Method for scoring dna damage repair based on adaptive heterogeneous graph attention network

Abstract

This invention discloses a DNA damage repair scoring method based on an adaptive heterogeneous graph attention network, comprising: S1. Acquiring genomic data and constructing a multidimensional omics feature matrix including gene expression scores, copy number variation scores, methylation scores, and somatic mutation scores; acquiring protein-protein interaction network data and constructing an adjacency matrix; S2. Capturing topological dependence features of the protein-protein interaction network through graph attention perception branches, screening high signal-to-noise ratio omics features through gating feature screening branches, and extracting deep semantic features through a multilayer perceptron; S3. Feature fusion to obtain high-dimensional fused features; S4. Outputting a predicted score. This invention solves the problem of functional directionality ambiguity, effectively preserves and enhances weakly expressed but functionally active gene signals, truncates false positive error edges, and overcomes the problem of "homogeneous redundancy".

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of bioinformatics, computational biology, and artificial intelligence. Specifically, it relates to a comprehensive scoring method for DNA damage repair that integrates multi-omics data features and fuses a heterogeneous graph attention network architecture. Background Technology

[0002] Cancer, a complex systemic disease caused by the accumulation of genomic abnormalities, is characterized by a significant increase in genomic instability. DNA damage repair (DDR) mechanisms, as a core defense for maintaining genomic integrity and stability, directly lead to the accelerated accumulation of somatic mutations when dysfunctional, thus driving tumorigenesis and development. Defects in the DDR pathway are not only initiating factors in tumorigenesis but also a key biological basis for tumor sensitivity to platinum-based drugs, PARP inhibitors, and other chemotherapy and targeted therapies. Therefore, accurately identifying key DDR genes closely related to cancer phenotypes and quantitatively assessing their repair efficacy are crucial prerequisites for a deeper understanding of tumorigenesis and development mechanisms, the development of highly effective targeted drugs, and the formulation of personalized treatment plans.

[0003] Despite some progress in gene function analysis and scoring, there are still many limitations to be addressed in the accurate scoring of DDR genes and the screening of key genes: (1) Lack of functional directionality and probabilistic interpretation in scoring: Traditional gene scoring models (such as FoldChange based on differential expression analysis, simple linear regression models, etc.) usually only output a scalar value without direction. This scalar can only characterize the “strength of association” between a gene and a disease phenotype, but cannot clearly distinguish whether the gene plays a “positive repair role” or a “negative damage promotion role” in the DDR pathway through mathematical symbols. In addition, linear scoring models cannot reflect the probability measure of a gene belonging to a certain functional category, and it is difficult to effectively deal with the threshold effect that is common in biological systems, resulting in unclear biological meaning of high-scoring genes and limited clinical guidance value.

[0004] (2) Single feature dimension and loss of weak signals: Most existing methods rely on only a single omics data (such as gene expression data) for analysis, ignoring the synergistic regulatory role of genes at different levels of molecular events, such as transcriptomics, genome, and epigenome. For example, some genes may be upregulated through copy number variation (CNV), and analyzing only expression data cannot accurately distinguish the source of regulation. More importantly, existing multi-omics integration methods mostly use simple linear weighting strategies, which cannot effectively capture the complex nonlinear relationships between features (such as the threshold inhibition effect of methylation on gene expression). This causes the weak functional signals of key genes at low abundance to be masked by background noise, resulting in the missed detection of a large number of potential key regulatory factors (especially those enzymes or scaffold proteins with low expression but high catalytic activity).

[0005] (3) Lack of a heterogeneous network topology perspective: Genes do not function in isolation, but rather form complex heterogeneous networks by participating in specific biological pathways or protein-protein interactions (PPIs). Abnormal function of a single gene can affect the functional state of the entire pathway through network transmission effects, thereby regulating tumor phenotypes. Most existing scoring methods analyze genes as independent individuals, ignoring the cascade regulatory effects between genes and the network topology dependence. For example, the expression level of a gene itself may not change much, but if its key interacting partners in the PPI network all experience significant functional abnormalities, according to the biological network principle of "one is influenced by one's surroundings," the functional state of the gene is also very likely to fluctuate drastically, and such potential functional abnormalities cannot be captured by traditional independent scoring methods.

[0006] (4) Insufficient ability of models to extract heterogeneous features: Traditional machine learning models (such as linear regression, random forest, support vector machine, etc.) have limited ability to extract features from high-dimensional, heterogeneous omics data, making it difficult to uncover deep nonlinear patterns and complex relationships in the data. In recent years, although deep learning technology has been applied to some extent in the field of bioinformatics, it still has obvious shortcomings: Convolutional Neural Networks (CNNs) are good at capturing local feature relationships, but they are difficult to handle graph data with non-Euclidean structures such as PPI networks; Graph Neural Networks (GNNs) can handle network structure data, but they often ignore the intensity differences of heterogeneous omics features within nodes; Transformer models can capture global dependencies, but they lack specificity when processing graph structure data.

[0007] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a DNA damage repair scoring method based on an adaptive heterogeneous graph attention network, enabling efficient and accurate screening of cancer driver genes. This invention is applicable to the precise discovery of key driver genes related to the maintenance of genome stability from massive high-throughput multi-omics sequencing data. It can be widely applied to various cutting-edge research and clinical applications, such as tumor molecular mechanism analysis, screening of synthetic lethal targets, drug sensitivity prediction, and precision medicine strategy development, providing core technical support for the precision diagnosis and treatment of cancer.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: a DNA damage repair scoring method based on an adaptive heterogeneous graph attention network, comprising the following steps: S1. Obtain genomic data and construct a multidimensional omics feature matrix. Multidimensional omics features include gene expression scores Copy number variation score methylation score Somatic mutation score Obtain protein-protein interaction (PPI) network data and construct an adjacency matrix. ; S2. Features are extracted using an adaptive heterogeneous graph attention network, specifically including: S2a. Multidimensional omics feature matrix and adjacency matrix As input, graph attention is used to perceive branches, and multi-head attention mechanisms are utilized to capture the topological dependence features of protein-protein interaction networks. , S2b. Multidimensional omics feature matrix As input, high signal-to-noise ratio omics features are selected through a gated feature selection branch using gated units. , S2c. Multidimensional omics feature matrix As input, semantic features of deep nonlinear patterns are extracted using a multilayer perceptron. ; S3. Topology dependency features High signal-to-noise ratio omics features semantic features Feature fusion is performed to obtain high-dimensional fused features. ; S4. Output Predicted Score .

[0010] Preferably, the attention coefficient of the multi-head attention mechanism The calculation formula is as follows:

[0011] in, This indicates the index of the attention head, and LeakyReLU is the activation function. This is a learnable attention weight vector. Indicates matrix transpose. This represents the concatenation of eigenvectors. and They are nodes and Post-projection features, This is the omics physical distance penalty term, in which To control hyperparameters, For Euclidean distance.

[0012] Preferably, step S3 is implemented using the following formula:

[0013]

[0014] in, For learnable weight parameters, Indicates bitwise Hadama product. This represents the matrix addition operator. It is a second-order cooperative crossover feature.

[0015] Preferably, the network is trained using a loss function, the formula of which is as follows:

[0016] in, This represents the number of genes involved in the loss function calculation during a single training iteration. For the first Predictive scores for each gene, For the first A gene efficacy monitoring tag; The regularization coefficient is used. For a set of learnable parameters, Performance monitoring labels are constructed in the following way:

[0017]

[0018]

[0019] in , , , These are the weighting coefficients. For the first Multidimensional omics features after gene standardization; For the first The linear synthesis strength of each gene; For all genes involved in computation Global mean; coefficient This is the value optimized through experience. This is the functional polarity coefficient.

[0020] In addition, this invention also provides a DNA damage repair scoring system based on an adaptive heterogeneous graph attention network, including a data acquisition module, an adaptive heterogeneous graph attention network module, a feature fusion module, and a prediction scoring module; the data acquisition module is used to acquire genomic data and construct a multidimensional omics feature matrix. The adaptive heterogeneous graph attention network module acquires protein-protein interaction (PPI) network data to construct an adjacency matrix. The module includes a graph attention perception branch, a gated feature selection branch, and an MLP branch. The graph attention perception branch utilizes a multi-head attention mechanism to capture the topological dependency features of the protein-protein interaction network. The gated feature screening branch uses gated units to screen high signal-to-noise ratio omics features. The MLP branch extracts semantic features of deep nonlinear patterns through nonlinear mapping. The feature fusion module is used to integrate topology dependency features. High signal-to-noise ratio omics features semantic features By fusing the components, high-dimensional fusion features are obtained. The prediction scoring module is used to output the predicted score. .

[0021] Meanwhile, the present invention also provides a DNA damage repair scoring device based on an adaptive heterogeneous graph attention network, comprising: a memory, a processor, and a DNA damage repair scoring program based on an adaptive heterogeneous graph attention network stored in the memory and executable on the processor, wherein the DNA damage repair scoring program based on the adaptive heterogeneous graph attention network is configured to implement a DNA damage repair scoring method based on the adaptive heterogeneous graph attention network.

[0022] The present invention also provides a storage medium storing a DNA damage repair scoring program based on an adaptive heterogeneous graph attention network. When executed, the DNA damage repair scoring program based on the adaptive heterogeneous graph attention network implements a DNA damage repair scoring method based on the adaptive heterogeneous graph attention network.

[0023] By adopting the above technical solution, the present invention has the following beneficial technical effects: (1) By using the hyperbolic tangent saturation mapping mechanism, the score is defined as a probability measure of the gene maintaining the stability of the genome, which avoids the subjectivity of artificially setting the equilibrium point in the traditional method and is more in line with the nonlinear law of biological systems. Furthermore, the functional polarity coefficient is introduced into the efficacy supervision label to realize bidirectional polarity mapping, so that the symbol of the predicted score is used to characterize the functional polarity of DNA damage repair-related genes, and the absolute value of the predicted score is used to characterize the strength of the corresponding functional efficacy. Thus, under the unified scoring framework, the positive repair effect and the negative damage promotion effect can be distinguished, and the directional clarity and biological interpretability of the scoring results can be improved.

[0024] (2) By constructing a scoring framework that combines multi-omics feature enhancement with heterogeneous graph modeling, on the basis of nonlinear enhancement of the original omics features, the gating screening mechanism and graph attention propagation mechanism are further combined to effectively preserve and enhance weakly expressed but functionally active gene signals, thereby improving the model’s sensitivity and scoring accuracy in identifying key DNA damage repair genes.

[0025] (3) A three-branch parallel architecture is adopted. The GAT module captures the topological association of genes in the PPI network, the Gate module filters omics intensity features with high signal-to-noise ratio, and the MLP module extracts nonlinear semantic patterns between features. The three are dynamically integrated through an adaptive weight fusion strategy, avoiding the bias of manually setting weights, realizing a comprehensive and multi-dimensional characterization of gene function, and significantly improving the accuracy and robustness of the scoring.

[0026] (4) This invention does not directly apply conventional deep learning operators, but instead innovatively injects an "omics physical distance penalty term" into the GAT attention calculation, effectively truncating false positive error edges in complex biological networks; at the same time, it abandons the conventional linear splicing method in the field during the feature fusion stage and introduces Hadamard product to construct a second-order collaborative crossover mechanism. This forces the model to learn the nonlinear physical feedback between omics strength, network topology and deep semantics from the mathematical bottom layer, completely overcoming the technical bottleneck of "homogeneous redundancy" of heterogeneous features.

[0027] (5) While outputting a comprehensive gene score, this invention introduces a graph attention mechanism and a gating feature screening mechanism into the model structure, so that the model can learn weight parameters that reflect the relationship between genes and the contribution of multi-omics features during the training process. The above weight parameters provide a basis for characterizing the source of the scoring results from the model structure, thereby improving the interpretability of the results while ensuring the accuracy of the scoring, and providing support for the analysis of DNA damage repair-related mechanisms. Attached Figure Description

[0028] Figure 1 This is a flowchart of the present invention.

[0029] Figure 2 This is a system structure diagram of the present invention.

[0030] Figure 3 This is a bar chart showing the bidirectional efficacy scores of the top 30 driver genes output by this invention. Detailed Implementation

[0031] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] Example 1: A DNA damage repair scoring method based on an adaptive heterogeneous graph attention network according to the present invention, such as... Figure 1 As shown, it includes the following steps: S1. Acquire genomic data and construct a multidimensional omics feature matrix based on the genomic data. The multidimensional omics features include gene expression scores. Copy number variation score methylation score Somatic mutation score Obtain protein-protein interaction (PPI) network data and construct an adjacency matrix. .

[0033] Specifically, S1a. In this embodiment, multi-omics data of a colorectal cancer cohort were obtained from the Cancer Genome Atlas (TCGA) database. This cohort includes 459 tumor samples and 41 paired adjacent normal control samples. All samples comply with ethical guidelines and are de-identified compliant data from the TCGA public database.

[0034] Inclusion criteria for samples include: complete clinical information, and ≥90% completeness of four types of omics data (gene expression, copy number variation, methylation, and somatic mutation). Samples with excessive missing data will be excluded.

[0035] Based on the above data, this embodiment integrates and constructs the complete DDR gene set, containing 487 genes. For the... One gene ( ,in (corresponding to the DDR gene set integrated and constructed in this embodiment), gene expression scores are calculated based on transcriptome sequencing (RNA-seq) data. Copy number variation score calculated based on GISTIC2.0 data Calculate methylation score based on methylation chip data Somatic mutation score calculated based on exon sequencing data Four types of features constitute a multidimensional omics feature vector. The Z-score is used to standardize the feature vectors to obtain the standardized feature vectors. The standardized feature vectors of all genes constitute a multidimensional omics feature matrix. .

[0036] The calculation methods for each feature are as follows: Gene expression score Fragments per kilobase of exon model per million mapped fragments (FPKM) is calculated based on RNA-seq data, reflecting transcriptional activity. The calculation formula is as follows:

[0037] in, Gene sequence number, This is the tumor sample number. This represents the total number of tumor samples. Indicates gene In the sample The amount of expression in; and They represent the first The formula calculates the mean and standard deviation of gene expression across all samples. First, it standardizes the gene expression level in each tumor sample using Z-scores. Then, it takes the absolute value of the standardized expression deviation and calculates the mean across all samples to quantify the degree to which the gene expression deviates from the overall level. A higher value indicates a more significant abnormality in gene expression.

[0038] Copy number variation score ( (This refers to data on copy number variation, reflecting the frequency and intensity of gene copy number abnormalities. The calculation formula is:)

[0039] in, Gene sequence number; and They represent the first The copy number amplification frequency and deletion frequency of each gene in all samples. The weights are as follows. The weighting is based on the following: considering that copy number amplification typically drives gene overexpression more strongly than heterozygous deletion, this embodiment assigns weights of 1.0 and 0.8 to amplification and deletion, respectively. Those skilled in the art can adjust these weights according to actual conditions.

[0040] Methylation score ( ): Based on the β value (reflecting the methylation level of CpG sites, ranging from [0,1]) of an Illumina methylation chip, this quantifies methylation abnormalities in gene promoter regions. The calculation formula is as follows:

[0041] in, Gene sequence number, This represents the number of CpG sites in the gene promoter region. The average methylation level of the k-th CpG site of gene i in the tumor sample; This represents the average methylation level of this site in normal control samples. The mean absolute value of the difference in methylation between tumor and normal samples reflects the degree of abnormal gene epigenetic regulation. The normal control samples used for methylation and gene expression characterization calculations were all paired adjacent normal tissue samples from the TCGA colorectal cancer cohort.

[0042] Somatic mutation score ( (This refers to variant recall data based on exon sequencing, combining mutation frequency and harmfulness. The calculation formula is:)

[0043] in, Gene sequence number; Indicates the first The frequency of nonsynonymous mutations in each gene in the sample population; Indicates the first The average harmfulness assessment score of all nonsynonymous mutations occurring in a gene. The frequency of nonsynonymous mutations. This refers to the number of valid tumor samples with nonsynonymous mutations in the gene divided by the total number of tumor samples. The average harmfulness assessment score can be obtained by extracting the mutation function influence prediction score (such as the PolyPhen-2 score) from variant call format (MAF) files in a publicly available database and calculating the mean.

[0044] The underlying sequencing and microarray data required for the calculation of the above four omics features can all be obtained from publicly available cancer databases such as TCGA. Those skilled in the art can perform standardized processing and parameter statistical calculations based on the above publicly available data.

[0045] Since the dimensions and distributions of the four types of features differ significantly, Z-score standardization is needed to eliminate the influence of dimensions. The formula is as follows:

[0046] In the formula, Gene sequence number, The total number of genes to be analyzed; For omics feature category numbers; and The first The first gene Scores of class features before and after standardization; and The first The global mean and standard deviation of class features across all genes involved in the calculation; To prevent extremely small constants with a denominator of zero, the eigenvectors are standardized. Each element is mapped to the same scale with a mean of 0 and a standard deviation of 1, thereby eliminating dimensional differences between features.

[0047] S1b. In this embodiment, an adjacency matrix is ​​constructed based on the PPI network obtained from the STRING (Search Tool for the Retrieval of InteractingGenes / Proteins) public database. When constructing the PPI adjacency matrix based on the STRING database, only high-confidence interaction edges with a comprehensive interaction score > 0.7 are retained, while low-confidence positive edges are filtered out; if there is an interaction edge between genes, the corresponding element is 1, otherwise it is 0, and diagonal self-loops are included to preserve their own characteristics.

[0048] S2. Extract features using an adaptive heterogeneous graph attention network.

[0049] Specifically, S2a. uses a multi-head attention mechanism through the GraphAttentionNetwork (GAT) branch to capture the topological dependency features of the protein-protein interaction (PPI) network. .

[0050] Traditional GAT operators rely entirely on the input network topology when aggregating features. However, the STRING database contains numerous "false positive" interaction edges generated from high-throughput experiments. Conventional GAT indiscriminately propagates information along these erroneous edges, causing features from irrelevant genes to interfere with the scoring of core genes, resulting in inaccurate model predictions. This invention reconstructs the core attention scoring function of GAT at a fundamental level, incorporating second-order physical constraints based on multi-omics phenotypes. The GAT branch models the topological manifold through a multi-head attention mechanism, with the specific steps as follows: Step 1: Input node features and graph structure.

[0051] The input layer contains a feature matrix. Adjacency Matrix Define nodes initial hidden features .

[0052] Step 2: Linear transformation.

[0053] Hidden features of each node Application No. Shared weight matrix for each attention head Transformation is performed (multi-head attention count in this embodiment) (The output dimension of a single attention head is 8, which can be adjusted by those skilled in the art according to actual needs.)

[0054] in, These are the projected feature vectors; For nodes Hidden layer features; is a learnable projection weight matrix.

[0055] Step 3: Calculate the nodes Attention coefficient between :

[0056] Among them, the LeakyRectifiedLinearUnit (LeakyReLU) is the activation function; This is a learnable attention weight vector. Indicates matrix transpose; This indicates the concatenation of eigenvectors; and They are nodes and Post-projection features; This is the omics physical distance penalty term, in which To control the hyperparameters, a value of 0.1 is used in this embodiment, which can be adjusted within the range of 0.05 to 0.2 based on data from different cancer types. For Euclidean distance.

[0057] This invention rigidly embeds an omics physical distance penalty term into its energy function. If two genes show significant phenotypic differences in the original omics data, then even if the database says they interact, the subtraction term will forcefully suppress their attention coefficients, thereby reducing potential erroneous association interference during information propagation and improving the ability to characterize the true functional associations.

[0058] Step 4: Softmax Normalization: Generate normalized attention weights :

[0059] in, Represents a node For neighboring nodes In the Normalized attention weights under each attention head; and Representing nodes respectively With nodes ,node With nodes In the The original attention score under each attention head; Represents a node The set of first-order neighbor nodes in a PPI topology network; For the natural constant An exponential function with base 0.

[0060] Step 5: Weighted aggregation of neighbor information: Output topological dependency features :

[0061] Among them, the Rectified Linear Unit (ReLU) is the activation function; The representative will The results of each attention point are spliced ​​together.

[0062] S2b. High signal-to-noise ratio omics features are selected through the gated feature selection branch (Gate branch). The details are as follows: Gating generation layer: Generates continuous gating weights :

[0063] Where Sigmoid is the activation function; and These are the weight matrix and bias term, and the feature matrix, respectively. It consists of the four-dimensional standardized feature vectors obtained after preprocessing in step S1.

[0064] Feature-weighted layer: Perform bitwise Hadamard Product (BFP) express):

[0065] in, This represents the weighted high signal-to-noise ratio feature. Feature projection layer: Maps to 32 dimensions, output :

[0066] Where ReLU is the activation function; and These are the projection layer weight matrix and the bias term, respectively.

[0067] By adaptively weighting different omics features, high signal-to-noise ratio information is preserved while the interference of noise features is weakened, thereby improving the ability to identify genes with low expression but active function.

[0068] S2c. The Multilayer Perceptron (MLP) branch extracts semantic features of deep nonlinear patterns through nonlinear mapping. .

[0069] MLP branches extract semantics through non-linear mappings, as follows: Hidden layer: Dimensions expanded to 128, using GELU activation function:

[0070] in, and These are the hidden layer weight matrix and the bias vector, respectively.

[0071] The GELU approximation formula is as follows:

[0072] Output layer: Compressed to 32 dimensions, output :

[0073] in, and These represent the output layer weight matrix and the bias vector, respectively. Through this mapping, the input feature matrix... Converted into a deep semantic feature matrix .

[0074] S3. Feature fusion, specifically including: S3a. Calculate second-order cooperative crossover features .

[0075] A second-order crossover mechanism is constructed by introducing bitwise Hadamard products. This mechanism forces the model to learn the nonlinear physical feedback and synergistic effects between heterogeneous features from the mathematical foundation, thus overcoming the homogenization and redundancy problem that easily occurs when heterogeneous features are fused.

[0076]

[0077] in, This represents the bitwise Hadamard Product operator, used to calculate the product of corresponding elements of two matrices of the same dimension. This represents the matrix addition operator, used to fuse and superimpose three pairs of intersecting features.

[0078] S3b. Adaptive feature fusion.

[0079] Using a set of learnable adaptive weight parameters, three basic branch features are dynamically measured and assigned. ) and second-order cross features ( The contribution of each factor to the final score. This avoids subjective bias caused by manually setting fixed weights based on experience.

[0080]

[0081] in, This represents the final high-dimensional fusion representation matrix output after adaptive weight integration. These are learnable weight parameters.

[0082] S4. Fully connected regression prediction output.

[0083] To make the scoring more consistent with real-world application scenarios, the high-dimensional final fused feature space is mapped to a one-dimensional continuous real number space, and the hyperbolic tangent function is used for boundary constraints to ensure that the predicted score matches the performance supervision label. The probability measure space is strictly aligned. The formula is as follows:

[0084] in, This represents the learnable weight matrix of the fully connected output layer in a linear regression. This represents the learnable bias vector of the fully connected output layer in a linear regression.

[0085] Preferably, a loss function can be used to train the network.

[0086] The loss function is as follows:

[0087] in, This represents the number of genes involved in the loss function calculation during a single training iteration. For the first Predictive scores for each gene, For the first A gene efficacy monitoring tag; The regularization coefficient is used. This is the set of learnable parameters. An early stopping mechanism is triggered when the validation set loss does not decrease for 10 consecutive rounds. Early stopping mechanism: When the validation set loss does not decrease for 10 consecutive rounds, training is stopped and the optimal model is saved.

[0088] The performance monitoring label Obtained through the following method: (1) Calculate the first Linear integration strength of individual genes The formula is as follows:

[0089] Among them, the weighting coefficient [ [Can be based on preferred empirical values] In practical applications, the weighting coefficients can be... Adaptive changes are made within the scope of [the relevant regulations / requirements]. For the first Multidimensional omics features after gene standardization.

[0090] Weighting is based on the following criteria: Upregulation of gene expression (0.25) is a direct manifestation of enhanced transcriptional activity and has a positive driving effect on maintaining genome stability; Copy number variation (0.3) directly promotes target gene expression through dose-effect, also providing a positive contribution; Methylation (-0.15) usually leads to epigenetic silencing in gene promoter regions, which has a negative inhibitory effect on repair function; Somatic mutation (-0.3) directly destroys the normal structure and repair catalytic function of proteins, and has a significant negative impact on repair efficiency.

[0091] (2) Introduce a variant of the sigmoid function (Sigmoid) as a nonlinear signal enhancement operator. The signal-to-noise ratio is amplified using the following formula:

[0092] In the formula, Gene sequence number; For the first The linear synthesis strength of each gene; For all genes involved in computation Global mean. Coefficient. This is an empirically optimized value, which can be adaptively adjusted based on the distribution of sample data within the range of 1 to 5. In this embodiment, it is 3.

[0093] (3) Construct a performance monitoring label system with a score range of -100 to 100. The formula is as follows:

[0094] In the formula, Gene sequence number; This is the functional polarity coefficient. If the gene... If it belongs to the core repair gene set, then If genes If it belongs to the core damage gene set, then If the function is unknown or unclear, then Genes with unknown functions are not included in the model training process; they are only used for predicting the full set of DDR gene scores after training is completed. This formula maps the scores to the interval [-100, 100], where the absolute value of the score corresponds to the posterior probability measure of the gene's function in maintaining genome stability, and the sign corresponds to the functional polarity of the gene.

[0095] In this embodiment, candidate genes were extracted and integrated and deduplicated from DNA damage repair-related pathways (including hsa03420 nucleotide excision repair, hsa03430 mismatch repair, hsa03440 homologous recombination repair, and hsa03450 non-homologous end joining) in the MSigDBv7.5 and KEGG databases to obtain a complete set of 487 DDR genes. Based on this, a core repair / damage gene set was constructed. This core repair / damage gene set is the experimentally validated gold standard gene set with clearly defined functional polarities from the complete set of genes. In this embodiment, the gold standard gene set contains 228 genes, including 186 positive core repair genes and 42 negative core damage genes. To overcome the technical shortcomings of existing scoring methods in terms of the ambiguity of gene function directionality, this invention defines the screening criteria for the gold standard genes based on cross-validation of Gene Ontology (GO) functional annotation and existing literature evidence: Positive core repair genes are limited to genes explicitly annotated in the GO database as participating in DNA damage repair or its positive regulatory processes, and whose encoded proteins are verified through direct functional experiments such as knockout, knockdown, inactivation, or loss of function to participate in the DNA repair cascade and maintain genome stability; negative core damage genes are limited to genes explicitly annotated in the GO database as participating in the negative regulatory processes of DNA repair, or whose abnormal activation, overexpression, or pathogenic alterations are verified through direct functional experiments to lead to increased DNA damage accumulation, increased mutational burden, or enhanced genome instability; genes lacking the above-mentioned direct functional directionality experimental evidence are not included in the gold standard gene set.

[0096] Dataset partitioning: The 228 gold standard core training genes that were rigorously selected and constructed above were divided into a sub-training set and a validation set in a 9:1 ratio. The sub-training set was used for model parameter learning, and the validation set was used for early stopping mechanism triggering and hyperparameter tuning. The core training gene set was used only throughout the entire model training process. The remaining 259 genes in the full set of 487 DDR genes were not used in the training process and were only used for model generalization verification.

[0097] The AdamW optimizer is used to train the adaptive heterogeneous graph attention network based on the aforementioned loss function. Early stopping is performed based on the validation set loss. The model is evaluated using mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). The model's pass / fail criterion is the coefficient of determination on the test set. Batch size B=64; AdamW optimizer initial learning rate 1e-4, using cosine annealing learning rate decay strategy, T_max=50, eta_min=1e-6. Output is based on prediction score. A list of genes sorted by absolute value.

[0098] After model training, generalization was first validated using 259 independent test genes not involved in the training within the complete DDR gene set. Once the model met the qualification criteria, it was applied to the complete set of 487 DDR genes to predict DDR functional scores across the entire spectrum. As shown in Figure 3, the system successfully identified classic high-frequency mutation core genes such as TP53 and TREX1 in the positive repair polarity, and also screened out key scaffold proteins with low abundance but extremely active function, such as SHLD2 / 3 and FAAP100. Simultaneously, it accurately identified classic oncogenic drivers such as EGFR and KRAS in the negative damage polarity. The scoring results highly matched the known functions of the genes, validating the high accuracy and sensitivity of this method. This indicates that the method proposed in this invention has high accuracy and provides an effective tool for screening key cancer genes.

[0099] This invention provides a novel solution for the precise screening of key genes for DNA damage repair by integrating multi-omics data with heterogeneous graph attention network technology. It has significant theoretical and clinical application value and is expected to be widely promoted and applied in the field of precision oncology.

[0100] Example 2: Additionally, this invention also provides a DNA damage repair scoring system based on an adaptive heterogeneous graph attention network, such as... Figure 2As shown, it includes a data acquisition module, an adaptive heterogeneous graph attention network module, a feature fusion module, and a prediction and scoring module.

[0101] The data acquisition module is used to acquire genomic data and construct a multidimensional omics feature matrix. To obtain protein-protein interaction (PPI) network data to construct an adjacency matrix. .

[0102] The adaptive heterogeneous graph attention network module includes a graph attention perception branch, a gated feature selection branch, and an MLP branch. The graph attention perception branch utilizes a multi-head attention mechanism to capture the topological dependence features of protein-protein interaction networks. The gated feature screening branch uses gated units to screen high signal-to-noise ratio omics features. The MLP branch extracts semantic features of deep nonlinear patterns through nonlinear mapping. .

[0103] The attention coefficient of the multi-head attention mechanism The calculation formula is as follows:

[0104] Where LeakyReLU is the activation function. This is a learnable attention weight vector. Indicates matrix transpose. This represents the concatenation of eigenvectors. and They are nodes and Post-projection features, Indicates the sequence number of the attention head. This is the omics physical distance penalty term, in which To control hyperparameters, For Euclidean distance.

[0105] The feature fusion module is used to integrate topology dependency features. High signal-to-noise ratio omics features semantic features By fusing the components, high-dimensional fusion features are obtained. The formula is as follows:

[0106]

[0107] in, For learnable weight parameters, Indicates bitwise Hadama product. This represents the matrix addition operator.

[0108] The prediction scoring module is used to output a prediction score. .

[0109] The system also includes a loss function used to train the network, and the formula for the loss function is as follows:

[0110] in, This represents the number of genes involved in the loss function calculation during a single training iteration. For the first Predictive scores for each gene, For the first A gene efficacy monitoring tag The regularization coefficient is . For a set of learnable parameters, Performance monitoring labels are constructed in the following way:

[0111]

[0112]

[0113] in , , , These are the weighting coefficients; For the first Multidimensional omics features after gene standardization; It is the functional polarity coefficient; For the first The linear synthesis strength of each gene; For all genes involved in computation Global mean; coefficient This is the value optimized through experience.

[0114] Furthermore, this embodiment is merely a basic description of the DNA damage repair scoring system based on adaptive heterogeneous graph attention network of the present invention. Technical details not described in detail in this embodiment can be found in the methods provided in any embodiment of the present invention, and will not be repeated here.

[0115] Example 3: Those skilled in the art will clearly understand that the systems and methods of the above embodiments can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, node packaging device, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0116] Therefore, the present invention also provides a DNA damage repair scoring device based on an adaptive heterogeneous graph attention network, comprising: a memory, a processor, and a DNA damage repair scoring program based on an adaptive heterogeneous graph attention network stored in the memory and executable on the processor, wherein the DNA damage repair scoring program based on the adaptive heterogeneous graph attention network is configured to implement a DNA damage repair scoring method based on the adaptive heterogeneous graph attention network.

[0117] In addition, the present invention also provides a storage medium storing a DNA damage repair scoring program based on an adaptive heterogeneous graph attention network.

[0118] In reality, when deploying equipment or programs, a program may execute all steps, or it may execute only one step, with multiple programs working together to achieve the full functionality. Therefore, the DNA damage repair scoring program based on adaptive heterogeneous graph attention network, when executed, implements all or some of the processes in the DNA damage repair scoring method based on adaptive heterogeneous graph attention network.

[0119] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A DNA damage repair scoring method based on adaptive heterogeneous graph attention network, characterized in that, Includes the following steps: S1. Obtain genomic data, construct a multi-dimensional omics feature matrix The multi-dimensional omics features include gene expression scores Copy number variation scores Methylation scores Somatic mutation scores Obtain protein-protein interaction (PPI) network data, construct an adjacency matrix ; S2. Features are extracted using an adaptive heterogeneous graph attention network, specifically including: S2a. The multi-omics feature matrix and the adjacency matrix As input, through the graph attention perception branch, the multi-head attention mechanism is utilized to capture protein-protein interaction network topology-dependent features , S2b. The multi-omics feature matrix is filtered by the gating feature filtering branch As input, the high signal-to-noise ratio omics features are filtered by the gating unit through the gating feature filtering branch , S2c. The multi-omics feature matrix is transformed into a latent space As input, deep nonlinear pattern semantic features are extracted by a multilayer perceptron ; S3. Topology dependency features High signal-to-noise ratio omics features semantic features Feature fusion is performed to obtain high-dimensional fused features. ; S4. Output Predicted Score .

2. The DNA damage repair scoring method based on an adaptive heterogeneous graph attention network as described in claim 1, characterized in that, The attention coefficient of the multi-head attention mechanism The calculation formula is as follows: in, This indicates the index of the attention head, and LeakyReLU is the activation function. This is a learnable attention weight vector. Indicates matrix transpose. This represents the concatenation of eigenvectors. and They are nodes and Post-projection features, For omics physical distance penalty term, among which To control hyperparameters, For Euclidean distance.

3. The DNA damage repair scoring method based on an adaptive heterogeneous graph attention network as described in claim 1, characterized in that, Step S3 is achieved through the following formula: in, For learnable weight parameters, This indicates the bitwise Hadama product. It is a second-order cooperative crossover feature.

4. The DNA damage repair scoring method based on an adaptive heterogeneous graph attention network as described in claim 1, characterized in that, The network is trained using a loss function, the formula of which is as follows: in, This represents the number of genes involved in the loss function calculation during a single training iteration. For the first Predictive scores for each gene, For the first A gene efficacy monitoring tag The regularization coefficient is . For a set of learnable parameters, Performance monitoring labels are constructed in the following way: in , , , These are the weighting coefficients. For the first Multidimensional omics features after gene standardization; For the first The linear synthesis strength of each gene; For all genes involved in computation Global mean; coefficient This is the value optimized through experience. This is the functional polarity coefficient.

5. A DNA damage repair scoring system based on an adaptive heterogeneous graph attention network, characterized in that, It includes a data acquisition module, an adaptive heterogeneous graph attention network module, a feature fusion module, and a prediction and scoring module; The data acquisition module is used to acquire genomic data and construct a multidimensional omics feature matrix. And obtain protein-protein interaction (PPI) network data to construct an adjacency matrix; The adaptive heterogeneous graph attention network module includes a graph attention perception branch, a gated feature selection branch, and an MLP branch. The graph attention perception branch uses a multi-head attention mechanism to capture the topological dependence features of protein-protein interaction networks. The gated feature screening branch uses gated units to screen high signal-to-noise ratio omics features. The MLP branch extracts semantic features of deep nonlinear patterns through nonlinear mapping. ; The feature fusion module is used to integrate topology dependency features. High signal-to-noise ratio omics features semantic features By fusing the components, high-dimensional fusion features are obtained. ; The prediction scoring module is used to output a prediction score. .

6. The DNA damage repair scoring system based on an adaptive heterogeneous graph attention network as described in claim 5, characterized in that, The attention coefficient of the multi-head attention mechanism The calculation formula is as follows: Where LeakyReLU is the activation function. This is a learnable attention weight vector. Indicates matrix transpose. This represents the concatenation of eigenvectors. and They are nodes and Post-projection features, Indicates the sequence number of the attention head. This is the omics physical distance penalty term, in which To control hyperparameters, For Euclidean distance.

7. The DNA damage repair scoring system based on an adaptive heterogeneous graph attention network as described in claim 5, characterized in that, The feature fusion module obtains high-dimensional fused features using the following formula. : in, For learnable weight parameters, This indicates the bitwise Hadama product. This represents the matrix addition operator.

8. The DNA damage repair scoring system based on an adaptive heterogeneous graph attention network as described in claim 5, characterized in that, The system also includes a loss function used to train the network, and the formula for the loss function is as follows: in, This represents the number of genes involved in the loss function calculation during a single training iteration. For the first Predictive scores for each gene, For the first A gene efficacy monitoring tag The regularization coefficient is . For a set of learnable parameters, Performance monitoring labels are constructed in the following way: in , , , These are the weighting coefficients; For the first Multidimensional omics features after gene standardization; It is the functional polarity coefficient; For the first The linear synthesis strength of each gene; For all genes involved in computation Global mean; coefficient This is the value optimized through experience.

9. A DNA damage repair scoring device based on an adaptive heterogeneous graph attention network, characterized in that, include: The system includes a memory, a processor, and a DNA damage repair scoring program based on an adaptive heterogeneous graph attention network stored in the memory and executable on the processor. The DNA damage repair scoring program based on the adaptive heterogeneous graph attention network is configured to implement the DNA damage repair scoring method based on the adaptive heterogeneous graph attention network as described in any one of claims 1-4.

10. A storage medium storing a DNA damage repair scoring program based on an adaptive heterogeneous graph attention network, wherein the DNA damage repair scoring program based on the adaptive heterogeneous graph attention network, when executed, implements the DNA damage repair scoring method based on an adaptive heterogeneous graph attention network as described in any one of claims 1-4.

Images