ncrna and protein interaction association prediction method based on mlp-mixer

Abstract

The application provides an ncRNA and protein interaction correlation prediction method based on an MLP-Mixer. First, protein / ncRNA sequences are encoded by a joint k-mer method; then, the sequences are input into a mixed network combined with an MLP-Mixer network to extract sequence features; finally, the interaction between nucleotides and amino acids is simulated by a double-sequence comparison attention module, and the output is classified by a classification part. The application comprehensively considers first-level sequence features, second-level structure sequence features and third-level structure features, introduces a multi-layer perception mixer module, and combines the module with a convolutional neural network and a bidirectional long short-term memory network, so that long-distance dependencies between features recognized by the CNN can be captured. Meanwhile, the complex interaction between nucleotides and amino acids is simulated by the double-sequence comparison attention module to explore subtle correlations between ncRNA and protein sequences.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for predicting the correlation between ncRNA and protein interactions based on MLP-Mixer. Background Technology

[0002] RNA-protein interactions play a crucial role in physiological and pathological processes, involving regulation at multiple levels, including RNA synthesis, modification, transport, translation, and degradation. The regulatory function of RNA largely depends on its binding proteins (RBPs). Many proteins, such as RNA polymerases, transcription factors, and regulatory proteins, require RNA binding to exert their biological functions.

[0003] A 2001 genome sequencing study revealed that only about 2% of RNA is responsible for encoding proteins, while a staggering 98% was identified as non-coding RNAs (ncRNAs). Research indicates that the interaction between proteins and ncRNAs is a crucial element in gene expression regulation and various biological processes. Although ncRNAs do not directly encode proteins, they play a key role at the post-transcriptional level through interactions with proteins. These interactions can regulate protein activity and stability, thereby influencing a variety of intracellular biological processes. These processes include translation, RNA splicing, chromatin remodeling, and gene regulation—key life activities. Therefore, accurately predicting ncRNA-protein interactions is essential for understanding the regulatory functions of ncRNAs, the pathogenesis of related diseases, and the development of new drugs and treatments.

[0004] Since the beginning of the 21st century, various high-throughput experimental and structural analysis techniques have been developed to elucidate the molecular mechanisms and structural basis of ncRNA-protein binding. These high-throughput techniques, such as RIP-Chip, HITS-CLIP, and PAR-CLIP, can simultaneously detect the interactions between a large number of ncRNAs and proteins, determining their binding sites and interaction strengths through steps such as cross-linking, immunoprecipitation, and sequencing. Structural analysis techniques include X-ray crystallography, nuclear magnetic resonance, and cryo-electron microscopy. While these techniques provide valuable data and information for studying ncRNA-protein binding, challenges remain, including difficulties in sample preparation, limitations in experimental conditions, and data noise and bias. Therefore, developing effective computational methods to predict ncRNA-protein binding is of great significance.

[0005] With the rapid development of artificial intelligence in the 21st century, computational methods have made significant progress in predicting ncRNA-protein interactions. These methods can be broadly classified into two categories: those based on sequence features as input and those based on structural features. For sequence feature-based methods, for example, IPMiner uses a stacked autoencoder to mine hidden features from the sequence composition features of protein and ncRNA sequences, then inputs the learned hidden features into a random forest model for prediction; RPISeq predicts ncRPI based solely on the sequences of ncRNAs and proteins using both support vector machines (SVM) and random forests (RF); CFRP extracts traditional k-mer features from ncRNA and protein pairs as basic features and constructs complex features through different types of nonlinear transformations to predict ncRPI. Compared to sequence feature-based methods, structural feature-based ncRPI prediction methods provide a new perspective for the study of ncRNA-protein interactions. This method can provide richer data by capturing the interaction patterns of molecules in three-dimensional space. Therefore, it performs well in terms of prediction accuracy and generalization ability, mainly due to the fact that structural features can provide details that sequence information cannot cover. In light of this, structure-based computational prediction methods have been extensively studied and applied. catRAPID calculates the secondary structural properties of protein and ncRNA sequences, combining physicochemical features such as hydrogen bonds, hydrophobicity, and van der Waals forces to determine binding propensity; RPI-Pred combines sequence and structural information of ncRNAs and proteins, using a support vector machine model to predict whether protein-ncRNA pairs interact; LPI-BLS utilizes a broad learning system, inputting the results of multiple independent broad learning systems into a stacked ensemble classifier to predict lncRNA-protein interactions.

[0006] In recent years, deep learning, a subfield of artificial intelligence, has significantly impacted research on ncRPI prediction. BGFE uses a position-specific scoring matrix (PSSM) to represent protein sequences and a k-mer sparse matrix to represent ncRNA sequences, employing a stacked autoencoder (SAE) network for ncRPI prediction. RPITER improves the joint triplet feature (CTF) encoding method, combining CNN and SAE to learn k-mer features of RNA and protein sequences. EDLMFC combines a convolutional neural network to learn dominant biological information and a Bi-LSTM network to capture long-range dependencies between features recognized by CNN. These advanced methods not only improve the accuracy and efficiency of ncRPI prediction but also provide invaluable tools for research in bioinformatics and systems biology.

[0007] While existing methods have made significant strides in predicting interactions between ncRNAs and proteins, there is still room for improvement in predictive performance. Traditional convolutional neural networks, due to their fixed structure, struggle to capture long-range interactions between sequences. In contrast, attention mechanisms are better suited to handling long-range dependencies in both protein and ncRNA sequences. Furthermore, current research methods often overlook subtle associations that may exist between ncRNA and protein sequences, and their feature extraction of both protein and ncRNA sequences is still insufficient. Summary of the Invention

[0008] This invention addresses the shortcomings of current biological experimental methods for identifying ncRNA-protein interactions (ncRPI), which are not only costly and time-consuming, but also fail to consider the subtle associations between ncRNA and protein sequences. Furthermore, the complex molecular interaction mechanisms between ncRNA and protein, and the limitations in the size and diversity of available datasets, further hinder the development of this method. Therefore, this invention provides an MLP-Mixer-based method for predicting ncRNA-protein interaction associations. By predicting ncRPI through a Multi-Layer Perceptron Mixer (MLP-Mixer) network, this method can accurately predict the interactions between ncRNA and protein.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for predicting the association between ncRNA and protein based on MLP-Mixer is proposed. First, the protein / ncRNA sequence is encoded by combining the k-mer method. Then, the sequence is input into a hybrid network combined with the MLP-Mixer network to extract sequence features. Finally, the interaction between nucleotides and amino acids is simulated by a double sequence comparison attention module, and the output is sent to the classification part for classification.

[0011] Furthermore, the hybrid network has two independent CNN Layer modules, one for ncRNA and one for protein data. Each CNN Layer module consists of a one-dimensional convolutional layer (Conv1D), a max pooling layer, and a batch normalization layer. As the convolutional filters move on the protein or ncRNA sequence, they can capture different combinations of amino acids or nucleotides, thereby generating a potential vector containing biological information.

[0012] Furthermore, the MLP-Mixer network mainly consists of a Per-patch fully connected layer and a Mixer Block. The Per-patch fully connected layer is responsible for generating a dense representation of the input features. During processing, the input is first converted into a feature matrix of a specific size, and then the Per-patch fully connected layer expands the channel dimension of each patch to 128. Subsequently, the Mixer Block further extracts features. The Mixer Block structure includes a fully connected layer, a global average pooling (GAP) layer, two MLP-Mixer blocks, and two layer normalization (LN) layers. The Mixer Block is defined as follows:

[0013] U S =X S +W3σ(W2LayerNorm(W1X S )) (1)

[0014]

[0015] In the formula, W1, W2, W3, W4, and W5 represent the inputs to the Mixer Block; W1, W2, W3, W4, and W5 are the weight matrices of the fully connected layers; σ(·) represents the GELU activation function; and LayerNorm is the layer normalization layer. and These are the outputs of the channel hybrid perceptron and the spatial hybrid perceptron, respectively.

[0016] Furthermore, the dual-sequence comparison attention module is implemented by creating an attention matrix that represents the spatial and channel interactions between the two sequences. The model then calculates the average of this matrix and applies a sigmoid function to obtain the final attention score. These scores are subsequently used to weight the original sequences to highlight the parts that the model deems important. The specific steps are as follows:

[0017] For the output R of Mixer Block S and P L They are then transformed into attention vectors Ra using a multilayer perceptron. S and Pa L ;

[0018] Ra S =σ(W R ·R S +b) (3)

[0019] Pa L =σ(W P ·P L +b) (4)

[0020] In the formula, σ(·) is the nonlinear activation function ReLU; and b is the weight matrix; b is the bias vector.

[0021] Then, an attention matrix A is created to represent the interaction between the two sequences in space and channels. S,L :

[0022] A S,L =σ(W A ·(Ra S +Pa L (5)

[0023] In the formula, This is the weight matrix;

[0024] After completing the above steps, the attention matrix Successfully constructed; subsequently, averaging was performed on each dimension to obtain the ncRNA sequence attention matrix. and protein sequence attention matrix

[0025] A R =α(Mean(A,2)) (6)

[0026] A P =α(Mean(A,1)) (7)

[0027] In the formula, α is the Sigmoid activation function, which is used to map the attention score to the range (0,1); Mean(A,axis) is used to obtain the average value of each row A in dimension axis;

[0028] Subsequently, based on the ncRNA sequence attention matrix and protein sequence attention matrix Obtain the characteristic matrix R A and P A :

[0029] R A =R S 0.6+R S e A R (8)

[0030] P A =P L 0.6+P L e A P (9)

[0031] In the formula, e represents element-wise multiplication; R S and P L This indicates the output of the Mixer Block;

[0032] Finally, the feature matrix R A and P A It is connected and fed into the classifier.

[0033] Furthermore, the steps of the MLP-Mixer-based method for predicting ncRNA-protein interaction associations are as follows:

[0034] First, the high-order features extracted from ncRNA, protein sequence features and their derived features through CNN layers are fed into the Bi-LSTM network.

[0035] Then, the network output is expanded and reshaped to the appropriate dimensions to fit the input of the Mixer Layer, and the MLP-Mixer network is used to further refine the feature representations of protein sequences and ncRNA sequences at a deeper level.

[0036] Subsequently, the complex interactions between nucleotides and amino acids were simulated in the Attention Moudule to uncover subtle associations between ncRNA and protein sequences, and the corresponding attention scores were calculated. These attention scores were then used to weight the original feature representations, and the weighted summed feature matrices were then concatenated.

[0037] Finally, the concatenated feature representation is input into a classifier consisting of two fully connected layers and a Softmax function. These layers work together to process the merged high-level features and ultimately generate a binary value as the basis for determining whether the objects interact.

[0038] This invention proposes a deep learning framework model (named RPI-HNBM) that integrates MLP-Mixer, CNN, Bi-LSTM, and a dual sequence comparison attention module. This model not only utilizes the primary sequence features of ncRNAs and proteins but also integrates secondary and tertiary structural features. Furthermore, to effectively extract deep network features and accurately identify ncRNA-protein interaction pairs, this invention employs MLP-Mixer to capture long-distance dependencies between protein and ncRNA sequences. Compared with existing technologies, the advantages of this invention are as follows:

[0039] (1) This invention is the first to use MLP-Mixer to capture the long-distance dependence between protein and ncRNA sequences extracted by CNN. This is the first effective application of MLP-Mixer in the field of ncRPI prediction to extract protein and ncRNA sequence features.

[0040] (2) This invention comprehensively considers primary sequence features, secondary structure sequence features, and tertiary structure features, and uses a combined k-mer method to extract sequence features of proteins and ncRNAs. In terms of model architecture, RPI-HNBM introduces a multilayer perceptron mixer (MLP-Mixer) module and combines it with a convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM), thereby capturing the long-range dependencies between features identified by the CNN. Simultaneously, this invention uses a dual sequence comparison attention module to simulate the complex interactions between nucleotides and amino acids to uncover subtle associations between ncRNA and protein sequences.

[0041] (3) This invention evaluates RPI-HNBM on three public datasets. Experimental results show that RPI-HNBM exhibits superior performance on most evaluation metrics compared to other state-of-the-art prediction methods, thus demonstrating the effectiveness of the RPI-HNBM method in ncRPI prediction. Attached Figure Description

[0042] Figure 1This section provides an overview of the RPI-HNBM model. Protein and non-coding RNA sequence features extracted by a CNN layer are processed by a Bi-LSTM network to capture long-range dependencies between input features from both directions. The output is expanded and reshaped to appropriate dimensions, then processed by an MLP-Mixer network to extract protein and non-coding RNA sequence features. A dual-sequence comparison attention mechanism is used to compare sequences and calculate attention scores, which are then used to weight the original feature representations. Finally, the weighted summation of the high-level features is fused and output to a fully connected layer, where the softmax activation function of the last layer is used for binary prediction.

[0043] Figure 2 This describes the specific structure of the Mixer Layer module. Each Mixer Layer module consists of two fully connected (FC) layers, one GELU activation function, and two Dropout layers.

[0044] Figure 3 This is a comparison of the average performance results of various methods using 5-fold cross-validation on the RPI1807 dataset.

[0045] Figure 4 This is a comparison of the average performance results of various methods using 5-fold cross-validation on the NPInterv2.0 dataset.

[0046] Figure 5 This is a comparison of the average performance results of various methods using 5-fold cross-validation on the RPI488 dataset. Detailed Implementation

[0047] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in 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 in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of this application.

[0048] Given an ncRNA sequence L and a corresponding protein sequence S, the main objective of this invention is to develop an efficient mapping function F(L,S). This function aims to map these two sequences to an abstract space defined by binary labels Y, where "1" indicates an interaction between the two sequences, and "0" indicates no interaction. Therefore, this invention proposes an MLP-Mixer-based method for predicting the association between ncRNA and protein interactions. First, the protein / ncRNA sequence is encoded using a combined k-mer method; then, it is input into a hybrid network incorporating an MLP-Mixer network to extract sequence features; finally, a dual-sequence comparison attention module simulates the interaction between nucleotides and amino acids, and the output is used for classification.

[0049] Example 1

[0050] This invention used three publicly available datasets to evaluate the performance of the method: RPI1807, NPInter v2.0, and RPI488. The RPI1807 dataset was created based on information from the Nucleic Acid Database (NDB) and the Protein-RNA Interface Database (PRIDB), and a non-redundant RPI composite training dataset was constructed by parsing these databases. It contains 1807 positive pairs (composed of 1807 proteins and 1078 RNA strands) and 1436 negative pairs (containing 1436 proteins and 493 RNA strands).

[0051] The RPI488 dataset (Fan XN, Zhang S W. LPI-BLS: Predicting lncRNA–protein interactions with abroad learning system-based stacked ensemble classifier[J]. Neurocomputing, 2019, 370: 88-93.) was extracted from the Protein Database (PDB). It consists of ncRNA-protein interaction pairs longer than 200 characters, covering 488 ncRNA-protein pairs, including 243 positive interaction pairs and 245 negative interaction pairs. The construction of the RPI1807 and RPI488 datasets followed the minimum atomic spacing criterion. This criterion stipulates that if any atomic pair between a protein and RNA is less than a predetermined threshold, the protein-RNA pair is classified as an interaction pair.

[0052] NPInter v2.0 is derived from the literature (Wang J, Zhao Y, Gong W, et al. EDLMFC: an ensemble deep learning framework with multi-scale features combination for ncRNA–protein interaction prediction[J]. BMC bioinformatics, 2021, 22: 1-19.), extracted from the NPInter dataset. The NPInter dataset mainly covers experimentally validated physical interactions, and its judgment criteria do not rely on the distance measurement between atoms in the RNA-protein complex.

[0053] Given the insufficient number of negative samples in the NPInter dataset, following the method described in the literature (Pan X, Fan YX, Yan J, et al. IPMiner: hidden ncRNA-protein interaction sequential pattern mining with stacked autoencoder for accurate computational prediction[J]. BMC Genomics, 2016, 17: 1-14.), an equal number of non-interacting pairs were artificially constructed as negative samples by randomly pairing ncRNAs and proteins from the positive samples. Detailed dataset data is shown in Table 1.

[0054] Table 1. Statistics of ncRNA and protein sequence data from three publicly available datasets.

[0055]

[0056] This invention employs the SPOT-RNA method to predict the secondary structure of ncRNAs. Since the SPOT-RNA method has a limit on sequence length (maximum input length is 500 nucleotides), all ncRNA sequences exceeding 500 nucleotides in length were excluded. The preprocessed dataset information is shown in Table 2. Furthermore, because the RPI488 dataset was too imbalanced after removing some ncRNAs, the method described in the literature (Wang J, Zhao Y, Gong W, et al. EDLMFC: an ensemble deep learning framework with multi-scale features combination for ncRNA–protein interaction prediction[J]. BMC bioinformatics, 2021, 22: 1-19.) was used to divide it into 5 subsets, and the average of these subsets was taken as the final result.

[0057] Table 2. Statistics of ncRNA and protein sequence data in the preprocessed dataset.

[0058]

[0059] In this invention, a total of four sets of features were selected for training the deep learning model. These features are all derived from protein sequences and ncRNA sequences. The following sections will elaborate on these feature sets.

[0060] 1. Original ncRNA sequences and protein sequences

[0061] For ncRNAs, the original ncRNA sequence includes its nucleotide sequence, which is the most basic biological information unit; for proteins, the amino acid sequence determines its linear structure, from which functional regions can be identified. This invention uses a combined k-mer encoding method to represent the types of amino acids in protein sequences and the types of ribonucleotides in ncRNA sequences.

[0062] 2ncRNAs and Protein Secondary Structure

[0063] In this invention, the secondary structure information of ncRNA is predicted using the SPOT-RNA method. In this method, the secondary structure of ncRNA can be described by seven different structural types, each represented by a specific character: stem (S), hairpin loop (H), multiloop (M), inner loop (I), protrusion (B), outer loop (X), and terminal (E).

[0064] Furthermore, this invention obtains the secondary structure of proteins using the SPIDER3 method. Given that the secondary structure of a protein sequence can be represented as three structural types—α-helix, β-sheet, and coil—three single-character identifiers are used to represent them.

[0065] 3. Protein tertiary structure

[0066] For protein tertiary structure prediction, this invention also uses the SPIDER3 method, which specifically includes: solvent accessible surface area (ASA), contact number (CN), upper hemisphere exposure (HSEα-up), and lower hemisphere exposure (HSEα-down).

[0067] 4 Interface preferences

[0068] Interface propensity (IP) is the probability of a specific residue appearing at a protein-nucleic acid interface. This invention uses the method described in the literature (Li CH, Cao LB, Su JG, et al. A new residue-nucleotide propensity potential with structural information considered for discriminating protein-RNA docking decoys[J]. Proteins: Structure, Function, and Bioinformatics, 2012, 80(1):14-24.) to obtain residue-nucleotide propensity.

[0069] To meet the demands of deep learning models, ncRNAs and protein sequences must be converted into numerical vector forms. This step is crucial for processing biological sequence data because it allows algorithms to identify and learn sequence patterns and characteristics.

[0070] This invention uses a combined k-mer encoding method to extract ncRNA and protein sequence features, that is, combining the frequency information of 1-mer to 3-mer to represent the primary sequence and secondary structure information of proteins; and combining the frequency information of 1-mer to 4-mer to represent the primary sequence and secondary structure information of ncRNAs.

[0071] In protein classification, based on properties such as dipole moment and side chain volume, 20 standard amino acids can be grouped into 7 groups: G1 (alanine (A), glycine (G), valine (V)); G2 (isoleucine (I), leucine (L), phenylalanine (F), and proline (P)); G3 (tyrosine (Y), methionine (M), threonine (T), and serine (S)); G4 (histidine (H), asparagine (N), glutamine (Q), and tryptophan (W)); G5 (arginine (R), lysine (K)); G6 (aspartic acid (D), glutamic acid (E)); and G7 (cysteine ​​(C)). Therefore, each protein sequence can be represented by seven different letters. By calculating the frequencies of 3-mers, 2-mers, and 1-mers, the protein sequence can be converted into a group containing 399... A numerical vector with 39 elements. Furthermore, based on their structural characteristics, the secondary structure of proteins can be divided into three basic types: α-helix, β-sheet, and random coil, thus it can be encoded as a 39-element vector. A numerical vector of 100 elements.

[0072] ncRNAs are composed of four ribonucleotides: adenine (A), uridine (U), guanine (G), and cytosine (C). This means that each ncRNA sequence can be represented by four different letters. By calculating the frequency distribution of 4-mer, 3-mer, 2-mer, and 1-mer, the ncRNA sequence can be converted into a 340-character nucleotide sequence. A numerical vector with 2800 elements. Furthermore, the secondary structure of ncRNA can be subdivided into seven types (X, E, S, I, H, B, M), therefore its secondary structure information can be represented by a 2800-element numerical vector. It is described by a numerical vector of 10 elements.

[0073] Figure 1 An overview of the RPI-HNBM model is presented, which includes a CNN Layer, Bi-LSTM, MLP-Mixer, and Attention Module.

[0074] First, high-order features extracted from ncRNA, protein sequence features, and their derived features through CNN layers are fed into a Bi-LSTM network. This network can capture the long-range dependencies of the input features from both forward and backward perspectives.

[0075] Then, the network output is expanded and reshaped to the appropriate dimensions to fit the input of the Mixer Layer, and the MLP-Mixer network is used to further refine the feature representations of protein sequences and ncRNA sequences at a deeper level.

[0076] Subsequently, the complex interactions between nucleotides and amino acids are simulated in the Attention Moudule to uncover subtle associations between ncRNA and protein sequences. The corresponding attention scores are calculated, and these attention scores are used to weight the original feature representations. The weighted summed feature matrices are then concatenated.

[0077] Finally, the present invention inputs the concatenated feature representation into a classifier, which consists of two fully connected layers and a Softmax function. These layers work together to process the merged high-level features and finally generate a binary value as the basis for determining whether the objects interact.

[0078] The structure of each module of the model is described below:

[0079] 1 CNN layer

[0080] In the model design of this invention, there are two independent CNN Layer modules, one for ncRNA and one for protein data. Each CNN Layer module consists of a one-dimensional convolutional layer (Conv1D), a max-pooling layer, and a batch normalization layer. These layers are used to extract semantic information from the sequence data. The one-dimensional convolutional layer can identify and capture local features in the entire sequence. As the convolutional filters move across the protein or ncRNA sequence, they can capture different combinations of amino acids or nucleotides, thereby generating a potential vector containing biological information. The max-pooling layer is used to focus on the most important parts, and the batch normalization layer provides a regularization effect, effectively improving the training efficiency and generalization ability of the model.

[0081] 2MLP-Mixer Network

[0082] To delve deeper into the features of ncRNA and protein sequences and compensate for the limitations of CNNs in capturing long-range dependencies, this invention introduces an MLP-Mixer network structure. The feature input to this network is provided by a multi-layer network consisting of a CNN layer and a Bi-LSTM network. The MLP-Mixer network mainly consists of per-patch fully connected layers and Mixer Blocks. The per-patch fully connected layers are responsible for generating a dense representation of the input features. During processing, the input is first transformed into a feature matrix of a specific size, and then the per-patch fully connected layers expand the channel dimension of each patch to 128. Subsequently, the Mixer Block further extracts features. The Mixer Block structure includes a fully connected layer, a global average pooling (GAP) layer, two MLP-Mixer blocks, and two layer normalization (LN) layers. The specific structure of the Mixer Layer is as follows: Figure 2 As shown, a Mixer Block can be defined as:

[0083] U S =X S +W3σ(W2LayerNorm(W1X S )) (1)

[0084]

[0085] In the formula, W1, W2, W3, W4, and W5 represent the inputs to the Mixer Block; W1, W2, W3, W4, and W5 are the weight matrices of the fully connected layers; σ(·) represents the GELU activation function; and LayerNorm is the layer normalization layer. and These are the outputs of the channel hybrid perceptron and the spatial hybrid perceptron, respectively.

[0086] 3. Two-sequence comparison attention module

[0087] To effectively uncover subtle associations between ncRNAs and protein sequences and simulate the complex interactions between nucleotides and amino acids, this invention introduces a special attention mechanism: the dual-sequence comparison attention module. This mechanism is achieved by creating an attention matrix that represents the spatial and channel-wise interactions between the two sequences. The model then calculates the average of this matrix and applies a sigmoid function to obtain a final attention score, which is subsequently used to weight the original sequences to highlight the parts that the model deems important.

[0088] More specifically, for the output R of the Mixer Block S and P LThis invention uses a multilayer perceptron to convert them into attention vectors Ra. S and Pa L .

[0089] Ra S =σ(W R ·R S +b) (3)

[0090] Pa L =σ(W P ·P L +b) (4)

[0091] In the formula, σ(·) is the nonlinear activation function ReLU; and is the weight matrix; b is the bias vector.

[0092] Then, an attention matrix A is created to represent the interaction between the two sequences in space and channels. S,L :

[0093] A S,L =σ(W A ·(Ra S +Pa L (5)

[0094] In the formula, This is the weight matrix.

[0095] After completing the above steps, the attention matrix Successfully constructed. Subsequently, this invention performs an averaging operation on each dimension to obtain the ncRNA sequence attention matrix. and protein sequence attention matrix

[0096] A R =α(Mean(A,2)) (6)

[0097] A P =α(Mean(A,1)) (7)

[0098] In the formula, α is the Sigmoid activation function, which is used to map the attention score to the range (0,1); Mean(A,axis) is used to obtain the average value of each row A in dimension axis.

[0099] Subsequently, based on the ncRNA sequence attention matrix and protein sequence attention matrix Obtain the characteristic matrix R A and P A :

[0100] RA =R S 0.6+R S e A R (8)

[0101] P A =P L 0.6+P L e A P (9)

[0102] In the formula, e represents element-wise multiplication; R S and P L This indicates the output of the Mixer Block.

[0103] Finally, the feature matrix R A and P A It is connected and fed into the classifier.

[0104] Example 2

[0105] Model evaluation metrics and experiments:

[0106] In this invention, interacting sample pairs are classified as positive samples, while non-interacting sample pairs are considered negative samples. To more comprehensively compare the performance of the proposed RPI-HNBM model with other state-of-the-art ncRPI prediction methods, this invention employs seven evaluation metrics: ACC, TPR, TNR, PPV, MCC, F1, and AUC. These metrics collectively construct a comprehensive evaluation system aimed at fully evaluating and reflecting the model's predictive performance.

[0107]

[0108] In the formula, TP (true positive) is the number of interaction pairs correctly identified by the model; FP (false positive) is the number of interaction pairs incorrectly predicted by the model; TN (true negative) is the number of non-interaction pairs correctly identified by the model; and FN (false negative) is the number of non-interaction pairs incorrectly predicted by the model.

[0109] In addition, ACC (Accuracy) represents the proportion of samples correctly classified by the classifier, and is a comprehensive performance metric. TPR (True Positive Rate) represents the proportion of samples that the classifier correctly identifies as actually positive. TNR (True Negative Rate) reflects the proportion of samples that the classifier correctly identifies as actually negative. PPV (Precision) shows the proportion of samples predicted as positive that are actually positive. MCC (Matthews Correlation Coefficient) is a comprehensive performance metric that takes into account class imbalance. F1 (F1 Score) is the harmonic mean of PPV and TPR, reflecting their balance. AUC (Area Under the Curve) measures the classifier's ability to rank positive classes ahead of negative classes. These metrics collectively measure the model's predictive accuracy and reliability.

[0110] In this invention, a GeForce RTX 3070 was used as the experimental device, and a deep learning model based on the Tensorflow 2.6 framework was built, combined with the CUDA 11.3 acceleration library and Python 3.8 to complete the experiments. During model training, this invention selected Adaptive Moment Estimation (Adam) and Stochastic Gradient Descent (SGD) as the main optimization algorithms.

[0111] In the configuration of the deep learning architecture, this invention sets the batch size to 128. This setting ensures a sufficient number of samples for effective learning while maintaining reasonable utilization of computational resources. To prevent overfitting, when implementing the MLP-Mixer structure, the linear mapping dimension of the per-patch fully connected layer is set to 128, and the dropout rate is set to 0.2 to enhance the model's generalization ability.

[0112] Experimental results:

[0113] 1. Comparison with other advanced methods

[0114] To comprehensively evaluate the performance of the RPI-HNBM model, this invention selected four state-of-the-art methods—EDLMFC, RPITER, IPMiner, and CFRP—as comparison objects. Five-fold cross-validation was performed on three public datasets: RPI1807, NPInter v2.0, and RPI488, to obtain more robust evaluation results. To comprehensively evaluate the performance of these methods from different perspectives, seven performance metrics were used: ACC (accuracy), TPR (true positive rate), TNR (true negative rate), PPV (precision), MCC (Matthews correlation coefficient), F1 (F1 score), and AUC (area under the curve). These metrics can comprehensively reflect the model's performance on classification problems.

[0115] Figure 3-5 Experimental results on the RPI1807, NPInter v2.0, and RPI488 datasets are presented, with detailed numerical results recorded in Tables 3-5. Table 4 shows the comparison results of the RPI-HNBM method of this invention with other state-of-the-art research methods on the NPInter v2.0 dataset, highlighting the contribution of this invention to the field of ncRNA binding protein prediction.

[0116] like Figure 4 As shown in Table 4, RPI-HNBM outperforms all other methods across seven performance metrics on the NPInter v2.0 dataset. Furthermore, Table 4 indicates that RPI-HNBM's ACC is 93.2%, which is 3.5%, 4.2%, and 10.4% higher than EDLMFC, RPITER, IPMiner, and CFRP, respectively. RPI-HNBM's accuracy (ACC) is 93.2%, which is 11.1% higher than CFRP, 10.4% higher than IPMinter, 4.2% higher than RPITER, and 3.5% higher than EDLMFC. Its true positive rate (TPR) is 95.3%, which is 18.1% higher than CFRP, 11.0% higher than IPMinter, 3.7% higher than RPITER, and 3.6% higher than EDLMFC. Its true negative rate (TNR) is 91.1%, which is 4.2% higher than CFRP, 9.8% higher than IPMinter, 4.9% higher than RPITER, and 3.4% higher than EDLMFC. Its precision (PPV) is 91.5%, which is 6.0% higher than CFRP and 10.4% higher than IPMinter. The RPI-HNBM score is 93.3%, which is 12.2% higher than CFRP, 10.1% higher than IPMinter, 4.0% higher than RPITER, and 3.4% higher than EDLMFC. The Matthews correlation coefficient (MCC) of RPI-HNBM is 86.5%, which is 22.1% higher than CFRP, 20.9% higher than IPMinter, 8.4% higher than RPITER, and 7.0% higher than EDLMFC. The area under the curve (AUC) of RPI-HNBM is 98.1%, which is 9.7% higher than CFRP, 15.4% higher than IPMinter, 2.4% higher than RPITER, and 2.2% higher than EDLMFC.

[0117] Compared to the four methods mentioned above, the RPI-HNBM proposed in this invention demonstrates superior performance in predicting ncRPI on the NPInter v2.0 dataset. Figure 3 As shown, RPI-HNBM achieved the highest ACC, F1 scores, and MCC on the RPI1807 dataset. And as... Figure 5 As shown, except for TNR and PPV, RPI-HNBM outperforms other state-of-the-art methods on the RPI488 dataset. This indicates that RPI-HNBM, based on the MLP-Mixer network and incorporating a dual-sequence comparison attention mechanism, is an efficient ncRPI prediction method.

[0118] Table 3 shows the performance comparison results of RPI-HNBM with other state-of-the-art methods on the RPI1807 dataset.

[0119]

[0120] Table 4 shows the performance comparison results of RPI-HNBM with other state-of-the-art methods on the NPInterv2.0 dataset.

[0121]

[0122] Table 5 shows the performance comparison results of RPI-HNBM with other state-of-the-art methods on the RPI488 dataset.

[0123]

[0124] 2 Ablation Experiment

[0125] 2.1 Impact of MLP-Mixer Networks

[0126] In this invention, the contribution of the MLP-Mixer network to model performance is verified by removing it. Table 6 shows the comparative experimental results on the NPInter v2.0 dataset. Specifically, without the MLP-Mixer network, the overall performance of the model decreases. The various metrics ACC, TPR, TNR, PPV, MCC, F1, and AUC decrease from 93.2%, 95.3%, 91.1%, 91.5%, 93.3%, 86.5%, and 98.1% to 92.3%, 94.9%, 89.8%, 90.3%, 92.5%, 84.8%, and 96.8%, respectively. In addition to the overall performance decline, the decrease in TNR is more significant, indicating that the model's effectiveness in identifying negative samples has weakened to some extent.

[0127] Meanwhile, this invention explores the impact of using per-patch linear mapping layers of different dimensions [128, 256, 512] in the MLP-Mixer network on embedded features. Table 7 shows that when the representation dimension of the linear mapping layer is set to 128, all evaluation metrics reach their optimal levels.

[0128] Table 6 shows the performance comparison of the RPI-HNBM model before and after applying the MLP-Mixer network on the NPInterv2.0 dataset.

[0129]

[0130] Table 7 shows the performance comparison of different dimensional linear mapping layers in the MLP-Mixer network applied to the RPI-HNBM dataset on the NPInterv2.0 dataset.

[0131]

[0132] 2.2 The impact of the attention module in two-sequence comparison

[0133] In this study, to verify the importance of the two-sequence comparison attention module, it was removed to assess its specific impact on the overall model performance. According to the data in Table 8, with the two-sequence comparison attention module, the performance metrics ACC, TPR, TNR, PPV, MCC, F1, and AUC were 93.2%, 95.3%, 91.1%, 91.5%, 93.3%, 86.5%, and 98.1%, respectively. However, once the attention module was removed, these metrics immediately dropped to 92.1%, 94.4%, 89.8%, 90.3%, 92.3%, 84.3%, and 96.7%. These data show that the F1 score decreased the most, reaching 2.2%. This indicates that the model's ability to balance TPR and PPV decreased most significantly, followed by a 1.4% decrease in AUC, indicating a weakening of the model's overall ability to distinguish between positive and negative samples. This suggests that the two-sequence comparison attention module plays a role in improving classification performance.

[0134] Furthermore, this invention analyzes the computation of the feature matrix R in the dual-sequence comparison attention module. A and P A The impact of different weighting values ​​[0.4, 0.5, 0.6] on the classification results is shown in Table 9. It can be concluded that when the weight is set to 0.6, all seven evaluation metrics—ACC, TPR, TNR, PPV, MCC, F1, and AUC—reach optimal levels. The results indicate that a weighting of 0.6 may provide a better balance in the calculation of the feature matrix, enabling the model to more effectively capture the correlations and differences between sequences.

[0135] Table 8 shows the performance comparison of the RPI-HNBM model before and after applying the dual-sequence comparison attention module on the NPInterv2.0 dataset.

[0136]

[0137] Table 9 shows the performance comparison results of the two-sequence comparison attention module with different weight values ​​in RPI-HNBM on the NPInterv2.0 dataset.

[0138]

[0139] In summary, this invention constructs a hybrid network model based on MLP-Mixer—RPI-HNBM. This model integrates the MLP-Mixer network to explore long-range dependencies between ncRNA sequences and protein sequences, and introduces a two-sequence comparison attention mechanism to capture complex interactions between nucleotides and amino acids, revealing more subtle connections between ncRNA and protein sequences. On three publicly available ncRPI datasets, the model of this invention demonstrates superior performance compared to other methods.

Claims

1. A method for predicting the association between ncRNA and protein interactions based on MLP-Mixer, characterized in that, First, protein and ncRNA sequences are encoded using a combined k-mer method; then, they are input into a hybrid network incorporating an MLP-Mixer network to extract sequence features; finally, a double sequence comparison attention module simulates the interaction between nucleotides and amino acids, and the output is sent to the classification part for classification. The hybrid network has two independent CNN Layer modules, one for ncRNA and one for protein data. Each CNN Layer module consists of a one-dimensional convolutional layer (Conv1D), a max pooling layer (Maxpooling), and a batch normalization layer (BatchNormalization). As the convolutional filters move on the protein or ncRNA sequence, they can capture different combinations of amino acids or nucleotides, thereby generating a potential vector containing biological information. The MLP-Mixer network mainly consists of a Per-patch fully connected layer and a Mixer Block. The Per-patch fully connected layer is responsible for generating a dense representation of the input features. During processing, the input is first transformed into a feature matrix of a specific size, and then the Per-patch fully connected layer expands the channel dimension of each patch to 128. Subsequently, the Mixer Block further extracts features. The Mixer Block structure includes a fully connected layer, a global average pooling (GAP) layer, two MLP-Mixer blocks, and two layer normalization (LN) layers. The Mixer Block is defined as follows: (1) (2) In the formula, This represents the input to the Mixer Block. , , , , This is the weight matrix of the fully connected layer; This represents the GELU activation function; It is a normalized layer; and These are the outputs of the channel hybrid perceptron and the spatial hybrid perceptron, respectively.

2. The prediction method as described in claim 1, characterized in that, The dual-sequence comparison attention module is implemented by creating an attention matrix that represents the spatial and channel interactions between the two sequences. The model then calculates the average of this matrix and applies a sigmoid function to obtain the final attention score. These scores are subsequently used to weight the original sequences to highlight the parts that the model deems important. The specific steps are as follows: For the output of Mixer Block and They are then transformed into attention vectors using a multilayer perceptron. and ; (3) (4) In the formula, ReLU is a non-linear activation function. and b is the weight matrix; b is the bias vector. Then, an attention matrix is ​​created to represent the interaction between the two sequences in space and channels. : (5) In the formula, This is the weight matrix; After completing the above steps, the attention matrix Successfully constructed; subsequently, averaging was performed on each dimension to obtain the ncRNA sequence attention matrix. and protein sequence attention matrix ; (6) (7) In the formula, It is the Sigmoid activation function, used to map attention scores to the (0,1) range; Mean( , ) used to obtain dimensions Each line The average value; Subsequently, based on the ncRNA sequence attention matrix and protein sequence attention matrix Obtain the feature matrix and : (8) (9) In the formula, This indicates element-wise multiplication; and This indicates the output of the Mixer Block; Finally, the feature matrix and It is connected and fed into the classifier.

3. The prediction method as described in claim 2, characterized in that, The steps are as follows: First, the high-order features extracted from ncRNA, protein sequence features and their derived features through CNN layers are fed into the Bi-LSTM network. Then, the network output is expanded and reshaped to the appropriate dimensions to fit the input of the Mixer Layer, and the MLP-Mixer network is used to further refine the feature representations of protein sequences and ncRNA sequences at a deeper level. Subsequently, the complex interactions between nucleotides and amino acids were simulated in the Attention Moudule to uncover subtle associations between ncRNA and protein sequences, and the corresponding attention scores were calculated. These attention scores were then used to weight the original feature representations, and the weighted summed feature matrices were then concatenated. Finally, the concatenated feature representation is input into a classifier consisting of two fully connected layers and a Softmax function. These layers work together to process the merged high-level features and ultimately generate a binary value as the basis for determining whether the objects interact.

Images