Cross-species RBP-RNA interaction prediction method, system and application

Abstract

The invention discloses a cross-species RBP-RNA interaction prediction method and system and application, and relates to the technical field of RNA post-transcription regulation and deep learning. The method comprises the following steps: acquiring a to-be-predicted RNA primary sequence, and performing structure prediction on the RNA primary sequence to obtain a secondary structure sequence of RNA; constructing an interaction prediction model, and training the interaction prediction model by using a data set formed by CLIP data of multiple species; and predicting the secondary structure sequence of the RNA and the primary sequence of the RNA by using the trained interaction prediction model to obtain the binding condition of the RNA binding protein and the RNA. According to the method, a deep learning model is constructed by utilizing an RBP evolutionary conservative signal and a tag smoothing regularization strategy, so that accurate prediction of multi-species RBP-RNA interaction can be realized, the cross-species prediction capability can be effectively improved, and the limitation of a traditional method is made up.

Description

Technical Field

[0001] This invention relates to the fields of RNA posttranscriptional regulation and deep learning technology, and in particular to a method, system and application for predicting cross-species RBP-RNA interactions. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] RNA-binding proteins (RBPs) play a crucial role in post-transcriptional gene expression regulation. By binding to specific RNA primary sequences or motifs, RBPs can influence RNA transcription, splicing, transport, and metabolism. The function of RBPs is vital in many biological processes; therefore, in-depth research into their interactions with RNA has significant biomedical implications. Numerous studies have confirmed that abnormal RBP-RNA interactions are closely related to the development and progression of various diseases, including cancer, neurodegenerative diseases, and cardiovascular diseases. Due to ethical controversies and cost issues involved in conducting disease mechanism research and therapeutic exploration in humans, it is necessary to first elucidate disease mechanisms and develop drugs using RBP-RNA interaction research techniques in model organisms such as mice and zebrafish. Currently, methods for detecting RBP-RNA binding sites (CLIP: Cross-linking immunoprecipitation) mainly rely on high-throughput experimental techniques such as eCLIP and PAR-CLIP. These techniques, by capturing the RBP-RNA binding sites, provide important experimental evidence for elucidating RBP function. However, since such techniques usually require specific antibodies and have problems such as long experimental cycles and high background noise, obtaining comprehensive RBP binding information in animal models still faces certain challenges.

[0004] With the advancement of artificial intelligence technology, computational models based on large amounts of RBP-RNA binding data have gradually become an effective research tool. Traditional machine learning and deep learning methods have been used to predict RBP binding sites and RNA sequence preferences, successfully incorporating the original sequence and structural information of RNA into the models and improving the accuracy of cross-species RBP-RNA interaction prediction. However, most existing computational methods are limited to data from specific species, especially since human RBP-CLIP data is far more abundant than data from other species, resulting in a significant limitation in the ability of computational models to predict RBP-RNA interactions for other species.

[0005] Numerous studies have found that RBPs are highly conserved in their core RNA-binding domains (RBDs) across species. Furthermore, RBPs with highly conserved RBDs exhibit highly similar binding RNA motifs. Current research has not fully explored how to leverage RBP evolutionary conservation to improve the predictive ability of cross-species RBP-RNA interactions. Therefore, how to utilize the conservation signals of RBPs across different species and the existing CLIP dataset to improve the predictive accuracy of models, especially in species with scarce data, remains a pressing challenge. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and application for predicting cross-species RBP-RNA interactions. By utilizing RBP evolutionary conservation signals and tag smoothing regularization strategies to construct a deep learning model, it is possible to accurately predict multi-species RBP-RNA interactions, effectively improving cross-species prediction capabilities and overcoming the limitations of traditional methods.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a method for predicting cross-species RBP-RNA interactions, comprising the following steps: Obtain the primary sequence of the RNA to be predicted, and perform structural prediction on the primary sequence to obtain the secondary structure sequence of the RNA; An interaction prediction model was constructed and trained using a dataset consisting of CLIP data from multiple species. The trained interaction prediction model was used to predict the secondary structure sequence and primary sequence of RNA to obtain the binding status of RNA-binding proteins with RNA.

[0008] Furthermore, RNAfold is used to predict the primary RNA sequence to obtain the secondary RNA sequence, and the secondary RNA sequence and the primary RNA sequence are one-hot encoded as input to the interaction prediction model.

[0009] Furthermore, the specific steps for training the interaction prediction model using a dataset composed of CLIP data from multiple species are as follows: A dataset was constructed using CLIP data from multiple species that were labeled. The labels are smoothed and regularized based on the RBP similarity between species. The dataset is divided into a training set and a validation set; The interaction prediction model is trained using the training set and its hyperparameters are fine-tuned using the validation set.

[0010] Furthermore, the multi-species CLIP data includes human RNA primary sequence data, mouse RNA primary sequence data, and zebrafish RNA primary sequence data.

[0011] Furthermore, when dividing the dataset into training and validation sets, human RNA primary sequence data is used as the training set, while mouse RNA primary sequence data and zebrafish RNA primary sequence data are used as the validation set.

[0012] Furthermore, the specific steps for smoothing and regularizing labels based on inter-species RBP similarity are as follows: The hard tags of the RNA primary sequence are smoothed and regularized to be converted into soft tags weighted by RBP similarity. The cross-entropy loss of hard labels and soft labels is calculated separately, and an adaptive gradient weight mechanism is adopted to balance the contribution of the cross-entropy loss of hard labels and soft labels to the total loss.

[0013] Furthermore, when training the interaction prediction model using the training set, convolutional layers and residual blocks are used to capture key features related to the combination of RBP, and the SE module is used to adaptively adjust the channel feature weights.

[0014] Furthermore, during the training of the interaction prediction model, Dropout and early stopping strategies are used to prevent overfitting, and smoothing gradients and guided backpropagation techniques are used to visualize important feature regions, further helping to identify high-response regions associated with RBP binding.

[0015] A second aspect of the present invention provides a cross-species RBP-RNA interaction prediction system, comprising: The data acquisition module is configured to acquire the primary sequence of the RNA to be predicted, and to perform structural prediction on the primary sequence of the RNA to obtain the secondary structure sequence of the RNA. The model training module is configured to build an interaction prediction model and train the interaction prediction model using a dataset consisting of CLIP data from multiple species. The interaction prediction module is configured to use the trained interaction prediction model to predict the secondary structure sequence and primary sequence of RNA, thereby obtaining the binding status of RNA-binding proteins with RNA.

[0016] The third aspect of this invention provides an application of the cross-species RBP-RNA interaction prediction system of the second aspect in predicting single nucleotide variations in mice.

[0017] The above one or more technical solutions have the following beneficial effects: This invention discloses a method, system, and application for predicting cross-species RBP-RNA interactions. Through deep learning, a model capable of accurately predicting RBP-RNA interactions across multiple species is constructed. By integrating primary RNA sequences and predicted secondary structures from multiple species, and training the deep learning model using RBP evolutionary conservation signals and tag smoothing regularization strategies, RNA binding site prediction can be achieved for species with scarce data. This technology can effectively improve cross-species prediction capabilities, overcome the limitations of traditional methods, and has significant application potential, particularly in in-depth research on RBP-RNA interactions in species other than humans.

[0018] The MuSIC model proposed in this invention demonstrates superior performance in predicting cross-species RNA-binding protein (RBP)-RNA interactions, such as... Figure 3 As shown, MuSIC demonstrates a significant advantage over existing benchmark methods (such as HDRNet, PrismNet, DeepBind, and GraphProt) in cross-species prediction tasks. Traditional methods typically train and test only on datasets of the same species, while MuSIC successfully improves cross-species generalization by training the model on one species and testing it on another. This approach not only improves prediction accuracy but also effectively addresses inter-species differences, enhancing cross-species learning ability and universality. In applications, MuSIC significantly outperforms other methods in AUC, PRAUC, and accuracy across multiple species (such as humans, mice, and zebrafish), showcasing its advantages in distinguishing positive and negative samples, feature representation capabilities, and detail capture.

[0019] The MuSIC model proposed in this invention demonstrates extremely high accuracy in predicting RBP binding sites. Comparison with experimental CLIP data shows that MuSIC's prediction results in RBP datasets from species such as mice and zebrafish significantly outperform existing methods, proving its high accuracy and robustness in cross-species RBP-RNA interaction prediction. Experimental validation shows that the overlap between the binding sites predicted by MuSIC and the actual data is significantly higher than other methods, demonstrating its unique advantage in binding site prediction. This achievement is of great significance for in-depth research on RBP function and RNA interactions, providing a more reliable reference for further biological research.

[0020] The MuSIC model proposed in this invention can not only predict RBPs in a single species, but also discover and annotate RBP binding motifs across species, such as... Figure 4As shown in the figure. By analyzing RBP binding patterns across different species, MuSIC identified cross-species conserved features, revealing differences in RBP binding preferences among different species. This innovation gives MuSIC unique cross-species annotation capabilities in RBP research. When predicting RBP binding motifs in species such as humans, mice, and zebrafish, MuSIC successfully identified motifs consistent with known binding preferences and demonstrated the evolutionary trends of cross-species motifs, providing profound insights into the evolution of RBP binding.

[0021] The MuSIC model proposed in this invention demonstrates unique advantages in predicting the impact of single nucleotide variants (SNVs) on RBP-RNA interactions across species. SNVs are key factors contributing to various genetic diseases, cancers, and neurodegenerative diseases. They affect RNA recognition and regulation by altering RBP binding specificity, ultimately leading to disease development. MuSIC can accurately predict changes in RBP-RNA binding affinity caused by different SNV variants, providing important evidence for disease mechanism research. In vitro and in vivo experiments have validated the accuracy of the MuSIC model in predicting changes in RBP-RNA binding capacity, particularly its reliability in cross-species RBP-RNA interaction studies. MuSIC provides strong support for disease mechanism research, drug development, and targeted therapy.

[0022] The MuSIC model proposed in this invention successfully identified multiple potential drug targets and pathogenic mechanisms through large-scale screening of SNV-RBP binding abnormalities in mice, providing new directions for target discovery and clinical research, particularly in cancer and neurodegenerative diseases. By combining ubiquitin-proteasome system (UPS)-related genes, MuSIC revealed the impact of SNVs on RBP function and linked it to clinical diseases such as cancer and neurodegenerative diseases in humans. MuSIC can not only efficiently screen RBPs related to disease mechanisms but also verify their mechanisms in human diseases, providing important support for the early diagnosis and targeted therapy of more diseases.

[0023] The MuSIC model proposed in this invention has broad application potential. It can not only accurately predict cross-species RBP-RNA interactions in more species, but also be extended to a wider range of species, especially those with limited genomic and CLIP experimental data. MuSIC's cross-species predictive capabilities make it a powerful tool for cross-species research, accurately identifying similar variant sites between different species and supporting disease mechanism research globally. Furthermore, MuSIC's predictive capabilities provide precise targets and treatment strategies for drug development, drug screening, and precision medicine, driving progress in new drug discovery, treatment design, and personalized disease treatment. The enormous potential of MuSIC in basic research, drug development, and clinical applications provides innovative technological support for future disease treatment and prevention.

[0024] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the improved deep learning model MuSIC architecture in this invention; Figure 2 This is a diagram showing the results of the conservation analysis of RNA-binding proteins (RBPs) in 11 species in this invention; Figure 3 This is a schematic diagram comparing the ROC curves and other performance evaluation metrics of the MuSIC model in this invention with existing models on cross-species problems; Figure 4 The image shows the RBP binding motifs of mice, rats, chickens, zebrafish, and fruit flies identified by the MuSIC model in this invention, and compared with the actual motifs. Figure 5 This is a flowchart illustrating the impact of SNV on CLIP peak signals and MuSIC model prediction results in this invention. Figure 6 This figure shows the consistency between the MuSIC model and the CLIP experimental results in identifying the influence of SNVs on RBP-RNA interactions in this invention. Figure 7This is a flowchart illustrating the effect of the MuSIC model on RBP-RNA interaction in humans and mice in this invention. Figure 8 This is a visualization of the effects of homologous SNVs on RBP-RNA interactions in humans and mice in this invention. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0029] This invention provides a method for predicting cross-species RBP-RNA interactions, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the primary sequence of the RNA to be predicted, and perform structural prediction on the primary sequence of the RNA to obtain the secondary structure sequence of the RNA.

[0030] Studies such as GraphProt, PrismNet, and HDRNet have demonstrated the crucial role of RNA secondary structure in enhancing model learning capabilities. RNA secondary structure exhibits higher cross-species conservation than primary sequence, reducing sequence variance interference and becoming a core stable feature. This enhances the model's generalization ability across different species, overcoming the barriers of diverse genomic backgrounds. Therefore, this invention uses RNAfold to predict RNA primary sequence, obtaining the RNA secondary structure sequence. The secondary structure sequence and the primary sequence are then one-hot encoded and used as input to an interaction prediction model.

[0031] Step 2: Construct an interaction prediction model and train it using a dataset consisting of CLIP data from multiple species.

[0032] Step 2.1: Construct an interaction prediction model.

[0033] Step 2.2: Train the interaction prediction model using a dataset consisting of CLIP data from multiple species.

[0034] This invention constructs an interaction prediction model based on a convolutional neural network framework, and optimizes the existing MuSIC model in multiple dimensions, covering aspects such as input feature length, feature information, and model framework selection.

[0035] The input to the MuSIC model is a set of RNA sequences containing N samples, each sample being a 200nt RNA sequence. After encoding, these sequences form a (6,200) matrix. The overall input is a (N,6,200) tensor containing the sequence and structural features of the RNA, which serves as the raw data for model processing.

[0036] The model obtains the binding probability of RNA and RBP through multi-layer processing. The process consists of convolutional blocks, squeeze-activation blocks, residual blocks, and fully connected layers, and finally outputs the results through the sigmoid activation function.

[0037] The convolutional block is the first processing layer of the model, receiving the raw feature tensor of the input. It contains 2D convolutional layers, batch normalization, and the ReLU activation function: the 2D convolutional layers are responsible for extracting local sequence patterns in the input, such as the base combination patterns of short segments; batch normalization stabilizes the training process by standardizing the feature distribution and accelerates model convergence; the ReLU activation function introduces non-linearity, enhancing the model's ability to express complex features while filtering out invalid negative feature values.

[0038] The squeeze-excitation block receives features from the output of the convolutional block. Its role is to optimize the feature representation by recalibrating the feature responses in the channel dimension using global information, emphasizing the features most relevant to the RBP and suppressing irrelevant features, allowing the model to focus more on key information.

[0039] Residual blocks receive features processed by squeeze-excitation blocks to capture long-range sequence and structural patterns. There are two types: ResidualBlock2D directly operates on the 2D feature map, while ResidualBlock1D first performs average pooling on the output of ResidualBlock2D before processing the 1D feature map. Both types of residual blocks avoid the vanishing gradient problem through skip connections, enabling the network to learn deeper features. Furthermore, batch normalization and ReLU activation functions are used to further optimize the features.

[0040] The fully connected layer receives features from the residual block output, processes them through two linear transformations (with ReLU activation in between), and then outputs the result after passing it through the sigmoid activation function. Its function is to integrate the previously extracted features and finally output a value between 0 and 1, representing the binding probability of the RNA sequence to RBP.

[0041] Step 2.2.1: Construct a dataset using the labeled multi-species CLIP data.

[0042] First, this invention downloaded and analyzed six-dimensional CLIP data of 216 RNA-binding proteins (RBPs) from 11 species. This CLIP data was used for smoothing coefficients in the subsequent tag smoothing process, such as... Figure 2 As shown in the figure. Analysis revealed that the highly conserved RBP across different species also exhibited significant conservation in its RNA binding pattern. Then, CLIP sequencing data from multiple species were collected from the POSTAR3 database to construct a training and validation set containing abundant labeled species data. This was used to validate the performance of the MuSIC model in both cross-species and intra-species capabilities and to explore its application in other species biology problems.

[0043] Specifically, to fully validate the cross-species capability of MuSIC, for a given RBP, human data needs to be used as the training set and data from another species as the validation set to evaluate MuSIC's cross-species capability. Due to data limitations in the POSTAR3 database, 11 species were selected based on the overlap of RBP data to determine the multi-species CLIP data, which included human RNA primary sequence data, mouse RNA primary sequence data, and zebrafish RNA primary sequence data.

[0044] Step 2.2.2: Perform smoothing regularization on the labels based on the RBP similarity between species.

[0045] In the field of computer vision, similarity exists between different types of images. For example, a cartoon image of a cat and a simple line drawing can be converted from hard labels (0, 1) to soft labels using similarity relationships. Based on this principle, this invention also smooths and regularizes the hard labels of RNA sequences, transforming them into soft labels weighted by RBP similarity. In the hard labels, 1 represents a positive sample, corresponding to a sequence in CLIP data that binds to RBP; 0 represents a negative sample, corresponding to a sequence outside CLIP data that does not bind to RBP. The soft labels are the smoothed result after RBP similarity weighting of the hard labels. For example, the original hard label 1 might be adjusted to 0.9, and the original hard label 0 might be adjusted to 0.1.

[0046] Step 2.2.2.1: Smooth and regularize the hard tags of the RNA primary sequence to convert them into soft tags weighted by RBP similarity.

[0047] Step 2.2.2.2: Calculate the cross-entropy loss of hard labels and soft labels respectively, and adopt a gradient weight adaptive mechanism to balance the contribution of the cross-entropy loss of hard labels and soft labels to the total loss, thereby avoiding one loss from dominating the optimization process and improving the robustness of the model.

[0048] Step 2.2.3: Divide the dataset into a training set and a validation set.

[0049] In this invention, when dividing the dataset into training and validation sets, human RNA primary sequence data is used as the training set, while mouse and zebrafish RNA primary sequence data are used as the validation sets. This division, using human RNA primary sequence data as the training set and mouse and zebrafish RNA primary sequence data as the validation set, aims to rigorously evaluate the model's cross-species generalization ability. Its design principles and advantages are closely aligned with the core requirement of cross-species prediction. From a molecular evolutionary perspective, RNA sequences from different species exhibit a clear "evolutionary distance gradient"—humans and mice belong to the same mammal family and are evolutionarily closely related, while humans and zebrafish are evolutionarily distant. This division, through complete species isolation between the training and validation sets, simulates the real-world scenario of the model learning patterns from a "known species (humans)" and migrating predictions to "unknown species (mice, zebrafish)." If the model performs well on mouse and zebrafish data, it indicates that it captures not human-specific sequence features, but rather conserved features universal across species, thus effectively validating the accuracy and robustness of the model's cross-species prediction ability. This classification method avoids over-reliance on human-specific traits by isolating species, ensuring that the assessment results accurately reflect cross-species migration capabilities. By utilizing the evolutionary distance difference between mice (closely related) and zebrafish (distantly related), the generalization boundary of the model is examined hierarchically, clarifying which species' RBPs can be used for cross-species predictions. This classification method aligns with practical application scenarios, directly demonstrating the model's practical value in cross-species research and enhancing its scientific explanatory power.

[0050] Step 2.2.4: Train the interaction prediction model using the training set, and fine-tune the hyperparameters of the interaction prediction model using the validation set.

[0051] This invention integrates convolutional layers and residual blocks into the interaction prediction model architecture to further improve model performance during training using a training set. These convolutional layers and residual blocks capture key features related to RBP binding, and the Squeeze-and-Excitation (SE) module adaptively adjusts channel feature weights, thereby enhancing the representational power of key features. During training, Dropout and early stopping strategies are used to prevent overfitting, and gradient smoothing and guided backpropagation techniques are employed to visualize important feature regions, further aiding in the identification of high-response regions related to RBP binding. Finally, the trained model maps the learned features to the prediction output layer.

[0052] The input data preprocessing for the MuSIC model revolves around the sequence information and secondary structure information of RNA sequences. For each RNA sequence in the dataset, it is first converted into a 4-dimensional matrix (size 4×200) using one-hot encoding, where each dimension corresponds to nucleotides A, C, G, and U, respectively. Simultaneously, RNAfold (version 2.1.0) is used to predict the secondary structure of the RNA sequence under default settings, and the structural information is encoded into a 2-dimensional matrix (size 2×200), where "P" represents paired bases and "U" represents unpaired bases. Finally, the 4-dimensional matrix of sequence information and the 2-dimensional matrix of secondary structure are concatenated to form a 6×200 combined matrix, which contains both sequence and structural features and serves as the input to the MuSIC model.

[0053] The input data is a tensor that integrates sequence and structural information, with a shape of (N, 6, 200), where N represents the number of samples in the batch, 6 corresponds to the feature dimension (containing sequence information of 4 nucleotides and secondary structure state information of 2), and 200 is the fixed length of the RNA sequence.

[0054] First, the data enters the convolutional block for initial feature extraction. This module employs padded 2D convolution operations, effectively capturing local sequence patterns and structural associations between adjacent nucleotides while maintaining the feature map size by using preset kernel size and padding strategies. Immediately after the convolution operation, batch normalization is performed to standardize the feature distribution, stabilizing the training process and accelerating model convergence. Then, a non-linear transformation is introduced using the ReLU activation function to filter out invalid features, further enhancing the model's feature representation capability.

[0055] The feature maps processed by the convolutional blocks then enter the squeeze-activation block for channel-level feature recalibration. This module achieves adaptive enhancement of key features through a "squeeze-activation" mechanism: first, global pooling is used to compress the two-dimensional feature map of each channel into a single value, thereby aggregating global information; then, a non-linear transformation is performed through a fully connected layer containing ReLU activation to learn the dependencies between channels; finally, the channel weights in the range of 0-1 are output by the sigmoid function, and these weights are multiplied element-wise with the original feature map to dynamically enhance the channel feature weights that are crucial for RBP prediction, while suppressing redundant information.

[0056] After optimization by the squeeze-activation block, the features are fed into the residual block for deep feature learning. The model includes two types of residual blocks: ResidualBlock2D directly operates on the 2D feature map, further extracting local and mid-range feature associations through multiple convolutional operations; ResidualBlock1D processes the 1D features obtained after average pooling of the feature map output from the 2D residual block, focusing more on capturing long sequence patterns. Both residual blocks employ a skip connection design, directly adding the input features to the output features after convolution and activation processing, effectively alleviating the gradient vanishing problem in deep networks, enabling the model to learn hierarchical long sequence patterns and structural patterns, taking into account both local details and global features.

[0057] Finally, the features processed by the residual blocks are fed into the fully connected layer. The fully connected layer performs feature dimensionality reduction and nonlinear processing through two linear transformations: the first linear transformation is followed by ReLU activation to introduce nonlinearity, and the second linear transformation maps the features to a single dimension. Finally, the output value is compressed into the range of 0-1 by the sigmoid activation function, and this value is the predicted binding probability of the RNA sequence with RBP.

[0058] The model captures key features through a multi-layered collaborative structure: convolutional blocks directly extract local sequences and structural patterns through 2D convolution, which is the foundation of feature extraction; residual blocks achieve deep network training through skip connections, gradually learning long-range dependencies and hierarchical features, taking into account both local details and global trends; and SE blocks enhance the representation of key features through a "squeeze-excitation mechanism"—first, the feature map is globally compressed to obtain global information, then channel weights are generated through nonlinear transformation and sigmoid, and finally multiplied element-wise with the original features, so that the model adaptively focuses on the channel features that are more important to the RBP combination, thereby improving feature utilization.

[0059] To prevent overfitting, the model incorporates Dropout technology at key locations in its multi-layer structure (convolutional layers, squeeze-excitation blocks, and one-dimensional residual blocks). By randomly deactivating some neurons, excessive dependence between neurons is reduced, enhancing the model's generalization ability. The model also employs an early stopping strategy: during training, performance (AUC and ACC metrics) is periodically evaluated on the validation set. If no significant improvement is observed after a specified number of consecutive epochs, training is terminated, and the current optimal model weights are retained.

[0060] Label smoothing achieves regularization by softening the distribution of target labels. For the original hard labels, the smoothed label distribution consists of two parts: one part is the original label multiplied by a smoothing coefficient, and the other part is the uniform distribution of positive and negative samples in the dataset multiplied by (1 minus the smoothing coefficient). The value of the smoothing coefficient is determined based on the interspecies conservation of RNA-binding proteins. The smoothed loss is calculated using cross-entropy, which can be divided into two parts: a regular classification loss (i.e., the loss between the predicted result and the hard label) and a smoothing loss (i.e., the loss between the predicted result and the soft label). The smoothing loss helps the model reduce the confidence of predictions when facing ambiguous categories, avoiding overfitting to a single label. The final total loss is a weighted sum of the classification loss and the smoothing loss, with two weighting parameters used to balance their effects.

[0061] Gradient adaptation is primarily used to resolve potential conflicts between classification loss and smoothing loss in terms of gradient direction. In its implementation, it first calculates the gradients of each loss with respect to the encoder and classifier parameters, respectively. Next, it determines the importance weights of the two losses by comparing the current loss to the previous loss: the weight of the original loss is the reciprocal of its ratio, and the weight of the smoothing loss is the reciprocal of its ratio. These weights are then normalized so that their sum equals 1. Finally, the normalized weights are multiplied by their corresponding gradients and summed to obtain the total gradient used for parameter updates. This dynamic adjustment method allows the gradient to flow in a direction more conducive to model optimization, reasonably balancing the contributions of the two losses in parameter updates, and improving the stability and effectiveness of training.

[0062] The softened target distribution (i.e., the introduction of soft labels) makes the loss function more flexible, preventing the model from making overconfident predictions about a single label, thus guiding a more robust and reliable gradient update process. Backpropagation is guided by a gradient adaptation strategy, dynamically adjusting the gradient weights of the original loss and the smoothed loss to ensure that the gradient contributions of these two losses are reasonably distributed according to the current training state, effectively avoiding training instability caused by conflicting gradient directions. In actual training, combined with an adaptive momentum optimizer and a gradient preheating scheduler, the model can update parameters more efficiently through the backpropagation algorithm. This process minimizes the total loss while effectively balancing the model's fitting ability and generalization ability, ultimately improving prediction performance.

[0063] By improving and optimizing in multiple aspects, the interaction prediction model of this invention effectively integrates and optimizes multi-dimensional features, showing significant advantages compared with existing RBP-RNA interaction prediction deep learning or machine learning models.

[0064] Step 3: Use the trained interaction prediction model to predict the secondary structure sequence and primary sequence of RNA to obtain the binding status of RNA-binding proteins with RNA.

[0065] Previous deep learning models, such as DeepBind, PrismNet, and HDRNet, typically use sequence input lengths of 101 nt (DeepBind, PrismNet, HDRNet) or 400 nt (GraphProt). Existing research has shown that excessively short sequence lengths lead to insufficient information learned by the model, affecting its cross-species generalization ability; while excessively long sequences may contain redundant information, thus reducing the model's learning efficiency. Against this backdrop, this invention selects 200 nt as the model's sequence input length. The advantage of a 200 nt sequence input length lies in balancing information completeness and learning efficiency: compared to the shorter 101 nt sequence, it can contain richer RNA sequence and structural information, reducing the problem of limited cross-species generalization ability due to insufficient information; while compared to the longer 400 nt sequence, it can effectively avoid interference from redundant information, reducing the model's learning burden and improving training efficiency, thereby achieving a better balance between capturing key features and optimizing learning effects.

[0066] This invention provides a cross-species RBP-RNA interaction prediction system, comprising: The data acquisition module is configured to acquire the primary sequence of the RNA to be predicted, and to perform structural prediction on the primary sequence of the RNA to obtain the secondary structure sequence of the RNA. The model training module is configured to build an interaction prediction model and train the interaction prediction model using a dataset consisting of CLIP data from multiple species. The interaction prediction module is configured to use the trained interaction prediction model to predict the secondary structure sequence and primary sequence of RNA, thereby obtaining the binding status of RNA-binding proteins with RNA.

[0067] This invention provides an application of the above-mentioned cross-species RBP-RNA interaction prediction system in predicting single nucleotide variations in mice.

[0068] Example 1: This embodiment uses single nucleotide variants (SNVs) as an example to illustrate the application steps of the above method using the obtained RNA primary sequence.

[0069] Single nucleotide variants (SNVs) are the most common type of variation in the genome, widely present in the genomes of humans and other species. These variants typically occur in coding or regulatory regions of genes, thereby affecting the interaction between RNA-binding proteins (RBPs) and RNA. RBPs regulate important biological processes such as RNA transcription, splicing, stability, and translation by binding to specific regions of the RNA sequence. Therefore, SNVs, by altering RBP-RNA interactions, may lead to abnormal gene expression, thereby triggering various diseases.

[0070] While existing methods can analyze the impact of SNVs on RBP-RNA interactions, these techniques are typically limited to single species or specific mutation sites and have limitations in predictive accuracy, cross-species applicability, and comprehensive consideration of genomic background. Therefore, this embodiment proposes a novel cross-species approach that combines the deep learning model MuSIC with CLIP-seq data, RNA primary sequences, and RNA secondary structure information from multiple species to comprehensively predict the impact of SNVs on RBP-RNA interactions, particularly the variation effects across different species.

[0071] 1. Data preprocessing and peak file acquisition like Figure 5 The steps shown include dbSNP data preprocessing, eCLIP data acquisition, and screening of high-confidence SNV sites.

[0072] 1.1 dbSNP data preprocessing: This embodiment first obtains single nucleotide variant (SNV) data from the human genome from the dbSNP database. dbSNP is an important public database containing genome-wide variation information, providing extensive SNV data covering genetic variations in different individuals and populations. SNV sites are extracted and these data are standardized. Specifically, the format of each SNV site is adjusted to "chr:position-base1>base2", where "chr" represents the chromosome number, "position" represents the specific location of the variant site on the chromosome, "base1" is the base in the reference genome, and "base2" is the detected variant base. For example, "chr10:74000-T>G" represents position 74000 on chromosome 10, where the T base is replaced by a G base. This standardized format accurately expresses the SNV site information, providing a reliable foundation for subsequent analysis and calculations. This format ensures the accuracy and consistency of the variation information and facilitates integration and analysis with other data sources such as eCLIP experimental data.

[0073] 1.2 eCLIP Data Acquisition: After preprocessing the dbSNP data, eCLIP experimental data for 103 RNA-binding proteins (RBPs) in the HepG2 cell line were downloaded from the ENCODE database. eCLIP (Enhanced Crosslinking and Immunoprecipitation) is a high-throughput technique for studying RBP-RNA interactions, providing accurate localization of RNA-binding regions and their corresponding binding strengths. Each RBP dataset contains three independent sets of replicate peak files, recording the RBP binding sites and signal intensities within the cells.

[0074] To ensure the reliability and consistency of the data, this embodiment performs rigorous quality control and filtering on the raw data. First, statistical tests are used to screen out significant peaks with p-values ​​less than 0.05 after p-adjustment, ensuring that the retained data has high statistical significance. Next, to improve the reproducibility and reliability of the data, the data from the three sets of replicate experiments are merged, and the peak regions shared by all replicate experiments are extracted. These overlapping portions represent peaks with high reproducibility and high signal-to-noise ratio, ensuring the quality of the selected data. Finally, these high-quality peak files are saved as standard BED format files as the basis for subsequent analysis.

[0075] 1.3 Screening of high-confidence SNV sites The bcftools tool was used to extract the SNV sites annotated in the BAM files corresponding to the eCLIP data (eCLIP data typically identifies a large number of SNV sites). To ensure the reliability of the sites, these SNV sites were compared with the SNV sites in the dbSNP database after preprocessing in section 1.1, and only the SNV sites already annotated in dbSNP were retained as high-confidence SNV sites. Specifically, if the eCLIP data showed an A→G mutation at position 1000 of chr1, but this mutation was not annotated in the dbSNP database, it was considered a possible sequencing error and excluded; if the eCLIP data detected a C→U mutation at position 100 of chr2, and this mutation had been included in dbSNP, it was considered a true mutation. Based on the high-quality peak files obtained from the eCLIP data, the impact of the mutation on the protein binding signal could be further analyzed based on each BED file—that is, comparing the binding peaks before and after the mutation (the binding peak reflects the binding strength of RNA-binding proteins on the RNA fragment).

[0076] The changes in SNV combined with the peak value of RBP, as analyzed in this step, can be used to further verify whether the trend predicted by MuSIC is consistent with the actual trend.

[0077] 2. Consistency Analysis of MuSIC Model Predicted Trends and Actual Trends This embodiment uses the MuSIC model, an interaction prediction model, to predict the impact of SNVs within a specific region, such as... Figure 6 As shown, the results indicate that the MuSIC model can accurately predict the impact of SNVs on RBP-RNA interactions, especially for proteins such as AGGF1, DDX3X, and SRSF1. For example, in the eCLIP experiment, the peak signal of AGGF1 in the chr19:47507208-47507493 region was 5.68 without SNVs, while it dropped to 2.04 with SNVs. MuSIC showed the same trend, with a prediction score of 0.94 without SNVs and 0.69 with SNVs. Using data from the dbSNP database, RNAseq SNVs were used as background data, and eCLIP SNVs were used as the true distribution. This allowed for the calculation of the diff. ratio This study quantifies the actual impact of SNVs on RBP-RNA interactions in the cellular environment. Simultaneously, the MuSIC model was also used to predict the effect of SNVs on RBP-RNA interactions. The significance index (-log) was used to... 10The consistency between the predicted and actual results was assessed using the FDR (Field-Dependent Ratio) method, and the results showed that the affected sites predicted by the MuSIC model were highly consistent with the experimental data. These results indicate that the MuSIC model has high effectiveness in detecting the impact of SNVs on RBP-RNA interactions.

[0078] 3. Cross-species analysis Studies have shown that many human disease mechanisms are explored through disease models based on mice or other model organisms, and then validated through clinical studies. Based on the ability of MuSIC to accurately detect the impact of SNVs on RBP-RNA interactions in humans, this embodiment performed a cross-species analysis to assess the effect of mouse SNVs (human homologous regions) on RBP-RNA interactions, such as... Figure 7 and Figure 8 As shown, human dbSNP data were BLASTed into the mouse genome, identifying SNV sites in 85,599 homologous regions. MuSIC's 186 mouse RBP models predicted that 107,639 of these SNV sites might affect RBP-RNA interactions (∆SNV>0.5). For example, the mutation NC_000006.11:g.4059022A>G disrupts PUM2 binding in the human NM_003913.52820-3021 region, with a binding score of 0.971 before the mutation decreasing to 0.073 afterward. In the mouse homologous region ENSMUST00000220965.2:1446A>G, ​​this mutation also disrupts PUM2 binding, with a score of 0.836 before the mutation decreasing to 0.031 afterward.

[0079] 4. Experimentally verify the effect of SNV on RBP-RNA interaction. This embodiment verifies the design of in vitro and in vivo experiments. The in vitro experiment examines whether RNA can bind to RBP to determine the amount of RBP. The in vivo experiment uses an in vivo CLIP assay to check whether RBP-bound RNA sequences show enrichment. Both in vitro and in vivo experiments are commonly used in this field, such as POND-seq and pull-down assays in existing technologies; the specific steps are not detailed here.

[0080] This embodiment selected three SNV sites that simultaneously affect PUM2-RNA interactions in both humans and mice. According to the predictions of the MuSIC model, the binding scores of the SNV sequences before and after mutation showed significant changes in both humans and mice.

[0081] This embodiment verified the binding of the SNV sequence to PUM2 using an in vitro pull-down experiment. The experimental results were consistent with the in vivo data, showing that the RNA sequence before SNV could bind to PUM2, while the binding of the sequence after SNV to PUM2 was significantly reduced.

[0082] This embodiment used POND-seq experiments to verify the enrichment of PUM2-binding RNA fragments in vivo. The results showed that the RNA sequence before SNV had a strong binding affinity to PUM2, while the binding affinity after SNV was significantly reduced. Statistical analysis indicated that the difference in SNV enrichment before and after the mutation was statistically significant.

[0083] These results validate the accuracy of the MuSIC model, demonstrating its consistent ability to predict the impact of SNVs on PUM2-RNA interactions, and its cross-species predictive power in both in vivo and in vitro experiments. These validation results further confirm the accuracy of the MuSIC model in detecting the effects of SNVs on RBP-RNA interactions, indicating its broad application potential in cross-species RBP-RNA interaction studies.

[0084] 5. Using model animals to elucidate the mechanisms of human diseases across species One of the core advantages of the MuSIC model lies in its powerful cross-species predictive ability, which can transcend species boundaries and effectively extrapolate findings from model organisms (such as mice) to the analysis of human disease mechanisms. This section uses mice as a research example to explore the interaction between RNA-binding proteins (RBPs) and RNA through the MuSIC model. This not only validates the model's feasibility in cross-species research but also reveals its significant value in the analysis of human disease mechanisms, particularly neurodegenerative diseases and cancer. Some results are consistent with previous studies, further corroborating the reliability of cross-species research.

[0085] 5.1 Effects of TARDBP-RNA binding: TARDBP (TDP-43) plays a crucial role in neurodegenerative diseases, particularly amyotrophic lateral sclerosis (ALS) and Alzheimer's disease (AD). The MuSIC model, with its ability to predict RBP-RNA interactions across species, provides a powerful tool for inferring human disease mechanisms from mouse studies. Sections 2-4 demonstrate the accuracy of MuSIC's cross-species analysis of the impact of SNVs on RBP-RNA.

[0086] MuSIC analysis revealed that the binding of TARDBP to ubiquitinated transcripts in mice is significantly affected by single nucleotide variants (SNVs). These variants alter the binding properties of TARDBP to its RNA substrates, leading to dysfunction. Specifically, multiple SNV sites were identified as altered in key regions of TARDBP binding. These changes may lead to misrecognition of RNA by TARDBP. This mechanism is highly consistent with cellular dysfunction observed in human neurodegenerative diseases, demonstrating the consistency of cross-species research.

[0087] Further analysis revealed that MuSIC predicted 573 genes in mice whose binding to TARDBP was influenced by SNVs. These genes were highly correlated with neural development, ubiquitination, and deubiquitination functions (FDR: 8.9e-06, 0.0123, 0.0253). These genes also play important roles in protein metabolism (FDR: 5.5e-04). These pathways have also been confirmed as core mechanisms in human neurodegenerative diseases and cancers, indicating that the TARDBP-RNA interaction patterns discovered in mouse models can be extrapolated to the analysis of human diseases, providing important insights into the pathogenic mechanisms of SNVs in human diseases.

[0088] 5.2 Cross-species analysis of the association between protein metabolism and cancer: The ubiquitin-proteasome system (UPS) is a core system for intracellular protein degradation, and its dysfunction is closely related to human cancers and neurodegenerative diseases. The MuSIC model, by analyzing the effects of SNVs on RBP-RNA interactions in mice, further validates the effectiveness of cross-species studies in elucidating the association between protein metabolism and disease.

[0089] The study found that 11.3% of RBPs in mice were strongly correlated with UPS (Proteinized Protein Blends), and 42.5% of these RBPs had been previously confirmed to be highly correlated with human cancer and neurodegenerative diseases. This result is consistent with existing research and directly proves the reliability of the MuSIC model's cross-species prediction. These UPS-related RBPs not only participate in protein metabolism, degradation, and homeostasis regulation in mouse cells, but their dysfunction patterns can also be mapped across species to human diseases: in cancer, they affect disease progression by regulating tumor cell proliferation, invasion, and metastasis; in neurodegenerative diseases such as Alzheimer's disease (AD), Huntington's disease (HD), and amyotrophic lateral sclerosis (ALS), their abnormal expression is closely related to protein folding disorders, aggregate formation, and cellular stress responses, which highly coincides with the pathological characteristics of human diseases.

[0090] KEGG enrichment analysis further confirmed that the pathways involved by these mouse UPS-related RBPs are also core regulatory pathways in human cancers and neurodegenerative diseases. This finding indicates that the functional patterns of RBPs identified in mice using the MuSIC model can be applied across species to elucidate human disease mechanisms, bridging the gap between model animal research and the development of therapeutic targets for human diseases.

[0091] In summary, the MuSIC model, using mice as a model organism, not only validates its feasibility in cross-species research by analyzing RBP-RNA interactions across species, but also effectively translates findings from model organism studies into key clues for elucidating the mechanisms of human diseases. This cross-species research strategy provides a novel perspective for a deeper understanding of the nature of diseases and the development of potential therapeutic targets, with significant application value, particularly in the fields of cancer and neurodegenerative diseases.

[0092] The steps and methods involved in Examples 2 and 3 above correspond to those in Example 1. For specific implementation details, please refer to the relevant description section of Example 1.

[0093] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)). The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting cross-species RBP-RNA interactions, characterized in that, Includes the following steps: Obtain the primary sequence of the RNA to be predicted, and perform structural prediction on the primary sequence to obtain the secondary structure sequence of the RNA; An interaction prediction model was constructed and trained using a dataset consisting of CLIP data from multiple species. The trained interaction prediction model was used to predict the secondary structure sequence and primary sequence of RNA to obtain the binding status of RNA-binding proteins with RNA.

2. The method for predicting cross-species RBP-RNA interactions as described in claim 1, characterized in that, RNAfold is used to predict the primary sequence of RNA to obtain the secondary structure sequence of RNA. The secondary structure sequence of RNA and the primary sequence of RNA are then one-hot encoded and used as input for the interaction prediction model.

3. The method for predicting cross-species RBP-RNA interactions as described in claim 1, characterized in that, The specific steps for training the interaction prediction model using a dataset composed of CLIP data from multiple species are as follows: A dataset was constructed using CLIP data from multiple species that were labeled. The labels are smoothed and regularized based on the RBP similarity between species. The dataset is divided into a training set and a validation set; The interaction prediction model is trained using the training set and its hyperparameters are fine-tuned using the validation set.

4. The method for predicting cross-species RBP-RNA interactions as described in claim 3, characterized in that, The multi-species CLIP data includes human RNA primary sequence data, mouse RNA primary sequence data, and zebrafish RNA primary sequence data.

5. The method for predicting cross-species RBP-RNA interactions as described in claim 4, characterized in that, When the dataset is divided into training and validation sets, human RNA primary sequence data is used as the training set, and mouse RNA primary sequence data and zebrafish RNA primary sequence data are used as the validation set.

6. The method for predicting cross-species RBP-RNA interactions as described in claim 3, characterized in that, The specific steps for smoothing and regularizing tags based on inter-species RBP similarity are as follows: The hard tags of the RNA primary sequence are smoothed and regularized to be converted into soft tags weighted by RBP similarity. The cross-entropy loss of hard labels and soft labels is calculated separately, and an adaptive gradient weight mechanism is adopted to balance the contribution of the cross-entropy loss of hard labels and soft labels to the total loss.

7. The method for predicting cross-species RBP-RNA interactions as described in claim 1, characterized in that, When training the interaction prediction model using the training set, convolutional layers and residual blocks are used to capture key features related to RBP combination, and the SE module is used to adaptively adjust the channel feature weights.

8. The method for predicting cross-species RBP-RNA interactions as described in claim 1, characterized in that, During the training of the interaction prediction model, Dropout and early stopping strategies are used to prevent overfitting, and smoothing gradients and guided backpropagation techniques are used to visualize important feature regions, further helping to identify high-response regions associated with RBP binding.

9. A cross-species RBP-RNA interaction prediction system, characterized in that, include: The data acquisition module is configured to acquire the primary sequence of the RNA to be predicted, and to perform structural prediction on the primary sequence of the RNA to obtain the secondary structure sequence of the RNA. The model training module is configured to build an interaction prediction model and train the interaction prediction model using a dataset consisting of CLIP data from multiple species. The interaction prediction module is configured to use the trained interaction prediction model to predict the secondary structure sequence and primary sequence of RNA, thereby obtaining the binding status of RNA-binding proteins with RNA.

10. The application of the cross-species RBP-RNA interaction prediction system of claim 9 in predicting single nucleotide variations in mice.

Images