Prediction method of mRNA sequence ribosome occupancy, model training method and related device

Abstract

The invention relates to an mRNA sequence ribosome occupancy prediction method, a model training method and a related device. The method comprises the following steps: acquiring an mRNA sequence and a secondary structure of the mRNA sequence in a target cell type; inputting the mRNA sequence and the secondary structure thereof into a pre-trained prediction model, respectively obtaining a first embedded representation of the mRNA sequence and a second embedded representation of the secondary structure through different encoders in the prediction model, fusing the first embedded representation and the second embedded representation based on a multi-head self-attention mechanism to obtain a fused embedded representation, and obtaining a multi-head self-attention prediction model; classification is carried out based on the fused embedded representation; and based on classified output of the prediction model, obtaining a prediction result of the ribosome occupancy of the mRNA sequence in the target cell type. Compared with the prior art, after the prediction model is trained in any cell type, ribosome occupancy prediction is carried out on any other cell type, the prediction precision is high, and the prediction model is high in mobility.

Description

Technical Field

[0001] The present application relates to the technical field of gene translation prediction, and in particular to a method for predicting ribosome occupancy of an mRNA sequence, a model training method and related devices. Background Art

[0002] The central law of genetic information is that DNA is transcribed into mRNA, which is then translated into protein. The process of mRNA being translated into protein is that ribosomes bind to mRNA.

[0003] Thanks to the development of translationomics, by measuring the number of ribosomes bound to each mRNA in the cell, it is possible to determine the ribosome occupancy on the mRNA and quantify the translation level of the mRNA, thereby providing data support for the design of mRNA sequences and predicting translation events on mRNA.

[0004] Currently, the prediction of ribosome occupancy on mRNA sequences usually uses mRNA sequences as model input, resulting in low prediction accuracy of ribosome occupancy and low transferability of the prediction process between different cell lines. Summary of the invention

[0005] Based on this, it is necessary to provide a prediction method, model training method and related devices for the ribosome occupancy of mRNA sequences that can predict with high accuracy and realize migration prediction between different cells in order to address the above technical problems.

[0006] In a first aspect, the present application provides a method for predicting ribosome occupancy of an mRNA sequence, the method comprising:

[0007] Obtaining an mRNA sequence and a secondary structure of the mRNA sequence in a target cell type;

[0008] Inputting the mRNA sequence and its secondary structure into a pre-trained prediction model, obtaining a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure through different encoders in the prediction model, fusing the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation, and performing classification based on the fused embedding representation;

[0009] Based on the classification output of the prediction model, a prediction result of the ribosome occupancy of the mRNA sequence in the target cell type is obtained.

[0010] In one embodiment, the mRNA sequence and its secondary structure are input into a pre-trained prediction model, a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure are obtained respectively through different encoders in the prediction model, the first embedding representation and the second embedding representation are fused based on a multi-head self-attention mechanism to obtain a fused embedding representation, and classification is performed based on the fused embedding representation, including:

[0011] Inputting the mRNA sequence and its secondary structure into a pre-trained prediction model;

[0012] The mRNA sequence is encoded by a first encoder in the prediction model to obtain a first embedded representation of the mRNA sequence; the first encoder encodes the mRNA sequence based on a self-attention mechanism;

[0013] The second encoder in the prediction model encodes the secondary structure to obtain a second embedding representation of the secondary structure; the second encoder encodes the secondary structure based on a masked self-attention mechanism;

[0014] fusing the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation;

[0015] The fused embedded representation is subjected to a maximum pooling operation to reduce the dimension of the fused embedded representation, and the embedded representation after the dimension reduction is input into a fully connected layer for classification.

[0016] In one embodiment, encoding the mRNA sequence by a first encoder in the prediction model to obtain a first embedded representation of the mRNA sequence includes:

[0017] Inputting the mRNA sequence into the first encoder;

[0018] Vectorizing the mRNA sequence through a first embedding layer of the first encoder to obtain a first vector sequence;

[0019] Iteratively extracting first embedded information from the first vector sequence through multiple first encoding layers of the first encoder; each of the first encoding layers processes the first vector sequence using a self-attention mechanism, and the input of the first first encoding layer is the first vector sequence, and the input of each subsequent first encoding layer is the first embedded information output by the previous first encoding layer;

[0020] The first embedding information output by the last first encoding layer is pooled to obtain a first embedding representation of the mRNA sequence.

[0021] In one embodiment, encoding the secondary structure by a second encoder in the prediction model to obtain a second embedded representation of the secondary structure includes:

[0022] inputting the secondary structure into the second encoder;

[0023] vectorizing the secondary structure through a second embedding layer of the second encoder to obtain a second vector sequence;

[0024] Iteratively extracting the second embedding information from the second vector sequence through multiple second encoding layers of the second encoder; each of the second encoding layers uses a masked self-attention mechanism to process the second vector sequence, the mask position of the masked self-attention mechanism is the position where the secondary structure on the mRNA sequence is missing, the input of the first second encoding layer is the second vector sequence, and the input of each subsequent second encoding layer is the second embedding information output by the previous second encoding layer;

[0025] Pooling is performed on the second embedding information output by the last second encoding layer to obtain a second embedding representation of the secondary structure.

[0026] In one embodiment, the fusing the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation includes:

[0027] Determining a key matrix and a value matrix of the multi-head self-attention mechanism based on the first embedding representation, and determining a query matrix of the multi-head self-attention mechanism based on the second embedding representation;

[0028] The elements in the key matrix, the value matrix, and the query matrix are divided according to the number of attention heads preset by the multi-head self-attention mechanism, and the attention weight of each attention head is determined;

[0029] The output of the corresponding attention head is determined based on each attention weight, and the outputs of each attention head are concatenated to obtain a fused embedded representation.

[0030] In a second aspect, the present application also provides a method for training an mRNA ribosome occupancy prediction model, the method comprising:

[0031] Constructing a multimodal database; the multimodal database includes mRNA sequence samples and secondary structure samples obtained by sequencing the mRNA sequence samples in the current cell type;

[0032] Inputting the mRNA sequence samples in the multimodal database and the secondary structure samples corresponding to the mRNA sequence samples into the prediction model to be trained;

[0033] Encoding the currently input mRNA sequence sample by a first encoder in the prediction model to be trained to obtain a current first embedding representation;

[0034] Encoding the currently input secondary structure sample by a second encoder in the prediction model to be trained to obtain a current second embedding representation;

[0035] Inputting the current first embedding representation and the current second embedding representation into the multi-head self-attention module in the prediction model to be trained to fuse and classify the current first embedding representation and the current second embedding representation, and obtaining ribosome occupancy prediction data on the current mRNA sequence sample in the current cell type based on the classification result; the current cell type is the cell type corresponding to the secondary structure sample currently input;

[0036] Based on the ribosome occupancy prediction data and the actual ribosome occupancy data of the current mRNA sequence sample in the current cell type, the model parameters of the prediction model to be trained are updated until a trained prediction model is obtained.

[0037] In one of the embodiments, the multimodal database also includes actual ribosome occupancy data of the mRNA sequence samples;

[0038] Before constructing the multimodal database, the method further includes:

[0039] Obtain mRNA sequence samples;

[0040] Determine the secondary structure sample of the mRNA sequence sample in the current cell type, detect the structural strength of the secondary structure sample, and obtain the RNA structure probe score of the secondary structure sample;

[0041] Determine the actual ribosome occupancy data of the mRNA sequence sample in the current cell type;

[0042] The constructing of a multimodal database comprises:

[0043] A multimodal database is constructed based on the mRNA sequence samples, secondary structure samples and actual ribosome occupancy data.

[0044] In one embodiment, the determining of the actual ribosome occupancy data of the mRNA sequence sample in the current cell type comprises:

[0045] Determine the mRNA fragments in the translation state on the mRNA sequence sample to obtain original Ribo-Seq data;

[0046] Extracting the original Ribo-Seq data to obtain sequencing fragments that meet preset quality conditions;

[0047] Matching the sequenced fragment with a predetermined reference genome to obtain the initial position of the sequenced fragment;

[0048] Correcting the initial position of the sequencing fragment;

[0049] An exon reference genome is constructed, and the corrected sequencing fragments are matched with the exon reference genome to obtain ribosome position information with single-base accuracy, and actual ribosome occupancy data is determined based on the ribosome position information with single-base accuracy.

[0050] In one embodiment, before inputting the mRNA sequence samples in the multimodal database and the secondary structure samples corresponding to the mRNA sequence samples into the prediction model to be trained, the method further includes:

[0051] Aligning the RNA structure probe scores of the secondary structure samples and the actual ribosome occupancy data with a predetermined reference genome at the transcript level to obtain a number of transcripts carrying sequence information, structure information, and ribosome occupancy information;

[0052] Taking the transcript as a unit, denoising the transcript based on the actual data of the ribosome occupancy rate and the length of the transcript to obtain a training sample, and constructing a training data set based on the training sample;

[0053] Extracting sequence information and structural information corresponding to each training sample in the training data set, determining the mRNA sequence corresponding to the extracted sequence information as the currently input mRNA sequence sample, and determining the secondary structure sample corresponding to the extracted structural information as the currently input secondary structure sample.

[0054] In one embodiment, the step of performing denoising on the transcript based on the actual ribosome occupancy data and the length of the transcript to obtain a training sample and constructing a training data set based on the training sample includes:

[0055] Detect the length of each transcript and remove transcripts shorter than a preset length;

[0056] Extracting the first several bases and the last several bases of each transcript remaining after removing the transcript of the preset length, and splicing the first several bases and the last several bases of each transcript to form a prediction sample;

[0057] Determining the actual data of the ribosome occupancy of each predicted sample, and removing the predicted samples whose actual data of the ribosome occupancy is zero;

[0058] After removing the sequencing samples whose actual ribosome occupancy data is zero, the remaining predicted samples are sorted according to the size of the actual ribosome occupancy data, and a number of predicted samples with higher rankings in the actual ribosome occupancy data are extracted to form positive samples, and a number of predicted samples with lower rankings in the actual ribosome occupancy data are extracted to form negative samples;

[0059] The positive samples and negative samples are integrated to obtain a training data set.

[0060] In a third aspect, the present application also provides a device for predicting ribosome occupancy of an mRNA sequence, the device comprising:

[0061] A data acquisition module, used to acquire mRNA sequences and secondary structures of the mRNA sequences in target cell types;

[0062] A model calling module, used for inputting the mRNA sequence and its secondary structure into a pre-trained prediction model, obtaining a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure through different encoders in the prediction model, fusing the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation, and performing classification based on the fused embedding representation;

[0063] A prediction module is used to obtain a prediction result of the ribosome occupancy of the mRNA sequence in the target cell type based on the classification output of the prediction model.

[0064] In a fourth aspect, the present application also provides a training device for an mRNA ribosome occupancy prediction model, the device comprising:

[0065] A database construction module, used to construct a multimodal database; the multimodal database includes an mRNA sequence and a secondary structure of the mRNA sequence in any cell type, and the mRNA sequence is determined as a current mRNA sequence and the secondary structure of the mRNA sequence in any cell type is determined as a current structure sequence;

[0066] A data input module, used for inputting the current mRNA sequence and the current structure sequence into the prediction model to be trained; encoding the current mRNA sequence through the first encoder in the prediction model to be trained to obtain a current first embedding representation; encoding the current structure sequence through the second encoder in the prediction model to be trained to obtain a current second embedding representation; inputting the current first embedding representation and the current second embedding representation into the multi-head self-attention module in the prediction model to be trained to fuse and classify the first embedding representation and the second embedding representation, and obtaining ribosome occupancy prediction data on the current mRNA sequence based on the classification structure of the embedding representation;

[0067] The parameter updating module is used to update the model parameters of the prediction model to be trained based on the ribosome occupancy prediction data and the actual ribosome occupancy data on the current mRNA sequence until a trained prediction model is obtained.

[0068] In a fifth aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-described method when executing the computer program.

[0069] In a sixth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and the computer program, when executed by a processor, implements the steps of the method described above.

[0070] In a seventh aspect, the present application also provides a computer program product, including a computer program, which implements the steps of the method described above when executed by a processor.

[0071] In the above-mentioned method for predicting the ribosome occupancy of the mRNA sequence, the mRNA sequence and the secondary structure of the mRNA sequence in the target cell type are obtained, and then the mRNA sequence and its secondary structure are input into a pre-trained prediction model, and the first embedding representation of the mRNA sequence and the second embedding representation of the secondary structure are respectively obtained through different encoders in the prediction model, and the first embedding representation and the second embedding representation are fused based on the multi-head self-attention mechanism to obtain the fused embedding representation, and classification is performed based on the fused embedding representation. Finally, based on the classification output of the prediction model, the prediction result of the ribosome occupancy of the mRNA sequence in the target cell type is obtained. Since the mRNA sequence is basically consistent in different cell types, and the secondary structure on the mRNA sequence shows different states in different cell types, the prediction model takes the mRNA sequence and the secondary structure in the corresponding target cell type as input, uses a fixed mRNA sequence, and fuses the secondary structure with cell type specificity to predict the ribosome occupancy, comprehensively considers the mRNA sequence and secondary structure information, and can accurately predict the cell type specific ribosome occupancy data. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0073] Figure 1 Schematic diagram of a method for predicting ribosome occupancy of an mRNA sequence in one embodiment;

[0074] Figure 2 A schematic diagram of a process for extracting embedded representations of mRNA sequences and secondary structures in one embodiment;

[0075] Figure 3 A schematic diagram of a process for encoding an mRNA sequence in one embodiment;

[0076] Figure 4 A schematic diagram of a process for encoding a secondary structure in one embodiment;

[0077] Figure 5 is a schematic diagram of a process of fusing a first embedded representation and a second embedded representation in one embodiment;

[0078] Figure 6 Schematic diagram of a training method for an mRNA ribosome occupancy prediction model in one embodiment;

[0079] Figure 7 A schematic diagram of a process for determining actual data of mRNA sequence samples, secondary structure samples, and ribosome occupancy in one embodiment;

[0080] Figure 8 A schematic diagram of a process for determining actual ribosome occupancy data in one embodiment;

[0081] Fig. 9 A schematic diagram of a process for processing training data before inputting the training data into a prediction model in one embodiment;

[0082] Fig.10 This is a structural diagram of a prediction model in an application example;

[0083] Fig.11 A comparison chart of the prediction effect of different inputs when the prediction model is trained and predicted in the same cell line in an application example;

[0084] Fig.12 A comparison chart of the prediction effect evaluation of different inputs when the prediction model is trained and predicted in different types of cell lines in an application example;

[0085] Fig.13 is a structural block diagram of a device for predicting ribosome occupancy of mRNA sequences in one embodiment;

[0086] Fig.14 1 is a structural block diagram of a training device for an mRNA ribosome occupancy prediction model in one embodiment. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0088] In one embodiment, Figure 1 As shown, a method for predicting the ribosome occupancy of an mRNA sequence is provided. This embodiment uses the method applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices, portable wearable devices, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services, etc.

[0089] In this embodiment, the above-mentioned method for predicting the ribosome occupancy of the mRNA sequence includes the following steps S101 to S103.

[0090] in:

[0091] Step S101, obtaining the mRNA sequence and the secondary structure of the mRNA sequence in the target cell type.

[0092] Among them, the mRNA sequence refers to messenger RNA (mRNA), which is a single-stranded ribonucleic acid transcribed from a strand of DNA as a template, which carries genetic information and can guide protein synthesis.

[0093] Among them, the target cell type refers to the cell type for which the ribosome occupancy of the mRNA sequence needs to be predicted to determine the translation efficiency of the mRNA sequence in this cell type. The translation efficiency can be quantified by how many ribosomes are bound to the mRNA sequence, that is, the ribosome occupancy. The ribosome occupancy of mRNA refers to the proportion of mRNA fragments that are bound and translated by ribosomes in a specific cell or tissue sample.

[0094] Among them, the secondary structure of the mRNA sequence refers to the stable structure formed by base pairing inside the mRNA sequence molecule, such as the hairpin structure, stem-loop structure, etc. The secondary structure provides sufficient binding energy for the mRNA molecule through hydrogen bond interactions between bases, thereby maintaining the stability and integrity of its structure. In addition, the secondary structure can affect the stability, variability and foldability of the mRNA molecule, thereby regulating the function and activity of the mRNA.

[0095] Exemplarily, an mRNA sequence is obtained, a target cell type for which ribosome occupancy prediction is required is determined, and based on the determined target cell type, a secondary structure corresponding to the mRNA sequence in the target cell type is determined.

[0096] Step S102, input the mRNA sequence and its secondary structure into a pre-trained prediction model, obtain a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure through different encoders in the prediction model, fuse the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism, obtain a fused embedding representation, and perform classification based on the fused embedding representation.

[0097] The encoder refers to a module that can encode an input and obtain an embedded representation corresponding to the input. In this embodiment, different encoders can encode different types of inputs and obtain embedded representations corresponding to each input. Different encoders are distinguished by a first encoder and a second encoder. The first encoder encodes the mRNA sequence to obtain a first embedded representation, and the second encoder encodes the secondary structure to obtain a second embedded representation.

[0098] Among them, the first embedded representation refers to the context information carrying the association relationship between each mRNA in the mRNA sequence obtained by encoding with the mRNA sequence as input. The second embedded representation refers to the context information carrying the association relationship between each secondary structure obtained by encoding with the secondary structure as input. The fused embedded representation refers to the information obtained by splicing and fusing the first embedded representation obtained based on the mRNA sequence and the second embedded representation obtained based on the secondary structure in some way (such as weighted addition, neural network, attention extraction, etc.), which can represent the association relationship between the first embedded representation and the second embedded representation.

[0099] Exemplarily, the mRNA sequence and its secondary structure are input into a pre-trained prediction model, and a first embedding representation of the mRNA sequence is obtained through a first encoder in the prediction model, and a second embedding representation of the secondary structure is obtained through a second encoder in the prediction model; the first embedding representation and the second embedding representation are fused based on a multi-head self-attention mechanism to obtain a fused embedding representation; finally, classification is performed based on the fused embedding representation.

[0100] Step S103, based on the classification output of the prediction model, obtain the prediction result of the ribosome occupancy of the mRNA sequence in the target cell type.

[0101] Exemplarily, a classification result of the prediction model for the embedded representation is obtained, and based on the classification result, a prediction result of the ribosome occupancy of the mRNA sequence in the target cell type is obtained.

[0102] In the above-mentioned method for predicting the ribosome occupancy of the mRNA sequence, the mRNA sequence and the secondary structure of the mRNA sequence in the target cell type are obtained, and then the mRNA sequence and its secondary structure are input into a pre-trained prediction model, and the first embedding representation of the mRNA sequence and the second embedding representation of the secondary structure are respectively obtained through different encoders in the prediction model, and the first embedding representation and the second embedding representation are fused based on the multi-head self-attention mechanism to obtain the fused embedding representation, and classification is performed based on the fused embedding representation. Finally, based on the classification output of the prediction model, the prediction result of the ribosome occupancy of the mRNA sequence in the target cell type is obtained. Since the mRNA sequence is basically consistent in different cell types, and the secondary structure on the mRNA sequence shows different states in different cell types, the prediction model takes the mRNA sequence and the secondary structure in the corresponding target cell type as input, uses a fixed mRNA sequence, and fuses the secondary structure with cell type specificity to predict the ribosome occupancy, comprehensively considers the mRNA sequence and secondary structure information, and can accurately predict the cell type specific ribosome occupancy data.

[0103] In one embodiment, Figure 2 As shown, step S102 encodes the mRNA sequence and the secondary structure to obtain the corresponding embedded representation, including the following steps S201 to S205. Among them:

[0104] Step S201, inputting the mRNA sequence and its secondary structure into a pre-trained prediction model.

[0105] Exemplarily, the mRNA sequence is input into a first encoder and the secondary structure of the mRNA sequence is input into a second encoder, and the parameters of the first encoder and the second encoder have been pre-trained and determined.

[0106] Step S202: Encode the mRNA sequence by using a first encoder in the prediction model to obtain a first embedded representation of the mRNA sequence.

[0107] The first encoder encodes the mRNA sequence based on the self-attention mechanism. The first encoder has a first embedding layer, a plurality of first encoding layers and a first Pooler layer connected in sequence. The first embedding layer is used to vectorize the input mRNA sequence. The first encoding layer is used to convert the vectorized mRNA sequence into a high-dimensional vector representation. The first Pooler layer is used to obtain the high-dimensional vector representation corresponding to the mRNA sequence to obtain the context information of the mRNA sequence, that is, the first embedding representation.

[0108] Exemplarily, the mRNA sequence is input into a first encoder, and the mRNA sequence is encoded by the first encoder to obtain a first embedded representation of the mRNA sequence.

[0109] Step S203: Encode the secondary structure by using a second encoder in the prediction model to obtain a second embedding representation of the secondary structure.

[0110] The second encoder encodes the secondary structure based on the masked self-attention mechanism. The second encoder has a second embedding layer, multiple second encoding layers and a second Pooler layer connected in sequence. The second embedding layer is used to vectorize the input secondary structure. The second encoding layer is used to convert the vectorized secondary structure into a high-dimensional vector representation. The second Pooler layer is used to obtain the high-dimensional vector representation corresponding to the secondary structure to obtain the context information of the secondary structure, that is, the second embedding representation.

[0111] Exemplarily, the secondary structure is input into a second encoder, and the secondary structure is encoded by the second encoder to obtain a second embedding representation of the secondary structure.

[0112] Step S204: fuse the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation.

[0113] Exemplarily, the first embedding representation and the second embedding representation are used as inputs of a multi-head self-attention mechanism, and the first embedding representation and the second embedding representation are fused based on the multi-head self-attention mechanism to obtain a fused embedding representation.

[0114] In step S205, a maximum pooling operation is performed on the fused embedding representation to reduce the dimension of the fused embedding representation, and the embedding representation after the dimension reduction is input into a fully connected layer for classification.

[0115] Among them, performing a maximum pooling operation on the fused embedded representation can reduce the dimension of the fused embedded representation while ensuring that the embedded representation after the reduced dimension still carries key information.

[0116] Exemplarily, the fused embedded representation is input into a maximum pooling layer for a maximum pooling operation, and feature information in the fused embedded representation is further extracted to reduce the dimension of the fused embedded representation. The embedded representation after the reduced dimension is then input into a fully connected layer for fusion with the global features (or all features) of the input sequence (i.e., the first embedded representation and the second embedded representation), and then input into an MLP layer. Classification prediction is performed on the embedded representation after the reduced dimension based on the MLP layer to obtain a classification result of the fused embedded representation.

[0117] In this embodiment, by inputting the mRNA sequence and its secondary structure into the prediction model, and encoding the mRNA sequence through the prediction model to obtain a first embedded representation, and encoding the secondary structure to obtain a second embedded representation, the first embedded representation and the second embedded representation are fused based on a multi-head self-attention mechanism to obtain a fused embedded representation, and then the fused embedded representation is classified to obtain a classification result, which can extract the embedded representations of the mRNA sequence and the secondary structure respectively, and fuse the two embedded representations, so as to accurately predict the ribosome occupancy of the mRNA sequence in various cell types in the subsequent prediction process.

[0118] In one embodiment, Figure 3 As shown, step S202 encodes the mRNA sequence including the following steps S301 to S304. Among them:

[0119] Step S301, input the mRNA sequence into the first encoder.

[0120] Exemplarily, the mRNA sequence is input into a first embedding layer of a first encoder.

[0121] Step S302, vectorize the mRNA sequence through the first embedding layer of the first encoder to obtain a first vector sequence.

[0122] The vectorization of the mRNA sequence refers to segmenting the input mRNA sequence, adding tags, and converting it into vector information that can be recognized by the first encoder.

[0123] Exemplarily, the mRNA sequence is segmented by the first embedding layer of the first encoder, and tags are added to the segmented mRNA sequences one by one, and a plurality of first vector information are integrated to obtain a first vector sequence.

[0124] Step S303: Iteratively extract first embedded information from the first vector sequence through multiple first encoding layers of the first encoder.

[0125] Among them, each first coding layer uses a self-attention mechanism to process the first vector sequence, and the input of the first first coding layer is the first vector sequence, and the input of each subsequent first coding layer is the first embedded information output by the previous first coding layer.

[0126] Exemplarily, the first vector sequence is input into the first first coding layer, and the first first coding layer extracts the embedding information of the first vector sequence based on the self-attention mechanism to obtain a first embedding information. Then, the first embedding information is input into the second first coding layer, and the second first coding layer extracts the first embedding information again based on the self-attention mechanism to obtain a new first embedding information. The new first embedding information is then input into the third first coding layer until the last first coding layer outputs the final first embedding information.

[0127] Step S304: Process the first embedding information output by the last first encoding layer to obtain a first embedding representation of the mRNA sequence.

[0128] Exemplarily, the final first embedding information output by the last first encoding layer is input into the first Pooler layer, and the first embedding information is linearized and activated by the first Pooler layer to obtain the first embedding representation of the mRNA sequence.

[0129] In this embodiment, the mRNA sequence is encoded by the first encoder, and the mRNA sequence is processed multiple times using the multi-layer first encoding layer of the first encoder, so that more embedded information in the mRNA sequence can be obtained to ensure that the first embedded representation carries more feature information.

[0130] In one embodiment, Figure 4 As shown, step S203 of encoding the secondary structure includes the following steps S401 to S404. Among them:

[0131] Step S401, input the secondary structure into the second encoder.

[0132] Exemplarily, the secondary structure is input into a second embedding layer of a second encoder.

[0133] Step S402: vectorize the secondary structure through a second embedding layer of a second encoder to obtain a second vector sequence.

[0134] The vectorization of the secondary structure refers to segmenting the input secondary structure, adding tags, and converting it into vector information that can be recognized by the second encoder.

[0135] Exemplarily, the second embedding layer of the second encoder is used to segment the secondary structure, and tags are added to the segmented secondary structures one by one, and a plurality of second vector information are integrated to obtain a second vector sequence.

[0136] Step S403: Iteratively extract second embedded information from the second vector sequence through multiple second encoding layers of the second encoder.

[0137] Among them, each second encoding layer uses a masked self-attention mechanism to process the second vector sequence. The mask position of the masked self-attention mechanism is the position where the secondary structure is missing on the mRNA sequence. The input of the first second encoding layer is the second vector sequence, and the input of each subsequent second encoding layer is the second embedded information output by the previous second encoding layer.

[0138] Exemplarily, the second vector sequence is input into the first second coding layer, and the first second coding layer extracts the embedding information of the second vector sequence based on the masked self-attention mechanism to obtain a second embedding information. Then, the second embedding information is input into the second second coding layer, and the second second coding layer extracts the second embedding information again based on the masked self-attention mechanism to obtain a new second embedding information. The new second embedding information is then input into the third second coding layer until the last second coding layer outputs the final second embedding information.

[0139] Step S404: Process the second embedding information output by the last second coding layer to obtain a second embedding representation of the secondary structure.

[0140] Exemplarily, the final second embedding information output by the last second encoding layer is input into the second Pooler layer, and the second embedding information is linearized and activated by the second Pooler layer to obtain a second embedding representation of the secondary structure.

[0141] In this embodiment, the secondary structure is encoded by the second encoder, and the secondary structure is processed multiple times by the multi-layer second encoding layer of the second encoder, so that more embedded information in the secondary structure can be obtained to ensure that the second embedded representation carries more feature information. At the same time, since the icSHAPE (invivo click selective 2-hydroxyl acylation and profiling experiment) sequencing technology is used for measurement in the process of obtaining the secondary structure, the secondary structure obtained by this method is not complete and there are many invalid value sites on it. When extracting the second embedded information, this embodiment adopts a masked self-attention mechanism, and uses the invalid value site as the mask position of the masked self-attention mechanism. The attention of the mask position is extracted through the positions of other valid sites to obtain the embedded information of the position, thereby completing the secondary structure and ensuring the accuracy of subsequent predictions.

[0142] In one embodiment, Figure 5 As shown, step S204 of fusing the first embedded representation and the second embedded representation includes the following steps S501 to S503. Among them:

[0143] Step S501, determining a key matrix and a value matrix of a multi-head self-attention mechanism based on the first embedding representation, and determining a query matrix of the multi-head self-attention mechanism based on the second embedding representation.

[0144] Among them, the query matrix has multiple query vectors, which are obtained by linear transformation based on each second embedded information in the second embedded representation, and are used to represent the information concerned at the current position on the mRNA sequence, that is, the specific situation of the secondary structure. The key matrix has multiple key vectors, which are obtained by linear transformation based on the first embedded information in the first embedded representation, and are used to represent the information at other positions on the mRNA sequence. The value matrix has multiple value vectors, which are obtained by linear transformation based on the first embedded information in the first embedded representation, and are used to represent the actual content at other positions on the mRNA sequence.

[0145] Exemplarily, a key weight matrix, a value weight matrix, and a query weight matrix of a multi-head self-attention mechanism are determined, the first embedding representation is multiplied by the key weight matrix to perform a corresponding linear transformation on the first embedding representation to obtain a key matrix of the multi-head self-attention mechanism, the first embedding representation is multiplied by the value weight matrix to perform a corresponding linear transformation on the first embedding representation to obtain a value matrix of the multi-head self-attention mechanism, and the second embedding representation is multiplied by the query weight matrix to perform a corresponding linear transformation on the second embedding representation to obtain a query matrix of the multi-head self-attention mechanism.

[0146] Step S502, divide the elements in the key matrix, value matrix and query matrix according to the number of attention heads preset by the multi-head self-attention mechanism, and determine the attention weight of each attention head.

[0147] Exemplarily, the multi-head attention mechanism of this embodiment is integrated in a multi-head self-attention module, and the key matrix, value matrix and query matrix are input into the multi-head self-attention module. The key matrix, value matrix and query matrix are divided according to a preset number of attention heads through the multi-head self-attention module to obtain multiple sub-key matrices, multiple sub-value matrices and multiple sub-query matrices that match the number of attention heads. The attention weights are then calculated based on each sub-key matrix, sub-value matrix and sub-query matrix to obtain the attention weight of each attention head.

[0148] Step S503: determine the output of the corresponding attention head based on each attention weight, and concatenate the outputs of each attention head to obtain a fused embedded representation.

[0149] Exemplarily, the output of each attention head is determined based on the attention weight of each attention head for weighted calculation, and the weighted outputs are concatenated to obtain a fused embedded representation.

[0150] In this embodiment, the first embedding representation and the second embedding representation are jointly input into the multi-head attention module, and the first embedding representation and the second embedding representation are respectively processed into the key, value and query of the multi-head attention module, so that when the multi-head attention module performs attention representation extraction, the embedded information in the first embedding representation and the second embedding representation are fused to obtain a fused embedding representation. In this way, the information of the mRNA sequence is fused with the corresponding secondary structure information, so that the predicted ribosome occupancy data has cell type specificity, thereby improving the accuracy of the prediction.

[0151] In one embodiment, Figure 6 As shown, a training method for an mRNA ribosome occupancy prediction model is provided. This embodiment uses the method applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices, portable wearable devices, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services, etc.

[0152] In this embodiment, the training method of the mRNA ribosome occupancy prediction model can be used to train the ribosome occupancy prediction model in the above embodiment; the method includes the following steps S601 to S606. Among them:

[0153] Step S601, constructing a multimodal database.

[0154] Among them, the multimodal database refers to a database of omics data samples with multiple data modalities. The multimodal database of this embodiment includes mRNA sequence samples and secondary structure samples obtained by sequencing the mRNA sequence samples in the current cell type.

[0155] Among them, the mRNA sequence sample refers to the mRNA sequence used for model training; the secondary structure sample refers to the secondary structure used for model training.

[0156] Exemplarily, an mRNA sequence sample and a secondary structure sample obtained by sequencing the mRNA sequence sample in the current cell type are obtained, and a multimodal database is constructed based on the mRNA sequence sample and its corresponding secondary structure sample.

[0157] Step S602: inputting the mRNA sequence samples in the multimodal database and the secondary structure samples corresponding to the mRNA sequence samples into the prediction model to be trained.

[0158] Step S603: encode the currently input mRNA sequence sample through the first encoder in the prediction model to be trained to obtain the current first embedding representation.

[0159] Exemplarily, the mRNA sequence sample is input into a first encoder in a prediction model to be trained, and the first encoder encodes the currently input mRNA sequence sample to obtain a current first embedding representation of the currently input mRNA sequence sample.

[0160] Among them, the process in which the first encoder in this embodiment encodes the currently input mRNA sequence sample to obtain the current first embedded representation is the same or similar to the process in which the first encoder in the above embodiment encodes the mRNA sequence to obtain the first embedded representation. For details, please refer to the records in the above embodiment, which will not be repeated in this embodiment.

[0161] Step S604: encode the currently input secondary structure sample through the second encoder in the prediction model to be trained to obtain the current second embedding representation.

[0162] Exemplarily, the secondary structure sample is input into the second encoder in the prediction model to be trained, and the second encoder encodes the currently input secondary structure sample to obtain a current second embedding representation of the currently input secondary structure sample.

[0163] Among them, the process in which the second encoder in this embodiment encodes the current input secondary structure sample to obtain the current second embedded representation is the same or similar to the process in which the second encoder in the above embodiment encodes the secondary structure to obtain the second embedded representation. For details, please refer to the records in the above embodiment, which will not be repeated in this embodiment.

[0164] Step S605, input the current first embedding representation and the current second embedding representation into the multi-head self-attention module in the prediction model to be trained to fuse and classify the current first embedding representation and the current second embedding representation, and based on the classification result, obtain the ribosome occupancy prediction data on the current mRNA sequence sample in the current cell type.

[0165] Among them, the current cell type is the cell type corresponding to the secondary structure sample currently input; that is, the cell type corresponding to the secondary structure sample used for training; the current cell type of this embodiment can be any cell type, and the current cell type can be different from the cell type (that is, the target cell type) of the prediction model in the actual prediction process.

[0166] Exemplarily, the current first embedding representation and the current second embedding representation are input into the multi-head self-attention module of the prediction model to be trained, and the current first embedding representation and the current second embedding representation are fused based on the multi-head self-attention module to obtain the fused current embedding representation, and then the fused current embedding representation is classified, and based on the classification result of the current embedding representation, the ribosome occupancy prediction data on the current mRNA sequence sample in the current cell type is obtained.

[0167] Step S606, based on the ribosome occupancy prediction data and the actual ribosome occupancy data on the current mRNA sequence in the current cell type, the model parameters of the prediction model to be trained are updated until a trained prediction model is obtained.

[0168] The actual data of ribosome occupancy on the current mRNA sequence in the current cell type refers to the ribosome occupancy data obtained by sequencing the acquired mRNA sequence samples using the translation group sequencing technology, that is, the single nucleotide precision ribosome distribution data.

[0169] Exemplarily, based on the predicted ribosome occupancy data and the actual ribosome occupancy data on the current mRNA sequence in the current cell type, the deviation between the predicted ribosome occupancy data and the actual ribosome occupancy data is determined, and the model parameters of the prediction model are updated based on the deviation until a trained prediction model is obtained.

[0170] Among them, when the deviation is determined and the model parameters are updated, it can be determined based on a loss function, that is, a loss function is determined, and the deviation between the predicted ribosome occupancy data and the actual ribosome occupancy data is determined based on the loss function, and the model parameters of the prediction model are updated according to the deviation until the loss function converges or reaches the maximum number of iterations, thereby obtaining a trained prediction model.

[0171] Exemplarily, the loss function of the prediction model is determined, and the loss function is solved according to the predicted ribosome occupancy data and the actual ribosome occupancy data to obtain the value of the loss function (i.e., the deviation between the predicted ribosome occupancy data and the actual ribosome occupancy data), and it is determined whether the deviation converges or reaches the set maximum number of iterations. If so, the training is terminated to obtain a trained prediction model. Otherwise, the model parameters of the prediction model are updated and optimized according to the deviation, and a new round of training is performed on the prediction model until the deviation converges or the maximum number of iterations is reached to obtain a trained prediction model.

[0172] Among them, the loss function may include but is not limited to logarithmic loss function, mean square error, mean absolute error, etc. When determining the loss function, a loss function can be defined for the first encoder, the second encoder and the multi-head attention module respectively, and then an overall loss function is defined for the prediction model. After solving the loss functions of the first encoder, the second encoder and the multi-head attention module respectively, the loss function of the overall prediction model is calculated based on the results of the loss functions of the three to determine the deviation between the predicted ribosome occupancy data and the actual ribosome occupancy data.

[0173] The training method of the mRNA ribosome occupancy prediction model is as follows: a multimodal database is constructed, and then the mRNA sequence samples and the corresponding secondary structure samples in the input multimodal database are respectively encoded by the first encoder and the second encoder in the prediction model to be trained, and after obtaining the current first embedding representation and the current second embedding representation, the current first embedding representation and the current second embedding representation are fused and classified, and based on the classification result, the ribosome occupancy prediction data on the current mRNA sequence sample in the current cell type is obtained, and based on the ribosome occupancy prediction data and the actual ribosome occupancy data on the current mRNA sequence sample in the current cell type, the training model to be trained is trained. The model parameters of the trained prediction model are updated until a trained prediction model is obtained. The constructed multimodal database contains multi-modal omics data such as mRNA sequence samples with commonality of cell types and secondary structure samples with cell type specificity, and the prediction model is trained based on this input, integrating the common characteristics and specific characteristics of each cell type, so that the trained prediction model can not only improve the accuracy of the prediction of ribosome occupancy of a single cell type, but also show advantages in the prediction of ribosome occupancy with cell type specificity, thereby making the trained prediction model transferable in the prediction of ribosome occupancy of different cell types.

[0174] In one embodiment, the multimodal database also includes actual data on ribosome occupancy of mRNA sequence samples. Figure 7 As shown, the training method of the mRNA ribosome occupancy prediction model further includes the following steps S701 to S703 before constructing the multimodal database. Among them:

[0175] Step S701, obtaining mRNA sequence samples.

[0176] Step S702, determining the secondary structure sample of the mRNA sequence sample in the current cell type, and detecting the structural strength of the secondary structure sample to obtain the RNA structure probe score of the secondary structure sample.

[0177] Among them, the structural strength of the secondary structure sample refers to the interaction between bases within the molecule of the mRNA sequence sample, the formation of hydrogen bonds, and the stability of the overall structure, which can be characterized by the RNA structure probe score with single nucleotide accuracy. The RNA structure probe score refers to an indicator for evaluating the stability of a specific structural region of the molecule of the mRNA sequence sample. The larger the RNA structure probe score, the greater the strength of the secondary structure, that is, the interaction between bases, the formation of hydrogen bonds, and the overall structure are more stable. The RNA structure probe score of this embodiment can be measured using icSHAPE (Selective 2'Hydroxyl Acylation analyzed by Primer Extension) sequencing technology, that is, selective 2'-hydroxy acylation analysis combined with primer extension sequencing.

[0178] Exemplarily, based on the current cell type, the secondary structure sample of the acquired mRNA sequence sample in the current cell type is determined, and the structural strength of the secondary structure sample is detected based on the icSHAPE sequencing technology to obtain the RNA structure probe score of the secondary structure sample.

[0179] Step S703, determining the actual data of ribosome occupancy of the mRNA sequence sample in the current cell type.

[0180] Exemplarily, the translation group sequencing technology is used to sequence the mRNA fragments in the translation state on the obtained mRNA sequence sample to determine the number of all translatable ribosomes on the mRNA sequence sample, so as to obtain the distribution data of ribosomes on the mRNA sequence sample, and then obtain the actual data of ribosome occupancy.

[0181] In one embodiment, based on the above process, constructing a multimodal database includes:

[0182] A multimodal database is constructed based on mRNA sequence samples, secondary structure samples and actual ribosome occupancy data.

[0183] In this embodiment, by determining the mRNA sequence samples and their corresponding secondary structure samples, the RNA structure probe scores of the secondary structure samples and the actual data of the ribosome occupancy of the mRNA sequence samples in the current cell type are obtained, thereby constructing a multimodal database, which can comprehensively consider various omics data during the prediction model training process, enrich the training samples, and provide the prediction accuracy of the trained prediction model.

[0184] In one embodiment, Figure 8 As shown, step S703 determines the actual data of ribosome occupancy of the mRNA sequence sample in the current cell type, including the following steps S801 to S804, wherein:

[0185] Step S801, determining the mRNA fragments in the translation state on the mRNA sequence sample to obtain the original Ribo-Seq data.

[0186] Among them, the mRNA fragment in the translation state refers to the mRNA fragment that is bound to the ribosome and is being translated on the mRNA sequence sample.

[0187] Among them, the original Ribo-Seq data refers to the data obtained by sequencing mRNA sequence samples using Ribo-Seq sequencing technology (Ribosome profiling, i.e. ribosome footprint analysis). Ribo-seq (Ribosome profiling) uses RNA to digest RNA in cells to obtain mRNA fragments that are being translated and protected by ribosomes (ribosome footprint), and then enriches, deeply sequences, and analyzes these mRNA fragments of a certain length protected by ribosomes.

[0188] Exemplarily, Ribo-Seq sequencing technology is used to identify mRNA in a translation state on an mRNA sequence sample to obtain raw Ribo-Seq data.

[0189] Step S802, extracting the original Ribo-Seq data to obtain sequencing fragments that meet preset quality conditions; matching the sequencing fragments with a predetermined reference genome to obtain the initial positions of the sequencing fragments.

[0190] Among them, the reference genome refers to a standardized, digital nucleotide sequence database, which is an idealized model generated by summarizing and optimizing the genome sequences of multiple biological individuals from the same population, representing the typical genome sequence of a species. In this embodiment, the reference genome selects the GENCODE Hg38 V26 version of the human database, which can provide annotation information on the location and structure of genes, transcripts, exons, introns, etc. in the human genome, and includes updates on gene structure, discovery of new genes, and correction information on known gene annotations.

[0191] The preset quality condition refers to a threshold condition used to measure whether the original Ribo-Seq data meets the matching requirements.

[0192] Among them, sequencing fragments refer to the fragments that meet the requirements when matching the original Ribo-Seq data with the reference genome; the initial positions of sequencing fragments refer to the corresponding positions of each sequencing fragment on the reference genome when mapping the original Ribo-Seq data to the reference genome.

[0193] Exemplarily, original Ribo-Seq data is obtained, and the quality of mRNA fragments (i.e., sequencing fragments) in the original Ribo-Seq data is detected and controlled based on a quality control program to obtain sequencing fragments that meet preset quality conditions; then, the sequencing fragments are matched with a predetermined reference genome to obtain the initial position of the sequencing fragments.

[0194] In specific implementation, the quality control program of this embodiment adopts FastQC (A Quality Control tool for High Throughput Sequence Data, i.e., a quality control program for high-throughput sequence data). The quality of each mRNA fragment in the original Ribo-Seq data is detected and controlled by FastQC, and then the low-quality sequencing fragments that do not meet the conditions in each mRNA fragment and the Adapter sequence fragments in the quality control process are cut out by the Cutadapt algorithm to obtain sequencing fragments that meet the preset quality conditions; then, the sequencing fragments are matched with a predetermined reference genome using BWA (Burrows-Wheeler-Alignment) software to obtain the initial position of the sequencing fragments.

[0195] The Adapter sequence fragment is a short nucleic acid chain with a known sequence, which is used to link different target fragments with unknown sequences during the quality control process.

[0196] Step S803, correcting the initial position of the sequencing fragment.

[0197] Among them, since the movement of ribosomes on the mRNA sequence sample is not completely uniform, the original Ribo-Seq data obtained by sequencing may have a certain deviation from the actual translation position. Correcting the initial position of the sequencing fragment is helpful to obtain accurate position information.

[0198] Exemplarily, the Metagene analysis mode (i.e., macrogene analysis mode) in the Plastid script is used to read the original Ribo-Seq data and gene annotation files, determine the ribosome density distribution near the start codon and the stop codon of each gene, and obtain the density distribution map of ribosomes at different positions of the gene (including near the start codon and the stop codon). Based on the density distribution map, the position of the A-site (ribosome A-site, i.e., the position where the mRNA codon is being read) and the P-site (ribosome P-site, i.e., the position where the previous codon is read) relative to the start codon and the stop codon is determined, and then according to the position of the A-site and the P-site, the distance between the A-site and the P-site is determined, and then the deviation between the mRNA fragment protected by the ribosome and the actual translation position is determined. Finally, the sequencing fragments in the original Ribo-Seq data are translated or adjusted based on the deviation to achieve the correction of the initial position of the sequencing fragments.

[0199] Among them, gene annotation files refer to files containing the locations and properties of genes, transcripts, exons, introns, promoters, terminators and other elements in the genome sequence. Gene annotation files include gene location information, transcript information, exon and intron information, and other functional elements, such as the locations and properties of promoters, terminators, regulatory elements, etc. Gene annotation files can be obtained from public databases (such as NCBI's GenBank, Ensembl, etc.) or genomics research. In addition, genomics tools or software can be used to generate custom gene annotation files.

[0200] The positions of the A-site and the P-site can be determined by analyzing the positions of the ribosome density peaks in the density distribution map.

[0201] Step S804, constructing an exon reference genome, matching the corrected sequencing fragments with the exon reference genome, obtaining ribosome position information with single-base accuracy, and determining actual ribosome occupancy data based on the ribosome position information with single-base accuracy.

[0202] Among them, the exon reference genome refers to a database or data set containing exon information constructed based on the reference genome annotation file. Exons are part of the gene of eukaryotic organisms and contain the core information required for protein synthesis. The gene sequence that appears in mature RNA is also called the expression sequence.

[0203] Exemplarily, an exon reference genome is constructed based on a gene annotation file, and the Plastid script is used to match the corrected sequencing fragments with the exon reference genome to obtain ribosome position information with single-base accuracy, and then the actual ribosome occupancy data is determined based on the ribosome position information with single-base accuracy.

[0204] In an optional embodiment, for the process of determining the actual ribosome occupancy data of the mRNA sequence sample in the current cell type, multi-cell type data containing multiple biological repetitions can be collected to perform the above steps, and the obtained actual ribosome occupancy data can be averaged to obtain the final actual ribosome occupancy data, so as to obtain reliable single nucleotide precision ribosome distribution data.

[0205] In this embodiment, the original Ribo-Seq data is obtained by sequencing the mRNA sequence sample, and after the quality control of the original Ribo-Seq data, the initial position of the sequencing fragment is corrected, and the position matching is performed based on the corrected sequencing fragment to obtain the ribosome position information with single-base accuracy, and the actual data of the ribosome occupancy is determined based on the ribosome position information with single-base accuracy, so that reliable single-nucleotide accuracy ribosome distribution data can be obtained, further ensuring the accuracy of the prediction results.

[0206] In one embodiment, Fig. 9 As shown, the training method of the mRNA ribosome occupancy prediction model further includes the following steps S901 to S903 before inputting the mRNA sequence samples and the secondary structure samples corresponding to the mRNA sequence samples in the multimodal database into the prediction model to be trained, wherein:

[0207] Step S901, aligning the RNA structure probe scores and actual ribosome occupancy data of the secondary structure sample with a predetermined reference genome at the transcript level to obtain a number of transcripts carrying sequence information, structure information and ribosome occupancy information.

[0208] Among them, transcript refers to one or more mature mRNAs that can encode proteins formed by a gene through the transcription process, that is, transcript specifically refers to the mature mRNA part that can encode proteins, which is a specific form or functional state of mRNA; in the transcription process, the genetic information of DNA is transcribed into RNA, a part of which is mature mRNA, namely transcript.

[0209] Among them, aligning the RNA structure probe score and the actual data of the ribosome occupancy of the secondary structure sample with the predetermined reference genome at the transcript level refers to the process of accurately matching and corresponding the RNA structure probe score and the actual data of the ribosome occupancy of the secondary structure sample with the gene information in the reference genome at the transcript level defined by the reference genome based on the similarity of the sequence and / or the positional relationship. Since the genome levels matched by the RNA structure probe score and the actual data of the ribosome occupancy of the secondary structure sample in this embodiment are different, the genome level matched by the RNA structure probe score of the secondary structure sample is the transcript sequence, while the actual data of the ribosome occupancy is matched to the coding region of the gene, but for the RNA structure probe score and the actual data of the ribosome occupancy of the secondary structure sample, both can determine their positions on the transcripts of the reference genome, and by comparison and analysis, the RNA structure probe score and the actual data of the ribosome occupancy of the secondary structure sample are associated with the corresponding transcript sequence, so as to align the RNA structure probe score and the actual data of the ribosome occupancy of the secondary structure sample to determine the intersection of the two modal omics data at the cell type level.

[0210] Exemplarily, the RNA structure probe score and the actual ribosome occupancy data of the secondary structure sample are respectively compared with the reference genome, and based on the comparison results, the RNA structure probe data and the ribosome occupancy data are aligned at the transcript level. Subsequently, the sequence information is extracted from the comparison results, including the base sequence and position on the mRNA sequence sample, and the structural information of the secondary structure sample is extracted based on the RNA structure probe data, including base pairing, loop structure, etc., and the position and quantity information of the ribosome on each transcript is extracted from the ribosome occupancy data, and the sequence information, structural information and ribosome occupancy information are integrated to form a transcript carrying sequence information, structural information and ribosome occupancy information.

[0211] Among them, when the RNA structure probe scores and actual ribosome occupancy data of the secondary structure samples are compared with the reference genome respectively, specialized bioinformatics tools (such as HISAT2, STAR, etc.) can be used to efficiently process the sequencing data and accurately find the sequences in the RNA structure probe scores and actual ribosome occupancy data of the secondary structure samples that match the reference genome.

[0212] Step S902 , taking transcripts as units, based on the actual data of ribosome occupancy and the length of the transcripts, the transcripts are denoised to obtain training samples, and a training data set is constructed based on the training samples.

[0213] Among them, denoising processing refers to processing the transcripts to a certain extent according to the set conditions to obtain transcripts that meet the conditions.

[0214] Exemplarily, the transcripts carrying sequence information, structural information, ribosome occupancy information and other information obtained in the aforementioned steps are taken as units, and the length of the transcripts is obtained, and the transcripts are denoised based on the actual ribosome occupancy data and the length of the transcripts; training samples are obtained based on the transcripts remaining after denoising, and the training samples are divided into a training set and a test set according to a set ratio to construct a training data set.

[0215] Step S903, extracting sequence information and structure information corresponding to each training sample in the training data set, determining the mRNA sequence corresponding to the extracted sequence information as the currently input mRNA sequence sample, and determining the secondary structure sample corresponding to the extracted structure information as the currently input secondary structure sample.

[0216] In this embodiment, by aligning the RNA structure probe score and the actual ribosome occupancy data of the secondary structure sample with a predetermined reference genome at the transcript level, a number of transcripts carrying sequence information, structural information and ribosome occupancy information are obtained, and the transcripts are denoised to extract the corresponding sequence information and structural information to obtain the currently input mRNA sequence sample and the currently input secondary structure sample, by aligning the two types of omics data at the transcript level, and integrating the sequence information of the transcripts and the expression information of the transcripts in different cell lines (i.e., structural information) in the final training sample, after multiple steps of screening and alignment, the intersection of the two types of omics data at the cell type level is taken, which can effectively realize the fusion prediction of multiple omics data.

[0217] In one embodiment, step S902 of performing denoising on the transcript includes:

[0218] Detect the length of each transcript and remove transcripts that are shorter than a preset length; extract a number of bases at the head and a number of bases at the tail of each transcript remaining after removing the transcripts of the preset length, and splice the number of bases at the head and a number of bases at the tail of each transcript to form a prediction sample; determine the actual data of ribosome occupancy of each prediction sample, and remove the prediction samples with the actual data of ribosome occupancy being zero; sort the remaining prediction samples after removing the prediction samples with the actual data of ribosome occupancy being zero according to the size of the actual data of ribosome occupancy, extract a number of prediction samples with the top ranking of the actual data of ribosome occupancy to form positive samples, and extract a number of prediction samples with the bottom ranking of the actual data of ribosome occupancy to form negative samples; integrate the positive samples and the negative samples to obtain a training data set.

[0219] The first several bases refer to the several bases located at the front end or the starting position of the transcript; the tail several bases refer to the several bases located at the back end or the end position of the transcript.

[0220] Exemplarily, the length of each transcript is detected, and the transcripts with a length of less than 500nt are removed in units of transcripts; among the remaining transcripts, the sequences of the first 500 bases and the last 500 bases on each transcript are extracted and spliced ​​to form a prediction sample with a length of 1000nt; based on the actual ribosome occupancy data of the spliced ​​prediction samples, the prediction samples with the actual ribosome occupancy data of zero are removed; based on the remaining sequencing samples after removing the prediction samples with the actual ribosome occupancy data of zero, the remaining sequencing samples are sorted according to the size of the actual ribosome occupancy data, and the top 10,000 prediction samples ranked by the actual ribosome occupancy data are extracted to form positive samples, and the bottom 10,000 prediction samples ranked by the actual ribosome occupancy data are extracted to form negative samples; after integrating the positive samples and the negative samples, 16,000 prediction samples are randomly selected as the training set, and the remaining 4,000 prediction samples are used as the test set to obtain the training data set.

[0221] It is understandable that the length, number of bases, and sample data involved in the above steps are only examples and are not intended to limit the present embodiment. In actual denoising, they can be adjusted according to the model training requirements.

[0222] In this embodiment, by denoising the transcripts, the noise in the ribosome distribution data can be removed, thereby reducing the interference of noise on model training and avoiding the influence of non-coding RNA on model training, thereby achieving the purpose of further improving the prediction accuracy.

[0223] In a specific application example, taking the RNA sequences of 293T cell line and HeLa cell line (i.e., cell types) and their corresponding single nucleotide accurate RNA structure probe scores (or secondary structures) as examples, the RNA sequences of the two cell lines are completely consistent, but the RNA structure probe scores change dynamically to reflect the cell line-specific RNA structure information. Based on the RNA sequences and their corresponding single nucleotide accurate RNA structure probe scores, this example integrates translation-related multimodal data, i.e., mRNA sequence information and secondary structure information. After the two are fused, not only can the prediction accuracy of ribosome occupancy under a single condition be improved, but also the prediction accuracy of ribosome occupancy under cell line-specific conditions can be improved. Advantages can also be shown in the task of predicting cell line-specific ribosome occupancy.

[0224] like Fig.10As shown in FIG, it is a structural block diagram of the prediction model in this example. For mRNA sequence information, the Bert module (i.e., the first encoder) is used for embedded encoding. For RNA structure probe scores, the Masked attention module (i.e., the second encoder) is used for embedded encoding to process the missing values ​​in the RNA structure probe scores, and the embedded representations of the two modal data (i.e., the first embedded representation and the second embedded representation) are obtained. Finally, the embedded representations of the two modal data are integrated through a Cross-attention module (i.e., the multi-head self-attention module), and then connected to the downstream classification module to finally predict the ribosome occupancy.

[0225] The following is an explanation of the training method of the mRNA ribosome occupancy prediction model of this example.

[0226] First, in the construction phase of the multimodal database, the translation-related multi-omics data mRNA sequence data (i.e., mRNA sequence samples), icSHAPE sequencing data (i.e., secondary structure samples or RNA structure probe scores), and Ribo-Seq ribosome occupancy data (i.e., actual ribosome occupancy data) were used. For icSHAPE sequencing data (intracellular RNA structure probe scores), two types of data were collected from Hela cell lines and 293T cell lines under the same batch of treatment. For Ribo-Seq ribosome occupancy data, the original data processing process constructed by ourselves was used to collect and standardize multi-cell line data containing more than three biological replicates to obtain reliable single-nucleotide precision ribosome distribution data, thereby determining the actual ribosome occupancy data.

[0227] icSHAPE sequencing data (RNA structure probe score) and Ribo-Seq ribosome occupancy data reflect the probability of mRNA double-strand pairing on the mRNA sequence and the distribution of ribosomes on the mRNA, respectively. In theory, both omics technologies can achieve quantitative measurement with single nucleotide accuracy. However, due to the different genomic levels matched by the two omics data, RNA structure probe data often matches the transcript sequence, while Ribo-Seq ribosome occupancy data mainly matches the coding region of the gene. Due to the technical difficulty, it is difficult to promote multi-omics related to simultaneous quantitative translation in experiments. Therefore, it is necessary to align the experimental conditions and cell lines of the aforementioned single-base-accurate omics data mRNA sequence data, icSHAPE sequencing data, and Ribo-Seq ribosome occupancy data.

[0228] Specifically, by screening out experimental samples under control conditions without special treatment, using a unified human reference genome version (GENCODE V26 version human reference genome), the two omics data were realigned at the transcript level, and the sequence information of the transcripts and the expression information of the transcripts in different cell lines were integrated into the final database. After multiple steps of screening and alignment, the intersection of the two omics data at the cell line level was taken, and finally the single nucleotide precision quantitative data of the HeLa cell line and the 293T cell line in the two omics were obtained. The two omics data were aligned according to the transcript sequence, which will be used to explore the prediction of RNA sequence properties and the integration of translation-related multimodal information.

[0229] Secondly, in terms of model prediction targets and model training details, since there is a certain amount of noise in the ribosome distribution data, this example uses transcripts as units, extracts the top 10,000 transcripts with the highest and lowest ribosome distributions on the transcripts, and converts the target into a binary classification task.

[0230] Specifically, for the mRNA sequences and RNA structure probe scores of 293T cell lines and HeLa cell lines, transcripts with a length of less than 500nt were deleted, and the sequences of the first 500 bases and the last 500 bases of each transcript were extracted and spliced ​​to form a sample with a length of 1000nt. In order to reduce the interference of noise and avoid the influence of non-coding RNA, transcripts with a ribosome occupancy rate of 0 were deleted, and then the first 10,000 transcripts and the last 10,000 transcripts of ribosome occupancy were extracted to form a positive and negative sample data set. 16,000 samples were randomly selected for training and 4,000 samples were used for prediction.

[0231] During training, for mRNA sequence data, the Bert module is used, and 4 layers of BERT attention layers (i.e., the first encoding layer) are used. Each BERT attention layer contains 4 attention heads and the hidden layer dimension is 128 to embed the mRNA sequence data.

[0232] For the RNA structure probe score, the Masked attention module is used, and a 4-layer BERT attention layer (i.e., the second encoding layer) is used. Each BERT attention layer contains 4 attention heads and the hidden dimension is 128. The RNA structure probe score is embedded and encoded. Since the RNA structure probe score is a continuous value from 0 to 1 and there are a considerable number of missing values, the position of the missing value is masked, and the position of the normal value is discretized by dividing it into 10 equal parts from 0 to 1 to process the missing values ​​in the RNA structure probe score and obtain the corresponding embedded information.

[0233] For the fusion of mRNA sequence information and RNA structure probe scores, the embedding representation corresponding to the mRNA sequence information (first embedding representation) is used as the K (key) and V (value) of the attention layer of the Cross-attention module, and the embedding representation corresponding to the RNA structure probe score (second embedding representation) is used as the Q (query) of the attention layer of the Cross-attention module. The Cross-attention module uses 8 attention heads and a hidden layer dimension of 128 to fuse the mRNA sequence information and RNA structure probe scores.

[0234] For the classification of the fused embedding representation, the maximum pooling operation is first used to compress the sequence length (i.e., the length of the fused embedding representation) dimension to 1, and then two fully connected layers are used for classification. The training process has a learning rate of 1e-3, a weight decay of 1e-3, a batch size of 64, and a dropout of 0.1. The model is trained on four 3090 graphics cards to obtain the trained prediction model.

[0235] Since the secondary structure sequence used in the model training process covers cell type-specific information, after the model is trained based on the mRNA sequence and its secondary structure in any type of cell line, the prediction of the ribosome occupancy of the mRNA sequence in any other cell line only requires replacing the secondary structure with the secondary structure in the corresponding cell line during the prediction.

[0236] To illustrate the advantages of this example, by screening the protein-coding mRNA transcript data, in order to reduce the impact of ribosome distribution data noise, the ribosome occupancy of each transcript is extracted on a transcript basis, and after sorting, the ribosome occupancy is divided into two categories, converting the prediction target into a binary classification problem. First, the training set and test set are divided on a single cell line data, and three different input training models are used respectively. The three inputs include using only the mRNA sequence as input, using only the RNA structure probe score as input, and integrating the mRNA sequence and RNA structure probe score as input. Fig.11As shown, when the model is trained on the 293T cell line with three different inputs and the ribosome occupancy is predicted on the 293T cell line, the accuracy, precision, recall rate (Recall), F1 score (F1Score), AUC value (Area Under Curve, area under the ROC curve), AUPRC value (Area Under the Precision-Recall Curve, area under the precision-recall curve) of the prediction model are compared under three inputs. It can be seen that the accuracy of using RNA structure probe scores alone is much better than that of using mRNA sequence data alone, and the prediction effect of using the fusion data of mRNA sequence and RNA structure probe scores is further improved.

[0237] Furthermore, to illustrate the transferability of the method in this example between different cell lines, based on the above process, the model was trained on 293T cell line data and predicted on HeLa cell line data. Fig.12 As shown in the figure, although the prediction effects of the three input types have declined, the advantage of using the fusion data of mRNA sequence and RNA structure probe score is still very obvious, and is better than the input using only one type of information in almost all indicators.

[0238] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0239] Based on the same inventive concept, the embodiment of the present application also provides a device for predicting the ribosome occupancy of an mRNA sequence for implementing the method for predicting the ribosome occupancy of an mRNA sequence involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of one or more devices for predicting the ribosome occupancy of an mRNA sequence provided below can refer to the limitations of the method for predicting the ribosome occupancy of an mRNA sequence above, and will not be repeated here.

[0240] In an exemplary embodiment, Fig.13As shown, a prediction device for mRNA sequence ribosome occupancy is provided, comprising: a data acquisition module 1301, a model calling module 1302 and a prediction module 1303, wherein:

[0241] The data acquisition module 1301 is used to acquire the mRNA sequence and the secondary structure of the mRNA sequence in the target cell type.

[0242] The model calling module 1302 is used to input the mRNA sequence and its secondary structure into a pre-trained prediction model, obtain a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure through different encoders in the prediction model, fuse the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation, and perform classification based on the fused embedding representation.

[0243] The prediction module 1303 is used to obtain the prediction result of the ribosome occupancy of the mRNA sequence in the target cell type based on the classification output of the prediction model.

[0244] Based on the same inventive concept, the embodiment of the present application also provides a training device for an mRNA ribosome occupancy prediction model for implementing the training method for the mRNA ribosome occupancy prediction model involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of the training device for one or more mRNA ribosome occupancy prediction models provided below can be referred to the limitations of the training method for the mRNA ribosome occupancy prediction model above, and will not be repeated here.

[0245] In an exemplary embodiment, Fig.14 As shown, a training device for an mRNA ribosome occupancy prediction model is provided, comprising: a database construction module 1401, a data input module 1402 and a parameter update module 1403, wherein:

[0246] A database construction module 1401 is used to construct a multimodal database; the multimodal database includes mRNA sequences and secondary structures of mRNA sequences in any cell type, and the mRNA sequence is determined as the current mRNA sequence and the secondary structure of the mRNA sequence in any cell type is determined as the current structure sequence;

[0247] The data input module 1402 is used to input the current mRNA sequence and the current structure sequence into the prediction model to be trained; encode the current mRNA sequence through the first encoder in the prediction model to be trained to obtain the current first embedding representation; encode the current structure sequence through the second encoder in the prediction model to be trained to obtain the current second embedding representation; input the current first embedding representation and the current second embedding representation into the multi-head self-attention module in the prediction model to be trained to fuse and classify the first embedding representation and the second embedding representation, and obtain the ribosome occupancy prediction data on the current mRNA sequence based on the classification structure of the embedding representation;

[0248] The parameter updating module 1403 is used to update the model parameters of the prediction model to be trained based on the ribosome occupancy prediction data and the actual ribosome occupancy data on the current mRNA sequence until a trained prediction model is obtained.

[0249] Each module in the above-mentioned mRNA sequence ribosome occupancy prediction device and mRNA ribosome occupancy prediction model training device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0250] In an exemplary embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface, the display unit and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented by WIFI, a mobile cellular network, a near field communication (Near Field Communication, NFC) or other technologies. When the computer program is executed by the processor, a method for predicting the ribosome occupancy of an mRNA sequence and / or a method for training a prediction model for the ribosome occupancy of an mRNA sequence are implemented. The display unit of the computer device is used to form a visually visible picture, and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball, or touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0251] It is understandable that in some other embodiments, the computer device may also be a server, including a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used for data such as mRNA sequences, secondary structures, RNA structure probe scores, and actual ribosome occupancy data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for predicting the ribosome occupancy of an mRNA sequence and / or a method for training a prediction model for the mRNA ribosome occupancy.

[0252] Those skilled in the art will appreciate that the structure of the computer device described above is merely a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than described, or combine certain components, or have a different arrangement of components.

[0253] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the method for predicting the mRNA sequence ribosome occupancy and / or the method for training the mRNA ribosome occupancy prediction model of the above-mentioned embodiment is implemented.

[0254] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for predicting the ribosome occupancy of an mRNA sequence and / or the method for training an mRNA ribosome occupancy prediction model of the above-mentioned embodiment is implemented.

[0255] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the method for predicting the ribosome occupancy of an mRNA sequence and / or the method for training an mRNA ribosome occupancy prediction model of the above-mentioned embodiment.

[0256] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0257] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0258] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0259] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for predicting ribosome occupancy of mRNA sequence, characterized in that: The method comprises: Obtaining an mRNA sequence and a secondary structure of the mRNA sequence in a target cell type; Inputting the mRNA sequence and its secondary structure into a pre-trained prediction model, obtaining a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure through different encoders in the prediction model, fusing the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation, and performing classification based on the fused embedding representation; Based on the classification output of the prediction model, a prediction result of the ribosome occupancy of the mRNA sequence in the target cell type is obtained.

2. The method according to claim 1, characterized in that The mRNA sequence and its secondary structure are input into a pre-trained prediction model, a first embedding representation of the mRNA sequence and a second embedding representation of the secondary structure are obtained respectively through different encoders in the prediction model, the first embedding representation and the second embedding representation are fused based on a multi-head self-attention mechanism to obtain a fused embedding representation, and classification is performed based on the fused embedding representation, including: Inputting the mRNA sequence and its secondary structure into a pre-trained prediction model; The mRNA sequence is encoded by a first encoder in the prediction model to obtain a first embedded representation of the mRNA sequence; the first encoder encodes the mRNA sequence based on a self-attention mechanism; The second encoder in the prediction model encodes the secondary structure to obtain a second embedding representation of the secondary structure; the second encoder encodes the secondary structure based on a masked self-attention mechanism; fusing the first embedding representation and the second embedding representation based on a multi-head self-attention mechanism to obtain a fused embedding representation; The fused embedded representation is subjected to a maximum pooling operation to reduce the dimension of the fused embedded representation, and the embedded representation after the dimension reduction is input into a fully connected layer for classification.

3. The method according to claim 2, characterized in that The encoding of the mRNA sequence by a first encoder in the prediction model to obtain a first embedded representation of the mRNA sequence includes: Inputting the mRNA sequence into the first encoder; Vectorizing the mRNA sequence through a first embedding layer of the first encoder to obtain a first vector sequence; Iteratively extracting first embedded information from the first vector sequence through multiple first encoding layers of the first encoder; each of the first encoding layers processes the first vector sequence using a self-attention mechanism, and the input of the first first encoding layer is the first vector sequence, and the input of each subsequent first encoding layer is the first embedded information output by the previous first encoding layer; The first embedding information output by the last first encoding layer is pooled to obtain a first embedding representation of the mRNA sequence.

4. The method according to claim 2, characterized in that: The encoding of the secondary structure by a second encoder in the prediction model to obtain a second embedding representation of the secondary structure includes: inputting the secondary structure into the second encoder; vectorizing the secondary structure through a second embedding layer of the second encoder to obtain a second vector sequence; Iteratively extracting the second embedding information from the second vector sequence through multiple second encoding layers of the second encoder; each of the second encoding layers uses a masked self-attention mechanism to process the second vector sequence, the mask position of the masked self-attention mechanism is the position where the secondary structure on the mRNA sequence is missing, the input of the first second encoding layer is the second vector sequence, and the input of each subsequent second encoding layer is the second embedding information output by the previous second encoding layer; Pooling is performed on the second embedding information output by the last second encoding layer to obtain a second embedding representation of the secondary structure.

5. The method according to any one of claims 1 to 4, characterized in that: The fusing the first embedding representation and the second embedding representation based on the multi-head self-attention mechanism to obtain a fused embedding representation includes: Determining a key matrix and a value matrix of the multi-head self-attention mechanism based on the first embedding representation, and determining a query matrix of the multi-head self-attention mechanism based on the second embedding representation; The elements in the key matrix, the value matrix, and the query matrix are divided according to the number of attention heads preset by the multi-head self-attention mechanism, and the attention weight of each attention head is determined; The output of the corresponding attention head is determined based on each attention weight, and the outputs of each attention head are concatenated to obtain a fused embedded representation.

6. A method for training an mRNA ribosome occupancy prediction model, characterized in that: The method comprises: Constructing a multimodal database; the multimodal database includes mRNA sequence samples and secondary structure samples obtained by sequencing the mRNA sequence samples in the current cell type; Inputting the mRNA sequence samples in the multimodal database and the secondary structure samples corresponding to the mRNA sequence samples into the prediction model to be trained; Encoding the currently input mRNA sequence sample by a first encoder in the prediction model to be trained to obtain a current first embedding representation; Encoding the currently input secondary structure sample by a second encoder in the prediction model to be trained to obtain a current second embedding representation; Inputting the current first embedding representation and the current second embedding representation into the multi-head self-attention module in the prediction model to be trained to fuse and classify the current first embedding representation and the current second embedding representation, and obtaining ribosome occupancy prediction data on the current mRNA sequence sample in the current cell type based on the classification result; the current cell type is the cell type corresponding to the secondary structure sample currently input; Based on the ribosome occupancy prediction data and the actual ribosome occupancy data of the current mRNA sequence sample in the current cell type, the model parameters of the prediction model to be trained are updated until a trained prediction model is obtained.

7. The method according to claim 6, characterized in that The multimodal database also includes actual data on ribosome occupancy of the mRNA sequence samples; Before constructing the multimodal database, the method further includes: Obtain mRNA sequence samples; Determine the secondary structure sample of the mRNA sequence sample in the current cell type, detect the structural strength of the secondary structure sample, and obtain the RNA structure probe score of the secondary structure sample; Determine the actual ribosome occupancy data of the mRNA sequence sample in the current cell type; The constructing of a multimodal database comprises: A multimodal database is constructed based on the mRNA sequence samples, secondary structure samples and actual ribosome occupancy data.

8. The method according to claim 7, characterized in that The method of determining the actual ribosome occupancy data of the mRNA sequence sample in the current cell type includes: Determine the mRNA fragments in the translation state on the mRNA sequence sample to obtain original Ribo-Seq data; Extracting the original Ribo-Seq data to obtain sequencing fragments that meet preset quality conditions; Matching the sequenced fragment with a predetermined reference genome to obtain the initial position of the sequenced fragment; Correcting the initial position of the sequencing fragment; An exon reference genome is constructed, and the corrected sequencing fragments are matched with the exon reference genome to obtain ribosome position information with single-base accuracy, and actual ribosome occupancy data is determined based on the ribosome position information with single-base accuracy.

9. The method according to claim 7, characterized in that: Before inputting the mRNA sequence samples in the multimodal database and the secondary structure samples corresponding to the mRNA sequence samples into the prediction model to be trained, the method further includes: Aligning the RNA structure probe scores of the secondary structure samples and the actual ribosome occupancy data with a predetermined reference genome at the transcript level to obtain a number of transcripts carrying sequence information, structure information, and ribosome occupancy information; Taking the transcript as a unit, denoising the transcript based on the actual data of the ribosome occupancy rate and the length of the transcript to obtain a training sample, and constructing a training data set based on the training sample; Extracting sequence information and structural information corresponding to each training sample in the training data set, determining the mRNA sequence corresponding to the extracted sequence information as the currently input mRNA sequence sample, and determining the secondary structure sample corresponding to the extracted structural information as the currently input secondary structure sample.

10. The method according to claim 9, characterized in that The step of taking the transcript as a unit, performing denoising processing on the transcript based on the actual data of the ribosome occupancy rate and the length of the transcript to obtain a training sample, and constructing a training data set based on the training sample includes: Detect the length of each transcript and remove transcripts shorter than a preset length; Extracting the first several bases and the last several bases of each transcript remaining after removing the transcript of the preset length, and splicing the first several bases and the last several bases of each transcript to form a prediction sample; Determining the actual data of the ribosome occupancy of each predicted sample, and removing the predicted samples whose actual data of the ribosome occupancy is zero; After removing the sequencing samples whose actual ribosome occupancy data is zero, the remaining predicted samples are sorted according to the size of the actual ribosome occupancy data, and a number of predicted samples with higher rankings in the actual ribosome occupancy data are extracted to form positive samples, and a number of predicted samples with lower rankings in the actual ribosome occupancy data are extracted to form negative samples; The positive samples and negative samples are integrated to obtain a training data set.

Images