A method and system for ncRNA-protein interaction prediction

Abstract

This invention belongs to the field of RNA drug target prediction technology, specifically relating to a method and system for predicting ncRNA-protein interactions. The method includes: acquiring the k-mer sequence encoding vectors of the ncRNA and protein to be tested, inputting them into an RPI prediction model to obtain RPI prediction results; the RPI prediction model includes a bidirectional dynamic interaction module; the bidirectional dynamic interaction module consists of cascaded interaction layers, including ncRNA-protein and protein-ncRNA paths; both paths include an improved cross-attention layer; the improvement lies in: pooling the query vector and key vector of the cross-attention in the sequence dimension, and determining the temperature parameter of the cross-attention based on the similarity between the two pooled vectors; the higher the similarity, the larger the temperature parameter. This invention solves the technical problem that existing cross-modal attention mechanisms, using a fixed-temperature attention distribution, are too rigid and cannot dynamically adjust their receptive field and focusing degree according to the specific interaction context.

Description

Technical Field

[0001] This invention belongs to the field of RNA drug target prediction technology, specifically relating to a method and system for predicting ncRNA-protein interactions. Background Technology

[0002] RNA-protein interaction (RPI) plays a central role in biological processes such as post-transcriptional regulation and RNA metabolism. Due to the high cost and limited throughput of traditional experimental validation techniques, computational prediction methods have become an important supplementary means for exploring large-scale RPIs.

[0003] Currently, deep learning-based representation learning methods have made significant progress in predicting molecular interactions, but they still suffer from the following technical limitations: First, existing feature extractors (such as convolutional neural networks (CNNs) mainly focus on local structural motifs, and due to their fixed receptive fields, they struggle to effectively capture long-range dependencies between sequences. While graph neural network (GNN) methods introduce topological information, their performance is highly dependent on the quality of graph construction, and their ability to simulate the dynamic evolution of biological sequences is limited. Second, existing cross-modal attention mechanisms generally employ fixed temperature scaling factors. However, the binding modes between biomolecules are highly heterogeneous (e.g., there is both broad-spectrum domain binding and precise residue site binding), and a fixed-temperature attention distribution is too rigid, failing to dynamically adjust its receptive field and focus based on the specific interaction context.

[0004] Furthermore, mainstream multi-view fusion strategies often employ simple feature concatenation or fixed weighted summation, lacking a flexible adaptive mechanism to dynamically coordinate local structural cues with global contextual information. Therefore, constructing a prediction framework that can effectively simulate dynamic features of a sequence and adaptively adjust feature alignment and fusion methods based on the interaction context is a pressing technical challenge in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting ncRNA-protein interactions, which solves the technical problem that the existing cross-modal attention mechanism, which uses a fixed-temperature attention distribution, is too rigid and cannot dynamically adjust its receptive field and focusing degree according to the specific interaction context.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a method for predicting ncRNA-protein interactions, the method comprising:

[0007] S1. Obtain the k-mer sequence coding vectors of the ncRNA and protein to be tested;

[0008] S2. Input the k-mer sequence encoding vectors of ncRNA and protein into the pre-trained RPI prediction model to obtain the RPI prediction results;

[0009] The RPI prediction model includes a two-way dynamic interaction module and an output prediction layer;

[0010] The bidirectional dynamic interaction module consists of cascaded interaction layers, which include an ncRNA-protein pathway with ncRNA characterization as input and protein feature expression as output, and a protein-ncRNA pathway with protein characterization as input and ncRNA feature expression as output. Both pathways include an improved cross-attention layer with the input of the other pathway as the target sequence to extract deep binding features of ncRNA and protein.

[0011] The improvement of the cross-attention layer lies in the fact that the query vector and key vector of the cross-attention are pooled in the sequence dimension, and the temperature parameter of the cross-attention is determined according to the similarity between the two pooled vectors; the higher the similarity, the larger the temperature parameter.

[0012] The output prediction layer outputs the final prediction result based on the output features of the bidirectional dynamic interaction module.

[0013] In one possible implementation, the temperature parameter Determined according to the following formula:

[0014]

[0015] in, This represents the similarity between the pooled query vector and the key vector. This represents the Sigmoid activation function; Represents a learnable scalar bias term; The dimension of the query vector; It is a numerically stable term.

[0016] In one possible implementation, the process of obtaining the protein feature expression based on the output features of the improved cross-attention layer includes: the protein attention features output by the improved cross-attention layer are linearly mapped and activated by an activation function to obtain corresponding projection features; these projection features are nonlinearly transformed and then weighted and added to the protein hidden state of the previous interaction layer using a learnable weight to obtain the candidate protein hidden state of the interaction layer; then, the candidate protein hidden state and the protein hidden state of the previous interaction layer are fused using a protein dynamic gating factor and a decay coefficient to obtain the protein hidden state of the interaction layer; the projected features and the protein hidden state of the interaction layer are concatenated and linearly mapped to obtain the protein feature expression of the interaction layer.

[0017] The protein dynamic gating factor is obtained as follows: the protein gating feature, protein attention feature, and the protein hidden state of the previous interaction layer are concatenated and then activated by a linear mapping and activation function to obtain the protein dynamic gating factor; the protein gating feature is obtained by fusing the protein attention feature and the protein representation as input to the interaction layer through directional gating fusion weights; the directional gating fusion weights are obtained by concatenating the protein attention feature and the protein representation and then activating by a linear mapping and activation function.

[0018] The process of obtaining the ncRNA feature expression based on the output features of the improved cross-attention layer includes: the ncRNA attention features output by the improved cross-attention layer are linearly mapped and activated by an activation function to obtain the corresponding projection features; the projection features are nonlinearly transformed and then weighted and added with the ncRNA hidden state of the previous interaction layer through a learnable weight to obtain the candidate ncRNA hidden state of the interaction layer; then, the candidate ncRNA hidden state and the ncRNA hidden state of the previous interaction layer are fused through an ncRNA dynamic gating factor and a decay coefficient to obtain the ncRNA hidden state of the interaction layer; the projection features and the ncRNA hidden state of the interaction layer are concatenated and linearly mapped to obtain the ncRNA feature expression of the interaction layer.

[0019] The ncRNA dynamic gating factor is obtained as follows: the ncRNA gating feature, the ncRNA attention feature, and the ncRNA hidden state of the previous interaction layer are concatenated and then activated by a linear mapping and an activation function to obtain the ncRNA dynamic gating factor; the ncRNA gating feature is obtained by fusing the ncRNA attention feature and the ncRNA representation as input to the interaction layer through a direction-gating fusion weight; the ncRNA attention feature and the ncRNA representation are concatenated and then activated by a linear mapping and an activation function to obtain the direction-gating fusion weight.

[0020] In one possible implementation, the RPI prediction model further includes a protein feature enhancement layer for feature enhancement of the protein's k-mer sequence encoding vector; the protein representation of the first interaction layer in the bidirectional dynamic interaction module is the output feature of the protein feature enhancement layer.

[0021] In one possible implementation, the protein feature enhancement layer includes a local convolutional path, a global LNN path, and a gated fusion layer;

[0022] The local convolution path is used to extract local features from the k-mer sequence encoding vector of a protein through a cascaded feature extraction module; the feature extraction module includes a one-dimensional convolutional layer, a batch normalization, and an activation function connected in sequence.

[0023] The global LNN path is used to extract global features from the k-mer sequence encoding vector of a protein through sequentially connected linear layers and LNN layers;

[0024] The gated fusion layer is used to weight and sum the local and global features using a gated weight factor to obtain the enhanced features, which are then used as the output features of the protein feature enhancement layer. The gated weight factor is obtained by concatenating the local and global features and then activating them using a linear mapping and activation function.

[0025] In one possible implementation, the output prediction layer includes a feature reconstruction layer, a feature fusion layer, and a fully connected prediction layer;

[0026] The feature reconstruction layer reconstructs the input by passing the k-mer sequence encoding vectors of ncRNA and protein through a multi-head attention mechanism and a feedforward neural network;

[0027] The feature fusion layer is used to splice the output features of the feature reconstruction layer, the k-mer sequence encoding vector of ncRNA, the k-mer sequence encoding vector of protein, and the ncRNA and protein features output by the bidirectional dynamic interaction module in the channel dimension.

[0028] The fully connected prediction layer is used to output RPI prediction results through a fully connected network based on the features concatenated by the feature fusion layer.

[0029] In one possible implementation, the process of training the RPI prediction model includes:

[0030] (1) Remove the bidirectional dynamic interaction module and retain the simplified network architecture that splices the k-mer sequence encoding vector of ncRNA and the output features of the protein feature enhancement layer and then directly feeds them into the output prediction layer; pre-train the simplified network architecture.

[0031] (2) Transfer the pre-trained network parameters to the complete RPI prediction model and randomly initialize the bidirectional dynamic interaction module in the RPI prediction model; then fine-tune the entire model to obtain the trained RPI prediction model.

[0032] In one possible implementation, the RPI prediction model is trained using a loss function that includes binary cross-entropy loss.

[0033] In one possible implementation, the k-mer sequence encoding vector of the ncRNA is a 4-mer vector; the k-mer sequence encoding vector of the protein is a 2-mer vector.

[0034] To address the aforementioned technical problems, a second aspect of the present invention provides an ncRNA-protein interaction prediction system, comprising a processor for executing a computer program to implement the steps of the method in any possible implementation of the first aspect of the present invention:

[0035] S1. Obtain the k-mer sequence coding vectors of the ncRNA and protein to be tested;

[0036] S2. Input the k-mer sequence encoding vectors of ncRNA and protein into the pre-trained RPI prediction model to obtain the RPI prediction results;

[0037] The RPI prediction model includes a two-way dynamic interaction module and an output prediction layer;

[0038] The bidirectional dynamic interaction module consists of cascaded interaction layers, which include an ncRNA-protein pathway with ncRNA characterization as input and protein feature expression as output, and a protein-ncRNA pathway with protein characterization as input and ncRNA feature expression as output. Both pathways include an improved cross-attention layer with the input of the other pathway as the target sequence to extract deep binding features of ncRNA and protein.

[0039] The improvement of the cross-attention layer lies in the fact that the query vector and key vector of the cross-attention are pooled in the sequence dimension, and the temperature parameter of the cross-attention is determined according to the similarity between the two pooled vectors; the higher the similarity, the larger the temperature parameter.

[0040] The output prediction layer outputs the final prediction result based on the output features of the bidirectional dynamic interaction module.

[0041] In one possible implementation, the temperature parameter Determined according to the following formula:

[0042]

[0043] in, This represents the similarity between the pooled query vector and the key vector. This represents the Sigmoid activation function; Represents a learnable scalar bias term; The dimension of the query vector; It is a numerically stable term.

[0044] In one possible implementation, the process of obtaining the protein feature expression based on the output features of the improved cross-attention layer includes: the protein attention features output by the improved cross-attention layer are linearly mapped and activated by an activation function to obtain corresponding projection features; these projection features are nonlinearly transformed and then weighted and added to the protein hidden state of the previous interaction layer using a learnable weight to obtain the candidate protein hidden state of the interaction layer; then, the candidate protein hidden state and the protein hidden state of the previous interaction layer are fused using a protein dynamic gating factor and a decay coefficient to obtain the protein hidden state of the interaction layer; the projected features and the protein hidden state of the interaction layer are concatenated and linearly mapped to obtain the protein feature expression of the interaction layer.

[0045] The protein dynamic gating factor is obtained as follows: the protein gating feature, protein attention feature, and the protein hidden state of the previous interaction layer are concatenated and then activated by a linear mapping and activation function to obtain the protein dynamic gating factor; the protein gating feature is obtained by fusing the protein attention feature and the protein representation as input to the interaction layer through directional gating fusion weights; the directional gating fusion weights are obtained by concatenating the protein attention feature and the protein representation and then activating by a linear mapping and activation function.

[0046] The process of obtaining the ncRNA feature expression based on the output features of the improved cross-attention layer includes: the ncRNA attention features output by the improved cross-attention layer are linearly mapped and activated by an activation function to obtain the corresponding projection features; the projection features are nonlinearly transformed and then weighted and added with the ncRNA hidden state of the previous interaction layer through a learnable weight to obtain the candidate ncRNA hidden state of the interaction layer; then, the candidate ncRNA hidden state and the ncRNA hidden state of the previous interaction layer are fused through an ncRNA dynamic gating factor and a decay coefficient to obtain the ncRNA hidden state of the interaction layer; the projection features and the ncRNA hidden state of the interaction layer are concatenated and linearly mapped to obtain the ncRNA feature expression of the interaction layer.

[0047] The ncRNA dynamic gating factor is obtained as follows: the ncRNA gating feature, the ncRNA attention feature, and the ncRNA hidden state of the previous interaction layer are concatenated and then activated by a linear mapping and an activation function to obtain the ncRNA dynamic gating factor; the ncRNA gating feature is obtained by fusing the ncRNA attention feature and the ncRNA representation as input to the interaction layer through a direction-gating fusion weight; the ncRNA attention feature and the ncRNA representation are concatenated and then activated by a linear mapping and an activation function to obtain the direction-gating fusion weight.

[0048] In one possible implementation, the RPI prediction model further includes a protein feature enhancement layer for feature enhancement of the protein's k-mer sequence encoding vector; the protein representation of the first interaction layer in the bidirectional dynamic interaction module is the output feature of the protein feature enhancement layer.

[0049] In one possible implementation, the protein feature enhancement layer includes a local convolutional path, a global LNN path, and a gated fusion layer;

[0050] The local convolution path is used to extract local features from the k-mer sequence encoding vector of a protein through a cascaded feature extraction module; the feature extraction module includes a one-dimensional convolutional layer, a batch normalization, and an activation function connected in sequence.

[0051] The global LNN path is used to extract global features from the k-mer sequence encoding vector of a protein through sequentially connected linear layers and LNN layers;

[0052] The gated fusion layer is used to weight and sum the local and global features using a gated weight factor to obtain the enhanced features, which are then used as the output features of the protein feature enhancement layer. The gated weight factor is obtained by concatenating the local and global features and then activating them using a linear mapping and activation function.

[0053] In one possible implementation, the output prediction layer includes a feature reconstruction layer, a feature fusion layer, and a fully connected prediction layer;

[0054] The feature reconstruction layer reconstructs the input by passing the k-mer sequence encoding vectors of ncRNA and protein through a multi-head attention mechanism and a feedforward neural network;

[0055] The feature fusion layer is used to splice the output features of the feature reconstruction layer, the k-mer sequence encoding vector of ncRNA, the k-mer sequence encoding vector of protein, and the ncRNA and protein features output by the bidirectional dynamic interaction module in the channel dimension.

[0056] The fully connected prediction layer is used to output RPI prediction results through a fully connected network based on the features concatenated by the feature fusion layer.

[0057] In one possible implementation, the process of training the RPI prediction model includes:

[0058] (1) Remove the bidirectional dynamic interaction module and retain the simplified network architecture that splices the k-mer sequence encoding vector of ncRNA and the output features of the protein feature enhancement layer and then directly feeds them into the output prediction layer; pre-train the simplified network architecture.

[0059] (2) Transfer the pre-trained network parameters to the complete RPI prediction model and randomly initialize the bidirectional dynamic interaction module in the RPI prediction model; then fine-tune the entire model to obtain the trained RPI prediction model.

[0060] In one possible implementation, the RPI prediction model is trained using a loss function that includes binary cross-entropy loss.

[0061] In one possible implementation, the k-mer sequence encoding vector of the ncRNA is a 4-mer vector; the k-mer sequence encoding vector of the protein is a 2-mer vector.

[0062] The beneficial effects of this invention are as follows: By introducing a bidirectional dynamic adaptive temperature cross-attention mechanism, the smoothness of the attention distribution is adjusted in real time based on the query-key similarity through an adaptive temperature branch, enabling the model to adaptively focus on broad-spectrum region binding or precise site binding modes. Experimental results show that this method has good prediction performance on multiple benchmark datasets, significantly improving the accuracy and cross-species generalization performance of non-coding RNA-protein interaction recognition, and providing a more stable and accurate computational means for studying complex intermolecular interaction mechanisms. This invention solves the technical problem that existing cross-modal attention mechanisms using fixed-temperature attention distribution are too rigid and cannot dynamically adjust their receptive field and focusing degree according to the specific interaction context. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the ncRNA-protein interaction prediction process in an embodiment of the ncRNA-protein interaction prediction method of the present invention.

[0064] Figure 2 This is a schematic diagram of protein feature enhancement in an embodiment of the ncRNA-protein interaction prediction method of the present invention.

[0065] Figure 3 This is a schematic diagram of the bidirectional dynamic adaptive temperature cross-attention mechanism in the implementation of the ncRNA-protein interaction prediction method of the present invention.

[0066] Figure 4 This is a schematic diagram of the system architecture in an embodiment of the ncRNA-protein interaction prediction system of the present invention. Detailed Implementation

[0067] This invention introduces a bidirectional dynamic adaptive temperature cross-attention mechanism. By adjusting the smoothness of the attention distribution in real time based on query-key similarity through an adaptive temperature branch, the model can adaptively focus on broad-spectrum binding or precise site binding modes. Experimental results show that this method has good prediction performance on multiple benchmark datasets, significantly improving the accuracy and cross-species generalization performance of non-coding RNA-protein interaction recognition, and providing a more stable and accurate computational tool for studying complex intermolecular interaction mechanisms. This invention solves the technical problem of existing cross-modal attention mechanisms using a fixed-temperature attention distribution that is too rigid and unable to dynamically adjust its receptive field and focusing degree according to the specific interaction context.

[0068] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

[0069] Implementation of methods for predicting ncRNA-protein interactions:

[0070] In recent years, with the development of deep learning technology in the field of bioinformatics, many identification methods have emerged that use neural networks to predict the interaction between non-coding RNA and proteins. These predictive models typically consist of two parts: feature extraction and interaction fusion. They utilize convolutional layers or attention mechanisms to capture molecular features and estimate the probability of non-coding RNA binding to proteins. By designing different network architectures and loss functions and training and optimizing using known biological datasets, better prediction results can be achieved. This modeling process relies on mining the deep patterns of biological sequences. However, existing static feature encoding and fixed-temperature attention mechanisms have limitations in handling biological interactions with dynamic characteristics and heterogeneity. They struggle to accurately capture long-range dynamic dependencies and context-dependent key information, resulting in a need for further improvement in the generalization ability and robustness of models in complex samples or cross-species tasks.

[0071] This invention proposes an ncRNA-protein interaction prediction method based on a liquid neural network and dynamic adaptive cross-attention. By constructing a dual-path dynamic feature processor and introducing a dynamic adaptive temperature cross-attention mechanism, it better captures complex multi-scale interaction patterns between molecules, further improving prediction accuracy and cross-species generalization ability. The main components include:

[0072] A. Extracting Dual-Path Dynamic Features: Obtain the ncRNA and protein sequences to be predicted and perform embedding encoding. Input the protein features into the dual-path processor, and simultaneously extract local convolutional motifs and global dynamic features based on liquid neural networks (LNN) and perform gated fusion.

[0073] B. Perform dynamic adaptive cross-modal interaction: Input the initial features of ncRNA and the fused protein features into the interaction module, and dynamically calculate and adjust the temperature parameter of cross attention based on the semantic similarity between the query vector and the key vector.

[0074] C. Cross-layer state update based on LNN: The update mechanism of LNN is used to iteratively process cross-modal interaction features, decouple the intra-layer output and inter-layer memory, and maintain the dynamic hidden state in the depth dimension of the interaction layer.

[0075] D. Multi-view feature fusion and test prediction: The deep fusion features after multiple iterations are spliced ​​and fused with the original sequence auxiliary features, input into the prediction network to obtain the interaction probability score, and the effect of the prediction model is evaluated.

[0076] Specifically, such as Figure 1 As shown, the ncRNA-protein interaction prediction method of this embodiment includes the following steps:

[0077] 1. A dual-path dynamic feature processing module based on LNN.

[0078] To obtain richer protein feature representations, such as Figure 2 As shown, this invention constructs a dual-path dynamic feature processor. This processor assigns the input protein sequence to two parallel branches (a local convolutional path and a global LNN path) for processing, enabling the model to capture fine-grained local patterns while also taking into account broader sequence-level global dependencies.

[0079] In this embodiment, both ncRNA and protein are k-mer sequence coding vectors (i.e., the 2-mer vector of protein and the 4-mer vector of ncRNA).

[0080] Local Convolution Path: In this embodiment, the local convolution path consists of two consecutive feature extraction modules (i.e., sequentially containing a one-dimensional convolutional layer Conv, a batch normalization layer BN, and a linear activation function ReLU), with a kernel size of 3. This path utilizes a window mechanism that slides along the amino acid sequence, specifically designed to detect short-range contextual features and local structural motifs generated by adjacent residues in the protein sequence. Features extracted through this path are denoted as local features. .

[0081] Global LNN Path: Unlike local convolutional extraction, the global LNN path aims to capture long-range temporal or sequential dependencies that cannot be effectively handled by relying solely on local convolutions. Specifically, the input features are first projected through a linear layer and then fed into a one-dimensional liquid neural network layer.

[0082] The LNN (Liquid Neural Network) architecture is based on the continuous-time dynamics of leaky integral discharge neurons, and the continuous evolution of its hidden states is described by the following ordinary differential equation:

[0083]

[0084] in, This indicates the time step of a neuron in a liquid neural network. The corresponding continuous hidden states are used to describe the network's internal representation of the dynamic changes in the input sequence; Indicates time step The input protein sequence feature vector (obtained from protein 2-mer encoding); It represents the time constant, used to control the degree of decay of the neuron's state over time; This represents the input weight matrix, which adjusts the input signal and is used to regulate the influence of input features on neuron state updates. The circular weight matrix represents the capture of state transition patterns and is used to characterize the information transmission relationship between the hidden states of neurons; Represents a non-linear activation function; This represents the set of parameters that need to be trained during the nonlinear transformation process, including relevant weight parameters and bias parameters.

[0085] In practical model deployment, to improve computational stability and training efficiency, this invention employs discretization and gated approximation to achieve the aforementioned continuous dynamic process. By endowing neurons with input-dependent dynamic characteristics, the LNN path can dynamically adjust memory retention and leakage based on the interaction context, thereby extracting global dynamic features. .

[0086] The parameter configuration of the dual-path dynamic feature processing module is shown in Table 1:

[0087] Table 1. Detailed parameter configuration table for the dual-path protein feature extraction network.

[0088]

[0089] In Table 1 above, the format of the local convolution path parameters is ( ) represent the number of input channels, kernel size, and output channels of a one-dimensional convolutional layer, respectively; for a global LNN path, the parameter format is ( () represents the input and output feature dimensions of a linear layer or a liquid neural network layer. In this example, the initial dense feature dimension of the input protein is 400.

[0090] Adaptive Gated Fusion Mechanism: To perfectly combine local details with global dynamic features, this invention introduces a learnable adaptive gated network that dynamically calculates the relative importance weights of local and global features for each independent sample. The mathematical expression of its fusion mechanism is as follows:

[0091] First, calculate the gating weight factor. :

[0092]

[0093] Then, dynamic feature fusion is performed to obtain an enhanced protein representation. :

[0094]

[0095] in,[ ] indicates local features With global features The components are stitched together along the channel dimension. This represents the weight matrix of the gated network; This represents the Sigmoid activation function, used to map the output to the range of 0 to 1; This refers to the generated adaptive fusion weights.

[0096] The advantage of this module lies in its gating mechanism, which endows the model with extremely strong biological adaptability. When faced with interaction samples dominated by long-range dynamics, the model will adaptively... Approaching 1, thus amplifying the global features. The dominant position; when faced with interaction samples dominated by local key motifs, the model will... Approaching 0, thus highlighting local features. This design significantly enhances the model's adaptability to multi-scale biological sequence distributions.

[0097] 2. Two-way dynamic adaptive temperature cross-modal interaction module.

[0098] After obtaining the enhanced feature representations of ncRNA and protein respectively, as follows Figure 3 As shown, this invention introduces a bidirectional dynamic adaptive temperature cross-modal interaction module. This module comprises two stacked interaction layers. In each layer, information exchange occurs in parallel along two directions: ncRNA to protein and protein to ncRNA. Specifically, ncRNA→protein represents using ncRNA features to guide protein feature updates, and protein→ncRNA represents using protein features to guide ncRNA feature updates.

[0099] In the "ncRNA→protein" interaction path, the initial ncRNA representation is input into the query mapping layer. A query vector Query is obtained through a linear transformation using a trainable query weight matrix. This Query is then used to interact with the features obtained from the initial protein representation after key-value mapping, thereby generating interactively enhanced ncRNA features. Similarly, in the "protein→ncRNA" interaction path, the initial protein representation is processed through the query mapping layer to obtain the Query, which is then used to interact with the features obtained from the initial ncRNA representation after key-value mapping, generating interactively enhanced protein features.

[0100] (1) Cross-attention layer. The interaction layer described above in this embodiment adopts a bidirectional symmetric architecture. In the "ncRNA→protein" interaction path, the initial ncRNA representation is input into the query mapping layer, and a query vector Query is obtained through a linear transformation using a trainable query weight matrix. This Query is then used to interact with the features obtained from the initial protein representation after key and value mapping, thereby generating interactively enhanced ncRNA features. Similarly, in the "protein→ncRNA" interaction path, the initial protein representation is processed through the query mapping layer to obtain Query, which is then used to interact with the features obtained from the initial ncRNA representation after key and value mapping, generating interactively enhanced protein features. This design ensures that the two molecular modalities achieve sufficient information penetration and feature alignment in deep space.

[0101] In traditional scaled dot product attention mechanisms, the attention distribution is typically calculated as follows:

[0102]

[0103] Among them, temperature parameter It is usually set to a fixed constant ( Take vector (Dimensions). However, the binding modes between biomolecules are highly diverse: some belong to broad-spectrum domain-region level binding, while others belong to highly specific residue-nucleotide level precise site binding. Fixed temperature parameters cannot accommodate this heterogeneous binding characteristics.

[0104] To overcome this limitation, combined with Figure 3 As shown in the intermediate calculation process, this invention innovatively designs an Adaptive Temperature Branch (ATB) in the cross-modal interaction module for dynamically estimating and adjusting temperature parameters. The core idea is to dynamically adjust the smoothness or sharpness of the Softmax distribution based on the semantic similarity between the current query vector and the key / value vector. The specific calculation steps are as follows:

[0105] 1) Calculate the global description vector: First, calculate the query vector separately. and key vector Perform average pooling on the sequence dimension and extract their respective global description vectors, denoted as follows: and Query vector and key vector Both are used to calculate the attention distribution query vector. and key vector .

[0106] 2) Calculate global semantic similarity: Subsequently, the similarity between the two global description vectors is calculated through inner product operation. :

[0107]

[0108] in, This represents the global semantic similarity score; The dimension of the feature vector; and These represent the first and second global description vectors, respectively. Elements of each dimension.

[0109] 3) Generate sample-specific dynamic temperature parameters The calculated similarity The input is fed into a monotonic mapping function to generate adaptive temperature parameters. :

[0110]

[0111] in, This represents the Sigmoid activation function; Represents a learnable scalar bias term; It serves as the basic scaling factor (maintaining the same dimensionality as traditional attention mechanisms); This is a numerically stable term used to prevent errors caused by division by zero in subsequent calculations.

[0112] 4) Perform dynamic adaptive temperature attention calculation: Finally, utilize the generated dynamic temperature parameters Replacing the traditional fixed parameters, the cross-attention distribution in each direction is recalculated (i.e., Figure 3 ncRNA and Protein characteristics):

[0113]

[0114] Each case is substituted according to the specific interaction pathways described above, namely "ncRNA→protein" or "protein→ncRNA". That is:

[0115]

[0116]

[0117] in, This represents the final output of the "protein → ncRNA" interaction pathway (i.e. Figure 3 ncRNA characteristics); The key vector Key represents the initial Protein representation mapped by the corresponding key matrix; it represents... This represents the value vector Value after the initial representation of Protein is mapped by the corresponding value matrix; The query vector Query represents the value query vector after the initial characterization of ncRNA has been mapped by the corresponding query matrix. Indicates based on query vector Sum value vector Obtain adaptive temperature parameters.

[0118] This represents the final output of the "ncRNA→protein" interaction pathway (i.e. Figure 3 Protein features in ( ). This represents the key vector Key, which is the result of mapping the initial characterization of ncRNA to the corresponding key matrix. Value represents the value vector after the initial characterization of ncRNA has been mapped by the corresponding value matrix; This represents the value query vector Query after the initial representation of Protein has been mapped by the corresponding query matrix; Indicates based on query vector Sum value vector Obtain adaptive temperature parameters.

[0119] The advantage of this module lies in its mechanism, which perfectly aligns with biological interaction logic. When a protein query feature exhibits broad high similarity to multiple fragments of ncRNA (corresponding to broad-spectrum domain binding), the ATB outputs a large τ value, generating a smooth attention distribution and encouraging the model to aggregate features from multiple relevant sites. Conversely, when a protein feature is only similar to a very few specific fragments of ncRNA (corresponding to highly specific residue binding), the ATB outputs a small τ value, generating a sharp attention distribution, thereby precisely focusing on key sites and suppressing background noise.

[0120] Directional gating and an LNN-based cross-layer iterative update module. After acquiring the cross-modal interaction features generated by dynamically adaptive temperature cross-attention, this implementation further designs a directional gating mechanism and an LNN-based cross-layer update module to capture the asymmetric dependencies between molecules and maintain dynamic memory in the network depth direction.

[0121] (2) Directional Gating: In biological systems, the regulatory relationship between non-coding RNA and proteins is usually asymmetric. To rationally control the information flow during cross-modal interactions, the model activates independent directional gating networks for different interaction directions ("ncRNA→protein" or "protein→ncRNA"). The calculation process is as follows:

[0122] First, calculate the directional gating fusion weights. :

[0123]

[0124] Subsequently, the adaptively fused gated features are calculated. :

[0125]

[0126] in, This represents the modal feature representation used as input to the query in the current interaction direction, i.e., the initial representation of the corresponding modal representation after processing by the feature enhancement module (Protein). This represents the cross-modal interaction enhancement features output by the adaptive temperature cross-attention module; This represents the learnable weight matrix corresponding to the interaction direction; This represents the Sigmoid activation function; This represents the fusion weighting factor that controls the proportion of cross-modal information inflow.

[0127] Taking the ncRNA→protein pathway as an example, the above formula and The process is as follows: using protein enhancement features as query features. (equivalent to) Figure 3 (Initial protein characterization), ncRNA-enhanced features are used as the retrieved cross-modal information source, and protein interaction enhancement features fused with ncRNA information are obtained through adaptive temperature cross-attention. (equivalent to) Figure 3 (Protein features in the text). Subsequently, according to formula (9), the gating weights are used. right and Weighted fusion is performed to obtain the final gating feature representation.

[0128] (3) LNN-based cross-layer update: Traditional models typically model temporal changes in the sequence dimension. However, to capture the feature evolution dependency across interaction layers, this invention introduces an LNN update mechanism at the end of each interaction module. The core of this mechanism lies in the clever decoupling of "intra-layer feature output" and "inter-layer memory state". The specific calculation and execution steps are as follows:

[0129] Feature projection: onto the original attention features Perform linear mapping (weight matrix is) ReLU activation yields projected features. :

[0130]

[0131] Dynamic gating calculation: combining current gating features Attention characteristics and the hidden state of the upper interaction layer Calculate the dynamic gating factor used to balance historical memory and the current state. :

[0132]

[0133] in, This represents the cross-modal enhancement features output by the cross-attention module in the current interaction layer. This represents the hidden state saved by the LNN update module of the previous interaction layer. Initial hidden state This represents the initial memory state of the LNN before the first interaction layer begins updating, used to provide initialization information for cross-layer state propagation. In this embodiment, it will... Initialize as zero vector .

[0134] Candidate State Update: Generating candidate hidden states using the unique "leakage-inertia" mechanism of LNNs. :

[0135]

[0136] in, This indicates the candidate hidden state that needs to be updated in the current interaction layer; This represents the hidden state output by the LNN module of the previous interaction layer; This represents the feature representation of the current interaction layer's interaction features after linear projection; This represents the candidate state generation function, used to perform nonlinear transformations on the projected features; This represents the learnable state update coefficient, used to adjust the fusion ratio between historical hidden states and current input information.

[0137] Inter-layer hidden state generation: based on dynamic gating factor With learnable attenuation coefficient The candidate hidden states generated by the current interaction layer Hidden historical state passed from the previous interaction layer Perform adaptive fusion to generate inter-layer hidden states for passing to the next interaction layer. The calculation process is as follows:

[0138]

[0139] in, This indicates the hidden state between layers after the current interaction layer has been updated, and is used to pass historical memory information to the next interaction layer; This represents the candidate hidden state generated from the current interaction features; This represents the historical hidden state output by the previous interaction layer; This represents the dynamic gating factor, used to control the fusion ratio of current candidate state information and historical state information; represents the learnable state decay coefficient, which is automatically updated through parameter optimization during model training to adjust the degree to which historical hidden states are preserved during cross-layer propagation; ⊙ represents element-wise multiplication.

[0140] Intra-layer output generation: Projecting features With the newly generated interlayer hidden state Perform concatenation (where "[;]" indicates feature concatenation operation), through an independent linear mapping matrix. Feature fusion is performed to obtain the final output representation of the current interaction layer. The calculation process is as follows:

[0141]

[0142] in, This represents the fused feature representation output by the current interaction layer, which is used as input for subsequent network modules; This represents hidden state information used for transmission across interaction layers; This represents the feature representation of the current layer's interactive features after projection.

[0143] By decoupling output and state, this module enables the model to achieve a "liquid"-like dynamic memory capability in the depth dimension. It is responsible for stably passing historical interaction context between layers, and This is responsible for combining the features of the current layer for immediate expression. This allows the model to capture extremely deep and dynamic cross-molecular dependencies after multiple rounds of bidirectional iteration.

[0144] Taking the ncRNA→protein pathway as an example, the protein feature expression output of the ncRNA→protein pathway in this interaction layer... (i.e., the above) The following formulas are used to obtain:

[0145]

[0146]

[0147]

[0148]

[0149]

[0150]

[0151]

[0152] in, The hidden state of the protein features output by this path in the interaction layer (equivalent to the above) and ); The interaction attention enhancement feature output by this path in the interaction layer (equivalent to the above) , equivalent to Figure 3 (Protein characteristics) for Corresponding projection features; Input features to the original proteins of this interaction layer (equivalent to the above) (i.e., the input characteristics of the protein→ncRNA pathway). Gated features (equivalent to the above) ).

[0153] 3. Multi-view feature fusion and prediction module.

[0154] After completing multi-scale feature extraction and cross-modal bidirectional interaction, the model generates deep representations of proteins and ncRNAs containing rich high-dimensional semantic information. To fully utilize these features and prevent the loss of underlying original sequence information, this invention designs a multi-view feature fusion strategy and a fully connected prediction network.

[0155] Feature Transformer: To further expand the feature representation space, this invention introduces a Feature Transformer module. This module concatenates the original k-mer sequence encoding vector (i.e., the 2-mer vector of protein and the 4-mer vector of ncRNA) as input. Subsequently, a multi-head attention mechanism and a feedforward neural network (FFN) are used to perform nonlinear mapping and reconstruction on the input, thereby unifying different types of original features into a more expressive representation space and uncovering nonlinear relationships that are difficult to reveal with single feature branches.

[0156] Multi-view fusion: During the fusion phase, the model concatenates the following five sets of feature vectors from different sources along the channel dimension to construct a unified multi-view fusion feature representation: (1) deeply fused protein features output by the cross-modal interaction module; (2) deeply fused ncRNA features output by the cross-modal interaction module; (3) the initial protein representation obtained after initial embedding of the coding layer, which represents the protein without cross-modal interaction and is used to preserve the original sequence coding information of the protein; (4) the initial ncRNA representation after initial embedding of the coding layer; and (5) the auxiliary features output after reconstruction by the Feature Transformer. This fusion feature integrates multi-level biological features from local sequence motifs and global dynamic modes to cross-modal interaction signals.

[0157] Fully Connected Prediction Network: The concatenated multi-view fusion features are then input into the fully connected prediction network to calculate the final interaction score. Specifically, the network sequentially comprises a first linear layer, a Dropout layer, and a second linear layer. The first linear layer, combined with the ReLU activation function, maps and reduces the dimensionality of the input high-dimensional fusion features to a 256-dimensional latent representation space. The dropout rate of the Dropout layer is set to 0.3 to suppress overfitting during training. The second linear layer further maps the 256-dimensional features into a one-dimensional scalar, which is then processed by the Sigmoid activation function to output a probability value in the [0,1] interval. This probability value represents the predicted interaction score between the ncRNA and the protein.

[0158] This module, through the aforementioned multi-view fusion and prediction architecture, fully leverages the deep cross-modal interaction features extracted by LNN and DACA while effectively preserving the basic feature information of the original sequence through direct concatenation. Combined with a prediction network featuring Dropout, the final decision-making process maintains good generalization ability and prediction robustness while integrating multi-scale features.

[0159] 4. Model training strategy and experimental verification.

[0160] Because this invention relates to complex dynamic differential equations and adaptive cross-modal attention, its parameter space is large. To avoid the model getting stuck in local optima in the early stages of training and to ensure that the underlying feature extractor is fully optimized, this invention adopts a two-stage (pre-training and fine-tuning) training strategy, combined with a binary cross-entropy loss function for parameter updates.

[0161] (1) Two-stage training strategy: Pre-training of the feature extraction network: First, a simplified network architecture is constructed, which retains the complete feature extraction backbone (i.e., the ncRNA sequence embedding layer and the protein dual-path dynamic feature processor), but removes the upper-layer cross-modal deep interaction module. In this simplified network, the ncRNA sequence embedding layer and the protein dual-path dynamic feature processor are retained, and training is performed using only the two basic modal feature representations. Specifically, the initially encoded ncRNA feature representation and the protein feature representation are concatenated and input into a lightweight prediction head for optimization training. The purpose of this stage is to force the underlying network to independently learn high-quality local motif and long-range dynamic representations.

[0162] (2) Fine-tuning of the complete model: After pre-training, the learned low-level weight parameters are loaded into the complete model of this invention. At this time, the upper-layer bidirectional dynamic adaptive temperature cross-modal interaction module and the multi-view fusion prediction network are randomly initialized. Then, the entire model is fine-tuned end-to-end. This strategy significantly improves the convergence speed and generalization ability of complex modules.

[0163] (3) Training Objectives and Optimization Parameters: During model training, binary cross-entropy is used as the loss function to quantify the difference between the predicted probability and the true label. The calculation formula is as follows:

[0164]

[0165] in, This represents the total loss value; This represents the total number of samples in each training batch; Indicates the first The true label of each sample is 1 if there is an interaction between ncRNA and protein, and 0 otherwise. The first output of the model represents the... The predicted probability of interaction between samples. During backpropagation, the Adam optimizer is used to iteratively update the learnable parameters in the network based on the calculated gradients until the validation set performance converges.

[0166] (4) To verify the effectiveness of the method of the present invention in the ncRNA-protein interaction prediction task, the 10-fold cross-validation method was used to evaluate the performance on four benchmark datasets (NPInter, RP1488, RP11807 and ZEA22133) and to compare it with a variety of existing ncRNA-protein interaction prediction methods.

[0167] The comparative methods employed are all existing predictive models, including traditional machine learning-based methods and deep learning-based methods. Each comparative model predicts the interaction between ncRNA and protein through different feature representation learning methods and classification prediction strategies. Specifically, machine learning-based methods typically construct a joint representation using RNA and protein sequence features and complete the prediction using a classifier; deep learning-based methods automatically learn local patterns and higher-order semantic features in the sequence through neural networks to predict potential interaction relationships. The method of this invention, however, achieves multi-source feature fusion and deep interaction relationship modeling through a liquid neural network dynamic feature processing module, a bidirectional dynamic adaptive temperature cross-modal interaction module, and a multi-view feature fusion module.

[0168] The evaluation metrics included the Area Under the Receiver Operating Characteristic Curve (AUC), the Area Under the Precision-Recall Curve (AUPR), Accuracy (ACC), Precision (PRE), Recall, Sensitivity (SEN), Specificity (SPE), Matthews Correlation Coefficient (MCC), and the F1 score. Specifically, AUC evaluates the model's ability to rank positive and negative samples overall; AUPR evaluates the model's ability to identify positive samples under class imbalance conditions; Accuracy represents the proportion of correctly predicted samples out of all samples; Precision represents the proportion of true positive samples among those predicted as positive; Recall represents the proportion of true positive samples correctly identified; and the F1 score comprehensively evaluates both Precision and Recall performance.

[0169] The experimental results of each comparison model and evaluation index are shown in Table 2, where LNN-DACA is the RPI prediction model proposed in this application:

[0170] Table 2. Performance comparison of the method of this invention with existing baseline models on four benchmark datasets.

[0171]

[0172] To verify the effectiveness of the method of the present invention, several existing ncRNA-protein interaction prediction models were selected as comparison methods, including LPIDF, RPI-EDLCN, MHAM, CCGNN, Graph-RPI, SeqMG-RPI, and ZHMolGraph.

[0173] Among them: (1) The LPIDF model is an ncRNA-protein interaction prediction method based on sequence feature encoding and feature fusion. It constructs a prediction model by extracting statistical features from RNA and protein sequences.

[0174] (2) The RPI-EDLCN model uses deep learning feature extraction to learn the potential representations in RNA and protein sequences through convolutional networks or deep neural networks.

[0175] (3) The MHAM model is based on the multi-head attention mechanism, which captures the association information between RNA and protein through different attention heads to achieve cross-modal feature modeling.

[0176] (4) The CCGNN model is based on a graph neural network structure. It learns RNA-protein interactions by constructing molecular feature maps and performing graph convolution propagation.

[0177] The above methods are all existing publicly available ncRNA-protein interaction prediction methods, and this invention uses them as benchmark models for performance comparison.

[0178] It is worth noting that ZEA22133 is a cross-species validation dataset derived from the plant *Zea mays*, and its interaction mechanism differs significantly from that of conventional animal data. The experimental data objectively and sufficiently demonstrate that, thanks to the dynamic evolutionary modeling of the liquid neural network and the adaptive temperature regulation mechanism of cross-attention, this invention not only achieves state-of-the-art performance on conventional datasets but also achieves significant leaps in all indicators when processing highly heterogeneous and widely distributed cross-species biological data. This demonstrates extremely excellent generalization ability and predictive robustness, possessing high practical application and industrial transformation value.

[0179] Implementation of the ncRNA-protein interaction prediction system:

[0180] A schematic diagram of the architecture of the ncRNA-protein interaction prediction system in this embodiment is shown below. Figure 4 As shown, the system specifically includes a memory, a processor, a system bus, and a computer program stored in the memory. The processor and the memory communicate and exchange data via the system bus. The processor executes the computer program to implement the steps of the ncRNA-protein interaction prediction method described in the embodiments of the present invention. The specific ncRNA-protein interaction prediction method has been described in sufficient detail in the above embodiments and will not be repeated here. The processor can be a microprocessor (MCU) or other processing system; the memory can be any type of memory that stores information using electrical energy, such as non-volatile storage media (including computer programs, databases), or other types of memory.

[0181] This invention has the following characteristics:

[0182] This invention proposes a prediction method, LNN-DACA, based on liquid neural networks and dynamic adaptive cross-attention. This method constructs a dual-path protein processor incorporating a liquid neural network (LNN) path, utilizing neural network differential equations to simulate the continuous-time evolution of biological sequences and capture long-range dynamic features. Simultaneously, a bidirectional dynamic adaptive temperature cross-attention mechanism is introduced, whereby the smoothness of the attention distribution is adjusted in real-time based on query-key similarity through an adaptive temperature branch, enabling the model to adaptively focus on broad-spectrum binding or precise site binding patterns. Experimental results show that this method exhibits excellent prediction performance on multiple benchmark datasets, significantly improving the accuracy and cross-species generalization performance of non-coding RNA-protein interaction recognition, and providing a more stable and accurate computational tool for studying complex intermolecular interaction mechanisms.

[0183] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still make modifications to the technical solutions described in the foregoing embodiments without creative effort, or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting ncRNA-protein interactions, characterized in that, The method includes: S1. Obtain the k-mer sequence coding vectors of the ncRNA and protein to be tested; S2. Input the k-mer sequence encoding vectors of ncRNA and protein into the pre-trained RPI prediction model to obtain the RPI prediction results; The RPI prediction model includes a two-way dynamic interaction module and an output prediction layer; The bidirectional dynamic interaction module consists of cascaded interaction layers, which include an ncRNA-protein pathway with ncRNA characterization as input and protein feature expression as output, and a protein-ncRNA pathway with protein characterization as input and ncRNA feature expression as output. Both pathways include an improved cross-attention layer with the input of the other pathway as the target sequence to extract deep binding features of ncRNA and protein. The improvement of the cross-attention layer lies in the fact that the query vector and key vector of the cross-attention are pooled in the sequence dimension, and the temperature parameter of the cross-attention is determined according to the similarity between the two pooled vectors; the higher the similarity, the larger the temperature parameter. The output prediction layer outputs the final prediction result based on the output features of the bidirectional dynamic interaction module.

2. The method for predicting ncRNA-protein interactions according to claim 1, characterized in that, The temperature parameter Determined according to the following formula: in, This represents the similarity between the pooled query vector and the key vector. This represents the Sigmoid activation function; Represents a learnable scalar bias term; The dimension of the query vector; It is a numerically stable term.

3. The method for predicting ncRNA-protein interactions according to claim 1, characterized in that, The process of obtaining the protein feature expression based on the output features of the improved cross-attention layer includes: the protein attention features output by the improved cross-attention layer are linearly mapped and activated by an activation function to obtain the corresponding projection features; the projection features are nonlinearly transformed and then weighted and added with the protein hidden state of the previous interaction layer through a learnable weight to obtain the candidate protein hidden state of the interaction layer; then the candidate protein hidden state and the protein hidden state of the previous interaction layer are fused through a protein dynamic gating factor and a decay coefficient to obtain the protein hidden state of the interaction layer; the projection features and the protein hidden state of the interaction layer are concatenated and linearly mapped to obtain the protein feature expression of the interaction layer. The protein dynamic gating factor is obtained as follows: the protein gating feature, protein attention feature, and the protein hidden state of the previous interaction layer are concatenated and then activated by a linear mapping and activation function to obtain the protein dynamic gating factor; the protein gating feature is obtained by fusing the protein attention feature and the protein representation as input to the interaction layer through directional gating fusion weights; the directional gating fusion weights are obtained by concatenating the protein attention feature and the protein representation and then activating by a linear mapping and activation function. The process of obtaining the ncRNA feature expression based on the output features of the improved cross-attention layer includes: the ncRNA attention features output by the improved cross-attention layer are linearly mapped and activated by an activation function to obtain the corresponding projection features; the projection features are nonlinearly transformed and then weighted and added with the ncRNA hidden state of the previous interaction layer through a learnable weight to obtain the candidate ncRNA hidden state of the interaction layer; then, the candidate ncRNA hidden state and the ncRNA hidden state of the previous interaction layer are fused through an ncRNA dynamic gating factor and a decay coefficient to obtain the ncRNA hidden state of the interaction layer; the projection features and the ncRNA hidden state of the interaction layer are concatenated and linearly mapped to obtain the ncRNA feature expression of the interaction layer. The ncRNA dynamic gating factor is obtained as follows: the ncRNA gating feature, the ncRNA attention feature, and the ncRNA hidden state of the previous interaction layer are concatenated and then activated by a linear mapping and an activation function to obtain the ncRNA dynamic gating factor; the ncRNA gating feature is obtained by fusing the ncRNA attention feature and the ncRNA representation as input to the interaction layer through a direction-gating fusion weight; the ncRNA attention feature and the ncRNA representation are concatenated and then activated by a linear mapping and an activation function to obtain the direction-gating fusion weight.

4. The method for predicting ncRNA-protein interactions according to claim 1, characterized in that, The RPI prediction model also includes a protein feature enhancement layer for enhancing the features of the protein's k-mer sequence encoding vector; the protein representation of the first interaction layer in the bidirectional dynamic interaction module is the output feature of the protein feature enhancement layer.

5. The method for predicting ncRNA-protein interactions according to claim 4, characterized in that, The protein feature enhancement layer includes a local convolutional path, a global LNN path, and a gated fusion layer; The local convolution path is used to extract local features from the k-mer sequence encoding vector of a protein through a cascaded feature extraction module; the feature extraction module includes a one-dimensional convolutional layer, a batch normalization, and an activation function connected in sequence. The global LNN path is used to extract global features from the k-mer sequence encoding vector of a protein through sequentially connected linear layers and LNN layers; The gated fusion layer is used to weight and sum the local and global features using a gated weight factor to obtain the enhanced features, which are then used as the output features of the protein feature enhancement layer. The gated weight factor is obtained by concatenating the local and global features and then activating them using a linear mapping and activation function.

6. The method for predicting ncRNA-protein interactions according to claim 1, characterized in that, The output prediction layer includes a feature reconstruction layer, a feature fusion layer, and a fully connected prediction layer; The feature reconstruction layer reconstructs the input by passing the k-mer sequence encoding vectors of ncRNA and protein through a multi-head attention mechanism and a feedforward neural network; The feature fusion layer is used to splice the output features of the feature reconstruction layer, the k-mer sequence encoding vector of ncRNA, the k-mer sequence encoding vector of protein, and the ncRNA and protein features output by the bidirectional dynamic interaction module in the channel dimension. The fully connected prediction layer is used to output RPI prediction results through a fully connected network based on the features concatenated by the feature fusion layer.

7. The method for predicting ncRNA-protein interactions according to claim 4, characterized in that, The process of training the RPI prediction model includes: (1) Remove the bidirectional dynamic interaction module and retain the simplified network architecture that splices the k-mer sequence encoding vector of ncRNA and the output features of the protein feature enhancement layer and then directly feeds them into the output prediction layer; pre-train the simplified network architecture. (2) Transfer the pre-trained network parameters to the complete RPI prediction model and randomly initialize the bidirectional dynamic interaction module in the RPI prediction model; then fine-tune the entire model to obtain the trained RPI prediction model.

8. The method for predicting ncRNA-protein interactions according to claim 1 or 7, characterized in that, The RPI prediction model is trained using loss functions including binary cross-entropy loss.

9. The method for predicting ncRNA-protein interactions according to claim 1, characterized in that, The k-mer sequence encoding vector of the ncRNA is a 4-mer vector; the k-mer sequence encoding vector of the protein is a 2-mer vector.

10. An ncRNA-protein interaction prediction system, comprising a processor, characterized in that, The processor is used to execute a computer program to implement the steps of the ncRNA-protein interaction prediction method as described in any one of claims 1 to 9.

Images