A neural network-based blood tumor targeted drug screening method and system

Abstract

The application belongs to the technical field of medical information processing, neural network calculation and drug screening data processing, and specifically discloses a blood tumor targeted drug screening method and system based on a neural network. The method obtains gene mutation, fusion gene, copy number change and transcription expression data of a blood tumor sample, constructs a heterogeneous graph containing a tumor clone node, a molecular abnormality node, a signal path node, a drug target node and a candidate drug node, performs hierarchical message passing through a graph neural network, generates a path gating value according to abnormal consistency and clone coverage state, and then combines the consistency constraint between the drug target and the signal path to score and sort the candidate drugs. The scheme can constrain the drug screening result by the clone source, abnormal path and target direction, reduce the interference of low correlation abnormalities on the score, and improve the structural consistency and traceability of the candidate targeted drug screening result.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing, neural network computing and drug screening data processing technology, and specifically discloses a method and system for screening targeted drugs for hematological malignancies based on neural networks. Background Technology

[0002] Screening for targeted drugs in hematological malignancies typically uses molecular testing results from patient samples as input, combined with known drug target relationships, gene mutation annotation, fusion gene interpretation, copy number alteration assessment, and transcriptional expression differential analysis to generate a ranking of candidate targeted drugs. Conventional technical solutions often employ rule matching or tabular knowledge base retrieval methods to map detected abnormal genes to existing drug targets. At the algorithmic processing level, some solutions convert features such as gene mutations, expression levels, copy numbers, and fusion genes into vectors, which are then input into classification or regression models to output drug sensitivity scores. The implementation path of these solutions is relatively straightforward: each type of molecular abnormality is treated as an independent input variable, and the statistical association between abnormal features and drugs is learned based on drug response labels in training samples. For samples with relatively simple molecular abnormalities and well-defined drug target pathways, the above solutions can provide a basic range of candidate drugs, but their processing logic still mainly remains at the point-to-point association between abnormal genes and drug targets.

[0003] In further improved techniques, multi-omics fusion models are used for drug screening in hematological malignancies. These models typically concatenate gene mutation matrices, fusion gene states, copy number fragments, transcriptional expression profiles, and pathway enrichment results to form a unified feature matrix, which is then used to predict drug response through deep neural networks, attention networks, or ensemble learning models. To enhance model interpretability, some approaches extract feature importance after training or map input features to known signaling pathways to determine the degree of association between candidate drugs and anomalous pathways. While these methods offer greater data coverage than single-gene matching, multi-source features are often calculated as parallel variables within the model, lacking unified constraints on clonal origin, pathway attribution, and drug target orientation among features. The biological relationships between different anomalous signals in the same sample are weakened to numerical correlations, and the drug scores output by the model are easily affected by high-dimensional feature noise and batch-specific differences in samples.

[0004] Hematologic malignancy samples also exhibit significant clonal heterogeneity. A single patient sample may contain a dominant clone, low-abundance subclones, and concomitant mutant cell populations. The contributions of mutations, fusion genes, copy number alterations, and expression abnormalities carried by different clones to drug response vary. Existing screening models often use methods such as allele frequency thresholding, mutation burden statistics, or clustering annotation to process clonal information, treating clonal information as an independent additional feature input to the model. This approach struggles to describe how clonal abnormalities are propagated along molecular aberration nodes, signaling pathway nodes, and drug target nodes to candidate drug nodes, and it also makes it difficult to determine whether a candidate drug's score originates from a key aberrant pathway in the dominant clone. For samples with low-abundance noisy mutations, concomitant mutations, and expression fluctuations, relying solely on feature splicing or thresholding can easily amplify the influence of non-dominant abnormalities, leading to inconsistencies between candidate drug ranking and the actual driving mechanisms of hematologic malignancies.

[0005] The main technical problem with existing technologies lies in the lack of transferable and constrained structured computational relationships among molecular abnormalities, tumor clones, signaling pathways, drug targets, and candidate drugs during the screening of targeted drugs for hematological malignancies. This makes it difficult for models to distinguish the true contribution of major clonal-driven abnormalities and low-correlation abnormalities to candidate drug scores. The technical reason for this is that conventional rule-based matching methods only focus on the direct correspondence between abnormal genes and drug targets, while multi-omics neural network methods tend to use multiple types of molecular detection data as general feature inputs, lacking a mechanism for clonal weight propagation along pathway levels and a lack of consistency constraints between drug targets and abnormal signaling pathways. Consequently, the model may misjudge low-abundance subclonal mutations, concomitant mutations, or irrelevant expression deviations as criteria for drug screening, or miss interventionable targets formed by the convergence of multiple weak abnormalities in the same signaling pathway, resulting in insufficient stability of targeted drug screening results. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for screening targeted drugs for hematologic malignancies based on neural networks, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for screening targeted drugs for hematologic malignancies based on neural networks includes: acquiring molecular detection data of hematologic malignancies samples, wherein the molecular detection data includes gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data; Based on the molecular detection data, a heterogeneous graph is constructed, which includes tumor clone nodes, molecular abnormality nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes, and edges carrying clonal contribution information and pathway association information are established between different types of nodes. The heterogeneous graph is input into the graph neural network and hierarchical message passing is performed so that the cloning weight of the tumor clone node is passed to the signaling pathway node through the molecular abnormality node, and then passed to the drug target node and the candidate drug node by the signaling pathway node. At the signal path node, a path gating value is generated based on abnormal consistency and clone coverage status; Based on the pathway gating value and the consistency constraints between the drug target node and the signaling pathway node, the candidate drug node is scored and the targeted drug screening results are output.

[0008] Preferably, when constructing the heterogeneity graph, the frequency of variant alleles, the level of mutational function impact, and the co-occurrence relationship of mutations in the gene mutation data are jointly encoded to form the mutational features of the molecular abnormal nodes; The fusion breakpoint location, fusion fragment retention status, and fusion gene expression status in the fusion gene data are jointly encoded to form the fusion features of the molecular abnormal node; The copy number change data and the transcriptional expression data are encoded to form the expression perturbation features of the molecular abnormal nodes; The mutation features, fusion features, and expression perturbation features are mapped to the tumor clone nodes according to their corresponding clonal origins, so that the same molecular abnormality node in the heterogeneous graph has both molecular abnormality attributes and clonal affiliation attributes.

[0009] Preferably, when performing the hierarchical message transmission, initial clonal weights are first assigned to adjacent molecular abnormality nodes based on the clonal abundance corresponding to the tumor clonal node, and then pathway propagation weights are generated based on the abnormality type matching relationship, expression direction relationship and pathway subordination relationship between the molecular abnormality node and the signaling pathway node. The initial cloning weights and the path propagation weights are merged into hierarchical propagation weights, and the path representations of the signal path nodes are updated using the hierarchical propagation weights. The target representation of the drug target node is updated jointly by the pathway representation of the adjacent signaling pathway nodes and the target attributes of the drug target node itself.

[0010] Preferably, when generating the pathway gating value, anomaly direction consistency features, clone coverage ratio features, and expression response direction features are extracted from multiple molecular anomaly nodes adjacent to the same signal pathway node, and the anomaly direction consistency features, clone coverage ratio features, and expression response direction features are input into the gating network; The gating network outputs the path gating value corresponding to the signal path node; The amount of information transmitted from the signal pathway node to the drug target node is weighted according to the pathway gating value, and the weighted information is used as the input feature for the drug target node to participate in the candidate drug scoring.

[0011] Preferably, when forming the tumor clone node, multiple molecular abnormalities are cloned and grouped according to the frequency of the variant alleles, the copy number change data and the expression status of the fusion gene, and a corresponding tumor clone node is established for each clone group. When establishing the edges between the tumor clone nodes and the molecular abnormality nodes, clonal abundance, abnormality detection confidence, and abnormal function impact level are encoded as clonal contribution edge features. When establishing the edge between the molecular anomaly node and the signal pathway node, the anomaly type, anomaly direction, and pathway affiliation confidence are encoded as pathway association edge features. The cloning contribution edge features and the path association edge features are used together as the edge inputs of the graph neural network.

[0012] Preferably, when updating the pathway characterization of the signaling pathway node, the first aggregate characterization from the molecular aberration node corresponding to the major clone, the second aggregate characterization from the molecular aberration node corresponding to the low abundance clone, and the third aggregate characterization from the molecular aberration node corresponding to the fusion gene aberration are calculated respectively. A pathway aggregation representation is generated based on the similarity and directional consistency among the first aggregation representation, the second aggregation representation, and the third aggregation representation; A residual connection is set between the path aggregation representation and the original path attributes of the signal path node, so that the signal path node retains the prior path attributes and receives the clone-weighted anomalous input during the graph neural network iteration process.

[0013] Preferably, when scoring candidate drugs based on the pathway gating value, target consistency constraints are established for the drug target nodes. The target consistency constraints include consistency constraints between the signaling pathway to which the target belongs and the direction of abnormal enhancement, consistency constraints between the signaling pathway to which the target belongs and the clonal coverage ratio, and consistency constraints between the downstream expression perturbation direction of the target and the inhibitory direction of the candidate drug. The drug target node representation that satisfies the target consistency constraint is passed to the adjacent candidate drug node. The drug target node representation that does not meet the target consistency constraint is attenuated according to the degree of constraint deviation before being passed to the adjacent candidate drug node.

[0014] Preferably, after scoring the candidate drug nodes, a counterfactual node blocking process is performed for each candidate drug node. The counterfactual node blocking process includes sequentially blocking the tumor clone node, the molecular abnormality node, and the signaling pathway node related to the candidate drug node score, and re-executing the message passing from the signaling pathway node to the candidate drug node after each blocking. Node contribution weights are generated based on the difference in candidate drug scores before and after shielding. The node contribution weights and the scores of the candidate drug nodes are used together to determine the supporting clones, supporting anomalies, and supporting signaling pathways in the targeted drug screening results.

[0015] Preferably, before outputting the targeted drug screening results, a consistency check is performed on the node contribution weights. The consistency check includes determining whether the tumor clone node with the highest contribution is connected to the signaling pathway node with the highest contribution through at least one of the molecular abnormality nodes, determining whether the signaling pathway node with the highest contribution is connected to the candidate drug node through a drug target node that satisfies the target consistency constraint, and determining whether the score of the candidate drug node originates from multiple supporting molecular abnormalities within the same clonal propagation chain. Only the ranking results of candidate drugs that meet the consistency check are written into the targeted drug screening results.

[0016] A neural network-based targeted drug screening system for hematologic malignancies includes: a data acquisition module for acquiring molecular detection data of hematologic malignancies samples, wherein the molecular detection data includes gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data; The heterogeneity graph construction module is used to construct a heterogeneity graph based on the molecular detection data. The heterogeneity graph includes tumor clone nodes, molecular abnormality nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes, and establishes edges carrying clonal contribution information and pathway association information between different types of nodes. The graph neural network processing module is used to perform hierarchical message passing on the heterogeneous graph, so that the cloning weight of the tumor clone node is passed to the drug target node and the candidate drug node through the molecular abnormality node and the signaling pathway node. The path gating processing module is used to generate path gating values ​​based on abnormal consistency and clone coverage status. The drug scoring module is used to score the candidate drug nodes and output the targeted drug screening results based on the pathway gating value and the consistency constraints between the drug target node and the signaling pathway node.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a heterogeneous graph containing tumor clonal nodes, molecular aberration nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes from gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data of hematological malignancies, clonal origin, aberration type, pathway attribution, and drug action relationship are all incorporated into the same graph structure for calculation. Clonal weights are passed from molecular aberration nodes to signaling pathway nodes, ensuring that the impact of different molecular aberrations on pathway characterization is constrained by clonal abundance and aberration attribution. Pathway gating values ​​are generated based on aberration direction consistency, clonal coverage status, and expression response direction, allowing mutually supporting aberration features within the same signaling pathway to participate in subsequent drug scoring. Consistency constraints between drug target nodes and signaling pathway nodes require candidate drug scores to simultaneously meet target attribution, aberration direction, and clonal coverage conditions. This processing chain ensures that drug screening results are no longer determined solely by a single aberrant gene or ordinary splicing features, reducing interference from low-abundance noise aberrations, accompanying mutations, and irrelevant expression deviations on candidate drug scores, and allowing drug targets formed by the convergence of multiple weak aberrations at the pathway level to be included in the screening process.

[0018] 2. By encoding the clonal contribution edge features between tumor clonal nodes and molecular aberration nodes, and the pathway association edge features between molecular aberration nodes and signaling pathway nodes, the graph neural network can retain limiting information such as clonal abundance, aberration detection confidence, aberration functional impact level, aberration direction, and pathway affiliation confidence when updating pathway and target representations. By forming aggregated representations for major clones, low-abundance clones, and fusion gene aberrations before updating pathway representations, direct mixing of aberrations from different sources in the same computational layer can be avoided. Through counterfactual node masking, candidate drug scores can establish a correspondence with supporting clones, supporting molecular aberrations, and supporting signaling pathways; through consistency verification, candidate drug ranking results need to form a closed association along the same clonal propagation link through molecular aberrations, signaling pathways, and drug targets. These processes provide clearer computational basis for the output results, facilitating the exclusion of abnormally high scores caused by isolated node contributions, and improving the consistency of hematological malignancy targeted drug screening results under different sample molecular complexities. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of a neural network-based targeted drug screening method for hematological malignancies. Figure 2 Flowchart for constructing heterogeneous graphs and encoding node and edge features; Figure 3 A flowchart illustrating hierarchical message passing, path gating, and target consistency constraints in a graph neural network. Figure 4 The flowchart shows the process of scoring candidate drugs, masking counterfactual nodes, verifying consistency, and outputting results. Detailed Implementation

[0020] In one embodiment, reference Figure 1 This paper provides a neural network-based method for screening targeted drugs for hematological malignancies. The method includes acquiring molecular detection data from hematological malignancies samples, including gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data. Based on the molecular detection data, a heterogeneous graph is constructed. This graph includes tumor clone nodes, molecular anomaly nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes. Edges carrying clonal contribution information and pathway association information are established between different types of nodes. The heterogeneous graph is then input into a graph neural network for hierarchical message passing, allowing the clonal weights of the tumor clone nodes to be passed to the signaling pathway nodes via the molecular anomaly nodes. The signaling pathway node transmits data to the drug target node and the candidate drug node. At the signaling pathway node, a pathway gating value is generated based on anomaly consistency and clonal coverage status. Based on the pathway gating value and the consistency constraint between the drug target node and the signaling pathway node, the candidate drug node is scored and the targeted drug screening result is output. In this embodiment, molecular anomalies, clonal origins, signaling pathways, drug targets, and candidate drugs that are scattered in blood tumor samples are placed in the same graph structure for calculation, so that the candidate drug score forms a closed processing relationship along the link from clone to anomaly, anomaly to pathway, pathway to target, and target to drug.

[0021] In this embodiment, the molecular detection data undergoes unified semantic encoding before entering the isomorphism graph. Gene mutation data is converted into mutation vectors containing mutation type, mutation functional impact, variant allele frequency, and mutation co-occurrence relationship. Fusion gene data is converted into fusion vectors containing fusion breakpoints, fusion fragment preservation states, and fusion expression states. Copy number change data is converted into copy number vectors containing amplification, deletion, and fragment stability states. Transcriptional expression data is converted into expression vectors containing expression deviation direction, expression deviation intensity, and pathway-related expression responses. The mutation vector, the fusion vector, the copy number vector, and the expression vector together form the... The initial features of molecular abnormality nodes are as follows: the initial features of tumor clone nodes are formed by the set of abnormalities associated with the clone, clonal abundance, and intraclonal abnormality consistency; the initial features of signaling pathway nodes are formed by the aggregation state of abnormalities within the pathway and prior attributes of the pathway; the initial features of drug target nodes are formed by the pathway to which the target belongs, the direction of target regulation, and the interaction relationship between the target and candidate drugs; and the initial features of candidate drug nodes are formed by the set of drug action targets and the training representation corresponding to known drug response labels. In this embodiment, the input data is not entered into the model in the ordinary feature splicing method, but the hierarchical relationship of the molecular mechanism of hematologic malignancies is maintained through node type and edge type.

[0022] In this embodiment, the system can be configured with a data acquisition module, a heterogeneous graph construction module, a graph neural network processing module, a pathway gating processing module, and a drug scoring module. The data acquisition module acquires molecular detection data from blood tumor samples, including gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data. The heterogeneous graph construction module constructs a heterogeneous graph based on the molecular detection data. This heterogeneous graph includes tumor clone nodes, molecular abnormality nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes, and establishes edges carrying clonal contribution information and pathway association information between different types of nodes. The graph neural network... The network processing module is used to perform hierarchical message passing on the heterogeneous graph, so that the clonal weight of the tumor clone node is passed to the drug target node and the candidate drug node through the molecular anomaly node and the signaling pathway node. The pathway gating processing module is used to generate a pathway gating value based on the anomaly consistency and clonal coverage status. The drug scoring module is used to score the candidate drug node and output the targeted drug screening result based on the pathway gating value and the consistency constraint between the drug target node and the signaling pathway node. The system structure and method processing link of this embodiment are consistent, so that each type of processing module corresponds to a necessary calculation relationship in the graph structure.

[0023] In this embodiment, the clonal contribution edge weight is used to represent the degree of contribution of a molecular abnormality to its corresponding tumor clone. During calculation, clonal abundance, abnormality detection confidence, and abnormal functional impact are all included in the normalization process. The normalized edge weight serves as the input to the graph neural network for transmitting information from the tumor clone node to the molecular abnormality node. The formula for calculating the clonal contribution edge weight is: ; Where ω represents the clonal contribution edge weight, φ represents clonal abundance, λ represents the anomaly detection confidence, μ represents the functional impact of the anomaly, and the denominator represents the sum of the products of all adjacent molecular anomalies at the same tumor clonal node. For example, φ is 0.70 for a major clone, λ is 0.90 and μ is 0.85 for a driver mutation anomaly, and λ is 0.60 and μ is 0.30 for another accompanying anomaly in the same clone. Under the same conditions, the driver mutation anomaly has a higher ω and occupies a higher input weight in subsequent pathway characterization. The resulting technical effect is that clonal abundance and functional impact of the anomaly are constrained simultaneously, and low-confidence accompanying anomalies cannot enter the candidate drug scoring main link solely based on the detection results.

[0024] In a preferred embodiment, reference Figure 2When constructing the heterogeneity graph, the frequency of variant alleles, the level of mutational function impact, and the co-occurrence relationship of mutations in the gene mutation data are jointly encoded to form the mutational features of the molecular abnormality nodes. The fusion breakpoint location, fusion fragment preservation status, and fusion gene expression status in the fusion gene data are jointly encoded to form the fusion features of the molecular abnormality nodes. The copy number change data and the transcriptional expression data are encoded to form the expression perturbation features of the molecular abnormality nodes. The mutational features, fusion features, and expression perturbation features are mapped to the tumor clone nodes according to their corresponding clonal origins, so that the same molecular abnormality node in the heterogeneity graph simultaneously possesses molecular abnormality attributes and clonal attribution attributes. Specifically, variants... Allele frequency is used to infer the attribution of molecular abnormalities to tumor clones; mutation functional impact level is used to describe the strength of the impact of the abnormality on protein-coding or regulatory sequences; mutation co-occurrence relationship is used to describe the co-occurrence of multiple abnormalities within the same clone; fusion breakpoint location is used to define the protein structural regions that may be altered by the fusion event; fusion fragment preservation status is used to define whether the functional domain is preserved after fusion; fusion gene expression status is used to define whether the fusion event forms a computable response at the transcriptional level; expression deviation coding is used to unify copy number changes and transcriptional expression direction to the same abnormal perturbation semantics. In this embodiment, the same molecular abnormality node has both abnormal content and clonal origin, enabling the heterogeneity map to distinguish different inputs from major clones and low-abundance subclones within the same pathway. The heterogeneity map data mapping for screening targeted drugs for hematologic malignancies is shown in Table 1 below.

[0025] Table 1. Heterogeneous Mapping Table for Screening Targeted Drugs for Hematologic Malignancies ; The data mapping relationships in Table 1 are used to define the node and edge positions of different molecular detection data after entering the heterogeneous graph. Gene mutation data and copy number change data mainly participate in clonal contribution calculation, while fused gene data and transcriptional expression data mainly participate in pathway association calculation. The pathway-target relationship and the drug-target relationship are used to constrain the subsequent propagation path of candidate drug scoring. In this embodiment, the mapping method shown in Table 1 avoids compressing multi-source detection data into unstructured planar vectors, so that the participation positions of different data objects in the graph neural network correspond to the calculation logic of hematological malignancy targeted drug screening.

[0026] In a preferred embodiment, when forming the tumor clone node, multiple molecular abnormalities are grouped into clone groups based on the frequency of the variant alleles, the copy number change data, and the expression status of the fusion gene, and a corresponding tumor clone node is established for each clone group. When establishing the edge between the tumor clone node and the molecular abnormality node, the clonal abundance, abnormality detection confidence, and abnormality functional impact level are encoded as clonal contribution edge features. When establishing the edge between the molecular abnormality node and the signaling pathway node, the abnormality type, abnormality direction, and pathway affiliation confidence are encoded as pathway association edge features. The clonal contribution edge features and the pathway association edge features together serve as the basis for the formation of the tumor clone node. Specifically, the clonal grouping of the graph neural network is not directly determined by a single variant allele frequency threshold. Instead, the variant allele frequency and copy number status are jointly corrected, and the expression status of the fusion gene is used to determine whether the fusion abnormality participates in the same clonal group. When multiple abnormalities have similar distributions and stable co-occurrence relationships within the corrected clonal ratio range, multiple abnormalities are assigned to the same clonal node. When a certain abnormality does not form a corresponding response at the expression level, its abnormality detection confidence is downgraded in the clonal contribution side features. In this embodiment, clonal grouping and side feature encoding together restrict low-correlation abnormalities from entering the main propagation link, avoiding the direct equivalence of a single detection signal to drug screening criteria.

[0027] In another preferred embodiment, reference Figure 3 When performing the hierarchical message passing, initial clonal weights are first assigned to adjacent molecular abnormality nodes based on the clonal abundance corresponding to the tumor clone node. Then, pathway propagation weights are generated based on the abnormality type matching relationship, expression direction relationship, and pathway dependency relationship between the molecular abnormality node and the signaling pathway node. The initial clonal weights and the pathway propagation weights are merged into hierarchical propagation weights, and the pathway representation of the signaling pathway node is updated using the hierarchical propagation weights. The target representation of the drug target node is updated jointly by the pathway representations of adjacent signaling pathway nodes and the target attributes of the drug target node itself. Specifically, the graph neural network processing module retains the node type identifier and edge type identifier in each iteration. The tumor clone node only outputs clonal weights to adjacent molecular abnormality nodes, the molecular abnormality node only outputs perturbation representations carrying abnormal directions to adjacent signaling pathway nodes, the signaling pathway node only outputs pathway representations after pathway aggregation to adjacent drug target nodes, and the drug target node then outputs target representations after target consistency constraints to candidate drug nodes. In this embodiment, the message passing direction is consistent with the causal link of drug screening, reducing unconstrained propagation between cross-layer nodes.

[0028] In this embodiment, the pathway propagation weight is used to limit the intensity of information transmission from molecular anomaly nodes to signal pathway nodes. During calculation, the anomaly type matching relationship, anomaly expression direction relationship, and pathway dependency confidence are combined into a propagation coefficient. The formula for calculating the pathway propagation weight is: ; Where π represents the pathway propagation weight, α represents the matching coefficient between the abnormality type and the signaling pathway, β represents the direction coefficient between the abnormal expression direction and the pathway activation or inhibition direction, χ represents the dependency confidence between the abnormality and the signaling pathway, and the denominator represents the sum of the products of all molecular abnormalities adjacent to the same signaling pathway node. For example, for a certain fusion gene abnormality, α is 0.80, β is 0.75, and χ is 0.90; for a certain expression deviation abnormality without a clear pathway affiliation, α is 0.30, β is 0.40, and χ is 0.35. After normalization, the previous abnormality obtains a higher π and provides a higher contribution to the pathway characterization. The resulting technical effect is that abnormal inputs must simultaneously meet three conditions: type relevance, direction consistency, and clear pathway affiliation in order to participate in pathway aggregation.

[0029] Furthermore, when updating the pathway representation of the signaling pathway node, a first aggregated representation from the molecular anomalous node corresponding to the major clone, a second aggregated representation from the molecular anomalous node corresponding to the low-abundance clone, and a third aggregated representation from the molecular anomalous node corresponding to the fusion gene anomalous are calculated respectively. A pathway aggregated representation is generated based on the similarity and directional consistency among the first, second, and third aggregated representations. A residual connection is established between the pathway aggregated representation and the original pathway attributes of the signaling pathway node, allowing the signaling pathway node to retain prior pathway attributes and receive clone-weighted anomalous inputs during the graph neural network iteration process. Specifically, the first aggregated characterization is formed by weighting the molecular anomalies of the major clone according to the clonal contribution edge weight and the pathway propagation weight. The second aggregated characterization is formed by weighting the molecular anomalies of the low-abundance clone according to the same category weight. The third aggregated characterization is formed by weighting the fusion gene anomalies according to the fusion breakpoint, the fusion fragment retention state, and the fusion expression state. When the first aggregated characterization and the third aggregated characterization are in the same direction, the pathway aggregated characterization is retained as an input that can be propagated to the target layer. When the second aggregated characterization conflicts with the first aggregated characterization in direction and the clonal coverage ratio is low, the second aggregated characterization only enters the pathway characterization as an additional perturbation. In this embodiment, the major clone, the low-abundance clone, and the fusion anomaly will not be mixed indiscriminately in the same aggregated layer.

[0030] In another preferred embodiment, when generating the pathway gating value, anomaly direction consistency features, clonal coverage ratio features, and expression response direction features are extracted from multiple molecular anomaly nodes adjacent to the same signaling pathway node. These features are then input into a gating network, which outputs the pathway gating value corresponding to the signaling pathway node. The amount of information transmitted from the signaling pathway node to the drug target node is weighted based on the pathway gating value, and this weighted information is used as the input feature for the drug target node in candidate drug scoring. Specifically, the anomaly direction consistency feature indicates whether multiple molecular anomalies in the same pathway jointly point to a pathway activation or inhibition state; the clonal coverage ratio feature indicates the coverage area of ​​clones carrying related anomalies in the sample; and the expression response direction feature indicates whether downstream expression changes in the pathway correspond to upstream anomaly directions. The gating network can be a multilayer sensing network with normalized output. The input end receives the above three types of features and their interaction terms, and the output end generates a gating value limited to a closed interval. In this embodiment, the information propagated from the signal pathway to the drug target is not directly triggered by a single anomaly, but is jointly determined by the directional relationship and clonal coverage relationship of multiple anomalies in the same pathway.

[0031] In this embodiment, the pathway gating value is used to control the amount of pathway information output from the signal pathway node to the drug target node. During calculation, the consistency of the abnormal direction, the clonal coverage ratio, and the expression response direction are all input into the nonlinear gating function. The formula for calculating the pathway gating value is: ; Where γ represents the path gating value, and σ represents the nonlinear function that maps the input to a closed interval. , , and The gating weights obtained from training are represented by δ, ψ, ε, and ε. The horizontal line represents the weighted average of adjacent anomalies in the same signaling pathway. For example, when δ is 0.82, ψ is 0.76, ε is 0.70 and the interaction term is positive in a certain signaling pathway, γ moves to a higher range and allows the pathway representation to enter the target layer. When δ is 0.20 and ψ is 0.15, even if a certain anomaly has a high expression deviation, γ is still in a lower range. The resulting technical effect is that isolated expression fluctuations cannot drive the candidate drug score alone.

[0032] Furthermore, when scoring candidate drugs based on the pathway gating value, target consistency constraints are established for the drug target nodes. These constraints include consistency constraints between the signaling pathway to which the target belongs and the direction of abnormal enhancement, consistency constraints between the signaling pathway to which the target belongs and the clonal coverage ratio, and consistency constraints between the downstream expression perturbation direction of the target and the inhibitory direction of the candidate drug. The representations of drug target nodes that satisfy the target consistency constraints are passed to adjacent candidate drug nodes. The representations of drug target nodes that do not satisfy the target consistency constraints are attenuated according to the degree of constraint deviation before being passed to adjacent candidate drug nodes. Specifically, for drug nodes, the consistency constraint between the signaling pathway to which the target belongs and the direction of abnormal enhancement is used to determine whether the candidate drug target is located in the pathway direction driven by molecular abnormality. The consistency constraint between the signaling pathway to which the target belongs and the clonal coverage ratio is used to determine whether the target score is supported by the main clone or an abnormal link with sufficient coverage. The consistency constraint between the downstream expression perturbation direction of the target and the candidate drug inhibition direction is used to determine whether the drug inhibition direction is opposite to or compatible with the downstream expression change in the sample. In this embodiment, the drug target cannot enter the high-scoring link simply because of the existence of a drug-action relationship, but must form a co-directional constraint with the abnormal pathway and clonal coverage.

[0033] In this embodiment, the target consistency coefficient is used to indicate whether a drug target node satisfies the joint constraints of pathway anomaly direction, clonal coverage status, and candidate drug action direction. The formula for calculating the target consistency coefficient is: ; Wherein, κ represents the target consistency coefficient, γ represents the pathway gating value corresponding to the signaling pathway node adjacent to the drug target, r-assignment represents the degree of matching between the target and the adjacent abnormal signaling pathway, r-direction represents the consistency between the target regulation direction and the abnormal enhancement direction, r-coverage represents the coverage of the related abnormality in the tumor clone, and r-inhibition represents the correspondence between the candidate drug action direction and the downstream expression perturbation direction. For example, when γ is 0.78, r-assignment is 0.90, r-direction is 0.85, r-coverage is 0.75, and r-inhibition is 0.80 for a certain target, κ remains in the range that can be transmitted to the candidate drug node. When r-direction is 0.20 or r-inhibition is 0.10, κ decreases accordingly and the corresponding target characterization is attenuated. The resulting technical effect is that the drug target score is consistent with the abnormal pathway direction within the sample. The consistency verification of candidate drug scores is shown in Table 2 below.

[0034] Table 2. Candidate Drug Scoring Consistency Verification Table ; Table 2 illustrates the verification objects, verification inputs, verification relationships, and output status before candidate drug scores are written into the output results. The clonal layer, anomaly layer, pathway layer, target layer, and drug layer are placed in the same link for layer-by-layer verification. If the score of a candidate drug node mainly comes from isolated targets or isolated expression deviations, the corresponding output status is marked as an incomplete link or excluded. If the score of a candidate drug node comes from multiple supporting molecule anomalies within the same clonal propagation link, the corresponding candidate drug can be written into the ranking results. In this embodiment, the output results correspond to the supporting links in the graph structure.

[0035] In another preferred embodiment, reference Figure 4 When scoring the candidate drug nodes, the drug target representation that satisfies the target consistency constraint is fused with the drug action relationship representation of the candidate drug node. The drug action relationship representation includes a training representation formed by the set of drug targets, the direction of action, and known drug response labels. During fusion, the target consistency coefficient is used as the input weight of the target representation, and the contributions of multiple drug targets to the same candidate drug are weighted and aggregated. The calculation formula for the candidate drug score is as follows: ; Where ρ represents the candidate drug score, κ represents the target consistency coefficient, η represents the strength of the interaction between the drug target and the candidate drug, τ deviation represents the target deviation penalty term that does not meet the consistency constraint, and σ represents the function that maps the aggregation result to a closed interval. For example, if a candidate drug connects two drug targets, where the first target has κ of 0.74 and η of 0.85, and the second target has κ of 0.30 and η of 0.50, and τ deviation is 0.10, then the first target makes the main contribution to ρ. If both targets have high τ deviation, then ρ is suppressed. The resulting technical effect is that the candidate drug ranking is bound to the target link that meets the consistency constraint.

[0036] In a further embodiment, after scoring the candidate drug nodes, a counterfactual node masking process is performed for each candidate drug node. The counterfactual node masking process includes sequentially masking the tumor clone nodes, molecular abnormality nodes, and signaling pathway nodes related to the candidate drug node score, and re-executing message passing from the signaling pathway node to the candidate drug node after each masking. A node contribution weight is generated based on the difference in candidate drug scores before and after masking. The node contribution weight and the candidate drug node score are used together to determine the supporting clones, supporting abnormalities, and supporting signaling pathways in the targeted drug screening results. Specifically, when masking tumor clone nodes, other clones and their abnormal links in the same sample are retained; when masking molecular abnormality nodes, other abnormal inputs in the same pathway are retained; when masking signaling pathway nodes, the connection between the drug target and other pathways is retained. The difference between the re-passed score and the unmasked score is calculated. The larger the difference, the higher the contribution of the corresponding node to the candidate drug score. In this embodiment, the candidate drug output results can correspond to specific clones, abnormalities, and pathway support links.

[0037] In this embodiment, the counterfactual node contribution weight is used to measure the impact of a node being blocked on the candidate drug score. The formula for calculating the counterfactual node contribution weight is as follows: ; Where Δ represents the contribution weight of the counterfactual node, ρ_drug_unmasked represents the score of the candidate drug node in the original heterogeneous graph, and ρ_drug_masked represents the candidate drug score recalculated after masking a certain tumor clone node, molecular abnormal node, or signaling pathway node. For example, if the unmasked ρ of a candidate drug is 0.81, and the ρ after masking the main clone node is 0.32, then the Δ of the corresponding main clone node is 0.49. If the ρ after masking the low abundance accompanied abnormal node is 0.77, then the Δ of the corresponding accompanied abnormal node is 0.04. The resulting technical effect is that the main supporting nodes and low contribution nodes can be distinguished through the same scoring link.

[0038] In a further embodiment, before outputting the targeted drug screening results, a consistency check is performed on the node contribution weights. The consistency check includes determining whether the tumor clone node with the highest contribution is connected to the signaling pathway node with the highest contribution through at least one of the molecular abnormality nodes; determining whether the signaling pathway node with the highest contribution is connected to the candidate drug node through a drug target node that satisfies the target consistency constraint; and determining whether the score of the candidate drug node originates from multiple supporting molecular abnormalities within the same clonal propagation chain. Only the candidate drug ranking results that satisfy the consistency check are written into the targeted drug screening results. Specifically, the consistency check is performed after the counterfactual node masking process is completed. If there is no traceable molecular abnormality connection between the highest contributing clone and the highest contributing signaling pathway, the corresponding candidate drug ranking result is not written into the final output. If the drug target between the highest contributing signaling pathway and the candidate drug does not satisfy the target consistency constraint, the corresponding candidate drug ranking result is stopped from output. If the candidate drug score is supported only by a single isolated abnormality and lacks other supporting molecular abnormalities within the same clonal propagation chain, the corresponding candidate drug ranking result is not included in the final ranking. In this embodiment, the output results must be supported by a complete chain consisting of clones, abnormalities, pathways, targets, and drugs.

[0039] In one embodiment for training and deployment, the graph neural network processing module can be trained using historical samples of hematologic malignancies with drug response labels. The training samples include molecular detection data, a set of candidate drugs, and corresponding drug response results. During training, the difference between candidate drug scores and drug response labels is used as the basic loss, the difference between pathway gating values ​​and pathway anomalous consistency is used as the pathway constraint loss, the deviation between the target consistency coefficient and the direction of drug target action is used as the target constraint loss, and the deviation between the counterfactual node contribution weight and the support link integrity is used as the link constraint loss. After training, the node encoding method, edge feature encoding method, graph neural network parameters, gating network parameters, and drug scoring parameters are fixed. When new hematologic malignancies are input, a heterogeneous graph is constructed according to the same data mapping rules, and the candidate drug ranking is output. In this embodiment, the training constraints and inference process use the same clone-to-drug propagation link, so that the model parameters are learned around the structured drug screening relationship.

[0040] In this embodiment, the basic loss can be the difference function between the candidate drug score and the drug response label; the pathway constraint loss can be the difference function between the pathway gating value and the abnormal consistency label; the target constraint loss can be the difference function between the target consistency coefficient and the target orientation label; and the link constraint loss can be the difference function between the node contribution weight and the link integrity label. Each loss term is weighted and summed to form the training objective. The weights in the training objective are determined by the stability evaluation on the validation samples. The stability evaluation includes the consistency of candidate drug ranking in duplicate samples within the same hematologic malignancy type, the retention of node contribution weights within the same clonal propagation link, and the fluctuation of pathway gating values ​​under different batches of molecular detection data. In this embodiment, the training phase not only fits the drug response label but also constrains pathway gating, target consistency, and link integrity, so that the candidate drug score in the inference phase has corresponding intermediate calculation basis.

[0041] In one embodiment addressing missing data, when some molecular detection data is missing in a blood tumor sample, the heterogeneous graph construction module does not directly delete the corresponding level nodes. Instead, it generates computable node representations based on the obtained data and sets missing identifiers for the source of the missing data. If fusion gene data is missing, no fusion abnormality representations are generated in the molecular abnormality nodes, but other mutation features and expression perturbation features are retained. If copy number change data is missing, the clonal grouping of the tumor clone nodes is based on the frequency of variant alleles and the expression response relationship, and a copy number missing identifier is added to the clonal contribution edge feature. If transcriptional expression data is missing, the expression response direction feature in the pathway gating value participates in the gating network calculation with the missing identifier instead of using the inferred value. The missing identifier participates in learning as an input dimension during the training phase, enabling the graph neural network to distinguish between true low response and missing data. In this embodiment, missing data will not destroy the heterogeneous layer level structure, nor will it misprocess the missing state as a no-abnormal state.

[0042] In one embodiment for updating the candidate drug set, when the candidate drug set changes, candidate drug nodes can be added to the existing heterogeneous graph based on the drug target set and direction of action. The drug scoring module does not change the construction method of tumor clone nodes, molecular abnormality nodes, signaling pathway nodes, and drug target nodes. It only generates new drug edge features based on the interaction relationship between the newly added candidate drug and the existing drug target nodes. If the drug target node connected to the newly added candidate drug already exists in the heterogeneous graph, the corresponding target representation and target consistency coefficient are directly used for scoring. If the drug target node connected to the newly added candidate drug does not exist, a new drug target node is generated based on the signaling pathway to which the target belongs, the target regulation direction, and the interaction relationship between the target and the candidate drug. The drug target node is then included in the consistency constraint. In this embodiment, the candidate drug set can be expanded without changing the clone-to-pathway propagation mechanism. The newly added drugs are still limited by pathway gating and target consistency constraints.

[0043] In one embodiment oriented towards output results organization, the targeted drug screening results include candidate drug ranking, candidate drug scores, supporting clones, supporting molecular anomalies, supporting signaling pathways, drug targets that meet consistency constraints, and counterfactual node contribution weights. The candidate drug ranking is generated from high to low by ρ drugs. Supporting clones are determined by tumor clone nodes with the highest counterfactual node contribution weights and that pass consistency verification. Supporting molecular anomalies are determined by molecular anomaly nodes connected to supporting clones and supporting signaling pathways within the same clonal propagation chain. Supporting signaling pathways are jointly determined by pathway gating values ​​and node contribution weights. Drug targets that meet consistency constraints are determined by drug target nodes where κ targets are in a reserved state. During output, candidate drug ranking results that do not meet consistency verification are excluded or marked as incomplete links. In this embodiment, the output content comes directly from the graph structure and scoring calculation process, which facilitates the review of the drug screening basis.

[0044] In one embodiment oriented towards system collaborative operation, the data acquisition module receives molecular detection data from hematological tumor samples and then transfers the data to the heterogeneous graph construction module. The heterogeneous graph construction module generates nodes, edges, and edge features according to the data mapping relationship shown in Table 1. The graph neural network processing module performs hierarchical message passing according to clonal contribution information and pathway association information. The pathway gating processing module generates pathway gating values ​​based on abnormal direction consistency features, clonal coverage ratio features, and expression response direction features. The drug scoring module generates targeted drug screening results based on target consistency constraints, candidate drug node representations, and counterfactual node contribution weights. The modules within the system transmit structured node representations, edge features, pathway gating values, target consistency coefficients, candidate drug scores, and node contribution weights. No external control objects unrelated to hematological tumor targeted drug screening are used. In this embodiment, the system processing objects, data flow, and computation links correspond to the method implementation process.

[0045] In one embodiment for controlling anomalous noise, the heterogeneous graph construction module sets an anomalous detection confidence level for each anomalous node when generating molecular anomalous nodes. The anomalous detection confidence level is jointly encoded by detection consistency, in-sample supporting read state, expression response state, and clonal co-occurrence state. The graph neural network processing module uses the anomalous detection confidence level when calculating clonal contribution edge weights and pathway propagation weights. The pathway gating processing module only includes anomalous nodes that meet the basic confidence conditions in the direction statistics when extracting anomalous direction consistency features. The drug scoring module applies a deviation penalty to the target representation formed by low-confidence anomalous nodes when calculating candidate drug scores. The counterfactual node masking process ensures that if the candidate drug score changes very little after masking low-confidence anomalous nodes, the low-confidence anomalous nodes are not written into supporting molecular anomalous nodes. In this embodiment, the impact of noise anomalous nodes on the output results is limited at multiple levels, including node generation, edge weight calculation, pathway gating, drug scoring, and counterfactual interpretation.

[0046] In an embodiment oriented towards multi-pathway intersection, if the same molecular anomaly node connects to multiple signaling pathway nodes, the pathway propagation weight is calculated separately according to the anomaly type matching relationship, expression direction relationship, and pathway subordination relationship of each signaling pathway. The same molecular anomaly node transmits different pathway propagation weights to different signaling pathway nodes. If multiple signaling pathway nodes connect to the same drug target node, the drug target node receives the pathway representations of multiple signaling pathway nodes and calculates the target consistency coefficient for each pathway. When scoring candidate drugs, only pathways that meet the consistency constraints and are connected to the target are included in the main aggregation. Pathway connections that do not meet the consistency constraints are attenuated according to the degree of deviation. In this embodiment, the same anomaly will not be uniformly amplified when participating in multiple pathways, and the same drug target will not deviate from the anomaly direction and clonal coverage conditions when connecting to multiple pathways.

[0047] In an embodiment targeting samples with strong clonal heterogeneity, if a hematologic tumor sample contains multiple tumor clone nodes, the system calculates the clonal contribution edge weights from each tumor clone node to the molecular anomalous node, and uses hierarchical message passing to allow the anomalous representations corresponding to each clone to enter the relevant signaling pathway nodes. When generating pathway gating values, the pathway gating module incorporates clonal coverage ratio features and anomalous direction consistency features. If the dominant clone and the low-abundance clone are consistent in the same pathway direction, the pathway gating value retains both types of clonal inputs. If the low-abundance clone and the dominant clone conflict in the same pathway direction and the counterfactual node contribution weight is low, the low-abundance clone input is weakened in the candidate drug scoring. If the low-abundance clone has a low proportion but a high counterfactual node contribution weight and connects to drug targets that meet the consistency constraints, the low-abundance clone can be written into the output result as a supporting clone. In this embodiment, clonal abundance and node contribution participate in the screening judgment simultaneously, so that a small number of key clones and ordinary low-correlation clones can be distinguished.

[0048] In one embodiment for comparing multiple candidate drugs, the drug scoring module independently calculates the candidate drug score, target consistency coefficient set, and counterfactual node contribution weight set for each candidate drug node. Then, it performs a ranking comparison within the same sample. During the comparison, the original score of the candidate drug node is not directly used. Instead, the candidate drug score and the integrity status of the supporting links are used together as the ranking criteria. When the candidate drug scores of two candidate drugs are close, the candidate drug with a complete clonal propagation link, multiple supporting molecular anomalies, and targets that meet the consistency constraints is given priority. When a candidate drug has a high score but mainly comes from isolated expression deviations or drug targets that do not meet the consistency constraints, the candidate drug is not included in the priority ranking interval. In this embodiment, the comparison between candidate drugs is jointly limited by the score value and the structured supporting links.

[0049] In this embodiment, the overall processing revolves around the lack of structured computational relationships between molecular abnormalities, tumor clones, signaling pathways, drug targets, and candidate drugs in the screening of targeted drugs for hematological malignancies. Data encoding ensures that each molecule detected is placed in its corresponding position within the heterogeneity graph. Clonal contribution edge weights allow clonal abundance and anomalous functional influence to be incorporated into message passing. Pathway propagation weights allow anomalous type, expression direction, and pathway affiliation to be incorporated into pathway aggregation. Pathway gating values ​​ensure that the consistency of anomalous direction, clonal coverage ratio, and expression response direction within the same pathway jointly limit the pathway output. Target consistency coefficients ensure that candidate drug scores are consistent with the anomalous pathway and target action direction. Counterfactual node contribution weights and consistency checks ensure that the output results correspond to traceable supporting clones, molecular abnormalities, and signaling pathways. The advantage of this embodiment is that each processing step constrains each other around the same propagation link, reducing the interference of low-relevance anomalous abnormalities on candidate drug ranking and preserving the possibility that multiple weak anomalous abnormalities can form screening criteria at the same pathway level.

Claims

1. A method for screening targeted drugs for hematologic malignancies based on neural networks, characterized in that, include: Molecular detection data of blood tumor samples are obtained, including gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data. Based on the molecular detection data, a heterogeneous graph is constructed, which includes tumor clone nodes, molecular abnormality nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes, and edges carrying clonal contribution information and pathway association information are established between different types of nodes. The heterogeneous graph is input into the graph neural network and hierarchical message passing is performed so that the cloning weight of the tumor clone node is passed to the signaling pathway node through the molecular abnormality node, and then passed to the drug target node and the candidate drug node by the signaling pathway node. At the signal path node, a path gating value is generated based on abnormal consistency and clone coverage status; Based on the pathway gating value and the consistency constraints between the drug target node and the signaling pathway node, the candidate drug node is scored and the targeted drug screening results are output.

2. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 1, characterized in that, When constructing the heterogeneity graph, the frequency of variant alleles, the level of mutational function, and the co-occurrence relationship of mutations in the gene mutation data are jointly encoded to form the mutational features of the molecular abnormal nodes; The fusion breakpoint location, fusion fragment retention status, and fusion gene expression status in the fusion gene data are jointly encoded to form the fusion features of the molecular abnormal node; The copy number change data and the transcriptional expression data are encoded to form the expression perturbation features of the molecular abnormal nodes; The mutation features, fusion features, and expression perturbation features are mapped to the tumor clone nodes according to their corresponding clonal origins, so that the same molecular abnormality node in the heterogeneous graph has both molecular abnormality attributes and clonal affiliation attributes.

3. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 1, characterized in that, When performing the hierarchical message transmission, the initial clonal weights are first assigned to the adjacent molecular abnormal nodes based on the clonal abundance corresponding to the tumor clonal node, and then the pathway propagation weights are generated based on the abnormality type matching relationship, expression direction relationship and pathway subordination relationship between the molecular abnormal node and the signaling pathway node. The initial cloning weights and the path propagation weights are merged into hierarchical propagation weights, and the path representations of the signal path nodes are updated using the hierarchical propagation weights. The target representation of the drug target node is updated jointly by the pathway representation of the adjacent signaling pathway nodes and the target attributes of the drug target node itself.

4. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 2, characterized in that, When generating the pathway gating value, the abnormal direction consistency feature, clone coverage ratio feature and expression response direction feature are extracted from multiple molecular abnormal nodes adjacent to the same signal pathway node, and the abnormal direction consistency feature, clone coverage ratio feature and expression response direction feature are input into the gating network. The gating network outputs the path gating value corresponding to the signal path node; The amount of information transmitted from the signal pathway node to the drug target node is weighted according to the pathway gating value, and the weighted information is used as the input feature for the drug target node to participate in the candidate drug scoring.

5. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 4, characterized in that, When forming the tumor clone node, multiple molecular abnormalities are cloned and grouped according to the frequency of the variant alleles, the copy number change data and the expression status of the fusion gene, and a corresponding tumor clone node is established for each clone group. When establishing the edges between the tumor clone nodes and the molecular abnormality nodes, clonal abundance, abnormality detection confidence, and abnormal function impact level are encoded as clonal contribution edge features. When establishing the edge between the molecular anomaly node and the signal pathway node, the anomaly type, anomaly direction, and pathway affiliation confidence are encoded as pathway association edge features. The cloning contribution edge features and the path association edge features are used together as the edge inputs of the graph neural network.

6. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 3, characterized in that, When updating the pathway characterization of the signaling pathway node, the first aggregate characterization from the molecular abnormal node corresponding to the major clone, the second aggregate characterization from the molecular abnormal node corresponding to the low abundance clone, and the third aggregate characterization from the molecular abnormal node corresponding to the fusion gene abnormality are calculated respectively. A pathway aggregation representation is generated based on the similarity and directional consistency among the first aggregation representation, the second aggregation representation, and the third aggregation representation; A residual connection is set between the path aggregation representation and the original path attributes of the signal path node, so that the signal path node retains the prior path attributes and receives the clone-weighted anomalous input during the graph neural network iteration process.

7. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 5, characterized in that, When scoring candidate drugs based on the pathway gating value, target consistency constraints are established for the drug target nodes. The target consistency constraints include consistency constraints between the signaling pathway to which the target belongs and the direction of abnormal enhancement, consistency constraints between the signaling pathway to which the target belongs and the clonal coverage ratio, and consistency constraints between the downstream expression perturbation direction of the target and the inhibitory direction of the candidate drug. The drug target node representation that satisfies the target consistency constraint is passed to the adjacent candidate drug node. The drug target node representation that does not meet the target consistency constraint is attenuated according to the degree of constraint deviation before being passed to the adjacent candidate drug node.

8. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 7, characterized in that, After scoring the candidate drug nodes, a counterfactual node blocking process is performed for each candidate drug node. The counterfactual node blocking process includes sequentially blocking the tumor clone node, the molecular abnormality node, and the signaling pathway node related to the candidate drug node score, and re-executing the message passing from the signaling pathway node to the candidate drug node after each blocking. Node contribution weights are generated based on the difference in candidate drug scores before and after shielding. The node contribution weights and the scores of the candidate drug nodes are used together to determine the supporting clones, supporting anomalies, and supporting signaling pathways in the targeted drug screening results.

9. The method for screening targeted drugs for hematological malignancies based on neural networks according to claim 8, characterized in that, Before outputting the targeted drug screening results, a consistency check is performed on the node contribution weights. The consistency check includes determining whether the tumor clone node with the highest contribution is connected to the signaling pathway node with the highest contribution through at least one of the molecular abnormality nodes, determining whether the signaling pathway node with the highest contribution is connected to the candidate drug node through a drug target node that satisfies the target consistency constraint, and determining whether the score of the candidate drug node comes from multiple supporting molecular abnormalities within the same clonal propagation chain. Only the ranking results of candidate drugs that meet the consistency check are written into the targeted drug screening results.

10. A neural network-based targeted drug screening system for hematologic malignancies, characterized in that, include: The data acquisition module is used to acquire molecular detection data of blood tumor samples, including gene mutation data, fusion gene data, copy number alteration data, and transcriptional expression data. The heterogeneity graph construction module is used to construct a heterogeneity graph based on the molecular detection data. The heterogeneity graph includes tumor clone nodes, molecular abnormality nodes, signaling pathway nodes, drug target nodes, and candidate drug nodes, and establishes edges carrying clonal contribution information and pathway association information between different types of nodes. The graph neural network processing module is used to perform hierarchical message passing on the heterogeneous graph, so that the cloning weight of the tumor clone node is passed to the drug target node and the candidate drug node through the molecular abnormality node and the signaling pathway node. The path gating processing module is used to generate path gating values ​​based on abnormal consistency and clone coverage status; The drug scoring module is used to score the candidate drug nodes and output the targeted drug screening results based on the pathway gating value and the consistency constraints between the drug target node and the signaling pathway node.

Images