Drug enzyme inhibition risk prediction system and prediction method based on structural characterization

Abstract

The invention discloses a drug enzyme inhibition risk prediction system and method based on structural characterization, relates to a drug enzyme inhibition risk prediction device and method, and aims to solve the problem that existing drug enzyme inhibition prediction lacks confidence quantification and credibility quantification. An input receiving module is used for generating an input file; the feature construction module is used for generating a feature vector and an atomic matrix; the structure prediction module is used for outputting structure prediction coordinates corresponding to one or more candidate compound conformations; the confidence estimation module is used for generating a confidence score; the result export module is used for outputting the sorted coordinates and scores; the rule evidence scoring module is used for analyzing the conformation of the candidate compound based on a preset geometric rule, and generating rule evidence directly related to the conformation and a corresponding rule score; and the fusion output module is used for outputting a risk prediction result. The method has the beneficial effect of improving the prediction stability and generalization ability.

Description

Technical Field

[0001] This invention relates to a device and method for predicting the risk of drug enzyme inhibition. Background Technology

[0002] With the increasing prevalence of multidrug combination therapy in clinical treatment, the risks of drug-related adverse reactions and drug-drug interactions are becoming increasingly prominent. The inhibitory effect of drugs on metabolism-related enzymes may lead to decreased drug clearance and increased in vivo exposure, thereby posing safety risks. Therefore, rapid, low-cost, and scalable prediction and assessment of the inhibitory risks of metabolism-related enzymes in candidate compounds in the early stages of new drug development is of significant practical importance.

[0003] In existing technologies, the assessment of the inhibition risk of metabolism-related enzymes mainly relies on in vitro experiments (such as enzyme inhibition assays, liver microsome / hepatocyte systems, etc.) and subsequent in vivo studies. While these methods provide relatively direct experimental evidence, they generally suffer from drawbacks such as long cycles, high costs, limited throughput, and reproducibility issues due to variations in experimental conditions, making them unsuitable for early-stage large-scale screening and rapid iteration. To improve screening efficiency, researchers have proposed computational models based on quantitative structure-activity relationship (QSAR) and structural methods based on molecular docking / scoring functions to infer inhibitory tendencies or binding abilities from molecular structural features. However, existing QSAR methods often rely on artificial features or molecular fingerprints, making it difficult to fully characterize complex three-dimensional interactions; docking and physical scoring methods are sensitive to conformation and parameters, have high computational costs, and limited applicability in high-throughput scenarios.

[0004] In recent years, the application of deep learning in drug discovery has developed rapidly, giving rise to various model frameworks for predicting drug-target binding strength and interactions. These include affinity prediction methods based on sequence or graph structures, and structural deep learning methods based on structural representations such as 3D meshes and point clouds (e.g., 3D convolutional models). Simultaneously, the development of high-precision protein structure prediction technologies has expanded the coverage of available structural information, providing richer data sources for structure-representation-driven predictions. Despite the progress made by these methods in various tasks, existing technologies still have several prominent shortcomings for high-risk decision-making tasks such as "predicting the inhibition risk of metabolism-related enzymes": First, training data often suffers from heterogeneity in origin, label noise, class imbalance, and insufficient chemical spatial coverage, leading to decreased generalization ability of models in novel compounds or distribution-biased scenarios. Second, most models only output point predictions or binary classification results, lacking a quantitative expression of the reliability of the prediction results, making it difficult to provide a reference for the "confidence level" of research and development decisions. Third, their robust modeling ability for a few "special samples / special interaction patterns" is insufficient, easily resulting in high-confidence false predictions. Existing research has shown that techniques such as model calibration and uncertainty estimation can improve the consistency between predicted probabilities and actual accuracy, and provide a basis for risk control in downstream decision-making. However, in the engineering implementation of prediction of metabolism-related enzyme inhibition risk, how to combine structural characterization, improve the robustness of special samples, and output usable confidence and credibility scores still needs further solutions.

[0005] Therefore, there is an urgent need for a computational method and system for predicting the effects of enzymes related to drug metabolism. This system should not only ensure predictive performance but also quantify and calibrate the confidence and reliability of the prediction results, and improve adaptability to special samples and distribution shift scenarios, thereby better serving the early screening and safety assessment of drug development. Summary of the Invention

[0006] The purpose of this invention is to address the lack of confidence and reliability metrics in existing drug enzyme inhibition prediction methods, and to propose a drug enzyme inhibition risk prediction system and method based on structural characterization.

[0007] The present invention provides a drug enzyme inhibition risk prediction system based on structural characterization, comprising an input receiving module, a feature construction module, a structure prediction module, a confidence estimation module, a result export module, a rule evidence scoring module, and a fusion output module.

[0008] The input receiving module is used to receive the raw data of the drug-enzyme complex to be evaluated and generate an input file;

[0009] The feature construction module is used to perform feature processing on the input file, generating feature vectors and atomic matrices;

[0010] The structure prediction module receives feature vectors and atomic matrices, performs structure prediction calculations using a generative architecture that includes a pairwise attention mechanism backbone and a diffusion model, and outputs the structure prediction coordinates corresponding to one or more candidate complex conformations.

[0011] The confidence estimation module receives the predicted coordinates of the structure and calculates and generates a confidence score based on the conformational consistency and distribution dispersion of the candidate complex.

[0012] The result export module is used to receive the predicted structural coordinates and confidence scores, summarize and sort the confidence scores corresponding to the multiple candidate complex conformations, and output the sorted coordinates and scores.

[0013] The rule-based evidence scoring module is used to receive the sorted coordinates and scores, and analyze the conformation of the candidate complex based on preset geometric rules to generate rule-based evidence and corresponding rule scores directly related to the conformation.

[0014] The fusion output module is used to fuse confidence scores and rule scores to output risk prediction results.

[0015] Furthermore, the raw input data includes structural files or multiple sequence alignment information of the drug-enzyme complex to be evaluated.

[0016] Furthermore, the structure prediction module includes a feature embedding unit, a backbone evolution unit, and a diffusion generation unit;

[0017] The feature embedding unit is used to encode the feature vector through an embedding layer and relative positions to generate initial paired features and single features;

[0018] The backbone evolutionary unit contains L stacked pairs of attention mechanism modules. Each layer interacts with information through triangular updates and self-attention mechanisms to output updated feature representations.

[0019] The diffusion generation unit is used to receive the updated feature representation output by the backbone evolution unit as a condition, start from Gaussian noise distribution, and output the three-dimensional coordinate matrix of the candidate conformation after T-step denoising sampling; the three-dimensional coordinate matrix of the candidate conformation is the structure prediction coordinate. Where L is an integer greater than 1 and T is an integer greater than 1.

[0020] Furthermore, the confidence estimation module includes a geometric feedback feature construction unit, a local confidence calculation unit, and a hybrid density network unit;

[0021] The geometric feedback feature construction unit is used to receive the predicted coordinates of the structure, extract the representative atoms, and calculate the Euclidean distance matrix between all atom pairs. After encoding the distance matrix with a one-hot code or mapping with a radial basis function, it is superimposed back into the pair representation of the pair attention mechanism backbone. The feature is refined through an additional pair attention module layer.

[0022] The local confidence calculation unit outputs the pLDDT score for each residue through the prediction head, and the pLDDT score is defined as the deviation statistics at different distance thresholds;

[0023] A hybrid density network unit is used to model the distance between atomic pairs as a Gaussian mixture distribution. The log-likelihood probability of the predicted distance under the Gaussian mixture distribution is used as a high-level confidence index of the interaction, and a confidence score is output.

[0024] Furthermore, the rule-based evidence scoring module includes a geometric distance determination unit, a contact and interaction determination unit, and a spatial distribution and span determination unit;

[0025] Geometric distance determination unit, used to determine the rigid geometric constraint between drug enzyme ligand and prosthetic group;

[0026] The contact and interaction determination unit is used to determine the contact density and interaction type between the drug enzyme ligand and a specific region of the protein pocket;

[0027] The spatial distribution and span determination unit is used to determine the overall topological morphology of the drug enzyme ligand in the binding pocket.

[0028] Furthermore, the fusion output module uses a weighted scoring method to fuse the confidence score and the rule score, and outputs the risk prediction result.

[0029] The calculation formula for the fusion process is: ;

[0030] in, This is a preset sensitivity threshold; The preset rule threshold.

[0031] A method for predicting the risk of drug enzyme inhibition based on structural characterization includes the following steps:

[0032] Input receiving step: The input receiving module receives the raw data of the drug-enzyme complex to be evaluated and generates an input file;

[0033] Feature construction steps: The feature construction module receives the input file and performs feature processing on the input file to generate feature vectors and atomic matrices for structure prediction;

[0034] Structure prediction step: The structure prediction module receives the feature vector and atomic matrix, performs structure prediction calculation on the complex in the input file, and outputs the structure prediction coordinates corresponding to one or more candidate complex conformations;

[0035] Confidence estimation steps: The confidence estimation module receives the predicted coordinates of the structure and calculates and generates a confidence score based on the conformational consistency and distribution dispersion of the candidate complex.

[0036] Conformation screening and sorting steps: Summarize and sort the confidence scores corresponding to the multiple candidate complex conformations to obtain the sorted coordinates and scores;

[0037] Rule-based evidence scoring steps: The rule-based evidence module receives the processed coordinates, analyzes the conformation of the candidate complex based on preset rules, and generates rule-based evidence directly related to the conformation and the corresponding rule score.

[0038] Fusion assessment and output steps: The fusion output module receives the confidence score and rule scoring results, performs fusion processing on the two, and outputs the risk prediction results.

[0039] Furthermore, the structure prediction step includes the following sub-steps:

[0040] Feature embedding: The feature vectors generated in the feature construction step are input into the embedding layer, and initial pairwise features are generated through relative position encoding. ) and monomer characteristics ( );

[0041] Main trunk evolution: the above paired features ( ) and monomer characteristics ( The input is fed into a pairwise attention mechanism module containing L stacked layers. Each layer interacts with information through triangular updates and self-attention mechanisms, and outputs the updated feature representation.

[0042] Diffusion generation: A diffusion denoising strategy is adopted; the updated feature representation of the backbone evolution output is received by the diffusion module as a condition, starting from the Gaussian noise distribution, and after T-step denoising sampling, the three-dimensional coordinate matrix of the candidate conformation is output.

[0043] Furthermore, the confidence estimation step includes the following sub-steps:

[0044] Geometric feedback feature construction: used to receive the predicted coordinates of the structure, extract representative atoms, and calculate the Euclidean distance matrix between all atom pairs; after encoding the distance matrix with a one-hot code or mapping it with a radial basis function, it is superimposed back into the pairwise representation of the backbone network and refined by an additional pairwise attention module layer;

[0045] Local confidence score calculation: The pLDDT score for each residue is output by the prediction head; this score is calculated using a formula defined as the statistical bias at different distance thresholds.

[0046] in, It is an indicator function; To predict the distances between local neighborhood atom pairs in the structure; This represents the distance between local neighboring atomic pairs in the real structure; This is the distance error tolerance threshold;

[0047] Distance distribution prediction based on hybrid density networks: A hybrid density network head is introduced to model the distance between atomic pairs as a Gaussian mixture distribution. The log-likelihood probability of the predicted distance under this Gaussian mixture distribution is calculated as a high-level confidence index of the interaction, and a confidence score is output. The formula for calculating the Gaussian mixture distribution is as follows:

[0048]

[0049] in, This represents the total number of components in the Gaussian mixture model. This is the index variable for the current component; It is a Gaussian probability density function; Mixed weights; The mean, Let Variance be the variance.

[0050] Furthermore, the rule evidence in the rule evidence scoring step includes evidence A, evidence B, and evidence C;

[0051] Evidence A is used to determine the rigid geometric constraint between the drug enzyme ligand and the prosthetic group; specifically, it includes metal coordination determination, axial blocking at the sixth coordination site, and geometric matching of the aromatic ring in the parallel plane of hemoglobin; among which, the specific process of metal coordination determination is: calculating the distance between the nitrogen or sulfur atom in the drug enzyme ligand and the iron atom at the center of the protein. ;like If determined to be strong coordination; It was determined to be a non-coordinating nearest neighbor;

[0052] Evidence B is used to determine the contact density and interaction type between the drug enzyme ligand and a specific region of the protein pocket;

[0053] Evidence C is used to determine the overall topological morphology of the drug enzyme ligand in the binding pocket.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] (a) Improve prediction stability and generalization ability;

[0056] Compared to existing methods based on single structures or single-point predictions, this invention utilizes a structure-characterization-driven complex modeling approach to generate multiple candidate conformations for the same complex. These candidate conformations are then ranked and comprehensively evaluated, thereby reducing the impact of random errors from single conformations on the prediction results. Furthermore, by introducing rule-based evidence directly related to the complex conformations as auxiliary signals, the system maintains relatively stable prediction performance even when faced with novel compounds or samples with shifted distributions, thus improving overall generalization ability.

[0057] (ii) Provide confidence and credibility information that can be used for decision-making;

[0058] This invention, while outputting risk prediction results, further outputs confidence or credibility information corresponding to the prediction results, so that the prediction results are no longer limited to a single numerical value or classification label. Through the confidence information, users can classify the prediction results according to their reliability, for example, directly accepting high-credibility results and initiating manual review or supplementary experiments for low-credibility results, thereby improving the controllability of the risk control and decision-making process.

[0059] (iii) Enhance the interpretability and auditability of forecast results;

[0060] This invention outputs evidence directly related to the complex conformation and corresponding sub-scoring results through a rule-based evidence module, providing a clear source for the final risk prediction result. The rule-based evidence is output in structured field format, corresponding one-to-one with the final prediction result, facilitating subsequent review, tracing, and interpretation, thereby improving the problems of opaque and difficult-to-audit prediction processes in existing technologies.

[0061] (iv) Improve applicability to special complex systems;

[0062] For complex systems containing prosthetic groups or metal coordination environments, this invention performs adaptation processing on relevant atom sets during confidence assessment and rule analysis, reducing the probability of outliers or unstable outputs in special systems. Therefore, this invention can maintain relatively consistent prediction behavior even under complex structural conditions, improving the system's applicability in practical applications.

[0063] (v) Reduce experimental costs and improve screening efficiency;

[0064] Compared to traditional methods that rely on in vitro inhibition or kinetic experiments, this invention automates the reasoning process using computer software, enabling batch processing of multiple candidate samples and significantly shortening the evaluation cycle. Furthermore, the output results are standardized and stored in the form of structured files and corresponding confidence and evidence files, facilitating direct integration into downstream screening, report generation, and auditing processes, reducing overall implementation costs and improving ease of use. Attached Figure Description

[0065] Figure 1 The diagram below shows a structural block diagram of a drug enzyme inhibition risk prediction system based on structural characterization, as described in Specific Implementation 1.

[0066] Figure 2 The flowchart of a drug enzyme inhibition risk prediction method based on structural characterization as described in Specific Implementation Method Seven;

[0067] Figure 3 This is a scoring flowchart for rule-based evidence in Implementation Method Ten. Detailed Implementation

[0068] Specific Implementation Method 1: Combination Figure 1 This embodiment describes a drug enzyme inhibition risk prediction system based on structural characterization, which includes an input receiving module 1, a feature construction module 2, a structure prediction module 3, a confidence estimation module 4, a result export module 5, a rule evidence scoring module 6, and a fusion output module 7.

[0069] Input receiving module 1 is used to receive the raw data of the drug-enzyme complex to be evaluated and generate an input file;

[0070] Feature construction module 2 is used to perform feature processing on the input file to generate feature vectors and atomic matrices;

[0071] The structure prediction module 3 is used to receive feature vectors and atomic matrices, and performs structure prediction calculations using a generative architecture that includes a pairwise attention mechanism backbone and a diffusion model, and outputs the structure prediction coordinates corresponding to one or more candidate complex conformations.

[0072] The confidence estimation module 4 is used to receive the predicted coordinates of the structure and calculate and generate a confidence score based on the conformational consistency and distribution dispersion of the candidate complex.

[0073] The result export module 5 is used to receive the predicted structural coordinates and confidence scores, summarize and sort the confidence scores corresponding to the multiple candidate complex conformations, and output the sorted coordinates and scores.

[0074] The rule evidence scoring module 6 is used to receive the sorted coordinates and scores, and analyze the conformation of the candidate complex based on preset geometric rules to generate rule evidence and corresponding rule scores directly related to the conformation.

[0075] The fusion output module 7 is used to fuse the confidence score and rule score to output the risk prediction result.

[0076] In this embodiment, a drug enzyme inhibition risk prediction system based on structural characterization can output both structural prediction results and confidence / credibility scores and interpretable evidence related to risk assessment. This embodiment is implemented using a computer device, which includes at least a processor, a memory, and a communication interface. The memory stores computer program instructions, and the processor, when executing these instructions, implements the inhibition risk prediction and confidence assessment method described in this embodiment. Each functional module of the system runs on the processor via program instructions and interacts with data through the memory. This embodiment also provides a computer-readable storage medium storing computer program instructions. When executed by the processor, these instructions implement any of the inhibition risk prediction and confidence assessment methods described below. Each module is stored in the memory as computer program instructions and is sequentially invoked and executed on the processor. Data between modules is transferred via memory data structures or files, including structural tensors, atomic arrays, confidence matrices, and rule-based evidence data objects. The confidence estimation module 4 does not directly judge the risk outcome, but calculates the consistency and distribution dispersion between conformations based on the conformation set output by the structure prediction module 3, thereby generating confidence parameters for risk fusion. In this embodiment, based on the output of the structure prediction module 3, a confidence estimation module 4 is introduced to analyze and process the structure prediction coordinates. In addition to outputting the corresponding prediction results, it further outputs confidence parameters to characterize the reliability of the prediction results. Preferably, the confidence parameters include statistics based on probability distributions to quantify the uncertainty of the prediction results, thereby providing a reliable basis for subsequent risk assessment that can be used for decision-making.

[0077] This embodiment describes a robust distance and masking mechanism for a drug-enzyme inhibition risk prediction system based on structural characterization, oriented towards prosthetic groups and metal environments. For complex systems containing prosthetic groups or metal coordination environments, this invention differentiates specific sets of atoms associated with prosthetic groups or metals during confidence estimation and rule analysis. By adapting the relevant distance calculation and masking strategies, the impact of special systems on the stability of confidence assessment is reduced, improving the system's applicability and reliability under complex structural conditions. Special processing logic for systems containing metal prosthetic groups (such as hemoglobin): Existing technologies often simply classify prosthetic groups as ligands or proteins, leading to confusion in distance supervision. This embodiment proposes an atomic-level identity decoupling masking strategy:

[0078] Identification: The auxiliary base is identified by a feature extractor, and its atomic index is obtained.

[0079] Dynamic mask rewriting: For metal atoms (such as iron atoms) at the prosthetic group center, the system forces the mask to be rewritten as "protein type", so that it can participate in the distance constraint calculation of the protein backbone.

[0080] For other atoms in the prosthetic group (such as carbon and nitrogen atoms on the porphyrin ring), their masks are forced to zero when calculating the ligand-protein interaction distance.

[0081] Results: This strategy eliminates the interference of the rigid structure inside the prosthetic group on the prediction of the flexibility of the binding pocket, allowing the model to focus on predicting the interaction between the "real small molecule ligand" and the "protein pocket containing the metal center".

[0082] The drug-enzyme inhibition risk prediction system based on structure characterization described in this embodiment has an interpretable rule-based evidence generation mechanism that is directly related to the complex conformation. A rule-based evidence scoring module 6 is set up to analyze the complex conformation output by the structure prediction module 3 and extract only the geometric or spatial evidence directly related to the conformation. The evidence includes, but is not limited to, atomic neighborhood distance, pocket contact relationship and spatial distribution characteristics, and is output in the form of structured evidence fields, so that the rule scoring results are traceable and auditable, which facilitates the description of the embodiment and subsequent interpretation and analysis.

[0083] This embodiment of the drug enzyme inhibition risk prediction system based on structural characterization employs a fusion-based risk decision-making mechanism of candidate conformation ranking. For the multiple candidate conformations output by the structural prediction module, this invention ranks the candidate conformations and generates reference indicators for risk assessment based on the ranking results. Furthermore, the confidence parameter is fused with the scoring results corresponding to rule-based evidence to output the final risk prediction result, ensuring that the prediction conclusion possesses both data-driven stability and rule-based interpretability.

[0084] Specific Implementation Method Two: This implementation method further defines the drug-enzyme inhibition risk prediction system based on structural characterization described in Specific Implementation Method One. In this implementation method, the original input data includes the structural file or multiple sequence alignment information of the drug-enzyme complex to be evaluated.

[0085] In this embodiment, the original input data includes complex structure files (PDB, CIF) or equivalent descriptions, which are converted into input files in a uniform format; optionally, multiple sequence alignment (MSA) information is generated or received and written into the input file.

[0086] Specific Implementation Method 3: This implementation method further defines the drug enzyme inhibition risk prediction system based on structural characterization described in Specific Implementation Method 2. In this implementation method, the structural prediction module 3 includes a feature embedding unit, a backbone evolution unit, and a diffusion generation unit.

[0087] The feature embedding unit is used to encode the feature vector through an embedding layer and relative positions to generate initial paired features and single features; The backbone evolutionary unit contains L stacked pairs of attention mechanism modules. Each layer interacts with information through triangular updates and self-attention mechanisms to output updated feature representations.

[0088] The diffusion generation unit is used to receive the updated feature representation output by the backbone evolution unit as a condition, start from Gaussian noise distribution, and output the three-dimensional coordinate matrix of the candidate conformation after T-step denoising sampling; the three-dimensional coordinate matrix of the candidate conformation is the structure prediction coordinate. Where L is an integer greater than 1 and T is an integer greater than 1.

[0089] In this embodiment, compared with existing methods based on single structure or single point prediction, the generative architecture that couples a pairwise attention mechanism backbone with a diffusion model improves prediction stability and generalization ability. This allows for the generation of multiple candidate conformations for the same complex, and subsequent evaluation is performed based on the candidate conformation set. This reduces the impact of random errors in a single conformation on the prediction results and improves the overall prediction stability. Furthermore, the deep learning architecture enables automated structure prediction calculation, which significantly shortens the evaluation cycle and reduces the overall implementation cost compared to traditional methods that rely on in vitro inhibition experiments.

[0090] Specific Implementation Method 4: This implementation method further defines the drug enzyme inhibition risk prediction system based on structural characterization described in Specific Implementation Method 3. In this implementation method, the confidence estimation module 4 includes a geometric feedback feature construction unit, a local confidence calculation unit, and a hybrid density network unit.

[0091] The geometric feedback feature construction unit is used to receive the predicted coordinates of the structure, extract the representative atoms, and calculate the Euclidean distance matrix between all atom pairs. After encoding the distance matrix with a one-hot code or mapping with a radial basis function, it is superimposed back into the pair representation of the pair attention mechanism backbone. The feature is refined through an additional pair attention module layer.

[0092] The local confidence calculation unit outputs the pLDDT score for each residue through the prediction head, and the pLDDT score is defined as the deviation statistics at different distance thresholds;

[0093] A hybrid density network unit is used to model the distance between atomic pairs as a Gaussian mixture distribution. The log-likelihood probability of the predicted distance under the Gaussian mixture distribution is used as a high-level confidence index of the interaction, and a confidence score is output.

[0094] In this embodiment, the local confidence calculation unit outputs the pLDDT score (defined as the deviation statistics under different distance thresholds) for each residue, enabling users to classify the prediction results according to their reliability. For example, high-confidence results can be directly accepted, while low-confidence results can be subject to manual review or supplementary experiments. The mixed density network unit models the atom pair distances as a Gaussian mixture distribution, and calculates the log-likelihood probability of the predicted distance under the Gaussian mixture distribution as a high-level confidence index of the interaction, thereby providing a reliable basis for subsequent risk assessment. The geometric feedback feature construction unit innovatively introduces a prediction geometric feedback mechanism, feeding back the predicted structural coordinates to the feature extraction process. By calculating the Euclidean distance matrix between all atom pairs and superimposing it back into the backbone network, a closed loop is formed between confidence assessment and structural prediction, improving the accuracy of confidence estimation.

[0095] Specific Implementation Method 5: This implementation method further defines the four drug enzyme inhibition risk prediction systems based on structural characterization described in Specific Implementation Method 3. In this implementation method, the rule evidence scoring module 6 includes a geometric distance determination unit, a contact and interaction determination unit, and a spatial distribution and span determination unit.

[0096] Geometric distance determination unit, used to determine the rigid geometric constraint between drug enzyme ligand and prosthetic group;

[0097] The contact and interaction determination unit is used to determine the contact density and interaction type between the drug enzyme ligand and a specific region of the protein pocket; The spatial distribution and span determination unit is used to determine the overall topological morphology of the drug enzyme ligand in the binding pocket.

[0098] In this embodiment, the interpretability and auditability of the prediction results are enhanced. By setting up a rule-based evidence scoring module 6, the complex conformation output by the structure prediction module is analyzed, and only geometric or spatial evidence directly related to the conformation is extracted (including but not limited to atomic neighborhood distance, pocket contact relationship, and spatial distribution characteristics). Evidence is traceable: it is output in the form of structured evidence fields, giving the rule scoring results a clear source basis. Results are auditable: the rule evidence and rule scoring results are used as inputs to the fusion output module, together with the confidence index, to generate the final risk prediction result and corresponding credibility and interpretation information. Verification is facilitated: the rule evidence is output in the form of structured fields, which can correspond one-to-one with the final prediction result, facilitating subsequent verification, tracing, and interpretation analysis, thereby improving the problems of opaque and difficult-to-audit prediction processes in existing technologies.

[0099] Improve the applicability to special complex systems; for complex systems containing prosthetic groups or metal coordination environments: determine the rigid geometric constraints between drug enzyme ligands and prosthetic groups through geometric distance determination units; combine contact and interaction determination units and spatial distribution and span determination units to differentiate specific sets of atoms related to prosthetic groups or metals; reduce the probability of outliers or unstable outputs in special systems, and improve the applicability and reliability of the system under complex structural conditions.

[0100] It provides complementary evaluation dimensions to supplement data-driven prediction results: Rule-based interpretability: It integrates data-driven confidence parameters with hard rules based on domain knowledge (geometric distance, contact density, etc.) to ensure that predictions possess both data-driven stability and rule-based interpretability. Risk control basis: By identifying potential "geometric taboos" (such as being too close or occupying a critical pocket) through rule-based evidence, it significantly increases the risk score regardless of model confidence, providing additional risk control basis for downstream decision-making.

[0101] Specific Implementation Method Six: This implementation method further defines the drug enzyme inhibition risk prediction system based on structural characterization described in Specific Implementation Method Three. In this implementation method, the fusion output module 7 uses a weighted scoring method to fuse the confidence score and the rule score, and outputs the risk prediction result.

[0102] The calculation formula for the fusion process is: ;

[0103] in, This is a preset sensitivity threshold; The preset rule threshold.

[0104] In this embodiment, the preset sensitivity threshold This represents the system's sensitivity to the risk factor that "data-driven predictions are unreliable." (1-Confidence) This term reflects how "uncertain" the model is about the structure it predicts. The larger the score, the more conservative the system; if the model is unsure about the structure prediction (low confidence), the system tends to classify the sample as high-risk (or requires a high score to trigger an alert / manual review); preset rule thresholds. This represents the system's emphasis on "hard rules based on domain knowledge". The larger the value, the more the system relies on "expert rules"; that is, as long as certain geometric taboos are violated (such as being too close or occupying a key pocket), the risk score will increase significantly regardless of the model confidence.

[0105] Detailed Implementation Method Seven: Combination Figure 2 This embodiment describes a drug enzyme inhibition risk prediction method based on structural characterization, which is implemented based on a drug enzyme inhibition risk prediction system based on structural characterization as described in Specific Embodiment Five. The prediction method includes the following steps:

[0106] Input receiving steps: Input receiving module 1 receives the raw data of the drug-enzyme complex to be evaluated and generates an input file;

[0107] Feature construction steps: The feature construction module 2 receives the input file, performs feature processing on the input file, generates feature vectors and atomic matrices for structure prediction, and records the corresponding dimension information;

[0108] Structure prediction step: The structure prediction module 3 receives the feature vector and atomic matrix, performs structure prediction calculation on the complex in the input file, and outputs the structure prediction coordinates corresponding to one or more candidate complex conformations;

[0109] Confidence estimation step: The confidence estimation module 4 receives the predicted coordinates of the structure and calculates and generates a confidence score based on the conformational consistency and distribution dispersion of the candidate complex.

[0110] Conformation screening and sorting steps: Summarize and sort the confidence scores corresponding to the multiple candidate complex conformations to obtain the sorted coordinates and scores;

[0111] Rule-based evidence scoring steps: The rule-based evidence module 6 receives the sorted coordinates, analyzes the conformation of the candidate complex based on preset rules, and generates rule-based evidence and corresponding rule scores directly related to the conformation.

[0112] Fusion assessment and output steps: The fusion output module 7 receives the confidence score and rule scoring results, performs fusion processing on the two, and outputs the risk prediction results.

[0113] In this embodiment, the original input data includes complex structure files (PDB, CIF) or equivalent descriptions, which are converted into input files in a unified format. Optionally, multiple sequence alignment (MSA) information is generated or received and written into the input file. In the structure prediction step, instead of a simple end-to-end regression, a generative architecture based on the coupling of a pairwise attention mechanism module backbone and a diffusion model is adopted. In the conformation screening and ranking step, the confidence estimation module 4 not only relies on feature vectors but also innovatively introduces a "predictive geometric feedback" mechanism. In the conformation screening and ranking step, the confidence estimation module 4 summarizes the confidence scores of multiple candidate conformations and generates ranking results according to preset rules to obtain stable confidence outputs for subsequent evaluation. The result export module 5 saves the predicted structure coordinates and corresponding confidence scores as structure files and confidence files for subsequent rule analysis or result output.

[0114] Specific Implementation Method Eight: This implementation method further defines the four drug enzyme inhibition risk prediction methods based on structure characterization described in Specific Implementation Method Seven. In this implementation method, the structure prediction step includes the following sub-steps:

[0115] Feature embedding: The feature vectors generated in the feature construction step are input into the embedding layer, and initial pairwise features are generated through relative position encoding. ) and monomer characteristics ( );

[0116] Main trunk evolution: the above paired features ( ) and monomer characteristics ( The input is fed into a pairwise attention mechanism module containing L stacked layers. Each layer interacts with information through triangular updates and self-attention mechanisms, and outputs the updated feature representation.

[0117] Diffusion generation: A diffusion denoising strategy is adopted; the updated feature representation of the backbone evolution output is received by the diffusion module as a condition, starting from the Gaussian noise distribution, and after T-step denoising sampling, the three-dimensional coordinate matrix of the candidate conformation is output.

[0118] In this embodiment, multiple sub-parallelisms are also included: in order to capture the uncertainty of prediction, the system sets different random seeds during inference and executes the above diffusion process in parallel to generate conformational clusters with different structures, providing a sample basis for subsequent confidence estimation.

[0119] Specific Implementation Method Nine: This implementation method further defines the four drug enzyme inhibition risk prediction methods based on structural characterization described in Specific Implementation Method Eight. In this implementation method, the confidence estimation step includes the following sub-steps:

[0120] Geometric feedback feature construction: used to receive the predicted coordinates of the structure, extract representative atoms, and calculate the Euclidean distance matrix between all atom pairs; after encoding the distance matrix with a one-hot code or mapping it with a radial basis function, it is superimposed back into the pairwise representation of the backbone network and refined by an additional pairwise attention module layer;

[0121] Local confidence score calculation: The pLDDT score for each residue is output by the prediction head; this score is calculated using a formula defined as the statistical bias at different distance thresholds.

[0122] in, It is an indicator function; To predict the distances between local neighborhood atom pairs in the structure; This represents the distance between local neighboring atomic pairs in the real structure; This represents the distance error tolerance threshold; 0.5, 1, 2, and 4 are four specific tolerance levels.

[0123] Distance distribution prediction based on hybrid density networks: A hybrid density network head is introduced to model the distance between atomic pairs as a Gaussian mixture distribution. The log-likelihood probability of the predicted distance under this Gaussian mixture distribution is calculated as a high-level confidence index of the interaction, and a confidence score is output. The formula for calculating the Gaussian mixture distribution is as follows:

[0124]

[0125] in, This represents the total number of components in the Gaussian mixture model. This is the index variable for the current component; It is a Gaussian probability density function; Mixed weights; The mean, Let Variance be the variance.

[0126] In this embodiment, This represents the total number of components in the Gaussian mixture model; it signifies the total number of "sub-distributions" the model uses to fit the complex distribution of inter-atomic distances. Physically, it reflects the richness of the different steady-state or conformational modes that may exist for the atomic pair distances (e.g., inter-atomic distances may be concentrated around a few different specific values). This is the index variable for the current component; it is a variable ranging from 1 to... The changing count variable is used to identify which specific Gaussian distribution component is currently being calculated; in the summation formula, it is used to iterate through each independent subdistribution and accumulate their contributions to the final probability density. Let be a Gaussian probability density function; this describes the first... A mathematical function of the shape of each component (i.e., a bell curve); it is based on the mean of that component. and variance Calculate at a specific distance The probability density value is used to quantify the likelihood that the distance belongs to that specific conformational mode.

[0127] Detailed Implementation Method 10: Combination Figure 3 This embodiment further defines the four drug enzyme inhibition risk prediction methods based on structural characterization described in Specific Embodiment Nine. In this embodiment, the rule evidence in the rule evidence scoring step includes evidence A, evidence B, and evidence C.

[0128] Evidence A is used to determine the rigid geometric constraint between the drug enzyme ligand and the prosthetic group; specifically, it includes metal coordination determination, axial blocking at the sixth coordination site, and geometric matching of the aromatic ring in the parallel plane of hemoglobin; among which, the specific process of metal coordination determination is: calculating the distance between the nitrogen or sulfur atom in the drug enzyme ligand and the iron atom at the center of the protein. ;like If determined to be strong coordination; It was determined to be a non-coordinating nearest neighbor;

[0129] Evidence B is used to determine the contact density and interaction type between the drug enzyme ligand and a specific region of the protein pocket;

[0130] Evidence C is used to determine the overall topological morphology of the drug enzyme ligand in the binding pocket.

[0131] In this embodiment, Evidence A: Geometric distance-related evidence. This type of evidence focuses on the rigid geometric constraints between the ligand and the prosthetic group (such as the iron atom center of hemoglobin), specifically including:

[0132] Metal coordination determination: Calculate the minimum distance between the nitrogen / sulfur (N / S) heavy atoms in the ligand and the central metal (such as an iron atom, Fe). .like If strong coordination is not formed, but the minimum distance between the coordinating heavy atom and Fe is [not specified], then [the relationship is considered strong coordination]. They were determined to be non-coordinated nearest neighbors. Both of these cases are supported by positive evidence.

[0133] Axial blocking at the sixth coordination site: When strong coordination is determined, further testing is conducted to determine whether the ligand blocks the sixth coordination direction of the Fe atom. The direction in which Fe points towards the coordinating atom is defined as the axis, and a dynamic conical / cylindrical envelope (radius 2.5~3.0) is constructed within the axial projection distance range of 1.8~4.8. If a non-coordinating atom of the ligand falls into this envelope region, it is determined that axial blocking exists, which is positive evidence.

[0134] Hemoglobin parallel-plane aromatic ring geometric matching: Detecting the geometric relationship between the ligand aromatic ring and the plane. Calculating the angle between the ligand aromatic ring normal and the plane normal, and the distance between their centroids. If the "angle" is satisfied... If the plane is parallel / antiparallel and the planar distance is between 3.0 and 5.0, and it forms contact with specific residues above the plane (such as Ile369 / Ala370, etc.), then it is considered a valid planar stacking, which is positive evidence.

[0135] Evidence B: Contact and interaction-related evidence; this type of evidence focuses on the contact density and interaction type between the ligand and specific regions of the protein pocket, specifically including:

[0136] Deep pocket multi-point contact: statistical analysis of the contact between ligands and key residue sets deep within the pocket (such as Phe108, Leu210, etc.) (distance). ); if the number of key residues contacted If there are one, then it is positive evidence.

[0137] Large-volume tail embedding: Statistical analysis of the contact between the ligand carbon backbone and the hydrophobic pocket residue set (e.g., Tyr53, Phe57). If the number of contact residues... If the large volume of the tail is effectively embedded, it provides positive evidence.

[0138] Global hydrophobicity and Depositional effects statistics: Calculate the number of hydrophobic contact pairs across the entire system and Quantity of deposits. Based on historical data and statistical distribution, if the number of hydrophobic contact pairs... or Stacked Pairs This provides positive evidence.

[0139] Evidence C: Evidence related to spatial distribution and span. This type of evidence focuses on the overall topological morphology of ligands within the binding pocket, specifically including: Multi-pocket span: Identify all ligand atoms that come into contact with the protein and calculate the maximum pairwise distance between these contacting atoms in three-dimensional space. If the maximum span... This indicates that the ligand has adopted an extended conformation spanning multiple sub-pockets, providing positive evidence. The rule-based evidence scoring module 6 takes the predicted complex structure file as input, extracts rule-based evidence directly related to the complex conformation, including evidence related to geometric distance, contact and interaction, and spatial distribution and span. Based on this rule-based evidence, it generates corresponding rule sub-scores and a total rule score. The rule-based evidence and rule scoring results serve as input to the fusion output module 7, working together with the credibility index to generate the final risk prediction result and corresponding credibility and explanatory information.

[0140] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A drug enzyme inhibition risk prediction system based on structural characterization, characterized in that, It includes an input receiving module (1), a feature construction module (2), a structure prediction module (3), a confidence estimation module (4), a result export module (5), a rule evidence scoring module (6), and a fusion output module (7). The input receiving module (1) is used to receive the raw data of the drug-enzyme complex to be evaluated and generate an input file; The feature construction module (2) is used to perform feature processing on the input file to generate feature vectors and atomic matrices; The structure prediction module (3) is used to receive feature vectors and atomic matrices, and to perform structure prediction calculations using a generative architecture that includes a pairwise attention mechanism backbone and a diffusion model, and output the structure prediction coordinates corresponding to one or more candidate complex conformations. The confidence estimation module (4) is used to receive the predicted coordinates of the structure and calculate and generate a confidence score based on the conformational consistency and distribution dispersion of the candidate complex. The result export module (5) is used to receive the predicted structural coordinates and confidence scores, summarize and sort the confidence scores corresponding to the multiple candidate complex conformations, and output the sorted coordinates and scores. The rule evidence scoring module (6) is used to receive the sorted coordinates and scores, and analyze the conformation of the candidate complex based on preset geometric rules to generate rule evidence and corresponding rule scores directly related to the conformation. The fusion output module (7) is used to fuse the confidence score and rule score to output the risk prediction result.

2. The drug enzyme inhibition risk prediction system based on structural characterization according to claim 1, characterized in that, The raw input data includes the structural file or multiple sequence alignment information of the drug-enzyme complex to be evaluated.

3. The drug enzyme inhibition risk prediction system based on structural characterization according to claim 2, characterized in that, The structure prediction module (3) includes a feature embedding unit, a backbone evolution unit, and a diffusion generation unit; The feature embedding unit is used to encode the feature vector through an embedding layer and relative positions to generate initial paired features and single features; The backbone evolutionary unit contains L stacked pairs of attention mechanism modules. Each layer interacts with information through triangular updates and self-attention mechanisms to output updated feature representations. The diffusion generation unit is used to receive the updated feature representation output by the backbone evolution unit as a condition, start from Gaussian noise distribution, and output the three-dimensional coordinate matrix of the candidate conformation after T-step denoising sampling; the three-dimensional coordinate matrix of the candidate conformation is the structure prediction coordinate. Where L is an integer greater than 1 and T is an integer greater than 1.

4. The drug enzyme inhibition risk prediction system based on structural characterization according to claim 3, characterized in that, The confidence estimation module (4) includes a geometric feedback feature construction unit, a local confidence calculation unit, and a hybrid density network unit; The geometric feedback feature construction unit is used to receive the predicted coordinates of the structure, extract the representative atoms, and calculate the Euclidean distance matrix between all atom pairs. After encoding the distance matrix with a one-hot code or mapping with a radial basis function, it is superimposed back into the pair representation of the pair attention mechanism backbone. The feature is refined through an additional pair attention module layer. The local confidence calculation unit outputs the pLDDT score for each residue through the prediction head, and the pLDDT score is defined as the deviation statistics at different distance thresholds; A hybrid density network unit is used to model the distance between atomic pairs as a Gaussian mixture distribution. The log-likelihood probability of the predicted distance under the Gaussian mixture distribution is used as a high-level confidence index of the interaction, and a confidence score is output.

5. The drug enzyme inhibition risk prediction system based on structural characterization according to claim 4, characterized in that, The rule-based evidence scoring module (6) includes a geometric distance determination unit, a contact and interaction determination unit, and a spatial distribution and span determination unit; Geometric distance determination unit, used to determine the rigid geometric constraint between drug enzyme ligand and prosthetic group; The contact and interaction determination unit is used to determine the contact density and interaction type between the drug enzyme ligand and a specific region of the protein pocket; The spatial distribution and span determination unit is used to determine the overall topological morphology of the drug enzyme ligand in the binding pocket.

6. The drug enzyme inhibition risk prediction system based on structural characterization according to claim 1, characterized in that, The fusion output module (7) uses a weighted scoring method to fuse the confidence score and the rule score, and outputs the risk prediction result. The calculation formula for the fusion process is: ; in, This is a preset sensitivity threshold; The preset rule threshold.

7. A method for predicting the risk of drug enzyme inhibition based on structural characterization, characterized in that, The prediction method is based on the drug enzyme inhibition risk prediction system based on structural characterization as described in claim 5, and the prediction method includes the following steps: Input receiving steps: The input receiving module (1) receives the raw data of the drug-enzyme complex to be evaluated and generates an input file; Feature construction steps: The feature construction module (2) receives the input file and performs feature processing on the input file to generate feature vectors and atomic matrices for structure prediction; Structure prediction steps: The structure prediction module (3) receives the feature vector and atomic matrix, performs structure prediction calculation on the complex in the input file, and outputs the structure prediction coordinates corresponding to one or more candidate complex conformations; Confidence estimation steps: The confidence estimation module (4) receives the predicted coordinates of the structure and calculates the confidence score based on the conformational consistency and distribution dispersion of the candidate complex. Conformation screening and sorting steps: Summarize and sort the confidence scores corresponding to the multiple candidate complex conformations to obtain the sorted coordinates and scores; Rule-based evidence scoring steps: The rule-based evidence module (6) receives the sorted coordinates, analyzes the conformation of the candidate complex based on preset rules, and generates rule-based evidence and corresponding rule scores directly related to the conformation. Fusion assessment and output steps: The fusion output module (7) receives the confidence score and rule scoring results, performs fusion processing on the two, and outputs the risk prediction results.

8. The method for predicting drug enzyme inhibition risk based on structural characterization according to claim 7, characterized in that, The structural prediction step includes the following sub-steps: Feature embedding: The feature vectors generated in the feature construction step are input into the embedding layer, and initial pairwise features are generated through relative position encoding. ) and monomer characteristics ( ); Main trunk evolution: the above paired features ( ) and monomer characteristics ( The input is fed into a pairwise attention mechanism module containing L stacked layers. Each layer interacts with information through triangular updates and self-attention mechanisms, and outputs the updated feature representation. Diffusion generation: A diffusion denoising strategy is adopted; The updated feature representations received from the backbone evolution output by the diffusion module are used as conditions. Starting from the Gaussian noise distribution, after T-step denoising sampling, the three-dimensional coordinate matrix of the candidate conformation is output.

9. The method for predicting drug enzyme inhibition risk based on structural characterization according to claim 8, characterized in that, The confidence estimation step includes the following sub-steps: Geometric feedback feature construction: used to receive the predicted coordinates of the structure, extract representative atoms, and calculate the Euclidean distance matrix between all atom pairs; after encoding the distance matrix with a one-hot code or mapping it with a radial basis function, it is superimposed back into the pairwise representation of the backbone network and refined by an additional pairwise attention module layer; Local confidence score calculation: The pLDDT score for each residue is output by the prediction head; The score is calculated using a formula defined as a statistical deviation at different distance thresholds: in, It is an indicator function; To predict the distances between local neighborhood atom pairs in the structure; This represents the distance between local neighboring atomic pairs in the real structure; This is the distance error tolerance threshold; Distance distribution prediction based on hybrid density networks: A hybrid density network head is introduced to model the distance between atomic pairs as a Gaussian mixture distribution. The log-likelihood probability of the predicted distance under this Gaussian mixture distribution is calculated as a high-level confidence index of the interaction, and a confidence score is output. The formula for calculating the Gaussian mixture distribution is as follows: in, This represents the total number of components in the Gaussian mixture model. This is the index variable for the current component; It is a Gaussian probability density function; Mixed weights; The mean, Let Variance be the variance.

10. The method for predicting drug enzyme inhibition risk based on structural characterization according to claim 9, characterized in that, The rule-based evidence in the rule-based evidence scoring step includes evidence A, evidence B, and evidence C; Evidence A is used to determine the rigid geometric constraint between the drug enzyme ligand and the prosthetic group; specifically, it includes metal coordination determination, axial blocking at the sixth coordination site, and geometric matching of the aromatic ring in the parallel plane of hemoglobin; among which, the specific process of metal coordination determination is: calculating the distance between the nitrogen or sulfur atom in the drug enzyme ligand and the iron atom at the center of the protein. ;like If determined to be strong coordination; It was determined to be a non-coordinating nearest neighbor; Evidence B is used to determine the contact density and interaction type between the drug enzyme ligand and a specific region of the protein pocket; Evidence C is used to determine the overall topological morphology of the drug enzyme ligand in the binding pocket.

Images