Cancer driver gene interpretable identification method based on trust calibration and prototype learning

Abstract

The application relates to a cancer driver gene explainable identification method based on trust calibration and prototype learning, and relates to the technical field of biological information identification. A gene graph is constructed by fusing a protein interaction network and gene multi-omics characteristics, and part of nodes are labeled. Label-aware message passing is performed through a trust calibration encoder, the neighborhood is split into a labeled part and a non-labeled part for independent calibration, and node embedding is adaptively fused. An angle margin prototype classifier is used to construct a class prototype on a hypersphere, the decision boundary is expanded, and a prediction result is output. A pivot node self-supervised regularizer is introduced, center nodes are screened from labeled driver genes, positive constraints are applied to neighbor non-labeled nodes, negative penalties are applied to non-neighbors, and a supervised boundary is maintained when non-labeled data is used. Finally, a structured explanation module is used to reuse the internal evidence of the model, a verifiable explanation is provided for prediction, and the unification of high precision and credible explanation is realized.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics identification technology, and more specifically, to an interpretable identification method for cancer driver genes based on trust calibration and prototype learning. Background Technology

[0002] Identifying cancer driver genes is crucial for understanding tumorigenesis and achieving precision oncology. In tumor genomics research, accurately identifying driver genes faces two major challenges: extremely sparse driver gene annotation and a highly imbalanced positive-negative sample ratio. This results in most gene nodes having unlabeled nodes dominating their neighbors during information propagation. This data characteristic makes it difficult for traditional methods to effectively capture the subtle features of rare driver genes, necessitating new technological solutions capable of stably handling unlabeled node-dominated graph structures.

[0003] Existing technologies can be mainly categorized into three types, each with significant drawbacks. Methods based on mutation frequency heavily rely on high-frequency mutation features, failing to identify low-frequency driver genes and completely ignoring functional relationships between genes, leading to numerous false negatives. Methods based on traditional machine learning, while integrating multi-omics data, evaluate each gene in isolation, failing to leverage network structure for information sharing and exhibiting insufficient ability to identify context-dependent driver genes. Methods based on graph neural networks, while utilizing protein-protein interaction networks, employ indiscriminate aggregation strategies when unlabeled neighbor nodes dominate, severely diluting the supervisory signal and causing representation drift. Furthermore, existing self-supervised strategies are mostly task-independent, failing to enhance the discriminative boundaries for rare driver genes.

[0004] Current methods also suffer from fundamental deficiencies in interpretability. Most graph neural network methods produce black-box predictions, lacking inherent interpretability mechanisms and failing to provide biologists with verifiable decision-making support. While some methods attempt to introduce post-hoc interpretation techniques, their interpretations deviate from the model's actual decision-making process, making it difficult to guarantee fidelity. Furthermore, these methods passively and unreliably utilize unlabeled data, failing to actively filter credible structural information and easily introducing noise interference. Ultimately, this results in significantly insufficient prediction accuracy and robustness in real-world biological data scenarios, making it difficult to meet the urgent need for reliable driver gene identification in precision medicine. Summary of the Invention

[0005] This invention provides an interpretable identification method for cancer driver genes based on trust calibration and prototype learning, in order to improve at least one of the above-mentioned technical problems.

[0006] This invention provides an interpretable identification method for cancer driver genes based on trust calibration and prototype learning, which includes the following steps S1 to S5.

[0007] S1. Obtain protein-protein interaction networks and multi-omics features of genes to construct a gene hierarchy graph structure. During the training phase, some nodes are labeled.

[0008] S2. The input gene hierarchy graph structure is labeled and message-passed through a trust calibration encoder. The node neighborhood is explicitly split into labeled neighbors and unlabeled neighbors and is independently mapped and calibrated. Then, the corrected neighborhood message and the node self-retained features are adaptively fused according to the node local structure statistics to obtain the node embedding representation.

[0009] S3. Using an angle-margin-based prototype classifier, class prototypes are constructed for labeled nodes on the hypersphere. The similarity between the node embedding representation and the class prototype is calculated. An angle margin is introduced to forcibly expand the decision boundary, thereby outputting the prediction results of cancer driver genes.

[0010] S4. Introduce a self-supervised regularizer centered on hub nodes, select center nodes from labeled driver gene nodes, apply positive similarity constraints to unlabeled nodes in the neighbors of the center nodes, and apply negative sample contrast penalties to unlabeled nodes that are not adjacent to the center nodes, so as to maintain the learned supervision boundary while utilizing unlabeled information.

[0011] S5. Through the structured interpretation module based on internal model evidence, the internal evidence generated during the forward computation of the trust calibration encoder and the prototype classifier based on angle margin is directly reused to output a verifiable structured interpretation for the prediction results.

[0012] Preferably, S2 specifically includes steps S21 to S24.

[0013] S21. At each layer of the gene hierarchy graph structure, perform edge-aware attention message passing on each node to obtain attention-weighted neighborhood messages.

[0014] S22. Based on the label, the neighborhood of a node is split into labeled neighbors and unlabeled neighbors, thereby decomposing the attention-weighted neighborhood message into two independent part messages.

[0015] S23. Use independent learnable mappings to perform calibration operations on two independent partial messages respectively, and sum the calibration results to obtain the corrected neighborhood message.

[0016] S24. Based on the adaptive fusion and correction of the neighborhood messages and the self-preserving features of the nodes according to the local structural statistics of the nodes, the node embedding representation is obtained.

[0017] Attention-weighted neighborhood messages are:

[0018] .

[0019] In the formula, For the first Layer nodes Attention-weighted neighborhood messages. Let be any node in the graph. For neighboring nodes. For nodes In the The set of neighboring nodes of a layer, which depends on the set of edges used by the current layer. For the first From neighboring nodes in the layer Pointing to node The normalized attention coefficient. This represents the edge condition message function. for Learnable parameters. For nodes In the The node representation vector of the layer. For the edge The eigenvectors on.

[0020] The two independent partial messages are: .

[0021] In the formula, For nodes In the The layer only obtains a portion of the messages from its labeled neighbors. For nodes In the The layer only obtains a portion of the messages from its unlabeled neighbors.

[0022] The corrected neighborhood message is: .

[0023] .

[0024] .

[0025] In the formula, This is the corrected neighborhood message. For nodes In the The layer consists of corrected portions of messages aggregated from their labeled neighbors and processed by calibration mapping. For nodes In the The layer consists of corrected portions of messages aggregated from their unlabeled neighbors and processed by calibration mapping. This is the LeakyReLU activation function. For calibration Learnable parameters. For calibration Learnable parameters.

[0026] Preferably, the process of adaptively fusing and correcting the neighborhood messages and the node's self-retained features based on the node's local structural statistics is as follows: .

[0027] .

[0028] In the formula, For nodes In the The structural gating value of the layer. It is the sigmoid function. It is a multilayer perceptron. For nodes In the Lightweight structural statistical vectors of layers. For nodes In the The updated node representation of the layer. This is the corrected neighborhood message. For nodes In the Self-retaining path representation of a layer.

[0029] Preferably, S3 specifically includes steps S31 to S34.

[0030] S31. For the two categories of driving genes and non-driving genes, calculate the mean of the labeled node embeddings as the category prototypes, and perform L2 normalization on the node embeddings and class prototypes on the hypersphere.

[0031] .

[0032] .

[0033] .

[0034] In the formula, For nodes The L2-normalized embedding vector. For nodes Embedded representation. For category The prototype vector after L2 normalization. Indicate category The prototype. This represents the L2 norm. category A collection of labeled nodes The number of nodes.

[0035] S32. Based on the normalized node embeddings and class prototypes, the basic log odds are calculated using cosine similarity. Then, a learnable class-wise affine adjustment is introduced, which includes scale parameters and bias terms, to adapt to class-related calibration.

[0036] .

[0037] .

[0038] In the formula, Represents a node Normalized embeddings and categories The cosine similarity between normalized prototypes.

[0039] This indicates transpose. Represents a node For category The basic logarithmic odds. For category The scale parameter. For category The bias term.

[0040] S33. Based on learnable class-wise affine adjustment, the target logarithmic probability is modified based on the margin.

[0041] .

[0042] In the formula, For nodes For category The target logarithmic probability after margin adjustment. For nodes The true label. This represents the edge distance strength.

[0043] S34. Based on the modified target log odds, output the prediction results of cancer driver genes.

[0044] .

[0045] In the formula, For the model to nodes Predicted labels. This indicates the parameter corresponding to the maximum value. The number of categories.

[0046] Preferably, S4 specifically includes steps S41 to S44.

[0047] S41. From the labeled positive example nodes, sort them according to the degree centrality of the nodes, select the top-ranked nodes to form the central node set, and obtain the unlabeled neighbor nodes within the single hop range of each central node.

[0048] S42. Apply gradient stopping operation to the representation of the center node to fix the anchor point, treat the unlabeled neighbor nodes and the center node as positive sample pairs, calculate the similarity score and apply a consistency positive constraint to it.

[0049] S43. Negative sampling is performed on the unlabeled non-neighboring nodes of the central node. The sampled nodes and the central node are used as negative sample pairs. The similarity score is calculated and a high similarity penalty is applied to them.

[0050] S44. Calculate the average value of the positive and negative constraint losses within the local self-neighborhood, and use it as the overall goal of self-supervised regularization centered on the hub node.

[0051] Preferably, the application of the consistency positive constraint and the high similarity penalty is as follows.

[0052] .

[0053] .

[0054] .

[0055] In the formula, Indicates the central node Positive constraint loss. Indicates the central node The set of unlabeled neighbor nodes. express One of the unlabeled neighbor nodes. This represents the binary cross-entropy loss with logarithmic odds. Indicates positive sample pairs The similarity score. Indicates the central node The negative constraint penalty loss. As the central node The negative sample set. for The size of the set. Indicates from the negative sample set A non-neighbor negative sample node without a label is sampled from the middle. Indicates negative sample pairs The similarity score. The total loss is centered on the hub node and is the result of the self-supervised regularization. This is the set of center nodes selected in the current training round.

[0056] Preferably, the internal evidence generated during the forward computation of the trust calibration encoder and the angular margin-based prototype classifier includes at least: Based on the structural gating values ​​calculated by the trust calibration encoder in each layer, the dependence on neighborhood information and self-retaining paths is quantitatively predicted.

[0057] The confidence score is calculated based on the cosine similarity between the node normalized embedding and the normalized prototype of each category, as well as the log probability difference obtained after applying angular margins to the correct category. This score is used to demonstrate the semantic closeness between node features and various prototypes, as well as the degree of separation between categories.

[0058] Based on the normalized attention coefficients calculated by the trust calibration encoder in edge-aware message passing, and the scalar calibration factor derived from the learnable parameters in the label-aware mapping, the neighbor evidence score is calculated to identify the key neighbor nodes that contribute the most to the prediction.

[0059] A self-supervised consistency evidence score is calculated by multiplying the gradient-stopped normalized embedding of the center node with the normalized embedding of the unlabeled node. This score reflects the structural association strength between the unlabeled target node and the reliable positive example center node. The center node is selected from the labeled positive example nodes based on degree centrality.

[0060] Preferably, the evidence score of the key neighbor node and the self-supervised consistency evidence score are specifically as follows: .

[0061] .

[0062] In the formula, Representing neighboring nodes In the Layer to node The contribution score. Indicates the first From neighboring nodes in the layer Pointing to node The normalized attention coefficient. Indicates the first The scalar calibration factor derived from the corresponding label-aware mapping in the layer. Representing neighboring nodes For nodes The overall neighbor evidence score. This indicates the total number of layers in the trusted calibration encoder, i.e., the network depth.

[0063] .

[0064] In the formula, Represents a node Self-supervised consistency evidence score Represents the set of central nodes All central nodes Iterate through the expressions and take the maximum value of the calculated result of the subsequent expressions. Indicator function Represents the set of one-hop neighbors of node h. Indicates gradient stopping operation Indicates the central node L2 normalized embedding vector Indicates transpose operation Represents a node The L2-normalized embedding vector.

[0065] Preferably, the overall model of the identification method is trained end-to-end by jointly optimizing the prototype loss based on the margin and the self-supervised regularizer loss centered on the hub node.

[0066] The overall objective function is expressed as: .

[0067] In the formula, This is the overall objective function of the model. The prototype loss based on margins is calculated by applying cross-entropy to labeled nodes. The loss is centered on the hub node and is a self-supervised regularizer. for The weight hyperparameter.

[0068] By adopting the above technical solution, the present invention can achieve the following technical effects: The cancer driver gene identification method provided by this invention brings multi-dimensional and significant technological advancements and application value through its innovative trust calibration and prototype learning framework. The core beneficial effects are reflected in a substantial improvement in prediction accuracy and robustness, the proactive and secure utilization of massive amounts of unlabeled data, and the endowment of prediction results with inherent and verifiable interpretability. Specifically, this invention explicitly separates and calibrates information from labeled and unlabeled neighbors through a trust calibration encoder. Combined with adaptive structural gating, it effectively suppresses the information dilution and representation drift problems commonly found in graph structures dominated by unlabeled nodes, thus providing stable and reliable node representations for basic predictions. Building upon this, a prototype classifier based on angle margins forces the construction of wider and more stable decision boundaries for rare driver gene categories in the feature space, significantly enhancing the discrimination and recall capabilities for low-frequency key genes. Simultaneously, a self-supervised regularizer centered on hub nodes cleverly uses high-confidence positive examples as anchors to refine the representations of unlabeled nodes in the local network without disrupting the learned supervision boundaries, transforming passive data into resources for active learning. Ultimately, the entire framework is inherently interpretable. Its structured interpretation module can directly reuse internal evidence from the training process (such as gating values, prototype similarity, and key neighbor contributions) to provide a traceability basis for "why this decision was made" for each prediction, which greatly enhances the practical value and acceptability of the model in high-credibility scenarios such as biomedical research. Attached Figure Description

[0069] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0070] Figure 1 It is a technology roadmap.

[0071] Figure 2 This is a model network structure diagram of an interpretable identification method for cancer driver genes based on trust calibration and prototype learning. Detailed Implementation

[0072] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0073] Please see Figures 1 to 2 The first embodiment of the present invention provides an interpretable identification method for cancer driver genes based on trust calibration and prototype learning, which includes the following steps S1 to S5.

[0074] S1. Obtain protein-protein interaction networks and multi-omics features of genes to construct a gene hierarchy graph structure. During the training phase, some nodes are labeled.

[0075] S2. The input gene hierarchy graph structure is labeled and message-passed through a trust calibration encoder. The node neighborhood is explicitly split into labeled neighbors and unlabeled neighbors and is independently mapped and calibrated. Then, the corrected neighborhood message and the node self-retained features are adaptively fused according to the node local structure statistics to obtain the node embedding representation.

[0076] S3. Using an angle-margin-based prototype classifier, class prototypes are constructed for labeled nodes on the hypersphere. The similarity between the node embedding representation and the class prototype is calculated. An angle margin is introduced to forcibly expand the decision boundary, thereby outputting the prediction results of cancer driver genes.

[0077] S4. Introduce a self-supervised regularizer centered on hub nodes, select center nodes from labeled driver gene nodes, apply positive similarity constraints to unlabeled nodes in the neighbors of the center nodes, and apply negative sample contrast penalties to unlabeled nodes that are not adjacent to the center nodes, so as to maintain the learned supervision boundary while utilizing unlabeled information.

[0078] S5. Through the structured interpretation module based on internal model evidence, the internal evidence generated during the forward computation of the trust calibration encoder and the prototype classifier based on angle margin is directly reused to output a verifiable structured interpretation for the prediction results.

[0079] The technical roadmap of this invention is as follows: Figure 1 As shown. This invention addresses the class imbalance and label sparsity problems in cancer driver gene identification by reconstructing it as a trust-calibrated prototype learning problem on a graph dominated by unlabeled nodes. A trust-calibrated and unlabeled-aware structural gating with explainable prototypes framework (TRUST-EP) is proposed. This framework consists of four core modules working collaboratively.

[0080] First, the trust calibration encoder performs label-aware correction operations by explicitly separating aggregated messages from labeled and unlabeled neighbor nodes and processing them with independent calibration maps, thereby reducing the dominance of unlabeled neighborhoods on node representations.

[0081] Subsequently, interlayer structure gating adaptively fuses graph paths and self-retained paths using structural statistics to achieve implicit, node-level trust calibration without the need for explicit trust estimation.

[0082] Secondly, the prototype classifier based on angle margins constructs class prototypes for each category on labeled nodes and performs L2 normalization on node embeddings and prototypes on the hypersphere. By introducing adjustable angle margins, the separation boundary between the correct and incorrect categories is forcibly expanded, thereby improving the model's robustness under extreme class imbalance conditions and its discriminative learning ability for rare driving genes.

[0083] To further utilize unlabeled data while maintaining the learned supervision boundary, a self-supervised regularizer centered on a hub node is designed, using a local self-supervised scheme with a reliable positive example hub node as the anchor point. This scheme applies consistency constraints only between the hub node and its one-hop unlabeled neighbor nodes, and, in conjunction with gradient stopping anchoring and non-neighbor negative sampling strategies, ensures that the introduction of unlabeled information does not destroy the integrity of the supervision signal, thus achieving proactive and reliable utilization of massive amounts of unlabeled data.

[0084] Furthermore, the structured interpretation module, based on internal model evidence, directly reuses the gating values, prototype similarity, and neighbor contribution coefficients generated during the forward computation, outputting structured interpretations without additional training. It quantifies the model's dependence on neighborhood information through inter-layer gating values, demonstrates the semantic proximity of nodes to various prototypes through prototype similarity, identifies key neighbor nodes that contribute most to the prediction through calibrated attention coefficients, and reflects the association strength between unlabeled nodes and reliable positive examples through consistency scores with the central node. This inherently endows the model's prediction process with traceable and verifiable explanatory capabilities.

[0085] Experiments on multiple cancer type-specific and pan-cancer datasets demonstrate that the TRUST-EP framework significantly outperforms existing mainstream methods in core metrics such as AUPRC and F1, as well as robustness and sparse data adaptability. It effectively addresses key technical challenges such as neighborhood information dilution and representation drift under unlabeled node dominance, insufficient discriminative learning for rare driver genes under extreme class imbalance, unreliable passive use of unlabeled data, and lack of intrinsic fidelity in model interpretability. This provides an innovative solution for reliable identification of cancer driver genes in bioinformatics and automated assisted analysis in precision medicine scenarios.

[0086] This invention is an end-to-end graph neural network solution specifically designed for cancer driver gene identification in gene networks dominated by unlabeled nodes. Through the collaborative work of four core modules, this framework effectively solves key technical challenges faced by traditional methods in unlabeled node-dominated scenarios, such as neighborhood information dilution, insufficient discriminative learning of rare driver genes under extreme class imbalance, passive utilization of unlabeled data, and lack of model interpretability.

[0087] like Figure 2As shown, the overall process of this invention includes four stages: text encoding and graph construction, trust calibration message passing, prototype discrimination and self-supervised refinement, and structured interpretation output, forming a complete training and inference framework.

[0088] The operating mechanism of the architecture of this invention is as follows.

[0089] First, a gene hierarchy graph structure is constructed based on a protein-protein interaction network, where nodes represent genes and edges represent interactions between genes. Each gene node is associated with multiple omics feature vectors, including molecular features such as mutation frequency, copy number variation, DNA methylation, and gene expression profiles, which together constitute the initial feature matrix of the node. Simultaneously, driver gene annotation information is obtained from authoritative databases and divided into training, validation, and test sets. The training set contains only a very small number of labeled nodes, with the remainder being unlabeled nodes, forming a typical semi-supervised learning scenario dominated by unlabeled nodes.

[0090] Secondly, the trust calibration encoder performs label-aware message passing on the input graph. This encoder explicitly splits the neighborhood of each node into labeled and unlabeled neighbors, processing them separately through independent calibration maps to suppress noise interference from unlabeled nodes at the source. Based on this, a structure gating mechanism adaptively calculates the fusion weights according to the node's local structure statistics (such as degree and the proportion of labeled neighbors), dynamically balancing the contribution of calibrated neighborhood messages and node self-retained features. In low-label-rate neighborhoods dominated by unlabeled neighbors, the gating value is proactively lowered to reduce noise injection. In neighborhoods with a moderate proportion of labeled neighbors, the gating value is increased to fully utilize the limited supervision signals. Through this layer-by-layer gating fusion, the trust calibration encoder can effectively suppress representation drift, providing stable node embeddings for downstream tasks.

[0091] Subsequently, a prototype classifier based on angular margins performs class discrimination on the hypersphere. This module constructs class prototypes for both driver genes and non-driver genes from labeled nodes. By performing L2 normalization on node embeddings and class prototypes and introducing adjustable angular margins, it forces that the similarity between a node and the correct class prototype is significantly higher than its similarity with the incorrect class prototype. This expands the decision boundary for rare driver gene categories in the feature space, significantly enhancing the model's discriminative learning ability under extreme class imbalance conditions.

[0092] Meanwhile, a self-supervised regularizer centered on a central node actively mines structural information from massive amounts of unlabeled data. This module selects reliable central nodes with high degree centrality from labeled positive examples as anchor points, embeds them securely, and applies gradient stopping to maintain the learned supervised boundary. Positive constraints are applied to the single-hop unlabeled neighbors of each central node, causing them to embed closer to the central node. Simultaneously, negative examples are sampled from the non-neighbors of the central node, and negative constraints are applied to move them away from the central node. This asymmetric "push-pull" design not only utilizes the local network structure to refine the representation of unlabeled nodes but also avoids decision boundary shifts caused by conflicts between self-supervised signals and task objectives, achieving proactive and reliable utilization of unlabeled data.

[0093] Finally, the structured interpretation module, based on internal model evidence, provides verifiable predictive evidence during the inference phase. This module directly reuses the gating values, prototype similarity, and neighbor contribution coefficients generated during the forward computation, outputting structured interpretations without additional training: It quantifies the model's dependence on neighborhood information through inter-layer gating values; it demonstrates the semantic proximity of nodes to both driving and non-driving gene prototypes through prototype similarity; it identifies key neighbor nodes that contribute most to the prediction through calibrated attention coefficients; and it reflects the association strength between unlabeled nodes and reliable positive examples through consistency scores with the central node.

[0094] This structured evidence enables biologists to trace the basis of predictions and verify decision-making logic, significantly enhancing the practicality and credibility of the models in biomedical research.

[0095] The entire framework adopts an end-to-end joint optimization strategy, which simultaneously optimizes the trust calibration encoder, the prototype classifier based on angle margin, and the self-supervised regularizer with the central node as the core through a multi-task loss function, so that the four modules promote each other and improve synergistically during the training process.

[0096] Experimental results show that the framework of this invention significantly outperforms existing mainstream methods in core metrics such as AUPRC and F1 on multiple cancer type-specific datasets and pan-cancer datasets, and exhibits stronger robustness and adaptability to sparse data in sparse annotation scenarios dominated by unlabeled nodes.

[0097] Based on the above embodiments, in an optional embodiment of the present invention, S2 specifically includes steps S21 to S24, namely the steps performed in the trust calibration encoder.

[0098] S21. At each layer of the gene hierarchy graph structure, perform edge-aware attention message passing on each node to obtain attention-weighted neighborhood messages.

[0099] S22. Based on the label, the neighborhood of a node is split into labeled neighbors and unlabeled neighbors, thereby decomposing the attention-weighted neighborhood message into two independent part messages.

[0100] S23. Use independent learnable mappings to perform calibration operations on two independent partial messages respectively, and sum the calibration results to obtain the corrected neighborhood message.

[0101] S24. Based on the adaptive fusion and correction of the neighborhood messages and the self-preserving features of the nodes according to the local structural statistics of the nodes, the node embedding representation is obtained.

[0102] The Trust Calibration Encoder (TCE) aims to stabilize message passing on unlabeled node-dominated gene graphs by making message passing label-aware and structure-adaptive.

[0103] set up and Indicates the first The node representation of the layer. Among them, This is the node representation matrix for layer 0 (input layer). The original node feature matrix has the following shape: ( (Input feature dimension). For the first The node representation of the layer is a matrix, belonging to the real number field. This represents the total number of nodes in the graph. For the first The vector dimension (embedding dimension) represented by the layer node. This is the layer index, with values ​​ranging from 1, 2, ..., L, where L is the total number of layers in the model.

[0104] At each layer, through a graph with random perturbations Edge-aware attention-based message passing is performed to regularize the learning under conditions of noise and incomplete interactions. For the first The actual set of edges used by the layer (the graph structure after random perturbation). Represents the original set of edges China-Israel probability Randomly discard some edges to obtain This is a graph regularization technique used to enhance robustness.

[0105] The attention message passing in edge-aware architecture specifically involves: for nodes ,make For its in the The set of neighboring nodes of a layer. Let be any node in the graph. For nodes In the The set of neighboring nodes of a layer, which depends on the set of edges used by the current layer. For neighboring nodes. Represents a node and There exists an edge (usually undirected, but can also be understood as directed).

[0106] compute nodes An attention-weighted neighborhood message.

[0107] .

[0108] .

[0109] In the formula, For the first Layer nodes Attention-weighted neighborhood messages. Let be any node in the graph. For neighboring nodes. For nodes In the The set of neighboring nodes of a layer, which depends on the set of edges used by the current layer. For the first From neighboring nodes in the layer Pointing to node The normalized attention coefficient. This represents the edge condition message function. for Learnable parameters. For nodes In the The node representation vector of the layer. For the edge The eigenvectors on.

[0110] Although the attention mechanism is adaptive, it still performs relative weighting within the neighborhood. Therefore, when unlabeled neighbor nodes dominate and their signals are heterogeneous or weakly informative, attention normalization will still allocate a large number of weights to unreliable information sources, resulting in dilution of neighborhood information and layer-by-layer representation drift.

[0111] The label-aware decomposition and calibration specifically involves: to explicitly address scenarios dominated by unlabeled nodes, this invention uses a trained set of labeled nodes... Split the neighborhood.

[0112] Specifically, define and .

[0113] This decomposes the attention-weighted neighborhood message into two parts.

[0114] .

[0115] In the formula, For nodes In the The layer only obtains a portion of the messages from its labeled neighbors. For nodes In the The layer only obtains a portion of the messages from its unlabeled neighbors.

[0116] By analyzing the total neighborhood messages The summation term is restricted to the labeled neighbor set. Obtained from above. By analyzing the total neighborhood messages The summation term is restricted to the unlabeled neighbor set. Obtained from above.

[0117] Because the two types of information sources have systematic differences in reliability and distribution, they are calibrated using independent, learnable mappings: .

[0118] .

[0119] In the formula, For nodes In the The layer consists of corrected portions of messages aggregated from their labeled neighbors and processed by calibration mapping. For nodes In the The layer consists of corrected portions of messages aggregated from their unlabeled neighbors and processed by calibration mapping. This is the LeakyReLU activation function. For calibration Learnable parameters. For calibration Learnable parameters.

[0120] The corrected neighborhood message is .

[0121] In this way, evidence anchored by labeled nodes is preserved, while signals dominated by unlabeled nodes are corrected before being mixed.

[0122] The adaptive fusion with the self-retaining path structure specifically addresses this: even after calibration, the reliability of neighborhood aggregation still depends on node characteristics. Therefore, the trust calibration encoder introduces a layer-by-layer structural gating to adaptively fuse the calibrated neighborhood messages with the self-retaining transform.

[0123] make For nodes In the Lightweight structural statistical vectors for layers, and define: .

[0124] In the formula, For nodes In the The structural gating value of the layer. It is the sigmoid function. It is a multilayer perceptron.

[0125] It is a scalar between 0 and 1, used to dynamically balance the fusion weights of neighborhood information and self-retained information.

[0126] In the implementation of this invention, Includes statistics related to node degree and label availability, such as And the proportion of labeled neighbors And the difference term derived based on the features of the current layer. Furthermore, the self-preserving path is represented as... ,in for Learnable parameters.

[0127] The layer update model is as follows: .

[0128] In the formula, For nodes In the The updated node representation of the layer. This is the corrected neighborhood message. For nodes In the Self-retaining path representation of a layer.

[0129] Subsequently, LayerNorm normalization is applied to the node embedding representation. Therefore, when neighborhood evidence is weak due to the dominance of unlabeled nodes, the gating value decreases, and the encoder relies more on self-preserving paths, thus suppressing cross-layer representation drift. After the layer, set The model parameter set is as follows: The model parameter set and downstream modules are jointly optimized under the overall objective function.

[0130] This module addresses the issue of neighborhood information dilution caused by indiscriminate aggregation in traditional GNNs by explicitly splitting the neighborhood into labeled and unlabeled parts and applying independent calibration maps to process them separately. Experiments show that removing the label-aware split reduces the model's F1 score by 1.90 percentage points and AUPRC by 1.20 percentage points, validating the crucial role of this design in preserving the integrity of the supervision signal. Simultaneously, the structural gating mechanism adaptively balances neighborhood messages and self-preserving features based on the node's local structural statistics. In low-label-rate neighborhoods dominated by unlabeled nodes, it actively lowers the gating value to suppress noise injection, while in medium-label-rate neighborhoods, it raises the gating value to fully utilize limited supervision signals. Ablation experiments show that removing the gating mechanism reduces the model's F1 score by 3.79 percentage points, demonstrating its core contribution to stabilizing node representations and suppressing representation drift. Theoretical analysis further proves that the gating mechanism can rigorously attenuate representation drift caused by random perturbations, enhancing the model's robustness to structural noise.

[0131] Based on the above embodiments, in an optional embodiment of the present invention, S3 specifically includes steps S31 to S34, which are the steps performed in the prototype classifier based on angular margins.

[0132] Given node embeddings from a trust calibration encoder The Angular Margin-Based Prototype Classifier (MPH) performs prototype discrimination with angular margins to enhance the robustness of the decision boundary under class imbalance conditions.

[0133] make To train a set of labeled nodes, For category The set of labeled nodes. Here, "category" refers to the gene category that needs to be predicted in the cancer-driven gene identification task. To represent nodes The true label.

[0134] Specifically, this is a binary classification task, and therefore typically includes the following two categories: Driver genes: genes that play a key role in the occurrence and development of cancer; Non-driver genes: other genes that do not belong to the driver gene category.

[0135] S31. For the two categories of driving genes and non-driving genes, calculate the mean of the labeled node embeddings as the category prototypes, and perform L2 normalization on the node embeddings and class prototypes on the hypersphere.

[0136] Prototype building specifically involves: for each category The category prototype is calculated as the mean of the labeled node embeddings. The prototype summarizes the evidence at the category level and provides an anchor that is less sensitive to label sparsity than an instance-level classifier.

[0137] .

[0138] In the formula, Indicates category The prototype. category A collection of labeled nodes The number of nodes. For nodes Embedded representation. The node embedding dimension is The real vector space.

[0139] It is A real vector of dimension 1 represents the category in the embedding space. The center or average feature of all labeled nodes is used as the reference anchor point for that category. The final output is calculated by the upstream trust calibration encoder.

[0140] Cosine similarity with learnable class-wise affine terms: To reduce sensitivity to feature amplitude and align with angular margins, this invention operates on a hypersphere and performs L2 normalization on the embedding and prototype.

[0141] .

[0142] .

[0143] In the formula, For nodes The L2-normalized embedding vector. For nodes Embedded representation. This represents the L2 norm. Indicates category The prototype. For category The prototype vector after L2 normalization.

[0144] S32. Based on the normalized node embeddings and class prototypes, the basic log odds are calculated using cosine similarity. Then, a learnable class-wise affine adjustment is introduced, which includes scale parameters and bias terms, to adapt to class-related calibration.

[0145] Specifically, cosine similarity is used to calculate the basic logarithmic odds.

[0146] .

[0147] In the formula, Represents a node Normalized embeddings and categories The cosine similarity between normalized prototypes.

[0148] This indicates transpose.

[0149] The range is [ [1,1], the larger the value, the closer the directions are.

[0150] Introducing learnable class-wise affine adjustments, including scale parameters and bias terms To accommodate category-related calibration: .

[0151] In the formula, Represents a node For category The base log odds (or score). For category The scale parameter. For category The bias term.

[0152] This design aligns with the implementation of this invention, which exposes category-specific biases and (potentially learnable) global scale parameters for recording and calibration.

[0153] S33. Based on learnable class-wise affine adjustment, the target logarithmic probability is modified based on the margin.

[0154] Specifically, the target log odds with margin enforcement are as follows: although the cosine log odds have decoupled the magnitude information, they do not explicitly enforce a large decision margin for rare positive classes. Therefore, the prototype classifier based on angular margins uses a margin-based modification for the target log odds.

[0155] For the label is Labeled nodes The probability of modifying the target logarithm is as follows: .

[0156] In the formula, For nodes For category The target logarithmic probability after margin adjustment. For nodes The true label. This represents the edge distance strength.

[0157] It is the margin strength. This increases the required similarity to the correct prototype, resulting in a greater angular separation on the hypersphere, which is particularly beneficial when positive categories are rare.

[0158] S34. Based on the modified target log odds, output the prediction results of cancer driver genes.

[0159] Prototype-based prediction and optimization specifically involves: for any node The prototype classifier based on angle margins outputs a log-odds vector. And make predictions: .

[0160] In the formula, For the model to nodes Predicted labels. This indicates the parameter corresponding to the maximum value. The number of categories.

[0161] During training, cross-entropy loss is applied to labeled nodes using the margin-adjusted log-odds. : Then, it is jointly optimized with the trust calibration encoder to create an angle-margin-based prototype classifier. It is worth noting that the angle-margin-based prototype classifier only needs to be optimized from... The computation is based on a class prototype, thus maintaining stability even when the set of labeled nodes is extremely small.

[0162] This module constructs class prototypes from labeled nodes and introduces adjustable angular margins on the hypersphere to forcibly widen the separation boundary between the correct and incorrect classes, significantly enhancing the discriminative learning ability for rare driver genes. Experiments show that removing the angular margins (retaining only the prototype classification) reduces the model's F1 score by 5.35 percentage points and AUPRC by 0.33 percentage points. Replacing the prototype classifier with a traditional linear classifier reduces the F1 score by 5.46 percentage points and AUPRC by 1.19 percentage points. These results fully demonstrate that in extreme class imbalance scenarios, prototype learning based on angular margins can provide a more stable and discriminative decision boundary for minority classes, effectively improving the recognition recall rate of driver genes.

[0163] Based on the above embodiments, in an optional embodiment of the present invention, preferably, S4 specifically includes steps S41 to S44. That is, the steps executed in the self-supervised regularizer centered on the hub node.

[0164] S41. From the labeled positive example nodes, sort them according to the degree centrality of the nodes, select the top-ranked nodes to form the central node set, and obtain the unlabeled neighbor nodes within the single hop range of each central node.

[0165] S42. Apply gradient stopping operation to the representation of the center node to fix the anchor point, treat the unlabeled neighbor nodes and the center node as positive sample pairs, calculate the similarity score and apply a consistency positive constraint to it.

[0166] S43. Negative sampling is performed on the unlabeled non-neighboring nodes of the central node. The sampled nodes and the central node are used as negative sample pairs. The similarity score is calculated and a high similarity penalty is applied to them.

[0167] S44. Calculate the average value of the positive and negative constraint losses within the local self-neighborhood, and use it as the overall goal of self-supervised regularization centered on the hub node.

[0168] The self-supervised regularizer centered on hub nodes (ESSR) is a task-aligned self-supervised regularization method that leverages a massive number of unlabeled nodes while anchoring the supervised decision boundary learned by a trust-calibrated encoder and an angle-margin-based prototype classifier, allowing only controlled fine-tuning rather than boundary distortion.

[0169] The design is based on two features of this scenario.

[0170] First, most nodes are unlabeled, so the gradients provided by purely supervised training are sparse and noisy.

[0171] Secondly, performing naive self-supervised learning across the entire graph may introduce spurious constraints and lead to representation drift, especially under extreme class imbalance conditions.

[0172] Therefore, the self-supervised regularizer centered on the hub node restricts self-supervision to the local self-neighborhood centered on the reliable positive training node, and further prevents the center node from drifting through gradient stopping.

[0173] The unlabeled mask and positive example center nodes specifically refer to: defining a set of reliable positive example center nodes. for: That is, all labeled positive training nodes.

[0174] To reduce variance and highlight central nodes with structural influence, a small number of central nodes with high degree of influence on the protein-protein interaction network can be selected. Perform subsampling. Let For nodes The degree, This represents the set of center nodes selected in the current training round.

[0175] For each central node Consider its one-hop neighbor nodes And define the unlabeled neighbor nodes in the neighborhood of the self-network as: .in If it is a set of unlabeled nodes, Then the central node It does not participate in the computation of the self-supervised regularizer centered on the hub node. It is the standard symbol for the empty set in mathematics, representing a set that does not contain any elements.

[0176] The center node of gradient stopping and the positive similarity constraint are specifically: Let This is a normalized embedding used for a prototype classifier based on angle margins. For each center node... A self-supervised regularizer centered on a hub node will apply gradient stopping to... As a fixed central node, denoted as .

[0177] For each unlabeled neighbor node Define similarity score: And by using Consistency constraints are imposed by treating samples as positive pairs. This constraint is achieved using a binary logic loss with log odds.

[0178] .

[0179] In the formula, Indicates the central node Positive constraint loss. Indicates the central node The set of unlabeled neighbor nodes. express One of the unlabeled neighbor nodes. This represents the binary cross-entropy loss with logarithmic odds. Indicates positive sample pairs The similarity score.

[0180] By decoupling the central node representation, the self-supervised regularizer centered on the hub node regularizes unlabeled neighbor nodes toward the central node without affecting the supervised positive example central node itself, thereby preventing the supervision boundary from being violated.

[0181] Specifically, the negative samples of unlabeled non-neighbor nodes are as follows: To avoid degenerate collapse, a self-supervised regularizer centered on the hub node further compares each central node with unlabeled non-neighbor nodes.

[0182] For the central node Define the candidate negative sample pool as: And sample negative samples from it. Its size is ,in It is a negative sampling ratio.

[0183] For each ,calculate: And high similarity is penalized by labeling it 0.

[0184] .

[0185] In the formula, Indicates the central node The negative constraint penalty loss. As the central node The negative sample set. for The size of the set. Indicates from the negative sample set A non-neighbor negative sample node without a label is sampled from the middle. Indicates negative sample pairs The similarity score.

[0186] Notably, in the binary logistic loss with log odds and the penalty for high similarity via label 0, the hub-centric self-supervised regularizer employs an asymmetric "push-pull" design to utilize unlabeled nodes in a controlled manner. Single-hop unlabeled neighbors of reliable positive hub nodes are considered weak positive samples because, in protein-protein interaction networks, cancer driver genes tend to form local functional modules, and their direct interaction partners are more likely to be cooperating or co-driving genes than randomly selected genes. In contrast, unlabeled non-neighbors have weaker structural associations with hub nodes and thus serve as safer negative sample references.

[0187] Explicitly excluding single-hop unlabeled neighbors from the negative sample pool reduces the risk of sampling false negative samples that have a structural association with the positive example center node.

[0188] The combined effect of positive sample attraction and negative sample repulsion regularizes the local neighborhood while maintaining the supervisory decision boundary anchored by the labeled driving gene, allowing the representation to be finely adjusted without distortion of the boundary.

[0189] The objective function of the self-supervised regularizer centered on the hub node is as follows: the self-supervised regularizer averages the loss of the hub node.

[0190] .

[0191] In the formula, The total loss is centered on the hub node and is the result of the self-supervised regularization. This is the set of center nodes selected in the current training round.

[0192] A self-supervised regularizer centered on the hub node is jointly optimized with a trust-calibrated encoder and an angle-margin-based prototype classifier through an overall objective function.

[0193] This module utilizes a high-confidence positive example center node as an anchor point, applies positive constraints to its one-hop unlabeled neighbor nodes, samples negative constraints from non-neighbor nodes, and employs a gradient stopping anchoring strategy to achieve proactive and reliable utilization of massive amounts of unlabeled data. Experiments show that removing this regularizer reduces the model's F1 score by 0.81 percentage points and AUPRC by 0.36 percentage points, demonstrating that this module can refine the representation of unlabeled nodes using the local network structure without compromising the supervision boundary. In particular, this design employs a "push-pull" asymmetric mechanism to encourage unlabeled neighbors to gravitate towards reliable positive examples and to move non-neighbors away from the anchor point. This enriches the semantic representation of low-frequency nodes and avoids decision boundary shifts caused by conflicts between self-supervised signals and task objectives.

[0194] The internal evidence generated during the forward computation of the trust calibration encoder and the angle-margin-based prototype classifier in the evidence-based interpretation module includes at least the following: Based on the structural gating values ​​calculated by the trust calibration encoder in each layer, the dependence on neighborhood information and self-retaining paths is quantitatively predicted.

[0195] The confidence score is calculated based on the cosine similarity between the node normalized embedding and the normalized prototype of each category, as well as the log probability difference obtained after applying angular margins to the correct category. This score is used to demonstrate the semantic closeness between node features and various prototypes, as well as the degree of separation between categories.

[0196] Based on the normalized attention coefficients calculated by the trust calibration encoder in edge-aware message passing, and the scalar calibration factor derived from the learnable parameters in the label-aware mapping, the neighbor evidence score is calculated to identify the key neighbor nodes that contribute the most to the prediction.

[0197] A self-supervised consistency evidence score is calculated by multiplying the gradient-stopped normalized embedding of the center node with the normalized embedding of the unlabeled node. This score reflects the structural association strength between the unlabeled target node and the reliable positive example center node. The center node is selected from the labeled positive example nodes based on degree centrality.

[0198] The Structured Explanation Module (EBX) based on in-model evidence provides a structured and verifiable explanation of prediction results by reusing the in-model evidence generated by the trust calibration encoder and the prototype classifier based on angle margins.

[0199] It aims to meet two requirements for cancer-driving gene identification. First, the interpretation should be faithful to the trained predictor, rather than a retrospective heuristic. Second, the interpretation should be operationally verifiable for biological testing, revealing whether the prediction is supported by its own characteristics, neighborhood evidence, or class-level prototype similarity.

[0200] The structure-gated predictive decomposition specifically involves: reviewing the layer-by-layer gating of the neighborhood messages and node self-retained features after adaptive fusion correction by the trust-calibrated encoder. The corrected neighborhood messages are fused with the self-preserving transformation to update the node representation.

[0201] The structured interpretation module based on internal model evidence uses... This is used to quantify the model's dependence on neighborhood information and self-retained paths at each layer. The inter-layer structure gating value is the structure gating value calculated in each layer by the trust calibration encoder.

[0202] For the target node Define the hierarchical dependency score: .

[0203] The overall neighborhood dependency is summarized as follows: .

[0204] Larger This indicates that the prediction relies primarily on aggregated neighborhood evidence, while smaller... This indicates that due to weak or unreliable neighborhood information, the model places more trust in self-retaining paths. This gate-based summary is faithful because it directly corresponds to the forward computation process of trusting the calibrating encoder.

[0205] Specifically, the confidence level of prototype evidence and margin awareness is as follows: the prototype classifier based on angular margins uses normalized embedding... With normalized class prototype Comparisons are used to make predictions. Therefore, the structured interpretation module based on internal model evidence reports category evidence in the form of prototype similarity.

[0206] For category For a binary classification task, the prototype evidence score is defined as... and the evidence margin is .

[0207] .

[0208] .

[0209] The sign indicates which prototype a node is closer to, while its magnitude reflects separability on the hypersphere. To align with actual decision rules, this invention additionally reports the margin-adjusted log-odds in the prototype classifier based on angular margins.

[0210] make To modify the marginal-adjusted log-odds when modifying the target log-odds, the log-odds difference is: As a confidence indicator for decision-making consistency.

[0211] In summary, It explains the hierarchical evidence supporting the predictions and the extent to which the margin-enhanced classification head separates nodes from competing classes.

[0212] The neighbor evidence achieved through calibration contributions specifically involves: in addition to category-level evidence, the structured interpretation module based on internal model evidence also identifies which neighbors in the trust calibration encoder contributed to the neighborhood messages.

[0213] In each layer Corrected neighborhood messages It is obtained by summing the two parts of the attention-weighted neighborhood message, as well as the labeled and unlabeled parts of the message during calibration. The structured interpretation module based on internal model evidence attributes neighbor evidence by combining attention weights and calibration.

[0214] For neighboring nodes The contribution score is defined as .

[0215] .

[0216] In the formula, Representing neighboring nodes In the Layer to node The contribution score. Indicates the first From neighboring nodes in the layer Pointing to node The normalized attention coefficient. Indicates the first The scalar calibration factor derived from the corresponding label-aware mapping in the layer.

[0217] Specifically, if ,but Depend on Calculated, otherwise by The calculation yields the overall neighbor evidence score, which is then obtained through cross-layer aggregation.

[0218] .

[0219] In the formula, Representing neighboring nodes For nodes The overall neighbor evidence score. This indicates the total number of layers in the trusted calibration encoder, i.e., the network depth.

[0220] The structured interpretation module based on internal model evidence returns the module with the largest... The former The neighboring nodes serve as neighborhood evidence supporting the prediction. This interpretation is faithful because it derives from the same coefficients and calibrations used during the forward propagation process.

[0221] The self-supervised consistency evidence centered on the hub node specifically involves a structured interpretation module based on internal model evidence reporting whether the prediction results are consistent with the self-supervised regularizer centered on the hub node. The latter provides unlabeled data utilization signals anchored to reliable positive example hub nodes. For nodes... Define relative to the central node set Self-network consistency score.

[0222] .

[0223] In the formula, Represents a node Self-supervised consistency evidence score Represents the set of central nodes All central nodes Iterate through the expressions and take the maximum value of the calculated result of the subsequent expressions. Indicator function Represents the set of one-hop neighbors of node h. Indicates gradient stopping operation Indicates the central node L2 normalized embedding vector Indicates transpose operation Represents a node The L2-normalized embedding vector.

[0224] higher Indicate node Strong alignment with at least one reliable positive central node under a self-supervised regularizer objective centered on a hub node provides additional evidence for candidate gene discovery.

[0225] The structured interpretation module output based on internal model evidence specifically provides the following for each predicted node: The structured interpretation module, based on evidence within the model, outputs a structured interpretation tuple.

[0226] .

[0227] This tuple summarizes: the model's dependence on its neighborhood, the confidence level of prototype-level evidence and margin awareness, the most influential neighboring nodes in calibration propagation, and consistency with self-supervision centered on hub nodes.

[0228] This module requires no additional training targets and is calculated directly from within the model, making the interpretation both faithful and easy to verify.

[0229] The training objective function of the cancer driver gene interpretable identification method based on trust calibration and prototype learning of the present invention is as follows: by jointly optimizing the trust calibration encoder and the prototype classifier based on angle margin, and using a self-supervised regularizer centered on the hub node as an auxiliary regularization term.

[0230] The overall objective function is: .

[0231] In the formula, This is the overall objective function of the model. The prototype loss based on margins is calculated by applying cross-entropy to labeled nodes. The loss is centered on the hub node and is a self-supervised regularizer. for The weight hyperparameter.

[0232] The structured interpretation module based on internal model evidence is used for evidence extraction during reasoning without introducing additional training loss.

[0233] This module directly reuses the gating values, prototype similarity, and neighbor contribution coefficients generated during the forward computation, outputting structured interpretations without additional training. It quantifies the model's dependence on neighborhood information through inter-layer gating values, demonstrates the semantic proximity of nodes to various prototypes through prototype similarity, identifies key neighbor nodes that contribute most to the prediction through calibrated attention coefficients, and reflects the association strength between unlabeled nodes and reliable positive examples through consistency scores with the central node. This design inherently endows the model's prediction process with traceable and verifiable interpretability, providing biologists with decision-making support beyond "black box" predictions, significantly enhancing the model's practicality and credibility in biomedical research.

[0234] This invention addresses key technical issues such as the dilution of neighborhood information under unlabeled node dominance, insufficient discriminative learning of rare driver genes under extreme class imbalance, passive utilization of unlabeled data, and lack of model interpretability, bringing three key benefits.

[0235] 1. Significantly improved prediction accuracy and robustness: In the core metrics of cancer-driving gene identification, AUPRC and F1 significantly outperform existing mainstream methods. Especially in sparse label scenarios dominated by unlabeled nodes, the model performance declines more gradually, demonstrating stronger robustness and adaptability to sparse data, providing more reliable technical support for cancer genomics research.

[0236] 2. Significantly improved efficiency in utilizing unlabeled data: Through a local self-supervised scheme centered on a central node, the system enables proactive and reliable utilization of massive amounts of unlabeled gene data. Representation refinement is completed without compromising the supervision boundary, significantly alleviating the problem of insufficient learning caused by sparse labels and improving the ability to identify low-frequency driver genes.

[0237] 3. The prediction results have inherent interpretability: By reusing the gating values, prototype similarity and neighbor contribution coefficients within the model, structured and verifiable explanatory evidence is provided, enabling biologists to trace the basis of the prediction and verify the decision-making logic. This overcomes the limitations of insufficient fidelity in traditional ex post-hoc explanation methods and enhances the practical value of the model in the biomedical field.

[0238] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for interpretable identification of cancer driver genes based on trust calibration and prototype learning, characterized in that, Includes the following steps: S1. Obtain protein-protein interaction networks and multi-omics features of genes to construct a gene hierarchical graph structure; some nodes are labeled during the training phase. S2. The input gene hierarchy graph structure is labeled and message-passed through a trust calibration encoder. The node neighborhood is explicitly split into labeled neighbors and unlabeled neighbors and is independently mapped and calibrated. Then, the corrected neighborhood message and node self-retained features are adaptively fused according to the node local structure statistics to obtain the node embedding representation. S3. Using an angle-margin-based prototype classifier, class prototypes are constructed for labeled nodes on the hypersphere. The similarity between the node embedding representation and the class prototype is calculated. An angle margin is introduced to forcibly expand the decision boundary, thereby outputting the prediction results of cancer driver genes. S4. Introduce a self-supervised regularizer centered on hub nodes, select center nodes from labeled driver gene nodes, apply positive similarity constraints to unlabeled nodes in the neighbors of the center nodes, and apply negative sample contrast penalties to unlabeled nodes that are not adjacent to the center nodes, so as to maintain the learned supervision boundary while utilizing unlabeled information. S5. Through the structured interpretation module based on internal model evidence, the internal evidence generated during the forward computation of the trust calibration encoder and the prototype classifier based on angle margin is directly reused to output a verifiable structured interpretation for the prediction results.

2. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 1, characterized in that, S2 specifically includes: S21. At each layer of the gene hierarchy graph structure, perform edge-aware attention message passing on each node to obtain attention-weighted neighborhood messages. S22. Based on the label, the neighborhood of the node is split into labeled neighbors and unlabeled neighbors, thereby decomposing the attention-weighted neighborhood message into two independent part messages. S23. Use independent learnable mappings to perform calibration operations on two independent partial messages respectively, and sum the calibration results to obtain the corrected neighborhood message; S24. Based on the adaptive fusion and correction of the neighborhood messages and the self-preserving features of the nodes according to the local structural statistics of the nodes, the node embedding representation is obtained.

3. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 2, characterized in that, Attention-weighted neighborhood messages are: ； In the formula, For the first Layer nodes Attention-weighted neighborhood messages; Let be any node in the graph; For neighboring nodes; For nodes In the The set of neighboring nodes of a layer, which depends on the set of edges used by the current layer. For the first From neighboring nodes in the layer Pointing to node The normalized attention coefficient; Represents the edge condition message function; for Learnable parameters; For nodes In the The node representation of a layer is a vector; For the edge eigenvectors on; The two independent partial messages are: ； In the formula, For nodes In the The layer only obtains a portion of the messages from its labeled neighbors; For nodes In the The layer only obtains a portion of the messages from its unlabeled neighbors; The corrected neighborhood message is: ； ； ； In the formula, The corrected neighborhood message; For nodes In the The layer consists of corrected portions of messages aggregated from their labeled neighbors and processed by calibration mapping. For nodes In the The layer consists of corrected portions of messages aggregated from their unlabeled neighbors and processed by calibration mapping. Use the LeakyReLU activation function; For calibration Learnable parameters; For calibration Learnable parameters.

4. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 2, characterized in that, The process of adaptively fusing and correcting neighborhood messages and node self-retained features based on node local structural statistics is as follows: ； ； In the formula, For nodes In the Layer structure gating values; It is the sigmoid function; It is a multilayer perceptron; For nodes In the Lightweight structural statistical vectors of layers; For nodes In the The updated node representation after layer update; The corrected neighborhood message; For nodes In the Self-retaining path representation of a layer.

5. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 1, characterized in that, S3 specifically includes: S31. For the two categories of driver genes and non-driver genes, calculate the mean of the labeled node embeddings as the category prototypes, and perform L2 normalization on the node embeddings and class prototypes on the hypersphere. ； ； ； In the formula, For nodes The L2-normalized embedding vector; For nodes Embedded representation; For category The L2-normalized prototype vector; Indicate category The prototype; Represents the L2 norm; category A collection of labeled nodes The number of nodes; S32. Based on the normalized node embeddings and class prototypes, the basic log odds are calculated using cosine similarity. Then, learnable class-wise affine adjustments are introduced, including scale parameters and bias terms, to adapt to class-related calibration. ； ； In the formula, Represents a node Normalized embeddings and categories Cosine similarity between normalized prototypes; Indicates transpose; Represents a node For category The basic logarithmic odds; For category Scale parameters; For category The bias term; S33. Based on learnable class-wise affine adjustment, the target logarithmic probability is modified based on the margin. ； In the formula, For nodes For category The target logarithmic probability after margin adjustment; For nodes The true label; For edge distance strength; S34. Based on the modified target log odds, output the prediction results of cancer driver genes; ； In the formula, For the model to nodes Predicted labels; This indicates the parameter corresponding to the maximum value; The number of categories.

6. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 1, characterized in that, S4 specifically includes: S41. From the labeled positive example nodes, sort them according to the degree centrality of the nodes, select the top-ranked nodes to form the central node set, and obtain the unlabeled neighbor nodes within the single hop range of each central node. S42. Apply gradient stopping operation to the representation of the center node to fix the anchor point, treat the unlabeled neighbor nodes and the center node as positive sample pairs, calculate the similarity score and apply a consistency positive constraint to it. S43. Perform negative sampling on the unlabeled non-neighboring nodes of the central node, take the sampled nodes and the central node as negative sample pairs, calculate the similarity score and apply a high similarity penalty to them. S44. Calculate the average value of the positive and negative constraint losses within the local self-neighborhood, and use it as the overall goal of self-supervised regularization centered on the hub node.

7. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 6, characterized in that, The application of the consistency positive constraint and the high similarity penalty specifically refers to: ； ； ； In the formula, Indicates the central node Positive constraint loss; Indicates the central node The set of unlabeled neighbor nodes; express One of the unlabeled neighbor nodes; Represents the binary cross-entropy loss with log odds; Indicates positive sample pairs The similarity score; Indicates the central node Negative constraint penalty loss; As the central node The negative sample set sampled; for The size of the set; Indicates from the negative sample set A negative sample node without a label and not a neighbor sampled in the middle; Indicates negative sample pairs The similarity score; The total loss of the self-supervised regularizer centered on the hub node; This is the set of center nodes selected in the current training round.

8. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 1, characterized in that, The internal evidence generated during the forward computation of the trust calibration encoder and the prototype classifier based on angle margins includes at least: Based on the structural gating value calculated by the trust calibration encoder in each layer, the dependence on neighborhood information and self-retaining paths is quantitatively predicted. The confidence score is calculated based on the cosine similarity between the node normalized embedding and the normalized prototype of each category, as well as the log odds difference obtained after applying angular margins to the correct category. This score is used to demonstrate the semantic similarity between node features and prototypes of various categories, as well as the degree of separation between categories. Based on the normalized attention coefficients calculated by the trust calibration encoder in edge-aware message passing, and the scalar calibration factor derived from the learnable parameters in the label-aware mapping, the neighbor evidence score is calculated to identify the key neighbor nodes that contribute the most to the prediction. The self-supervised consistency evidence score is calculated by multiplying the gradient-stopped normalized embedding of the center node and the normalized embedding of the unlabeled node. This score reflects the structural association strength between the unlabeled target node and the reliable positive example center node. The center node is selected from the labeled positive example nodes based on degree centrality.

9. The interpretable identification method for cancer driver genes based on trust calibration and prototype learning according to claim 8, characterized in that, The evidence scores of the key neighbor nodes and the self-supervised consistency evidence scores are specifically as follows: ； ； In the formula, Representing neighboring nodes In the Layer to node Contribution score; Indicates the first From neighboring nodes in the layer Pointing to node The normalized attention coefficient; Indicates the first Scalar calibration factors derived from the corresponding label-aware mapping in the layer; Representing neighboring nodes For nodes Overall neighbor evidence score; This indicates the total number of layers in the trusted calibration encoder, i.e., the network depth. ； In the formula, Represents a node Self-supervised consistency evidence score Represents the set of central nodes All central nodes Iterate through the expressions and take the maximum value of the calculated result of the subsequent expressions. Indicator function Represents the set of one-hop neighbors of node h. Indicates gradient stopping operation Indicates the central node L2 normalized embedding vector Indicates transpose operation Represents a node The L2-normalized embedding vector.

10. A method for interpretable identification of cancer driver genes based on trust calibration and prototype learning according to any one of claims 1 to 9, characterized in that, The overall model of the identification method is trained end-to-end by jointly optimizing the prototype loss based on the margin and the self-supervised regularizer loss centered on the hub node. The overall objective function is expressed as: ； In the formula, The overall objective function of the model; The prototype loss based on margins is calculated by applying cross-entropy to labeled nodes; The loss is centered on the self-supervised regularizer with the hub node as the core. for The weight hyperparameter.

Images