Method for constructing rna modification prediction model, mrna and rna modification prediction method

Abstract

The present application relates to the technical field of RNA modification prediction, and particularly relates to an RNA modification prediction model construction method, an mRNA and an RNA modification prediction method. The present application constructs a prediction model successively comprising an embedding layer, a transformer encoder layer, a Bi LSTM layer, a convolution layer and a full connection layer by taking the RNA sequence with modification as a positive sample sequence and taking the RNA sequence without modification as a negative sample sequence. The prediction model is trained using the positive sample sequence and the negative sample sequence to determine the parameter values of the prediction model, so that the prediction model performs best. Feature learning is performed through the transformer encoder layer and the Bi LSTM layer, features are extracted through the convolution layer, and fitting is performed through the full connection layer. The combination of the transformer encoder layer and the Bi LSTM layer enables the prediction model to have better ability to learn sequence feature information, and improves the accuracy of the prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of RNA modification prediction, and in particular to an RNA modification prediction model construction method, an mRNA and an RNA modification prediction method. BACKGROUND

[0002] RNA modification is a biological phenomenon often observed in RNA sequences, including m6A, m7G and m5C. In terms of biological mechanisms, RNA modification plays a role in gene expression regulation, translation process efficiency and RNA stability. In addition, RNA modification has been found to be closely related to cancer, neurological disorders and cardiovascular system dysfunction, and as a result, many new ideas for treating cancer, neurological disorders and cardiovascular diseases have been developed. Therefore, accurately determining whether RNA sequences are modified will be beneficial to the subsequent study of RNA modification. In recent years, the emergence of sequencing technology has made many contributions to RNA modification prediction, and a large amount of RNA modification data has been obtained. For example, acetylated RNA immunoprecipitation (acRIP-seq) is one of the most widely used traditional experimental methods for identifying ac4C modification sites. This method is to co-incubate acetylated RNA-specific antibodies with randomly interrupted RNA fragments, and then select fragments with acetylation modification for sequencing. acRIP-seq can show the acetylation level of a specific region of the target sequence, but it cannot identify ac4C sites at base level resolution. Recently, a more complex method called N4-acetylcytosine sequencing (ac4C-seq) has been developed, which can identify ac4C modification sites at a single base level. However, traditional laboratory sequencing methods are complex, time-consuming and costly, so in the field of bioinformatics, RNA modification databases are combined with deep learning technology to obtain a prediction model for determining whether RNA sequences are modified. Although the existing method has achieved certain results, the prediction accuracy is low. SUMMARY

[0003] The present application provides an RNA modification prediction model construction method, which solves the technical problem of low prediction model accuracy in the prior art.

[0004] In order to achieve the above purpose, the technical scheme adopted by the embodiments of the present application is as follows:

[0005] In an embodiment of the first aspect, the present application provides an RNA modification prediction model construction method, which comprises:

[0006] Obtaining RNA sequence data, taking RNA sequences with modification as positive sample sequences, and taking RNA sequences without modification as negative sample sequences;

[0007] The prediction model comprises, in sequence, an embedding layer, a transformer encoder layer, a BiLSTM layer, a convolution layer, and a full connection layer.

[0008] The prediction model is trained using the positive sample sequences and the negative sample sequences, and the parameter values of the prediction model are determined to make the prediction model perform best.

[0009] In some embodiments, before the prediction model is trained, the positive sample sequences and the negative sample sequences need to be preprocessed, including:

[0010] Each base in each of the positive sample sequences and the negative sample sequences is iteratively converted into a word according to the sequence order of the base.

[0011] Different numbers are assigned to different words, so that the positive sample sequences and the negative sample sequences can be represented by one-dimensional matrices.

[0012] In some embodiments, the embedding layer converts the positive sample sequences and the negative sample sequences represented by one-dimensional matrices into high-dimensional matrices, and takes the high-dimensional matrices as the input of the transformer encoder layer.

[0013] In some embodiments, the transformer encoder layer comprises a multi-head self-attention module and a fully connected feedforward module, and both the multi-head self-attention module and the fully connected feedforward module adopt layer normalization and residual connection.

[0014] In some embodiments, in the multi-head self-attention module, for each of the positive sample sequences and the negative sample sequences, an attention weight matrix is obtained by calculating a query vector and a key vector, the attention weight matrix is a [heads, length, length] matrix, and a [length] vector is obtained by averaging the first dimension and the second dimension in sequence.

[0015] In some embodiments, the BiLSTM layer comprises, in sequence, an input layer, a forward LSTM layer, a backward LSTM layer, and an output layer, and the data processing process thereof comprises:

[0016] The input layer receives the output sequence data of the transformer encoder layer.

[0017] The forward LSTM layer receives forward input of the sequence data and outputs a first sequence.

[0018] The backward LSTM layer receives reverse input of the sequence data and outputs a second sequence.

[0019] The output layer receives a third sequence generated by splicing the first sequence and the second sequence, and performs prediction on the third sequence.

[0020] In some embodiments, the convolutional layer comprises a plurality of one-dimensional convolutional layers, each of which is followed by a Batchnorm layer, and the Batchnorm layer of the last one-dimensional convolutional layer is further followed by a maxpooling layer.

[0021] In an embodiment of the second aspect, the present application provides a method for predicting mRNA with N4-acetylcytidine modification, comprising:

[0022] Obtain a database of mRNA sequences with N4-acetylcytidine modification by acRIP-seq, take the sequences therein as first positive sample sequences, obtain a database of mRNA sequences without N4-acetylcytidine modification, take the sequences therein as first negative sample sequences;

[0023] Truncate the first positive sample sequences, and retain the sequence segments with N4-acetylcytidine modification during the truncation operation; after the truncation operation, the positive sample sequences are uniformly distributed within a preset length range; truncate the first negative sample sequences, and after the truncation operation, the negative sample sequences are uniformly distributed within the preset length range; the truncated first positive sample sequences and the truncated first negative sample sequences constitute a first database;

[0024] Use the first database to train the prediction model according to any one of claims 1 to 7, so that the prediction model performs best;

[0025] Input the mRNA sequence to be predicted into the prediction model to obtain a prediction result.

[0026] In an embodiment of the third aspect, the present application provides a method for predicting RNA with N4-acetylcytidine modification, comprising:

[0027] Obtain a database of RNA sequences with N4-acetylcytidine modification by ac4C-seq, take the sequences therein as second positive sample sequences, obtain a database of RNA sequences without N4-acetylcytidine modification, take the sequences therein as second negative sample sequences; the second positive sample sequences and the second negative sample sequences have the same length; the second positive sample sequences and the second negative sample sequences constitute a second database;

[0028] Use the second database to train the prediction model according to any one of claims 1 to 7, so that the prediction model performs best;

[0029] The RNA sequence to be predicted is iteratively cut into several subsequences of a preset length with C base sites as the center;

[0030] The prediction model is inputted with the several subsequences to obtain the prediction results of each subsequence.

[0031] In some embodiments, the prediction method of RNA N4-acetylcytidine modification further comprises:

[0032] The prediction results of all the subsequences are integrated to obtain the prediction results of the whole RNA sequence to be predicted.

[0033] Advantages

[0034] The positive sample sequence and the negative sample sequence are inputted into the prediction model by the embedding layer after encoding, sequentially pass through the transformer encoder layer and the BiLSTM layer for feature learning, then pass through the convolution layer for feature extraction, and finally pass through the full connection layer for fitting to obtain the prediction results, and the parameter values are adjusted according to the error between the prediction results and the sample data, so that the prediction model performs best. The combination of the transformer encoder layer and the BiLSTM layer makes the prediction model have better ability to learn sequence feature information, and improves the accuracy of the prediction model.

[0035] Additional aspects and advantages of embodiments of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following description and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0037] The methods, systems, and / or programs of the accompanying drawings will be further described according to exemplary embodiments. These exemplary embodiments will be described in detail with reference to the drawings. These exemplary embodiments are non-limiting exemplary embodiments, in which the example numbers represent similar mechanisms in each view of the drawings.

[0038] Figure 1 is a flow chart of the RNA modification prediction model construction method in the first aspect embodiment;

[0039] Figure 2 is a structural block diagram of the prediction model in an embodiment;

[0040] Figure 3Figure 1 is a flow chart of a sample sequence preprocessing step in an embodiment;

[0041] Figure 4 Figure 2 is a schematic diagram of a transformer encoder layer in an embodiment;

[0042] Figure 5 Figure 3 is a flow chart of a BiLSTM layer in an embodiment;

[0043] Figure 6 Figure 4 is a flow chart of a method for predicting mRNA N4-acetylcytidine modification in an embodiment of the second aspect;

[0044] Figure 7 Figure 5 is a flow chart of a method for predicting RNA N4-acetylcytidine modification in an embodiment of the third aspect;

[0045] Figure 8 Figure 6 is an ablation experiment data table provided in an embodiment;

[0046] Figure 9 Figure 7 is a ROC curve diagram provided in an embodiment. DETAILED DESCRIPTION

[0047] In order to make the objectives, 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 but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0048] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0049] In the description of the present application, it should be noted that the terms "first", "second", etc. are only used for differentiation and description, and cannot be understood as indicating or implying relative importance.

[0050] Please refer to Figure 1 and Figure 2 In an embodiment of the first aspect, the present application provides a method for constructing an RNA modification prediction model, comprising:

[0051] S1. Obtain RNA sequence data, and take the RNA sequence with modification as the positive sample sequence and the RNA sequence without modification as the negative sample sequence.

[0052] S2. Construct a prediction model, which comprises an embedding layer, a transformer encoder layer, a BiLSTM layer, a convolutional layer and a fully connected layer in sequence.

[0053] S3. Train the prediction model using the positive sample sequence and the negative sample sequence, and determine the parameter values of the prediction model to make the prediction model perform best.

[0054] The positive sample sequence and the negative sample sequence are input into the prediction model after encoding by the embedding layer, and are sequentially subjected to feature learning by the transformer encoder layer and the BiLSTM layer, feature extraction by the convolutional layer, and fitting by the fully connected layer to obtain a prediction result. The parameter values are adjusted according to the error between the prediction result and the sample data to make the prediction model perform best.

[0055] Please refer to Figure 3 In some embodiments, the positive sample sequence and the negative sample sequence need to be preprocessed before being input into the embedding layer. The preprocessing steps include:

[0056] A1. Iteratively convert each three bases in each positive sample sequence and negative sample sequence into a word according to their sequence order.

[0057] A2. Assign different numbers to different words so that the positive sample sequence and the negative sample sequence can be represented by a one-dimensional matrix.

[0058] It should be noted that iteratively converting each three bases into a word means converting the 1st, 2nd and 3rd bases in the sequence into the first word, the 2nd, 3rd and 4th bases into the second word, the 3rd, 4th and 5th bases into the third word, and so on. To ensure the consistency of the sequence length, the sequence data length can be set to a fixed value m, and the sequence with insufficient length can be padded with zeros at the end of the sequence.

[0059] Since the bases are divided into A, G, C and T, there are 65 possible combinations of three bases, and the calculation formula is: 1+4x4x4, where 1 represents the case without bases. Therefore, after using 65 different identifiers to number the words, each RNA sequence (including the positive sample sequence and the negative sample sequence) can be represented by a one-dimensional matrix [n, 1], where the value of n is m-2.

[0060] Further, the embedding layer converts the RNA sequence represented by a one-dimensional matrix into a high-dimensional matrix, and takes the high-dimensional matrix as the input of the transformer encoder layer. Specifically, the high-dimensional matrix converted from the one-dimensional matrix [n, 1] can be represented as [n, d], where d is the dimension, and the more sequence features, the larger the dimension. In this embodiment, the value of d is 300.

[0061] Referring to Figure 4 In some embodiments, the transformer encoder layer includes a multi-head self-attention module and a fully connected feed-forward module, both of which employ layer normalization and residual connections. The multi-head self-attention module employs parallel processing, so as to simultaneously focus on different input segments to capture context, where the number of parallel processing is the number of heads in the multi-head. The fully connected feed-forward module independently analyzes the attention output, so as to identify complex patterns in the data. Both of the two modules employ layer normalization and residual connections to prevent the occurrence of gradient vanishing problem.

[0062] In the transformer encoder, the input data is first calculated by the multi-head self-attention module, and the calculation formula of the multi-head self-attention module is as follows:

[0063]

[0064]

[0065] In the multi-head self-attention module, first, Q, K, and V three vectors are obtained by calculating the output according to formula 1, where Q, K, and V represent Query, Key, and Value respectively, and then the output attention is calculated by using the Q, K, and V three vectors according to formula 2, where is the attention weights (attention weight matrix), and its dimension is [length, length], i.e. both dimensions are the length of the input matrix. However, in order to more efficiently extract information, actually, the following formula is used for calculation:

[0066] MultiHead(Q, K, V) = Concat(head1,..., headh)W (formula 3);

[0067] It can be understood that the dimensions of the input matrix are evenly divided according to the number of heads, then parallel calculation is performed in each head, finally the results of multiple heads are spliced and multiplied by a learnable matrix W as a final output, and therefore the attention weights are also spliced into a [heads, length, length] matrix. In the present scheme, in order to provide redundancy for calculation, the number of heads is 6, that is, h = 6. There is no difference in dimensions between the input and the output of the multi-head self-attention, but the matrix expression of the output of the multi-head self-attention contains more appropriate features between the relationships of the sequence context. Considering that the multi-head self-attention module still lacks certain fitting ability, the output of the multi-head self-attention module needs to be input into the full-connection feedforward module after standardization and residual connection, and calculation is performed in the full-connection feedforward module, and the formula is:

[0068] FFN(x) = Relu(0, xW1 + b1)W2 + b2 (Formula 4);

[0069] Wherein, X is the result of the output of the multi-head self-attention module after standardization and residual connection, W1, W2, b1 and b2 are parameters of the full-connection feedforward module.

[0070] In order to calculate with the downstream module and avoid the problem of gradient disappearance, the output of the full-connection feedforward module will be standardized and connected again, and the final output of the transformer encoder is obtained. In the entire transformer encoder module, the matrix dimensions of the input and the output do not change, both are [n, d].

[0071] Further, in the multi-head self-attention module, for each positive sample sequence and negative sample sequence, the attention weight matrix is calculated by the query vector and the key vector (by formula 2). The attention weight matrix is a [heads, length, length] matrix, and a [length] vector is obtained by averaging the first dimension and the second dimension in turn. The size of each value in the vector corresponds to the contribution of each word to the prediction result. The larger the value, the greater the contribution of the word to the prediction result.

[0072] Referring to Figure 5 In some embodiments, the BiLSTM layer sequentially includes an input layer, a forward LSTM layer, a backward LSTM layer and an output layer, and the data processing process thereof includes:

[0073] B1. The input layer receives the output sequence data of the transformer encoder layer;

[0074] B2. The forward LSTM layer receives a forward input of the sequence data and outputs a first sequence;

[0075] B3. The backward LSTM layer receives a backward input of the sequence data and outputs a second sequence;

[0076] B4. The output layer receives a third sequence generated by concatenating the first sequence and the second sequence, and makes a prediction on the third sequence.

[0077] The formula of the BiLSTM model is as follows:

[0078] f_t = sigma_g(W_fx_t + U_fh_{t-1} + b_f) (Equation 5);

[0079] i_t = sigma_g(W_ix_t + U_ih_{t-1} + b_i) (Equation 6);

[0080] o_t = sigma_g(W_ox_t + U_oh_{t-1} + b_o) (Equation 7);

[0081] tilde{C_t} = tanh(W_Cx_t + U_Ch_{t-1} + b_C) (Equation 8);

[0082] C_t = f_t * C_{t-1} + i_t * tilde{C_t} (Equation 9);

[0083] h_t = o_t * tanh(C_t) (Equation 10);

[0084] where f_t, i_t, o_t, tilde{C_t}, C_t, h_t represent the forget gate, the input gate, the output gate, the current cell state, the current time state, and the current time output, respectively; sigma_g represents the sigmoid function; tanh represents the hyperbolic tangent function; W, U, b represent the weight, the bias term, respectively.

[0085] Please refer to Figure 2 In some embodiments, the convolutional layer includes a plurality of one-dimensional convolutional layers, each of which contains a Batchnorm layer, and the Batchnorm layer after the last one-dimensional convolutional layer further includes a maxpooling layer.

[0086] Specifically, in some embodiments, four one-dimensional convolution layers are adopted, the size of the convolution kernel is 7 in the first one-dimensional convolution layer, and the size of the convolution kernel is 3 in the last three one-dimensional convolution layers. Such a combination can reduce the operation amount of the model while obtaining better feature information extraction capability. In the one-dimensional convolution layer, each convolution layer will iteratively input the input matrix with the size of [convolution kernel size, d] for extraction, and then perform dot multiplication with the learnable matrix W to form a new matrix. Moreover, a Batchnorm layer is included after each 1D convolution layer, and a maxpooling layer is additionally included after the last convolution layer. Through the Batchnorm layer, each batch of input will be standardized, making the model more robust, and the maxpooling layer assists the model in extracting the most important features, making the prediction ability of the model stronger.

[0087] Referring to Figure 2 In some embodiments, the fully connected layer includes one to three layers, and the number of fully connected layers depends on the length of the sequence. Because a long sequence needs more neurons for calculation, more fully connected layers are needed to fit and prevent sudden loss of the number of neurons. The dimension of the final output will be a two-dimensional vector.

[0088] Currently, the prediction of mRNA with N4-acetylcytidine modification is generally obtained by acRIP-seq method, and the length of the obtained mRNA sequence is concentrated in the range of 400-415 nt. Therefore, when the mRNA sequence data obtained by acRIP-seq is used for model training, there will be a problem of overfitting of sequences in the length range of 400-415 nt, and the prediction accuracy of sequences outside this length range is low.

[0089] Referring to Figure 6 In an embodiment of the second aspect, the present application provides a method for predicting mRNA with N4-acetylcytidine modification, comprising:

[0090] N1. Obtain an mRNA sequence database with N4-acetylcytidine modification by acRIP-seq, take the sequences in it as first positive sample sequences, obtain an mRNA sequence database without N4-acetylcytidine modification, and take the sequences in it as first negative sample sequences.

[0091] N2. Truncate the first positive sample sequences, and retain the sequence segments with N4-acetylcytidine modification during the truncation operation; after the truncation operation, the positive sample sequences are uniformly distributed within a preset length range; truncate the first negative sample sequences, and after the truncation operation, the negative sample sequences are uniformly distributed within the preset length range; the truncated first positive sample sequences and the truncated first negative sample sequences constitute a first database.

[0092] N3. Use the first database to train the prediction model as any one of the preceding, so that the prediction model performs best.

[0093] N4. Input the mRNA sequence to be predicted into the prediction model to obtain a prediction result.

[0094] This embodiment truncates the mRNA sequence, so that the sequence length in the first database is uniformly distributed within a preset range. Specifically, the length range can be set to 200 nt to 415 nt, which is sufficient to meet most needs in actual applications. In this way, we can obtain sample sequences with a length range of 200 nt to 415 nt, and the sample sequences of each length are uniformly distributed in the total amount of samples, so that the trained prediction model can achieve more accurate prediction of mRNA sequences of various lengths.

[0095] The mRNA sequence to be predicted is converted into a one-dimensional matrix [n, 1], which is calculated by the trained prediction model to obtain an output. Therefore, the probability of acetylation modification of the mRNA sequence is P = (softmax(output) dot vector(0, 1)), and if this probability is greater than 0.5, then it is considered that the mRNA sequence will undergo N4-acetylcytidine modification.

[0096] At present, to realize single-base precision N4-acetylcytidine modification prediction, the ac4C-seq method is generally used to obtain it. This method uses induced mutation and sequencing detection principles to achieve single-base precision prediction, but this experimental technique takes a long time, and the detection capacity depends on the sequencing depth of 100X level, and the work efficiency is low. The present application combines the RNA sequence database obtained by the ac4C-seq method with the prediction model to propose a prediction method for RNA N4-acetylcytidine modification, which realizes efficient single-base precision N4-acetylcytidine modification prediction.

[0097] Please refer to Figure 7 In an embodiment of the third aspect, the present application provides a prediction method for RNA N4-acetylcytidine modification, comprising:

[0098] V1. Obtain an RNA sequence database with N4-acetylcytidine modification by ac4C-seq, take the sequences therein as second positive sample sequences, obtain an RNA sequence database without N4-acetylcytidine modification, and take the sequences therein as second negative sample sequences; the second positive sample sequences and the second negative sample sequences have the same length; the second positive sample sequences and the second negative sample sequences constitute a second database.

[0099] V2. Use the second database to train the prediction model as any one of the preceding, so that the prediction model performs best.

[0100] V3. The RNA sequence to be predicted is iteratively cut into several subsequences of a preset length with C base sites as the center.

[0101] V4. The aforementioned several subsequences are input into the prediction model to obtain the prediction result.

[0102] The ac4C-seq method generally cuts the RNA sequence into 21 nt subsequences with C base sites as the center for prediction, so the data obtained by the ac4C-seq method is mostly 21 nt RNA sequences. Taking this as an example, the length of the second negative sample sequence and the length of the first negative sample sequence in this embodiment are both 21 nt, and the preset length in V3 is less than 21 nt.

[0103] The RNA sequence to be predicted is iteratively cut into 21 nt subsequences with C base sites as the center. Each subsequence will be represented by a one-dimensional matrix calculated by a trained prediction model to obtain an output. Therefore, the probability of acetylation modification of each subsequence is P = (softmax(output) dot vector (0, 1)). If this probability is greater than 0.5, then it is considered that the C base site at the center of the subsequence will be modified by N4-acetylcytosine to achieve single-base accuracy prediction.

[0104] In some embodiments, the prediction method of RNA N4-acetylcytosine modification further comprises:

[0105] V5. The prediction results of all subsequences are integrated to obtain the prediction result of the whole RNA sequence to be predicted.

[0106] Specifically, if the possibility of N4-acetylcytosine modification of the whole RNA sequence to be predicted is needed, the probabilities P of all subsequences can be integrated and calculated by the following formula:

[0107]

[0108] wherein the numerator is the sum of the probability P of acetylation modification of each subsequence, and the denominator is the number of subsequences. If ac4Cscore is greater than 0.5, it is considered that the RNA sequence to be tested will be modified. The RNA sequence that can be predicted by this embodiment is not limited by the sequence length, and the possibility of RNA N4-acetylcytosine modification can be predicted for RNA sequences of any length.

[0109] In some embodiments, the obtained RNA sequence data can contain multiple races such as humans, yeasts, and archaea, and contain multiple RNA sequence information such as tRNA, mRNA, and rRNA, and the prediction model can predict multiple races and multiple types of RNA sequences.

[0110] It should be noted that the embedding layer, the transformer encoder, the BiLSTM, the convolution algorithm and the full connection algorithm are all algorithms that have been proposed, and the prediction model provided by the present application is a model carefully composed of these algorithms, which has both interpretability and prediction accuracy.

[0111] In the comparative documents with publication numbers CN115424663A and CN115662508A, only the transformer encoder layer is used for feature learning, and the BiLSTM layer and the convolution layer in the present application are missing. The BiLSTM layer can better capture the bidirectional semantic dependency based on the context, improve the feature learning effect of the prediction model, and then use the convolution layer for feature extraction, thereby improving the prediction accuracy of the model.

[0112] In the comparative document with publication number CN116525001A, the scheme first uses the convolution layer for feature extraction, and then uses the BiLSTM and the attention layer for feature learning. The problems of this scheme are: first, feature extraction may lose important information; the convolution layer is in the first layer of the model, which will change the dimension of the input matrix and lose the interpretability of the model for prediction; only using the attention layer lacks the normalization layer and the residual connection layer in the transformer encoder, which is prone to the problem of gradient vanishing and is difficult to deform and deepen the network. The present scheme uses the transformer encoder and the BiLSTM for feature learning and then uses the convolution layer for feature extraction, which avoids the above problems.

[0113] The prediction model provided by the present application is the highest prediction accuracy scheme at present, Figure 8 is the ablation experiment data, which can show that the present scheme (transAC4C) has higher prediction accuracy than the common prediction model. Figure 9 The ROC curve in the above shows the good effect of the present prediction model in the known experimental verification data set.

[0114] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for predicting mRNA N4-acetylcytidine modification, characterized by, The method comprises the following steps: Obtaining RNA sequence data, taking the RNA sequence with modification as a positive sample sequence, and taking the RNA sequence without modification as a negative sample sequence; Constructing a prediction model, which comprises an embedding layer, a transformer encoder layer, a BiLSTM layer, a convolution layer and a full connection layer in sequence; Training the prediction model using the positive sample sequence and the negative sample sequence to determine the parameter values of the prediction model, so that the prediction model performs best; Obtaining an mRNA sequence database with N4-acetylcytidine modification by acRIP-seq, taking the sequence in the database as a first positive sample sequence, and obtaining an mRNA sequence database without N4-acetylcytidine modification, taking the sequence in the database as a first negative sample sequence; Truncating the first positive sample sequence, and retaining the sequence segment with N4-acetylcytidine modification during the truncation operation; after the truncation operation is completed, the positive sample sequence is uniformly distributed within a preset length range; truncating the first negative sample sequence, and after the truncation operation is completed, the negative sample sequence is uniformly distributed within the preset length range; the first positive sample sequence after truncation and the first negative sample sequence after truncation constitute a first database; Training the prediction model using the first database to make the prediction model perform best; Inputting the mRNA sequence to be predicted into the prediction model to obtain a prediction result.

2. The method of predicting N4-acetylcytosine modification of mRNA of claim 1, wherein, Before training the prediction model, the positive sample sequence and the negative sample sequence need to be preprocessed, which comprises: Iteratively converting each three bases in each of the positive sample sequence and the negative sample sequence into a word according to the sequence order of the bases; Assigning different numbers to different words respectively, so that the positive sample sequence and the negative sample sequence can be represented by a one-dimensional matrix.

3. The method of predicting N4-acetylcytosine modification of an mRNA of claim 2, wherein, The embedding layer converts the positive sample sequence and the negative sample sequence represented by a one-dimensional matrix into a high-dimensional matrix, and takes the high-dimensional matrix as the input of the transformer encoder layer.

4. The method of predicting N4-acetylcytosine modification of mRNA of claim 1, wherein, The transformer encoder layer comprises a multi-head self-attention module and a full connection feedforward module, and both the multi-head self-attention module and the full connection feedforward module adopt layer normalization and residual connection.

5. The method of predicting N4-acetylcytosine modification of mRNA of claim 4, wherein, In the multi-head self-attention module, for each of the positive sample sequence and the negative sample sequence, an attention weight matrix is obtained by calculating a query vector and a key vector, the attention weight matrix is a [heads, length, length] matrix, and a [length] vector is obtained by sequentially averaging the first dimension and the second dimension.

6. The method of predicting N4-acetylcytosine modification of mRNA of claim 1, wherein, The BiLSTM layer comprises an input layer, a forward LSTM layer, a backward LSTM layer and an output layer in sequence, and its data processing process comprises: The input layer receives the output sequence data of the transformer encoder layer; The forward LSTM layer receives the forward input of the sequence data and outputs a first sequence; The backward LSTM layer receives a reverse input of the sequence data, and outputs a second sequence; The output layer receives a third sequence generated by splicing the first sequence and the second sequence, and performs prediction on the third sequence.

7. The method of predicting N4-acetylcytosine modification of mRNA of claim 1, wherein, The convolutional layer comprises a plurality of one-dimensional convolutional layers, each of which is followed by a Batchnorm layer, and the Batchnorm layer of the last one-dimensional convolutional layer is further followed by a maxpooling layer.

8. A method for predicting RNA N4-acetylcytidine modification, characterized by, Comprise: obtain the RNA sequence database of N4-acetylcytosine modification by ac4C-seq, take the sequence as the second positive sample sequence, obtain the RNA sequence database of N4-acetylcytosine modification, take the sequence as the second negative sample sequence; the length of the second positive sample sequence and the second negative sample sequence is the same; the second positive sample sequence and the second negative sample sequence constitute a second database; Use the second database to train the prediction model according to any one of claims 1 to 7, so that the prediction model performs best; Cut the RNA sequence to be predicted into several sub-sequences of a preset length iteratively with C base sites as the center; Input several sub-sequences into the prediction model to obtain the prediction result of each sub-sequence.

9. The method of predicting RNA N4-acetylcytosine modification of claim 8, wherein, Also include: Integrate the prediction results of all the sub-sequences to obtain the prediction result of the whole RNA sequence to be predicted.

Images