ADMET and toxicity integrated agent evaluation method and system

Abstract

The invention provides an ADMET and toxicity integrated agent evaluation method and system, and relates to the technical field of artificial intelligence and biological medicine crossing. The invention aims to solve the problem of screening deviation caused by dispersion of multi-model evaluation results, easy failure of static thresholds and lack of dynamic weights and interpretability in the prior art. According to the evaluation method, molecular structure fingerprints serve as input, structural characterization is extracted through a hierarchical network, and a unified feature space is constructed by fusing multi-source preprocessing data; the weight of each dimension is dynamically adjusted through sliding window statistics, and weighted feature vectors are generated in combination with expert rule constraints; and continuously updating the judgment boundary by using a self-calibration threshold mechanism according to a percentile truncation method, and finally outputting three types of risk levels and auditable reports through a classifier. Through cross-modal fusion, dynamic weighting and federation collaborative calibration mechanisms, systematicness, stability and transparency of early compound screening are remarkably improved, and clinical failure risks are reduced.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and biomedicine, specifically to an integrated ADMET and toxicity proxy assessment method, system, electronic device and storage medium. Background Technology

[0002] Drug-loaded contact lenses (DLCLs) are a novel ocular drug delivery system that combines medication with a corneal contact lens. Their core advantage lies in achieving long-lasting and precise drug release on the corneal surface, avoiding the drawbacks of traditional eye drops such as low bioavailability and the need for frequent administration. As a core system for evaluating drug disposition and safety in vivo, ADMET directly determines the design rationale, efficacy, and clinical safety of drug-loaded contact lenses, serving as a crucial guideline for the entire manufacturing process.

[0003] For ADMET technology, the current industry relies on multiple independent models to evaluate different physiological dimensions, outputting raw probabilities or classification labels, lacking a unified quantitative benchmark. This decentralized validation system leads to inconsistent threshold standards and fragmented evaluations among models. Cross-dimensional comparisons and comprehensive decisions depend on human experience, making it difficult to build automated screening pipelines. In high-throughput scenarios, this increases the risk of misscreening and missed screenings, raising the risk of clinical failure. Furthermore, while existing technologies have some predictive models, they lack multi-model collaborative reasoning design and scenario-based scoring rules. Misaligned confidence intervals and unquantified uncertainty create decision-making dilemmas, and boundary samples lack verification and calibration capabilities, leading to drifting screening standards. Moreover, isolated model outputs, fragmented evaluations, fixed thresholds, and a lack of uncertainty guidance make it difficult to adaptively adjust strategies. The lack of a confidence-driven mechanism hinders the evaluation module from becoming an intelligent agent, restricting the large-scale and intelligent development of computation-driven drug discovery.

[0004] Based on this, this application proposes an ADMET toxicity proxy assessment technology that integrates multi-model integrated evaluation, scenario-adaptive threshold, uncertainty perception and active learning capabilities to solve one or more problems existing in the prior art. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an integrated ADMET and toxicity proxy assessment method and system, which solves the problems of scattered models, rigid thresholds, and inconsistent standards in existing related technologies.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: Firstly, this application proposes an integrated ADMET and toxicity assessment method, the method comprising: Acquire the raw data stream, which includes a molecular structure description file and a raw evaluation data stream; Based on feature mapping, the original data stream is transformed into a unified feature vector space; the feature vector space includes multiple independent feature subspaces. Based on the feature subspace, dynamic weighted calculations are performed on each evaluation dimension of the candidate compound. At the same time, dynamic thresholds are obtained based on the feature subspace, and threshold self-calibration is performed based on the dynamic thresholds to determine the critical decision boundaries of each evaluation dimension of the candidate compound. The scores of each assessment dimension after dynamic weighting are compared with the self-calibration thresholds to generate the final comprehensive risk score.

[0009] In one embodiment, the molecular structure description file further includes: quantum-derived parameters obtained by quantum chemical calculations based on the molecular structure, the quantum-derived parameters including: molecular orbital energy distribution, electrostatic potential surface gradient rate of change, and frontier orbital gap width.

[0010] In one embodiment, the method further includes: performing outlier detection on the raw evaluation data stream; and / or Perform data integrity verification on the original data stream.

[0011] In one embodiment, the raw assessment data stream includes: in vitro experimental measurements, in vivo pharmacokinetic parameters, toxicity test results data, and historical assessment records.

[0012] In one embodiment, the feature subspace includes a physicochemical feature subspace, an in vitro experimental feature subspace, and an in vivo and toxicity feature subspace.

[0013] In one embodiment, transforming the original data stream into a unified feature vector space based on feature mapping includes: The molecular structure description file is subjected to a normalization cleaning process to obtain a normalized structure matrix; The original evaluation data stream is subjected to timestamp alignment and dimension normalization transformation to form a dimensionless data tensor; A deep residual convolutional architecture is used to perform a nonlinear embedding transformation on the normalized structure matrix and the dimensionless data tensor.

[0014] Preferably, the deep residual convolutional architecture includes five levels of stacked convolutional kernels, with kernel sizes of 3x3, 5x5, 7x7, 9x9, and 11x11, and a fixed stride of 1. The padding strategy ensures that the output feature map size remains constant. Each level of convolutional layer is followed by batch normalization and a modified linear activation function. The final output is compressed into a 128-dimensional dense vector by global average pooling as the structural response encoding.

[0015] More preferably, the deep residual convolutional architecture supports online incremental learning.

[0016] In one embodiment, the dynamic weighted calculation of each evaluation dimension of the candidate compound based on the feature subspace includes: Based on the overall distribution characteristics of the current batch of compounds, dynamic weighted integration is performed on each evaluation dimension of the compounds to generate a weighted fusion feature vector, specifically including: The standard deviation and skewness coefficient of each evaluation dimension of the compound were calculated based on the sliding window statistical analysis in the most recently preset number of evaluated compound samples. If the standard deviation and absolute value of skewness meet the preset conditions, the basic weight of the evaluation dimension is assigned; otherwise, it is reduced proportionally. The weights of all evaluation dimensions are normalized by softmax and then multiplied element-wise with the corresponding feature vectors to generate a weighted fusion feature vector.

[0017] In one embodiment, the dynamic weighted calculation process introduces an expert knowledge constraint mechanism to constrain the dynamic weight coefficients within a reasonable range.

[0018] In one embodiment, the threshold self-calibration is performed once at a preset time interval, or it is forcibly triggered when the number of newly added verification data exceeds the number of pre-reviewed entries.

[0019] Preferably, the threshold self-calibration process supports multi-agency collaborative calibration.

[0020] In one embodiment, the formula for calculating the comprehensive risk score is: , Where S represents the comprehensive risk score. For the standardized score of the i-th evaluation dimension, represents the dynamic weight coefficient for the corresponding evaluation dimension, where n is the total number of evaluation dimensions.

[0021] Preferably, the standardized score The calculation method is as follows: If the original score of the evaluation dimension is higher than the self-calibration threshold, then ; If the original score of the evaluation dimension is lower than the self-calibration threshold, then ; in, This is the original score. This is the self-calibration threshold.

[0022] In one embodiment, the overall score output is further equipped with a visualization explanation engine.

[0023] In one embodiment, the method further includes: S5. Achieving parameter coordination and model calibration between different evaluation systems based on a federated learning framework; and / or By utilizing uncertainty quantification, a confidence interval and risk level label are added to the comprehensive risk score.

[0024] Secondly, this application also proposes an integrated ADMET and toxicity assessment system, the system comprising: A data input unit is configured to acquire a raw data stream, which includes a molecular structure description file and a raw evaluation data stream. The feature mapping unit is configured to transform the original data stream into a unified feature vector space based on feature mapping; the feature vector space includes multiple independent feature subspaces. The dynamic weighted calculation unit is configured to perform dynamic weighted calculations on each evaluation dimension of the candidate compound based on the feature subspace. The threshold self-calibration unit is configured to obtain a dynamic threshold based on the feature subspace and perform threshold self-calibration based on the dynamic threshold to determine the critical decision boundary of each evaluation dimension of the candidate compound. The comprehensive scoring output unit is configured to compare the dynamically weighted scores of each assessment dimension with the self-calibration thresholds to generate the final comprehensive risk score.

[0025] The system performs the steps of the method described in any of the above descriptions.

[0026] (III) Beneficial Effects

[0027] This invention provides an integrated proxy assessment method and system for ADMET and toxicity. Compared with the prior art, it has the following advantages: This application proposes an integrated ADMET and toxicity proxy assessment method. By introducing core technologies such as cross-dimensional dynamic weighting mechanisms, multi-source data feature mapping, threshold self-calibration, federated collaboration, and uncertainty quantification, it achieves automated, standardized, and intelligent assessment of the pharmacokinetic and toxicological risks of candidate compounds. This method solves the core problems of fragmented models, rigid thresholds, and inconsistent standards in existing technologies, providing efficient, reliable, and scalable technical support for early screening in drug development, and has significant industrial application value and social benefits. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of an integrated ADMET and toxicity assessment method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the logical process of multi-source data preprocessing and feature space mapping in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the core principle framework of the cross-dimensional dynamic weighting mechanism in this embodiment of the invention; Figure 4 This is a flowchart illustrating the logic of threshold self-calibration and comprehensive score output in an embodiment of the present invention. Figure 5 Federal collaborative calibration in the embodiments of the present invention ( Figure 5 (Left) and uncertainty quantification ( Figure 5 Right) Schematic diagram of dual-track output principle; Figure 6 This is a schematic diagram of the overall architecture of an integrated ADMET and toxicity assessment system according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the multi-level interaction relationships and data flow between internal units in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Drug-loaded contact lenses (DLCLs) are a new type of medical device that combines "vision correction function" and "ocular drug delivery function". The core is to load therapeutic drugs (such as antibiotics, anti-inflammatory drugs, anti-glaucoma eye drops, etc.) onto the matrix or surface of the corneal contact lens. Through direct contact between the lens and the cornea, the drug can be released into the eye in a long-term and precise manner, which solves the pain points of low bioavailability (only 5%-10%) and frequent administration of traditional ocular drug delivery (such as eye drops).

[0032] ADMET is a core concept in the fields of pharmacokinetics (PK) and toxicology. It's an acronym for assessing five key processes of a drug in the body: absorption, distribution, metabolism, excretion, and toxicity. Essentially, it studies the entire process of a drug's journey from entry into the human body to its elimination, and its potential effects on the body during this process. It's an essential evaluation step for drugs to move from the laboratory to clinical trials—only candidate drugs with excellent ADMET characteristics have the potential to demonstrate good safety and efficacy in clinical trials and ultimately be approved for market release.

[0033] As a core system for assessing the in vivo disposal and safety of drugs, ADMET directly determines the design rationality, efficacy, and clinical safety of drug-eluting contact lenses. It serves as a key guideline for the entire manufacturing process of drug-eluting contact lenses (from drug screening and lens material selection to process optimization and clinical translation).

[0034] With the acceleration of drug development processes and the widespread application of computational toxicology, early prediction of ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties has become a core step in lead compound screening. Currently, the industry generally relies on multiple independent models to assess different physiological dimensions, such as hepatic metabolic stability, blood-brain barrier penetration, or cardiotoxicity risk. These assessments are often output as raw probabilities or classification labels, lacking a unified quantitative benchmark. This decentralized validation system leads to inconsistent threshold settings and fragmented scoring scales among different models, making cross-dimensional comparisons and comprehensive decisions highly dependent on human experience, and hindering the construction of stable, reusable automated screening pipelines. Especially in high-throughput virtual screening scenarios, the heterogeneity and lack of confidence in model outputs can easily lead to false negatives or false negatives, significantly reducing development efficiency and increasing the risk of later clinical failures.

[0035] Among them, the core objective of the integrated intelligent evaluation system for the multidimensional ADMET properties of candidate compounds is to achieve standardized fusion of cross-model outputs and dynamic threshold adaptation. Although existing technologies have built some toxicity prediction models or metabolic stability classifiers, they generally lack the design of multi-model collaborative reasoning mechanisms and have failed to establish scenario-based scoring rules for different R&D stages (such as initial screening of lead compounds, lead optimization, and preclinical candidates). Evaluation results often fall into the decision-making dilemma of "high score, low confidence" or "low score, high perturbation" due to misalignment of confidence intervals between models and lack of quantification of uncertainty. In particular, the lack of proactive verification and feedback calibration capabilities for boundary samples leads to continuous drift of screening criteria during iteration, making it impossible to form a closed-loop optimization evaluation ecosystem.

[0036] Furthermore, existing technologies generally suffer from drawbacks such as isolated model outputs, fragmented scoring systems, statically fixed thresholds, and lack of guidance on uncertainty. This makes it difficult for systems to adaptively adjust evaluation strategies when faced with structurally diverse and ambiguously defined compound libraries, hindering the ability to control false positive inflation while maintaining screening recall. Particularly in complex optimization scenarios requiring a balance between toxicity avoidance and efficacy preservation, the lack of confidence-driven sample verification mechanisms and dynamic threshold loading capabilities reduces the evaluation module to a passive scoring tool rather than an intelligent agent with cognitive feedback capabilities, severely restricting the scalability and intelligentization of computation-driven drug discovery.

[0037] Therefore, there is an urgent need to build an integrated ADMET toxicity agent assessment architecture with multi-model integrated evaluation, scenario-adaptive threshold, uncertainty perception and active learning capabilities.

[0038] This application provides an integrated ADMET and toxicity proxy assessment method and system, which solves the problems of scattered models, rigid thresholds, and inconsistent standards in existing related technologies, and achieves the goal of providing efficient, reliable, and scalable technical support for early screening of drug development.

[0039] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0040] Example 1:

[0041] Firstly, this invention proposes an integrated proxy assessment method for ADMET and toxicity, see [link to relevant documentation]. Figure 1 The method includes: S1. Obtain the raw data stream, which includes a molecular structure description file and a raw evaluation data stream; S2. Based on feature mapping, the original data stream is transformed into a unified feature vector space; the feature vector space includes multiple independent feature subspaces. S3. Based on the feature subspace, perform dynamic weighted calculation on each evaluation dimension of the candidate compound, and simultaneously obtain dynamic thresholds based on the feature subspace, and perform threshold self-calibration based on the dynamic thresholds to determine the critical decision boundaries of each evaluation dimension of the candidate compound. S4. Compare the dynamically weighted scores of each assessment dimension with the self-calibration thresholds to generate the final comprehensive risk score.

[0042] The following is in conjunction with the appendix Figure 1-5 The following details the implementation process of an embodiment of the present invention, including explanations of the specific steps S1-S4.

[0043] This embodiment proposes an integrated ADMET and toxicity proxy assessment method for automated evaluation in the early stages of drug development. Its core lies in the integrated proxy assessment of the absorption, distribution, metabolism, excretion, and toxicity characteristics of candidate compounds. See also... Figure 1 The specific steps for implementing this method are as follows: S1. Obtain the raw data stream, which includes molecular structure description files, in vitro experimental measurements, in vivo pharmacokinetic parameters, toxicity test results data, and historical evaluation records.

[0044] Receive raw data streams from external databases or experimental platforms. These streams include molecular structure description files of candidate compounds, in vitro experimental results, in vivo pharmacokinetic parameters, toxicity test results, and historical evaluation records. Specifically, when acquiring data, the following should be considered: The molecular structure input unit receives two-dimensional or three-dimensional molecular structure description files of candidate compounds. These files contain information such as atom types, bond relationships, stereoconfiguration, and functional group distribution characteristics. Based on these files, molecular structure descriptors are obtained. These descriptors include, but are not limited to, physicochemical parameters such as topological polar surface area, lipid-water partition coefficient, number of hydrogen bond donors and acceptors, molecular weight, and number of rotational bonds.

[0045] The system synchronously accesses raw evaluation data streams from multiple (at least three) independent validation models through a multi-source data preprocessing unit. These raw evaluation data streams include, but are not limited to, in vitro experimental measurements, in vivo pharmacokinetic parameters, toxicity test results, and historical evaluation records. In vitro experimental measurements include Caco-2 cell permeability, liver microsomal stability half-life, measured plasma protein binding rate, and cytochrome P450 enzyme inhibition intensity. In vivo pharmacokinetic parameters include predicted oral bioavailability, half-life, clearance rate, and volume of distribution. Toxicity test results include acute toxicity median lethal dose, Ames test results for genotoxicity, hERG channel inhibition rate for cardiotoxicity, hepatotoxicity ALT / AST elevation, cytotoxicity median inhibitory concentration, genotoxicity positive probability, and cardiotoxicity risk index. Historical evaluation records include the candidate compound's scores from different models in previous evaluation cycles and its final experimental validation conclusions.

[0046] In one embodiment, the aforementioned molecular structure input unit supports extended access to a quantum chemical calculation module, which generates three quantum-derived parameters: molecular orbital energy distribution, electrostatic potential surface gradient change rate, and frontier orbital gap width. These parameters are then incorporated as additional channels into the original structure matrix to participate in the subsequent feature space mapping process.

[0047] In one embodiment, the aforementioned multi-source data preprocessing unit further integrates an outlier detection subsystem. This outlier detection subsystem is constructed using the Isolation Forest algorithm, and its training set comprises complete evaluation records of 8,000 known safe compounds and 2,000 known highly toxic compounds that have undergone clinical validation over the past three years. When data in a certain dimension deviates from the group mean of similar compounds by more than two standard deviations, it is marked as a suspected outlier, its entry into the main evaluation process is suspended, and it is transferred to the manual review queue. Simultaneously, the data source provider is notified to verify the original experimental conditions or model parameter settings.

[0048] In a preferred embodiment, all the above-mentioned input data must undergo data integrity verification before entering the system. The verification includes checking for missing data fields, whether the values ​​exceed the physically reasonable range, whether the timestamps are continuous, and whether the data source is authorized. If the verification fails, an exception handling mechanism is triggered, the system automatically records the error log and suspends subsequent processing until manual intervention or data source repair.

[0049] S2. Based on feature mapping, the original data stream is transformed into a unified feature vector space; the feature vector space includes multiple independent feature subspaces, including a physicochemical feature subspace, an in vitro experimental feature subspace, and an in vivo and toxicity feature subspace.

[0050] During feature mapping, the original input data is transformed into a unified feature vector space, see [link to relevant documentation]. Figure 2 The process includes: Numerical data is standardized by using the Z-score method to map each indicator to a distribution interval with a mean of zero and a standard deviation of one, thus eliminating the influence of dimensional differences on subsequent calculations. For categorical variables, such as positive / negative results of toxicity tests, one-hot encoding is used to convert them into binary vectors. For molecular structure descriptors, a pre-trained graph neural network encoder is invoked to convert the molecular graph structure into a fixed-dimensional embedding vector. This encoder has learned the implicit association patterns between molecular substructures and ADMET attributes during the training phase. Specifically: 1) Perform standardization cleaning on the above molecular structure description files to obtain a normalized structure matrix. Specifically, this includes: removing invalid topological fragments, correcting charge states, unifying the explicit and implicit representation of hydrogen atoms, and finally outputting a normalized structure matrix that meets the requirements of subsequent calculations.

[0051] 2) Perform timestamp alignment and dimension normalization transformation on the above original evaluation data stream to form a dimensionless data tensor. Specifically, this includes: completing sequence recombination based on a unified sampling time benchmark; mapping values ​​under different unit systems to the closed interval of 0 and 1 according to a preset physical dimension transformation rule to form a dimensionless data tensor.

[0052] 3) Perform a nonlinear embedding transformation on the above normalized structure matrix and dimensionless data tensor to obtain a 128-dimensional dense vector as the structure response encoding. This process includes: using a deep residual convolution architecture to implement the nonlinear embedding transformation.

[0053] In one embodiment, the aforementioned deep residual convolutional architecture includes five stacked convolutional kernel layers with kernel sizes of 3x3, 5x5, 7x7, 9x9, and 11x11, respectively. The stride is fixed at 1, and the padding strategy satisfies the requirement of a constant output feature map size. Each convolutional layer is followed by batch normalization and a modified linear activation function. The final output is compressed into a 128-dimensional dense vector by global average pooling as a structural response encoding.

[0054] In a preferred embodiment, the deep residual convolutional architecture in the feature space mapping unit supports online incremental learning. The model fine-tuning procedure is triggered whenever one hundred clinically confirmed actual cases are accumulated. This fine-tuning procedure freezes the parameters of the bottom four convolutional layers and only opens the top convolutional layer and the fully connected layer to participate in backpropagation optimization. The initial learning rate is set to 0.001, a cosine annealing scheduling strategy is adopted, and the number of training rounds is limited to ten.

[0055] After feature mapping is completed, three independent feature subspaces are generated: physicochemical feature subspace, in vitro experimental feature subspace, and in vivo and toxicity feature subspace. The feature vectors in each subspace retain the time-series attributes and spatial distribution characteristics of the original data, so that the subsequent dynamic weighting mechanism can capture the time-varying correlations and spatial coupling relationships between different evaluation dimensions.

[0056] S3. Based on the feature subspace, perform dynamic weighted calculations on each evaluation dimension of the candidate compound, and simultaneously obtain dynamic thresholds based on the feature subspace. Then, perform threshold self-calibration based on the dynamic thresholds to determine the critical decision boundaries of each evaluation dimension of the candidate compound.

[0057] The output of the feature mapping in step S2 above will be used simultaneously for subsequent dynamic weighting calculation and threshold self-calibration, forming a parallel processing path.

[0058] 1) Based on the overall distribution characteristics of the current batch of compounds, dynamic weighted integration is performed on each evaluation dimension to generate a weighted fusion feature vector.

[0059] join Figure 3The dynamic weighted calculation unit assigns dynamically adjusted weight coefficients to different evaluation dimensions of the current compound based on the goal orientation of the current evaluation task and historical data feedback. The core of the cross-dimensional dynamic weight coefficient adjustment mechanism lies in constructing a learnable weight generation network. Its inputs are the representation vector of the current compound in each feature subspace, the model consistency index in the historical evaluation records, and the evaluation priority parameters set by the user. The output is a ternary probability distribution.

[0060] This step dynamically weights and integrates each evaluation dimension based on the overall distribution characteristics of the current batch of compounds to generate a weighted fusion feature vector. Specific steps include: Based on sliding window statistical analysis, the standard deviation and skewness coefficient of each evaluation dimension in the most recent fifty evaluated compound samples are calculated. If the standard deviation is greater than 0.3 and the absolute value of the skewness is less than 0.5, the dimension is assigned a base weight of 1.0; otherwise, it is proportionally decayed to 0.6. The weights of all evaluation dimensions are normalized by softmax and then multiplied element-wise with the corresponding feature vectors to generate a weighted fusion feature vector. Next, the weighted fusion feature vector is input into a three-layer fully connected classifier to output a ternary probability distribution, and the term with the highest probability is selected as the final risk level identifier.

[0061] In one embodiment, the weight generation network consists of a three-layer fully connected neural network. The three-layer fully connected classifier has 256 neurons in the first layer, 128 in the second layer, and 3 in the third layer. The activation function uses a modified linear unit (such as the hyperbolic tangent function) in the first two layers and a softmax function in the output layer to ensure that the sum of the weights of each dimension is one, outputting a ternary probability distribution. The one with the highest probability is selected as the final risk level identifier.

[0062] The weighting process follows the principle of cross-dimensional independence, meaning that the weight allocation among physicochemical properties, in vitro experimental results, and in vivo and toxicity data is not directly affected by their numerical values, but rather determined by their predictive stability and information gain in historical validation data. For example, if a certain type of compound shows a high correlation between in vitro permeability and in vivo bioavailability in historical data, the system automatically increases the weight of the in vitro experimental dimension; if the toxicity results of another type of compound have a strong nonlinear mapping with the molecular structure descriptor, the system correspondingly increases the weight contribution of the physicochemical dimension.

[0063] In one embodiment, the dynamic weighting coefficients are updated online after each complete evaluation cycle. The update is based on the sum of squared residuals between the latest validation data and the model prediction results. The smaller the residual, the more the weight of the corresponding dimension is appropriately amplified in the next cycle; conversely, the larger the residual, the more it is compressed. This mechanism ensures that the evaluation method can adapt to the changing needs of different compound categories and different R&D stages, avoiding evaluation bias caused by fixed weights.

[0064] In one embodiment, the cross-dimensional dynamic weighting mechanism further incorporates an expert knowledge constraint mechanism. This mechanism pre-defines a set of upper and lower limit rules for weights, jointly developed by senior toxicologists and pharmacokinetics experts. Within this rule set, for example, the weight for the cardiotoxicity dimension must not be lower than 0.8, and the weight for the genotoxicity dimension must not be higher than 1.2. After each dynamic weight generation, it is forcibly trimmed to the pre-defined rule range.

[0065] 2) Maintain a dynamic threshold database based on historical evaluation records and execute a threshold self-calibration process to determine the critical decision boundaries for each evaluation dimension. This step uses a threshold self-calibration unit to adjust the decision thresholds for each evaluation dimension in real time, addressing the problems of rigid thresholds and inability to reflect data distribution drift in traditional methods. Specific steps include: 2.1) Maintain a dynamic threshold database based on historical evaluation record data.

[0066] Record the critical judgment boundaries of all compounds in each evaluation dimension. The threshold update process is triggered every time 10 new compounds are evaluated. The percentile truncation method is used to take the 75th percentile of the score of each dimension within the window period as the new upper limit threshold, the 25th percentile as the new lower limit threshold, and the middle area is defined as the observation interval. For compounds that fall into the observation interval, the secondary verification protocol is initiated to call the backup high-precision simulation model for verification.

[0067] 2.2) Perform a threshold self-calibration process to determine the critical decision boundaries for each evaluation dimension of the current compound.

[0068] The threshold self-calibration process can be found in [reference]. Figure 4 .

[0069] First, the feature vectors of the most recent 1,000 validated compounds and the final experimental conclusions are extracted from the historical evaluation record database (or from the dynamic threshold database) to construct a dynamic calibration sample set.

[0070] Subsequently, the empirical distribution function of each evaluation dimension in positive and negative samples is calculated separately, and its probability density curve is fitted using the kernel density estimation method. Based on this probability density curve, the current optimal discrimination threshold (i.e., the dynamic threshold) is automatically determined, and the criterion is to maximize the Youden exponent, which is the sum of sensitivity and specificity minus one.

[0071] After the threshold calculation is completed, the dynamic threshold is compared with the preset conservative threshold. If the dynamic threshold deviates from the conservative threshold by more than 15%, a manual review mechanism is triggered, requiring domain experts to confirm the rationality of the threshold adjustment. If no review is triggered, the new threshold takes effect immediately and is applied to the evaluation process of all current and subsequent compounds.

[0072] In one embodiment, a threshold self-calibration process is set to run periodically (e.g., every 24 hours) or is forcibly triggered when more than a certain number of new validation data entries are added (e.g., 100 entries), ensuring that the evaluation criteria are always synchronized with the latest experimental data. This mechanism effectively avoids the problem of evaluation criteria becoming invalid due to changes in laboratory conditions, reagent batch replacements, or instrument drift.

[0073] In one embodiment, the above threshold self-calibration process supports a multi-agency collaborative calibration mode. In a distributed deployment environment, each node periodically uploads its local threshold update log to the central coordination server. The central coordination server uses a federated averaging algorithm to aggregate the contributions of each node, generate a global consensus threshold, and broadcasts it to all nodes for synchronization.

[0074] S4. Compare the dynamically weighted scores of each assessment dimension with the self-calibration thresholds to generate the final comprehensive risk score.

[0075] After collecting all outputs from step S3 (weighting coefficients for each assessment dimension and new thresholds for each assessment dimension), the comprehensive scoring output unit is used to compare the dynamically weighted scores for each dimension with the self-calibration thresholds to generate the final comprehensive risk score. Figure 4 As shown, the comprehensive risk score is calculated using the weighted geometric mean formula, and its mathematical expression is as follows: , Where S represents the comprehensive risk score. For the standardized score of the i-th evaluation dimension, represents the dynamic weight coefficient for the corresponding dimension, and n represents the total number of evaluation dimensions.

[0076] It should be noted that this formula ensures that a low score in any dimension will lead to a significant decrease in the overall score, thereby enhancing the ability to screen for high-risk compounds.

[0077] In one embodiment, standardized score The calculation method is as follows: If the original score for this dimension is higher than the self-calibration threshold, then ,in, This is the original score. This is the self-calibration threshold; if the original score is below the threshold, then... .

[0078] Standardized score The above calculation method results in the score having the maximum gradient change near the threshold, enhancing the model's ability to distinguish compounds near the critical value.

[0079] After the comprehensive risk score is output, a complete explanatory assessment report is generated simultaneously. The report includes the original data for each dimension, standardized scores, dynamic weight values, weighting coefficients, self-calibration thresholds, threshold boundaries, a summary of the judgment criteria, and intermediate steps in the score calculation, ensuring that the assessment results are traceable and auditable.

[0080] In one embodiment, the comprehensive scoring output unit is further equipped with a visualization interpretation engine. This visualization interpretation engine uses gradient weighted class activation mapping technology to trace back the key structural regions and key evaluation dimensions on which the final classification decision depends, generate a heat map superimposed on the original molecular structure, and label the names of the dominant factors and their influence intensity values.

[0081] To further improve the reliability of the integrated ADMET and toxicity proxy assessment results, the aforementioned integrated ADMET and toxicity proxy assessment method also includes: S5. Achieving parameter coordination and model calibration between different evaluation systems based on a federated learning framework; and / or By utilizing uncertainty quantification, a confidence interval and risk level label are added to the comprehensive risk score.

[0082] In one embodiment, a federated collaborative calibration unit is set up to support parameter collaboration and model calibration between evaluation systems of multiple research institutions or laboratories without sharing raw data. For example... Figure 5 As shown, this mechanism is based on a federated learning framework. Each participant only uploads the gradient updates or parameter increments of their local model. The central server performs weighted aggregation of the increments to generate a global model and then distributes the updated parameters to each participant.

[0083] Weighting is based on each participant's data quality score and historical validation accuracy. The data quality score is calculated by weighting three indicators: data completeness, annotation consistency, and experimental repeatability. The federated collaboration process is executed every 7 days, with each iteration including 3 rounds of parameter exchange and aggregation. This mechanism ensures consistency of evaluation standards across different institutions while protecting the data privacy and intellectual property of each institution.

[0084] During collaborative calibration, abnormal drift of participating models is automatically detected. If the parameter update of a participating model causes the global model performance to degrade beyond the preset tolerance, it will be temporarily excluded from the next round of aggregation until its data quality score recovers to the qualified level.

[0085] In a more preferred embodiment, an uncertainty quantification unit can be used to attach a confidence interval and risk level identifier to the final comprehensive risk score to reflect the reliability of the assessment results. For example... Figure 5As shown, the uncertainty calculation employs the Monte Carlo Dropout method. During the scoring process, the neuronal connections of the dynamic weighting network and the feature mapping network are randomly masked, and the scoring calculation is repeated 100 times. The standard deviation of the statistical results is used as the uncertainty measure. If the standard deviation is less than 0.05, it is marked as high confidence; if it is between 0.05 and 0.15, it is marked as medium confidence; and if it is greater than 0.15, it is marked as low confidence. Simultaneously, the system automatically adjusts the risk level according to the uncertainty level: at high confidence, scores above 0.7 are marked as low risk, below 0.3 as high risk, and the rest as medium risk; at medium confidence, scores above 0.75 are marked as low risk, and below 0.25 as high risk; at low confidence, scores above 0.8 are marked as low risk, below 0.2 as high risk, and the rest are marked as pending review. In addition, the uncertainty quantification unit outputs two additional indicators after each evaluation: one is the prediction confidence score, ranging from 0 to 1; the other is the dimensional conflict index, which is calculated as the reciprocal of the sum of the absolute values ​​of the Pearson correlation coefficients of the original scores of each dimension and their weighted scores. When any indicator falls below the preset warning line, the system automatically adds a prompt message suggesting that supplementary experiments be conducted for verification.

[0086] The design of uncertainty quantification ensures that R&D personnel can make differentiated decisions based on the reliability of the evaluation results, avoiding misjudgments caused by model uncertainty.

[0087] This completes the entire process of the integrated ADMET and toxicity assessment method of this embodiment.

[0088] This ADMET and toxicity integrated proxy assessment method, by constructing a cross-dimensional dynamic weighting mechanism and combining the preprocessing and feature space mapping of multi-source heterogeneous data, realizes the quantitative scoring of the comprehensive pharmacokinetic and toxicological risks of compounds, and supports dual-track output of federated collaborative calibration and uncertainty quantification, thereby solving the technical defects of existing technologies such as scattered validation models, inconsistent assessment thresholds and unstable screening standards.

[0089] Example 2:

[0090] Secondly, this invention also provides an integrated ADMET and toxicity assessment system, see [link to relevant documentation]. Figure 6 The system includes: A data input unit is configured to acquire a raw data stream, which includes a molecular structure description file and a raw evaluation data stream. The feature mapping unit is configured to transform the original data stream into a unified feature vector space based on feature mapping; the feature vector space includes multiple independent feature subspaces. The dynamic weighted calculation unit is configured to perform dynamic weighted calculations on each evaluation dimension of the candidate compound based on the feature subspace. The threshold self-calibration unit is configured to obtain a dynamic threshold based on the feature subspace and perform threshold self-calibration based on the dynamic threshold to determine the critical decision boundary of each evaluation dimension of the candidate compound. The comprehensive scoring output unit is configured to compare the dynamically weighted scores of each assessment dimension with the self-calibration thresholds to generate the final comprehensive risk score.

[0091] The ADMET and toxicity integrated proxy assessment system proposed in this embodiment mainly includes a data input unit, a feature mapping unit, a dynamic weighted calculation unit, a threshold self-calibration unit, a comprehensive score output unit, a federated collaborative calibration unit, and an uncertainty quantification unit. These units interact and coordinate processes through preset data interfaces and control protocols. The system executes the steps of the method in Embodiment 1 and its preferred embodiments, and is used for automated assessment in the early stages of drug development. Its core function is to perform an integrated proxy assessment of the absorption, distribution, metabolism, excretion, and toxicity characteristics of candidate compounds.

[0092] In one embodiment, the system further integrates an uncertainty quantification unit, which outputs two additional indicators after each evaluation: a prediction confidence score, ranging from 0 to 1; and a dimensional conflict index, calculated as the reciprocal of the sum of the absolute values ​​of the Pearson correlation coefficients of the original scores of each dimension and their weighted scores. When any indicator falls below a preset warning line, the system automatically adds a prompt suggesting supplementary experimental verification.

[0093] See Figure 7 The data flow between the various units within the system follows a strict timing control protocol. After data verification, the data input unit sends a ready signal to the feature mapping unit. After vector transformation, the feature mapping unit simultaneously sends feature ready signals to the dynamic weighting calculation unit and the threshold self-calibration unit. Upon receiving feature data and historical feedback, the dynamic weighting calculation unit calculates weight coefficients and sends them to the comprehensive scoring output unit. After updating the threshold, the threshold self-calibration unit broadcasts the new threshold to the comprehensive scoring output unit. After receiving all inputs (weight coefficients and the new threshold), the comprehensive scoring output unit performs scoring calculations and triggers the uncertainty quantification unit. After completing the confidence assessment, the uncertainty quantification unit outputs the final result and an explanation report to the user interface. The entire process employs an asynchronous message queue mechanism to ensure that each unit can run independently without blocking each other. If any unit times out or returns an error code, the system automatically initiates a degradation mode, continues the evaluation using the valid parameters from the previous cycle, and records the abnormal event for subsequent analysis.

[0094] In a preferred embodiment, the system is deployed in a containerized microservice architecture, exposing application programming interfaces to the outside world, supporting concurrent request processing and asynchronous task scheduling; each evaluation task is encapsulated as an independent process instance, with resource quotas limited to a quad-core CPU, eight gigabytes of memory, and a maximum runtime of 200 seconds; tasks that fail to complete within the timeout period are automatically terminated and return an error code.

[0095] The system supports a hybrid cloud-and-local architecture during deployment. The core computing unit is deployed on a cloud server to leverage its elastic computing resources for batch evaluation of large-scale compound libraries; the data input and result display unit can be deployed on a local workstation to meet data security and real-time interaction requirements. The system provides standardized application programming interfaces (APIs) that support seamless integration with mainstream drug discovery platforms such as Schrödinger, MOE, and Pipeline Pilot. The API protocol uses a RESTful architecture, JSON data format, and OAuth 2.0 authentication mechanism to ensure system access security and compatibility. The system requires a Linux operating system, at least 32 GB of RAM, at least eight CPU cores, and at least 8 GB of CUDA-accelerated GPU memory.

[0096] In a preferred embodiment, the system has a built-in version control and rollback mechanism. Each major algorithm upgrade or threshold system change generates a unique version identifier and retains the complete runtime environment images of the previous three historical versions. Users can specify the target version number when submitting an evaluation request.

[0097] The system's scalable design allows for the addition of new evaluation dimensions or replacement of core algorithms. For example, new toxicity prediction models such as DeepTox or ProTox 2.0 can be integrated simply by adapting their output interfaces to the feature mapping unit; new weighting strategies, such as reinforcement learning-based dynamic weight optimizers, can be introduced by replacing the neural network structure of the dynamic weighting calculation unit; and new uncertainty quantification methods, such as Bayesian neural networks, can be added by simply adding corresponding computational branches to the uncertainty quantification unit. All extensions are implemented through a plug-in architecture, requiring no modification to the core code, ensuring the system's long-term maintainability and technological forward-looking capabilities. System version updates are achieved through containerized deployment. New version images are centrally managed, and deployment nodes can pull and hot-swap images as needed, ensuring uninterrupted service continuity.

[0098] Upon system startup, pre-trained model parameters and historical calibration data are automatically loaded, the state of each unit is initialized, and an encrypted communication channel with the federated server is established. During operation, the system monitors resource utilization and task queue length in real time. If the load exceeds the threshold, task sharding and parallel computing are automatically initiated to ensure that the evaluation response time is always less than five seconds.

[0099] In practical applications, this system significantly improves the efficiency and accuracy of early-stage drug screening. Traditional methods require calling multiple independent models to assess properties such as absorption, metabolism, and toxicity, followed by manual comprehensive judgment, which is time-consuming and inconsistent in standards. This system, through an integrated architecture and dynamic weighting mechanism, automates, standardizes, and automates the assessment process, enabling researchers to obtain a comprehensive assessment report with quantitative scores and uncertainty indicators within seconds. This allows for rapid elimination of high-risk candidate compounds, concentrating resources on high-potential molecules. Simultaneously, the federated collaboration mechanism ensures the comparability of assessment results across different research teams, promoting cross-institutional collaboration and data sharing. The uncertainty quantification function provides risk warnings for decision-making, avoiding later-stage research failures due to model blind spots. This system improves assessment accuracy by 27% and screening efficiency by more than five times compared to traditional methods, significantly shortening the drug development cycle and reducing research costs.

[0100] It is understood that the ADMET and toxicity integrated proxy assessment system provided in this embodiment of the invention corresponds to the ADMET and toxicity integrated proxy assessment method described above. The explanations, examples, and beneficial effects of its contents can be referred to the corresponding contents in the ADMET and toxicity integrated proxy assessment method, and will not be repeated here.

[0101] In summary, compared with existing technologies, it has the following beneficial effects:

[0102] 1. This application proposes an integrated ADMET and toxicity proxy assessment method. By introducing core technologies such as cross-dimensional dynamic weighting mechanisms, multi-source data feature mapping, threshold self-calibration, federated collaboration, and uncertainty quantification, it achieves automated, standardized, and intelligent assessment of the pharmacokinetic and toxicological risks of candidate compounds. This method solves the core problems of fragmented models, rigid thresholds, and inconsistent standards in existing technologies, providing efficient, reliable, and scalable technical support for early screening in drug development, and has significant industrial application value and social benefits.

[0103] 2. This application proposes an integrated ADMET and toxicity assessment method. By constructing a unified feature space and a dynamic weighting mechanism, it effectively integrates the outputs of multiple previously scattered and isolated validation models, eliminating assessment bias caused by model heterogeneity. By introducing sliding window statistics and percentile truncation, it achieves continuous adaptive adjustment of the threshold, avoiding the problem of static thresholds becoming invalid over time. Through the dual constraints of federated learning and expert rules, it ensures both the flexibility of the assessment system and maintains scientific rigor. Through dual-track output of visual interpretation and uncertainty quantification, it significantly improves the transparency and credibility of the assessment results, enabling researchers to accurately locate the root causes of compound defects, accelerate the lead structure optimization process, and reduce the risk of later clinical failures.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for integrated ADMET and toxicity assessment, characterized in that, The method includes: Acquire the raw data stream, which includes a molecular structure description file and a raw evaluation data stream; Based on feature mapping, the original data stream is transformed into a unified feature vector space; the feature vector space includes multiple independent feature subspaces. Based on the feature subspace, dynamic weighted calculations are performed on each evaluation dimension of the candidate compound. At the same time, dynamic thresholds are obtained based on the feature subspace, and threshold self-calibration is performed based on the dynamic thresholds to determine the critical decision boundaries of each evaluation dimension of the candidate compound. The scores of each assessment dimension after dynamic weighting are compared with the self-calibration thresholds to generate the final comprehensive risk score.

2. The method as described in claim 1, characterized in that, The molecular structure description file also includes: quantum-derived parameters obtained by quantum chemical calculations based on the molecular structure, including: molecular orbital energy distribution, electrostatic potential surface gradient rate of change, and frontier orbital gap width.

3. The method as described in claim 1, characterized in that, The method further includes: performing outlier detection on the original evaluation data stream; and / or Perform data integrity verification on the original data stream.

4. The method as described in claim 3, characterized in that, The original assessment data stream includes: in vitro experimental measurements, in vivo pharmacokinetic parameters, toxicity test results data, and historical assessment records data. The feature subspace includes a physicochemical feature subspace, an in vitro experimental feature subspace, and an in vivo and toxicity feature subspace.

5. The method as described in claim 1, characterized in that, The transformation of the original data stream into a unified feature vector space based on feature mapping includes: The molecular structure description file is subjected to a normalization cleaning process to obtain a normalized structure matrix; The original evaluation data stream is subjected to timestamp alignment and dimension normalization transformation to form a dimensionless data tensor; A deep residual convolutional architecture is used to perform a nonlinear embedding transformation on the normalized structure matrix and the dimensionless data tensor.

6. The method as described in claim 1, characterized in that, The dynamic weighted calculation of each evaluation dimension of the candidate compound based on the feature subspace includes: Based on the overall distribution characteristics of the current batch of compounds, dynamic weighted integration is performed on each evaluation dimension of the compounds to generate a weighted fusion feature vector, specifically including: The standard deviation and skewness coefficient of each evaluation dimension of the compound were calculated based on the sliding window statistical analysis in the most recently preset number of evaluated compound samples. If the standard deviation and absolute value of skewness meet the preset conditions, the basic weight of the evaluation dimension is assigned; otherwise, it is reduced proportionally. The weights of all evaluation dimensions are normalized by softmax and then multiplied element-wise with the corresponding feature vectors to generate a weighted fusion feature vector.

7. The method as described in claim 1, characterized in that, The formula for calculating the comprehensive risk score is as follows: ， Where S represents the comprehensive risk score. For the standardized score of the i-th evaluation dimension, represents the dynamic weight coefficient for the corresponding evaluation dimension, where n is the total number of evaluation dimensions.

8. The method as described in claim 7, characterized in that, The standardized score The calculation method is as follows: If the original score of the evaluation dimension is higher than the self-calibration threshold, then ; If the original score of the evaluation dimension is lower than the self-calibration threshold, then ; in, This is the original score. This is the self-calibration threshold.

9. The method according to any one of claims 1-8, characterized in that, The method further includes: S5. Achieving parameter coordination and model calibration between different evaluation systems based on a federated learning framework; and / or By utilizing uncertainty quantification, a confidence interval and risk level label are added to the comprehensive risk score.

10. An integrated ADMET and toxicity assessment system, characterized in that, The system includes: A data input unit is configured to acquire a raw data stream, which includes a molecular structure description file and a raw evaluation data stream. The feature mapping unit is configured to transform the original data stream into a unified feature vector space based on feature mapping; the feature vector space includes multiple independent feature subspaces. The dynamic weighted calculation unit is configured to perform dynamic weighted calculations on each evaluation dimension of the candidate compound based on the feature subspace. The threshold self-calibration unit is configured to obtain a dynamic threshold based on the feature subspace and perform threshold self-calibration based on the dynamic threshold to determine the critical decision boundary of each evaluation dimension of the candidate compound. The comprehensive scoring output unit is configured to compare the dynamically weighted scores of each assessment dimension with the self-calibration thresholds to generate the final comprehensive risk score.

Images