General protein-RNA (Ribonucleic Acid) binding prediction method based on multi-molecular modal fusion

Abstract

The invention relates to a general protein-RNA (Ribonucleic Acid) combination prediction method based on multi-molecular modal fusion, which comprises the following steps: carrying out initial characterization on preprocessed protein and RNA sequences through a generalized biological basic model, and then training a constructed neural network model based on an improved cross attention mechanism; in the online stage, the trained model is adopted to output a predicted value of the combination possibility of the two molecular sequences according to the arbitrarily input protein sequence and RNA molecular sequence. According to the method, a general biological basic model is used for carrying out initial characterization on two molecules of protein and RNA, effective information extraction and fusion are carried out on characterization of the two molecules through an improved cross attention mechanism, and an artificial neural network trained based on the effective fusion information can more accurately predict a protein and RNA binding result; and the method has stronger generalization ability, and can cope with combination prediction scenes and tasks of various proteins.

Description

Technical Field

[0001] The present invention relates to a technology in the field of genetic engineering, specifically a universal protein-RNA binding prediction method based on multi-molecular modality fusion. Background Art

[0002] Predicting the binding of proteins to RNA is of great significance for exploring RNA functions and treating diseases. Most current prediction methods are based on traditional machine learning methods, which are based on single protein data and cannot effectively predict the binding of unknown proteins to RNA, and lack generalization capabilities. Summary of the invention

[0003] In view of the deficiency that the prior art cannot predict the binding results of any given protein and RNA, the present invention proposes a universal protein-RNA binding prediction method based on multi-molecular modality fusion, which uses a universal biological basic model to perform initial characterization of both protein and RNA molecules, and extracts and fuses effective information from the characterizations of the two molecules through an improved cross-attention mechanism. The artificial neural network trained based on the effective fusion information can more accurately predict the binding results of proteins and RNA, and has a stronger generalization ability, which can cope with the binding prediction scenarios and tasks of multiple types of proteins.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a universal protein-RNA binding prediction method based on multi-molecular modality fusion. After pre-processed protein and RNA sequences are initially characterized by a generalized biological basis model, a neural network model based on an improved cross-attention mechanism is constructed for training. In the online stage, the trained model is used to output a predicted value of the binding possibility of the two molecular sequences according to any input protein sequence and RNA molecular sequence.

[0006] The preprocessing refers to: constructing the binding signal spectrum of various proteins on the full-length RNA, traversing the full-length signal spectrum using a sliding window method, obtaining positive samples of RNA sequences bound by each protein and setting the label to 1, as well as the same number of negative samples that are not bound and setting the label to 0 according to a pre-set window signal value division standard, and finally dividing the samples into different training sets, validation sets and test sets according to chromosomes.

[0007] The generalized biological basis model is adopted but not limited to the general biological basis model LucaOne (https: / / github.com / LucaOne), a pre-trained basic model for unified learning of nucleic acid and protein languages, performs initial characterization of pre-processed protein molecule sequences and RNA molecule sequences respectively.

[0008] The neural network model based on the improved cross-attention mechanism includes: a multi-molecule representation fusion module and a multi-layer perceptron prediction module, wherein: the multi-molecule representation fusion module is based on the initial representation of the protein and RNA molecular sequences after pre-representation, and through the improved cross-attention mechanism based on the two molecular features of protein and RNA, that is, each attention head calculates the attention coefficient between the two modal features according to the following formula Where Q r The feature matrix representing the initial characterization of RNA, The feature matrix representing the initial representation of the protein is transposed, Used for scaling the size of attention. The model dynamically distributes information weights between the representations of the two molecular modalities of RNA and protein, captures the interaction rules between RNA and protein sequences, and provides fusion features for fine-grained information fusion of the two molecular features, thereby helping the model enhance its understanding of protein-RNA binding tasks; the multi-layer perceptron prediction module uses the multi-modal information fused in the early stage based on the fused multi-modal representation, deeply mines the nonlinear features in the data through its own multi-layer structure, and outputs the predicted value of protein-RNA binding.

[0009] The multi-layer perceptron prediction module includes: an average pooling layer, a linear layer and a GELU activation function layer. Technical Effects

[0010] The present invention uses a cross-attention module based on the two molecular features of protein and RNA, which can capture important molecular information that affects the binding process of various proteins and RNA, and provide more sufficient prior knowledge for prediction tasks; the neural network model constructed based on this can accurately predict the binding results of various unknown proteins and RNA. Compared with the prior art, the present invention can effectively complete the task of predicting the binding of various unknown proteins and RNA, and give more accurate and reliable prediction results, eliminating the limitations of existing methods based on protein specificity, and has strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flow chart of the present invention;

[0012] Figure 2 Schematic diagram of the neural network model of the present invention;

[0013] Figure 3 This is a comparison chart of the data chart of the present invention. DETAILED DESCRIPTION

[0014] like Figure 1 As shown, this embodiment relates to a universal protein-RNA binding prediction method based on multi-molecular modality fusion, which specifically includes:

[0015] Step 1) Data collection and preprocessing, including:

[0016] Step 1.1) Obtain the eCLIP-seq alignment result file and the identified binding site file of 168 human RNA binding proteins from the ENCODE project website.

[0017] Step 1.2) Based on the above files, the binding signal spectrum of each protein on the full-length RNA is constructed.

[0018] Step 1.3) Use a fixed-length sliding window to traverse the full-length RNA binding signal spectrum of each protein, obtain a fixed-length fragment at each signal peak, and construct an RNA positive sample set with each protein label as 1.

[0019] Step 1.4) According to the number of positive samples of each protein, the same sliding window is used to obtain the same number of RNA fragments at the signal-free peaks on the respective full-length RNA binding signal spectra to construct an RNA negative sample set with each protein label as 0.

[0020] Step 2) After data preprocessing, all protein sequences and RNA positive and negative samples are pre-characterized, and a general biological basic model is used to give initial characterizations of important biological information contained in protein sequences and RNA sequences respectively. The initial characterizations of each protein and all RNA samples are saved to facilitate direct calls in subsequent model training processes, greatly speeding up model training.

[0021] Step 3) After data preparation is completed, the neural network model is constructed using the improved cross-attention mechanism, such as Figure 2 As shown, specifically including:

[0022] Step 3.1) Input the initial representation of protein and RNA sequences into the improved multi-head cross attention module, calculate each attention coefficient and the weighted sum of the protein feature matrix, merge and add to the RNA initial representation, and obtain a multimodal representation that integrates the two molecular information of protein and RNA.

[0023] Step 3.2) The fused multimodal representation is input into a multi-layer perceptron prediction module consisting of an average pooling layer, a linear layer, and a GELU activation function to predict the input protein-RNA binding prediction value.

[0024] Step 4) Perform model training on the preprocessed data. Considering the generalization of the prediction ability of the verification model, we divide all the data into the following parts: randomly select 10 proteins and their corresponding RNA samples as the test set, and the remaining proteins and their corresponding RNA samples as the training set to verify the model's ability to predict the binding of unknown proteins to RNA. Separate all samples corresponding to chromosome 1 in the training set to verify the model's ability to predict the binding of RNA samples corresponding to unknown chromosome categories to proteins. The AUROC of each protein test set is used to measure the generalization performance and prediction accuracy of the model.

[0025] Step 5) After the model parameters are adjusted and the training converges, the model is used to predict all samples of each protein in the test set.

[0026] After specific practical experiments, under the specific environment settings of Python3.9 in Linux environment, using 134 proteins and a total of 807362 RNA sample data, running the above method to train the model on an A100 graphics card, it is possible to obtain a universal RNA binding prediction model for various proteins. Prediction is performed on the chromosome 1 test set and other chromosome test sets of another 10 unknown proteins. The experimental data obtained show that the average AUROC of the chromosome 1 test set prediction of various proteins is above 0.77, and the average AUROC of the other chromosome test sets is above 0.8, indicating that this method can accurately predict the binding results of various unknown proteins and RNA, and has strong generalization ability.

[0027] The prediction results are shown in Table 1. In the prediction of the binding of many unknown proteins and RNAs, the AUROC of the model for various chromosomes is above 0.8, achieving good generalization ability and prediction accuracy.

[0028] Table 1. Model performance on different protein test sets

[0029] Compared with the existing technology, this method uses a cross-attention module based on the improved molecular features of protein and RNA, effectively capturing important molecular information that affects the binding process. It not only has strong versatility, but also significantly improves the accuracy and precision of various unknown protein-RNA binding predictions. As far as we know, this is the first general method that can cope with various protein-RNA binding prediction tasks.

[0030] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A universal protein-RNA binding prediction method based on multi-molecular modality fusion, characterized in that: After the preprocessed protein and RNA sequences are initially characterized by the generalized biological basis model, they are used to train the constructed neural network model based on the improved cross-attention mechanism. In the online stage, the trained model is used to output the predicted value of the binding possibility of the two molecular sequences based on any input protein sequence and RNA molecular sequence.

2. The universal protein-RNA binding prediction method based on multi-molecular modality fusion according to claim 1, characterized in that: The preprocessing refers to: constructing the binding signal spectrum of various proteins on the full-length RNA, traversing the full-length signal spectrum using a sliding window method, obtaining positive samples of RNA sequences bound by each protein and setting the label to 1, as well as the same number of negative samples that are not bound and setting the label to 0 according to a pre-set window signal value division standard, and finally dividing the samples into different training sets, validation sets and test sets according to chromosomes.

3. The universal protein-RNA binding prediction method based on multi-molecular modality fusion according to claim 1, characterized in that: The neural network model based on the improved cross-attention mechanism includes: a multi-molecule representation fusion module and a multi-layer perceptron prediction module, wherein: the multi-molecule representation fusion module is based on the initial representation of the protein and RNA molecular sequences after pre-representation, and through the improved cross-attention mechanism based on the two molecular features of protein and RNA, that is, each attention head calculates the attention coefficient between the two modal features according to the following formula Where Q r The feature matrix representing the initial characterization of RNA, The feature matrix representing the initial representation of the protein is transposed, The attention size is scaled so that the model can dynamically distribute information weights between the representations of the two molecular modalities of RNA and protein, capture the interaction rules between RNA and protein sequences, and provide fusion features for fine-grained information fusion of the two molecular features, thereby helping the model to enhance its understanding of protein-RNA binding tasks; the multi-layer perceptron prediction module uses the multi-modal information fused in the early stage based on the fused multi-modal representation, deeply mines the nonlinear features in the data through its own multi-layer structure, and outputs the predicted value of protein-RNA binding; The multi-layer perceptron prediction module includes: an average pooling layer, a linear layer and a GELU activation function layer.

4. The universal protein-RNA binding prediction method based on multi-molecular modality fusion according to any one of claims 1 to 3, characterized in that: The protein and RNA sequences are obtained by the following method: Step 1.1) Obtain the eCLIP-seq alignment result file and the identified binding site file of 168 human RNA binding proteins from the ENCODE project website; Step 1.2) constructing the binding signal spectrum of each protein on the full-length RNA based on the above files; Step 1.3) Use a fixed-length sliding window to traverse the full-length RNA binding signal spectrum of each protein, obtain a fixed-length fragment at each signal peak, and construct an RNA positive sample set with each protein label as 1; Step 1.4) According to the number of positive samples of each protein, the same sliding window is used to obtain the same number of RNA fragments at the signal-free peaks on the respective full-length RNA binding signal spectra to construct an RNA negative sample set with each protein label as 0.

5. The universal protein-RNA binding prediction method based on multi-molecular modality fusion according to any one of claims 1 to 3, characterized in that: The initial characterization refers to: pre-characterizing all protein sequences and RNA positive and negative samples after data preprocessing, using a general biological basic model to give initial characterizations of important biological information contained in protein sequences and RNA sequences respectively, and saving the initial characterizations of each protein and all RNA samples, so as to facilitate direct calling in the subsequent model training process, greatly speeding up the model training speed.

6. The universal protein-RNA binding prediction method based on multi-molecular modality fusion according to any one of claims 1 to 3, characterized in that: The neural network model based on the improved cross-attention mechanism is obtained in the following way: Step 3.1) Input the initial representation of protein and RNA sequences into the improved multi-head cross attention module, calculate each attention coefficient and the weighted sum of the protein feature matrix, merge and add to the RNA initial representation, and obtain a multimodal representation that integrates the information of both protein and RNA molecules; Step 3.2) The fused multimodal representation is input into a multi-layer perceptron prediction module consisting of an average pooling layer, a linear layer, and a GELU activation function to predict the input protein-RNA binding prediction value.

7. The universal protein-RNA binding prediction method based on multi-molecular modality fusion according to claim 6, characterized in that: The training is performed on preprocessed data: Taking into account the generalization of the prediction ability of the verification model, we divide all the data into: 10 proteins and their corresponding RNA samples are randomly selected as the test set, and the remaining other proteins and their corresponding RNA samples are used as the training set to verify the model's ability to predict the binding of unknown proteins to RNA. All samples corresponding to chromosome 1 in the training set are divided out to verify the model's ability to predict the binding of RNA samples corresponding to unknown chromosome categories to proteins. The AUROC of each protein test set is used to measure the generalization performance and prediction accuracy of the model.

Images