRNA sequence to be predicted, including:

[0093] The node data corresponding to the molecular dynamics trajectory is obtained, the node being a residue in the target mRNA sequence to be predicted, and the node data including the spatial coordinates and velocity of the node at multiple time steps.

[0094] The node data is input into the encoder structure in the neural relation inference model, residue features are extracted, and hidden variables are generated based on the residue features.

[0095] The latent variables are input into the decoder structure of the neural relation inference model to reconstruct the molecular dynamics trajectory. The reconstruction error is calculated based on the reconstructed molecular dynamics trajectory and the original molecular dynamics trajectory. The encoder parameters, latent variables and decoder parameters are updated by backpropagation of the reconstruction error. The training is completed after multiple iterations.

[0096] Obtain the neural relationship inference model after training. The latent variables generated by the neural relationship inference model after training are the interaction relationship diagrams corresponding to the mRNA sequences of the target to be predicted.

[0097] In this embodiment, the target mRNA sequence to be predicted is taken as EGFP as an example. Molecular dynamics simulations of the target mRNA are performed using Gromacs molecular dynamics software, and the molecular dynamic trajectory is reconstructed using a neural relationship inference model to obtain the interaction relationship diagram of the target mRNA. (Refer to...) Figure 5 The diagram illustrates a target-relationship inference model constructed using a variational autoencoder (VAE) model. The encoder of the VAE model extracts features from the input target mRNA molecule's dynamic trajectory and generates latent variables. The decoder then reconstructs the target mRNA molecule's dynamic trajectory based on these latent variables. Subsequently, by analyzing the reconstructed trajectory, an interaction map between target mRNA molecule residues is generated.

[0098] Specifically, the dynamic trajectory of the input target mRNA consists of N nodes, and the feature vector of each node is composed of its position and velocity in the x, y, and s dimensions. Let i represent the characteristic of node i at time t. Therefore, the characteristic of all nodes at time t can be represented as follows: The trajectory of node i can be represented as Where T represents the number of time steps. The neural relation inference model learns the interaction graph of target mRNAs and reconstructs the future trajectory of the target mRNA dynamic system based on the interaction graph in an unsupervised manner. The interaction between node i and node j takes the form of the latent variable z. i,j ∈{1,2,…,K}, where K is the number of interaction types being modeled; in this embodiment, K=2. These interaction types have no predefined meanings; rather, the model learns to assign meanings to each type.

[0099] The target-relationship reasoning model is constructed using a variational autoencoder model, which maximizes the lower bound of evidence (ELBO) through the following formula:

[0100]

[0101] Where Φ and θ are trainable parameters, q Φ (z|x) represents the encoder's return value as the hidden variable z, p θ(x|z) is a decoder that reconstructs the dynamics system given a set of latent variables z. p θ (z) is a prior probability distribution that is a uniform independent categorical distribution for each relational variable.

[0102] Further, the encoder is shown as follows:

[0103] q Φ (z ij |x)=softmax(f enc,Φ (x) ij,1:K )

[0104] where f enc,Φ (x) is a graph neural network without self-connection, which is used to predict the latent target mRNA interaction relationship graph. The distribution of z is learned according to the encoder. The decoder is shown as follows:

[0105]

[0106] It reconstructs the dynamics system trajectory p θ (x t+1 |x t ,…,x 1 ,z) according to the given latent target mRNA interaction relationship graph hidden variable z.

[0107] The interaction relationship graph of the target mRNA includes an N*N adjacency matrix, where N is the number of residues of the target mRNA, and the elements in the adjacency matrix represent the interaction score of the corresponding two residues.

[0108] In this embodiment, the variational autoencoder structure can effectively extract the features of the residues in the molecular dynamics trajectory, capture the complex molecular interaction information, and further effectively improve the accuracy of siRNA screening.

[0109] In some optional embodiments, based on the interaction relationship graph, the preliminary siRNA screening result is re-screened to obtain a final siRNA screening result, including:

[0110] Based on the adjacency matrix included in the interaction relationship graph, the degree of each node is determined, wherein the degree is the number of connections of the node to other nodes.

[0111] A relationship curve graph of the node and the degree is generated, and the abscissa of the relationship curve graph is the residue number and the ordinate is the degree.

[0112] It is judged whether the residue number of the target mRNA sequence corresponding to the siRNA sequence in the preliminary siRNA screening result is within the residue number range corresponding to the trough region of the relationship curve graph.

[0113] If the residue number range corresponding to the trough region of the relationship curve graph is in the residue number range, the corresponding siRNA sequence is retained.

[0114] If the residue number range corresponding to the trough region of the relationship curve graph is not in the residue number range, the corresponding siRNA sequence is deleted from the preliminary siRNA screening result.

[0115] Specifically, the N x N adjacency matrix in the target mRNA interaction graph can be added by row or by column to obtain the degree of each node, and a relationship curve graph of nodes and degrees is drawn, as shown in FIG. 2. Figure 6 Then, the residues corresponding to the trough region of the relationship curve graph are selected, and the siRNA screened by the siRNA inhibition rate prediction model is further screened. If the residue number of the target mRNA corresponding to the siRNA is in the trough region of the relationship curve graph, the siRNA is retained, otherwise, the siRNA is removed. The retained siRNA is the siRNA obtained by the final screening.

[0116] In this embodiment, the neural relationship inference model is combined with the siRNA inhibition rate prediction model, which can further improve the accuracy of siRNA screening based on the interaction relationship graph.

[0117] In addition, in the present application, the accuracy of EGFP target siRNA screening using only a deep learning model and using a deep learning model and a neural relationship inference model is compared. In the screening result, if the inhibition rate of the screened siRNA is greater than 0.7, the screening result is considered positive, otherwise, it is considered negative. Further, in order to explore the application of the neural relationship inference model in various deep learning model screenings and the ability of the deep learning model to screen siRNA, the present application also uses the AttSiOff model (a gene expression inhibition prediction model combining attention mechanism and siRNA interference) to screen siRNA for the EGFP target. The results are shown in Table 1:

[0118] Table 1 Accuracy of screening results by different methods

[0119] Model Neural relation inference model (no) Neural relation inference model (yes) The invention 58.6％ 68.6％ AttSiOff 62.8％ 78.9％

[0120] As shown in Table 1, if only the siRNA inhibition rate deep learning prediction model constructed by the present application is used to screen siRNA for the EGFP target, the accuracy is 58.6%, and if the neural relationship inference model is further screened, the accuracy can be improved by 10%. In the screening of the AttSiOff model, the accuracy is 62.8%, but on this basis, the neural relationship inference model is further screened, and the accuracy can reach 78.9%, which is improved by 16.1%.

[0121] Therefore, the screening method based on the neural relation inference model can improve the accuracy of the deep learning model in screening siRNA, and the screening method can be used for various deep learning models, and is not limited to the deep learning model provided in the embodiment.

[0122] In the embodiment, a deep learning-based siRNA screening device is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware is also possible and is contemplated.

[0123] The embodiment provides a deep learning-based siRNA screening device, as shown in the following Figure 7 The device comprises:

[0124] The construction module 201 is configured to construct an siRNA inhibition rate prediction model.

[0125] The siRNA inhibition rate prediction module 202 is configured to obtain a preliminary siRNA screening result corresponding to a target mRNA sequence to be predicted based on the siRNA inhibition rate prediction model.

[0126] The molecular dynamics trajectory generation module 203 is configured to perform molecular dynamics simulation on the target mRNA sequence to be predicted, and generate a molecular dynamics trajectory corresponding to the target mRNA sequence to be predicted.

[0127] The relation graph generation module 204 is configured to construct a neural relation inference model, and generate an interaction relation graph corresponding to the target mRNA sequence to be predicted.

[0128] The screening module 205 is configured to perform re-screening on the preliminary siRNA screening result based on the interaction relation graph, and obtain a final siRNA screening result.

[0129] In some optional embodiments, the device further comprises:

[0130] The construction module is configured to obtain an siRNA data set, and the siRNA data set comprises: an siRNA sequence, a target mRNA sequence corresponding to the siRNA sequence, and an inhibition rate of the target mRNA sequence corresponding to the siRNA sequence.

[0131] obtain the siRNA sequence corresponding to the thermodynamic property characteristics and the expert experience characteristics; splice the siRNA sequence in the siRNA data set and the local target mRNA sequence corresponding to the siRNA sequence, and add the corresponding position code on the spliced sequence to obtain the coded spliced sequence; wherein the local target mRNA sequence is the sequence matched with the siRNA sequence plus the preset number of residue sequences before and after the matching with the siRNA sequence; input the coded spliced sequence into the pre-constructed feature network structure for feature extraction to obtain the sequence feature; train the preset fully connected neural network based on the inhibition rate of the siRNA sequence on the corresponding target mRNA sequence, the sequence feature, the siRNA thermodynamic property characteristics and the expert experience characteristics to obtain the siRNA inhibition rate prediction model after training.

[0132] In some optional embodiments, the siRNA inhibition rate prediction module 202 includes:

[0133] constructing an antisense sequence of the siRNA sequence of the target mRNA sequence to be predicted; inputting the antisense sequence of the siRNA sequence into the siRNA inhibition rate prediction model to obtain the siRNA predicted inhibition rate corresponding to the siRNA sequence; sequentially sorting the siRNA predicted inhibition rates from high to low; sequentially selecting siRNA sequences of a preset proportion as the preliminary siRNA screening result from the highest siRNA predicted inhibition rate. Wherein, the 5' end of the local target mRNA sequence is taken as the starting point, and sequentially translated to the 3' end until the end of the antisense sequence of the siRNA sequence is flush with the 3' end of the target mRNA; the length of the antisense sequence of the siRNA sequence ranges from 19nt to 21nt.

[0134] In some optional embodiments, the relationship graph generation module 204 includes:

[0135] The generation unit is configured to obtain node data corresponding to the molecular dynamics trajectory, the node being a residue in the target mRNA sequence to be predicted, the node data including spatial coordinates and velocity of the node at multiple time steps; input the node data into the encoder structure in the neural relationship inference model to extract residue features and generate hidden variables based on the residue features; input the hidden variables into the decoder structure in the neural relationship inference model to generate the interaction relationship graph.

[0136] In some optional embodiments, the screening module 205 includes:

[0137] The screening unit is used to determine the degree of each node based on the adjacency matrix included in the interaction graph, where the degree is the number of other nodes connected to the node; generate a relationship curve between nodes and degrees, with the horizontal axis representing the residue number and the vertical axis representing the degree; determine whether the residue number of the target mRNA sequence corresponding to the siRNA sequence in the preliminary siRNA screening results is within the range of residue numbers corresponding to the trough region in the relationship curve; if it is within the range of residue numbers corresponding to the trough region in the relationship curve, the corresponding siRNA sequence is retained; if it is not within the range of residue numbers corresponding to the trough region in the relationship curve, the corresponding siRNA sequence is deleted from the preliminary siRNA screening results.

[0138] In this embodiment, the deep learning-based siRNA screening device is presented in the form of a functional unit. Here, a unit refers to an ASIC circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0139] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0140] This invention also provides a computer device having the above-described features. Figure 7 The deep learning-based siRNA screening device shown is illustrated.

[0141] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 8 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 8 Take a processor 10 as an example.

[0142] The processor 10 can be a central processor, a network processor, or a combination thereof. The processor 10 can further include hardware chips. The hardware chips can be application specific integrated circuits, programmable logic devices, or a combination thereof. The programmable logic devices can be complex programmable logic devices, field programmable logic gate arrays, general array logic, or any combination thereof.

[0143] The memory 20 stores instructions executable by the at least one processor 10 to cause the at least one processor 10 to perform the methods illustrated in the above embodiments.

[0144] The memory 20 can include a program storage area and a data storage area. The program storage area can store an operating system, application programs required by at least one function, and the like. The data storage area can store data created according to the use of the computer device, and the like. In addition, the memory 20 can include a high-speed random access memory, and can further include a non-transitory memory such as at least one disk storage device, a flash memory device, or other non-transitory solid state memory device. In some alternative embodiments, the memory 20 can optionally include a memory disposed remotely from the processor 10, which can be connected to the computer device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0145] The memory 20 can include a volatile memory such as a random access memory, and can further include a non-volatile memory such as a flash memory, a hard disk, or a solid state disk, and a combination thereof.

[0146] The computer device further includes a communication interface 30 for communication of the computer device with other devices or communication networks.

[0147] The embodiments of the present application further provide a computer readable storage medium. The above method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or stored in a remote storage medium or a non-transitory machine readable storage medium and downloaded from a network and stored in a local storage medium by computer code, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special purpose hardware. The storage medium can be a disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can further include a combination of the above-mentioned memories. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method illustrated in the above embodiments is implemented.

[0148] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, can invoke or provide the method and / or technical solutions according to the present application. Those skilled in the art should understand that the form of computer program instructions in computer readable medium includes but is not limited to source files, executable files, installation package files, etc., and accordingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0149] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A deep learning-based siRNA screening method, characterized by, The method comprises: constructing an siRNA inhibition rate prediction model; obtaining a preliminary siRNA screening result corresponding to a target mRNA sequence to be predicted based on the siRNA inhibition rate prediction model; performing molecular dynamics simulation on the target mRNA sequence to be predicted to generate a molecular dynamics trajectory corresponding to the target mRNA sequence to be predicted; constructing a neural relation reasoning model to generate an interaction relationship graph corresponding to the target mRNA sequence to be predicted; based on the interaction relationship graph, re-screening the preliminary siRNA screening result to obtain a final siRNA screening result; wherein, comprising: determining the degree of each node based on the adjacency matrix included in the interaction relationship graph, wherein the node is a residue in the target mRNA sequence to be predicted, and the degree is the number of connections of the node to other nodes; generate a relationship curve graph of the node and the degree, the horizontal coordinate of the relationship curve graph is the residue number, and the vertical coordinate is the degree; determine whether the residue number of the target mRNA sequence corresponding to the siRNA sequence in the preliminary siRNA screening result is within the residue number range corresponding to the trough region in the relationship curve graph; if it is within the residue number range corresponding to the trough region in the relationship curve graph, the corresponding siRNA sequence is retained; if it is not within the residue number range corresponding to the trough region in the relationship curve graph, the corresponding siRNA sequence is deleted from the preliminary siRNA screening result.

2. The method of claim 1, wherein, The construction of the siRNA inhibition rate prediction model comprises: obtaining an siRNA data set, the siRNA data set comprising: siRNA sequences, target mRNA sequences corresponding to the siRNA sequences, and inhibition rates of the siRNA sequences corresponding to the target mRNA sequences; obtaining thermodynamic property characteristics and expert experience characteristics corresponding to the siRNA sequences; splice the siRNA sequences in the siRNA data set and the local target mRNA sequences corresponding to the siRNA sequences, and add corresponding position encodings on the spliced sequences to obtain encoded spliced sequences; wherein the local target mRNA sequence is a sequence matched with the siRNA sequence plus a preset number of residue sequences before and after the siRNA sequence is matched; input the encoded spliced sequence into a pre-constructed feature network structure for feature extraction to obtain sequence features; train a preset fully connected neural network based on the inhibition rates of the siRNA sequences corresponding to the target mRNA sequences, the sequence features, the thermodynamic property characteristics, and the expert experience characteristics to obtain the siRNA inhibition rate prediction model after training.

3. The method of claim 1, wherein, The siRNA inhibition rate prediction model is used to obtain a preliminary siRNA screening result corresponding to a target mRNA sequence to be predicted, comprising: constructing the antisense sequence of the siRNA sequence of the target mRNA sequence to be predicted; input the antisense sequence of the siRNA sequence into the siRNA inhibition rate prediction model to obtain the siRNA prediction inhibition rate corresponding to the siRNA sequence. The siRNA predicted inhibition rates are sorted from high to low; From the highest siRNA predicted inhibition rate, siRNA sequences of a preset proportion are selected as the preliminary siRNA screening result.

4. The method of claim 3, wherein, The antisense sequence of the siRNA sequence of the target mRNA sequence to be predicted is constructed, including: Taking the 5' end of the target mRNA sequence as the starting point, translating to the 3' end in turn until the end of the antisense sequence of the siRNA sequence is flush with the 3' end of the target mRNA; the length of the antisense sequence of the siRNA sequence ranges from 19nt to 21nt.

5. The method of claim 1, wherein, The neural relationship inference model is constructed based on a variational autoencoder model, and the construction of the neural relationship inference model generates the interaction relationship graph corresponding to the target mRNA sequence to be predicted, including: Obtain the node data corresponding to the molecular dynamics trajectory, the node data including the spatial coordinates and velocities of the nodes at multiple time steps; Input the node data into the encoder structure in the neural relationship inference model, extract residue features and generate hidden variables based on the residue features; Input the hidden variables into the decoder structure in the neural relationship inference model, reconstruct the molecular dynamics trajectory, and calculate the reconstruction error according to the reconstructed molecular dynamics trajectory and the original molecular dynamics trajectory, update the encoder parameters, hidden variables and decoder parameters in the reverse direction, and complete the training after multiple iterations; Obtain the trained neural relationship inference model, and the hidden variables generated by the trained neural relationship inference model are the interaction relationship graph corresponding to the target mRNA sequence to be predicted. 6.A siRNA screening device based on deep learning, characterized by, The device comprises: A construction module for constructing an siRNA inhibition rate prediction model; An siRNA inhibition rate prediction module for obtaining a preliminary siRNA screening result corresponding to a target mRNA sequence to be predicted based on the siRNA inhibition rate prediction model; A molecular dynamics trajectory generation module for performing molecular dynamics simulation on the target mRNA sequence to be predicted to generate a molecular dynamics trajectory corresponding to the target mRNA sequence to be predicted; A relationship graph generation module for constructing a neural relationship inference model to generate an interaction relationship graph corresponding to the target mRNA sequence to be predicted; A screening module for re-screening the preliminary siRNA screening result based on the interaction relationship graph to obtain a final siRNA screening result; The screening module comprises a screening unit configured to determine a degree of each node based on an adjacency matrix included in the interaction relationship graph, wherein the node is a residue in the target mRNA sequence to be predicted, and the degree is a number of other nodes connected to the node; generate a relationship graph of the node and the degree, wherein an abscissa of the relationship graph is a residue number, and an ordinate of the relationship graph is the degree; determine whether a residue number of a target mRNA sequence corresponding to an siRNA sequence in the preliminary siRNA screening result is within a residue number range corresponding to a trough region in the relationship graph; if the residue number is within the residue number range corresponding to the trough region in the relationship graph, retain the corresponding siRNA sequence; and if the residue number is not within the residue number range corresponding to the trough region in the relationship graph, delete the corresponding siRNA sequence from the preliminary siRNA screening result.

7. A computer device, comprising: The memory and the processor are communicatively connected, and the memory stores computer instructions. The processor executes the computer instructions to perform the deep learning-based siRNA screening method according to any one of claims 1-5. The computer readable storage medium stores computer instructions for causing the computer to perform the deep learning-based siRNA screening method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing the computer to perform the deep learning-based siRNA screening method according to any one of claims 1-5.

9. A computer program product, characterised in that, The computer readable storage medium stores computer instructions for causing the computer to perform the deep learning-based siRNA screening method according to any one of claims 1-5.

Images