Protein-RNA interaction prediction method based on multi-modal deep learning

Abstract

The invention relates to the technical field of multi-modal RNA feature fusion and high-confidence data set construction, and particularly discloses a protein-RNA interaction prediction method based on multi-modal deep learning, which models and predicts an RBP binding mode, classifies cell line specific RBP-RNA interaction by applying a computing technology, and provides a method for predicting protein-RNA interaction based on multi-modal deep learning. An interpretable group of binding specificity between different cell lines is determined; establishing a cell line specific RNA structure similarity matrix by adopting a structure comparison and clustering technology; predicting the exact nucleotide binding positions of these sites as markers RBP; mRNA (messenger ribonucleic acid) and lincRNA are used as controls; the positive correlation between the RBP-RNA interaction and RNA translation, especially the positive correlation between the RBP-RNA interaction and the RNA translation, is established, a theory and a strategy are provided for cancer targeted therapy, the solution of data heterogeneity and the calculation complexity are reduced, and the prediction efficiency of the protein-RNA interaction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal RNA feature fusion and high-confidence dataset construction, and in particular to a protein-RNA interaction prediction method based on multimodal deep learning. Background Art

[0002] Protein-RNA interactions (PRIs) are central to life, participating in key physiological functions such as gene expression regulation, post-transcriptional modification, and protein synthesis. Their dysfunction is closely associated with a variety of diseases, such as cancer and neurodegenerative diseases. Early research on PRIs progressed slowly due to limitations in experimental techniques. In recent years, breakthroughs in high-throughput experimental techniques, such as RNA immunoprecipitation (RIP)-seq and crosslinking and immunoprecipitation (CLIP)-seq, have driven data accumulation. Combined with advances in computational methods, PRIs research has entered a period of rapid development.

[0003] Early methods relied on hand-crafted features (such as sequence encoding and physicochemical properties) and shallow learning models such as support vector machines (SVM) and logistic regression (LR). The k-mer model and position weight matrix (PWM) reveal binding patterns by counting nucleotide frequencies. Limited by the expressive power of hand-crafted features, these models can capture nucleotide preferences but are unable to handle long-range dependencies. In recent years, deep learning technology has overcome the limitations of traditional methods, extracting features directly from raw sequence or structural data through end-to-end learning. For example, DeepBind uses a convolutional neural network (CNN) to predict the sequence preferences of RNA-binding proteins (RBPs); DeepCLIP uses only sequence information, using a bidirectional long short-term memory (BiLSTM) network to capture long-range dependencies and improve performance; and RPI-Net integrates RNA secondary structure information through graph neural networks (GNNs), significantly improving prediction accuracy.

[0004] Three-dimensional structural features are crucial for PRIs prediction. Relying solely on traditional structural encoding methods, such as secondary structure classification, can limit performance due to information loss. In recent years, the introduction of point clouds, voxel grids, and GNNs has brought new opportunities for structural modeling. For example, 3DCNNs, which utilize voxel encoding to preserve geometric details through three-dimensional grids, excel in predicting protein functional domains. GraphBind, which utilizes GNNs to simulate RNA base interactions through node features and edge connections, has improved structural elucidation capabilities. With the advent of AlphaFold, AI-based prediction has achieved prediction accuracy and resolution comparable to cryo-electron microscopy for certain proteins. This means that RBP structures that have not yet been successfully solved experimentally can now be accurately predicted by AlphaFold. This large amount of accurate structural data makes it possible to construct a closed-loop "sequence → structure → interaction" system, achieving accurate predictions independent of experimental data. Furthermore, static PDB structures cannot reflect the conformational changes that occur during protein-RNA binding. Molecular dynamics simulation data must be combined with the development of temporal models (such as spatiotemporal convolutional networks) to capture dynamic processes. Traditional methods are mostly based on single-sided molecular modeling (only protein or RNA), ignoring the cooperative nature of interactions. Multimodal fusion models and the development of unified embedding frameworks are becoming increasingly popular, but these approaches require addressing data heterogeneity and computational complexity. Therefore, a protein-RNA interaction prediction method based on multimodal deep learning is needed. Summary of the Invention

[0005] In response to the problems existing in the prior art, the purpose of the present invention is to provide a protein-RNA interaction prediction method based on multimodal deep learning.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a protein-RNA interaction prediction method based on multimodal deep learning, comprising the following steps:

[0007] S1, NeuroPRIS integrates multi-source data and combines traditional RNA sequence data with advanced tertiary structure information from deep learning to build models and predict RBP binding patterns;

[0008] S2. We applied computational techniques to classify cell line-specific RBP-RNA interactions and identified interpretable groups of binding specificities across cell lines. We used structural alignment and clustering techniques to create a cell line-specific RNA structural similarity matrix, demonstrating the structural dependence of RBP-RNA binding.

[0009] S3, NeuroPRIS visualizes binding sites and predicts these sites as the exact nucleotide binding positions of the tagged RBP;

[0010] S4. NeuroPRIS was used to elucidate the relationship between these lincRNAs and RBPs. Using mRNA and untranslatable small peptide lincRNA as controls, a positive correlation between RBP-RNA interactions and RNA translation was established, especially between RBP-RNA interactions and lincRNAs.

[0011] Specifically, the NeuroPRIS in step S1 integrates multi-source data to download complete sequences of lincRNA and mRNA from the ENCODE and ENCORI databases, removes RNA longer than 7500nt, uses RNAfold and 3dRNA to obtain the secondary and tertiary structures of the complete RNA, selects binding sites with a length greater than 10 and less than 150 nucleotides from the CLIP experiment, extracts the primary sequence, secondary structure and tertiary structure of the RNA fragments at these sites from the complete RNA, and fixes the RNA fragments to 150nt in length, marks positive samples as "1" and negative samples as "0", and constructs a NeuroPRIS model for each RBP dataset for each cell line. By default, 80% of the input data is used for training, 10% for validation, and 10% for testing. NeuroPRIS trains the model for each RBP with CLIP experimental data through multiple iterations of the training set and validation set, evaluates the score of each nucleotide position in the binding site, and finally outputs the score of the entire binding site.

[0012] Specifically, the RBP training model input includes sequence input, secondary structure input and 3D coordinate input. The 3D coordinate input is feature extracted through the T-net architecture. The T-net consists of multiple convolutional layers and a global maximum pooling layer, which is specifically implemented by three convolutional layers. The three convolutional layers use 32, 64 and 128 filters respectively, and the kernel size is 1. It also includes an output layer that obtains feature transformation by linear mapping of input features, so that the model can autonomously learn feature representations that are more suitable for downstream tasks. The features processed by T-net are combined with the embedding of sequence and structural features, and the sequence is embedded through an embedding with an input dimension of 5 and an output dimension of 127. The input layer is converted into a high-dimensional representation, and the secondary structure is converted into a high-dimensional representation through an embedding layer with an input dimension of 3 and an output dimension of 128. The embedded features are then combined with the output of T-net through dot operations to create a multimodal input. These fused features are input into a bidirectional LSTM network layer containing 48 LSTM units to capture the temporal dependencies in the input sequence. The return sequence mode of the BiLSTM layer facilitates the subsequent output layer to process sequence information. Finally, the BiLSTM output is classified through a fully connected layer with two output classes, and the softmax activation function is applied to generate the final class probability. Batch Normalization is used for all convolutional and fully connected layers in the model to enhance training stability. A dropout layer with a dropout rate of 0.1 is introduced after each activation layer to reduce the risk of overfitting. The model uses the Adam optimizer, a batch size of 64, and an initial learning rate of 0.001.

[0013] Specifically, the computational technology in step S2 uses NeuroPRIS to reveal unique RBP-RNA binding patterns in different cell lines, and finds that RBPs in the IGF2BPs family have high specificity when binding to RNA, and these highly specific binding RBPs play a key role in specific cell lines.

[0014] Specifically, the cell line-specific RNA structural similarity matrix in step S2 adopts RNA clustering and analysis, uses the internal qTMclust algorithm of US-align to perform multiple structural alignments, and adopts two cutoff values ​​to cluster the RNA binding site structure; through structural cluster analysis, the unique core structure of the RNA binding site and the relationship between different cell line specificities and the RNA binding site structure are revealed, and the cell line-specific RNA structural similarity matrix is ​​constructed to reveal the "sequence-independent, structure-driven" interaction rules.

[0015] Specifically, the NeuroPRIS visualization binding sites in step S3 are spliced ​​along the sequence dimension by three-dimensional point cloud features and one-dimensional sequence information, and the context dependency is captured using bidirectional LSTM. The sum of the hidden states of the bidirectional LSTM time steps is extracted to generate an interpretable RNA-binding protein motif, which clarifies that the RNA-RBP interaction is completed by a small number of "core binding bases".

[0016] Specifically, the lincRNA in step S4 is a long intergenic non-coding RNA. The lincRNA was downloaded from LncPepAtlas to encode translatable peptides, and the lincRNAs marked as verified by MC-related literature were screened out. From all the lincRNAs in the GENCODE database, the lincRNAs that were verified as translatable peptides by multi-dimensional verification were eliminated. The multi-dimensional verification included MC-related literature verification, m6A, Ribo, IRES, CPAT, CPC2, Pfam, and CNCI. The lincRNAs that were not translatable peptides were retained. mRNAs were randomly selected from a total of 111,868 mRNAs in the GENCODE database. RNAs with a length of more than 7,500 nt were eliminated from the three data sets, and a small amount of RNAs that were not recognized by RNAfold and 3dRNA ultimately obtained 787 lincRNAs of potential translatable small peptides, 1784 lincRNAs of untranslatable small peptides, and 1763 mRNAs. NeuroPRIS was used to predict the binding of each group of RNAs to RNA-binding protein RBP. In the three data sets, the binding trends of lincRNAs and mRNAs of potential translatable small peptides to RBP were similar, while the binding probability of lincRNAs of untranslatable small peptides to RBP was significantly different from the former two. This difference is due to the fact that lincRNAs and mRNAs of potential translatable small peptides have specific functional binding sites for RBP, thereby promoting translation, clarifying the positive correlation between lincRNA-RBP binding and lincRNA translation.

[0017] The present invention has the following beneficial effects:

[0018] This multimodal deep learning-based protein-RNA interaction prediction method, designed in this study, uses NeuroPRIS to integrate multi-source data. It fuses traditional RNA sequence data with potential tertiary structure information from RBP CLIP-seq data from 23 cell lines and 202 in vivo binding events. It operates through a multi-component framework, extracting binding site information from multiple sources and dimensions. RNA tertiary structure data is processed into a point cloud dataset, and tertiary structural features are extracted using a double-transform alignment network (containing two sets of T-net modules), improving the network's robustness to different viewing angles and permutations of the point cloud. The study also used NeuroPRIS to reveal unique RBP-RNA binding patterns in different cell lines. Structural clustering analysis revealed that clustering methods based on three-dimensional RNA structure are more effective than traditional one-dimensional sequence clustering, revealing the unique core structure of RNA binding sites and the relationship between different cell line specificities and RNA binding site structure. A cell line-specific RNA structural similarity matrix was constructed, revealing "sequence-independent, structure-driven" interaction patterns. Furthermore, NeuroPRIS can infer biologically significant motifs, and the constructed binding profiles visually demonstrate the differences in binding strength between different bases. The study also provides interpretable RNA-binding protein motifs, demonstrating that RNA-RBP interactions are accomplished by a small number of "core binding bases." Finally, by constructing a dataset, the study clarified the relationship between RNA-RBP interactions and RNA translation, finding that translatable long noncoding RNAs (lincRNAs) and mRNAs have similar binding trends with RBPs. Furthermore, NeuroPRIS can effectively capture the dynamic changes in RBP-RNA interactions in different cell lines, identifying potential targets for cancer research. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the network architecture of the NeuroPRIS algorithm.

[0020] Figure 2 Schematic diagram of the main framework for processing tertiary structures.

[0021] Figure 3 Schematic diagram of inputting the fusion features of RNA structure and sequence into the BiLSTM layer.

[0022] Figure 4 The Circos heatmap shows the AUC score comparison results of NeuroPRIS with Graphprot, PRIESSTESS and Deepclip methods on all 202 RNA-RBP datasets.

[0023] Figure 5The violin plot shows the overall AUC score comparison results of the NeuroPRIS model and other methods in all 202 different cell lines.

[0024] Figure 6 The figure shows the ROC curve comparison results of NeuroPRIS and other methods on the RBP dataset.

[0025] Figure 7 This is the t-SNE clustering result diagram of the output features of NeuroPRIS in predicting RNA-RBP binding.

[0026] Figure 8 Performance comparison of different deep network architectures for NeuroPRIS ablation studies.

[0027] Figure 9 The classification accuracy curves of different deep learning frameworks of NeuroPRIS from 0 to 100 training epochs.

[0028] Figure 10 A comparison of the AUC scores of 35 NeuroPRIS models in their original cell lines and other cell lines is shown. Based on the degree of difference in AUC scores between the original cell lines and other cell lines, the specificity levels are divided into three groups: high, medium, and low. The specificity levels are sorted from left to right based on the magnitude of the difference.

[0029] Figure 11 Upset diagram for specific grouping in different cell lines.

[0030] Figure 12 The Venn diagram shows the distribution of different RBPs in the high specificity group, medium specificity group and low specificity group.

[0031] Figure 13 Figure 2 is a representative RNA profile of a representative RBP: IGF2BP3, specifically bound in three cell lines.

[0032] Figure 14 The figure is a stacked bar chart showing the structural similarity percentage of the binding sites of 35 RBPs on RNA.

[0033] Figure 15 This is the structural clustering diagram of IGF2BP3 in HEK293, K562 and HepG2 cell lines.

[0034] Figure 16 The three-dimensional structure of a representative RNA binding site of IGF2BP3.

[0035] Figure 17 Figure 4 is a Circos diagram of motifs extracted from all 202 RBP datasets.

[0036] Figure 18 Comparison chart of motifs extracted by NeuroPRIS and those in the literature.

[0037] Figure 19 This is the binding map of SRSF1 on RNA sequences.

[0038] Figure 20 Binding probability diagram of mRNA, translatable lincRNA and untranslatable lincRNA with RBP under different cell line conditions.

[0039] Figure 21 Figure 2 shows the binding probability of mRNA, translatable lincRNA, and untranslatable lincRNA to YTHDF1 and YTHDF2. DETAILED DESCRIPTION

[0040] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely further describe the technical solutions in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0041] like Figures 1-21 As shown, Figures 1-3 This is the network architecture of the NeuroPRIS algorithm. Figure 1 Comprehensive modeling and prediction of RBP binding for NeuroPRIS using RNA sequence and structure information and RBP binding sites in matching cell types. Figure 2 It is the main framework for processing three-level structures. The joint use of input transformation and feature transformation ensures the model's ability to handle local and global spatial transformations. Figure 3 The fused features of RNA structure and sequence are input into the BiLSTM layer to model the relationships between individual bases in RNA molecules, the contextual information is used to identify RNA regions related to RBP binding, and interpretable RNA-binding protein motifs are generated by summing the hidden states of time steps.

[0042] Figures 4-7 NeuroPRIS showed higher accuracy than other methods in predicting RBP binding sites on RNA. Figure 4 AUC scores of NeuroPRIS and other methods (including Graphprot, PRIESSTESS, and Deepclip) across all 202 RNA-RBP datasets are shown for the Circos heatmap. Figure 5 Violin plots show the overall AUC scores of the NeuroPRIS model and other methods across all 202 different cell lines. ****P < 0.001 (one-sided paired t-test).Figure 6 ROC curves for RBP dataset using HDRNet and other methods. Figure 7 t-SNE clustering results of NeuroPRIS output features.

[0043] Figures 8-9 For the ablation study of NeuroPRIS. Figure 8 Performance comparison of various deep network architectures for NeuroPRIS (center line, median; box limits, upper and lower quartiles). Figure 9 Class accuracy curves for different deep learning frameworks from 0 to 100 training epochs. The thick line represents the average value, and the thin lines represent the range of maximum and minimum values.

[0044] Figures 10-13 The RBP-RNA binding events showed differences in different cell lines. Figure 10 Comparison of the AUC scores of 35 NeuroPRIS models between the original cell line and other cell lines. Based on the difference in AUC scores between the original cell line and other cell lines, the specificity levels are divided into three groups: high, medium, and low. The scores are sorted from left to right in order of difference. Figure 11 Inversion plots of specific groups within different cell lines. Figure 12 Inversion plots of high, medium, and low specificity groups in different cell lines. Red dots represent bound RNAs, blue bars indicate the number of specifically bound RNAs, and red lines connecting dots represent RNAs bound by the same RBP. Venn diagram showing the prevalence of different RBPs in the high, medium, and low specificity groups. Figure 13 Representative RNAs specifically bound by IGF2BP3 in three cell lines are shown.

[0045] Figure 14 This is a stacked bar chart showing the percentage of structural similarity among the RNA binding sites of 35 RBPs. Different colors represent the degree of clustering: 1 indicates the number of groups with no identical RNA binding sites, 2 indicates the number of groups with two identical RNA binding sites, and so on. 11 indicates the number of groups with 11-20 identical RNA binding sites, and 12 indicates the number of groups with 21-78 identical RNA binding sites. The categories at the bottom represent different levels of cell line specificity. Figure 15 Figure 2. Structural clustering of IGF2BP3 in HEK293, K562, and HepG2 cell lines. The horizontal and vertical axes represent representative RNA binding sites of IGF2BP3, and the intersection points on the diagonal indicate that multiple RNA binding sites have the same tertiary structure. Figure 16 The three-dimensional structure of a representative IGF2BP3 RNA binding site. The ID of the RNA containing this structure is shown on the right.

[0046] Figures 17-19 For the interpretability study of NeuroPRIS. Figure 17 Figure 4 is a Circos diagram of motifs extracted from all 202 RBP datasets. Figure 18 The motifs extracted for NeuroPRIS were compared with binding specificities described in the literature. Figure 19 The SRSF1 binding map on RNA sequences. The horizontal axis represents the RNA sequence, the vertical axis represents the score calculated by NeuroPRIS, the red dots represent the binding sites from the CLIP-seq experimental results, and the blue dots represent the binding sites calculated by NeuroPRIS (highest score range).

[0047] Figures 20-21 The relationship between RBP interaction and RNA translation. Figure 20 The binding probabilities of mRNA, translatable lincRNA and untranslatable lincRNA with RBP under different cell line conditions. Figure 21 is the binding probability of mRNA, translatable lincRNA and untranslatable lincRNA to YTHDF1 and YTHDF2.

[0048] This study introduces NeuroPRIS method to study protein-RNA interaction from single sequence analysis to multi-modal structure modeling, the cross of deep learning and structural biology, and promotes the breakthrough of the field. Focus on structural dynamics, multi-modal fusion and interpretability to realize the transformation from basic research to precision medicine. NeuroPRIS integrates multi-source data, from 23 cell lines RBP CLIP-seq data and 202 in vivo binding events, combines traditional RNA sequence data and potential tertiary structure information. It works through a multi-component framework: extract binding site information from multi-source and multi-dimensional; process RNA tertiary structure data into point cloud dataset, use double transformation alignment network (contains two T-net modules) to extract tertiary structure features, improve the network's robustness to different perspectives and arrangement of point cloud; the study also uses NeuroPRIS to reveal the unique RBP-RNA binding mode in different cell lines, and finds that RBPs such as IGF2BPs family have high specificity in binding RNA, and these high-specificity binding RBPs play a key role in specific cell lines. Through structural clustering analysis, it is found that the clustering method based on RNA three-dimensional structure is more effective than the traditional one-dimensional sequence clustering, which can reveal the unique core structure of RNA binding site and the relationship between different cell line specificity and RNA binding site structure. Constructing cell line-specific RNA structure similarity matrix reveals the interaction rule of "sequence-independent, structure-driven". In addition, NeuroPRIS can also infer biologically meaningful motifs, and the constructed binding spectrum can intuitively display the binding strength difference of different bases. And provide interpretable RNA binding protein motifs, and clarify that the interaction between RNA-RBP is completed by a few "core binding bases". Finally, the study constructs the dataset to clarify the relationship between RNA-RBP interaction and RNA translation, and finds that the long non-coding RNA (lincRNAs) and mRNA have similar binding trends with RBPs, and NeuroPRIS can effectively capture the dynamic changes of RBP and RNA interaction in different cell lines, and identify potential targets for cancer research.

[0049] 1. Data processing: Construction of a general model dataset. Complete sequences of lincRNA and mRNA were downloaded from the ENCODE and ENCORI databases. RNAs longer than 7500 nt were removed. The secondary and tertiary structures of the complete RNA were obtained using RNAfold and 3dRNA. Binding sites with a length greater than 10 and less than 150 nucleotides were selected from the CLIP experiment. The primary sequence, secondary structure, and tertiary structure of the RNA fragments at these sites were extracted from the complete RNA, and the RNA fragments were fixed to 150 nt in length. Positive samples were marked as "1" and negative samples were marked as "0". For each RBP dataset, a NeuroPRIS model was constructed for each cell line. By default, 80% of the input data was used for training, 10% for validation, and 10% for testing. NeuroPRIS trained the model for each RBP with CLIP experimental data through multiple iterations of the training and validation sets, evaluated the score of each nucleotide position within the binding site, and finally output the score of the entire binding site.

[0050] 2. Model construction includes the following steps:

[0051] 2.1. The T-net network forms the object's skeleton. 1426 carbon atoms (the average number of carbon atoms in a 150-nucleotide RNA molecule) are randomly selected from a pool of tertiary structural coordinates. The spatial positional information of these carbon atoms is generated into a point cloud dataset, represented as a set of three-dimensional points {Ci|i=1,...,n}. The T-net network processes the three-dimensional point cloud data, effectively capturing the spatial structural information of the molecule and the interactions between points. Global features are normalized to the range 0-1. The global features are composed of K local features. The interval [0,1] is evenly divided into K parts, each representing a local feature space. The values ​​within each local feature are also normalized to 0-1. After multi-layer convolutional dimensionality expansion, each point corresponds to K values ​​of different dimensions, corresponding to K local feature spaces. The model learns to capture local structure from nearby points and the combined interactions between these local structures, effectively extracting the tertiary structural features of the RNA molecule. Experiments have shown that key points form the object's skeleton.

[0052] 2.2 Encoding: Sequence data is processed into linearized one-hot vectors. In this one-hot representation, each term in the vocabulary v = (A, C, G, U) is represented by a vector with a 1 in a unique dimension. Secondary and tertiary structural information are represented as the fifth and sixth vectors, respectively. All input sequences are zero-padded to the same length, reaching the maximum length allowed by "full convolution" as defined in the Lasagne document. After vectorizing the sequence bases, the combined vector is linearized, and the resulting one-dimensional data is used as the input to the neural network.

[0053] 2.3. The fused features of RNA structure and sequence are input into a bidirectional long short-term memory (BiLSTM) layer to model the relationships between bases in the RNA molecule and identify RNA regions associated with RBP binding using contextual information. In the BiLSTM layer, the input sequence is presented to two independent LSTM layers: one processing the sequence forward and the other processing it backward, both connected to the same output layer. Formulas are used to calculate the forget gate, input gate, output gate values, and state values ​​at each time step, ultimately determining the hidden layer state. NeuroPRIS's BiLSTM layer is used to generate nucleotide-level binding profiles. This bidirectional network structure is capable of capturing hidden correlations between preceding and following bases.

[0054]

[0055] At time step t, the values ​​of the forget gate, input gate, and output gate are represented as f t 、i t and o t . a t Indicates that at time t h t-1 and preliminary feature extraction of x, where x t represents the input at time t, h t-1 Represents the hidden state value at time t-1. x in the forget gate, input gate, output gate and feature extraction process t The weight coefficients are expressed as w i 、w o and w a . Similar, u f 、u i 、u a and u o Represents h in these gates and feature extraction processes respectively t-1 The weight coefficient of the forget gate, input gate, output gate and feature extraction process are represented by b f 、b i 、b a and b o The function tanh represents the hyperbolic tangent function, and the sigmoid function is represented by σ (Formula (1)-Formula (6)). The calculation results of the forget gate and the input gate are applied to c t-1 , forming the state c at time t t , where ⊙ represents the Hadamard product (element-wise multiplication). Finally, the hidden layer state h t By output gate o t and the current state c tTogether, they determine (Formula (7)-Formula (8)). The NeuroPRIS bidirectional long short-term memory (BiLSTM) layer is used to generate nucleotide-level binding maps. In addition, this bidirectional network structure can capture hidden correlations between previous and next bases.

[0056] 2.4 Parameter Setup. The model inputs include sequence input, secondary structure input, and 3D coordinate input. The 3D coordinate input is subjected to feature extraction using the T-Net architecture. The T-Net consists of multiple convolutional layers and a global max pooling layer. Specifically, it is implemented using three convolutional layers, using 32, 64, and 128 filters, respectively, with a kernel size of 1. An output layer also transforms the input features by linearly mapping them, enabling the model to autonomously learn feature representations more suitable for downstream tasks. The T-Net processed features are combined with embeddings of sequence and structural features. The sequence is converted to a high-dimensional representation using an embedding layer with an input dimension of 5 and an output dimension of 127. The secondary structure is converted to a high-dimensional representation using an embedding layer with an input dimension of 3 and an output dimension of 128. The embedded features are then combined with the T-Net output through a dot operation to create a multimodal input. These fused features are fed into a bidirectional LSTM network layer consisting of 48 LSTM units to capture temporal dependencies in the input sequence. The return sequence mode of the LSTM layer facilitates processing of the sequence information by the subsequent output layer. Finally, the LSTM output is classified through a fully connected layer with two output classes, and a softmax activation function is applied to produce the final class probabilities. Batch Normalization is used in all convolutional and fully connected layers in the model to enhance training stability, and a dropout layer with a dropout rate of 0.1 is introduced after each activation layer to mitigate the risk of overfitting. The model uses the Adam optimizer with a batch size of 64 and an initial learning rate of 0.001. An early stopping strategy that monitors validation loss is used to prevent overfitting and save resources. The model was trained on an NVIDIA GeForce RTX 3090 GPU (24GB of memory) on 202 RBP datasets, with a total training time of 70 hours, ultimately resulting in a model that effectively classifies sequence and structural features.

[0057] 3. RNA clustering and analysis: Multiple structural alignments were performed using the internal qTMclust algorithm of US-align. Two cutoffs were used to cluster the RNA binding site structures: cutoff = 0.45 and cutoff = 0.3. Clusters with a TM-score of 0.45 or greater were classified as high-precision groups. The average TM-score RNA (approximately 0.273, approximately 0.3) was used as the second cutoff, and clusters with a TM-score greater than 0.3 were classified as average-precision groups.

[0058] 4. Motif generation: For a given input sequence, the input to the output layer is a vector consisting of the row sums of the matrix H, where the padding is masked to ensure that predictions only come from valid one-hot encoding embedding regions. This design enables the model to focus on meaningful segments in the sequence, thereby improving classification accuracy. The prediction of a specific RNA sequence is generated by aggregating the output values ​​of the bidirectional long short-term memory network (BiLSTM) layer: the total LSTM output at each time step is calculated to obtain a subarray of the final prediction score. A sliding window method is used to identify subarrays of length 6mer in a given score in order to determine the subsequence with the highest sum and its position, as shown in formula (9):

[0059]

[0060] The extracted BiLSTM hidden state can generate the final prediction score by summing the output of each time step, where y represents the output prediction score and T represents the total number of time steps.

[0061]

[0062] In formula (10), we use a sliding window technique to identify the subarray with the largest sum, where N represents the total number of prediction scores and n represents the length of the subarray (6mer). During the summation phase, the sliding window systematically traverses all possible subarrays and records the maximum value. Regarding the output of the BiLSTM, if the hidden state at a particular time step is primarily positive, then the base at that position is likely associated with the binding site identified by the CLIP experiment; conversely, if it is primarily negative, the base is more likely associated with random background sequences in the genome. If it is close to zero, the base at that position may belong to both categories.

[0063] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the guidance of the present invention, and any technical solutions that are the same or similar to those of the present invention, fall within the scope of protection of the present invention.

[0064] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A protein-RNA interaction prediction method based on multimodal deep learning, characterized in that: The following steps are involved: S1, NeuroPRIS integrates multi-source data and combines traditional RNA sequence data with advanced tertiary structure information from deep learning to build models and predict RBP binding patterns; S2. We applied computational techniques to classify cell line-specific RBP-RNA interactions and identified interpretable groups of binding specificities across cell lines. We used structural alignment and clustering techniques to create a cell line-specific RNA structural similarity matrix, demonstrating the structural dependence of RBP-RNA binding. S3, NeuroPRIS visualizes binding sites and predicts these sites as the exact nucleotide binding positions of the tagged RBP; S4. NeuroPRIS was used to elucidate the relationship between these lincRNAs and RBPs. Using mRNA and untranslatable small peptide lincRNA as controls, a positive correlation between RBP-RNA interactions and RNA translation was established, especially between RBP-RNA interactions and lincRNAs.

2. The protein-RNA interaction prediction method based on multimodal deep learning according to claim 1, characterized in that The NeuroPRIS in step S1 integrates multi-source data to download complete sequences of lincRNA and mRNA from the ENCODE and ENCORI databases, removes RNA longer than 7500 nt, uses RNAfold and 3dRNA to obtain the secondary and tertiary structures of the complete RNA, selects binding sites with a length greater than 10 and less than 150 nucleotides from the CLIP experiment, extracts the primary sequence, secondary structure, and tertiary structure of the RNA fragments at these sites from the complete RNA, and fixes the RNA fragments to 150 nt in length. Positive samples are marked as "1" and negative samples are marked as "0". For each RBP dataset, a NeuroPRIS model is constructed for each cell line. By default, 80% of the input data is used for training, 10% for validation, and 10% for testing. NeuroPRIS trains the model for each RBP with CLIP experimental data through multiple iterations of the training set and validation set, evaluates the score of each nucleotide position within the binding site, and finally outputs the score of the entire binding site.

3. The protein-RNA interaction prediction method based on multimodal deep learning according to claim 2, characterized in that The RBP training model input includes sequence input, secondary structure input and 3D coordinate input. The 3D coordinate input is feature extracted through the T-net architecture. T-net consists of multiple convolutional layers and a global maximum pooling layer, which is specifically implemented by three convolutional layers. The three convolutional layers use 32, 64 and 128 filters respectively, and the kernel size is 1. It also includes an output layer that obtains feature transformation by linear mapping of input features, so that the model can autonomously learn feature representations that are more suitable for downstream tasks. The features processed by T-net are combined with the embedding of sequence and structural features. The sequence is embedded through an embedding layer with an input dimension of 5 and an output dimension of 127. Converted to a high-dimensional representation, the secondary structure is converted to a high-dimensional representation through an embedding layer with an input dimension of 3 and an output dimension of 128. The embedded features are then combined with the output of T-net through a dot operation to create a multimodal input. These fused features are input to a bidirectional LSTM network layer containing 48 LSTM units to capture the temporal dependencies in the input sequence. The return sequence mode of the BiLSTM layer facilitates the subsequent output layer to process sequence information. Finally, the BiLSTM output is classified through a fully connected layer with two output classes, and the softmax activation function is applied to generate the final class probability. All convolutional and fully connected layers in the model use Batch Normalization to enhance training stability. A dropout layer with a dropout rate of 0.1 is introduced after each activation layer to reduce the risk of overfitting. The model uses the Adam optimizer, a batch size of 64, and an initial learning rate of 0.

001.

4. The protein-RNA interaction prediction method based on multimodal deep learning according to claim 2, characterized in that The computational technology in step S2 uses NeuroPRIS to reveal unique RBP-RNA binding patterns in different cell lines, and finds that RBPs of the IGF2BPs family have high specificity when binding to RNA, and these highly specific binding RBPs play a key role in specific cell lines.

5. The protein-RNA interaction prediction method based on multimodal deep learning according to claim 1, characterized in that The cell line-specific RNA structural similarity matrix in step S2 is constructed using RNA clustering and analysis, using the internal qTMclust algorithm of US-align to perform multiple structural alignments, and using two cutoff values ​​to cluster the RNA binding site structures. Through structural cluster analysis, the unique core structure of the RNA binding site and the relationship between different cell line specificities and the RNA binding site structure are revealed, and a cell line-specific RNA structural similarity matrix is ​​constructed to reveal the "sequence-independent, structure-driven" interaction rules.

6. The protein-RNA interaction prediction method based on multimodal deep learning according to claim 1, characterized in that The NeuroPRIS visualization binding sites in step S3 are spliced ​​along the sequence dimension by three-dimensional point cloud features and one-dimensional sequence information, and the context dependency is captured using bidirectional LSTM. The bidirectional LSTM time step hidden state is extracted and summed to generate an interpretable RNA-binding protein motif, which clarifies that the RNA-RBP interaction is completed by a small number of "core binding bases".

7. The protein-RNA interaction prediction method based on multimodal deep learning according to claim 1, characterized in that The lincRNA in step S4 is a long intergenic non-coding RNA. The lincRNA was downloaded from LncPepAtlas, and the lincRNAs marked as verified by MC-related literature were screened out. From all the lincRNAs in the GENCODE database, the lincRNAs that were verified as translatable peptides by multiple dimensions were eliminated. The multi-dimensional verification included MC-related literature verification, m6A, Ribo, IRES, CPAT, CPC2, Pfam, and CNCI. The lincRNAs that were not translatable peptides were retained. mRNAs were randomly selected from a total of 111,868 mRNAs in the GENCODE database. RNAs with a length of more than 7,500 nt were eliminated from the three data sets, and a small amount of RNAs that were not verified by R were removed. The RNAs identified by NAfold and 3dRNA ultimately yielded 787 lincRNAs of potential translatable small peptides, 1784 lincRNAs of untranslatable small peptides, and 1763 mRNAs. NeuroPRIS was used to predict the binding of each group of RNAs to RNA-binding protein RBP. In the three data sets, the binding trends of lincRNAs and mRNAs of potential translatable small peptides to RBP were similar, while the binding probability of lincRNAs of untranslatable small peptides to RBP was significantly different from the former two. This difference is due to the fact that lincRNAs and mRNAs of potential translatable small peptides have specific functional binding sites for RBP, thereby promoting translation, clarifying the positive correlation between lincRNA-RBP binding and lincRNA translation.

Images