A 2'-o-methylation site prediction method based on multi-modal adaptive graph contrastive learning

Abstract

The present application relates to a kind of 2 '-O-methylation site prediction method based on multi-modal adaptive graph contrast learning, comprising: obtaining RNA molecule, based on RNA molecule, construct multi-modal graph;Local graph structure enhancement and node feature coding are carried out to multi-modal graph, obtain node embedding feature, input adaptive attention fusion model to node embedding feature, obtain 2 '-O-methylation site prediction result;Adaptive attention fusion model is obtained using first training set training;Wherein, adaptive attention fusion model calculates the specific weight of each node in node embedding feature in different modalities by adaptive attention fusion module, fuses node feature using specific weight, carries out classification prediction based on fully connected layer to the node feature after fusion, obtains 2 '-O-methylation site prediction result.The present application further optimizes the prediction efficiency of 2OM site, provides more reliable tool for relevant research.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and artificial intelligence, in particular to a 2'-O-methylation site prediction method based on multi-modal adaptive graph contrast learning. BACKGROUND

[0002] 2'-O-methylation (2OM) is a ubiquitous post-transcriptional modification in RNA molecules, which is widely involved in regulating the stability of RNA, mRNA splicing and translation process, and also has an important influence on innate immune response. Studies have shown that natural 2OM modification can effectively inhibit the secretion of pro-inflammatory factors in human monocytes through immune silencing mechanism, which provides a new perspective for understanding immune regulation.

[0003] At present, a variety of experimental methods have been developed for accurately identifying 2'-O-methylation (2OM) sites in RNA, including perchloric acid (HClO4) hydrolysis, periodic acid oxidation hydrolysis, chromatography and mass spectrometry. Although these experimental methods have achieved certain results in identifying 2OM sites, they all have common technical defects, such as complex operation, time-consuming and laborious, need for expensive professional equipment, and may cause damage to RNA samples. In addition, when the amount of RNA sample is extremely small, it also increases the difficulty of operation, limiting its application in some studies.

[0004] In order to overcome the above limitations, high-throughput technologies based on deep sequencing have been proposed, such as Nm-seq, RiboMeth-seq, 2OMe-seq, RibOxi-Seq and Nm-seq. These technical means can identify 2OM sites at the level of transcriptome, providing more efficient tools. However, these methods still have problems such as high cost, complex operation and need for professional technical support, which limit their wide application. SUMMARY

[0005] In order to solve the problems existing in the prior art, the purpose of the present application is to provide a 2'-O-methylation site prediction method based on multi-modal adaptive graph contrast learning, which can construct a spatial proximity relationship graph, a chemical action relationship graph and a dynamic mechanical information Figure ThreeA modal graph, the local centrality index of the node is used to realize the adaptive enhancement of the node feature, and then a cross-modal node-level graph contrast learning strategy is designed to strengthen the complementarity and discrimination ability between nodes of different modalities. In addition, an adaptive attention fusion mechanism is proposed to automatically learn the weight of different modal graphs, and efficient node feature fusion is realized. The synergistic effect of the above strategies effectively improves the expression ability of the node feature, significantly improves the prediction accuracy and reliability of the 2'-O-methylation site in the RNA sequence, and provides an efficient, stable and easy-to-implement innovative theoretical method and technical means for accurate identification and in-depth study of RNA methylation modification sites.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] A 2'-O-methylation site prediction method based on multi-modal adaptive graph contrast learning, comprising:

[0008] An RNA molecule is obtained, and based on the RNA molecule, a multi-modal graph is constructed; the multi-modal graph includes a spatial proximity graph, a chemical action graph and a dynamic mechanics graph;

[0009] The multi-modal graph is subjected to local graph structure enhancement and node feature encoding to obtain node embedding features, the node embedding features are input into an adaptive attention fusion model to obtain 2'-O-methylation site prediction results; the adaptive attention fusion model is obtained by training with a first training set;

[0010] The adaptive attention fusion model calculates the specific weight of each node in the node embedding features under different modalities through an adaptive attention fusion module, fuses the node features using the specific weight, classifies and predicts the fused node features based on a full connection layer, and obtains the 2'-O-methylation site prediction results.

[0011] Optionally, constructing the spatial proximity graph comprises:

[0012] The three-dimensional coordinates of each nucleotide node in the RNA molecule are obtained, the Euclidean distance between each pair of nucleotide nodes is calculated according to the three-dimensional coordinates, when the Euclidean distance is less than a preset threshold, the each pair of nucleotide nodes is connected to form a spatial proximity edge, and the weight of the spatial proximity edge is calculated:

[0013]

[0014] wherein ω(e spatial ) is the weight of the spatial proximity edge, d ij is the Euclidean distance between each pair of nucleotide nodes, and σ is a target constant;

[0015] Based on the spatial proximity edges and the corresponding weights, the spatial proximity graph is constructed.

[0016] Optionally, constructing the chemical interaction graph comprises:

[0017] Extracting the chemical connection relationship between nucleotides in the RNA molecule, the chemical connection relationship comprising: covalent bond, hydrogen bond, π-π stacking and van der Waals force non-covalent interaction;

[0018] Setting different initial strength values for different chemical interaction types, if the chemical connection relationship exists between the bases, the initial strength value of the corresponding chemical interaction type is given to the bases, and the bases are connected to form chemical interaction edges, and the weights of the chemical interaction edges are calculated:

[0019]

[0020] Wherein, ω(e chemical ) is the weight of the chemical interaction edge, is the covalent bond between base i and base j, is the hydrogen bond between base i and base j, is the π-π stacking interaction between base i and base j, is the van der Waals interaction between base i and base j, d i,j is the Euclidean distance of base i and base j in space;

[0021] Based on the chemical interaction edges and the weights, the chemical interaction graph is constructed.

[0022] Optionally, constructing the dynamic mechanical graph comprises:

[0023] Molecular dynamics simulation is performed on the RNA sequence in the RNA molecule to obtain the three-dimensional conformation of the RNA molecule at different time points;

[0024] According to the three-dimensional conformation, the exchange times of each pair of nucleotide nodes in each time point are obtained when the spatial distance is less than a preset dynamic threshold, and based on the exchange times, the comprehensive interaction frequency is obtained, if the comprehensive interaction frequency is greater than a preset dynamic frequency threshold, the nucleotide nodes are connected to form dynamic mechanical edges, and the weights of the dynamic mechanical edges are calculated:

[0025]

[0026] Wherein, ω(e dynamic ) is the weight of the dynamic mechanical edge, F dyn (i,j) is the comprehensive interaction frequency of base i and base j, F dyn (p,q) is the comprehensive interaction frequency of base p and base q.

[0027] Based on the dynamic mechanics edge and the weight thereof, the dynamic mechanics graph is constituted.

[0028] Optionally, the obtaining the node embedding feature comprises:

[0029] The multi-modal graph is locally enhanced in graph structure, and the enhanced multi-modal graph is input into a graph neural network model for node feature coding to obtain the node embedding feature; the graph neural network model is trained by using a second training set;

[0030] The enhanced multi-modal graph is input into a graph neural network model for node feature coding, comprising:

[0031]

[0032] wherein H (m) is the node embedding feature, G (m) is each modal graph, is the enhanced multi-modal graph.

[0033] Optionally, the multi-modal graph is locally enhanced in graph structure, comprising:

[0034] The node local centrality index in the multi-modal graph is calculated, comprising:

[0035]

[0036] wherein, is the node v i is the set of neighbor nodes in the modal graph m, is the edge weight;

[0037] The nodes in the multi-modal graph are enhanced according to the node local centrality index value, comprising:

[0038]

[0039] wherein, α (m) , β (m) are modal-specific learnable parameters, σ(·) represents a sigmoid function, and represents element-wise multiplication, represents the node local centrality index value.

[0040] Optionally, the graph neural network model comprises:

[0041] The original enhanced modal graph in the second training set is input into an original graph neural network model, and an overall cross-modal graph contrast loss function is set:

[0042]

[0043] wherein, a contrast loss function for the spatial proximity graph features and the chemical interaction graph features, a contrast loss function for the spatial proximity graph features and the dynamic mechanics graph features, a contrast loss function for the chemical interaction graph features and the dynamic mechanics graph features;

[0044] optimizing model parameters of the original graph neural network model to obtain the graph neural network model, taking minimization of the overall cross-modal graph contrast loss function as an objective.

[0045] Optionally, obtaining the fused node features comprises:

[0046] calculating, by the adaptive attention fusion module, a specific weight of each node in the node embedding features under different modalities comprises:

[0047]

[0048] wherein, are modal weights, is a vector representation of node i under m modalities, is a vector representation of node i under n modalities;

[0049] fusing node features by using the specific weight comprises:

[0050]

[0051] wherein, is a fused node feature, is a specific weight of each node under different modalities.

[0052] Optionally, obtaining the adaptive attention fusion model comprises:

[0053] inputting original node embedding features in the first training set into an original adaptive attention fusion model, and optimizing model parameters of the original adaptive attention fusion model by using a classification cross-entropy loss function to obtain the adaptive attention fusion model;

[0054] the classification cross-entropy loss function comprises:

[0055]

[0056] wherein, y i is a node true label, is a predicted probability value.

[0057] Optionally, the method further comprises:

[0058] In the process of training the model by using the first training set and the second training set, a global optimization objective and a training strategy are further adopted to optimize the model parameters:

[0059]

[0060] Wherein, lambda is a balance parameter.

[0061] The present application has the following advantages:

[0062] The present application starts from the three-dimensional molecular hypergraph of RNA sequence, combines the advantages of hypergraph convolution network (HGCN) and spiking neural network (SNN), and proposes a new hybrid deep learning model, "hypergraph-spiking neural network". By comprehensively capturing the static structure and dynamic characteristics of RNA molecules, the prediction efficiency is significantly improved:

[0063] (1) The accuracy and stability of 2'-O-methylation site prediction are improved:

[0064] The present application constructs a multi-modal graph based on spatial proximity relationship, chemical interaction relationship and dynamic mechanical information, and proposes a method of adaptive node feature enhancement and cross-modal graph contrast learning, which effectively improves the expression ability of node features, and significantly improves the accuracy and stability of 2'-O-methylation site prediction in RNA sequence.

[0065] (2) Efficient fusion and complementarity of different modal graph information is realized:

[0066] The present application proposes an adaptive attention mechanism to automatically weigh the information contribution of space, chemistry and dynamics, which can fully exploit the internal complementarity and difference between different modalities, and avoid the defects of insufficient modal information fusion in traditional methods.

[0067] (3) The local structure expression ability and interpretability of node features are enhanced:

[0068] The present application designs an adaptive feature enhancement strategy based on node local centrality, which strengthens the local structure features of key nodes, effectively highlights the importance of nodes, and enhances the biological interpretability of model prediction results. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0070] Figure 1This is a flowchart of a 2′-O-methylation site prediction method based on multimodal adaptive graph contrastive learning according to an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram comparing the AUROC curves of HS2OMNet and other comparison methods on the Am, Um, Cm, and Gm datasets according to an embodiment of the present invention.

[0072] Figure 3 This is a schematic diagram comparing the PRROC curves of HS2OMNet and other comparative methods on the Am, Um, Cm, and Gm datasets according to an embodiment of the present invention.

[0073] Figure 4 This is a schematic diagram comparing the AUROC and PRROC curves of HS2OMNet and other comparison methods on the ALLm dataset according to an embodiment of the present invention.

[0074] Figure 5 This is a schematic diagram showing the comparison of MCC values ​​between HS2OMNet and other comparison methods on the ALLm dataset according to an embodiment of the present invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0077] like Figure 1 As shown, this embodiment discloses a method for predicting 2′-O-methylation sites based on multimodal adaptive graph contrastive learning, including: obtaining RNA molecules; constructing a multimodal graph based on RNA molecules; the multimodal graph includes: a spatial proximity graph, a chemical interaction graph, and a dynamic mechanical graph; performing local graph structure enhancement and node feature encoding on the multimodal graph to obtain node embedding features; inputting the node embedding features into an adaptive attention fusion model to obtain 2′-O-methylation site prediction results; the adaptive attention fusion model is trained using a first training set; wherein, the adaptive attention fusion model calculates the specific weights of each node in the node embedding features under different modalities through an adaptive attention fusion module, fuses the node features using the specific weights, and performs classification prediction on the fused node features based on a fully connected layer to obtain 2′-O-methylation site prediction results.

[0078] The embodiment discloses a 2'-O-methylation site prediction method based on multi-modal adaptive graph contrast learning, which improves the accuracy and reliability of 2'-O-methylation (2OM) site prediction in RNA. The innovative method has wide potential application in the field of RNA-related research and bioinformatics, and the specific implementation process is Figure 1 .

[0079] Further, the collection and preprocessing of the data set:

[0080] Integrate Nm site information in human RNA from multiple data sources verified by experiments, including RMBase v3.0, Nm-seq, Nm-Mut-seq and Nm-REP-seq databases. Set the similarity threshold to 0.6 by CD-HIT software, delete sequences with similarity greater than 0.6, and finally obtain 3648 2OM sequences (i.e. positive samples) and 15402 non-2OM sequences (i.e. negative samples). In addition, since the data contains modification of 4 bases (A, U, C and G), the data set can be divided into 4 sub-sets. In addition, the data set is divided into training set and test set by random segmentation, with training set accounting for 80% and test set accounting for 20%, and the specific number of each data set is shown in Table 1. Am represents A base modification 2OM subset, Um represents U base modification 2OM subset, Cm represents C base modification 2OM subset, Gm represents G base modification 2OM subset, and Allm represents the sum of Am, Um, Cm and Gm.

[0081] Table 1 Specific number of each data set

[0082]

[0083] Further, the feature encoding construction method:

[0084] The RNA sequence is constructed into a multi-modal graph data structure from the base space position, the chemical bond action between bases, and the dynamic process between bases.

[0085] Specifically, the spatial proximity graph G is constructed from the base space position based on the RNA sequence spatial , wherein the node feature is representing the spatial coordinates of base i. The chemical action graph G is constructed based on the interaction relationship between bases in the RNA sequence chemical , wherein the node feature is representing the chemical bond strength of base i, wherein covalent is covalent bond, hydrogen is hydrogen bond, π-π is π-π stacking, and vdw is van der Waals force. Finally, the dynamic mechanics graph G is constructed based on the dynamics of bases in the RNA sequence dynamic , wherein the node feature is F iThis represents the probability that base i is touched.

[0086] Furthermore, constructing a spatial proximity graph includes: obtaining the three-dimensional coordinates of each nucleotide node in the RNA molecule; calculating the Euclidean distance of each pair of nucleotide nodes based on the three-dimensional coordinates; connecting each pair of nucleotide nodes to form a spatial proximity edge when the Euclidean distance is less than a preset threshold; calculating the weight of the spatial proximity edge; and constructing a spatial proximity graph based on the spatial proximity edge and its corresponding weight.

[0087] Specifically, the method for constructing spatially adjacent edges:

[0088] Spatial proximity edges are used to represent the spatial relationships between adjacent nucleotides in the three-dimensional structure of RNA. First, the three-dimensional coordinates (x, y, z) of each nucleotide in the RNA molecule are obtained; then, the spatial proximity edges for each pair of nucleotide nodes (v...) are calculated. i ,v j The Euclidean distance d between them ij , Secondly, when the distance d ij Less than the preset threshold At that time, the nucleotide node v i ,v j Connect, forming a spatially adjacent edge e spatial The formula for calculating the weight of the last spatial neighbor edge is as follows: σ = 3.0; where, In atomic scale, representing 10 -10 rice.

[0089] Furthermore, constructing the chemical interaction graph involves: extracting the chemical linkages between nucleotides in RNA molecules, including covalent bonds, hydrogen bonds, π-π stacking, and van der Waals forces (non-covalent interactions); setting different initial strength values ​​for different types of chemical interactions; if there is a chemical linkage between bases, assigning the corresponding initial strength value to the bases and connecting the bases to form chemical interaction edges; calculating the weights of the chemical interaction edges; and constructing the chemical interaction graph based on the chemical interaction edges and their weights.

[0090] Specifically, the method for constructing chemical interaction edges:

[0091] Chemical interactions represent the chemical interactions within RNA molecules. First, the chemical connections between nucleotides are extracted from the RNA 3D structure file, including covalent bonds, hydrogen bonds, π-π stacking, and van der Waals forces (VDW) non-covalent interactions. Then, a specific initial strength value is assigned to each type of chemical interaction. If base i and base j are covalently bonded, i.e.:

[0092] 1.0 if base i and base j have hydrogen bond, otherwise 0; 0.7 if base i and base j have π-π stacking, otherwise 0; 0.5 if base i and base j have van der Waals interaction, otherwise 0; 0.3 if base i and base j have π-π stacking, otherwise 0. Finally, the weight of the chemical interaction edge is normalized by distance:

[0093]

[0094] where, x∈[covalent, hydrogen, π-π, vdw] is the intensity value of the above corresponding, d i,j is the Euclidean distance of base i and base j in space.

[0095] Further, the constructing the dynamic graph comprises: performing molecular dynamics simulation on the RNA sequence in the RNA molecule to obtain the three-dimensional conformation of the RNA molecule at different time points; obtaining the exchange frequency of each pair of nucleotide nodes in each time point according to the three-dimensional conformation, the exchange frequency being the number of exchanges when the spatial distance is less than a preset dynamic threshold; connecting the nucleotide nodes based on the exchange frequency to form a dynamic mechanical edge and calculate the weight of the dynamic mechanical edge if the comprehensive interaction frequency is greater than a preset dynamic frequency threshold; and constructing the dynamic graph based on the dynamic mechanical edge and the weight thereof.

[0096] Specifically, the constructing method of the dynamic mechanical edge comprises:

[0097] The dynamic mechanical edge captures the dynamic changes of the RNA molecule in the time dimension. First, the molecular dynamics simulation is performed on the RNA sequence to obtain the three-dimensional conformation of the RNA molecule at N different time points (t=1, 2, …, N); then, for each time point t, the interaction frequency of each pair of nucleotide nodes (v i ,v j ) when the spatial distance is less than a dynamic threshold is counted, denoted as f dyn (i,j,t); secondly, the interaction frequency of all time points is accumulated to obtain the comprehensive interaction frequency ; finally, if the accumulated interaction frequency F dyn (i,j) of 100 times exceeds a dynamic frequency threshold, a dynamic mechanical edge is constructed between the nucleotide nodes v i ,v j and the weight w

[0098] Finally, the weight of the dynamic mechanical edge is calculated according to the weight of the spatial proximity edge ω(e spatial ) and the weight of the chemical interaction edge ω(echemical ) and dynamic mechanics edge weight ω(e dynamic ), the spatial proximity graph G spatial , the chemical interaction graph G chemical , and the dynamic mechanics graph G dynamic are constructed by a minimum connected graph construction algorithm, respectively.

[0099] Further, the node embedding feature acquisition includes: performing local graph structure enhancement on the multi-modal graph, inputting the enhanced multi-modal graph into a graph neural network model for node feature coding to obtain the node embedding feature; and the graph neural network model is trained by using the second training set.

[0100] Further, the local graph structure enhancement on the multi-modal graph includes: calculating a node local centrality index in the multi-modal graph; and enhancing the nodes in the multi-modal graph according to the node local centrality index value.

[0101] Further, the graph neural network model includes: inputting the original enhanced multi-modal graph in the second training set into an original graph neural network model, setting an overall cross-modal graph contrast loss function, optimizing the model parameters of the original graph neural network model to obtain the graph neural network model.

[0102] Specifically, the model is constructed as follows:

[0103] The present application proposes a method based on multi-modal adaptive graph contrastive learning (MAGCL) to realize the accurate prediction of 2'-O-methylation sites in RNA sequences. Specifically, the MAGCL method includes the following four main steps:

[0104] 1) Adaptive enhancement strategy of multi-modal graph (Adaptive Graph Augmentation):

[0105] To highlight the importance of key nodes in each modal graph and reduce the interference of redundant information, the present application proposes an adaptive node feature enhancement strategy:

[0106] First, for each modal graph G (m) (m∈{spatial,chemical,dynamic}), the node local centrality index The degree centrality of the node is taken as the node importance index wherein is the neighbor node set of node v i in the modal graph m, is the edge weight. Then, the original node features are enhanced based on the node centrality values: where α (m) , β (m) are modal-specific learnable parameters, which are set to 0.5 respectively, and σ(·) denotes the sigmoid function, and denotes element-wise multiplication.

[0107] 2) Cross-modal Graph Contrastive Learning:

[0108] To further exploit the complementarity and difference of node representations between different modal graphs, the application proposes a cross-modal node-level contrastive learning method.

[0109] First, the graph neural network GNN is used to encode the node features of the enhanced modal graph, respectively, to obtain the node embedding representation, For any two different modal graphs m, n ∈ {spatial, chemical, dynamic}, the node-level cross-modal contrastive loss function is defined as:

[0110]

[0111] where, denotes the cosine similarity function, and τ is the temperature hyperparameter.

[0112] The overall cross-modal graph contrastive loss function is defined as the sum of the contrastive loss between the three modalities:

[0113]

[0114] By minimizing the above loss, the model forces the same node to have more similar representations in different modalities, and the representations of different nodes to have greater differences, significantly improving the expression ability of node features.

[0115] Further, obtaining the adaptive attention fusion model comprises: inputting the original node embedding features in the first training set into the original adaptive attention fusion model, and using the classification cross-entropy loss function to optimize the model parameters of the original adaptive attention fusion model, to obtain the adaptive attention fusion model.

[0116] Specifically, adaptive attention multi-modal feature fusion and 2'-O-methylation site prediction:

[0117] To efficiently fuse each modal information, the application proposes an adaptive attention fusion module, which automatically learns the importance of modalities and fuses according to the modal embedding representation of each node:

[0118] First, a multi-layer perceptron (MLP) structure is defined to calculate the modality weight: MLP(x) = W2(ReLU(W1x + b1)) + b2, where W1, W2, b1, b2 are learnable parameters, and ReLU is an activation function. Then the specific importance weight of node i in modality m is calculated:

[0119]

[0120] Then, the fused node feature representation is defined as:

[0121]

[0122] Finally, the classification prediction of each node (base) is made through the fully connected layer (Linear) and the softmax function: where W and b are the learnable parameters of the Linear layer, and the classification cross-entropy loss function is defined as:

[0123]

[0124] where y i is the true label of the node (0 or 1), is the predicted probability value.

[0125] Further, the method further comprises: in the process of training the model by using the first training set and the second training set, further adopting an overall optimization target and a training strategy to optimize the model parameters.

[0126] Specifically, the overall optimization target and the training strategy are:

[0127] The overall optimization target of the model is the combination of the classification loss and the cross-modal contrast loss:

[0128]

[0129] where λ is a balance parameter, set to 0.5, the Adam optimizer is used for parameter optimization in the training stage, the learning rate is initially set to 0.001, the batch data size is 32, and when the loss function is less than 0.0001 or the training iteration number exceeds 256, the training model is stopped.

[0130] Model evaluation and verification of the invention:

[0131] The model of the present application is compared with a plurality of 2'-O-methylation (2OM) site prediction models, and significant performance improvement is achieved in a plurality of evaluation indexes. The evaluation indexes include specificity (Specificity, SEP), sensitivity (Sensitivity, SEN), accuracy (Accuracy, ACC), Matthews correlation coefficient (Matthews correlation coefficient, MCC), F1-score (F1), area under the curve (AUC), and average accuracy (Average precise, AP).

[0132] (4) Experimental comparison results of HS2OMNet and existing methods

[0133] The performance of the present application HS2OMNet is compared with H2Opred, i2OM, Meta-2OM and Nmix; HS2OMNet and H2Opred, i2OM, Meta-2OM and Nmix are respectively trained in Am, Um, Gm, Cm and ALLm training data sets, and tested in the corresponding test data sets. The corresponding comparison results are shown in Tables 2, 3, 4, 5 and 6.

[0134] Table 2 Comparison of evaluation indexes of HS2OMNet and other comparison models on Am data set

[0135]

[0136] Table 3 Comparison of evaluation indexes of HS2OMNet and other comparison models on Um data set

[0137]

[0138] Table 4 Comparison of evaluation indexes of HS2OMNet and other comparison models on Gm data set

[0139]

[0140] Table 5 Comparison of evaluation indexes of HS2OMNet and other comparison models on Cm data set

[0141]

[0142] Table 6 Comparison of evaluation indexes of HS2OMNet and other comparison models on ALLm data set

[0143]

[0144] The performance of HS2OMNet was benchmarked against four baseline models (H2Opred, i2OM, Meta-2OM, Nmix) on five different datasets (Am, Um, Gm, Cm, and the aggregated ALLm). Evaluation was based on seven standard metrics: Sensitivity (SE), Specificity (SP), Accuracy (ACC), F1-score (F1), Matthews Correlation Coefficient (MCC), Area Under the ROC Curve (AUC), and Average Precision (AP). Quantitative results are detailed in Tables 2-6.

[0145] On all evaluated datasets, HS2OMNet consistently achieved the highest scores on the comprehensive performance metrics ACC, F1, and MCC. On the ALLm dataset (Table 6), HS2OMNet recorded an ACC of 0.9318, an F1 of 0.9318, and a MCC of 0.8706. These values surpassed the highest scores achieved by any baseline model on this dataset (ACC≤0.9024, F1≤0.9020, MCC≤0.8176). Similar top rankings for HS2OMNet on ACC, F1, and MCC were observed on the individual Am, Um, Gm, and Cm datasets (Tables 2-5).

[0146] Evaluation of discriminative ability using AUC and AP also demonstrated the performance advantage of HS2OMNet. This model produced the highest AUC and AP values on all five datasets. On the Am dataset (Table 2), HS2OMNet's AUC was 0.9397 and AP was 0.8696, surpassing the suboptimal baseline model values (AUC≤0.8821, AP≤0.7484). This leading trend for HS2OMNet on AUC and AP persisted across the Um, Gm, Cm, and ALLm datasets.

[0147] Regarding the identification of balance between positive and negative classes, HS2OMNet exhibited high performance on both SE and SP. It consistently maintained near-top or top SE values across datasets (Am dataset SE = 0.9989, Table 2; ALLm dataset SE = 0.9920, Table 6). More notably, HS2OMNet reported the highest SP values among all models on each evaluated dataset (Am dataset SP = 0.9095, Table 2; ALLm dataset SP = 0.8787, Table 6), indicating a stronger ability to correctly identify true negative instances compared to the baselines.

[0148] The performance advantage of HS2OMNet is consistent across individual datasets (Am, Um, Gm, Cm) and the aggregated ALLm dataset. This consistent ranking across multiple data partitions indicates that, under test conditions, the HS2OMNet model exhibits superior robustness and generalization ability compared to the compared methods. In summary, the empirical results in Tables 2-6 demonstrate that, on the test dataset, HS2OMNet outperforms H2Opred, i2OM, Meta-2OM, and Nmix on all evaluation metrics. Its corresponding AUROC and PRROC are as follows: Figure 2 As shown in Figures 3 and 4.

[0149] Furthermore, to further demonstrate the performance of HS2OMNet on the entire dataset, the ALLm dataset was randomly divided into training and test sets with different seeds in an 8:2 ratio. Using the same training set, the MCC values ​​of HS2OMNet and the comparison method were compared on the test set. The MCC value of the comparison method (MCC1) was used as the x-axis, and the MCC value of HS2OMNet (MCC2) was used as the y-axis. Plotting the coordinates on the axis showed that HS2OMNet performed better when the coordinates were on the right diagonal, and better when they were below the right diagonal. This method was repeated 100 times, and the performance of HS2OMNet was observed. Figure 5 As shown, from Figure 5 It can be observed that the MCC value obtained by HS2OMNet is still greater than that of other comparative methods under different data partitions, which further illustrates that the HS2OMNet method is superior to other comparative methods in predicting the anti-stress peptide problem.

[0150] (5) Performance robustness verification of the HS2OMNet model:

[0151] To further evaluate the performance stability and robustness of HS2OMNet compared to baseline methods under different data partitioning conditions, 100 repeated random subsampling cross-validations were performed on the ALLm dataset. In each iteration, the ALLm dataset was divided into training and test sets at an 80%:20% ratio using independent random seeds. To ensure fairness, the training set generated in each iteration was used to train HS2OMNet and all comparison models, and the Matthews correlation coefficient (MCC) of each model was evaluated on the corresponding test set. For visualization, a scatter plot was created using the MCC value (MCC2) obtained by HS2OMNet in each iteration as the ordinate and the MCC value (MCC1) of each baseline method in the corresponding iteration as the abscissa. In this coordinate system, if a data point is above the diagonal (MCC1, MCC2), it indicates that HS2OMNet outperforms the comparison method in that data partitioning; conversely, if it is below the diagonal, the comparison method performs better.Figure 5 All pairwise comparison results from 100 iterations are shown. As Figure 5 shown, the vast majority of data points are significantly distributed in the upper region of the main diagonal. This result clearly indicates that, in the vast majority of random data split scenarios, the MCC values obtained by HS2OMNet are consistently higher than all the compared baseline methods. It shows that the HS2OMNet model not only has superior performance in the prediction of 2'-O-methylation tasks, but also shows good robustness and consistency relative to the changes of data division.

[0152] The above-described embodiments are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning, characterized in that, The method comprises the following steps: obtaining an RNA molecule, and constructing a multi-modal graph based on the RNA molecule; the multi-modal graph comprises a spatial proximity graph, a chemical interaction graph, and a dynamic mechanics graph; performing local graph structure enhancement and node feature coding on the multi-modal graph to obtain node embedding features, inputting the node embedding features into an adaptive attention fusion model to obtain a 2'-O-methylation site prediction result; the adaptive attention fusion model is trained by using a first training set; wherein the adaptive attention fusion model calculates a specific weight of each node in the node embedding features under different modalities through an adaptive attention fusion module, fuses the node features by using the specific weight, classifies and predicts the fused node features based on a full connection layer to obtain the 2'-O-methylation site prediction result.

2. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning according to claim 1, characterized in that, constructing the spatial proximity graph comprises the following steps: obtaining the three-dimensional coordinates of each nucleotide node in the RNA molecule, calculating the Euclidean distance between each pair of nucleotide nodes according to the three-dimensional coordinates, connecting each pair of nucleotide nodes when the Euclidean distance is less than a preset threshold to form a spatial proximity edge, and calculating the weight of the spatial proximity edge: where ω(e spatial ) is the weight of the spatially adjacent edge, d ij is the Euclidean distance between each pair of nucleotide nodes, and σ is a target constant. based on the spatial proximity edge and the corresponding weight, the spatial proximity graph is constructed.

3. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning according to claim 1, characterized in that, constructing the chemical interaction graph comprises the following steps: extracting the chemical connection relationship between nucleotides in the RNA molecule, the chemical connection relationship comprising covalent bond, hydrogen bond, π-π stacking, and van der Waals force non-covalent interaction; setting different initial intensity values for different chemical interaction types, if the chemical connection relationship exists between the bases, assigning the initial intensity value of the corresponding chemical interaction type to the bases, connecting the bases to form a chemical interaction edge, and calculating the weight of the chemical interaction edge: wherein ω(e chemical ) is the weight of the chemical interaction edge, is the presence of a covalent bond between base i and base j, is the presence of a hydrogen bond between base i and base j, is the presence of a π-π stacking interaction between base i and base j, is the presence of a van der Waals interaction between base i and base j, d i,j is the Euclidean distance in space between base i and base j; based on the chemical interaction edge and the weight, the chemical interaction graph is constructed.

4. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 1, wherein, constructing the dynamic mechanics graph comprises the following steps: performing molecular dynamics simulation on the RNA sequence in the RNA molecule to obtain the three-dimensional conformation of the RNA molecule at different time points; obtaining the exchange frequency of each pair of nucleotide nodes in each time point when the spatial distance is less than a preset dynamic threshold according to the three-dimensional conformation, obtaining the comprehensive interaction frequency based on the exchange frequency, connecting the nucleotide nodes when the comprehensive interaction frequency is greater than a preset dynamic frequency threshold to form a dynamic mechanics edge, and calculating the weight of the dynamic mechanics edge: where ω(e dynamic ) is the weight of the dynamic force, F dyn (i,j) is the comprehensive interaction frequency of base i and base j, F dyn (p,q) is the comprehensive interaction frequency of base p and base q; based on the dynamic mechanics edge and the weight, the dynamic mechanics graph is constructed.

5. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 1, wherein, obtaining the node embedding features comprises the following steps: performing local graph structure enhancement on the multi-modal graph, inputting the enhanced multi-modal graph into a graph neural network model for node feature coding to obtain the node embedding features; the graph neural network model is trained by using a second training set; inputting the enhanced multi-modal graph into the graph neural network model for node feature coding comprises the following steps: Among them, H (m) G is the node embedding feature. (m) For each modality diagram, This is the enhanced multimodal graph.

6. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 5, wherein, performing local graph structure enhancement on the multi-modal graph comprises the following steps: calculating the node local centrality index in the multi-modal graph: wherein, is a node v i a set of neighbor nodes in the modality graph m, is an edge weight; enhancing the nodes in the multi-modal graph according to the node local centrality index value: where α (m) , β (m) are modal-specific learnable parameters, σ(·) denotes the sigmoid function, ⊙ denotes element-wise multiplication, denotes the node local centrality indicator value.

7. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 5, wherein, obtaining the graph neural network model comprises the following steps: inputting original enhanced modality graphs in the second training set into an original graph neural network model, and setting an overall cross-modality graph contrast loss function: wherein, a contrast loss function for the spatial proximity graph features and the chemical interaction graph features, a contrast loss function for the spatial proximity graph features and the dynamic mechanics graph features, a contrast loss function for the chemical interaction graph features and the dynamic mechanics graph features; optimizing model parameters of the original graph neural network model with the overall cross-modality graph contrast loss function as a target, and obtaining the graph neural network model.

8. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 1, wherein, obtaining the fused node features includes: calculating specific weights of each node in different modalities in the node embedding features by the adaptive attention fusion module includes: wherein, are modal weights of the modality, is a vector representation of node i in the m-th modality, is a vector representation of node i in the n-th modality; fusing the node features by the specific weights includes: wherein, is the fused node feature, is a specific weight for each node under different modalities.

9. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 1, wherein, obtaining the adaptive attention fusion model includes: inputting original node embedding features in the first training set into an original adaptive attention fusion model, and optimizing model parameters of the original adaptive attention fusion model by a classification cross-entropy loss function, to obtain the adaptive attention fusion model; the classification cross-entropy loss function includes: where y i is the node true label, is the predicted probability value.

10. The 2'-O-methylation site prediction method based on multi-modal adaptive graph contrastive learning of claim 5, wherein, the method further includes: in the process of training the model by using the first training set and the second training set, an overall optimization target and a training strategy are further adopted to optimize the model parameters: wherein λ is a balance parameter.

Images