Esophageal cancer RNA-protein interaction space prediction method and device based on graph neural network

Abstract

The invention relates to the technical field of bioinformatics, in particular to an esophageal cancer RNA-protein interaction space prediction method and device based on a graph neural network, and can solve the problem that it is difficult to accurately distinguish the difference of cancer tissue and normal tissue on RPI in a traditional method to a certain extent. The method comprises the steps of obtaining multi-stage space transcriptome data related to esophageal cancer, and performing preprocessing to obtain a multi-modal feature tensor and a canceration stage tag; using the multi-modal feature tensor and the canceration stage tag, using RNA and protein as nodes, defining a static edge based on sequence complementarity and structure matching degree, defining a dynamic edge based on spatial co-expression correlation, and constructing a space-time dynamic graph updated along with the canceration stage; the space-time dynamic graph serves as input, interaction existence, site coordinates and binding energy are predicted through space-time double-branch coding and attention mechanism fusion features, and an interaction prediction result is output.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology, and more specifically, to a method and apparatus for spatial prediction of esophageal cancer RNA-protein interactions based on graph neural networks. Background Technology

[0002] Esophageal cancer is a major disease threatening human health worldwide. Abnormal remodeling of RNA-protein interaction (RPI) networks plays a key role in the malignant progression of esophageal cancer. For example, abnormal binding of long non-coding RNAs to oncogenic proteins can drive epithelial-mesenchymal transition in esophageal squamous cell carcinoma; dysregulation of the interaction between circular RNAs and RNA-binding proteins can help the tumor microenvironment achieve immune escape. Therefore, accurately analyzing the spatiotemporal dynamic characteristics of esophageal cancer-specific RPIs is crucial for revealing the molecular mechanisms of carcinogenesis and screening therapeutic targets.

[0003] Currently, the closest existing technologies for predicting RNA-protein interactions (RPI) in esophageal cancer mainly revolve around traditional molecular docking methods. These methods predict the interaction between RNA and protein by constructing static three-dimensional structural models. For example, in some early studies, researchers used classic molecular docking algorithms to simulate the possible binding modes of RNA and protein based on their structural complementarity, thereby inferring potential RPIs.

[0004] At the data application level, existing technologies are also trying to integrate multiple types of data. For example, some studies combine the primary sequence information of RNA with the amino acid sequence of proteins, and use bioinformatics algorithms to analyze the association between sequences and then predict RPI. After the development of spatial transcriptomics technology, some studies have also tried to integrate the gene expression data obtained from it with traditional molecular structure data, attempting to explore more clues related to RPI from the perspective of the spatial distribution of gene expression.

[0005] However, these existing technologies have many obvious shortcomings. Traditional molecular docking methods only focus on the static structure of RNA and proteins, completely ignoring the key factor that the intensity and location of RPIs change significantly in the multi-stage evolution of esophageal cancer from normal mucosa to invasive carcinoma. In terms of data fusion, although there have been attempts to integrate different types of data, key information such as RNA modification, dynamic changes in protein conformation, and cell localization have not been effectively integrated, resulting in many key RPIs being missed. Furthermore, existing algorithms generally lack the ability to effectively encode the specific characteristics of the esophageal cancer microenvironment, making it difficult to accurately distinguish the differences in RPIs between cancerous and normal tissues, which greatly limits the in-depth understanding of the pathogenesis of esophageal cancer and the accurate screening of related therapeutic targets. Summary of the Invention

[0006] To address the problems of traditional methods failing to capture spatiotemporal dynamic features, lacking sufficient data fusion, and lacking effective encoding capabilities for specific features of the esophageal cancer microenvironment, making it difficult to accurately distinguish the differences in RPI between cancerous and normal tissues, this application provides a spatial prediction method and device for esophageal cancer RNA-protein interaction based on graph neural networks.

[0007] The embodiments of this application are implemented as follows:

[0008] In a first aspect, this application provides a spatial prediction method for esophageal cancer RNA-protein interactions based on graph neural networks, including:

[0009] We acquired multi-stage spatial transcriptome data related to esophageal cancer and preprocessed it to obtain multimodal feature tensors and cancer stage labels.

[0010] Using the multimodal feature tensor and the cancer stage label, with RNA and protein as nodes, static edges are defined based on sequence complementarity and structural matching degree, and dynamic edges are defined based on spatial co-expression correlation, to construct a spatiotemporal dynamic graph in which the edge weights are updated with the cancer stage.

[0011] Using the spatiotemporal dynamic graph as input, the interaction existence, site coordinates, and binding energy are predicted by fusing features through spatiotemporal bi-branch coding and attention mechanism, and the interaction prediction results are output.

[0012] In one possible implementation, the acquisition of multi-stage spatial transcriptome data related to esophageal cancer, and the preprocessing to obtain multimodal feature tensors and cancer stage labels, further includes:

[0013] Input RNA data, protein data, spatial data, and tag data. The RNA data includes esophageal cancer tissue RNA-seq, RNA secondary structure prediction, and m5C modification sites. The protein data includes quantitative proteomics data and AlphaFold2 predicted three-dimensional structures. The spatial data includes Xenium spatial transcriptome data. The tag data includes known RNA-protein interactions.

[0014] Feature extraction was performed on RNA data, protein data, and spatial data respectively;

[0015] Output a multimodal feature tensor and cancer stage labels. The multimodal feature tensor includes an RNA feature matrix, a protein feature matrix, and a spatial feature matrix. The cancer stage labels are t∈{1,2,3,4}, corresponding to 4 levels of cancer progression.

[0016] In one possible implementation, the feature extraction of RNA data, protein data, and spatial data, respectively, further includes:

[0017] Sequence-level, structural-level, and modification-level feature extraction were performed on the RNA data, wherein:

[0018] Sequence level: One-Hot coding + 3-mer frequency statistics to capture nucleotide sequence patterns;

[0019] Structural level: RNAfold calculates the free energy of secondary structures, and the stem-loop structure is encoded as an adjacency matrix;

[0020] Modification level: DeepM6A predicts m5C methylation sites as a functional regulatory feature;

[0021] Sequence-level and structural-level feature extraction is performed on the protein data, wherein:

[0022] Sequence level: ConSurf evolutionary conservation score, PFAM functional domain annotation;

[0023] Structural level: AlphaFold2 coordinate PCA dimensionality reduction, retaining the top 50 principal components, DSSP secondary structure proportion;

[0024] The spatial data is used to extract location features and micro-environment features, wherein:

[0025] Localization characteristics: single-cell three-dimensional coordinates, subcellular localization;

[0026] Microenvironment characteristics: proportion of cell types within a 50μm radius, Spearma coefficient matrix of gene co-expression.

[0027] In one possible implementation, the step of utilizing the multimodal feature tensor and the cancer stage label, using RNA and protein as nodes, defining static edges based on sequence complementarity and structural matching, and defining dynamic edges based on spatial co-expression correlation to construct a spatiotemporal dynamic graph that updates with the cancer stage, further includes:

[0028] Input RNA feature matrix, protein feature matrix, spatial feature matrix, and cancer stage labels;

[0029] Define RNA nodes and protein nodes. The RNA nodes integrate RNA sequence, modification, spatial coordinates and stage information, while the protein nodes integrate protein structure, functional domains, location and stage information.

[0030] Static edge generation: Based on sequence complementarity, the PWM matching score between RNA binding motif and protein binding domain is calculated. When the threshold is ≥0.8, initial edges are generated. Then, based on the structural matching degree, the docking free energy is predicted by ClusPro, and the first 10% of the initial edges are retained as static edges.

[0031] Dynamic edge generation: Based on spatial co-expression correlation, the Spearman correlation coefficient between RNA expression level and protein abundance at carcinogenesis stage t is calculated. Only RNA-protein pairs with Spearman coefficient > 0.5 are retained to generate edges. The initial weight is the coefficient value, which is dynamically adjusted over time. The edge weight changes with the carcinogenesis stage t.

[0032] The output is a spatiotemporal dynamic graph, whose node set carries multimodal features and whose edge set includes static and dynamic edges.

[0033] In one possible implementation, the step of taking the spatiotemporal dynamic graph as input, fusing features through spatiotemporal bi-branch coding and an attention mechanism, predicting interaction existence, site coordinates, and binding energy, and outputting interaction prediction results, further includes:

[0034] The spatiotemporal dynamic graph is input into a dynamic spatiotemporal graph neural network model, which includes a spatial branch, a temporal branch, a spatiotemporal attention fusion module, and a multi-task prediction head.

[0035] The spatial branch outputs spatial features, and the temporal branch outputs temporal features.

[0036] The spatiotemporal attention fusion module outputs fused features by calculating the interaction weights between the spatial features and the temporal features;

[0037] The multi-task prediction module includes a classification task and a regression task. Based on the fusion feature, the classification task outputs the interaction probability through the Sigmoid function, and the loss function is the binary cross-entropy. Based on the fusion feature, the regression task outputs three-dimensional coordinates through the Transformer decoder, and the loss function is the mean squared error.

[0038] Output interaction prediction results, which include an existence probability matrix and a set of interaction site coordinates.

[0039] In one possible implementation, the spatial branch comprises a two-layer graph convolutional neural network:

[0040] The first layer is at the residue level. The atomic coordinates of RNA nucleotides and protein amino acids are input. Through a graph attention network, the van der Waals forces and hydrogen bond interactions between atoms are calculated to generate the interaction scores of residue pairs. Then, the features of neighboring atoms are aggregated according to the attention weights.

[0041] The second layer is at the molecular level. It inputs RNA / protein molecular features aggregated at the residue level, and generates the overall molecular embedding through GraphSAGE mean aggregation. It captures the matching degree between RNA secondary structure domains and protein functional domains, generates spatial proximity tags, and aggregates cellular neighborhood features.

[0042] The time branch input cancer stage sequence is encoded one-hot and then input into a gated recurrent unit. The stage evolution time dependence is captured by the hidden state update to obtain stage-specific features.

[0043] In one possible implementation, after outputting the interaction prediction results, the following is also included:

[0044] Based on the interaction prediction results, molecular dynamics simulations were performed, graph neural networks were used to predict the initial conformation, the structural stability was optimized through electrostatic interactions, and key regulatory pathways were analyzed.

[0045] In one possible implementation, based on the prediction results, molecular dynamics simulations are performed, a graph neural network is used to predict the conformation and generate the initial conformation, structural stability is optimized through electrostatic interactions, and key regulatory pathways are analyzed, further including:

[0046] The interaction prediction results are input into the molecular kinematics optimization layer for optimization.

[0047] Initial conformations were generated. Based on the predicted sites, rigid docking was performed using ClusPro2.0 to generate multiple initial conformations. The initial conformations were then subjected to GROMACS energy minimization until the gradient was <1000kJ / mol / nm.

[0048] Self-consistent field neural network electrostatic optimization was performed to accelerate long-range electrostatic interaction calculations. Atomic coordinates and charges were input, and electrostatic potential was output. The initial conformation was optimized iteratively to stabilize the complex binding free energy ΔG at -9±0.5kcal / mol.

[0049] Molecular dynamics trajectory simulations were performed using CHARMM36 as the force field, TIP3P as the solvent, with a simulation duration of 100 ns, a step size of 2 fs, and the trajectory was saved every 10 ps. The analytical parameters included root mean square deviation, root mean square fluctuation, and number of contact residues.

[0050] The neural relational reasoning model is used to analyze the regulatory pathways and output a mechanism verification report, which includes a stable conformation file and a list of key regulatory pathways.

[0051] In one possible implementation, the use of a neural relational reasoning model to parse regulatory pathways and output a mechanism verification report, the mechanism verification report including a stable conformation file and a list of key regulatory pathways, further including:

[0052] Trajectory feature extraction: For the trajectory generated by molecular dynamics trajectory simulation, the dynamic features of the interaction points are extracted and encoded as node states. The dynamic features include root mean square deviation, root mean square fluctuation and number of contact residues.

[0053] Relational reasoning: Using graph attention networks to infer dynamic dependencies between nodes, generate time-varying adjacency matrices, and identify key control edges;

[0054] Pathway validation: Enrichment analysis was used to validate the correlation between the inferred pathway and esophageal cancer glycolysis and epithelial-mesenchymal transition pathways, and key pathways with FDR < 0.05 were screened.

[0055] Secondly, this application provides a spatial prediction device for esophageal cancer RNA-protein interactions based on graph neural networks, comprising:

[0056] The data acquisition and processing module is used to acquire multi-stage spatial transcriptome data related to esophageal cancer and perform preprocessing to obtain multimodal feature tensors and cancer stage labels.

[0057] The dynamic spatiotemporal graph construction module is used to construct a spatiotemporal dynamic graph that updates with the cancer stage by utilizing the multimodal feature tensor and the cancer stage label, with RNA and protein as nodes, defining static edges based on sequence complementarity and structural matching degree, and defining dynamic edges based on spatial co-expression correlation.

[0058] The model calculation module, using the spatiotemporal dynamic graph as input, fuses features through spatiotemporal bi-branch coding and attention mechanism to predict the existence of interactions, site coordinates, and binding energy, and obtains the interaction prediction results.

[0059] The technical solution provided in this application can achieve at least the following beneficial effects:

[0060] This application provides a spatial prediction method for esophageal cancer RNA-protein interaction (RPI) based on graph neural networks. By integrating multi-stage spatial transcriptome data, constructing a dynamic graph structure, and incorporating a spatiotemporal attention mechanism, it achieves accurate capture of the spatiotemporal dynamic features of RPI. A three-level feature encoding system is established, combined with a hierarchical graph neural network, to complete deep fusion and feature extraction of multimodal data. Molecular dynamics simulation optimization is combined to enhance the learning of esophageal cancer-specific features, analyze the RPI regulatory mechanism, and overcome the limitations of traditional methods. This method achieves in-depth analysis of esophageal cancer RPI from the structural level to the spatiotemporal functional mechanism, providing an innovative and effective technical solution for esophageal cancer pathogenesis research, therapeutic target discovery, and clinical precision medicine. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic flowchart illustrating an exemplary embodiment of the present application of a spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks;

[0063] Figure 2 This is a schematic diagram illustrating the overall implementation process of another spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks, as shown in an exemplary embodiment of this application.

[0064] Figure 3 This is a schematic diagram illustrating the process of constructing a dynamic spatiotemporal graph according to an exemplary embodiment of this application;

[0065] Figure 4 This is a schematic diagram illustrating the model training and optimization process of an exemplary embodiment of this application;

[0066] Figure 5 This is a schematic flowchart illustrating a molecular kinetic simulation according to an exemplary embodiment of this application;

[0067] Figure 6 This is a schematic diagram of the structure of an esophageal cancer RNA-protein interaction spatial prediction device based on a graph neural network, as illustrated in an exemplary embodiment of this application.

[0068] Figure label:

[0069] 1. Data acquisition and processing module; 2. Dynamic spatiotemporal graph construction module; 3. Model calculation module; 4. Molecular dynamics optimization module. Detailed Implementation

[0070] To make the objectives, implementation methods and advantages of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0071] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0072] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0073] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0074] To facilitate the explanation of the technical solution of this application, some concepts involved in this application will be explained first below.

[0075] I. Theoretical Basis of Biomolecular Interactions

[0076] 1. Molecular mechanisms of RNA-protein interaction (RPI)

[0077] Sequence complementarity: Sequence matching between RNA binding motifs (such as the YXXY motif of YBX1) and protein domains (such as the RRM domain) is the basis of interaction, which conforms to the molecular recognition principle of the lock-and-key model.

[0078] Structural adaptability: The spatial complementarity between RNA secondary structures (stem-loop, convex-loop) and protein three-dimensional conformations follows the induced fit model, that is, molecular conformations can be dynamically adjusted to optimize binding energy during interaction.

[0079] Microenvironment regulation: Cellular neighborhoods in the tumor microenvironment (such as the CAF-Epi niche) affect the spatiotemporal specificity of RPI by secreting cytokines and regulating molecular expression, which is consistent with cell communication theory.

[0080] 2. The theory of multi-stage evolution of esophageal cancer

[0081] Spatiotemporal heterogeneity of cancer development: The evolution of esophageal cancer from normal mucosa (NOR) to invasive carcinoma (ESCC) follows the theory of clonal evolution, and the RPI network is dynamically reconstructed with gene mutations and epigenetic modifications (such as m5C methylation).

[0082] Biological significance of spatial transcriptomics: Single-cell spatial localization data obtained by Xenium technology supports the spatial partitioning theory of the tumor microenvironment, that is, the spatial proximity of different cell subpopulations (CAF, Epi, immune cells) directly affects the probability of RPI occurrence.

[0083] II. Graph Theory and Graph Neural Networks (Graph Neural Networks) Theory

[0084] 1. Mathematical basis of dynamic graph modeling

[0085] Time-varying graphs (GTGraphs) have edge weights that change dynamically with time (cancer stage) and space (cell location), conforming to dynamic graph theory. Their formal definition is a quintuple G(t) = (V, E(t), X(t), A(t), τ(t)), where G(t) is the time-varying graph evolving with cancer stage t, V is the set of all nodes in the graph, E(t) is the dynamic edge set, X(t) is the multimodal feature tensor of all nodes at time t, A(t) is the weighted matrix of the connections between nodes, and τ(t) is the timestamp function.

[0086] Heterogeneous graphs: Node types include RNA and protein, and edge types include static (sequence / structure matching) and dynamic (spatial co-expression). Following the theory of heterogeneous graph neural networks, it is necessary to design type-specific aggregation functions (such as relational attention in HGT).

[0087] 2. Theoretical Framework of Spatiotemporal Attention Mechanism

[0088] Spatial dependency modeling: The neighborhood aggregation operation of Graph Convolutional Network (GCN) is based on the spectral domain analysis of the graph Laplacian operator, which extends the convolution operation in Euclidean space to graph structures and achieves spatial local smoothing of node features.

[0089] Time series modeling: Gated recurrent units (GRUs) capture long-distance temporal dependencies through gating mechanisms (update gate, reset gate), which are theoretically equivalent to the discretized approximation of the state-space model (SSM) and are suitable for modeling the sequence evolution of cancer stages.

[0090] Attention mechanism: Query-Key-Value computation of multi-head attention is essentially a generalization of similarity measurement theories (such as cosine similarity and dot product), and achieves effective fusion of cross-modal features (spatial + temporal) through adaptive weight allocation.

[0091] III. Multimodal Machine Learning Theory

[0092] 1. Theoretical basis of three-level feature coding

[0093] Sequence level: Transformer's self-attention mechanism uses position encoding to capture long-distance sequence dependencies, which is suitable for extracting contextual features of RNA m5C modification sites; evolutionary conservation analysis of protein sequences is based on the theory of co-evolution, that is, there are co-evolutionary signals at amino acid sites of interacting molecules.

[0094] At the structural level: The success of AlphaFold2 validates the deep learning paradigm for protein structure prediction. Its core is to treat proteins as graph structures (residues as nodes and distances as edges) and predict their three-dimensional coordinates through graph neural networks. The free energy calculation of RNA secondary structure is based on the thermodynamic folding theory, and the structure with the minimum free energy corresponds to the most stable conformation.

[0095] Spatial level: The three-dimensional modeling of cell coordinates follows spatial statistical theory. It quantifies the spatial correlation of the microenvironment through distance matrix and neighborhood analysis (such as cell pairs within a 50μm radius), which is consistent with the local modeling idea of ​​geographically weighted regression (GWR).

[0096] 2. Theoretical advantages of hierarchical feature learning

[0097] Cross-scale representation theory: a hierarchical aggregation from the residue level (atomic interactions) to the molecular level (domain matching) and then to the cellular level (microenvironment interactions), which is consistent with the hierarchical representation hypothesis in cognitive science, that is, the understanding of complex systems requires the extraction of features from the specific to the general through multi-level abstraction.

[0098] Heterogeneous Information Network (HIN) theory: Features at different scales (sequence, structure, space) constitute heterogeneous information sources, and feature fusion strategies guided by meta-paths need to be designed (such as meta-path modeling from RNA sequence → secondary structure → cell location).

[0099] IV. Molecular Dynamics and Physical Simulation Theory

[0100] 1. Physical basis of molecular dynamics (MD) simulations

[0101] Newtonian mechanics: MD trajectory calculation is based on Newton's second law F=ma. It predicts molecular conformational changes by solving the interatomic interaction forces (covalent bond forces, van der Waals forces, electrostatic forces). The force field (such as CHARMM36) is a specific realization of the empirical potential energy function.

[0102] Statistical thermodynamics: The simulation process follows the Boltzmann distribution and maintains system equilibrium through a canonical ensemble (NVT) or isothermal-isobaric ensemble (NPT) to ensure that conformational sampling covers possible thermodynamic states.

[0103] 2. Neural Network-Enhanced Physical Simulation

[0104] Self-Consistent Field Neural Network (SCF-NN): It replaces the traditional Poisson-Boltzmann equation for solving long-range electrostatic forces. Based on data-driven physical modeling theory, it fits the mapping relationship between electric potential and atomic coordinates through neural networks, thereby improving computational efficiency while maintaining accuracy.

[0105] Neural Relation Inference (NRI): Inferring the dynamic dependencies of interaction networks from MD trajectories, based on the causal reasoning theory of graph neural networks, and capturing the dynamic interactions between molecules (such as the formation and breaking of hydrogen bonds) through time-varying adjacency matrices.

[0106] V. Machine Learning Optimization Theory

[0107] 1. The theoretical advantages of multi-task learning

[0108] Inductive transfer learning: Classification tasks (interaction existence) and regression tasks (site coordinate prediction) share the underlying feature extraction module, which conforms to the hard parameter sharing theory. Through joint optimization, the generalization ability of the model is improved, which is especially suitable for small sample scenarios of biological data.

[0109] The theoretical basis of contrastive loss: Triplet Loss is based on metric learning theory. It enhances the model's ability to discriminate interaction specificity by bringing positive sample pairs closer together and pushing negative sample pairs further apart.

[0110] 2. Theoretical support for dynamic weight updates

[0111] Adaptive system theory: The dynamic edge weights are adjusted according to the cancer stage and spatial location. In essence, it is the adaptive control of time-varying systems. The nonlinear dynamic modeling of edge weights is realized through MLP and gating mechanism to adapt to the complex evolution law of RPI network.

[0112] Learning in non-Euclidean spaces: The non-Euclidean properties of graph structures (such as arbitrary node degree and variable connectivity) pose a challenge to traditional machine learning. Graph neural networks extend convolution operations to the graph domain through spectral graph theory, providing theoretical support for representation learning of dynamic graphs.

[0113] Before explaining the spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks provided in the embodiments of this application, the application scenarios and implementation environment of the embodiments of this application will be introduced first.

[0114] Esophageal cancer is a major disease threatening human health worldwide. According to GLOBOCAN data in 2023, there were approximately 626,000 new cases and as many as 544,000 deaths. Research has found that abnormal remodeling of RNA-protein interaction (RPI) networks plays a key role in the malignant progression of esophageal cancer. For example, abnormal binding of long non-coding RNAs to oncogenic proteins can drive epithelial-mesenchymal transition in esophageal squamous cell carcinoma; and dysregulation of the interaction between circular RNAs and RNA-binding proteins can help the tumor microenvironment achieve immune escape.

[0115] Therefore, accurately analyzing the spatiotemporal dynamic characteristics of esophageal cancer-specific RPIs is crucial for revealing the molecular mechanisms of carcinogenesis and screening therapeutic targets.

[0116] However, current RPI prediction techniques have significant shortcomings.

[0117] Currently, the closest existing technologies for predicting RNA-protein interactions (RPI) in esophageal cancer mainly revolve around traditional molecular docking methods. These methods predict the interaction between RNA and protein by constructing static three-dimensional structural models. For example, in some early studies, researchers used classic molecular docking algorithms to simulate the possible binding modes of RNA and protein based on their structural complementarity, thereby inferring potential RPIs.

[0118] Traditional molecular docking and other methods, which are based solely on static three-dimensional structure modeling of RNA and proteins, completely ignore the crucial factor that the intensity and location of RPI undergo significant spatiotemporal changes during the multi-stage evolution of esophageal cancer from normal mucosa to invasive carcinoma. They fail to capture the spatiotemporal changes of RPI intensity and location during the multi-stage evolution of esophageal cancer from normal mucosa to invasive carcinoma.

[0119] In terms of data application, existing technologies are also attempting to integrate various types of data.

[0120] For example, some studies combine the primary sequence information of RNA with the amino acid sequence of proteins, and use bioinformatics algorithms to analyze the association between sequences, thereby predicting RPI.

[0121] Despite attempts to integrate different types of data, key information such as RNA modification, protein conformational dynamics, and cellular localization has not been effectively integrated, resulting in many critical RPIs being missed.

[0122] Furthermore, existing algorithms generally lack the ability to effectively encode the specific characteristics of the esophageal cancer microenvironment, making it difficult to accurately distinguish the differences in RPI between cancerous and normal tissues. This greatly limits the in-depth understanding of the pathogenesis of esophageal cancer and the accurate screening of related therapeutic targets.

[0123] Following the development of spatial transcriptomics technology, some studies have attempted to integrate the gene expression data obtained from it with traditional molecular structure data, trying to uncover more clues related to RPI from the perspective of the spatial distribution of gene expression.

[0124] It is worth noting that the development of spatial transcriptomics technology has made it possible to obtain single-cell spatial localization and gene expression data of esophageal cancer tissue, and tools such as AlphaFold2 have also improved the accuracy of protein three-dimensional conformation prediction.

[0125] Against this backdrop, this invention proposes a prediction algorithm based on dynamic spatiotemporal graph neural networks and molecular dynamics simulations, which deeply integrates carcinogenesis stages, cell localization, and molecular structural features. This algorithm is expected to break through the bottlenecks of traditional technologies and provide strong support for precision medicine for esophageal cancer.

[0126] Based on this, this application provides a method and device for spatial prediction of RNA-protein interactions in esophageal cancer based on graph neural networks. It constructs a closed-loop system of "data integration - dynamic modeling - multimodal fusion - mechanism verification": integrating multi-source data to extract sequence, structural, and spatial features; using RNA and protein as nodes, constructing a dynamic graph that updates with the cancerous stage based on sequence / structure matching and spatial co-expression; predicting interaction existence, site coordinates, and binding energy by fusing features through spatiotemporal bi-branch coding (GCN extracts spatial dependence, gated recurrent units capture time series) and attention mechanisms; and using graph neural networks to predict conformations to generate initial conformations, optimizing the structure through molecular dynamics simulations and analyzing regulatory pathways.

[0127] This enabled the accurate capture of RPI spatiotemporal dynamic features, deep fusion of multimodal data, and feature extraction, thereby enhancing the learning of esophageal cancer-specific features and the analysis of mechanisms.

[0128] In tests on public datasets, the AUROC reached 0.90, the AUPRC reached 0.85, and the spatial localization RMSE was 0.8nm, providing an innovative and effective technical solution for the study of esophageal cancer pathogenesis, the discovery of therapeutic targets, and clinical precision medicine.

[0129] Next, the technical solutions of this application and how they solve the aforementioned technical problems will be described in detail through embodiments and in conjunction with the accompanying drawings. The embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0130] Figure 1 This is a schematic flowchart illustrating an exemplary embodiment of the present application of a spatial prediction method for esophageal cancer RNA-protein interactions based on graph neural networks. Figure 2 This is a schematic diagram illustrating the overall implementation process of another spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks, as shown in an exemplary embodiment of this application.

[0131] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a spatial prediction method for RNA-protein interactions in esophageal cancer based on graph neural networks is provided. In this embodiment, the method may include the following steps:

[0132] Step 100: Obtain multi-stage spatial transcriptome data related to esophageal cancer and preprocess it to obtain multimodal feature tensors and cancer stage labels.

[0133] Figure 3 This is a schematic diagram illustrating the process of constructing a dynamic spatiotemporal graph according to an exemplary embodiment of this application.

[0134] In some embodiments of the present invention, the data preprocessing layer is used for the extraction and integration of multimodal features, and this step 100 can be implemented through the following steps:

[0135] Step 110: Input RNA data, protein data, spatial data, and tag data. The RNA data includes esophageal cancer tissue RNA-seq, RNA secondary structure prediction, and m5C modification sites. The protein data includes quantitative proteomics data and AlphaFold2 predicted three-dimensional structures. The spatial data includes Xenium spatial transcriptome data. The tag data includes known RNA-protein interactions.

[0136] In some embodiments of the present invention, RNA data includes esophageal cancer tissue RNA-seq (TCGA-ESCC), RNA secondary structure prediction (RNAfold), and m5C modification sites predicted from DeepM6A; protein data includes quantitative proteomics data (TMT markers) and AlphaFold2 predicted three-dimensional structures (PDB format); spatial data includes Xenium spatial transcriptome data, i.e., single-cell coordinates, cell type annotation, and NOR / LGIN / HGIN / ESCC of carcinogenesis stages; and tag data includes known RNA-protein interactions, which are derived from RPI-Seq, StarBase database, and CLIP-seq experimental validation set.

[0137] Step 120: Extract features from RNA data, protein data, and spatial data respectively.

[0138] Furthermore, feature extraction from RNA data, protein data, and spatial data includes the following steps:

[0139] Step 121: RNA Feature Extraction

[0140] Sequence level: One-Hot coding + 3-mer frequency statistics to capture nucleotide sequence patterns;

[0141] Structural level: RNAfold calculates the free energy of secondary structures, and the stem-loop structure is encoded as an adjacency matrix;

[0142] Modification level: DeepM6A predicts m5C methylation sites as a functional regulatory feature.

[0143] Step 122: Protein Feature Extraction

[0144] Sequence level: ConSurf evolutionary conservation score, PFAM functional domain annotation (e.g., RNA binding domain RRM);

[0145] Structural level: AlphaFold2 coordinate PCA dimensionality reduction (preserving the top 50 principal components), DSSP secondary structure proportion (α-helix / β-sheet).

[0146] Step 123: Spatial Feature Extraction:

[0147] Localization characteristics: single-cell three-dimensional coordinates (x, y, z), subcellular localization (secretory protein / nuclear protein tags);

[0148] Microenvironment characteristics: proportion of cell types within a 50 μm radius (CAF / Epi / Immune), Spearma coefficient matrix of gene co-expression.

[0149] Step 130: Output the multimodal feature tensor and cancer stage labels. The multimodal feature tensor includes an RNA feature matrix, a protein feature matrix, and a spatial feature matrix. The cancer stage labels are t∈{1,2,3,4}, corresponding to 4 levels of cancer progression.

[0150] In some embodiments of the present invention, the output RNA feature matrix is ​​as follows: (n is the number of RNA molecules, 200 dimensions including sequence / structure / modification features); the protein feature matrix is... (m represents the number of protein molecules, 150 dimensions including sequence / structural features); the spatial feature matrix is... (Coordinates + Stage + Microenvironment Characteristics, m is the number of protein molecules, n is the number of RNA molecules, and 50 is the spatial neighborhood radius).

[0151] Step 200: Using the multimodal feature tensor and the cancer stage label, with RNA and protein as nodes, define static edges based on sequence complementarity and structural matching degree, and define dynamic edges based on spatial co-expression correlation, and construct a spatiotemporal dynamic graph in which the edge weights are updated with the cancer stage.

[0152] In some embodiments of the present invention, the dynamic spatiotemporal graph construction layer is used for spatiotemporal modeling of interaction networks. Its goal is to construct a spatiotemporal graph that dynamically updates with the stages of carcinogenesis, using RNA / protein as nodes, to characterize the static potential and dynamic regulation of interactions. Its input data is the RNA feature matrix output from the data preprocessing layer. The protein feature matrix is The spatial feature matrix is And the cancer stage label t∈{1,2,3,4}.

[0153] like Figure 3 As shown, step 200 can be achieved through the following steps:

[0154] Step 210: Define a node, which includes an RNA node and a protein node. The RNA node integrates RNA sequence, modification, spatial coordinates and stage information, and the protein node integrates protein structure, functional domain, location and stage information.

[0155] In some embodiments of the present invention, the RNA node features include:

[0156] Sequence-level features:

[0157] One-Hot encoding: The RNA sequence is encoded into a matrix (L is the sequence length, 4 is the nucleotide type), with the addition of a One-Hot marker for the m5C modification site (1 indicates the presence of modification, 0 indicates the absence of modification).

[0158] 3-mer features: Extract 3-mer frequency vectors and normalize them using TF-IDF to capture local sequence patterns of binding motifs.

[0159] Structural features:

[0160] Secondary structure prediction: RNAfold generates minimum free energy structures, converts them into dotted bracket notation, generates adjacency matrices through structure encoding models (such as RNA2DGraph), and extracts structural features such as loop length and stem length.

[0161] Solvent-accessible surface area (ASA): The ASA of nucleotides is calculated using RNAComposer, reflecting the degree of exposure to potential binding sites.

[0162] Spatial-level features:

[0163] Cell localization: The three-dimensional coordinates (x,y,z) of the cell are normalized to [0,1], and a cell type label is attached (e.g., Epi cells are encoded as a one-hot vector [1,0,0], and CAF cells are encoded as [0,1,0]).

[0164] Cancer stage: One-hot encoding of the current stage t∈{1,2,3,4} (corresponding to NOR / LGIN / HEIN / ESCC).

[0165] In some embodiments of the present invention, the protein node features include:

[0166] Sequence-level features:

[0167] Evolutionary conservation: Conservation scores of amino acid sites are calculated using ConSurf, generating a vector (N is the number of amino acid residues).

[0168] Functional domain annotation: Based on the PFAM database, the position of functional domains (such as the RNA binding domain RRM) is encoded as a binary mask (the position of the domain is 1 if it exists, otherwise it is 0).

[0169] Structural features:

[0170] 3D coordinates: CA atom coordinates predicted by AlphaFold2 Dimensionality was reduced to 128 dimensions using principal component analysis (PCA).

[0171] Secondary structure: The DSSP algorithm extracts the proportion of structures such as α-helices and β-folds as the basis for predicting flexible regions.

[0172] Spatial-level features:

[0173] Subcellular localization: secreted proteins (such as JAG1) are labeled in extracellular space, and nuclear proteins are labeled in nuclear coordinates, calibrated using the CellAtlas database.

[0174] Neighboring cell types: Based on spatial transcriptomics, the proportion of neighboring cell types within 50 μm was statistically analyzed (e.g., CAF accounted for 60%, Epi accounted for 30%).

[0175] Step 220: Static edge generation. Based on sequence complementarity, calculate the PWM matching score between RNA binding motif and protein binding domain. When the threshold is ≥0.8, generate initial edges. Then, based on the structural matching degree, predict the docking free energy using ClusPro and retain the first 10% of the initial edges as static edges (free energy ≤-8kcal / mol).

[0176] In some embodiments of the present invention, the static edges are driven by prior knowledge.

[0177] The formula for sequence complementarity score is expressed as:

[0178]

[0179] Wherein, match_count is the number of sequence matches between RNA binding motifs (such as the YXXY motif of YBX1) and protein binding domains (such as the RRM domain), max_possible_match is the maximum possible match number (the theoretical maximum number of residues that the motif and the binding domain can match), and motif_score is the matching score calculated based on the position weight matrix (PWM) trained on the RPI-Seq database (reflecting the degree of specific matching between the motif and the binding domain, ranging from 0 to 1).

[0180] The formula for the structure matching score is expressed as:

[0181]

[0182] Among them, Sim struct (i,j) is the structure matching score, which converts the binding energy into a score in the range of 0-1 using an exponential function (the lower the binding energy, the closer the score is to 1, and the better the structure matching). ClusPro_score(i,j) is the lowest energy conformation binding energy for ClusPro molecular docking, and kT is the thermal energy at room temperature (1.987 cal / mol·K × 310 k). The lowest energy binding energy of the RNA-protein complex calculated by the ClusPro tool indicates a more stable structure matching.

[0183] Molecular docking was performed using ClusPro, and the binding energy of the lowest energy conformation was taken (unit: kcal / mol). kT is the thermal energy at room temperature (1.987 cal / mol·K×310k).

[0184] The initial weight of a static edge is represented as: λ1 = 0.6, λ2 = 0.4.

[0185] Step 230: Dynamic edge generation. Based on spatial co-expression correlation, calculate the Spearman correlation coefficient between RNA expression level and protein abundance at carcinogenesis stage t. Only RNA-protein pairs with Spearman coefficient > 0.5 are retained to generate edges. The initial weight is the coefficient value, which is dynamically adjusted over time. The edge weight changes with stage t.

[0186] In some embodiments of the present invention, the dynamic edges are driven by spatiotemporal data.

[0187] The formula for spatial co-expression correlation is expressed as:

[0188] Corr t (i,j)=Spearman(ExprRNA i (t),ExprProtein j (t));

[0189] Corrt(i,j) is the Spearman correlation coefficient, based on RNA (ExprRNA). i (t) and protein (ExprProtein) j (t) Expression level at stage t, Spearman rank correlation coefficient (range -1 to -1), quantitative trend of co-expression (the closer the value is to 1, the stronger the co-expression) ExprRNA i (t) represents the expression level of RNA t during the carcinogenesis stage (TPM value), ExprProtein j (t) represents the abundance of protein t in the carcinogenesis stage (iBAQ value).

[0190] Calculate the Spearman correlation coefficient between RNA expression level and protein abundance at the carcinogenesis stage (t), retaining only |Corr t Edges with a correlation of ≥0.5 (only edges with strong co-expression correlation are retained, and weak correlation noise is filtered out).

[0191] The formula for the dynamic weight update mechanism is expressed as follows:

[0192]

[0193] in, It is a dynamic weight, c i,j =(x i -x j ,y i -y j ,z i -z j ) represents the three-dimensional coordinate difference (in nm) between the RNA and protein cells, reflecting spatial distance. The MLP contains two fully connected layers (input dimension 4: 3 coordinate differences + 1 cancer stage t, hidden layer 8 neurons, output dimension 1) to fit the influence of coordinate differences and stages on weights. The Gating function is Sigmoid to ensure that the weights are in the [0,1] interval, serving as the weights of dynamic edges in stage t (the values ​​are dynamically updated with stage and spatial distance).

[0194] Step 240: Output a spatiotemporal dynamic graph, whose node set carries multimodal features and whose edge set includes static and dynamic edges.

[0195] Among them, static edges represent stable interaction potential, while dynamic edges represent spatiotemporally specific regulation.

[0196] Step 300: Using the spatiotemporal dynamic graph as input, the interaction existence, site coordinates and binding energy are predicted by fusing features through spatiotemporal bi-branch coding and attention mechanism, and the interaction prediction results are output.

[0197] In some embodiments of the present invention, the dynamic spatiotemporal graph neural network model computation layer is used to predict the existence, location and energy of RNA-protein interactions through the spatiotemporal graph neural network and extract key features. Its input data is the spatiotemporal dynamic graph output by the dynamic spatiotemporal graph construction layer.

[0198] Figure 4 This is a schematic diagram illustrating the model training and optimization process of an exemplary embodiment of this application.

[0199] like Figure 4 As shown, step 300 can be achieved through the following steps:

[0200] Step 310: Input the spatiotemporal dynamic graph into the dynamic spatiotemporal graph neural network model, which includes a spatial branch, a spatiotemporal attention fusion module, and a multi-task prediction head.

[0201] In some embodiments of the present invention, the spatial branch comprises a two-layer graph convolutional neural network:

[0202] The first layer is at the residue level. The atomic coordinates of RNA nucleotides and protein amino acids are input. Through a graph attention network, the van der Waals forces and hydrogen bond interactions between atoms are calculated to generate the interaction scores of residue pairs. Then, the features of neighboring atoms are aggregated according to the attention weights.

[0203] The second layer is at the molecular level. It takes RNA / protein molecular features aggregated at the residue level as input, and generates the whole molecular embedding through GraphSAGE mean aggregation. It captures the matching degree between RNA secondary structural domains and protein functional domains, generates spatial proximity tags, and aggregates cellular neighborhood features.

[0204] In some embodiments of the present invention, the time branch input cancer stage sequence is encoded one-hot and input into a gated recurrent unit, and then the stage evolution time dependence is captured by hidden state update to obtain stage-specific features.

[0205] Step 320: The spatial branch outputs spatial features. The time branch outputs time features.

[0206] In some embodiments of the present invention, spatial branches obtain spatial features through spatiotemporal dynamic diagrams.

[0207] This spatial branch aggregates node neighborhood features through graph convolution operations to generate node features in the (l+1)th layer (preserving spatial dependencies). Its inter-layer propagation formula is as follows:

[0208]

[0209] Where H(1+1) is the feature matrix of the nodes in the (1+1)th layer. This is an adjacency matrix with self-loops (diagonal elements are 1, and the remaining elements are the same as the original adjacency matrix, ensuring that the node's own information is preserved). for The degree matrix (diagonal matrix, diagonal elements are) (The sum of the elements in the corresponding row) For the standardization operation of the adjacency matrix (to avoid the impact of node degree wiping on feature aggregation), H (l) Let be the feature matrix of the nodes in the l-th layer (each row corresponds to the feature vector of a node). The weight matrix is ​​d (where d is the feature dimension), and the activation function σ uses ReLU (which introduces non-linearity and filters negative eigenvalues).

[0210] The first layer (residue level) of this spatial branch: in The adjacency matrix is ​​constructed through inner atomic contacts. The feature hi(l+1) of node i at layer l+1 is generated by the feature hj(l) of all its neighboring nodes j at layer l through "normalized weighted summation + nonlinear activation". Essentially, it is an aggregation and abstraction of neighbor information, and its message passing formula is expressed as:

[0211]

[0212] in, This represents the feature vector (output feature) of node i in layer (l+1) (after update). The dimension d encodes the interaction pattern between node i and its neighbors (e.g., the new feature generated after RNA node i aggregates the features of its bound protein neighbor j), and l is the network layer index, indicating the stage of feature update. j is a neighboring node of node i (belonging to the neighborhood N(i) of i). In a biological context, a neighbor is defined as a node that interacts with i (e.g., the neighbor of RNA node i is the protein node j that it binds to, or other spatially adjacent RNA nodes). |N(i)| represents the number (degree) of the neighbors of node i, that is, the total number of nodes contained in N(i) (e.g., if RNA node i has unrelated protein neighbors, then |N(i)| = 5). |N(j)| represents the number (degree) of the neighbors of node j, that is, the total number of nodes contained in N(j) (e.g., if protein node j binds to 3 other RNAs besides i, then |N(j)| = 4). W represents the normalization coefficient, which balances the influence of nodes with different degrees. (l) This represents the learnable weight matrix of the l-th layer, with dimensions d×d (input feature dimension × output feature dimension), and its function is to weight the features of neighbor node j. Mapping to a feature space compatible with node i (e.g., converting protein structural features into a representation that can be fused with RNA sequence features). This represents the feature vector (input feature) of neighbor node j in layer l (before update). It has dimension d and encodes the original attributes of node j (such as the functional domain, spatial location, etc. of protein node j).

[0213] The second layer (molecular level) of this spatial branch: molecular interactions within 50nm construct an adjacency matrix, aggregating cellular neighborhood features.

[0214] In some embodiments of the present invention, time branches obtain time features through a spatiotemporal dynamic graph.

[0215] This time branch uses a gated loop unit, and its state update equation is expressed as:

[0216]

[0217] Enter x t One-hot encoding (4-dimensional) for the cancer stage, with hidden state h t Dimension 128 captures the temporal dependencies of stage evolution (such as the interaction enhancement signal of the HGIN -> ESCC stage).

[0218] Step 330: The spatiotemporal attention fusion module outputs fused features by calculating the interaction weights of the spatial features and the temporal features.

[0219] In some embodiments of the present invention, the interaction weights between spatial features and temporal features are calculated using the following formula:

[0220]

[0221] Among them, H space For GCN output (dimensions), H time For the hidden state (dimension) of the gated loop unit, the number of heads K=8, and the dimension d of each head. k =32, and the final fused feature is Concat(Head1,…,Head8), which outputs the fused feature.

[0222] Step 340: The multi-task prediction module includes a classification task and a regression task. The classification task outputs the interaction probability through the Sigmoid function, and the loss function is the binary cross-entropy. The regression task outputs three-dimensional coordinates through the Transformer decoder, and the loss function is the mean squared error.

[0223] In some embodiments of the present invention, the formula for the loss function binary cross-entropy for the classification task is expressed as:

[0224]

[0225] Where N is the total number of samples, y i ∈{0,1} represents the true label, p i To predict probabilities.

[0226] The formula for the mean squared error of the loss function in the regression task is expressed as:

[0227]

[0228] Where K is the number of positive samples, p k To determine the site coordinates for the experiment, For predicted coordinates.

[0229] Step 350: Output the interaction prediction results, which include the existence probability matrix and the coordinate set of the interaction sites.

[0230] Figure 5 This is a schematic flowchart illustrating a molecular kinematics simulation according to an exemplary embodiment of this application.

[0231] In some embodiments of the present invention, the prediction method provided in this application further includes, after step 300:

[0232] Step 400: Based on the prediction results, molecular dynamics is used for simulation, graph neural network is used to predict the conformation to generate the initial conformation, the structural stability is optimized through electrostatic interactions, and key regulatory pathways are analyzed.

[0233] In some embodiments of the present invention, the molecular dynamics optimization layer is used to realize a closed loop from prediction to mechanism verification, verify the structural stability of predicted interactions through molecular simulation, and analyze key regulatory pathways. Its input is the existence probability matrix and the set of interaction site coordinates in the interaction prediction results.

[0234] like Figure 5 As shown, step 400 can be achieved through the following steps:

[0235] Step 410: Generate initial conformations. Based on the predicted sites, perform rigid docking using ClusPro2.0 to generate the first 10 initial conformations. Perform GROMACS energy minimization on the initial conformations until the gradient is <1000kJ / mol / nm.

[0236] In some embodiments of the present invention, the constraint condition for rigid body docking is the radius of the interaction site predicted by the graph neural network (e.g., 1 nm around the m5C modification site of RNA, 1 nm around the RRM domain residues of protein), and the first 10 initial conformations generated are the first 10 lowest energy conformations; by using the Steepest Descent algorithm in GROMACS, selecting CHARMM36 as the force field, using TIP3P as the solvent model, iterating 1000 steps until the gradient is less than 1000 kJ / mol / nm, its energy is minimized.

[0237] Step 420: Perform electrostatic optimization using a self-consistent field neural network. The self-consistent field neural network is used to accelerate the calculation of long-range electrostatic interactions. Atomic coordinates and charges are input, and electrostatic potential is output. The conformation is optimized iteratively to stabilize the binding free energy ΔG of the complex at -9±0.5kcal / mol.

[0238] In some embodiments of the present invention, the self-consistent field neural network is input via atomic coordinates. and charge (M is the number of atoms in the complex), after passing through 3 fully connected layers (512 neurons per layer, ReLU activated), a long-range electrostatic potential is output. This neural network can minimize the mean square error between the predicted potential and the value calculated by the PME (Particle Mesh Ewald) method. The training data comes from the MD trajectories of 100 known RNA-protein complexes. Its electrostatic computation time is reduced from 200 ms / step of the traditional method to 10 ms / step, while maintaining the accuracy RMAE < 0.1 kcal / mol.

[0239] Step 430: Perform molecular dynamics trajectory simulation. The force field for the trajectory simulation is CHARMM36, the solvent is TIP3P, the simulation duration is 100 ns, the step size is 2 fs, and the trajectory is saved every 10 ps. The analysis indicators include root mean square deviation, root mean square fluctuation, and number of contact residues.

[0240] Step 440: Use a neural relationship inference model to analyze the regulatory pathways and output a mechanism verification report. The mechanism verification report includes a stable conformation PDB file and a list of key regulatory pathways P. pathway (Including pathway enrichment scores and core interaction pair list).

[0241] Furthermore, step 440 can be achieved through the following steps:

[0242] Step 441: Trajectory feature extraction. For the trajectories generated by molecular dynamics trajectory simulation, extract the dynamic features of the interaction points and encode them as node states. The dynamic characteristics include root mean square deviation, root mean square fluctuation, and number of contact residues;

[0243] Step 442: Relationship Reasoning: Use a graph attention network to infer the dynamic dependencies between nodes and generate a time-varying adjacency matrix A. t Identify key regulatory edges (such as the persistent contact between the SMOX m5C modification site and amino acids 98-105 of YBX1);

[0244] Step 443: Pathway validation: Validate the correlation between the inferred pathway and the esophageal cancer glycolysis and epithelial-mesenchymal transition pathway through enrichment analysis (such as the KEGG pathway), and screen key pathways with FDR < 0.05.

[0245] In some embodiments of the present invention, after path validation, a clinical application layer can be set up to realize the transformation from mechanism to precision medicine. This layer enables the development of clinical tools based on the validated interaction network, forming a "prediction-validation-application" closed loop. Its input data uses the P-value output from the molecular dynamics optimization layer. pathway .

[0246] In terms of application, the specific implementation steps are as follows:

[0247] Therapeutic target screening:

[0248] Screening criteria: Interaction probability > 0.9, microenvironment correlation degree ≥ 0.7 (e.g., JAG1-NOTCH1, the pivot molecule in the CAF-Epi interaction network);

[0249] Target validation: The binding constant was determined using SPR technology (e.g., predicted Kd = 2.3 nM, experimental value 2.5 ± 0.3 nM), which served as the starting point for the design of small molecule inhibitors.

[0250] Predictive model construction:

[0251] Risk scoring formula: Score = ∑ (i,j)∈CAF-Epi p(i,j)×CD8+T cell infiltration;

[0252] Clinical validation: This score was negatively correlated with 5-year survival (HR = 1.82, 95% CI: 1.51-2.18, p < 0.001) and can be used for preoperative survival prediction.

[0253] Data closed-loop feedback:

[0254] Experimental validation data (such as CLIP-seq interaction sites and SPR binding constants) are used as new samples to supplement the training set of the data preprocessing layer.

[0255] For every 50 new clinical samples, the model is retrained to dynamically optimize prediction accuracy.

[0256] The final output is a clinical toolset, including a list of candidate drug targets and prognostic assessment software. The list of candidate drug targets includes a mechanism of action and a structural diagram, while the prognostic assessment software includes input spatial transcriptome data and outputs risk scores and survival curves.

[0257] It can be seen that the overall architecture of this application has the following core advantages:

[0258] Multimodal fusion modeling: integrates sequence, structure, and spatial three-dimensional features, breaking through the limitations of traditional single-dimensional feature analysis;

[0259] Dynamic spatiotemporal representation: Capturing dynamic changes in interactions during cancer evolution through stage-sensitive edge weight updates;

[0260] Computational experimental closed loop: From AI prediction to molecular simulation verification and then to clinical application, an interpretable technology chain is formed;

[0261] Clinically-oriented design: Directly outputs prognostic biomarkers and therapeutic targets, accelerating the translation of basic research into precision medicine.

[0262] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially as indicated, these steps are not necessarily executed in the indicated order. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0263] Corresponding to the aforementioned embodiments of the spatial prediction method for esophageal cancer RNA-protein interactions based on graph neural networks, and employing the same technical concept, this application also provides embodiments of a spatial prediction device for esophageal cancer RNA-protein interactions based on graph neural networks.

[0264] Figure 6 This is a schematic diagram of the structure of an esophageal cancer RNA-protein interaction spatial prediction device based on a graph neural network, as illustrated in an exemplary embodiment of this application.

[0265] In one exemplary embodiment, such as Figure 6 As shown, the spatial prediction device for esophageal cancer RNA-protein interactions based on graph neural networks includes:

[0266] Data acquisition and processing module 1 is used to acquire multi-stage spatial transcriptome data related to esophageal cancer and perform preprocessing to obtain multimodal feature tensors and cancer stage labels.

[0267] The dynamic spatiotemporal graph construction module 2 is used to construct a spatiotemporal dynamic graph that updates with the cancer stage by using the multimodal feature tensor and the cancer stage label, with RNA and protein as nodes, defining static edges based on sequence complementarity and structural matching degree, and defining dynamic edges based on spatial co-expression correlation.

[0268] Model calculation module 3 is used to take the spatiotemporal dynamic graph as input, fuse features through spatiotemporal bi-branch coding and attention mechanism, predict the existence of interaction, site coordinates and binding energy, and output the interaction prediction results;

[0269] In one exemplary embodiment, the esophageal cancer RNA-protein interaction spatial prediction device based on graph neural networks provided in this application further includes:

[0270] The molecular dynamics optimization module 4 is used to perform molecular dynamics simulations based on the prediction results, use graph neural networks to predict conformations to generate initial conformations, optimize structural stability through electrostatic interactions, and analyze key regulatory pathways.

[0271] To verify this application, it was applied to clinical scenarios such as esophageal cancer RNA-protein interaction prediction, prognostic assessment biomarker screening, and targeted therapy target discovery. The specific application scenarios are described below.

[0272] In one exemplary embodiment, its application scenario is the prediction of esophageal cancer RNA-protein interactions based on publicly available datasets, and the specific steps are as follows:

[0273] 1. Data Preparation

[0274] RNA-seq and proteomics data from 100 esophageal cancer patients were obtained from the TCGA database. Spatial transcriptome data from 5 esophageal cancer tissue samples were downloaded from the Xenium spatial transcriptome database, covering four stages: normal mucosa, low-grade intraepithelial neoplasia, high-grade intraepithelial neoplasia, and invasive carcinoma. Additionally, known RNA-protein interaction data were obtained from the RPI-Seq database as tags.

[0275] 2. Feature Engineering

[0276] RNA characteristics: One-hot encoding of RNA sequences, prediction of m5C modification sites using the DeepM6A tool, and calculation of secondary structure free energy using RNAfold.

[0277] Protein characterization: AlphaFold2 was used to predict the three-dimensional structure of proteins, extract functional domain annotation information, and calculate the evolutionary conservation score of amino acids.

[0278] Spatial characteristics: Based on spatial transcriptome data, the three-dimensional coordinates of each RNA and protein in the cell were determined, the distance matrix of the cell neighborhood within a 50 μm radius was calculated, and the carcinogenesis stage was marked.

[0279] 3. Construction of dynamic spatiotemporal graphs

[0280] Using RNA and protein as nodes, static edges are generated based on sequence complementarity and structural matching. The edge weights are determined by the formula... Calculate, where λ1 = 0.6, λ2 = 0.4, Sim seq For sequence complementarity, Sim struct To achieve structural matching, dynamic edges are generated based on spatial co-representation correlation, and edge weights are determined by... Dynamically updated.

[0281] 4. Calculation of Dynamic Spatiotemporal Graph Neural Network Model

[0282] Spatiotemporal coding layer: The spatial branch uses GCN to perform two-layer feature aggregation, and the temporal branch processes cancer stage sequences through gated recurrent units.

[0283] Attention Fusion Layer: Employs an 8-head attention mechanism to fuse spatiotemporal features.

[0284] Interaction prediction layer: The classification head predicts the interaction probability, and the regression head predicts the three-dimensional coordinates of the interaction sites.

[0285] 5. Molecular Dynamics Post-processing

[0286] The predicted interaction site coordinates are input into GROMACS for initial conformation generation. The long-range electrostatic interaction is optimized using SCF-NN. Finally, the MD trajectory is analyzed using the NRI-MD model to infer the regulatory path.

[0287] 6. Result Verification

[0288] On the test set, the model achieved an AUROC of 0.90, an AUPRC of 0.85, and a spatial localization RMSE of 0.8 nm, predicting the key interaction between YBX1 and SMOX mRNA in the invasive cancer stage.

[0289] In one exemplary embodiment, its application scenario is the prediction of RPI (Reproductive Risk Index) for esophageal cancer in multicenter clinical samples, and the specific steps are as follows:

[0290] 1. Data Collection

[0291] In collaboration with three hospitals, fresh tissue samples were collected from 200 esophageal cancer patients for RNA-seq, proteomics, and spatial transcriptome sequencing. Simultaneously, immunohistochemical experiments were used to validate partial RNA-protein interactions as the gold standard.

[0292] 2. Data Preprocessing

[0293] Unify the sequencing platform and analysis process for data from different hospitals to eliminate batch effects.

[0294] Quality control and annotation of RNA and protein sequences.

[0295] 3. Model Training and Optimization

[0296] Transfer learning is employed, using the model trained in Example 1 as a pre-trained model for fine-tuning on a new dataset.

[0297] Adjust the model parameters, such as changing the learning rate from 1e-4 to 5e-5, and increasing the number of training epochs to 300.

[0298] 4. Prediction and Analysis

[0299] Using the optimized model to predict RNA-protein interactions, we found that a new protein secreted by CAF cells specifically interacts with Epi cell RNA, and the stability of this interaction was verified by molecular dynamics simulations.

[0300] 5. Clinical Application

[0301] By combining the prediction results with the patients' clinicopathological characteristics, a prognostic model was constructed, revealing that a specific RNA-protein interaction network was significantly associated with the patients' five-year survival rate.

[0302] In an exemplary embodiment, an algorithm component replacement and comparison experiment is conducted, and the specific steps are as follows:

[0303] 1. Dynamic graph construction and replacement

[0304] The original MLP-based dynamic edge weight update method in the dynamic graph construction is replaced with a Transformer-based update method. Specifically, the cell coordinate difference vector and cancer stage information are used as inputs to the Transformer, and the dynamic edge weights are calculated through a multi-head attention mechanism.

[0305] 2. Spatiotemporal coding layer replacement

[0306] Replace GCN with GraphSAGE, replace the gated recurrent unit with LSTM, and rebuild the spatiotemporal coding layer.

[0307] 3. Comparative Experimental Design

[0308] On the same dataset, RNA-protein interaction predictions were performed using the original algorithm and the algorithm with replaced components, and the following metrics were compared:

[0309] Precision, recall, and F1 score of interaction prediction.

[0310] RMSE and TM-score for spatial positioning.

[0311] The time consumed in model training and prediction.

[0312] 4. Experimental Results

[0313] Experiments show that the dynamic edge weight update based on Transformer improves AUPRC by 2%, but increases the computation time by 15%; the spatiotemporal coding layer combining GraphSAGE and LSTM is similar to the original algorithm in accuracy, but is slightly inferior in capturing long-distance dependencies.

[0314] As can be seen from the above different application scenarios, this application has the following significant advantages:

[0315] 1. Significantly improved prediction performance

[0316] In the first application scenario, tests based on publicly available datasets show that this application achieves excellent results in predicting RNA-protein interactions in esophageal cancer, with AUROC of 0.90 and AUPRC of 0.85, and a spatial localization RMSE of 0.8 nm.

[0317] Compared to traditional molecular docking methods (such as HADDOCK, whose AUROC is typically between 0.7 and 0.8) and existing graph neural network models (such as ZHMolGraph, whose AUROC is 0.798), this application significantly improves the accuracy of interaction existence prediction and spatial positioning precision.

[0318] This is thanks to dynamic spatiotemporal graph modeling and multimodal feature fusion technology, which can accurately capture the spatiotemporal dynamic changes of RNA-protein interactions during esophageal cancer carcinogenesis, avoiding the limitations of static modeling and single-dimensional feature analysis.

[0319] 2. Outstanding clinical application value

[0320] In the second application scenario, through multi-center clinical sample validation, this application not only predicts known RNA-protein interactions, but also discovers specific interactions between CAF cell secreted proteins and Epi cell RNA. By combining the prediction results with the patient's clinicopathological characteristics, a prognostic model was successfully constructed, confirming that specific RNA-protein interaction networks are significantly associated with the patient's five-year survival rate.

[0321] This achievement provides new biomarkers for the prognostic assessment of esophageal cancer patients and offers precise directions for the screening of targeted therapy targets, which will help achieve personalized precision medicine for esophageal cancer and improve the level of clinical diagnosis and treatment.

[0322] 3. The algorithm is highly flexible and scalable.

[0323] The algorithm component replacement experiment in the third application scenario shows that the architecture of this application has good flexibility and scalability. After replacing the MLP in dynamic graph construction with Transformer, AUPRC improved by 2%; although the computation time increased, it proves that the algorithm's performance can be further optimized by improving the components.

[0324] Meanwhile, experiments on replacing the spatiotemporal coding layer show that the combination of GraphSAGE and LSTM achieves similar accuracy to the original algorithm, indicating that the core architecture of this invention is compatible with different graph neural network components and temporal models, making it convenient for researchers to adjust the algorithm according to actual needs and adapt to esophageal cancer RNA-protein interaction prediction tasks in different scenarios.

[0325] 4. Breakthroughs in Theoretical and Technological Innovation

[0326] This application is the first to apply dynamic spatiotemporal graph neural networks to predict RNA-protein interactions in esophageal cancer, and constructs a closed-loop technical system of "data integration - dynamic modeling - multimodal fusion - mechanism verification".

[0327] Through a three-level feature coding system and a hierarchical graph neural network, deep fusion of multi-dimensional data in sequence, structure, and space was achieved; by combining molecular dynamics simulation and neural network optimization, the leap from prediction to physical mechanism verification was completed.

[0328] These innovations have broken through the bottlenecks of traditional methods in terms of dynamics, multimodal processing, and mechanism analysis, providing a brand-new technical path and theoretical framework for the study of cancer molecular mechanisms, and promoting the technological development of the intersection of bioinformatics and artificial intelligence.

[0329] It can be seen that:

[0330] This application constructs a dynamic spatiotemporal graph structure. By integrating multi-stage spatial transcriptome data of esophageal cancer, the edge weights of graph nodes are adjusted in real time according to the cancer stage. Furthermore, by utilizing a spatiotemporal bi-branch attention mechanism, it simultaneously captures the spatial dependence of cell neighborhoods and the time-series information of cancer stages, accurately characterizing the spatiotemporal dynamic changes of RNA-protein interactions.

[0331] This application also establishes a three-level feature coding system at the sequence level, structure level, and spatial level, integrating sequence information, three-dimensional structural information, and cellular spatial positioning information of RNA and protein. It uses a hierarchical graph neural network to achieve cross-scale feature learning and aggregation from the residue level to the cellular level, comprehensively improving the model's ability to represent complex biological information.

[0332] This application combines graph neural network prediction results with molecular dynamics simulations, uses self-consistent field neural networks to accelerate long-range electrostatic interaction calculations, optimizes complex structural stability, and uses a neural relationship inference model to analyze the esophageal cancer-specific RNA-protein interaction regulatory pathway from the simulated trajectory, thus achieving a closed loop from prediction to mechanism verification.

[0333] In the task of predicting RNA-protein interactions in esophageal cancer, this application achieved excellent results with AUROC of 0.90, AUPRC of 0.85, and spatial localization RMSE of 0.8nm. The algorithm architecture is highly flexible and its performance can be further optimized by replacing core components to adapt to different application scenarios.

[0334] Specific limitations regarding the graph neural network-based spatial prediction device for esophageal cancer RNA-protein interaction can be found in the above description of the limitations of the graph neural network-based spatial prediction method for esophageal cancer RNA-protein interaction, and will not be repeated here. Each module in the aforementioned graph neural network-based spatial prediction device for esophageal cancer RNA-protein interaction can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0335] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0336] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A spatial prediction method for RNA-protein interactions in esophageal cancer based on graph neural networks, characterized in that, include: We acquired multi-stage spatial transcriptome data related to esophageal cancer and preprocessed it to obtain multimodal feature tensors and cancer stage labels. Using the multimodal feature tensor and the cancer stage label, with RNA and protein as nodes, static edges are defined based on sequence complementarity and structural matching degree, and dynamic edges are defined based on spatial co-expression correlation, to construct a spatiotemporal dynamic graph in which the edge weights are updated with the cancer stage. Using the spatiotemporal dynamic graph as input, the interaction existence, site coordinates, and binding energy are predicted by fusing features through spatiotemporal bi-branch coding and attention mechanism, and the interaction prediction results are output.

2. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 1, characterized in that, The acquisition of multi-stage spatial transcriptome data related to esophageal cancer, followed by preprocessing to obtain multimodal feature tensors and cancer stage labels, further includes: Input RNA data, protein data, spatial data, and tag data. The RNA data includes esophageal cancer tissue RNA-seq, RNA secondary structure prediction, and m5C modification sites. The protein data includes quantitative proteomics data and AlphaFold2 predicted three-dimensional structures. The spatial data includes Xenium spatial transcriptome data. The tag data includes known RNA-protein interactions. Feature extraction was performed on RNA data, protein data, and spatial data respectively; Output a multimodal feature tensor and cancer stage labels. The multimodal feature tensor includes an RNA feature matrix, a protein feature matrix, and a spatial feature matrix. The cancer stage labels are t∈{1,2,3,4}, corresponding to 4 levels of cancer progression.

3. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 2, characterized in that, The feature extraction of RNA data, protein data, and spatial data respectively further includes: Sequence-level, structural-level, and modification-level feature extraction were performed on the RNA data, wherein: Sequence level: One-Hot coding + 3-mer frequency statistics to capture nucleotide sequence patterns; Structural level: RNAfold calculates the free energy of secondary structures, and the stem-loop structure is encoded as an adjacency matrix; Modification level: DeepM6A predicts m5C methylation sites as a functional regulatory feature; Sequence-level and structural-level feature extraction is performed on the protein data, wherein: Sequence level: ConSurf evolutionary conservation score, PFAM functional domain annotation; Structural level: AlphaFold2 coordinate PCA dimensionality reduction, retaining the top 50 principal components, DSSP secondary structure proportion; The spatial data is used to extract location features and micro-environment features, wherein: Localization characteristics: single-cell three-dimensional coordinates, subcellular localization; Microenvironment characteristics: proportion of cell types within a 50μm radius, Spearma coefficient matrix of gene co-expression.

4. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 2, characterized in that, The step of constructing a spatiotemporal dynamic graph that updates with the cancer stage, using the multimodal feature tensor and the cancer stage labels, with RNA and protein as nodes, defining static edges based on sequence complementarity and structural matching, and defining dynamic edges based on spatial co-expression correlation, further includes: Input RNA feature matrix, protein feature matrix, spatial feature matrix, and cancer stage labels; Define RNA nodes and protein nodes. The RNA nodes integrate RNA sequence, modification, spatial coordinates and stage information, while the protein nodes integrate protein structure, functional domains, location and stage information. Static edge generation: Based on sequence complementarity, the PWM matching score between RNA binding motif and protein binding domain is calculated. When the threshold is ≥0.8, initial edges are generated. Then, based on the structural matching degree, the docking free energy is predicted by ClusPro, and the first 10% of the initial edges are retained as static edges. Dynamic edge generation: Based on spatial co-expression correlation, the Spearman correlation coefficient between RNA expression level and protein abundance at carcinogenesis stage t is calculated. Only RNA-protein pairs with Spearman coefficient > 0.5 are retained to generate edges. The initial weight is the coefficient value, which is dynamically adjusted over time. The edge weight changes with the carcinogenesis stage t. The output is a spatiotemporal dynamic graph, whose node set carries multimodal features and whose edge set includes static and dynamic edges.

5. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 4, characterized in that, The step of taking the spatiotemporal dynamic graph as input, fusing features through spatiotemporal bi-branch coding and attention mechanism, predicting interaction existence, site coordinates, and binding energy, and outputting interaction prediction results further includes: The spatiotemporal dynamic graph is input into a dynamic spatiotemporal graph neural network model, which includes a spatial branch, a temporal branch, a spatiotemporal attention fusion module, and a multi-task prediction head. The spatial branch outputs spatial features, and the temporal branch outputs temporal features. The spatiotemporal attention fusion module outputs fused features by calculating the interaction weights between the spatial features and the temporal features; The multi-task prediction module includes a classification task and a regression task. Based on the fusion feature, the classification task outputs the interaction probability through the Sigmoid function, and the loss function is the binary cross-entropy. Based on the fusion feature, the regression task outputs three-dimensional coordinates through the Transformer decoder, and the loss function is the mean squared error. Output interaction prediction results, which include an existence probability matrix and a set of interaction site coordinates.

6. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 5, characterized in that, The spatial branch comprises a two-layer graph convolutional neural network: The first layer is at the residue level. The atomic coordinates of RNA nucleotides and protein amino acids are input. Through a graph attention network, the van der Waals forces and hydrogen bond interactions between atoms are calculated to generate the interaction scores of residue pairs. Then, the features of neighboring atoms are aggregated according to the attention weights. The second layer is at the molecular level. It inputs RNA / protein molecular features aggregated at the residue level, and generates the overall molecular embedding through GraphSAGE mean aggregation. It captures the matching degree between RNA secondary structure domains and protein functional domains, generates spatial proximity tags, and aggregates cellular neighborhood features. The time branch input cancer stage sequence is encoded one-hot and then input into a gated recurrent unit. The stage evolution time dependence is captured by the hidden state update to obtain stage-specific features.

7. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 5, characterized in that, After outputting the interaction prediction results, the following is also included: Based on the interaction prediction results, molecular dynamics simulations were performed, graph neural networks were used to predict the initial conformation, the structural stability was optimized through electrostatic interactions, and key regulatory pathways were analyzed.

8. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 7, characterized in that, Based on the predicted results, molecular dynamics simulations are performed, graph neural networks are used to predict conformations and generate initial conformations, structural stability is optimized through electrostatic interactions, and key regulatory pathways are analyzed. This further includes: The interaction prediction results are input into the molecular kinematics optimization layer for optimization. Initial conformations were generated. Based on the predicted sites, rigid docking was performed using ClusPro2.0 to generate multiple initial conformations. The initial conformations were then subjected to GROMACS energy minimization until the gradient was <1000kJ / mol / nm. Electrostatic optimization was performed using a self-consistent field neural network to accelerate long-range electrostatic interaction calculations. Atomic coordinates and charges were input, and electrostatic potential was output. The initial conformation was iteratively optimized to stabilize the complex binding free energy ΔG at -9±0.5kcal / mol. Molecular dynamics trajectory simulations were performed using CHARMM36 as the force field, TIP3P as the solvent, with a simulation duration of 100 ns, a step size of 2 fs, and the trajectory was saved every 10 ps. The analytical parameters included root mean square deviation, root mean square fluctuation, and number of contact residues. The regulatory pathways are analyzed using a neural relational reasoning model, and a mechanism verification report is output, which includes a stable conformation file and a list of key regulatory pathways.

9. The spatial prediction method for esophageal cancer RNA-protein interaction based on graph neural networks as described in claim 8, characterized in that, The process uses a neural relational reasoning model to analyze regulatory pathways and outputs a mechanism verification report. This report includes a stable conformation file and a list of key regulatory pathways, and further includes: Trajectory feature extraction: For the trajectory generated by molecular dynamics trajectory simulation, the dynamic features of the interaction points are extracted and encoded as node states. The dynamic features include root mean square deviation, root mean square fluctuation and number of contact residues. Relational reasoning: Using graph attention networks to infer dynamic dependencies between nodes, generate time-varying adjacency matrices, and identify key control edges; Pathway validation: Enrichment analysis was used to validate the correlation between the inferred pathway and esophageal cancer glycolysis and epithelial-mesenchymal transition pathways, and key pathways with FDR < 0.05 were screened.

10. A spatial prediction device for esophageal cancer RNA-protein interactions based on graph neural networks, characterized in that, include: The data acquisition and processing module is used to acquire multi-stage spatial transcriptome data related to esophageal cancer and perform preprocessing to obtain multimodal feature tensors and cancer stage labels. The dynamic spatiotemporal graph construction module is used to construct a spatiotemporal dynamic graph that updates with the cancer stage by utilizing the multimodal feature tensor and the cancer stage label, with RNA and protein as nodes, defining static edges based on sequence complementarity and structural matching degree, and defining dynamic edges based on spatial co-expression correlation. The model calculation module, using the spatiotemporal dynamic graph as input, fuses features through spatiotemporal bi-branch coding and attention mechanism to predict the existence of interactions, site coordinates, and binding energy, and obtains the interaction prediction results.

Images