Multi-modal adaptive fusion synthetic lethal prediction method based on non-common forgetting and low-rank interaction

Abstract

The invention discloses a synthetic lethal prediction method based on non-generality forgetting and low-rank interaction multi-mode adaptive fusion, and relates to the technical field of bioinformatics, computational biology and drug target screening. The method comprises the following steps: firstly, constructing a multi-view gene embedding expression based on a gene graph structure, a hypergraph structure and a knowledge graph; then constructing a difference vector and an interaction vector for any gene pair so as to characterize the difference and potential complementary relationship between the genes; multi-source information fusion is realized by using a deep learning model containing interactive attention, a forgetting gate and a multi-modal fusion mechanism; on this basis, multi-task optimization is carried out through joint classification loss, modal interaction loss and redundancy suppression loss, and optimal discrimination is realized through a dynamic threshold search strategy; and finally, outputting a sorting result of the candidate synthetic lethal gene pairs according to the prediction score. According to the method, the accuracy and generalization ability of synthetic lethal relationship prediction can be effectively improved, and a high-reliability calculation auxiliary tool is provided for anti-cancer drug target screening and gene therapy strategy design.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics, computational biology and drug target screening technology, and specifically relates to a synthetic lethal prediction method based on non-common forgetting and low-rank interaction multimodal adaptive fusion, which is used to identify gene pairs with synthetic lethal relationships with high accuracy. Background Technology

[0002] Synthetic lethality (SL) refers to the phenomenon where individual mutations in two genes do not affect cell survival, but simultaneous mutations in both genes lead to cell death. Based on this mechanism, gene pairs with a synthetic lethal relationship are considered important targets for developing anticancer drugs, precision medicine, and combined immunotherapy, and have become a core direction in precision oncology research.

[0003] In existing technologies, the identification of SL gene pairs mainly relies on wet experimental verification, statistical modeling analysis, and computational prediction methods based on machine learning or deep learning. Among them, wet experimental methods directly verify the lethal relationship between genes through CRISPR / Cas9 gene knockout, RNAi, genome-scale synthetic mutation experiments, etc. Although the results are highly reliable, they generally suffer from problems such as high experimental costs, long cycles, strong dependence on experimental conditions, and difficulty in covering the entire genome, which makes it difficult to meet the practical needs of rapid screening of large-scale candidate targets.

[0004] Statistical modeling methods typically infer potential SL gene pairs based on statistical indicators such as gene expression levels, mutation frequencies, and pathway enrichment. These methods have certain advantages in terms of computational complexity and interpretability. However, due to the high complexity, multi-scale regulatory characteristics, and significant nonlinear dependencies of the tumor system itself, traditional statistical models often struggle to characterize deep-seated interaction mechanisms between genes. The prediction results are highly sensitive to feature selection and assumptions, and their generalization ability and stability are significantly limited.

[0005] With the accumulation of multi-omics data and the improvement of computing power, SL prediction methods based on machine learning and deep learning have gradually become a research hotspot. These methods can integrate multi-source heterogeneous information such as knowledge graphs, protein-protein interaction networks, and functional association data to uncover complex nonlinear relationships between genes. However, existing technologies still have the following shortcomings in multimodal data fusion and feature representation:

[0006] At the data fusion level, existing methods typically employ only a single modality (such as knowledge graphs) or simply concatenate features from multiple modalities, neglecting the complementarity and high heterogeneity between different data sources. For example, knowledge graphs provide knowledge reasoning relationships, protein interaction networks capture molecular interaction structural information, and functional association sets reflect higher-order functional cluster dependencies; each of these modalities possesses unique biological meaning. Simple concatenation fails to alleviate modality conflicts and statistical distribution differences, and also fails to extract the most expressive and unique information from each modality.

[0007] At the feature fusion level, existing models tend to amplify common features among different modalities while neglecting the implicit differences and complementarities between modalities. This leads to the problem of "over-sharing," where the unique deep biological logic of each modality is diluted or even lost. Moreover, interactions between multimodalities often rely on simple concatenation or linear transformations, lacking fine-grained modeling of complex nonlinear interactions between modalities. Furthermore, the models lack effective modality selection capabilities, failing to adaptively determine the importance of each modality at the individual sample level; simultaneously, the high correlation between multimodal features may also cause information redundancy and noise accumulation.

[0008] In summary, although existing studies have improved the predictive performance of SL (Single-Legion) data by integrating multi-omics data, the high cost of experimental validation, simple modality fusion strategies, insufficient utilization of differences and complementarities, limited interaction modeling capabilities, and lack of decorrelation constraints still constrain the development of this field. Therefore, it is necessary to further improve the fusion methods and feature representation mechanisms of multi-source heterogeneous data based on existing technologies to address the long-standing technical bottlenecks in SL gene pair prediction, thereby providing more reliable technical support for the efficient discovery of anticancer drug targets. Summary of the Invention

[0009] This invention aims to overcome these technical bottlenecks and proposes a synthetic lethal prediction method based on non-common forgetting and low-rank interaction multimodal adaptive fusion. By introducing a non-common forgetting mechanism, low-rank bilinear interaction, sample adaptive modality gating, and correlation constraints, it achieves dynamic decomposition of modal common information and unique information, efficient interaction modeling between multiple modalities, and fine-grained control of modality contributions, thereby significantly improving the accuracy and interpretability of SL gene prediction.

[0010] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0011] The core objective of this application is to disclose a synthetic lethality prediction method based on multimodal adaptive fusion of non-common forgetting and low-rank interactions. This method involves constructing a multimodal adaptive fusion model based on non-common forgetting and low-rank interactions, comprising four processes: multimodal gene representation construction, gene pair differential feature construction and cross-modal representation enhancement of non-common forgetting, multimodal fusion optimization based on low-rank interactions and adaptive gating, and supervised learning-based SL gene pair prediction and ranking selection.

[0012] Step S1: Construction of multimodal gene representation. Based on three types of data, namely knowledge graph, protein interaction network and functional association data, gene semantic feature extraction, graph structure feature learning and hypergraph high-order relation modeling are carried out to obtain multi-source heterogeneous single-modal gene embedding representation, which provides a foundation for subsequent fusion.

[0013] First, a knowledge graph embedding learning model based on multi-hop attention is used to embed the SynLethKG knowledge graph constructed from multi-source biomedical databases, obtaining deep semantic vectors that reflect gene semantic associations and cross-entity multi-hop reasoning capabilities. Second, a graph structure is constructed based on protein interaction data, with genes as nodes and protein interaction relationships as edges. A multi-order Chebyshev convolutional network is used to aggregate the topological information of gene neighborhoods, achieving unified modeling of gene local structural features and multi-hop neighborhood dependencies, and obtaining a graph structure embedding representing gene interaction patterns. Simultaneously, a high-order gene function hypergraph is constructed based on gene function association data, with genes as nodes and sets of genes sharing the same biological function, pathway, or complex process as hyperedges, characterizing high-order group relationships between genes. A hypergraph neural network is used to perform message propagation and feature aggregation on multi-node interactions within the hyperedges, thereby generating a hypergraph structure feature representation containing high-order biological semantics.

[0014] Step S2: Gene pair differential feature construction and cross-modal representation enhancement through non-common forgetting. The single-modal gene embedding is extended into gene pair representation. Multimodal cross-gene pair features are constructed through differential operations, Hadamard interactions, etc., and the non-common forgetting mechanism is used to achieve compact representation fusion between modalities, thereby improving the interactive feature expression ability.

[0015] The knowledge graph embedding, graph structure embedding, and hypergraph structure embedding obtained in step S1 are subjected to differential processing and element-level interaction operations to construct multimodal cross-gene pair features that can represent the differential features and synergistic relationships between gene pairs. Simultaneously, a lightweight non-common forgetting mechanism is introduced to process the gene pair representations of the three modalities modally, automatically suppressing unique features that are inconsistent, irrelevant, or highly noisy between modalities, retaining only common features and key differential features that are stable and beneficial for subsequent prediction tasks. This mechanism achieves preliminary purification of cross-modal features without changing the original structural information, making the gene pair features output in step S2 more compact and laying a good foundation for subsequent finer-grained interactions and fusions.

[0016] Step S3: Multimodal fusion optimization based on low-rank interaction and adaptive gating. The low-rank linear interaction mechanism compresses the modal representation dimension and strengthens high-order cross-modal associations. A sample-level adaptive gating network is used to dynamically adjust the contribution of each modality, while cross-modal consistency constraints are combined to optimize multimodal collaborative representation, achieving fine-grained adaptive fusion of multimodal features.

[0017] First, the three modal features from step S2 are input into the multimodal interaction module, where a low-rank interaction mechanism models high-order relationships between modalities in the compressed representation space. This mechanism utilizes low-rank decomposition to reduce parameter size and models cross-modal second-order or bilinear relationships in a compact form, thereby strengthening the alignable parts between modalities and supplementing complementary structural information. After obtaining the fused intermediate representation, a learnable modality selection gating network is further employed to dynamically assign importance weights to different modalities based on the feature quality of the current sample, and the multi-source features are weighted and fused to generate a low-dimensional, stable, unified representation, enhancing the discriminative power of the fused representation in subsequent classification prediction and ranking tasks.

[0018] Step S4: Supervised learning-based prediction and ranking of SL gene pairs. The model is trained through multi-loss joint supervision, and the optimal threshold is automatically searched during the validation phase. Finally, all candidate gene pairs are predicted and ranked to output potential SL gene pairs, supporting the discovery and screening of anti-cancer targets.

[0019] The cross-modal fusion features obtained in step S3 are used as input to the final predictor to estimate the probability of whether any gene pair constitutes a synthetic lethal relationship. The predictor consists of a multilayer perceptron, including linear mapping, nonlinear activation, and normalization structures, and its goal is to learn a stable and generalizable discriminative boundary from the fusion vector.

[0020] To ensure the model can simultaneously utilize classification information, modality difference information, and structural feature alignment information, this step employs a multi-loss joint training mechanism; let the true label be... The model prediction results are The classification loss is then defined as:

[0021]

[0022] Combining the constraints of modal non-common forgetting loss and cross-modal consistency, the overall training objective can be expressed as:

[0023]

[0024] in, Corresponding to intermodal interaction consistency constraints. Modeling and penalizing cross-modal non-common forgetting. , , These are the multi-task weights;

[0025] By jointly optimizing the above three losses, this step achieves the following training objectives: (1) improve the accuracy and stability of the final classification prediction; (2) effectively utilize the intermodal difference features to enhance the distinguishability of samples; (3) suppress redundant modal noise and improve the robustness of cross-modal information fusion; (4) ensure that graph structure, hypergraph structure and knowledge graph semantics are optimized in a unified space; and (5) improve the generalization performance of the model on unseen gene pairs.

[0026] In the prediction phase, any gene and The final fusion feature is input into the predictor to obtain the probability that it is a synthetic lethal gene pair.

[0027] This application effectively addresses the problem that existing methods struggle to simultaneously cover multi-level mechanisms of action in biological systems by collaboratively modeling gene semantic information, interaction structure information, and functional module information within the same framework. The knowledge graph channel is used to characterize the semantic associations between genes and diseases, pathways, and biological processes; the interaction network channel is used to capture the local connectivity and long-range topological dependencies of genes in protein interaction systems; and the functional hypergraph channel is used to express the collaborative patterns of multiple genes within the same functional unit. Through complementary modeling of these three types of information, gene representation is no longer limited to a single perspective, thus improving the model's ability to characterize real biological regulatory relationships and providing more comprehensive and consistent basic information support for predicting synthetic lethal relationships.

[0028] At the gene pair level, this application combines differential construction with element-level interaction to naturally elevate single-gene features into paired features. It also introduces a non-common forgetting mechanism to automatically decouple shared information and modality-specific differences in multimodal features, thereby avoiding information duplication and mutual interference caused by simple splicing. At the same time, it uses low-rank interaction structure to characterize the cooperative relationship between modalities in a compact parameter space, enabling the model to capture key interaction patterns with lower complexity. Furthermore, it combines a sample-level adaptive gating mechanism to dynamically allocate modal weights according to the feature distribution of different gene pairs, making the fusion process more flexible and reliable. Overall, it significantly improves the stability, discriminativeness, and noise resistance of multimodal fusion representation.

[0029] Through the synergistic effect of the above-mentioned structured modeling and fusion strategies, this application can obtain a unified gene pair representation that is sensitive to differences, semantically consistent, and structurally robust in the prediction stage. Under the joint loss constraint, it can effectively balance classification performance and cross-modal consistency, thereby maintaining high accuracy and generalization ability in the prediction of unknown gene pairs. It can still work stably, especially in practical application scenarios with imbalanced samples and large biological data noise. Ultimately, it can reliably sort and screen potential synthetic lethal gene pairs, providing a scalable, reusable, and practically valuable technical means for the discovery of anticancer drug targets and the design of combination drug strategies. Attached Figure Description

[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0031] Figure 1 This is a flowchart of the synthetic lethal prediction method based on non-common forgetting and low-rank interactive multimodal adaptive fusion as described in an embodiment of the present invention;

[0032] Figure 2 This is an example diagram of the graph structure constructed for the protein interaction data described in the embodiments of the present invention;

[0033] Figure 3 This is an example diagram of the hypergraph structure constructed for the functionally related data described in the embodiments of the present invention;

[0034] Figure 4 This is a diagram illustrating the construction of differential features and non-common forgetting structures of gene pairs as described in the embodiments of the present invention;

[0035] Figure 5 This is a diagram of the multimodal fusion optimization structure described in the embodiments of the present invention;

[0036] Figure 6 This is a structural diagram of the predictor described in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0038] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are only used to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0039] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0040] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0041] like Figure 1 As shown, this application discloses a synthetic lethal prediction method based on non-common forgetting and low-rank interaction multimodal adaptive fusion, including the following steps:

[0042] S1: Multimodal feature construction;

[0043] Based on three heterogeneous sources—knowledge graphs, protein-protein interaction networks, and functional association data—we learn semantic representations, graph structure representations, and higher-order hypergraph representations of genes, respectively. The knowledge graph part models long-range semantic dependencies across entities through a multi-hop attention mechanism; the graph structure part aggregates neighborhood topology using multi-order convolution; and the hypergraph part captures higher-order group relationships through hyperedge message propagation. Finally, we obtain three complementary single-modal gene embedding features, laying the foundation for subsequent cross-modal learning.

[0044] S2: Gene pair differential construction;

[0045] Each gene embedding in step S1 is expanded into gene pair features. The differences and synergies between two genes are characterized simultaneously through differential operations and element-level Hadamard interactions, achieving a structural improvement from single-point features to pairwise relationships. To further distinguish between shared structures and unique differences between modalities, a non-common forgetting mechanism is used to decouple the gene pair features of each modality from shared features to differences, automatically suppressing redundant or conflicting information. At the same time, cross-modal interaction units are used to align and complete the shared representations of the three modalities, enhancing semantic consistency.

[0046] S3: Low-rank interaction and adaptive fusion;

[0047] Building upon this foundation, a low-rank interaction structure is employed to model higher-order associations between modalities, expressing the bilinear coupling relationship between multimodalities in a compact parameter space. Furthermore, a sample-level adaptive gating network is used to dynamically adjust the contribution of the three modalities based on the quality of the input features, achieving fine-grained and weighted multimodal fusion. Ultimately, a unified gene pair representation that is sensitive to differences, semantically consistent, and structurally stable is obtained, providing robust input for downstream prediction tasks such as classification and ranking.

[0048] S4: Supervised Prediction and Ranking;

[0049] The unified fusion representation obtained in step S3 is input into the final predictor to achieve probabilistic inference of whether gene pairs constitute an SL relationship. Model training employs joint optimization of classification loss, non-common forgetting constraint, and cross-modal consistency loss to simultaneously improve classification stability, cross-modal feature quality, and model generalization ability. During the inference phase, the predictor scores all candidate gene pairs and sorts them by probability to screen potential SL gene pairs, supporting the discovery of anticancer drug targets.

[0050] The synthetic lethality prediction method disclosed in this application, based on the adaptive fusion of non-common forgetting and low-rank interaction multimodal approaches, utilizes the semantic structure of knowledge graphs, gene interaction graph structures, and higher-order functional hyperspectral mapping. Figure 3Heterogeneous data are uniformly incorporated into a multimodal learning framework to construct an adaptive fusion model based on non-common forgetting and low-rank interaction multimodality, achieving comprehensive modeling of complex semantic associations, local structural dependencies, and high-order functional synergies between genes. First, based on a knowledge graph covering multi-source biomedical facts, a multi-hop attention mechanism is used to learn the semantic representation of gene relationships. Simultaneously, multi-order topological features are extracted based on protein interaction networks, and high-order group interactions at the gene set level are captured using a biofunctional hypergraph, forming three complementary single-modal embeddings. Subsequently, through differential feature construction and a non-common forgetting mechanism, the three gene embeddings are extended into cross-gene pair representations. The non-common forgetting mechanism, based on a shared-difference decoupling framework, automatically reduces redundant common components among the three modalities and highlights key difference patterns in semantics, topology, and high-order relationships between gene pairs, providing a cleaner and more structured input for subsequent interaction modeling.

[0051] Building upon this foundation, a low-rank linear interaction mechanism is introduced to compactly model differential and collaborative features. Through a low-rank decomposition-based linear interaction structure, the collaborative relationships between gene pairs in the multimodal space are captured in a parameter-efficient manner, resulting in a more stable and generalizable cross-gene pair representation. Simultaneously, a sample-level adaptive gating network is combined to dynamically allocate modality weights based on sample quality. Furthermore, cross-modal consistency constraints achieve joint alignment of the three modal semantic spaces, enabling the fused features to maintain discriminative power while possessing higher robustness and generalization ability. Finally, the model employs a joint supervised learning strategy combining classification loss, non-common forgetting loss, and consistency loss to achieve accurate probability prediction of synthetic lethal gene pairs and automatically completes the optimal threshold search to ensure the reliability of the ranking and selection process.

[0052] The multimodal fusion strategy proposed in this application breaks through the limitations of existing methods in both theory and practice: First, the model integrates three key information sources—gene semantics, graph structure, and higher-order function—simultaneously for the first time, avoiding the problem that a single modality cannot cover the multi-level mechanisms of biological processes; second, low-rank interaction and non-common forgetting mechanisms effectively solve the pain points of multimodal redundancy, information conflict, and noise amplification; third, sample-level adaptive gating ensures that the model can dynamically select the optimal modality combination for different gene pairs, improving the robustness of prediction; fourth, multi-loss joint supervision and threshold adaptive optimization significantly enhance the model's generalization performance on unseen gene pairs. In summary, the method in this application not only significantly improves the accuracy, generalization ability, and noise robustness of synthetic lethality prediction, but also provides a scalable and lower-cost intelligent solution for anti-cancer target screening, possessing significant scientific research value and broad application prospects.

[0053] As a preferred example of this application, in step S1, semantic representations of genes are obtained using knowledge graph embedding learning based on multi-hop attention. First, an adjacency sampling structure of a gene-related knowledge graph is constructed, and for each entity in the knowledge graph... (Including genes and related entities), a fixed-length neighbor list and corresponding relationship sequence are pre-calculated, denoted as:

[0054]

[0055]

[0056] in, This represents the number of samples from single-hop neighbors. Indicates the first Skip neighbor set and relation representation; establish trainable embedding vector tables for entities and relations respectively:

[0057]

[0058] in For entity number, Number of relation types For the embedded dimension;

[0059] For each batch of entities, a multi-hop aggregation and self-attention mechanism are used to iteratively compute the final entity representation; let a certain entity... The initial vector is For the first The jump neighbor aggregation uses the following self-attention aggregation module:

[0060] First, for each neighbor and their relationship Construct spliced ​​vectors ;

[0061] Secondly, the attention energy is calculated and normalized into weights:

[0062]

[0063] in, For attention projection weights, For the activation function, this invention uses leaky-ReLU;

[0064] Then, the neighbor vectors are weighted and aggregated based on attention weights, and then linearly transformed after being concatenated with the residuals of their own vectors:

[0065]

[0066]

[0067] in, , These are trainable parameters;

[0068] right Repeat the above aggregation process to finally obtain... .

[0069] This application fully leverages the advantages of multi-hop self-attention mechanisms in structured knowledge representation, uniformly encoding neighborhood structures, long-range dependencies, and multi-layer semantic associations in knowledge graphs into continuous vector representations. Compared to traditional embedding methods that rely solely on one-hop relationships or fixed aggregation methods, this invention, through multi-hop neighborhood unfolding and layer-by-layer attention weighting, enables the model to adaptively identify the semantic paths and neighborhood entities that contribute most to the target gene. This effectively captures deep semantic dependencies across multiple hops while maintaining the structural integrity of the knowledge graph. The resulting gene embedding vector more accurately reflects the comprehensive expression of multi-hop semantic associations, neighborhood structural features, and cross-relational patterns, providing high-quality, structure-sensitive foundational features for subsequent hypergraph structure learning and multimodal fusion.

[0070] As a preferred example of this application, in step S1, a Chebyshev graphical convolutional network is used to comprehensively model the high-order topological information and local structural features of the protein interaction network. The basic form of its graphical convolution can be expressed as:

[0071]

[0072] in, Indicates the first Chebyshev polynomials This represents the scaled, normalized Laplace matrix. Indicates the first The node feature matrix of the layer, These are the learnable parameters for the corresponding order.

[0073] In the examples of this application, an undirected graph is constructed based on a protein interaction network. Figure 2 For an example of the constructed graph, its adjacency matrix is ​​defined as:

[0074]

[0075] And thus calculate the node degree matrix. ,in, It is a column vector of all 1s; further, we obtain the normalized Laplace matrix:

[0076]

[0077] To satisfy the convergence of the Chebyshev polynomial approximation, a scaling transformation is performed on the Laplace matrix:

[0078]

[0079] in, for The largest eigenvalue;

[0080] For a given gene node, its initial features are:

[0081]

[0082] F represents the initial feature dimension; in this application example, 128-dimensional features processed by PCA are used.

[0083] To extract low-order structural features, firstly utilize and The process of propagating information within a local neighborhood and updating it can be described as follows:

[0084]

[0085]

[0086] in, This represents the activation function, and Dropout is the random deactivation operation.

[0087] To further capture higher-order topological dependencies over longer distances, a second Chebyshev convolution layer is stacked on top of the first layer, and its update method is as follows:

[0088]

[0089]

[0090] In this application, to enhance the model's expressive power, a third Chebyshev convolution layer is further superimposed, and its output can be represented as:

[0091]

[0092] This result is the structural embedding representation of the gene, denoted as... .

[0093] This application, in extracting PPI graph structure information using the ChebNet graph convolutional network, achieves efficient propagation of node features across 0-hop, 1-hop, and 2-hop neighborhoods through multi-order Chebyshev polynomials, thereby simultaneously encoding local connectivity relationships and long-distance topological dependencies. Unlike traditional graph convolutions based solely on first-order propagation, this method can directly model complex relationships across multiple nodes, thus capturing the implicit cascade regulation and synergistic effects in protein-protein interaction networks. By stacking three layers of Chebyshev convolutions and combining them with Dropout and activation functions, this application can fully exploit the hierarchical structural features of protein-protein interaction networks, resulting in gene graph structures that embed both low-order and high-order topological information, more accurately reflecting the functional roles and interaction patterns of genes in protein-protein interaction networks.

[0094] As a preferred example of this application, in step S1, a hypergraph constructed based on functional association data is used to characterize high-order population relationships at the gene set level, and a hypergraph neural network is used to jointly model local information within hyperedges and global transmission between hyperedges. Figure 3 Here is an example of the constructed hypergraph.

[0095] First, starting with the functional association matrix, the original gene functional association matrix is ​​defined as follows: Where N is the number of genes and M is the number of candidate superedges; based on a set threshold T, the set of superedges whose coverage reaches the threshold is selected, and the active superedge index set is defined as... And thus construct the sparse hypergraph correlation matrix:

[0096]

[0097] in, .

[0098] In a hypergraph, every hyperedge Typically, each hyperedge corresponds to a functional item (such as a GO term, pathway, module, etc.), and a functional item may be associated with multiple genes simultaneously. Therefore, to measure the importance of different hyperedges, a weight needs to be assigned to each hyperedge. Based on this, the hyperedge weight vector is defined in this application example as follows:

[0099]

[0100]

[0101] in, This represents the set of genes involved in the computation. Indicates gene Does it belong to the superedge? To ensure numerical stability, node degree and hyperedge degree are defined as follows:

[0102]

[0103]

[0104] in It is a column vector with all elements being 1. If there are zero entries, they are replaced with the smallest positive number to avoid division by zero.

[0105] At the hypergraph convolution level, the following propagation formula based on hypergraph Laplacian is used to implement bidirectional message passing between nodes and hyperedges:

[0106]

[0107] in, For the first The node feature matrix of the layer, Indicates the first The feature dimension of the layer This is the hyperedge weight matrix. For learnable mappings, This represents the element-wise activation function.

[0108] In the specific implementation, the input features are first linearly projected to unify the dimension, then three layers of hypergraph convolution are stacked and activation with residual connections and Dropout operations are used to enhance training stability. The layer-by-layer update can be formalized as follows:

[0109]

[0110]

[0111]

[0112]

[0113] Operator The main body of the matrix product is:

[0114]

[0115] The final hypergraph embedding is:

[0116]

[0117] Each row corresponds to a high-order hypergraph structure vector for a single gene. This vector contains both global information after backpropagation through hyperedge aggregation and fine-grained interaction features within local hyperedges. Through the aforementioned hypergraph construction and feature extraction process, this application can effectively mine multi-gene collaborative signals and high-order interaction patterns contained in functional annotation data, providing high-quality hypergraph structure embedding support for subsequent gene pair differential construction, low-rank interactions, and adaptive multimodal fusion.

[0118] As a preferred example of this application, in step S1, protein interaction data is first obtained from the BioGRID database to construct a gene interaction graph. After PCA dimensionality reduction and normalization optimization, the graph is input into a Chebyshev graph convolutional network to extract the multi-order topological structure features of genes in the interaction network. Secondly, semantic association information of genes is obtained from the SynLethKG knowledge graph, and low-dimensional semantic vectors are generated through a knowledge graph embedding model. Then, feature alignment is performed through a linear mapping module to obtain an embedded representation representing the biological semantic relationship of genes. Finally, a high-order hyperedge structure is constructed from the functional association gene set obtained from MSigDB and input into a hypergraph neural network to encode the high-order co-occurrence relationship within the functional module and extract the functional association features of genes.

[0119] This application utilizes ChebNet's multi-order polynomial propagation mechanism in the PPI channel, enabling the model to simultaneously capture the local interaction structure and long-range topological dependencies of genes. In the knowledge graph channel, it achieves effective extraction and compression of entity-relation semantics through embedding models and linear mapping, making the semantic features of genes more regular and usable. In the functional association channel, the hypergraph neural network can aggregate features of unpaired high-order relationships between multiple genes within the same functional module, supplementing the group functional patterns that are difficult to express by traditional graph structures. By modeling and uniformly outputting the above three types of features, this application achieves a comprehensive representation of gene structural information, semantic information, and functional information in step S1, providing a more sufficient, stable, and complementary multimodal input foundation for subsequent fusion and drug target screening.

[0120] As a preferred example of this application, in step S2, a combination of differential construction and element-level Hadamard interaction is used to upgrade the dimensionality of single-point gene features to a pairwise relation space, such as... Figure 4 As shown. For any gene and The three-modal single-point embeddings are as follows:

[0121]

[0122]

[0123] in, Represents the modal features of a knowledge graph. Represents the modal characteristics of the graph. Represents hypergraph modal features. For the embedded dimension.

[0124] First, differential features and element-wise Hadamard interaction features are constructed to simultaneously characterize the differential and interaction structures of the two genes in each modality, as defined below:

[0125]

[0126]

[0127] in, Representing three modes, This indicates the absolute difference operation. This represents Hadamard multiplication;

[0128] This yields the initial characteristics of the three-modal gene pairs:

[0129]

[0130] in, This represents a vector-level concatenation operation.

[0131] Because the three modalities share certain structures and have modality-specific differences, direct fusion can lead to redundant information stacking or conflicting features interfering with each other. Therefore, this application introduces a non-common forgetting mechanism to decouple the shared structures and modality-specific differences of gene pair features from different modalities.

[0132] First, the three modal features are uniformly mapped to a shared feature space:

[0133]

[0134] in, To share the linear mapping matrix, For bias terms;

[0135] Then, the three-modal shared candidate features are stacked as follows:

[0136]

[0137] To obtain modality-shared summaries, learnable attention vectors are introduced. Calculate the importance of each mode:

[0138]

[0139] in, This represents the vector dot product; from this, a shared summary representation of gene pairs is obtained:

[0140]

[0141] To automatically suppress redundant components between modes while retaining mode-specific and prediction-beneficial parts, this application employs mode-adaptive residual units to compensate for differences in shared representations:

[0142]

[0143] in, The purification features are those obtained after modal projection. This is a modal-specific differential residual signal. It is a single-layer feedforward network that combines shared features to suppress commonalities between modes and enhance mode uniqueness.

[0144] The final three-modal cleanup characteristics are expressed as follows:

[0145]

[0146] To further improve the semantic consistency of trimodal pairs in relational relationships, shared summaries are utilized. As a cross-modal completion signal, it is injected into the cleaned features of each modality:

[0147]

[0148] in, These are learnable parameters;

[0149] The final three-modal fusion structure of the gene pair was obtained:

[0150]

[0151] This serves as the input for the next feature fusion module.

[0152] After the above steps, this application can achieve the following advantages: (1) It realizes the structural dimensionality upgrade from single gene to gene pair, the differential features characterize the dissimilarity between genes, and the Hadamard interaction captures the synergistic effect; (2) It automatically suppresses modal repetition information and avoids the redundant accumulation caused by simple splicing; (3) It retains the key signals in modal differences and improves the structural sensitivity of the model to specific modalities; (4) It shares semantic alignment across modalities, so that the three modalities have a consistent semantic space at the gene pair level and improve the fusion quality.

[0153] As a preferred example of this application, in step S3, the unified gene pair vector obtained in step S2 is further used to construct a low-rank linear interaction and implement sample-level adaptive gating fusion, thereby obtaining the final unified gene pair representation for downstream tasks such as classification or ranking. Figure 5 As shown.

[0154] First, the projection vector and residual vector are calculated based on the three modal sub-vectors:

[0155]

[0156]

[0157] in, The dimension of the gene pair vector;

[0158] Subsequently, for any two modes projection vector First, project both onto a rank of 1. Low-rank subspace:

[0159]

[0160]

[0161] in, , d; then element-wise multiplication is performed on the low-rank subspace to capture linear interactions:

[0162]

[0163] And the low-rank interaction vector is restored to the interaction output vector through a nonlinear mapping:

[0164]

[0165] in, For element-wise nonlinear functions, For learnable parameters, For bias terms;

[0166] Three sets of interaction vectors were obtained according to the above formula. , , ;

[0167] The projection vector, residual vector, and their two associated interaction vectors of each mode are concatenated into the mode-long vector:

[0168]

[0169] This long vector is then projected onto a unified fusion dimension using a dimensionality reduction mapping:

[0170]

[0171] The three modality fusion vectors are obtained in the manner described above. , , .

[0172] To achieve sample-level adaptive fusion, this application introduces a gating network. It receives the fusion vector as input and outputs normalized weights that act on the three modes.

[0173] Specifically, construct the merged vector:

[0174]

[0175] And calculate the modal weight vector:

[0176]

[0177] Finally, a weighted summation method is used to obtain the final unified gene pair representation:

[0178]

[0179] After the above steps, the gene pair vectors are fused with low-rank bilinear interaction and sample-level gating adaptively to obtain a final unified gene pair representation that is sensitive to differences, semantically consistent and structurally stable, and this representation is used as the input for subsequent classification or ranking modules.

[0180] As a preferred example of this application, in step S4, to further improve the stability, trainability, and cross-modal collaborative capability of the fused representation, this application introduces two types of regularization terms based on the principal discriminant loss: one is the modal residual energy constraint, used to suppress the excessive amplification of residual components of different modalities after decommonimetry, thereby ensuring the robustness of each modal contribution; the other is the inter-modal interaction consistency constraint, used to adjust the alignment degree and complementary diversity of the cross-modal fused representation, promoting the collaborative expression of multimodal information, such as... Figure 6 As shown. The residual energy constraint applies to the residual vectors of the three modes. , , The expectation of the second norm is penalized, defined as:

[0181]

[0182] in This is the scaling factor.

[0183] To further constrain the statistical correlation of the fused representations between modalities, this application constructs a normalized cross-correlation matrix to simultaneously enhance the consistency of the corresponding dimensions and suppress redundant dependencies between modalities.

[0184] Specifically, for any two sets of fused vectors ,in To determine the batch size, first normalize each dimension to zero mean and unit variance:

[0185]

[0186]

[0187] Then, its normalized correlation matrix is ​​calculated:

[0188]

[0189] Based on this correlation matrix, a consistency loss is constructed: diagonal terms should be close to 1 (to improve cross-modal consistency in the corresponding dimension), and off-diagonal terms should be as close to 0 as possible (to improve the independence of different dimensions).

[0190] Therefore, the correlation penalty is defined as:

[0191]

[0192] in, The weights of the off-diagonal terms. This indicates summing the off-diagonal elements of a matrix using its norm.

[0193] Applying this constraint simultaneously to the three modal combinations yields the final interaction consistency loss:

[0194]

[0195] During the model training phase, this application jointly optimizes the principal discriminant loss with the aforementioned residual energy constraint and interaction consistency loss according to weights; wherein, the fusion vector obtained in step S3 As input to the classifier, the classifier output is denoted as... The ultimate training objective is to effectively improve the stability, interpretability, and biological robustness of modality fusion representations while ensuring discriminative performance.

[0196] This application proposes a synthetic lethality prediction method based on multimodal adaptive fusion of non-common forgetting and low-rank interactions. By unifying the semantic information of knowledge graphs, the structural information of protein-protein interaction networks, and the population collaboration information in functional association hypergraphs into a single learning framework, it achieves a multi-level and comprehensive characterization of complex relationships between genes. Its core improvement lies in moving away from relying on a single data source or simply splicing multimodal features. Instead, it effectively distinguishes truly complementary information from redundant noise information in different modalities through mechanisms such as differential construction, non-common forgetting, and low-rank interactions. This allows the model to significantly improve stability and generalization ability while maintaining expressive power. Through a sample-level adaptive gating mechanism, the fusion weights are dynamically adjusted based on the information quality of different gene pairs in each modality. This avoids the dominant influence of a particular modality failure or noise amplification on the overall prediction results. As a result, the overall modeling of gene semantic associations, topological dependencies, and functional module synergistic effects is achieved. The resulting gene pair representations can reflect both local structural differences and global functional consistency, significantly improving the overall performance of synthetic lethality prediction in terms of accuracy, robustness, and generalizability. Furthermore, it provides an intelligent technical solution for anticancer drug target screening that utilizes information more fully, provides more reliable prediction results, and has lower practical application costs.

[0197] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A synthetic lethality prediction method based on non-identity forgetting and low-rank interaction multi-modal adaptive fusion, characterized by, Comprising the following steps: S1: Obtain single gene features of at least three modalities of genes, including knowledge graph modality features, graph modality features and hypergraph modality features, to form the basic embedding representation of the genes; S2: Construct difference features and element-level Hadamard interaction features for any gene pair to upgrade the single gene embedding to gene pair features, and combine shared abstract, modality difference extraction and non-common forgetting mechanism to realize three modality purification fusion, and obtain gene pair fusion representation; S3: Construct a low-rank linear interaction model based on the gene pair fusion representation of step S2, perform sample-level adaptive gating fusion, and generate the final unified gene pair feature vector; S4: Input the gene pair feature vector of step S3 into the downstream classification or ranking model to predict the correlation between the gene pairs or the ranking score; In the prediction phase, the final fused features of any gene and are input into the predictor to obtain the probability that it is a synthetic lethal gene pair.

2. The synthetic lethality prediction method according to claim 1, characterized in that, In step S1, the semantic representation of the gene is obtained based on multi-hop attention knowledge graph embedding learning, and trainable embeddings are established for entities and relationships respectively. The entity is aggregated by multi-hop self-attention, and the neighbor splicing vector is constructed. Through: wherein, denotes an aggregated vector representation based on attention weights, , are trainable parameters; and repeating the above aggregation, finally taking , effectively capturing long-range dependencies of the knowledge graph.

3. The synthetic lethality prediction method according to claim 2, characterized in that, In step S1, the high-order topological information and local structure features of the protein interaction network are modeled by using the Chebyshev graph convolutional network, and the basic form of the graph convolution can be expressed as: wherein, denotes the Chebyshev polynomial of the denotes the scaled normalized Laplacian matrix, denotes the node feature matrix of the is a learnable parameter of the corresponding order.

4. The synthetic lethality prediction method according to claim 3, characterized in that, In step S1, the hypergraph is constructed based on the functional association data, and the high-order functional features of the genes are extracted by the hypergraph neural network. The following hypergraph Laplace-based propagation formula is used to realize the bidirectional message passing between nodes and hyperedges: where, denotes the node degree, denotes the node degree, is the hypergraph incidence matrix, is the hyperedge weight matrix, is the node feature matrix of the layer, denotes the feature dimension of the layer, is the learnable mapping, denotes the ReLU activation function.

5. The synthetic lethality prediction method according to claim 4, characterized in that, In step S1, Protein interaction data is obtained from the BioGRID database to construct a gene interaction graph, which is optimized by PCA dimensionality reduction and normalization before being input into the Chebyshev graph convolutional network to extract the multi-order topological structure features of the genes in the interaction network; The semantic association information of the genes is obtained from the SynLethKG knowledge graph, and a low-dimensional semantic vector is generated by a knowledge graph embedding model. Then, the feature alignment is performed through a linear mapping module, so as to obtain the embedding representation of the biological semantic relationship of the genes; The functional association gene set obtained from MSigDB is used to construct a high-order hyperedge structure, which is input into a hypergraph neural network to encode the high-order co-occurrence relationship within the functional module and extract the functional association features of the genes.

6. The method of predicting synthetic lethality according to any one of claims 1 to 5, wherein, In step S2, first, difference features and element-level Hadamard interaction features are constructed to simultaneously depict the difference structure and interaction structure of two genes in each modality, which are defined as follows: wherein, denotes three modalities, denotes an absolute difference operation, denotes a Hadamard product; Thus, the initial features of the gene pair in three modalities are obtained: wherein, denotes a vector level concatenation operation.

7. The method of predicting synthetic lethality according to claim 6, wherein, In step S2, the three modality features are uniformly mapped to a shared feature space and stacked; Introducing learnable attention vectors , compute importance of each modality: where • denotes the vector inner product, is a learnable attention vector, denotes a single modality-shared candidate feature; At the same time, the parts unique to each modality and beneficial to prediction are retained, and a modality adaptive residual unit is used to compensate for the differences in the shared representation: wherein, is the purified feature after modal projection, is the modal unique difference residual signal, is a one-layer feedforward network, denotes the shared abstract representation of gene pairs, which suppresses the commonality between modalities and strengthens the uniqueness of modalities in combination with shared features.

8. The method of predicting synthetic lethality according to claim 7, wherein, In step S3, the obtained unified gene pair vector is further constructed into a low-rank linear interaction: wherein, is an element-wise nonlinear function, is a learnable parameter, is a bias term; Introducing a gating network Receiving fusion vectors as input and output normalized weights that act on the three modalities: After the above steps, the gene pair vector is processed by low-rank bilinear interaction and sample-level gating adaptive fusion to obtain the final unified gene pair representation which is sensitive to differences, consistent in semantics and stable in structure.

9. The method of predicting synthetic lethality according to claim 8, wherein, In step S4, two types of regularization terms are introduced based on the main discriminant loss: one is the modality residual energy constraint, which is used to suppress the excessive amplification of the residual components of different modalities after commonality removal: wherein are scaling coefficients; The second is the inter-modal consistency constraint, which is used to adjust the alignment degree and complementary diversity of the cross-modal fusion representation: wherein the weight of the off-diagonal element, denotes taking the off-diagonal elements of the matrix and summing with the norm; The constraint is applied to three groups of modal combinations to obtain the final inter-modal consistency loss: 。

Images