A drug target-disease association strength prediction method and system

Abstract

The present application relates to the technical field of data processing, in particular to a drug target-disease correlation strength prediction method and system. The present application converts heterogeneous records such as significance, effect size, expression fold, network distance, phenotype state and research and development stage into same-interval numerical values with source identification through the cooperation of source-specific evidence characterization, historical calibration parameters and single-source scoring with consistent direction; the subsequent steps read the same group of scores, grades and weights along the identification, form repeatable calculation inputs and improve the comparability between different sources, thereby reducing the judgment fluctuations caused by manual conversion caliber and processing sequence differences, and alleviating the technical problems of scattered multi-source evidence, different dimensions and difficult-to-reproduce judgment results.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for predicting the strength of drug target-disease association. Background Technology

[0002] Target selection in drug development requires determining whether inhibiting or regulating a specific target can affect the target disease. This determination typically utilizes evidence from various sources, including genetic associations, omics, pathways and networks, animal models, literature and knowledge graphs, and known drugs. Each source reflects different aspects such as genetic causality, molecular expression, mechanistic networks, phenotypic reproducibility, statistical association, or existing pharmacological validation.

[0003] This evidence is distributed across different databases, literature, or internal data records. The raw data includes both numerical indicators such as significance, effect size, fold increase, and network distance, and categorical indicators such as phenotypic recurrence and drug development stage. Differences in the units of evidence, range of values, and supporting directions require researchers to interpret and convert each item individually. The same set of evidence may lead to different judgments depending on the interpretation.

[0004] For the same target-disease pair, different sources may exhibit opposite orientations. For example, genetic evidence may support a strong association between the two, while animal models or literature evidence may show a weaker association. Directly summing, averaging, or placing all sources in the same weighting system will allow evidence with lower causality to participate in determining the conflict resolution orientation, thereby weakening the priority role of evidence with higher causality in that orientation. Increasing the weight of a single source alone is still a cross-level mixing, making it difficult to clearly distinguish between hierarchical priority relationships and reliability differences within the same level.

[0005] To address the above problems, this invention proposes a method and system for predicting the strength of drug target-disease association. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, the present invention aims to provide a method and system for predicting the strength of drug target-disease association.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for predicting the strength of drug target-disease association includes: Receive target-disease pairs, collect associated evidence from multiple sources and form evidence records grouped by source; generate feature vectors for each valid source according to the characteristic rules corresponding to the evidence source; input each feature vector into the corresponding single-source association scoring function to obtain single-source association scores with consistent value directions; Based on the strength of causality of the evidence sources, each source is classified into a causality level with high and low priority, and a normalized reliability weight is determined within each level; conflicts are detected according to the support threshold and the denial threshold, and the resolution orientation value is determined according to the highest causality level, so that lower-level conflict sources converge towards the resolution orientation value. The model input is formed according to the expected sources in a fixed order and fed into the association strength prediction model to obtain the association strength; the prediction confidence is calculated based on the coverage and conflict degree of the valid evidence sources; the association strength, prediction confidence and evidence source tracing details are collected to form a prediction result record.

[0008] Preferably, feature vectors for each valid source are generated according to the featureization rules corresponding to the source of evidence, including: For genetic association sources, extract association significance, effect size, physical distance from the site to the target or mapping confidence level, and the number of evidences of loss of function; for omics sources, extract differential expression fold change, differential significance, and expression specificity in disease-related cell types; for pathway and network sources, extract pathway overlap and network topological proximity; for animal model sources, extract phenotypic recurrence status, phenotypic similarity, and the number of evidence entries; for literature and knowledge graph sources, extract co-occurrence frequency, semantic association strength of co-occurrence contexts, and the shortest association path length; for known drug sources, extract the ontological distance between the drug indication and the target disease and the highest development stage. Write the features of each valid source into the feature vector in the order of the predetermined fields to form a feature vector set; generate valid labels in the fixed order of the expected source set, with the first value for valid sources and the second value for missing sources.

[0009] Preferably, each feature vector is input into the corresponding single-source association scoring function to obtain single-source association scores with consistent value directions, including: Each source is configured with a corresponding scoring function and historical calibration parameters. All functions are associated with the direction of support enhancement by increasing their numerical values. Each component in the feature vector is truncated and normalized using the lower and upper boundaries of the historical calibration data. For features with smaller values ​​and stronger correlation support, the corresponding normalized value is subtracted from one before participating in the weighting. When the lower and upper boundaries of the historical calibration data are equal, the normalized feature is directly set to the preset median value. The normalized components are weighted and summed according to their non-negative characteristic coefficients to obtain the single-source correlation score, which ranges from 0 to 1.

[0010] Preferably, based on the strength of causality of the evidence sources, each source is classified into a causality level with high and low priority, and a normalized reliability weight is determined within each level, including: The causality levels are ranked as follows: highest causality level, intermediate causality level, and lowest causality level, with the priority order being highest level > intermediate level > lowest level. Genetic association sources are configured as the highest causality level; omics, pathway and network, and animal model sources are configured as intermediate causality levels; literature and knowledge graph sources are configured as the lowest causality level; known drug sources are pre-configured as intermediate or lowest causality levels based on the nature of their evidence. For sources within the same level, the normalized reliability weight of each source within the level is determined by the ratio of the reliability parameter of each source to the sum of the reliability parameters within the same level.

[0011] Preferably, conflicts are detected based on support and rejection thresholds, and a resolution orientation value is determined according to the highest causality level, causing lower-level conflict sources to converge towards the resolution orientation value, including: Sources with a single-source association score greater than the support threshold are written into the support source set, and sources with a single-source association score less than the negation threshold are written into the negation source set. The support threshold is greater than the negation threshold. When both the support source set and the negation source set are not empty, a conflict source set is formed. From the causality levels involved in the conflict source set, the highest causality level is selected in priority to form the conflict source set within that level; only the single-source association score and reliability weight of the conflict source set within that level are used to calculate the weighted average as the resolution orientation value; For each source in the negative source set, the average of its single-source correlation score and the resolution orientation value is taken as the single-source correlation score after resolution; for sources that do not belong to the negative source set, the original single-source correlation score is retained; Calculate the difference between the average score of the supporting source set and the average score of the negative source set. Use the ratio of the number of conflicting sources to the number of valid sources as the conflict percentage, and use the product of the two as the conflict degree.

[0012] Preferably, the model input is formed in a fixed order according to the expected sources and fed into the association strength prediction model to obtain the association strength, including: According to the fixed order of the expected source set, for each source, write the valid label, the product of the valid label and the single-source association score after resolution, and the product of the valid label and the reliability weight in sequence to form the fixed-dimensional model input; The fixed-dimensional model is input into the association strength prediction model, and the output is the association strength with a value range of 0 to 1; the association strength is converted into the corresponding strength level according to the preset grading threshold.

[0013] Preferably, the predicted confidence level is calculated based on the coverage and conflict level of valid evidence sources, including: The ratio of the number of sources of valid evidence actually obtained to the total number of expected sources of evidence is used as the evidence coverage, and the evidence coverage is used as the basis for credibility. The predicted confidence level is obtained by multiplying the base confidence level by one and subtracting the product of the conflict downsizing factor and the conflict degree; when no conflict is detected, the conflict degree is zero, and the predicted confidence level remains the evidence coverage. The corresponding credibility level is formed based on the preset first credibility threshold and second credibility threshold.

[0014] This invention also provides a drug target-disease association strength prediction system, comprising: The evidence collection unit is used to receive target-disease pairs, collect relevant evidence from multiple sources, and form evidence records grouped by source. The feature scoring unit is used to generate feature vectors for each valid source according to the featureization rules corresponding to the evidence source. Each feature vector is input into the corresponding single-source association scoring function to obtain a single-source association score with consistent value direction. The grading configuration unit is used to classify each source into a causal grade with high and low priority based on the causal strength of the evidence sources, and to determine the normalized reliability weight within the grade. The conflict resolution unit is used to detect conflicts based on support thresholds and rejection thresholds, determine a resolution orientation value according to the highest causality level, and cause lower-level conflict sources to converge toward the resolution orientation value. The intensity prediction unit is used to form model inputs according to a fixed order of expected sources, and then feed them into the correlation intensity prediction model to obtain the correlation intensity. The credibility assessment unit is used to calculate the predicted credibility based on the coverage and degree of conflict of valid evidence sources; The result output unit is used to collect the correlation strength, the prediction confidence and evidence traceability details to form a prediction result record.

[0015] Preferably, the conflict resolution unit writes sources with single-source correlation scores greater than the support threshold into the support source set, and sources with single-source correlation scores less than the negation threshold into the negation source set, where the support threshold is greater than the negation threshold; a conflict source set is formed when both the support source set and the negation source set are not empty; the highest causality level is selected from the causality levels involved in the conflict source set in priority order to form a conflict source set within that level; and a weighted average is calculated using only the single-source correlation scores and reliability weights in the conflict source set within that level as the resolution orientation value. For each source in the negative source set, the average of its single-source association score and the resolution orientation value is taken as the single-source association score after resolution. The difference between the average score of the supporting source set and the average score of the negative source set is calculated. The proportion of conflicting sources to valid sources is taken as the conflict ratio, and the product of the two is taken as the conflict degree.

[0016] Preferably, each source in the feature scoring unit is configured with a corresponding scoring function and historical calibration parameters. All functions output in the direction of enhanced association support as the numerical value increases. Each component in the feature vector is truncated and normalized using the lower and upper boundaries of the historical calibration data. For features with smaller numerical values ​​and stronger association support, the corresponding normalized value is subtracted by one before participating in the weighting. The normalized components are weighted and summed according to the non-negative feature coefficients to obtain the single-source association score. The intensity prediction unit writes a valid label, the product of the valid label and the single-source association score after resolution, and the product of the valid label and the reliability weight for each source in a fixed order according to the expected source set. This forms a fixed-dimensional model input, which is then fed into the association intensity prediction model to output the association intensity with a value range of 0 to 1. The credibility assessment unit uses the ratio of the number of sources of valid evidence actually obtained to the total number of expected sources of evidence as the evidence coverage. The evidence coverage is used as the base value of credibility. The predicted credibility is obtained by multiplying the base value of credibility by one and subtracting the product of the conflict downsizing coefficient and the degree of conflict.

[0017] Beneficial effects 1. This invention combines source-specific evidence characterization, historical calibration parameters, and consistent single-source scoring to convert heterogeneous records such as significance, effect size, expression fold, network distance, phenotypic status, and R&D stage into values ​​within the same interval with source identifiers. Subsequent steps read the same set of scores, grades, and weights along this identifier, forming repeatable calculation inputs and improving the comparability between different sources. This reduces judgment fluctuations caused by differences in manual conversion methods and processing order, and alleviates the technical problems of scattered evidence from multiple sources, inconsistent dimensions, and difficulty in reproducing judgment results.

[0018] 2. This invention employs a combination mechanism of hard priority between levels and normalized weighting within levels. First, the highest causality level among the conflict sources determines the resolution orientation value, and then the conflict scores of lower levels converge toward this orientation value. This orientation value remains controlled by high causal evidence, and multiple sources within the same level form a stable compromise based on reliability. This effectively reduces the pull of opposite orientations across levels and keeps the source details verifiable, thereby improving the problem of uniform weighting or simple averaging diluting high causal conclusions.

[0019] 3. This invention generates association strength based on the resolved single-source association score and reliability weight, and simultaneously combines the proportion of valid evidence sources with the degree of conflict between support and denial scores to form an independent predicted credibility. The two types of results are output together with the source score, conflict participation status, highest conflict causality level and resolution orientation value, so that the strength and the sufficiency of the evidence basis can be observed separately, and the credibility of low coverage or high conflict situations is reduced accordingly, thereby alleviating the technical problem that it is difficult to identify insufficient evidence or evidence discrepancies by only giving the association strength. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the invention and are not intended to limit the scope of protection of the invention.

[0022] Example 1 Please see Figure 1 As shown in the figure, this embodiment provides a method for predicting the strength of drug target-disease association, which includes the following steps: S1. Receive target-disease pairs, collect relevant evidence from multiple sources, and form evidence records grouped by source; The system receives standard identifiers for target targets, standard identifiers for target diseases, and a set of expected evidence sources. Target targets can be represented by gene symbols or protein identifiers, and target diseases can be represented by disease ontology identifiers or medical subject term identifiers. The computing device performs queries based on the target pairs in genetic associations, omics, pathways and networks, animal models, literature and knowledge graphs, and known drug data sources, converting the obtained structured or semi-structured records into evidence records. Each evidence record contains at least a target pair identifier, a source identifier, the original value, the meaning of the value, and the validity status. This step forms a set of evidence records grouped by source and a set of valid sources. The evidence record set is used in subsequent feature extraction steps, while the set of valid sources is used in the credibility assessment step to calculate coverage. By fixing the target pair identifiers and source identifiers during the acquisition phase, feature extraction and coverage status can point to the same set of evidence.

[0023] S2. Generate feature vectors for each valid source according to the featureization rules corresponding to the source of evidence; For genetically related sources, extract association significance, effect size, physical distance from the locus to the target or mapping confidence, and the number of evidence entries for loss of function from the corresponding evidence records; for omics sources, extract differential expression fold, differential significance, and expression specificity in disease-related cell types; for pathway and network sources, extract pathway overlap and network topological proximity; for animal model sources, extract phenotypic recurrence status, phenotypic similarity, and the number of evidence entries; for literature and knowledge graph sources, extract co-occurrence frequency, semantic association strength of co-occurrence context, and the shortest association path length; for known drug sources, extract the ontological distance between the drug indication and the target disease and the highest development stage.

[0024] The features of each valid source are written into a feature vector in a predetermined field order to form a feature vector set. Valid labels are also generated in a fixed order according to the expected source set, with 1 for valid sources and 0 for missing sources. The feature vector set is used in subsequent single-source association calculation steps, the source identifier is used in subsequent level configuration steps to add levels, and the valid labels are used as input for subsequent intensity prediction steps to form a fixed-dimensional model.

[0025] S3. Input each feature vector into the corresponding single-source association scoring function to obtain single-source association scores with consistent value directions; Each source is configured with a corresponding scoring function and historical calibration parameters, and the calculation formula is as follows:

[0026]

[0027] in, Indicate source Single-source association segmentation of target-disease association, Indicate source The single-source association scoring function, This represents the feature vector of the source. The source index is indicated. The scoring function first truncates and normalizes each component in the feature vector according to the historical calibration boundary, and then performs a linear weighted sum on the normalized components. All functions output directions of association support enhancement by increasing the numerical value.

[0028] Furthermore, for the eigenvector of the th... Each component is truncated and normalized using the lower and upper boundaries of historical calibration data:

[0029]

[0030] in, Indicates the feature index within the source. Represents normalized features. and These represent the historical calibration boundaries, This represents a truncation operator that sets values ​​less than 0 to 0 and values ​​greater than 1 to 1. This indicates that the coefficients are non-negative and that the sum of all coefficients from the same source is 1. For features with smaller numerical values ​​indicating stronger association, the corresponding normalized value is subtracted from 1 before being included in the weighting. Direct order No division is performed. This step outputs a single-source associated subset with source identifiers, which is used by the subsequent level configuration step to bind levels and weights, and by the subsequent conflict detection step to perform conflict detection.

[0031] The lower and upper boundaries of the historical calibration data are determined based on the distribution of this feature in existing historical annotation data. Preferably, for the significance of association in the genetic association source, since a smaller value indicates a stronger association, its lower boundary is taken as the minimum value of 0 in the historical data, and its upper boundary is taken as the maximum value of 1 corresponding to the significance threshold in the historical data; For effect sizes in genetic association sources, if the common range of values ​​for the effect size in historical data is 0.5 to 5, then the lower boundary is 0.5 and the upper boundary is 5. Original values ​​outside this range are restricted to the boundary by the truncation operator. For differential expression folds from omics sources, the lower bound is: 3, with the upper boundary being +3.

[0032] S4. Based on the strength of causality of the evidence sources, classify each source into a causality level with high and low priority, and determine the normalized reliability weight within each level. The causality levels are in descending order of highest causality. Intermediate causal hierarchy and lowest causality level The priority order is Genetic association source configuration is as follows: The sources of omics, pathways and networks, and animal models are configured as follows: The literature and knowledge graph sources are configured as follows: Known drug sources are pre-configured based on the nature of their evidence. or For sources within the same level, the original reliability parameters are used. Calculate the normalized reliability weights:

[0033]

[0034] in, Indicate source Positive reliability parameters within the relevant level. This indicates the source index within the same level. Indicates the current level of causality. This represents the normalized reliability weights within the tier. Reliability parameters The data is derived from either a configuration value or historical calibration data. The configuration value is a pre-defined positive number based on the source type; historical calibration data consists of positive numbers determined based on the accuracy or consistency metrics of the source in historical labeled data. This weight is used only to compare sources within the same causality level and is not used to change the priority between levels. This step generates a set of score records containing source identifiers, single-source association scores, causality levels, and reliability weights. These records are used by subsequent conflict resolution steps to determine conflict levels and resolution orientations, and also as input for subsequent intensity prediction steps to construct the model.

[0035] S5. Detect conflicts based on support and rejection thresholds, determine the resolution orientation value according to the highest causality level, and bring lower-level conflict sources toward it. Read support threshold and negation threshold ,in Preferably, the support threshold can be set to 0.7, and the rejection threshold can be set to 0.3. This will satisfy... The source is written into the supporting source set. , will satisfy The source is written into the negative source set. A conflict source set is formed when both sets are non-empty. .

[0036] The first paragraph begins with a set of sources of conflict. Among the levels involved, Select the highest level of causality. This forms the set of conflict sources within that level. and the set of conflict sources below this level. The second paragraph only uses The single-source correlation score and reliability weight calculation resolves the orientation value:

[0037] in, This represents the resolution orientation value determined by the highest conflict causality level. This represents the set of highest-level conflict sources. The resolution orientation value... Pressed Reliability weights of each internal source Weighted calculation, for Each source in the calculation ; for those that do not belong Sources maintained That is, reliability weights are used to form resolution orientation values. The convergence step involves directing lower-level sources of negation towards the already formed... The two terms, "gather" and "collect," target different objects.

[0038] This step also calculates the degree of conflict. Let the number of valid sources be... ,but:

[0039]

[0040]

[0041] in, This represents the difference between the average score from supporting sources and the average score from opposing sources. This indicates the proportion of conflict sources to valid sources. Indicates the degree of conflict. This represents the total number of valid sources. The above calculations are only applicable to... and Execution is not performed when the system is empty; in other cases, each condition is maintained. And order =0, to avoid empty set mean calculation. This step outputs the resolved single-source association subset, conflict degree, highest conflict causality level, and resolution orientation value. The resolved single-source association subset is used for subsequent intensity prediction steps, the conflict degree is used for subsequent confidence assessment steps to adjust the confidence, and the highest conflict causality level and resolution orientation value are used for subsequent result output steps to form source tracing details.

[0042] S6. Form model inputs according to the expected source in a fixed order, and send them into the association strength prediction model to obtain the association strength; The computing device writes data sequentially to each source in a fixed order according to the expected set of sources. , and To form a fixed-dimensional model input ; source missing =0 and write three 0s in the corresponding position, valid source because =1, while retaining the score and weight after resolution. Wherein, Indicate source The valid marker is 1 for a valid source and 0 for a missing source. Indicate source The resolved single-source correlation; Indicate source The reliability weights are normalized within the rank. Training samples are encoded using the same source order and valid labels. The association strength prediction model outputs the following formula:

[0043]

[0044]

[0045] in, This represents the input of a fixed-dimensional model. Indicates a valid source marker. Represents the expected source set, The parameter is Predictive models, This indicates the strength of the association. The prediction model can employ logistic regression, gradient boosting decision trees, or multilayer perceptrons; the training set contains positive samples with confirmed strong associations and negative samples with known weak or no associations. This step also categorizes samples according to a preset grading threshold. The data is converted into strong association, moderate association, weak association, or no obvious association level, and then passed to the subsequent results output step. The classification threshold is preset based on the association strength range (0 to 1).

[0046] S7. Calculate the prediction confidence based on the coverage and conflict level of valid evidence sources; Let the total number of expected sources of evidence be... The actual number of sources from which valid evidence was obtained was Evidence coverage is The basic credibility value is taken as The prediction confidence level is then calculated using the following formula:

[0047]

[0048]

[0049]

[0050] in, Indicates the confidence level of the prediction. This represents the base value obtained from coverage. This indicates the conflict reduction coefficient. This indicates the degree of conflict in the output of step S5. and These represent the number of valid sources and the expected number of sources, respectively. Since 0 ≤ ≤1 and ,therefore Thus ensuring Always satisfy 0≤ ≤ ≤1. No conflict detected. =0, credibility remains at coverage; when the proportion of conflict increases, credibility is reduced proportionally.

[0051] Preferably, the conflict reduction coefficient The value is 0.5. No conflict detected. =0, the credibility remains at the coverage level; as the degree of conflict increases, the downward adjustment amount increases proportionally. This step forms high credibility, intermediate credibility, or low credibility levels based on preset first and second credibility thresholds, and passes the credibility and its level to subsequent result output steps, so that insufficient evidence coverage and evidence conflict are applied to credibility through the same repeatable formula. The first and second credibility thresholds are preset based on the range of predicted credibility values. Preferably, the first credibility threshold is 0.7 and the second credibility threshold is 0.4.

[0052] S8. Collect details of correlation strength, prediction confidence, and evidence tracing to form a prediction result record; The method aggregates the association strength and strength level, prediction confidence and confidence level from previous steps, and reads the source identifier, single-source association score, reliability weight, conflict participation status, highest conflict causality level, and resolution orientation value retained in steps S3 to S5 to form a prediction result record. The prediction result record includes at least the target pair identifier, association strength, association strength level, prediction confidence, confidence level, and evidence tracing details. The computing device outputs this record to a data file, database record, or API call as the final output of this method. This output structure allows the association strength calculation result to be traced back to the conflict resolution process along the source identifier, and allows the confidence level to be traced back to the corresponding input along the valid source set and conflict degree.

[0053] Example 2 Please see Figure 2 As shown, this embodiment provides a drug target-disease association strength prediction system. The system 100 includes an evidence collection unit 110 that transmits data sequentially, a feature scoring unit 120, a level configuration unit 130, a conflict resolution unit 140, an intensity prediction unit 150, a credibility assessment unit 160, and a result output unit 170.

[0054] The evidence collection unit 110 accesses the target marker, target disease marker, and expected evidence source set, queries each evidence source, and forms a set of evidence records grouped by source and a set of valid sources. The feature scoring unit 120 reads the evidence records, extracts feature vectors by source, applies the corresponding scoring function, generates a single-source association subset with source markers, and generates valid markers according to the fixed order of expected sources. The grading configuration unit 130 reads the single-source association subset, adds causality grades and reliability weights within each grade to each source, and sends the scoring record containing scores, grades, and weights to the conflict resolution unit 140.

[0055] The conflict resolution unit 140 forms a set of supporting sources and a set of opposing sources based on the scoring records. When both sets are not empty, it determines the highest conflict causality level according to the priority of the levels, calculates the resolution orientation value of that level, performs convergence on the conflict scores of lower levels, and generates a set of resolved single-source association scores and conflict degree. When at least one set is empty, it retains each single-source association score and sets the conflict degree to 0. The intensity prediction unit 150 reads the resolved single-source association score set, the reliability weight generated by the level configuration unit 130, and the valid labels generated by the feature scoring unit 120, forms the model input according to a fixed source order, and outputs the association strength and intensity level. The credibility assessment unit 160 reads the valid source set from the evidence collection unit 110 and the conflict degree from the conflict resolution unit 140, and calculates the coverage, credibility, and credibility level.

[0056] The result output unit 170 simultaneously receives the correlation strength result from the intensity prediction unit 150, the credibility result from the credibility assessment unit 160, and the tracing field retained by the conflict resolution unit 140, generates a prediction result record, and outputs it. Each unit can be implemented by one or more processors executing program instructions in memory, or it can be deployed in interconnected computing nodes according to the above data flow; regardless of the deployment form, the data objects and processing order passed between units remain consistent.

[0057] System 100 also stores a source configuration table, historical calibration parameters, grade priority, threshold parameters, and prediction model parameters. Evidence acquisition unit 110 determines the expected source according to the source configuration table, feature scoring unit 120 selects historical calibration parameters based on the same source identifier, and grade configuration unit 130 reads the grade and reliability parameters based on that identifier. When parameters are updated, the source identifier remains unchanged, ensuring that conflict resolution unit 140, intensity prediction unit 150, and result output unit 170 can still read the corresponding records according to the predetermined field positions. After the prediction model is retrained, intensity prediction unit 150 loads the new model parameters, while the data interfaces of other units maintain their original order and type.

[0058] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the strength of drug target-disease association, characterized in that, include: Receive target-disease pairs, collect associated evidence from multiple sources and form evidence records grouped by source; generate feature vectors for each valid source according to the characteristic rules corresponding to the evidence source; input each feature vector into the corresponding single-source association scoring function to obtain single-source association scores with consistent value directions; Based on the strength of causality of the evidence sources, each source is divided into a causality level with high and low priority, and a normalized reliability weight is determined within each level. Conflicts are detected based on support and rejection thresholds, and the resolution orientation value is determined according to the highest causality level, so that lower-level conflict sources converge towards the resolution orientation value; The model input is formed according to the expected sources in a fixed order and fed into the association strength prediction model to obtain the association strength; the prediction confidence is calculated based on the coverage and conflict degree of the valid evidence sources; the association strength, prediction confidence and evidence source tracing details are collected to form a prediction result record; The prediction confidence is calculated based on the coverage and conflict level of valid evidence sources, including: The ratio of the number of sources of valid evidence actually obtained to the total number of expected sources of evidence is used as the evidence coverage, and the evidence coverage is used as the basis for credibility. The predicted confidence level is obtained by multiplying the base confidence level by one and subtracting the product of the conflict downsizing factor and the conflict degree; when no conflict is detected, the conflict degree is zero, and the predicted confidence level remains the evidence coverage. The corresponding credibility level is formed based on the preset first credibility threshold and second credibility threshold.

2. The method for predicting the strength of drug target-disease association according to claim 1, characterized in that, Feature vectors for each valid source are generated according to the characteristic rules corresponding to the source of evidence, including: For genetic association sources, extract association significance, effect size, physical distance from the site to the target or mapping confidence level, and the number of evidences of loss of function; for omics sources, extract differential expression fold change, differential significance, and expression specificity in disease-related cell types; for pathway and network sources, extract pathway overlap and network topological proximity; for animal model sources, extract phenotypic recurrence status, phenotypic similarity, and the number of evidence entries; for literature and knowledge graph sources, extract co-occurrence frequency, semantic association strength of co-occurrence contexts, and the shortest association path length; for known drug sources, extract the ontological distance between the drug indication and the target disease and the highest development stage. Write the features of each valid source into the feature vector in the order of the predetermined fields to form a feature vector set; generate valid labels in the fixed order of the expected source set, with the first value for valid sources and the second value for missing sources.

3. The method for predicting the strength of drug target-disease association according to claim 1, characterized in that, Each feature vector is input into the corresponding single-source association scoring function to obtain single-source association scores with consistent value directions, including: Each source is configured with a corresponding scoring function and historical calibration parameters. All functions are associated with the direction of support enhancement by increasing their numerical values. Each component in the feature vector is truncated and normalized using the lower and upper boundaries of the historical calibration data. For features with smaller values ​​and stronger correlation support, the corresponding normalized value is subtracted from one before participating in the weighting. When the lower and upper boundaries of the historical calibration data are equal, the normalized feature is directly set to the preset median value. The normalized components are weighted and summed according to their non-negative characteristic coefficients to obtain the single-source correlation score, which ranges from 0 to 1.

4. The method for predicting the strength of drug target-disease association according to claim 1, characterized in that, Based on the strength of causality of the evidence sources, each source is classified into a causality level with high and low priority, and a normalized reliability weight is determined within each level, including: The causality levels are ranked as follows: highest causality level, intermediate causality level, and lowest causality level, with the priority order being highest level > intermediate level > lowest level. Genetic association sources are configured as the highest causality level; omics, pathway and network, and animal model sources are configured as intermediate causality levels; literature and knowledge graph sources are configured as the lowest causality level; known drug sources are pre-configured as intermediate or lowest causality levels based on the nature of their evidence. For sources within the same level, the normalized reliability weight of each source within the level is determined by the ratio of the reliability parameter of each source to the sum of the reliability parameters within the same level.

5. The method for predicting the strength of drug target-disease association according to claim 1, characterized in that, Conflicts are detected based on support and rejection thresholds. A resolution orientation value is determined according to the highest causality level, converging lower-level conflict sources towards the resolution orientation value, including: Sources with a single-source association score greater than the support threshold are written into the support source set, and sources with a single-source association score less than the negation threshold are written into the negation source set. The support threshold is greater than the negation threshold. When both the support source set and the negation source set are not empty, a conflict source set is formed. From the causality levels involved in the conflict source set, the highest causality level is selected in priority to form the conflict source set within that level; only the single-source association score and reliability weight of the conflict source set within that level are used to calculate the weighted average as the resolution orientation value; For each source in the negative source set, the average of its single-source correlation score and the resolution orientation value is taken as the single-source correlation score after resolution; for sources that do not belong to the negative source set, the original single-source correlation score is retained; Calculate the difference between the average score of the supporting source set and the average score of the negative source set. Use the ratio of the number of conflicting sources to the number of valid sources as the conflict percentage, and use the product of the two as the conflict degree.

6. The method for predicting the strength of drug target-disease association according to claim 1, characterized in that, The model input is formed according to the expected source in a fixed order and then fed into the association strength prediction model to obtain the association strength, including: According to the fixed order of the expected source set, for each source, write the valid label, the product of the valid label and the single-source association score after resolution, and the product of the valid label and the reliability weight in sequence to form the fixed-dimensional model input; The fixed-dimensional model is input into the association strength prediction model, and the output is the association strength with a value range of 0 to 1; the association strength is converted into the corresponding strength level according to the preset grading threshold.

7. A drug target-disease association strength prediction system, characterized in that, include: The evidence collection unit is used to receive target-disease pairs, collect relevant evidence from multiple sources, and form evidence records grouped by source. The feature scoring unit is used to generate feature vectors for each valid source according to the featureization rules corresponding to the evidence source. Each feature vector is input into the corresponding single-source association scoring function to obtain a single-source association score with consistent value direction. The grading configuration unit is used to classify each source into a causal grade with high and low priority based on the causal strength of the evidence sources, and to determine the normalized reliability weight within the grade. The conflict resolution unit is used to detect conflicts based on support thresholds and rejection thresholds, determine a resolution orientation value according to the highest causality level, and cause lower-level conflict sources to converge toward the resolution orientation value. The intensity prediction unit is used to form model inputs according to a fixed order of expected sources, and then feed them into the correlation intensity prediction model to obtain the correlation intensity. The credibility assessment unit is used to calculate the predicted credibility based on the coverage and degree of conflict of valid evidence sources; The result output unit is used to collect the correlation strength, the prediction confidence and evidence traceability details to form a prediction result record. The credibility assessment unit uses the ratio of the number of sources of valid evidence actually obtained to the total number of expected sources of evidence as the evidence coverage. The evidence coverage is used as the base value of credibility. The predicted credibility is obtained by multiplying the base value of credibility by one and subtracting the product of the conflict downsizing coefficient and the degree of conflict. When no conflict is detected, the degree of conflict is zero, and the predicted credibility remains the evidence coverage.

8. The drug target-disease association strength prediction system according to claim 7, characterized in that, The conflict resolution unit writes sources with single-source correlation scores greater than the support threshold into the support source set and sources with single-source correlation scores less than the negation threshold into the negation source set, with the support threshold being greater than the negation threshold. A conflict source set is formed when both the support source set and the negation source set are not empty. The highest causality level is selected from the causality levels involved in the conflict source set in priority order to form the conflict source set within that level. Only the single-source correlation scores and reliability weights in the conflict source set within that level are used to calculate the weighted average as the resolution orientation value. For each source in the set of negative sources, the average of its single-source correlation score and the resolution orientation value is taken as the single-source correlation score after resolution of that source; Calculate the difference between the average score of the supporting source set and the average score of the negative source set. Use the ratio of the number of conflicting sources to the number of valid sources as the conflict percentage, and use the product of the two as the conflict degree.

9. The drug target-disease association strength prediction system according to claim 7, characterized in that, Each source in the feature scoring unit is configured with a corresponding scoring function and historical calibration parameters. All functions output in the direction of enhanced association support as the numerical value increases. Each component in the feature vector is truncated and normalized using the lower and upper boundaries of the historical calibration data. For features with smaller numerical values ​​and stronger association support, the corresponding normalized value is subtracted from one before participating in the weighting. The normalized components are weighted and summed according to the non-negative feature coefficients to obtain the single-source association score. The intensity prediction unit writes a valid label, the product of the valid label and the single-source association score after resolution, and the product of the valid label and the reliability weight for each source in a fixed order according to the expected source set. This forms a fixed-dimensional model input, which is then fed into the association intensity prediction model to output the association intensity with a value range of 0 to 1.

Images