A candidate drug ranking method and system based on topological coding supervised reconstruction

Abstract

The application discloses a candidate drug ranking method and system based on topological coding supervision reconstruction, and belongs to the field of bioinformatics. The method comprises the following steps: step S1, obtaining drug design data; step S2, constructing a drug-biological entity association matrix; step S3, calculating a drug side topological matrix for describing the connection relationship between drugs and a biological entity side topological matrix for describing the connection relationship between biological entities according to the association matrix; step S4, constructing a finite-order topological filter and a spectral stable topological filter to form a bidirectional topological evidence field; step S5, obtaining a reconstruction supervision matrix through a supervised reconstruction objective function; and step S6, obtaining candidate drug ranking based on the reconstruction supervision matrix. The application can improve the high-order mechanism recognition ability, robustness and interpretability in drug-target prediction, drug repositioning, candidate molecule screening and molecular generation result ranking.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a candidate drug ranking method and system based on topological coding supervised reconstruction. Background Technology

[0002] Drug design and drug discovery typically require screening potentially active molecules from a large pool of candidate compounds and predicting their associations with target proteins, disease phenotypes, pathway modules, or cellular states. In recent years, methods based on graph neural networks, matrix factorization, molecular representation learning, and drug-target network propagation have been widely used for drug-target interaction prediction, drug relocation, virtual screening, and lead compound optimization.

[0003] Existing methods typically rely on local supervision signals. For example, in drug-target interaction prediction, edges in known drug-target bipartite graphs are often used as positive samples, while unknown edges are used as candidate samples; in drug relocation, known associations in the drug-disease association matrix are often used as supervision; and in molecular graph learning, molecular embeddings are often obtained through drug molecular graph encoders, and then dot products, similarity functions, or multilayer perceptrons are used to predict drug activity.

[0004] Existing local dot product or local similarity scores are easily influenced by drug exposure, target research activity, the number of known interactions, and data bias. A drug with a large number of known targets may achieve high scores on multiple candidate targets due to its high degree, but this does not necessarily indicate a true structural or mechanistic fit.

[0005] Even if two candidate drug-target pairs have the same local common neighbors or similar local scores, they may still exhibit significant differences in higher-order drug-target-drug, target-drug-target, drug-disease-drug, and drug-pathway-target topologies. Existing local supervision methods struggle to distinguish this type of higher-order mechanism evidence. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes a candidate drug ranking method and system based on topological coding-supervised reconstruction.

[0007] The technical solution adopted in this invention is: A candidate drug ranking method based on topological coding-supervised reconstruction includes the following steps: Step S1: Obtain drug design data; Step S2: Construct a drug-biological entity association matrix, where rows represent drugs or candidate compounds and columns represent targets, diseases, pathways, or cell states. Step S3: Calculate the drug-side topology matrix used to describe the connection relationship between drugs and the biological entity-side topology matrix used to describe the connection relationship between biological entities based on the correlation matrix. Step S4: Construct a finite-order topological filter and a spectrally stable topological filter to form a two-way topological evidence field; Step S5: Obtain the reconstruction supervision matrix through the supervised reconstruction objective function; Step S6: Obtain the candidate drug ranking based on the reconstructed supervision matrix.

[0008] This invention also proposes a candidate drug ranking system based on topological coding-supervised reconstruction to implement the method described above, comprising: Data input module: used to receive drug molecular maps, SMILES, drug-target interaction matrices, drug-disease association matrices, perturbation response data, target sequences, disease phenotypes, and ADMET tags; Molecular representation module: used to generate drug embeddings using molecular fingerprints, graph neural networks, Transformers, or other molecular encoders; The association matrix module is used to construct association matrices for drugs and targets, drugs and diseases, drugs and pathways, or drugs and cell states. Local observation module: used to calculate local dot product score, number of common neighbors, local molecular similarity, drug degree, and entity exposure features; Diagnostic module: Used to calculate the drug design topological discriminability index and determine whether to enable high-order topology supervised reconstruction; Finite Topology Module: Used to aggregate finite high-order drug topological paths; Spectral filtering module: used to obtain spectrally stable topology monitoring signals; Two-way topological field module: used to fuse evidence from the drug side and the biological entity side; Candidate ranking module: Used to output a ranking of candidate drugs based on the reconstruction supervision matrix, molecular properties, ADMET predictions, and task constraints; The explanation audit module is used to output contribution ratio, truncation error, disturbance drift, ranking indicators, and candidate explanation reports; Drug design output module: Used to output candidate drug ranking list, predicted target or indication for candidate drugs, topological evidence description, drug-side contribution ratio, biological entity-side contribution ratio, ADMET or task constraint filtering results, perturbation drift audit results and interpretation report, and provide the above results to virtual screening, drug relocation, generated molecule reordering or lead compound optimization process.

[0009] The beneficial technical effects of the present invention are as follows: This invention proposes a candidate drug ranking method and system based on topological coding-supervised reconstruction. The method includes: acquiring drug molecular structure, drug-target interaction, drug-disease association, drug pathway association, or drug perturbation response data; constructing a drug-biological entity association matrix; constructing drug-side and biological entity-side topological matrices; interpreting local dot products, local molecular similarity, or local drug-target scores as low-order topological observations; calculating a drug design topological discriminability index to determine the existence of non-local topological evidence; when non-local topological evidence exists, constructing finite high-order topological operators, infinite-spectrum stable topological filters, and a bidirectional drug-biological entity topological evidence field; generating a reconstruction supervision matrix through closed-form spectral solutions; and outputting candidate drug ranking, predicted targets, potential indications, topological evidence interpretation, and perturbation robustness audit results based on the reconstruction supervision matrix. This invention can improve the identification ability, robustness, and interpretability of high-order mechanisms in drug-target prediction, drug relocation, candidate molecule screening, and molecular generation result ranking.

[0010] Specifically, firstly, this invention interprets local drug design scores as low-order topological observations, thereby revealing the entanglement between candidate scores and drug degree, target popularity, and data exposure. Secondly, this invention introduces higher-order pathways such as drug-target, drug-disease, and drug-pathway through higher-order topological operators, enabling the discovery of candidate drugs with weak local evidence but strong higher-order mechanism evidence. Thirdly, this invention improves the mechanism explanation capability of candidate predictions by utilizing a bidirectional drug-biological entity topological evidence field, simultaneously leveraging common mechanisms on the drug side and common structures on the biological entity side. Fourthly, this invention uses a topological discriminability index to determine whether a specific drug design task is suitable for topological enhancement before training, avoiding the blind use of complex topological models on locally dominant data. Fifthly, this invention transforms bidirectional network propagation into an interpretable, reproducible, and efficient spectral solution process through closed-form spectral solutions. Sixthly, this invention provides an auditable evidence chain for candidate drug prediction through contribution ratio, truncation error, and perturbation drift, facilitating subsequent experimental verification and molecular optimization. This invention can serve as a supervised enhancement module for existing molecular graph neural networks, drug relocation models, drug-target prediction models, and molecular generation models, exhibiting strong compatibility. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the candidate drug ranking method based on topological coding supervision and reconstruction of the present invention. Figure 2 Design topological diagnostic maps for drugs; Figure 3 A dataset-level topology gain plot; Figure 4 This is a diagram illustrating local entanglement and stability auditing. Figure 5 This is a two-way topological field diagnostic diagram; Figure 6 This is a structural diagram of the candidate drug ranking system based on topological coding supervision and reconstruction of the present invention; Figure 7 The figure shows the five-fold cross-validation ROC-AUC results of the topology reconstruction score on the target side of this invention. Detailed Implementation

[0012] Existing drug design models rely excessively on local molecular similarity, local dot product scoring, or low-order drug-target associations, making them susceptible to the influence of drug degree, target popularity, and data exposure bias. Furthermore, existing drug design processes lack pre-training topological discriminability diagnosis, failing to determine whether a specific task is suitable for high-order topological supervised reconstruction. Therefore, this invention proposes a candidate drug ranking method based on topological coding supervised reconstruction, reconstructing local supervision signals in drug design tasks into high-order, spectrally stable, bidirectional, diagnostic, and auditable candidate drug design supervision signals. This invention constructs drug-side topological matrices, biological entity-side topological matrices, finite high-order topological operators, infinite spectrally stable topological filters, bidirectional drug-biological entity topological evidence fields, closed-form spectral solutions, and drug design topological discriminability indices for candidate drug screening, drug-target interaction prediction, drug relocation, lead compound selection, and molecular generation result ranking.

[0013] Specifically, this invention proposes a candidate drug ranking method based on topological coding-supervised reconstruction. This method acquires drug design-related data, including molecular structure, drug-target relationships, drug-disease relationships, drug-pathway relationships, perturbation expression profiles, and ADMET tags; constructs a drug-biological entity association matrix, and calculates drug-side and biological entity-side topological matrices from the association matrix; constructs finite-order and spectrally stable topological filters to form a bidirectional topological evidence field; obtains the reconstruction supervision matrix through a supervised reconstruction objective function; and outputs candidate drug ranking, contribution explanation, and perturbation drift report based on the reconstruction supervision matrix. This invention creatively extends local supervision signals to high-order topological supervision, reconstructing local supervision signals into high-order, bidirectional, and spectrally stable topological supervision signals, reducing bias caused by relying solely on local dot products. Simultaneously, by constructing drug-side and biological entity-side topological matrices from the drug-biological entity association matrix, and then forming a bidirectional topological evidence field through finite-order and spectrally stable filtering, the contributions of both drug-side and biological entity-side are preserved, making the candidate ranking interpretable. Furthermore, the ranking stability is controlled through spectral stability filtering and perturbation drift auditing.

[0014] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0015] Example 1: like Figure 1 As shown, a candidate drug ranking method based on topological coding-supervised reconstruction includes the following steps: Step S1: Obtain drug design-related data, including molecular structure, drug-target relationship, drug-disease relationship, drug-pathway relationship, perturbation expression profile, and ADMET tag.

[0016] Step S2: Construct a drug-biological entity association matrix, where rows represent drugs or candidate compounds and columns represent targets, diseases, pathways, or cell states.

[0017] The association matrix between drugs and biological entities is represented by the following formula: in: Represents the association matrix between drugs and biological entities; Number of drugs or candidate molecules; This represents the total number of biological entities uniformly incorporated into the columns of the association matrix. This total number, or set of biological entities, includes targets, diseases, pathways, cell states, or response states, and is represented by a unified numbering system in the matrix. The column; Indicates the first One drug, Indicates the first One biological entity; It is obtained by mapping confirmed drug-target interactions, drug-disease associations, drug-pathway associations, drug-cell state associations, perturbation expression profile tags or perturbation response tags, and manually verified tags. The existence of a relationship is recorded as 1 or real value intensity, and the absence or unknown is handled according to the preset missing rules. It is the root matrix of all subsequent topological matrices.

[0018] Step S3: Calculate the drug-side topology matrix used to describe the connection relationship between drugs and the biological entity-side topology matrix used to describe the connection relationship between biological entities based on the correlation matrix.

[0019] The calculation formula is as follows: in, Represents the drug-side topology matrix; Represents the topological matrix of the biological entity side; , indicating the first The drug and the first The strength of connections between biological entities sharing a drug; , indicating the first The biological entity and the first The drug connectivity strength shared by multiple biological entities. This connectivity calculation is a second-order co-occurrence projection calculation based on the correlation matrix. The two matrices characterize the common connectivity structure on the drug side and the entity side, respectively. Subsequent finite higher-order topological coding, spectral stable filtering, and bidirectional topological evidence field are based on this common connectivity structure.

[0020] Step S4: Construct a finite-order topological filter and a spectrally stable topological filter to form a two-way topological evidence field.

[0021] like Figure 2 As shown, the drug design task first enters the local observation unit, where it calculates the local dot product, common neighbors, local molecular similarity, drug degree, and entity exposure features. It then enters the exposure correction unit to reduce the dominance of high-exposure drugs or high-frequency entities on the local score. Finally, it enters the topology diagnostic unit to determine whether the task is locally dominant or topologically discernible. Locally dominant tasks directly output a diagnostic report, while topologically discernible tasks proceed to the enhancement process.

[0022] Finite-order topological filters are used to aggregate evidence from finite high-order paths, and the computational relationship is as follows: in, Represents a finite-order topological filter; Desirable or ; This is the highest propagation order; This is the spectral stability scale parameter. From the matrix The result of continuous multiplication is that, representing Topological path; Used to suppress excessive propagation along long paths.

[0023] The calculation formula for a spectrally stable topological filter is as follows: in, Indicates a spectrally stable topology filter; It is the identity matrix; This is the spectral stability scaling parameter, used to control convergence and stability. When infinite-order but stable topological propagation is required, it is determined by... The spectral stability analytical filtering result is calculated using the identity matrix, without directly expanding an infinitely long path. This result avoids over-diffusion caused by unconstrained propagation.

[0024] The computational formula for the two-way topological evidence field is as follows: in, Represents a two-way topological evidence field. This represents the drug side spectrum filtering function. Represents the side spectrum filtering function for biological entities; and These are not externally fixed function names, but rather the drug-side and biological entity-side spectral filtering functions defined in this scheme. Both can be instantiated from finite-order topological filters, spectrally stable analytical filters, or closed-loop spectral filters induced by supervised reconstruction targets; specifically, This is a drug-side topological evidence transfer item. For entity-side topological evidence transfer items, Indicates the first The drug and the first Strength of bidirectional topological evidence between biological entities.

[0025] like Figure 5 As shown, the drug-side topology and the biological entity-side topology together constitute a two-way topological field; the two-way topological field enters the spectral closed-form solution, forming contribution interpretations of drug-side dominance, entity-side dominance, or two-way balance, respectively, and finally merges into the reconstruction supervision matrix.

[0026] Step S5: Obtain the reconstruction supervision matrix through the supervised reconstruction objective function.

[0027] The formula for reconstructing the supervision matrix is ​​as follows: in, This represents the reconstruction of the supervision matrix; This is a drug side-closed spectrum analysis. For biological entities, the spectral solution is a closed-form spectral solution. and It is obtained by solving the supervised reconstruction objective function; Represents the association matrix between drugs and biological entities; Indicates the first The candidate drug and the first The reconstruction supervision strength among biological entities, which is determined by the drug-side topological reconstruction term. With entity-side topology reconstruction items Provided jointly. Used for candidate ranking, drug repositioning, target prediction, or candidate generation and screening.

[0028] The expression for the supervised reconstruction objective function is as follows: in, and The supervised reconstruction weights are defined by α and β, which are regularization parameters. ||·||F² represents the squared Frobenius norm. The supervised reconstruction objective function uses the original incidence matrix A as the supervised target, making... A and A The known correlation structure is reconstructed jointly. The supervised reconstruction objective function is used to obtain the reconstruction supervision signal for use by the trainable model.

[0029] Drug lateral closed spectrum analysis: Biological entity lateral closed spectral analysis: in, Represents the association matrix between drugs and biological entities; express Singular value decomposition; The singular vector matrix on the drug side; The singular vector matrix on the biological entity side; It is a singular value diagonal matrix; For the first One singular value; , For regularization parameters; Reconstruct weights for drug-side supervision; Reconstruct weights for supervision on the biological entity side; This indicates that the elements within the parentheses form a diagonal matrix. The supervised reconstruction objective function first uses... To obtain the target of reconstruction and Substitute The reconstructed supervision matrix is ​​obtained, thus forming a relationship chain of "supervised reconstruction objective function - closed spectral solution - reconstructed supervision matrix".

[0030] Step S6: Based on the reconstructed supervision matrix, obtain the candidate drug ranking, contribution explanation, and perturbation drift report.

[0031] The ranking score is calculated using the following formula: in, Indicating candidate drugs For biological entities The ranking score; These are the weights of the topological evidence field, used to adjust the contribution of the bidirectional topological evidence field to the final ranking score. The ranking criteria are determined by the validation set, a preset weight table, or task constraints, without using test results to determine the reverse. From the reconstructed supervision matrix The Line 1 Column elements; From a two-way topological evidence field The Line 1 Column elements.

[0032] Based on ranking score Output the candidate list, sort the candidate drugs, and include them in the report along with the perturbation and drift audit.

[0033] like Figure 3 As shown, candidate datasets are entered into the topology enhancement model after task-level diagnosis, followed by gain evaluation. The gain evaluation results correspond to interaction prediction, relocation prediction, pathway association prediction, and perturbation response prediction, respectively, and are summarized into the task screening conclusion.

[0034] This invention does not require all data to be present simultaneously. For drug-target tasks, the biological entity can be the target protein; for drug relocation tasks, the biological entity can be the disease; for pathway or cell state tasks, the biological entity can be a pathway, phenotype, cell state, or response module.

[0035] The data that can be received by this invention includes one or more of the following: drug molecular structure, SMILES, molecular diagram, drug-target interaction matrix, drug-disease association matrix, drug-pathway association matrix, drug perturbation expression profile, target protein sequence, disease phenotypic similarity, drug chemical similarity, or ADMET property data.

[0036] Example 2: A candidate drug ranking method based on topological coding-supervised reconstruction includes the following steps: Step S1: Obtain drug design-related data, including molecular structure, molecular diagram, drug-target interaction matrix, drug-disease association matrix, drug-pathway association matrix, perturbation expression profile, or ADMET tag.

[0037] Candidate drugs can be represented as molecular graphs, molecular fingerprints, molecular embeddings, or network embeddings; where nodes in the molecular graph represent atoms and edges represent chemical bonds.

[0038] Step S2: Construct a drug-biological entity association matrix, where rows represent drugs or candidate compounds and columns represent targets, diseases, pathways, or cell states.

[0039] Step S3: Construct drug-side topology matrix and biological entity-side topology matrix based on the association matrix, which are used to describe the common connection relationships between drugs and between biological entities, respectively.

[0040] Step S4, as follows Figure 2 As shown, local dot product, local molecular similarity, number of common neighbors, local association score or matrix factorization score are interpreted as low-order topological observations, and exposure correction is performed in combination with drug degree and entity exposure features; then the drug design topological discriminability index is calculated to determine whether there is non-local topological evidence in the current task that cannot be explained by local scores and exposure features.

[0041] In the specific calculation, the local observation score is first calculated on the training image. Topology score and exposure features Standardize the three to obtain , and Use again and right Performing linear residual decomposition yields residual topological components that cannot be explained by local scores and exposure features. .

[0042] Topological discriminability index Calculation, where The higher-order topology tail proportion is used to measure the portion of a finite higher-order or spectral filter score that exceeds the lower-order observations. To prevent the stability constant from having a denominator of zero, the following is used: With preset threshold Comparison, default ;when When it is determined to be a topologically discernible task, finite higher-order topological operators, spectral stable filters, and bidirectional topological evidence fields are activated; when When this is determined to be a locally dominant task, local observation or low-level supervision is retained.

[0043] This step extends finite high-order topology propagation to a spectrally stable filtering result, avoiding over-diffusion or numerical instability caused by infinite path propagation; it directly corresponds to the spectrally stable topology filter, and its output serves as... or One implementation form is entering a bidirectional topological evidence field. .

[0044] like Figure 4 As shown, when the diagnostic results indicate the presence of nonlocal topological evidence, a finite higher-order topological operator is constructed to aggregate higher-order path evidence such as drug-target-drug, target-drug-target, or drug-disease-drug. When the spectral convergence condition is met, an infinite spectrally stable topological filter is constructed to obtain a spectrally stable topological coding supervision signal. This signal is then combined with spectral convergence checks and perturbation drift audits to determine whether candidate results are stable.

[0045] Furthermore, a two-way drug-biological entity topological evidence field is constructed, which integrates drug-side topological evidence and biological entity-side topological evidence.

[0046] Step S5: Generate a reconstruction supervision matrix that can be used to train the drug design model by solving the closed-form spectral matrix or equivalent spectral filtering. The closed-form spectral matrix is ​​not a typical model training step, nor is it an arbitrary black-box optimization step; rather, it is an analytical solution method for the unconstrained Frobenius regularized supervised reconstruction objective. This method first addresses the correlation matrix... Perform spectral decomposition, and then apply the regularization parameter. , get and Finally generated Equivalent spectral filtering is used to solve low-rank approximations, random spectral decompositions, or fractional implementations covering large-scale matrices.

[0047] Step S6: Reconstruct the supervision matrix The reconstructed supervision signal and its derivatives are used for drug-target interaction prediction, drug relocation, candidate compound ranking, molecular generation result screening, lead compound optimization, or risk filtering. Reconstructed supervision matrix It is the matrix carrier for reconstructing the supervision signal, and the matrix contains... For the first The drug and the first The strength of reconstruction supervision among biological entities; when used for training. It can be converted into soft labels, ranking targets, distillation targets, or sample weights; when used for screening... and Merge into ranking score Outputs a list of candidate drugs, predicted targets or indications, sources of topological evidence, proportion of contribution from the drug side, proportion of contribution from the entity side, ranking metrics, truncation error, perturbation drift, and interpretability report.

[0048] This invention does not merely propose a new drug prediction model backbone, but rather a supervisory signal reconstruction mechanism for drug design. This mechanism reconstructs supervisory signals that were originally limited to local dot products, local molecular similarity, or low-order drug-target edges into topologically encoded supervisory signals that include high-order drug-target network paths, drug-side topology, target-side topology, and spectral stabilization filtering.

[0049] This supervised reconstruction mechanism can be located either before model training or after molecule generation. When located before training, the reconstruction supervision matrix can serve as training labels, soft labels, ranking targets, or distillation targets; when located after molecule generation, the reconstruction supervision scores can be used to re-rank the generated molecules at the network mechanism level.

[0050] This method can be used as an independent drug design model, or as a supervised enhancement module for existing molecular graph neural networks, drug-target prediction models, drug relocation models, molecular generation models, or virtual screening platforms.

[0051] The effectiveness of the candidate drug ranking method based on topological coding supervised reconstruction of this invention is verified below: The controlled topological relationship data is divided into five folds, with the comprehensive association label as the true category and the topological reconstruction score of the target side of this invention as the candidate ranking score; each time, one fold is used for verification, the judgment threshold is moved according to the ranking score, the true positive rate and false positive rate are calculated, and the ROC curve is plotted. The gray diagonal line in the figure represents the random judgment level.

[0052] like Figure 7 As shown, the five-fold AUCs were 0.918, 0.923, 0.917, 0.914, and 0.931, respectively, with a mean AUC of 0.921 and a standard deviation between folds of 0.007. Specifically, the five-fold ROC curves for the target-side topology reconstruction candidate ranking are shown. The five colored curves correspond to the five validation folds: green line represents fold 1 (AUC = 0.918); blue line represents fold 2 (AUC = 0.923); red line represents fold 3 (AUC = 0.917); yellow line represents fold 4 (AUC = 0.914); purple line represents fold 5 (AUC = 0.931); and the black line is the mean ROC curve. All fold ROC curves are clearly above the random decision line, indicating that the target-side topology reconstruction score can stably distinguish between highly correlated and low-correlation candidates. Figure 7 The results show that the topological encoding and supervised reconstruction of the method of the present invention can enable highly correlated candidates to obtain higher overall ranking scores and maintain consistent performance in five-fold verification; the five ROC curves and their AUC provide directly verifiable evidence of the effectiveness of this distinguishing ability, thus proving that the method is effective, feasible and stable.

[0053] Example 3: like Figure 6 As shown, a candidate drug ranking system based on topological coding-supervised reconstruction includes: Data input module: used to receive drug molecular maps, SMILES, drug-target interaction matrix, drug-disease association matrix, perturbation response data, target sequences, disease phenotypes and ADMET tags.

[0054] Molecular representation module: Used to generate drug embeddings using molecular fingerprints, graph neural networks, Transformers, or other molecular encoders.

[0055] The association matrix module is used to construct association matrices for drugs-targets, drugs-diseases, drugs-pathways, or drugs-cell states.

[0056] Local observation module: used to calculate local dot product score, number of common neighbors, local molecular similarity, drug degree, and entity exposure features.

[0057] Diagnostic module: Used to calculate the drug design topology discriminability index and determine whether to enable high-order topology supervised reconstruction.

[0058] Finite Topology Module: Used to aggregate finite high-order drug topology paths.

[0059] Spectral filtering module: used to obtain spectrally stable topology monitoring signals.

[0060] Two-way topological field module: used to fuse evidence from the drug side and the biological entity side.

[0061] Candidate ranking module: Used to rank candidate drugs based on the reconstruction supervision matrix, molecular properties, ADMET predictions and task constraints.

[0062] The Explanation Audit module is used to output contribution ratio, truncation error, disturbance drift, ranking indicators, and candidate explanation reports.

[0063] Drug design output module: Used to output candidate drug ranking list, predicted target or indication for candidate drugs, topological evidence description, drug-side contribution ratio, biological entity-side contribution ratio, ADMET or task constraint filtering results, perturbation drift audit results and interpretation report, and provide the above results to virtual screening, drug relocation, generated molecule reordering or lead compound optimization process.

[0064] This invention takes a drug set and a target set as input and constructs a drug-target interaction matrix. The system first calculates the local drug-target score, and then calculates the topological discriminability index. .when At that time, among them To preset the topologically discernible threshold, which is set to 0.5 by default, the system determines the current task as a topologically discernible task, constructs high-order topological supervision, and generates a bidirectional drug-target topological evidence field; when When the system determines that the current task is locally dominant, it retains local observations or low-order supervision, and does not force the use of high-order topology reconstruction. Finally, the supervision matrix is ​​reconstructed. elements With bidirectional topological evidence field elements Merging to form a ranking score The ranking score serves as a prediction score between the candidate drug and the target. This invention can discover potential interactions with few local common neighbors but supported by mechanisms through higher-order drug-target pathways.

[0065] The input is a drug-disease association matrix, where rows represent marketed drugs or candidate compounds, and columns represent diseases. The drug-side topology reflects multiple drugs sharing indications or disease profiles, while the disease-side topology reflects multiple diseases sharing therapeutic drugs or pathological mechanisms. The system outputs candidate drug-disease association scores and the proportion of drug-side contribution, used to explain whether relocation suggestions primarily stem from drug-like mechanisms or shared disease-side therapeutic structures.

[0066] The system takes candidate molecules generated by a molecular generative model as input. It obtains the candidate molecule embeddings through a molecular graph encoder and integrates them into existing drug-target or drug-disease networks to obtain an extended correlation matrix. The system calculates a topologically encoded supervision score on the extended correlation matrix and, combined with chemoeffectiveness, synthetic accessibility, ADMET prediction, and topologically encoded target or disease scores, outputs a ranking of candidate molecules.

[0067] Input a drug-adverse reaction matrix, a drug-toxicity target matrix, or a drug-ADMET tag matrix. The system constructs drug-side topology and risk entity-side topology, and aggregates risk paths using higher-order topology operators. For candidate drugs, if their topology encoding score on the toxicity entity is high, a risk warning is output.

[0068] The system performs random edge deletion, degree-preserving reconnection, long-tail target retention, or cold-start drug masking on the drug-target network to obtain a perturbation matrix. The system calculates the perturbation drift. If the candidate drug ranking remains stable after perturbation, robust support is output; if high-scoring candidate pairs drift significantly, boundary hints are output.

[0069] In this invention, the biological entity is not limited to target proteins, but can also be a disease, pathway, cell state, gene module, transcription factor, protein complex, phenotype, or adverse reaction. The drug representation is not limited to molecular diagrams, but can also be SMILES sequences, molecular fingerprints, drug perturbation expression profiles, three-dimensional conformations, drug text descriptions, or multimodal fusion representations. Output tasks may include drug-target interaction prediction, drug-disease association prediction, drug relocation, lead compound screening, molecular candidate generation and ranking, ADMET risk prediction, drug combination prediction, or personalized drug response prediction.

[0070] This invention proposes a candidate drug ranking method based on topological coding-supervised reconstruction. This method utilizes a drug design-oriented topological coding-supervised reconstruction mechanism to reconstruct local dot products, local molecular similarities, or low-order drug-target edges into high-order topological coding supervision signals. A drug design topological discriminability index is proposed to determine the existence of non-local topological evidence before training. A bidirectional evidence fusion method is proposed for drug-side and biological entity-side topological matrices, preserving common mechanisms on the drug side and common structures on the entity side. The application of finite high-order topological operators and infinite spectral stable filters in candidate drug screening, drug-target prediction, and drug relocation is proposed. A closed-form spectral solution is proposed to generate a reconstructed supervision matrix, improving the interpretability, reproducibility, and computational efficiency of bidirectional network propagation. An auditable report for candidate drugs combining contribution ratio, truncation error, perturbation drift, and ranking indicators is proposed.

[0071] First, this invention interprets local drug design scores as low-order topological observations, thereby revealing the entanglement between candidate scores and drug degree, target popularity, and data exposure. Second, this invention introduces higher-order drug-target, drug-disease, and drug-pathway pathways through higher-order topological operators, enabling the discovery of candidate drugs with weak local evidence but strong higher-order mechanism evidence. Third, this invention improves the mechanism explanation capability of candidate predictions by utilizing a bidirectional drug-biological entity topological evidence field, simultaneously leveraging common mechanisms on the drug side and common structures on the biological entity side. Fourth, this invention uses a topological discriminability index to determine whether a specific drug design task is suitable for topological enhancement before training, avoiding the blind use of complex topological models on locally dominant data. Fifth, this invention transforms bidirectional network propagation into an interpretable, reproducible, and efficient spectral solution process through closed-form spectral solutions. Sixth, this invention provides an auditable evidence chain for candidate drug prediction through contribution ratio, truncation error, and perturbation drift, facilitating subsequent experimental verification and molecular optimization. This invention can serve as a supervised enhancement module for existing molecular graph neural networks, drug relocation models, drug-target prediction models, and molecular generation models, exhibiting strong compatibility.

[0072] For any parts not mentioned above, existing technologies can be adopted or referenced.

[0073] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any improvements, equivalent substitutions, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A candidate drug ranking method based on topological coding supervised reconstruction, characterized in that... Includes the following steps: Step S1: Obtain drug design data; Step S2: Construct a drug-biological entity association matrix, where rows represent drugs or candidate compounds and columns represent targets, diseases, pathways, or cell states. Step S3: Calculate the drug-side topology matrix and the biological entity-side topology matrix to describe the connection relationship between drugs and biological entities based on the drug-biological entity association matrix. Step S4: Construct a finite-order topological filter and a spectrally stable topological filter to form a two-way topological evidence field; Step S5: Obtain the reconstruction supervision matrix through the supervised reconstruction objective function; Step S6: Obtain the candidate drug ranking based on the reconstructed supervision matrix.

2. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 1, characterized in that, In step S1: the drug design data includes molecular structure, drug-target relationship, drug-disease relationship, drug-pathway relationship, perturbation expression profile and ADMET tag.

3. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 2, characterized in that, In step S2, the drug-biological entity association matrix is ​​represented by the following formula: in: Represents the association matrix between drugs and biological entities; Number of drugs or candidate molecules; This represents the set of biological entities uniformly incorporated into the columns of the association matrix. This set of biological entities includes targets, diseases, pathways, and cell states or response states, and is represented by a matrix after being uniformly numbered. The column; Indicates the first One drug, Indicates the first One biological entity; It is obtained by mapping confirmed drug-target interactions, drug-disease associations, drug-pathway associations, drug-cell state associations, perturbation expression profile tags or perturbation response tags, and manually verified tags. The existence of a relationship is recorded as 1 or real value intensity, and the absence of a relationship or the unknown is handled according to the preset missing rules.

4. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 3, characterized in that, In step S3, the following relation is calculated: in, Represents the drug-side topology matrix; Represents the topological matrix of the biological entity side; , Right now The , Item, indicating the first The drug and the first The strength of connections between biological entities sharing a drug; , Right now The Item, indicating the first The biological entity and the first Drug connectivity strength shared by multiple biological entities.

5. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 4, characterized in that, In step S4, a finite-order topological filter is used to aggregate evidence from a finite number of higher-order paths, and the calculation relationship is as follows: in, Represents a finite-order topological filter; Pick or ; This is the highest propagation order; For spectral stability scaling parameters; From the matrix The result of continuous multiplication is that, representing Topological path; Used to suppress excessive propagation along long paths; The calculation formula for a spectrally stable topological filter is as follows: in, Indicates a spectral-stabilized topology filter; It is the identity matrix; This is the spectral stability scaling parameter, used to control convergence and stability; The computational formula for the two-way topological evidence field is as follows: in, Represents a two-way topological evidence field. This represents the drug side spectrum filtering function. This represents the spectral filtering function for biological entities; specifically, For drug-side topological evidence transfer terms, it means that filtering is first performed on the drug-side topological matrix, and then the evidence is passed back to the drug-entity relationship space through the association matrix. For entity-side topological evidence transfer terms, it means that filtering is first performed on the biological entity-side topological matrix, and then backpropagated to the drug-entity relationship space; Indicates the first The drug and the first Strength of bidirectional topological evidence between biological entities.

6. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 5, characterized in that, Step S4 also includes the following steps: Step S41: Interpret local dot product, local molecular similarity, number of common neighbors, local association score, or matrix factorization score as low-order topological observations; Step S42: Calculate the drug design topological discriminability index to determine whether there is non-local topological evidence in the current task that cannot be explained by local scores and exposure features, and obtain the diagnostic result; Step S43: When the diagnostic results indicate the existence of nonlocal topological evidence, construct a finite number of higher-order topological operators to aggregate higher-order path evidence such as drug-target-drug, target-drug-target, or drug-disease-drug. Step S44: When the spectral convergence condition is met, construct an infinite spectral stable topology filter to obtain a spectral stable topology-coded monitoring signal.

7. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 6, characterized in that, In step S5, the calculation formula for the reconstructed supervision matrix is ​​as follows: in, This indicates the reconstruction of the supervision matrix; This is a drug side-closed spectrum analysis. For biological entities, the spectral solution is a closed-form spectral solution. and It is obtained by solving the supervised reconstruction objective function; Represents the association matrix between drugs and biological entities; Indicates the first The candidate drug and the first The reconstruction supervision strength among biological entities, which is determined by the drug-side topological reconstruction term. With entity-side topology reconstruction items Provided jointly; Used for candidate ranking, drug repositioning, target prediction, or candidate generation and screening; The expression for the supervised reconstruction objective function is as follows: in, and The supervised reconstruction weights are defined by α and β, which are regularization parameters. ||·||F² represents the squared Frobenius norm. The supervised reconstruction objective function uses the original incidence matrix A as the supervised target, making... A and A The known correlation structure is jointly reconstructed; the supervised reconstruction objective function is used to obtain the reconstruction supervision signal that can be used by the trainable model. Drug lateral closed spectrum analysis: Biological entity lateral closed spectral analysis: in, Represents the association matrix between drugs and biological entities; express Singular value decomposition; The singular vector matrix on the drug side; The singular vector matrix of the biological entity side; It is a singular value diagonal matrix; For the first One singular value; , For regularization parameters; Reconstruct weights for drug-side supervision; Reconstruct weights for supervision on the biological entity side; This indicates that the elements within the brackets form a diagonal matrix; Supervised reconstruction objective function first To obtain the target of reconstruction and Substitute The reconstructed supervision matrix is ​​obtained.

8. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 7, characterized in that, In step S6, the ranking score is calculated using the following formula: in, Indicating candidate drugs For biological entities The ranking score; These are the weights of the topological evidence field, used to adjust the contribution of the bidirectional topological evidence field to the final ranking score. Determined by validation set sorting criteria, preset weight table, or task constraints; From the reconstructed supervision matrix The Line 1 Column elements; From a two-way topological evidence field The Line 1 Column elements; Based on ranking score Output the candidate list to sort the candidate drugs.

9. The candidate drug ranking method based on topological coding supervised reconstruction according to claim 7, characterized in that, Step S6 also includes obtaining contribution explanations and perturbation drift reports based on the reconstructed supervision matrix.

10. A candidate drug ranking system based on topological coding supervised reconstruction, used to implement the method as described in any one of claims 1-9, characterized in that, include: Data input module: used to receive drug molecular maps, SMILES, drug-target interaction matrices, drug-disease association matrices, perturbation response data, target sequences, disease phenotypes, and ADMET tags; Molecular representation module: used to generate drug embeddings using molecular fingerprints, graph neural networks, Transformers, or other molecular encoders; The association matrix module is used to construct association matrices for drugs and targets, drugs and diseases, drugs and pathways, or drugs and cell states. Local observation module: used to calculate local dot product score, number of common neighbors, local molecular similarity, drug degree, and entity exposure features; Diagnostic module: Used to calculate the drug design topological discriminability index and determine whether to enable high-order topology supervised reconstruction; Finite Topology Module: Used to aggregate finite high-order drug topological paths; Spectral filtering module: used to obtain spectrally stable topology monitoring signals; Two-way topological field module: used to fuse evidence from the drug side and the biological entity side; Candidate ranking module: Used to output a ranking of candidate drugs based on the reconstruction supervision matrix, molecular properties, ADMET predictions, and task constraints; The explanation audit module is used to output contribution ratio, truncation error, disturbance drift, ranking indicators, and candidate explanation reports; Drug design output module: It is used to output a candidate drug ranking list, the predicted target or indication corresponding to the candidate drug, topological evidence description, drug-side contribution ratio, biological entity-side contribution ratio, ADMET or task constraint filtering results, perturbation drift audit results and interpretation report, and provide the above results to the virtual screening, drug relocation, generated molecule reordering or lead compound optimization process.

Images