A machine learning-based preclinical drug experiment quality assessment method
By constructing a dose-effect trend function and optimizing the LOF algorithm with multiple neighborhood parameters, the problems of misjudgment and missed judgment in drug experiments by traditional machine learning LOF algorithms are solved, and the accuracy and stability of drug experiment quality assessment are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG XINBO PHARMA R&D
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional machine learning LOF algorithms fail to understand pharmacological trends in preclinical drug trials, leading to misjudgments of data points. Furthermore, the neighborhood parameter settings are not adapted to uneven data density, resulting in unstable evaluation results.
By constructing a dose-response trend function, calculating the trend residual and biological indicator evaluation weights, and combining composite distance and multi-neighborhood parameters, the LOF algorithm is optimized to identify outlier data points.
It improves the accuracy and stability of drug trial quality assessment, avoids misjudgment of valid data and omission of abnormal data, and provides reliable assessment results.
Smart Images
Figure CN122158189A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology. More specifically, this invention relates to a machine learning-based method for quality assessment of preclinical drug trials. Background Technology
[0002] In the innovative drug development process, preclinical trials are a crucial step in evaluating the efficacy and safety of candidate drugs. The quality of data generated from preclinical trials directly affects the accuracy of subsequent research and development decisions, such as whether to advance the candidate drug to the more costly clinical trial stage. Therefore, it plays a vital role in the entire drug development project.
[0003] Currently, quality control in preclinical trials primarily relies on researchers strictly adhering to standard operating procedures, their professional experience and judgment, and the application of basic statistical tests. To pursue more objective and automated evaluation methods, it is necessary to explore the application of machine learning and data mining algorithms to improve the efficiency and accuracy of experimental quality assessment. The Local Outlier Factor (LOF) algorithm, a classic unsupervised machine learning anomaly detection technique, has been attempted to identify outlier data points in experiments, thereby assessing the quality of drug trials. The core idea of the LOF algorithm is to determine whether a data point is an outlier by comparing the density of a data point with the density of other data points in its neighborhood.
[0004] However, when traditional machine learning LOF algorithms are directly applied to the evaluation of complex preclinical experimental data, they often use the universal Euclidean distance to measure the proximity of data points. This metric treats all biological indicators equally and fails to understand and integrate the inherent laws in pharmacology. For example, there is usually a specific trend relationship between drug dose and biological effect. A valid data point that deviates numerically from the population but perfectly matches the dose-response curve trend may be incorrectly judged as abnormal. Conversely, a data point that appears to be within the normal range but seriously violates the trend that should exist at its dose level may be incorrectly judged as abnormal. Abnormal data points may be overlooked; at the same time, the performance of traditional LOF algorithms is highly dependent on the setting of neighborhood parameters. In preclinical experimental data, the distribution density of data points is often uneven. For example, the data of the control group may be densely distributed due to consistent effects, while the data of the high-dose group may become sparse due to individual differences or the occurrence of toxic reactions. If a single, globally fixed neighborhood parameter is used, it is difficult to adapt to subgroups of different densities at the same time. This can easily lead to missing real abnormal points in dense data areas and misjudging normal points as abnormal in sparse data areas, making the results of drug trial quality assessment unstable and reducing reliability. Summary of the Invention
[0005] To address the shortcomings of traditional machine learning-based LOF algorithms in drug trial quality assessment, such as insufficient accuracy due to a lack of domain knowledge in distance metrics and instability caused by reliance on a single neighborhood parameter, this invention proposes a machine learning-based method for preclinical drug trial quality assessment. This method includes the following steps: Multiple experimental samples are acquired, each including a dose level and observed values of at least one biological indicator. Based on the dose level, curve fitting is performed on each biological indicator to obtain a dose-response trend function. The trend residual of each biological indicator in each experimental sample is determined based on the observed value of each biological indicator and the fitted value of the corresponding dose-response trend function. Based on the trend residual of each biological indicator, a biological indicator evaluation weight sequence composed of the quality evaluation weights of all biological indicators is determined. Based on the Euclidean distance between any two experimental samples, the trend residual of each biological indicator in the two experimental samples, and the quality evaluation weight of each biological indicator in the biological indicator evaluation weight sequence, a composite distance between two experimental samples is determined. Based on a preset neighborhood parameter sequence and the composite distance between each experimental sample, a quality evaluation index for each experimental sample is determined using the LOF algorithm. Based on the quality evaluation index of each experimental sample, the quality of the preclinical drug trial is evaluated.
[0006] This invention constructs a dose-response trend function by curve fitting each biological indicator based on dose level, achieving a scientific model of the intrinsic laws of pharmacology. By calculating the trend residual, it can effectively identify abnormal data points deviating from the dose-response curve, overcoming the limitation of traditional machine learning LOF algorithms in understanding pharmacological trend relationships. By determining the evaluation weight sequence of biological indicators through trend residuals, it achieves differentiated evaluation of the importance of different biological indicators, and can dynamically adjust their weights according to the contribution of each indicator to anomaly detection, improving the accuracy and specificity of quality assessment. By constructing a composite distance calculation model including Euclidean distance, trend residuals, and weight sequences, it achieves a comprehensive evaluation of the similarity between experimental samples, considering both numerical similarity and integrating pharmacological trend information, optimizing the distance measurement method of machine learning algorithms, and improving the scientificity and accuracy of anomaly detection. Through the adaptive selection of preset neighborhood parameter sequences, it overcomes the limitations of traditional machine learning LOF algorithms in cases of uneven data distribution, and can dynamically adjust the neighborhood size according to the local characteristics of data density, effectively avoiding missed detections in dense areas and misjudgments in sparse areas, improving the stability and reliability of preclinical drug trial quality assessment.
[0007] Furthermore, the biological indicators include at least one of tumor volume, specific protein expression levels, or body weight.
[0008] Furthermore, the observed values are those that have undergone Z-score normalization.
[0009] Furthermore, the trend residual satisfies: In the formula, For experimental samples The The trend residuals of each biological indicator For experimental samples The Observed values of biological indicators, For experimental samples dosage level, For experimental samples The The dose-response trend function corresponding to each biological indicator It is an absolute value function.
[0010] This invention constructs an evaluation model of trend residuals by calculating the absolute difference between observed values and fitted values of the dose-response trend function. This model effectively measures the degree to which experimental samples deviate from the expected dose-response relationship, providing a scientific basis based on pharmacological principles for anomaly detection in machine learning.
[0011] Furthermore, the dose-effect trend function is obtained by dividing all experimental samples with the same dose level into a dose group; for each dose group, the dose-effect trend function is obtained by curve fitting of each biological indicator using spline regression or a four-parameter logistic model.
[0012] Furthermore, the biological indicator evaluation weight sequence satisfies: In the formula, This is a sequence of biological indicator assessment weights, composed of the quality assessment weights of all biological indicators. For all experimental samples The sum of the trend residuals of each biological indicator For all experimental samples Kurtosis of the trend residual distribution of a biological indicator It is a normalized exponential function.
[0013] This invention constructs a composite function comprising the sum of trend residuals and kurtosis, and utilizes... Normalization enables the evaluation of the weights of biological indicators, taking into account the degree of anomaly and distribution characteristics of the indicators, and ensuring the rationality and accuracy of the weight allocation; the sum of trend residuals reflects the overall degree of anomaly of the biological indicator. The larger the sum, the more important the biological indicator is in machine learning anomaly detection; kurtosis reflects the sharpness of the residual distribution. The larger the kurtosis, the more obvious the anomalous characteristics of the biological indicator. The introduction ensures that all weight values are within The sum of all values within the specified range is 1, which facilitates weighted processing in subsequent composite distance calculations and improves the relevance and accuracy of quality assessment.
[0014] Furthermore, the composite distance satisfies: In the formula, For experimental samples With experimental samples The composite distance between them For experimental samples With experimental samples The Euclidean distance between them and Experimental samples With experimental samples The The trend residuals of each biological indicator The weighted sequence of biological indicators is used to evaluate the first... The quality assessment weight of each biological indicator The number of biological indicators.
[0015] This invention constructs a comprehensive evaluation model of composite distance by combining the weighted sum of Euclidean distance and trend residual differences. This model retains the traditional numerical similarity assessment while incorporating pharmacological trend information, achieving multi-dimensional similarity evaluation and optimizing the core distance metric of machine learning algorithms. The trend residual difference term reflects the similarity between two samples in the dose-response trend, and the weights ensure that important indicators receive a higher contribution. When the trend residual differences between two samples in key biological indicators are large, the composite distance increases accordingly, indicating lower similarity. By incorporating trend information, the applicability and accuracy of the machine learning LOF algorithm in pharmacological data analysis are improved, providing a more scientific basis for similarity measurement in preclinical drug trial quality assessment.
[0016] Furthermore, the quality assessment index satisfies: In the formula, For experimental samples The quality assessment index To use the preset neighborhood parameter sequence of the first Experimental samples obtained using the LOF algorithm with neighborhood parameters. Local anomalous factors, This refers to the number of neighborhood parameters in the preset neighborhood parameter sequence. It is a function with maximum value. It is a minimum value function.
[0017] This invention achieves a comprehensive evaluation of the quality assessment index by constructing a composite function containing the maximum and range of local anomaly factors under multiple neighborhood parameters. It fully utilizes the anomaly detection results of machine learning models under different neighborhood parameters, effectively addressing the problem of uneven data distribution. The maximum value term reflects the degree of anomaly of the experimental sample under the most sensitive neighborhood parameter, while the range term reflects the variation of the anomaly factor under different neighborhood parameters. The combination of these two terms can accurately identify samples that exhibit anomalies under multiple neighborhood settings. Through the comprehensive evaluation of multiple neighborhood parameters, it effectively avoids the possibility of missed detections or misjudgments by machine learning models that may be caused by a single parameter setting, thus improving the stability and reliability of the quality assessment.
[0018] Furthermore, the neighborhood parameter sequence is as follows: .
[0019] Furthermore, the assessment of the quality of preclinical drug trials includes: If the quality assessment index of any experimental sample exceeds a preset quality threshold, the drug corresponding to that experimental sample is deemed to have a quality problem, and the preclinical drug trial quality assessment is completed.
[0020] The present invention has the following beneficial effects: (1) Breaking through the limitations of traditional machine learning LOF algorithm which only relies on general Euclidean distance and ignores the trend of drug dose effect, this paper evaluates the fit between observed values and pharmacological laws through dose effect trend function and incorporates it into composite distance calculation. For valid data whose values deviate from the population but conform to the dose effect curve, the trend residual is small and the composite distance will weaken the influence of numerical deviation and avoid being misjudged as abnormal. For abnormal data whose values are in the normal range but violate the dose trend, the trend residual is large and the composite distance will amplify the abnormal feature and ensure accurate identification. This solves the problem of misjudging valid data and missing abnormal data in traditional methods and improves the accuracy of abnormal detection of experimental data.
[0021] (2) Instead of setting fixed neighborhood parameters globally in the traditional machine learning LOF algorithm, a preset neighborhood parameter sequence is used to adapt to the non-uniform density characteristics of preclinical experimental data. For example, for the data-dense control group, a smaller neighborhood parameter is selected to accurately capture subtle abnormalities, and for the data-sparse high-dose group, a larger neighborhood parameter is selected to avoid misjudging normal sparse data as abnormal. This solves the problem of missed detection in dense areas and misjudgment in sparse areas in non-uniform data by the traditional method, and makes the quality assessment results stable and reliable in different dose groups and different data density scenarios.
[0022] (3) By evaluating the weight sequence of biological indicators, different weights can be assigned to different indicators. For example, core indicators that are directly related to drug safety and efficacy can be given higher weights, while secondary metabolic indicators can be given lower weights. This avoids the shift in the evaluation focus caused by the traditional method of treating all indicators equally. When calculating the composite distance, the trend residuals and numerical differences of high-weight indicators will be given priority consideration to ensure that the quality assessment is more in line with the core objectives of preclinical experiments and to improve the pertinence and rationality of the assessment.
[0023] (4) By constructing a quality assessment index, we can avoid misjudging effective data that conforms to the pharmacological rules as abnormal and removing it, and accurately identify low-quality samples that violate the rules. Based on reliable assessment results, we can more accurately judge the safety and effectiveness of drugs, reduce experimental conclusion deviations caused by data misjudgment, reduce the research and development risk of ineffective drugs entering the clinical stage or effective drugs being misscreened, and provide reliable data support for preclinical drug evaluation and subsequent research and development decisions. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating the steps of a machine learning-based preclinical drug trial quality assessment method according to an embodiment of the present invention.
[0025] Figure 2 This is a schematic diagram of the anomaly detection results of the traditional LOF algorithm in a machine learning-based preclinical drug trial quality assessment method according to an embodiment of the present invention.
[0026] Figure 3 This is a schematic diagram of the abnormal detection results of a machine learning-based preclinical drug trial quality assessment method according to an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below. The described embodiments are only a part of the embodiments of the present invention. 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.
[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0029] Please see Figure 1 The diagram illustrates a flowchart of a machine learning-based preclinical drug trial quality assessment method according to an embodiment of the present invention. The method includes the following steps: S01: Obtain multiple experimental samples, each including dose levels and observations of at least one biological indicator.
[0030] It should be noted that, in order to ensure the accuracy and comparability of subsequent machine learning model analysis, it is first necessary to unify and structure the original experimental data from different sources to eliminate interference caused by non-experimental factors such as inconsistent data scales and units, and to organize the data according to the experimental design to establish a clear data structure for subsequent trend fitting and locality analysis.
[0031] Specifically, during the relevant experiments, raw experimental data including dose level, biological indicators, experiment number, and drug batch number were collected for each experimental sample. The biological indicators included tumor volume, expression level of specific proteins, body weight, etc., and the observed values of the biological indicators were Z-score standardized.
[0032] S02: Determine the trend residuals for each biological indicator for each experimental sample.
[0033] It's important to note that the LOF algorithm in machine learning determines anomalies by judging the distance between the absolute values of data. However, the absolute value of data cannot indicate whether it stems from a genuine biological effect or from operational errors or instrument noise during the experiment. Directly relying on raw observations for anomaly detection might overlook implicit errors, such as values that, while not extreme, deviate from scientific principles. Therefore, this step involves dose-effect trend fitting and trend residual calculation, providing the machine learning model with core features incorporating domain knowledge.
[0034] Based on the dose level, curve fitting is performed on each biological indicator to obtain the dose-effect trend function. The trend residual of each biological indicator for each experimental sample is determined according to the observed value of each biological indicator and the fitted value of the corresponding dose-effect trend function.
[0035] Specifically, the dose-effect trend function is obtained as follows: All experimental samples with the same dose level were grouped into one dose group; For each dose group, the dose-response trend function for each biological indicator is obtained by curve fitting for each biological indicator, such as using spline regression or a four-parameter logistic model.
[0036] Specifically, the trend residuals satisfy: ; In the formula, For experimental samples The The trend residuals of each biological indicator For experimental samples The Observed values of biological indicators, For experimental samples dosage level, For experimental samples The The dose-response trend function corresponding to each biological indicator It is an absolute value function.
[0037] in, Indicates experimental sample The The theoretical expected value of each biological indicator Indicates experimental sample The The difference between the observed value and the theoretical expected value of a biological indicator; the larger the value, the better the experimental sample. The more likely the actual observed value deviates from its expected biological trend, the greater the likelihood that it is an outlier; the smaller the value, the more likely the experimental sample is to be an outlier. The closer the actual observed value is to its expected biological trend, the less likely it is to be an outlier.
[0038] S03: Determine the sequence of biological indicator assessment weights, which consists of the quality assessment weights of all biological indicators.
[0039] It should be noted that an ideal biological indicator should possess both strong regularity and sensitivity to anomalies. Different biological indicators exhibit varying degrees of regularity and sensitivity to anomalies, thus requiring different levels of importance in experimental quality assessment. To measure the differences in importance among different biological indicators and optimize the feature weights of the machine learning model, this step calculates the quality assessment weights for each biological indicator based on the trend residuals.
[0040] Based on the trend residuals of each biological indicator, a sequence of biological indicator assessment weights is determined, consisting of the quality assessment weights of all biological indicators.
[0041] Specifically, the biological indicator evaluation weight sequence satisfies: ; In the formula, This is a sequence of biological indicator assessment weights, composed of the quality assessment weights of all biological indicators. For all experimental samples The sum of the trend residuals of each biological indicator For all experimental samples Kurtosis of the trend residual distribution of a biological indicator For normalized exponential functions, The input is a sequence, and the output is also a sequence, which is the evaluation weight of each biological indicator. Indicates will , ... The constructed sequence is input to Thus, the weight sequence for evaluating biological indicators is obtained.
[0042] in, The larger the value, the more likely it is to affect the first... The more accurate the fit of the first biological indicator, the better. The stronger the regularity of the first biological indicator, the more likely the second one is to be found. The greater the quality assessment weight of a biological indicator, the better; The smaller the value, the more difficult it is to determine the first... To accurately fit the biological indicators, the first... The weaker the regularity of the first biological indicator, the more likely the second one is to be found. The smaller the quality assessment weight of each biological indicator. The larger it is, the more likely it is to be the first The more concentrated the distribution of the biological indicators, the better. The sharper the trend residual distribution of a biological indicator, the smaller the difference between the actual observed value of most data points and the fitted value of its dose-effect trend function, thus making a few abnormal experimental samples stand out and easy to identify, and their quality assessment weight is greater. The smaller the value, the more evenly the distribution of the trend residuals. Even the actual observed values of normal experimental samples generally deviate to some extent from the fitted values of their dose-effect trend function. This means that abnormal experimental samples are likely to be submerged in the general data fluctuations and difficult to identify clearly. Therefore, the... The weaker the discriminative power of a biological indicator, the smaller its weight in quality assessment.
[0043] S04: Determine the composite distance between any two experimental samples.
[0044] It should be noted that the machine learning LOF algorithm measures the differences between data using Euclidean distance. However, Euclidean distance cannot perceive pharmacological trends, which can lead to the misclassification of outliers that conform to the pattern. In order to improve this defect and optimize the distance measurement method of the machine learning model, this step calculates the composite distance based on the trend residual and the quality assessment weight.
[0045] The composite distance between two experimental samples is determined based on the Euclidean distance between any two experimental samples, the trend residual of each biological indicator in the two experimental samples, and the quality assessment weight of each biological indicator in the biological indicator assessment weight sequence.
[0046] Specifically, the composite distance satisfies: ; In the formula, For experimental samples With experimental samples The composite distance between them For experimental samples With experimental samples The Euclidean distance between them and Experimental samples With experimental samples The The trend residuals of each biological indicator The weighted sequence of biological indicators is used to evaluate the first... The quality assessment weight of each biological indicator The number of biological indicators.
[0047] in, This represents the basic distance between two experimental samples. The larger the value, the farther apart the two experimental samples are in the feature space, and the larger the composite distance between the two experimental samples. The smaller the value, the closer the two experimental samples are in the feature space, and the smaller the composite distance between the two experimental samples. Indicates that the two experimental samples are at the 1st... The difference between trend residuals of two biological indicators indicates the degree of deviation of the two experimental samples from the expected trend. For example, if one experimental sample conforms to the trend while the other deviates significantly, the difference between the two experimental samples is greater, and the composite distance between the two experimental samples is larger. The smaller the value, the more consistent the behavior of the two points. For example, if both experimental samples conform to the trend or deviate in a similar way, the difference between the two experimental samples is smaller, and the composite distance between the two experimental samples is smaller.
[0048] S05: Determine the quality assessment index for each experimental sample.
[0049] Based on the preset neighborhood parameter sequence and the composite distance between each experimental sample, the quality assessment index of each experimental sample is determined by the LOF algorithm.
[0050] It is important to note that in preclinical experimental data, the data density is naturally uneven across different groups. If a small neighborhood parameter is chosen, the machine learning LOF algorithm may become overly sensitive to sparsely dense regions, easily misclassifying normal but discrete data points as abnormal. Conversely, if a large neighborhood parameter is chosen, the machine learning LOF algorithm may ignore minor local anomalies within densely dense regions, leading to missed detections. Any single, fixed neighborhood parameter represents a compromise that cannot adequately consider the overall picture. Therefore, this step simultaneously examines the data points at multiple scales, calculating LOF scores under different neighborhood parameters, and obtaining a quality assessment index based on the LOF scores obtained under different neighborhood parameters.
[0051] Implementers can set the neighborhood parameter sequence according to the specific implementation situation, for example, .
[0052] Specifically, ; In the formula, For experimental samples The quality assessment index To use the preset neighborhood parameter sequence of the first Experimental samples obtained using the LOF algorithm with neighborhood parameters. Local anomalous factors, This refers to the number of neighborhood parameters in the preset neighborhood parameter sequence. It is a function with maximum value. It is a minimum value function.
[0053] in, The larger the sample size, the better. The more likely it is to be an outlier, the better the experimental sample. The higher the quality assessment index, the better; The smaller the value, the better the experimental sample size. The more likely it is to be a normal sample, the better the experimental sample. The smaller the quality assessment index, the better. The larger the value, the more experimental samples are under different scales of the LOF algorithm. The greater the difference between local outliers, the better for the experimental samples. The more likely an outlier a sample is to be easily overlooked, the more appropriate the experimental sample should be. The higher the quality assessment index, the greater the chance of success for the experimental samples. , therefore, through right Make an upward correction.
[0054] S06: Assess the quality of preclinical drug trials based on the quality assessment index for each experimental sample.
[0055] Specifically, the assessment of the quality of preclinical drug trials includes: If the quality assessment index of any experimental sample exceeds the preset quality threshold, it is determined that the drug corresponding to that experimental sample has a quality problem. At this time, special attention should be paid to this batch of drugs, and the preclinical drug trial quality assessment should be completed.
[0056] Implementers can set quality thresholds based on specific implementation conditions, such as the average quality assessment index of experimental samples corresponding to multiple normal drugs in history.
[0057] like Figure 2 and Figure 3 As shown, PCA dimensionality reduction was used to reduce the biological indicators from multiple dimensions to two dimensions for easier display. Figure 2 and Figure 3 The results of the traditional machine learning LOF algorithm and the improved machine learning LOF algorithm of this invention in the detection of abnormal samples in preclinical drug trial samples are presented respectively. The figure shows the abnormal identification of sample points by the two algorithms. It can be seen from the figure that the improved machine learning LOF algorithm of this invention can more accurately identify real abnormal samples and reduce the misjudgment or missed judgment problems that may occur in the traditional machine learning LOF algorithm.
[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A machine learning-based method for quality assessment of preclinical drug trials, characterized in that, include: Multiple experimental samples were obtained, each including dose levels and observations of at least one biological indicator; Based on the dose level, curve fitting is performed on each biological indicator to obtain the dose-effect trend function. The trend residual of each biological indicator of each experimental sample is determined according to the observed value of each biological indicator of each experimental sample and the fitted value of the corresponding dose-effect trend function. Based on the trend residual of each biological indicator, a sequence of biological indicator assessment weights, consisting of the quality assessment weights of all biological indicators, is determined. The composite distance between two experimental samples is determined based on the Euclidean distance between any two experimental samples, the trend residual of each biological indicator in the two experimental samples, and the quality assessment weight of each biological indicator in the biological indicator assessment weight sequence. Based on the preset neighborhood parameter sequence and the composite distance between each experimental sample, the quality assessment index of each experimental sample is determined by the LOF algorithm. The quality of preclinical drug trials is assessed based on the quality assessment index of each experimental sample.
2. The method for quality assessment of preclinical drug trials based on machine learning according to claim 1, characterized in that, The biological indicators include at least one of tumor volume, expression level of a specific protein, or body weight.
3. The method for quality assessment of preclinical drug trials based on machine learning according to claim 1, characterized in that, The observed values are those that have undergone Z-score normalization.
4. The machine learning-based preclinical drug trial quality assessment method according to claim 1, characterized in that, The trend residual satisfies: ； In the formula, For experimental samples The The trend residuals of each biological indicator For experimental samples The Observed values of biological indicators, For experimental samples dosage level, For experimental samples The The dose-response trend function corresponding to each biological indicator It is an absolute value function.
5. A machine learning-based method for preclinical drug trial quality assessment according to claim 1 or 4, characterized in that, The dose-effect trend function is obtained as follows: All experimental samples with the same dose level were grouped into one dose group; For each dose group, the dose-response trend function is obtained by curve fitting to each biological indicator using spline regression or a four-parameter logistic model.
6. The method for quality assessment of preclinical drug trials based on machine learning according to claim 1, characterized in that, The weight sequence for evaluating the biological indicators satisfies: ； In the formula, This is a sequence of biological indicator assessment weights, composed of the quality assessment weights of all biological indicators. For all experimental samples The sum of the trend residuals of each biological indicator For all experimental samples Kurtosis of the trend residual distribution of a biological indicator It is a normalized exponential function.
7. The method for quality assessment of preclinical drug trials based on machine learning according to claim 1, characterized in that, The composite distance satisfies: ； In the formula, For experimental samples With experimental samples The composite distance between them For experimental samples With experimental samples The Euclidean distance between them and Experimental samples With experimental samples The The trend residuals of each biological indicator The weighted sequence of biological indicators is used to evaluate the first... The quality assessment weight of each biological indicator The number of biological indicators.
8. The method for quality assessment of preclinical drug trials based on machine learning according to claim 1, characterized in that, The quality assessment index satisfies: ； In the formula, For experimental samples The quality assessment index To use the preset neighborhood parameter sequence of the first Experimental samples obtained using the LOF algorithm with neighborhood parameters. Local anomalous factors, This refers to the number of neighborhood parameters in the preset neighborhood parameter sequence. It is a function with maximum value. It is a minimum value function.
9. A machine learning-based method for preclinical drug trial quality assessment according to claim 1 or 8, characterized in that, The neighborhood parameter sequence is .
10. The machine learning-based preclinical drug trial quality assessment method according to claim 1, characterized in that, The assessment of the quality of preclinical drug trials includes: If the quality assessment index of any experimental sample exceeds a preset quality threshold, the drug corresponding to that experimental sample is deemed to have a quality problem, and the preclinical drug trial quality assessment is completed.