An aie fluorescent probe multi-index weighted scoring screening method and system

Abstract

The application provides a multi-index weighted scoring screening method and system for AIE fluorescent probes, and the method comprises the following steps: obtaining probe molecular structure, experimental environment and other data, and preprocessing to obtain a feature data set; for thermal stability, light stability, Stokes shift and AIE enhancement factor, the best molecular descriptor and binary classification prediction model are respectively optimized; the probe to be evaluated is input into the target model to obtain a prediction result, and a synthetic accessibility score is calculated; the prediction result is combined with the standardized synthetic accessibility score to form index data; the weight of each index is determined through mixed data factor analysis, and a comprehensive score is calculated to realize probe sorting and screening. The application can realize efficient screening of AIE probes, reduce trial and error cost, and enhance the stability and applicability of the score.

Description

Technical Field

[0001] This invention relates to the field of machine learning material design and screening technology, and in particular to a multi-index weighted scoring screening method and system for AIE fluorescent probes. Background Technology

[0002] AIE (Aggregation-Induced Emission) fluorescent probes, due to their enhanced luminescence in aggregated states, low background signal, and high detection sensitivity, show promising applications in chemical sensing, bioimaging, and environmental monitoring. With the continuous development of AIE molecular design, synthesis, and performance research, the number of candidate probe molecules continues to increase. How to rapidly screen target probes with superior overall performance and practical application value from a large number of candidate molecules has become one of the urgent problems to be solved in this field.

[0003] Current methods for screening AIE fluorescent probes primarily rely on experimental testing and human experience, typically requiring individual experimental verification and condition optimization for each target molecule's luminescence properties, stability, and other indicators. While these methods yield relatively direct experimental results, they generally suffer from long screening cycles, high experimental costs, and significant trial-and-error expenses when dealing with a large number of candidate molecules, numerous evaluation indicators, and complex experimental conditions. This makes them unsuitable for the practical needs of rapid design and iterative optimization of AIE fluorescent probes. Especially in real-world applications, the optimal selection of AIE fluorescent probes usually does not depend on a single performance indicator but requires comprehensive consideration of factors such as thermal stability, photostability, Stokes shift, AIE enhancement factor, and synthetic accessibility. Traditional screening methods relying on single tests or empirical rules struggle to provide a unified, objective, and efficient comprehensive evaluation of multiple key indicators.

[0004] With the gradual accumulation of experimental data and publicly available literature related to AIE (Alternative Electron Imaging) research, structure-to-property prediction methods based on machine learning have provided a new technical path for AIE fluorescent probe screening. However, existing technologies mostly focus on single-property prediction or local performance analysis, making it difficult to form a comprehensive multi-index scoring screening mechanism for AIE fluorescent probe selection. Furthermore, different property tasks have varying degrees of adaptability to molecular characterization methods and modeling strategies. If a unified descriptor or modeling method is used, the prediction effect of some properties may be limited, thus affecting the reliability of the comprehensive screening results.

[0005] Furthermore, when constructing a comprehensive evaluation system for AIE fluorescent probes, the scoring indicators typically include both continuous and discrete data. For example, synthesis accessibility scores are usually continuous variables, while whether key properties meet standards or corresponding prediction results may be discrete data or classification results. For such mixed-type indicator data, directly applying traditional weighting analysis methods applicable only to single data types can easily lead to insufficient utilization of some indicator information, thereby affecting the rationality of weight allocation and the stability and interpretability of the comprehensive scoring results.

[0006] Therefore, there is an urgent need for a multi-index weighted scoring method and system for AIE probe selection, which can achieve efficient sorting and screening of candidate AIE probes, reduce experimental trial and error costs, and enhance the stability and applicability of the scoring results. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-index weighted scoring method and system for AIE probe selection, aiming to solve the problems of difficulty in unified evaluation of multiple indicators, unreasonable weight allocation, and low screening efficiency in existing AIE fluorescent probe screening.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a multi-index weighted scoring screening method for AIE fluorescent probes, comprising the following steps:

[0009] S1. Obtain raw data including candidate AIE probe molecular structure information, experimental environment information and AIE property information, and perform data preprocessing and feature engineering on the raw data to obtain a feature dataset for modeling.

[0010] S2. For the four key properties of thermal stability, optical stability, Stokes shift and AIE enhancement factor, multiple molecular descriptor characterization schemes are constructed, and candidate binary classification prediction models are established based on the characterization results of different molecular descriptors. According to the model performance evaluation results, the corresponding optimal molecular descriptor and target property prediction model are determined for each key property.

[0011] S3. Input the candidate AIE probe to be evaluated into each of the target property prediction models to obtain the prediction results of the candidate AIE probe to be evaluated on the four key properties, and calculate the synthesis accessibility score of the candidate AIE probe to be evaluated.

[0012] S4. Construct comprehensive scoring index data based on the prediction results of the four key properties and the synthetic accessibility score, and standardize the synthetic accessibility score so that it has a consistent scoring direction with the prediction results of the four key properties.

[0013] S5. Perform mixed data factor analysis on the comprehensive scoring index data to determine the weight of each scoring index, and perform comprehensive scoring on the candidate AIE probes to be evaluated according to the weight of each scoring index, so as to sort and screen the candidate AIE probes according to the comprehensive score.

[0014] As a further improvement to the above technical solution, the molecular structure information includes at least the SMILES representation of the molecule, the experimental environment information includes at least solvent information, and the AIE property information includes at least thermal stability, photostability, and property data related to absorption or emission spectra.

[0015] As a further improvement to the above technical solution, in step S1, the feature engineering process includes at least molecular structure feature construction, solvent information numerical processing, and key property label construction.

[0016] As a further improvement to the above technical solution, the solvent information numerical processing is as follows: the solvent name is mapped to a preset solvent parameter to form a numerical feature that can be used for machine learning modeling.

[0017] The preset solvent parameters include at least one or more of Et(30), SP, SdP, SA and SB.

[0018] As a further improvement to the above technical solution, the construction of the key property tags includes:

[0019] Thermal stability is classified into two categories based on decomposition temperature. Preferably, when the decomposition temperature is... If the condition is met, it is classified as a positive class; otherwise, it is classified as a negative class.

[0020] Photostability is classified into two categories based on whether it decomposes, changes its spectrum, or changes its molecular structure under light conditions. When it decomposes upon exposure to light or changes its spectrum or molecular structure after exposure to light, it is classified as negative; otherwise, it is classified as positive.

[0021] The Stokes shift is calculated based on the absorption wavelength and the emission wavelength, and binary classification is performed according to a preset threshold. Preferably, the Stokes shift... According to the emission peak wavelength With absorption peak wavelength The difference is calculated, and 100nm is used as the binary classification threshold.

[0022] The AIE enhancement factor was calculated based on the luminescence intensity in the aggregated and solution states. And perform binary classification labeling based on a preset threshold. Preferably, the AIE enhancement factor... The luminescence intensity was calculated based on the ratio of the luminescence intensity in the aggregated state to that in the dispersed state, and a binary classification threshold of 10 was used.

[0023] As a further improvement to the above scheme, step S1 also includes: based on the data availability of the fields corresponding to the four key properties, the preprocessed data is organized into multiple modeling data subsets corresponding to thermal stability, optical stability, Stokes displacement and AIE enhancement factor, respectively.

[0024] As a further improvement to the above scheme, in step S2, the multiple molecular descriptor characterization schemes include at least two of the following: Morgan fingerprint, Daylight fingerprint, Atom-pair fingerprint, MACCS fingerprint, and the Description molecular descriptor set.

[0025] As a further improvement to the above scheme, in step S2, the performance of different molecular descriptor characterization schemes is compared through multiple random data partitioning conditions, and the target random data partitioning conditions for modeling various key properties are determined based on the comprehensive performance index.

[0026] As a further improvement to the above scheme, in step S2, the candidate binary classification prediction model includes at least two of the following: support vector machine classification model, logistic regression model, gradient boosting tree classification model, multilayer perceptron classification model, random forest classification model, and extreme gradient boosting classification model.

[0027] As a further improvement to the above scheme, in step S2, a unified data partitioning and verification process is adopted when evaluating the performance of the candidate binary classification prediction model to ensure that the evaluation results of different models and different molecular descriptor characterization schemes are comparable.

[0028] The performance evaluation metrics include at least accuracy (ACC), precision (Precision), recall (Recall), F1 score, area under the curve (AUC), and overall evaluation metric (OAI).

[0029] Preferably, the formulas for calculating accuracy (ACC), precision (Precision), recall (Recall), and F1 score are as follows:

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] Where TP represents the number of true positives, TN represents the number of true negatives, FP represents the number of false positives, and FN represents the number of false negatives; AUC is the area under the receiver operating characteristic (ROC) curve, used to characterize the overall discriminative ability of the model at different classification thresholds; the comprehensive evaluation index OAI satisfies:

[0035] .

[0036] As a further improvement to the above scheme, in step S2, the four key properties correspond to different optimal molecular descriptors and / or different target property prediction models.

[0037] As a further improvement to the above scheme, in step S3, the prediction results for the four key properties are the prediction probabilities output by the prediction models for each target property. The prediction probabilities include the prediction probability of thermal stability compliance, the prediction probability of optical stability compliance, the prediction probability of Stokes displacement compliance, and the prediction probability of AIE enhancement factor compliance.

[0038] As a further improvement to the above scheme, in step S3, the synthesis accessibility score is obtained by calculating the candidate AIE probe using cheminformatics tools, and is used as one of the comprehensive scoring indicators in the subsequent weighted scoring.

[0039] As a further improvement to the above scheme, in step S4, when standardizing the synthesis accessibility score, it includes normalization and direction alignment, so that the standardized synthesis accessibility score satisfies the prediction results of the four key properties. The larger the value, the better the overall performance of the candidate AIE probe.

[0040] As a further improvement to the above scheme, in step S5, the comprehensive scoring index data includes both continuous and discrete scoring variables, and the mixed data factor analysis is used to determine the weight of each scoring index under the mixed data conditions of continuous and discrete scoring variables.

[0041] As a further improvement to the above scheme, in step S5, the weight of each scoring index is determined according to the loading of each scoring index on the principal component and the explanation rate of the corresponding principal component, and the prediction results of the four key properties and the standardized synthesis accessibility score are linearly weighted based on the weight to obtain the comprehensive score of the candidate AIE probe.

[0042] As a further improvement to the above scheme, in step S5, the comprehensive score satisfies the following relationship:

[0043] The comprehensive score of the candidate AIE probe is obtained by multiplying the predicted probability of thermal stability compliance, the predicted probability of optical stability compliance, the predicted probability of Stokes shift compliance, the predicted probability of AIE enhancement factor compliance, and the standardized synthesis accessibility score by their respective weights and then summing them.

[0044] Secondly, the present invention also provides a multi-index weighted scoring system for AIE fluorescent probes, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the multi-index weighted scoring screening method for AIE fluorescent probes as described in the first aspect.

[0045] Because the present invention adopts the above technical solutions, the beneficial effects of the present invention are as follows:

[0046] This invention provides a multi-index weighted scoring screening method for AIE fluorescent probes. By predicting multiple key properties of AIE probes and combining them with synthesis accessibility for comprehensive evaluation, the method achieves the ranking and screening of candidate AIE probes. This improves screening efficiency and reduces experimental trial and error costs while taking into account both key performance and synthesis feasibility.

[0047] Specifically, firstly, this invention acquires the molecular structure information, experimental environment information, and AIE property information of candidate AIE probes, and performs data preprocessing and feature engineering on the raw data to form a feature dataset for modeling. This enables the unified organization and effective utilization of scattered structural, environmental, and property information, providing a more consistent data foundation for subsequent multi-index prediction and comprehensive screening, which is beneficial to improving the standardization and feasibility of the screening process.

[0048] Secondly, this invention constructs multiple molecular descriptor characterization schemes for four key properties: thermal stability, optical stability, Stokes shift, and AIE enhancement factor, and establishes candidate binary classification prediction models for each. Based on the model performance evaluation results, the optimal molecular descriptor and target property prediction model are determined for each key property. Since different properties are not entirely adaptable to molecular characterization methods and modeling strategies, by optimizing descriptors and models separately, it is possible to avoid the limitation of a unified descriptor or model on the prediction of some properties, thereby improving the relevance of the prediction results for multiple key properties and the reliability of the comprehensive evaluation.

[0049] Furthermore, this invention incorporates a synthesis accessibility score into the comprehensive evaluation while obtaining prediction results for four key properties. Since the optimal selection of AIE probes involves not only performance characteristics but also practical synthesis and application feasibility, combining the key property prediction results with synthesis accessibility information ensures that the screening results not only focus on simple performance differences but also consider the actual feasibility of candidate molecules, thereby enhancing the application reference value of the screening results.

[0050] Furthermore, this invention constructs comprehensive scoring index data based on the prediction results of four key properties and the synthetic accessibility score, and standardizes the synthetic accessibility score to ensure it has a consistent scoring direction with the prediction results of each key property. Through the above processing, indicators from different sources, with different dimensions, and with different evaluation directions can be compared and integrated under a unified evaluation framework, thereby helping to solve the problem of difficulty in unified evaluation of multiple indicators in the prior art.

[0051] Furthermore, this invention employs hybrid data factor analysis on the comprehensive scoring index data to determine the weights corresponding to each scoring index, and accordingly performs comprehensive scoring and ranking screening of candidate AIE probes. Since the data involved in the evaluation simultaneously includes property prediction results and synthetic accessibility scores, representing comprehensive indicators with different data characteristics, using hybrid data factor analysis to determine weights can, to some extent, reduce the arbitrariness caused by relying solely on subjective experience for weight assignment. This makes the weight allocation more aligned with the actual role of each indicator in the comprehensive evaluation, thereby improving the rationality and stability of the comprehensive scoring results.

[0052] This invention, through a technical solution of "separate modeling of multiple properties, introduction of synthetic accessibility, standardization of indicators, weighting of mixed data, and comprehensive scoring screening," can effectively achieve unified evaluation of multiple indicators of AIE probes and ranking and screening of candidate molecules, thereby improving screening efficiency and reducing the experimental trial-and-error burden. Attached Figure Description

[0053] 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 the structures shown in these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating a multi-index weighted scoring screening method for AIE fluorescent probes disclosed in this invention.

[0055] Figure 2This is a schematic diagram showing the performance comparison of molecular descriptors for four key properties tasks under the default random forest model condition, based on the multi-index weighted scoring screening method for AIE fluorescent probes disclosed in this invention.

[0056] Figure 3 This is a schematic diagram comparing the performance of candidate models under four key property tasks disclosed in this invention, wherein... Figure 3 (a) represents the thermal stability of the candidate AIE probe. Predictive tasks, Figure 3 (b) represents the photostability of the candidate AIE probe. Predictive tasks, Figure 3 (c) represents the Stokes displacement of the candidate AIE probe. Predictive tasks, Figure 3 (d) represents the AIE enhancement factor of the candidate AIE probe. Predictive tasks;

[0057] Figure 4 This is a line graph showing the cumulative explanatory power of the FAMD weighted analysis disclosed in this invention.

[0058] Figure 5 This is a schematic diagram showing the relative weight distribution of each comprehensive scoring index in mixed data factor analysis under different principal component numbers (K=1-5) conditions disclosed in this invention.

[0059] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0061] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0062] Example 1

[0063] See Figures 1-5This invention provides a multi-index weighted scoring screening method for AIE fluorescent probes, suitable for ranking and optimizing candidate AIE fluorescent probes, to solve the problems of difficulty in uniformly evaluating multiple indicators, unreasonable weight allocation, and low screening efficiency in existing AIE fluorescent probe screening processes. The specific implementation steps of the screening method are as follows:

[0064] S1. Raw Data Acquisition and Feature Dataset Construction:

[0065] First, obtain the raw data for the candidate AIE fluorescent probe. This raw data can be derived from public databases, published literature, experimental records, or a combination of these sources. The raw data should include at least the following three types of information:

[0066] Molecular structural information, such as the SMILES expression, structural formula, or molecular representation information that can be converted into structural features of candidate AIE fluorescent probes;

[0067] Experimental environment information, such as the type of solvent used in the test, solvent parameters, or environmental information related to the luminescence test conditions;

[0068] AIE property information, such as thermal stability, photostability, absorption peak position, emission peak position, aggregated emission intensity, and dispersed emission intensity, can characterize probe performance.

[0069] After acquiring the raw data, data preprocessing is performed. This preprocessing may include: removing duplicate samples, standardizing data formats, handling missing items, identifying outliers, and cleaning up invalid records. By standardizing and organizing the raw data, the interference of data noise on subsequent modeling can be reduced, providing a more consistent data foundation for subsequent multi-indicator modeling.

[0070] Based on the preprocessing, further feature engineering is performed. This feature engineering process may include:

[0071] Convert molecular structure information into structural features that can be used for machine learning modeling;

[0072] Convert experimental environment information into numerical environmental features;

[0073] Extract or calculate key evaluation indicators from raw property information;

[0074] Label data required to build subsequent classification models.

[0075] For example, smiles of candidate AIE fluorescent probes can be input into cheminformatics tools to generate various molecular descriptors or molecular fingerprints related to molecular structure. For solvent information, numerical mapping can be performed based on solvent polarity parameters, hydrogen bond acceptor parameters, or other characterization parameters to incorporate experimental environment differences into the model. By incorporating both molecular structure information and experimental environment information into the modeling features, subsequent models can consider not only the molecular structure itself but also the influence of the testing environment on property performance, thereby improving the applicability of property predictions.

[0076] S2, Construction of Key Property Labels:

[0077] For the task of optimizing AIE fluorescent probes, four key properties are selected as the basis for comprehensive evaluation: thermal stability, photostability, Stokes shift, and AIE enhancement factor.

[0078] The thermal stability label can be constructed based on decomposition temperature or thermogravimetric indicators. For example, a sample can be marked as compliant or non-compliant based on whether it exceeds a preset thermal stability threshold. This process facilitates the conversion of raw thermal test results into a unified binary evaluation result, making it easier to integrate with other properties into a comprehensive scoring system.

[0079] Photostable labels can be constructed based on whether a sample undergoes significant decomposition, luminescence attenuation, spectral changes, or structural changes under illumination. For example, samples that meet preset photostable conditions can be labeled as stable, while samples that do not meet the conditions can be labeled as unstable. By labeling photostable performance, this indicator can be used in subsequent modeling in a unified manner.

[0080] Stokes shift can be calculated based on the difference between the emission peak wavelength and the absorption peak wavelength, and then classified into two categories according to a preset threshold. This method can convert the original spectral parameters into property labels more suitable for the optimization task, making the model output closer to the goal of whether it meets the application requirements.

[0081] The AIE enhancement factor can be calculated based on the ratio of aggregated emission intensity to dispersed emission intensity, and labels are classified according to a preset threshold. This step transforms the core indicator of AIE characteristics into a directly modelable discrimination result, which is beneficial for the rapid screening of candidate probes in the subsequent process.

[0082] Through the above processing, raw property data from different sources and with different forms of expression can be uniformly organized into modelable property labels or property indicators, thus providing a foundation for subsequent parallel modeling of multiple properties.

[0083] S3. Construction of property descriptors and selection of target model:

[0084] Since different key properties are not equally sensitive to molecular structural features and environmental features, this embodiment does not use a single descriptor and a single model to uniformly process all properties. Instead, it constructs multiple descriptor characterization schemes for different properties and conducts model training and performance comparisons separately.

[0085] Specifically, for each key property, multiple candidate molecular descriptor characterization schemes can be constructed. These descriptor characterization schemes may include, but are not limited to, Morgan fingerprints, Daylight fingerprints, Atom-pair fingerprints, MACCS fingerprints, two-dimensional molecular descriptors, three-dimensional molecular descriptors, or combinations thereof. Comparing multiple molecular characterization schemes for the same property helps to find a feature representation method more suitable for the prediction task of that property, avoiding the problem of a unified descriptor being insufficient for representing some properties.

[0086] After obtaining the representation results of various descriptors, candidate binary classification prediction models are established. These candidate models may include one or more classification models such as Support Vector Machine, Logistic Regression, Random Forest, Gradient Boosting Tree, Extreme Gradient Boosting, and Multilayer Perceptron. For each key property, different descriptor-model combinations can be trained, validated, and their performance evaluated under the same data partitioning conditions.

[0087] The performance evaluation can employ one or more of the following metrics: accuracy, precision, recall, F1 score, AUC score, or a comprehensive evaluation index. Preferably, under the same property task, after uniformly evaluating multiple descriptor-model combinations, the one with the best performance is selected as the optimal molecular descriptor and target property prediction model corresponding to that property.

[0088] S4. Prediction of the properties of candidate probes to be evaluated and calculation of their synthetic accessibility:

[0089] After optimizing the target model for the four key properties, the candidate AIE fluorescent probes to be evaluated are input into the corresponding target property prediction model to obtain their prediction results for thermal stability, photostability, Stokes shift, and AIE enhancement factor.

[0090] The prediction results can be expressed as classification labels, or preferably as predicted probabilities. When using predicted probabilities as model output, it can reflect the degree to which candidate molecules meet the corresponding property requirements in more detail, rather than being limited to a simple "yes / no" judgment, and is therefore more suitable as a comprehensive scoring input.

[0091] In addition to the predicted results of four key properties, this embodiment also calculates the synthetic accessibility score of the candidate AIE fluorescent probes to be evaluated. The synthetic accessibility score can be obtained using existing cheminformatics evaluation tools or rule-based algorithms, and is used to characterize the ease or difficulty of actually synthesizing the candidate molecules. By incorporating synthetic accessibility into the evaluation system, the comprehensive score no longer only reflects theoretical performance but also, to a certain extent, takes into account practical preparation feasibility, which helps to improve the application reference value of the screening results.

[0092] S5. Construction and Standardization of Comprehensive Scoring Indicators:

[0093] After obtaining the prediction results for four key properties and the synthesis accessibility score, a comprehensive scoring index data is constructed. The comprehensive scoring index data includes at least: thermal stability prediction results, optical stability prediction results, Stokes displacement prediction results, AIE enhancement factor prediction results, and synthesis accessibility score.

[0094] Because the numerical sources, dimensions, and evaluation directions of the aforementioned indicators differ, this embodiment standardizes the synthetic accessibility score and aligns it with the scoring direction of the four key property prediction results. In other words, the processed index values ​​can all be uniformly represented as "the larger the value, the better the overall evaluation" or other consistent directions.

[0095] S6. Weight determination based on factor analysis of mixed data:

[0096] In this embodiment, the comprehensive scoring index data contains different types of data. For example, the prediction results of the four key properties can be expressed as probability values, classification results, or evaluation values ​​derived from them, while the synthetic accessibility score is usually continuous data. Therefore, the comprehensive scoring index data belongs to a dataset with mixed-type characteristics.

[0097] To address this characteristic, this embodiment employs mixed data factor analysis to perform weight analysis on the comprehensive scoring index data. Mixed data factor analysis can be understood as a method for factor extraction and contribution analysis when both continuous and categorical variables exist simultaneously, used to evaluate the relative roles of each index in the comprehensive evaluation within a unified analytical framework.

[0098] In practice, the standardized indicators can be input into the mixed data factor analysis module. The weights of each scoring indicator can be determined using principal component contribution rates, loading coefficients, or other statistical measures. Compared to directly using empirical weighting or simple average weighting, this method determines indicator weights based on the structural relationships of the data itself, thereby reducing the arbitrariness of subjective weighting and making the weight allocation more closely reflect the actual contribution of each indicator in the comprehensive evaluation.

[0099] By adopting this weighting method, the rationality, stability and interpretability of the comprehensive scoring results can be improved to a certain extent, which is especially suitable for AIE fluorescent probe selection scenarios that simultaneously include different types of evaluation information.

[0100] S7. Comprehensive score calculation and sorting:

[0101] After determining the weights of each scoring indicator, a comprehensive score is calculated for the candidate AIE fluorescent probes to be evaluated. This comprehensive score can be obtained by weighted summation of each indicator with its corresponding weight, or by using other equivalent weighted aggregation methods.

[0102] For example, if the prediction results of the four key properties are denoted as the first score, the second score, the third score, and the fourth score, respectively, and the standardized synthesis accessibility score is denoted as the fifth score, with the corresponding weights denoted as the first weight, the second weight, the third weight, the fourth weight, and the fifth weight, respectively, then the comprehensive score of the candidate AIE fluorescent probe can be expressed as:

[0103] The overall score is calculated as follows: First weight × First score value + Second weight × Second score value + Third weight × Third score value + Fourth weight × Fourth score value + Fifth weight × Fifth score value.

[0104] After ranking multiple candidate AIE fluorescent probes based on their comprehensive scores, the candidate molecules with higher scores can be selected for subsequent experimental verification. This approach transforms the screening process, which previously required comparing multiple indicators one by one, into a priority selection process that first calculates and ranks the probes and then performs focused verification. This helps narrow down the scope of experimental verification, improves screening efficiency, and reduces trial-and-error costs.

[0105] As a preferred embodiment, the composition of the original data and the feature engineering process in step S1 will be further explained.

[0106] In this embodiment, the molecular structure information includes at least the SMILES representation of the candidate AIE fluorescent probe molecule, the experimental environment information includes at least solvent information, and the AIE property information includes at least thermal stability, photostability, and property data related to absorption or emission spectra.

[0107] SMILES indicates that the connection relationships and basic structural features of candidate molecules can be characterized in a unified and standardized string format, which facilitates batch reading, storage, and subsequent computer processing. Solvent information reflects the differences in external conditions of candidate AIE fluorescent probes in the testing or application environment. Since the luminescence performance, spectral position, and some stability indicators of AIE probes are usually affected by environmental factors, including solvent information in the raw data helps to improve the completeness of subsequent property analysis and modeling. Thermal stability, photostability, and absorption or emission spectra correspond to the stability and spectral performance indicators that are of great concern in the actual selection process of AIE probes. Using these as basic property information inputs is beneficial for providing a data foundation for subsequent unified evaluation of multiple indicators.

[0108] Furthermore, in step S1, the feature engineering process includes at least molecular structure feature construction, solvent information numerical processing, and key property label construction.

[0109] The construction of molecular structural features refers to converting the SMILES representation of a molecule into structural features suitable for processing by machine learning models. Specifically, cheminformatics tools can be used to parse the SMILES, extracting structural characterization information such as the molecular skeleton, atomic connections, functional group composition, topological relationships, molecular fingerprints, or molecular descriptors, and converting this information into vectorized features. This step transforms the raw molecular structural information from string format into numerical features that the model can recognize and compute, thus providing a unified input basis for subsequent predictive modeling of different properties. Compared to directly using unstructured raw representations, this approach improves the utilization of structural information, enabling subsequent models to more effectively identify potential structural patterns related to thermal stability, optical stability, and spectral properties.

[0110] The numerical processing of solvent information refers to further mapping the original solvent name to preset solvent parameters to form numerical features that can be used for machine learning modeling. Preferably, the preset solvent parameters include at least one or more of Et(30), SP, SdP, SA, and SB. Among them, Et(30) can be used to characterize solvent polarity-related features, while SP, SdP, SA, and SB can reflect the solvent's solubility parameters or acceptor properties from different perspectives. By converting solvent names into quantifiable parameters, rather than simply treating them as category labels, the model can more finely distinguish the impact of different solvent environments on the performance of AIE probes. In this way, on the one hand, it can avoid the situation where similar solvent names are only represented as discrete symbols in the model, making it difficult to reflect environmental differences; on the other hand, it is also beneficial to incorporate experimental environmental factors and molecular structure factors into the analysis framework, thereby improving the consistency between subsequent property prediction results and actual application environments.

[0111] Furthermore, the construction of key property labels refers to the labeling of key properties related to the optimization task under unified rules based on the original AIE property information. Specifically, thermal stability labels can be constructed based on thermal stability test results, photostability labels can be constructed based on property changes before and after illumination, and indices such as Stokes shift can be further calculated based on absorption and emission spectrum data. If necessary, AIE enhancement-related evaluation results can also be constructed by combining aggregated and dispersed luminescence intensities. Through this step, original experimental results from different sources, with different dimensions, and different forms of expression can be converted into standardized property representations required for subsequent model training and unified scoring. Such processing helps to solve the problem in existing technologies where different property indices are difficult to directly incorporate into the same evaluation process, enabling multiple properties to be subsequently predicted and comprehensively analyzed within a unified data framework.

[0112] In this embodiment, molecular structure feature construction, solvent information numerical processing, and key property label construction work in tandem. Specifically, molecular structure feature construction focuses on extracting the intrinsic structural information of the candidate AIE fluorescent probe itself, solvent information numerical processing focuses on introducing external experimental environmental factors, and key property label construction transforms the raw test results into target outputs for the screening task. By using the above three types of processing results together for subsequent modeling, the feature dataset can simultaneously contain information at three levels: "molecular intrinsic structure—external environmental conditions—target property performance." Compared to simple screening based solely on molecular structure, this processing method is beneficial for improving the adaptability of the subsequent model to actual screening scenarios and provides more complete data support for separate modeling of different properties and subsequent comprehensive scoring.

[0113] By standardizing the original data format and constructing structural and environmental features, data dispersion and invalid differences in the subsequent multi-property modeling process can be reduced, which is beneficial to improving the comparability of prediction results for each property. Secondly, by labeling and structuring the original property data, the originally scattered thermal, optical, and environmental data can be incorporated into a unified evaluation process, providing a basis for subsequent comprehensive scoring. Thirdly, the parameterized expression of solvent information allows experimental environmental factors to participate in the analysis in a quantitative manner, thus making the screening results not limited to the judgment at the ideal structural level, but closer to the actual testing and application conditions. Therefore, the feature dataset formed in step S1 can provide a relatively stable data foundation for subsequent multi-index prediction, unified scoring, and ranking screening, which helps to improve the efficiency and feasibility of the entire screening process.

[0114] In a preferred embodiment, the construction of key property labels includes at least the construction of thermal stability labels, optical stability labels, Stokes displacement labels, and AIE enhancement factor labels.

[0115] Thermal stability labeling: For thermal stability, binary labeling based on the decomposition temperature of candidate AIE fluorescent probes is preferred. Specifically, when the decomposition temperature... At temperatures above 200℃, it is classified as a positive class; when the decomposition temperature... Temperatures less than or equal to 200℃ are classified as negative.

[0116] Decomposition temperature directly reflects the stability of candidate AIE fluorescent probes under heated conditions, and thermal stability is one of the most important fundamental properties in practical applications. By converting the raw decomposition temperature into a binary label, on the one hand, continuous thermal test results can be transformed into target outputs that can be directly used by subsequent classification models; on the other hand, it also helps to distinguish between candidate molecules that "meet the requirements" and "do not meet the requirements" according to a unified standard, thereby improving the clarity of the subsequent screening process.

[0117] Photostability labeling: For photostability, a binary classification labeling method is preferred based on whether candidate AIE fluorescent probes undergo decomposition, spectral changes, or molecular structural changes under illumination. Specifically, samples that decompose upon exposure to light, or exhibit significant spectral or molecular structural changes after illumination, are classified as negative; samples that do not undergo these changes under illumination are classified as positive.

[0118] Photostability not only affects the continued usability of AIE fluorescent probes in actual detection and imaging processes, but also directly impacts the reliability of their application results. Judging solely based on a single optical parameter often fails to comprehensively reflect the stability of candidate molecules under illumination. Therefore, this embodiment uses photodecomposition, spectral changes, and molecular structure changes as unified criteria; any adverse change among these is considered a failure to meet photostability requirements. This approach makes the photostability criteria more aligned with practical screening needs. Furthermore, uniformly labeling photostability as positive or negative transforms previously scattered experimental observations into clear machine learning labels, reducing reliance on unstructured experimental records during subsequent modeling, improving data processing efficiency, and enhancing the comparability of photostability results with other property indicators.

[0119] Stokes shift label construction: For the Stokes shift, it is preferable to calculate it based on the absorption wavelength and the emission wavelength, and then perform binary classification labeling according to a preset threshold. Preferably, the Stokes shift Δλ is calculated according to the difference between the emission peak wavelength λem and the absorption peak wavelength λabs, that is:

[0120] Δλ = λem - λabs.

[0121] After obtaining the Stokes shift, 100 nm is preferably used as the binary classification threshold; when Δλ is greater than or equal to 100 nm, it is determined to be a positive class; when Δλ is less than 100 nm, it is determined to be a negative class.

[0122] Stokes shift reflects the spectral spacing between absorption and emission, and its magnitude typically affects self-absorption interference, signal resolution, and actual detection performance. In the selection of AIE fluorescent probes, a larger Stokes shift is generally more beneficial for improving signal discrimination. Therefore, in this embodiment, the difference between the absorption and emission peaks is used as a unified calculation index, and a binary classification label is constructed using a threshold method to directly reflect whether candidate AIE fluorescent probes meet the predetermined requirements in this property. By further calculating the Stokes shift from the raw spectral data and labeling it according to a unified threshold, spectral experimental results from different sources can be transformed into standardized evaluation results. This avoids the inefficient method of manually analyzing the raw spectral data item by item in subsequent comprehensive evaluations, and helps improve the standardization and processing efficiency of the screening process.

[0123] AIE Enhancement Factor Label Construction: For the AIE enhancement factor, it is preferable to calculate it based on the luminescence intensity in the aggregated state and the solution state, and then perform binary classification labeling according to a preset threshold. Preferably, the AIE enhancement factor α is calculated as the ratio of the luminescence intensity in the aggregated state to the luminescence intensity in the dispersed state, i.e.:

[0124] α=I agg / I sol ;

[0125] Among them, I agg I represents the luminescence intensity in the aggregated state. sol This indicates the intensity of light emitted in the dispersed or solution state.

[0126] After obtaining the AIE enhancement factor, it is preferred to use 10 as the binary classification threshold; when α is greater than or equal to 10, it is determined to be a positive class; when α is less than 10, it is determined to be a negative class.

[0127] The aggregation-induced emission (AIE) enhancement factor directly characterizes the degree of luminescence enhancement of candidate molecules under aggregated conditions relative to dispersed conditions, and is an important indicator reflecting AIE properties. By calculating the ratio of luminescence intensity in aggregated and dispersed states, the original luminescence experimental results can be transformed into evaluation parameters with clear physical meaning. Further, by using a threshold for binary classification, this parameter can be converted into a unified label format suitable for classification model training and subsequent comprehensive scoring. This approach incorporates the core aggregation-induced emission characteristics of AIE fluorescent probes into a unified screening framework, ensuring that subsequent comprehensive evaluation considers not only stability and spectral characteristics but also the intrinsic performance of AIE, thereby improving the consistency between the screening results and the actual preferred targets.

[0128] In this embodiment, thermal stability, optical stability, Stokes shift, and AIE enhancement factor are derived from thermal tests, illumination experiments, absorption / emission spectral data, and luminescence intensity data, respectively, and their original expressions differ significantly. Directly using the original experimental values ​​for comprehensive analysis not only makes it difficult to form a unified model input target but also easily affects the rationality of subsequent scoring results due to differences in dimensions and recording methods. Through the aforementioned unified processing, the original experimental data can be transformed into label results with consistent rules and clear semantics, enabling different properties to enter the subsequent model training and scoring analysis process at the same data expression level. This approach not only improves the comparability between multi-property data but also helps reduce the subjectivity and inefficiency caused by manual item-by-item judgment, thus providing a foundation for subsequent unified evaluation of multiple indicators and efficient ranking and screening.

[0129] In a preferred embodiment, step S1 further includes: based on the data availability of the fields corresponding to the four key properties of thermal stability, optical stability, Stokes displacement and AIE enhancement factor, the preprocessed data is organized into multiple modeling data subsets corresponding to thermal stability, optical stability, Stokes displacement and AIE enhancement factor, respectively.

[0130] Specifically, after deduplication, format standardization, missing item cleanup, and basic feature construction of the original data, each field related to the four key properties in each sample was checked item by item. Samples with decomposition temperature and the information required for its determination were assigned to the thermal stability modeling data subset; samples with records of decomposition, spectral changes, or structural changes under illumination were assigned to the photostability modeling data subset; samples with absorption peak wavelength and emission peak wavelength information and capable of calculating the Stokes shift were assigned to the Stokes shift modeling data subset; and samples with aggregated and dispersed luminescence intensity information and capable of calculating the AIE enhancement factor were assigned to the AIE enhancement factor modeling data subset.

[0131] In other words, this embodiment does not require every sample to have all four key properties simultaneously. Instead, it organizes the samples into the corresponding task datasets based on the actual data availability for different types of tasks. The advantage of this approach is that it can fully utilize the existing effective information in the original data, avoid discarding samples entirely due to missing individual fields, and thus improve data utilization.

[0132] As a preferred embodiment, the process of constructing molecular descriptors, comparing data partitions, and establishing candidate models in step S2 will be further explained.

[0133] In this embodiment, considering that the structural influencing factors corresponding to the four key properties—thermal stability, optical stability, Stokes shift, and AIE enhancement factor—are not entirely the same, using a single molecular characterization method and a single modeling condition for all properties could easily lead to insufficient expression of the structural features of some properties, thereby affecting the reliability of the subsequent comprehensive evaluation results. Therefore, in step S2, it is preferable to construct multiple molecular descriptor characterization schemes for different key properties and compare them under different data partitioning conditions and different model conditions to select a more suitable characterization method and prediction model for each key property.

[0134] Construction of multiple molecular descriptor characterization schemes: The multiple molecular descriptor characterization schemes include at least two of the following: Morgan fingerprint, Daylight fingerprint, Atom-pair fingerprint, MACCS fingerprint, and Description molecular descriptor set.

[0135] Among them, Morgan fingerprints can be used to characterize local structural features centered on atomic neighborhoods, suitable for reflecting structural information related to local substitution environments in candidate AIE fluorescent probes; Daylight fingerprints can be used to characterize path features, helping to depict connection paths and local fragment combination relationships in the molecular skeleton; Atom-pair fingerprints can be used to reflect atomic pairs and their topological distance relationships, which is beneficial for describing structural associations between distant positions within a molecule; MACCS fingerprints can encode common functional groups or substructures in a molecule based on preset structural bond values, facilitating the rapid acquisition of a unified format of structural feature representations. The Description molecular descriptor set may include one or more of the following: molecular weight, topological polar surface area, number of hydrogen bond donors, number of hydrogen bond acceptors, number of aromatic rings, number of rotatable bonds, LogP value, or other two-dimensional and three-dimensional descriptors.

[0136] In practical implementation, for each candidate AIE fluorescent probe sample's SMILES representation, cheminformatics computing tools can be used to generate the aforementioned different types of fingerprint features or descriptor features, thereby forming multiple sets of molecular characterization results. If necessary, two or more molecular fingerprints or descriptors can be spliced ​​or combined to form a composite descriptor characterization scheme. By simultaneously constructing multiple molecular descriptor characterization schemes, molecular information can be examined from different structural representation perspectives under the same property task, helping to reduce the loss of structural information caused by the limitations of a single characterization method.

[0137] Different key properties exhibit varying sensitivities to molecular structural features; some rely more heavily on local functional group information, while others depend more on the overall conjugated framework or topological relationships. Therefore, introducing and comparing multiple descriptor schemes can improve the sufficiency of structural feature extraction, providing a foundation for building more adaptable prediction models for different properties. This approach helps address the problem in existing technologies where a unified descriptor struggles to adequately predict multiple properties.

[0138] See Figure 2 This is a diagram showing the performance comparison of different molecular descriptors on four key property tasks under the default random forest model. The horizontal axis represents the classification and prediction tasks of the four key properties of candidate AIE probes, in order: thermal stability. Light stability Stokes displacement and AIE enhancement factor The vertical axis represents the median of the overall evaluation index; the higher the value, the better the predictive performance of the molecular descriptor in this property task.

[0139] Figure 2 The bar charts in the text compare six different molecular descriptors and fingerprint characterization schemes, specifically: Morgan fingerprint (Morgan, light blue bar), MACCS fingerprint (MACCS, green bar), Torsion fingerprint (Torsion, dark blue bar), Daylight fingerprint (RD, red bar), Atom-pair fingerprint (AP, gray bar), and Description molecular descriptor set (DES, yellow bar).

[0140] By comparison Figure 2 The height distribution of each column leads to the following conclusions: For different key property tasks, the performance of each molecular descriptor varies significantly. Specifically: in terms of thermal stability... In the prediction task, Morgan and Torsion fingerprints showed relatively high median OAI, demonstrating superior characterization ability; in terms of photostability... In prediction tasks, Daylight fingerprints (RD) significantly outperform other descriptors; in Stokes shift... In the prediction task, the combined performance of MACCS fingerprint and Daylight fingerprint (RD) is leading; in the AIE enhancement factor... In the prediction task, despite the overall low OAI baseline, Morgan fingerprint still demonstrated the best prediction performance. This performance comparison chart intuitively proves that different optical or physicochemical properties of AIE probes have varying sensitivities to characterizing molecular structures, and that using a single molecular descriptor cannot achieve optimal results in all property predictions.

[0141] Random data partitioning condition comparison: In order to reduce the random impact of a single data partitioning method on the model evaluation results, in step S2, it is preferable to compare the performance of different molecular descriptor characterization schemes through multiple random data partitioning conditions, and determine the target random data partitioning conditions for modeling various key properties based on the comprehensive performance index.

[0142] Specifically, multiple random data partitioning conditions can be set for each subset of data used to model key properties. For example, different random seeds, different training and test set ratios can be used, or different random sampling methods can be used to form multiple sets of training and test data under the same partitioning ratio. For each data partitioning condition, model training and validation are performed based on different molecular descriptor representation schemes, and the corresponding performance evaluation results are recorded.

[0143] Preferably, 100 random seeds are selected by sampling at intervals of 10 within the range of 1 to 1001. For each random seed, different molecular descriptors are modeled under the default random forest model conditions, and the comprehensive evaluation index (OAI) is calculated. After summarizing the performance statistics of each descriptor under multiple random seed conditions, the random seed that is closest to the average comprehensive performance index of multiple random seeds is selected as the target random seed for subsequent model construction and performance comparison for this key property task.

[0144] By first comparing different data partitioning conditions and then optimizing the model under the better partitioning conditions, subsequent comparisons of molecular descriptors and models can be based on relatively stable data, thereby improving the reliability of the modeling results for various key properties. This approach plays a positive role in addressing the problem that screening results are significantly affected by a single sample partitioning.

[0145] Construction of candidate binary classification prediction models: The candidate binary classification prediction models include at least two of the following: support vector machine classification model, logistic regression model, gradient boosting tree classification model, multilayer perceptron classification model, random forest classification model, and extreme gradient boosting classification model.

[0146] Specifically, after obtaining the representation results of various molecular descriptors corresponding to each property task, these can be input into different candidate binary classification prediction models for training. Support vector machine (SVM) classification models are suitable for handling classification problems in high-dimensional feature spaces and are beneficial for identifying the boundaries between samples of different classes; logistic regression models have a relatively clear structure, making it easy to obtain basic classification results; gradient boosting tree (GPRS) classification models and extreme gradient boosting (EPR) classification models can, to some extent, characterize the nonlinear relationships between features; random forest (RMR) classification models can improve classification robustness through the ensemble of multiple decision trees; and multilayer perceptron (MPP) classification models are suitable for learning more complex nonlinear mapping relationships.

[0147] In practical implementation, for each key property, under the condition of randomized partitioning of the target data, different molecular descriptor representation schemes can be input into the aforementioned candidate models for training and validation. For example, for the thermal stability task, support vector machine classification models, logistic regression models, random forest classification models, and extreme gradient boosting classification models can be built based on Morgan fingerprints; simultaneously, corresponding models can also be built based on Daylight fingerprints, Atom-pair fingerprints, MACCS fingerprints, and molecular descriptor sets. For the photostability, Stokes shift, and AIE enhancement factor tasks, multiple sets of candidate models can also be constructed in the same way. Preferably, the training set / test set partition ratio is 0.75 / 0.25, and 5-fold cross-validation is used to evaluate the model performance.

[0148] See Figure 3 The horizontal axis of each subplot lists the six candidate binary classification prediction models participating in the performance comparison, specifically: Multilayer Perceptron (MLP), Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Regression Tree (GBRT), and Extreme Gradient Boosting (XGB). The vertical axis of each subplot represents the overall performance of the candidate models, measured by the Overall Evaluation Index (OAI). Higher bars indicate better predictive performance in the task. By comparing the bar height distribution in each subplot, it can be seen that: in terms of thermal stability... In prediction tasks, the Support Vector Machine (SVM) classification model performed best in overall evaluation metrics, followed by the Multilayer Perceptron (MLP); in terms of light stability... In the prediction task, the Support Vector Machine (SVM) classification model showed a significant performance advantage, clearly outperforming the other five candidate models; in the Stokes shift... In the prediction task, the Support Vector Machine (SVM) classification model still achieved the highest score, followed closely by the Random Forest (RF) classification model; in the AIE enhancement factor... In prediction tasks, performance trends change, and the Random Forest (RF) classification model performs better, becoming the highest-scoring classification model. Figure 3 The performance comparison chart clearly illustrates that the optimal model type differs for different key property tasks. This is because the data distribution characteristics and mapping rules to the feature space differ fundamentally for different property tasks. This invention does not adopt the conventional approach of "a single model covering all tasks," but instead determines the most suitable target property prediction model based on the actual model performance evaluation results for each task. For example, the SVM model is used for the first three properties, while the RF model is used for the AIE enhancement factor. This selection strategy effectively avoids performance compromises during multi-task modeling and provides the highest confidence prediction input results for the subsequent comprehensive scoring module.

[0149] As a preferred embodiment, in step S2, the candidate binary classification prediction model is preferably evaluated using a unified data partitioning and verification process to ensure that the evaluation results of different models and different molecular descriptor characterization schemes are comparable.

[0150] Specifically, for different molecular descriptor characterization schemes and different candidate binary classification prediction models under the same key property task, it is preferable to use the same training set, validation set and test set partitioning rules, or to use the same cross-validation rules for evaluation.

[0151] For example, once a certain random data partitioning condition has been determined, all descriptor schemes and candidate models for that task can be trained and validated using the same training and test sets. Alternatively, when using cross-validation, all candidate schemes can be trained using the same number of folds, the same data stratification method, and the same random seed. This approach allows different models and descriptor schemes to be compared under as consistent data conditions as possible, avoiding overestimation of performance due to a particular scheme simply corresponding to a more easily classified data partition.

[0152] Furthermore, under a unified data partitioning and validation process, the input features, sample partitioning, and evaluation steps of each candidate model remain consistent for the same property task, with only the molecular descriptor representation method or model type itself changing. In this way, the evaluation results better reflect the fit between the descriptor scheme and the model itself, rather than fluctuations caused by differences in external validation conditions. For this invention, this approach helps improve the fairness of comparisons between different modeling schemes, thus providing a more consistent basis for subsequently determining the optimal molecular descriptor and target property prediction model for thermal stability, optical stability, Stokes shift, and AIE enhancement factor.

[0153] In this embodiment, the performance evaluation metrics include at least accuracy (ACC), precision (Precision), recall (Recall), F1 score, area under the curve (AUC), and overall evaluation metric (OAI).

[0154] Among them, accuracy (ACC) represents the proportion of the model that correctly classifies the entire model; precision represents the proportion of samples predicted as positive that are actually positive; recall represents the proportion of actual positive samples that are correctly identified; the F1 score reflects the balance between precision and recall; and AUC represents the overall discriminative ability of the model under different classification thresholds. Since a single evaluation metric is usually insufficient to fully reflect model performance, this embodiment introduces multiple metrics and further constructs a comprehensive evaluation metric, OAI, for unified evaluation of candidate models.

[0155] Preferably, the formulas for calculating accuracy (ACC), precision (Precision), recall (Recall), and F1 score are as follows:

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] Where TP represents the number of true positives, TN represents the number of true negatives, FP represents the number of false positives, and FN represents the number of false negatives.

[0161] Among the aforementioned metrics, ACC reflects the overall classification results of the model, but when the class distribution is imbalanced, using ACC alone may not be sufficient to fully characterize the model's ability to identify positive samples; Precision reflects the accuracy of the model's predictions, but if used alone, it may overlook the coverage of actual positive samples; Recall reflects the ability to detect positive samples, but if used alone, it may fail to reflect false positives; the F1 score, by combining Precision and Recall, balances these two factors to some extent. Therefore, a multi-metric joint evaluation can more comprehensively reflect the actual performance of candidate models in tasks of different natures.

[0162] In this embodiment, AUC is the area under the ROC curve of the receiver operating characteristic (ROC) curve, used to characterize the overall discriminative ability of the model under different classification thresholds. Compared with ACC, Precision, Recall, and F1 values ​​obtained based on fixed classification thresholds, AUC can examine the model's ability to distinguish between positive and negative samples within a wider threshold range, and therefore can serve as an important supplementary indicator for evaluating the model's ranking and discriminative ability.

[0163] For this invention, since it is necessary not only to determine whether the candidate AIE fluorescent probes meet the standards, but also to participate in the comprehensive scoring based on the model output results, introducing AUC helps to examine the model's ability to distinguish between good and bad samples as a whole, thereby providing a reference for the stability of the subsequent screening results.

[0164] To further comprehensively consider different evaluation indicators, this embodiment preferably sets a comprehensive evaluation indicator OAI, which satisfies:

[0165] ;

[0166] The reason for constructing the OAI using the above method is that ACC, Precision, Recall, F1, and AUC reflect the classification performance of candidate models from different perspectives. If only one single metric is used to determine the quality of a model, it's easy for a model to perform well on one metric but weakly on others, thus affecting the final selection result. By incorporating all these metrics into the OAI, overall accuracy, positive class recognition ability, classification balance, and overall discriminative ability can be considered simultaneously within a unified evaluation value.

[0167] Furthermore, since OAI is composed of multiple metrics, its value will be affected if a model is significantly deficient in one of the key metrics. Therefore, this metric can, to some extent, avoid the problem of model misselection caused by the "one-sided superiority" of a single metric. For this invention, using OAI as a comprehensive performance evaluation criterion is beneficial for comparing different molecular descriptor schemes and candidate models from multiple perspectives, providing quantitative support for selecting more suitable modeling schemes for each key property.

[0168] It should be noted that the product form is preferred in this embodiment to construct the OAI, so that all performance indicators can play a constraining role in the comprehensive evaluation. However, without departing from the core concept of this invention, those skilled in the art can also make conventional adjustments to the combination of OAI according to the specific data distribution and task requirements. As long as it is essentially still based on a unified comprehensive evaluation of candidate models using multiple performance indicators, it should be understood as an extension of the technical concept of this invention.

[0169] As a preferred embodiment, this embodiment further explains the process of determining the optimal molecular descriptor and / or target property prediction model for the four key properties in step S2.

[0170] In this embodiment, thermal stability, optical stability, Stokes shift, and AIE enhancement factor are modeled and analyzed as independent key properties. Since these four properties reflect different objects, their correlation with molecular structural features and experimental environmental factors may also differ. Therefore, in step S2, the four key properties are not pre-defined to share the same molecular descriptor characterization scheme and the same target property prediction model. Instead, different optimal molecular descriptors and / or different target property prediction models are allowed for each of the four key properties.

[0171] In other words, for the thermal stability task, the first optimal molecular descriptor can be determined from multiple candidate molecular descriptors based on the training results of the task, and the first target property prediction model can be determined from multiple candidate binary classification prediction models; for the photostability task, the second optimal molecular descriptor and the second target property prediction model can be determined independently; for the Stokes shift task and the AIE enhancement factor task, the same approach is used to determine the third optimal molecular descriptor, the third target property prediction model, the fourth optimal molecular descriptor, and the fourth target property prediction model, respectively.

[0172] In a specific implementation example, for a subset of data used for thermal stability modeling, Morgan fingerprints, Daylight fingerprints, Atom-pair fingerprints, MACCS fingerprints, and the Description molecular descriptor set can be used as input features. Under a unified data partitioning and validation process, support vector machine classification models, logistic regression models, gradient boosting tree classification models, multilayer perceptron classification models, random forest classification models, and extreme gradient boosting classification models can be trained respectively. Performance metrics such as ACC, Precision, Recall, F1, AUC, and OAI for each scheme are calculated. Based on the comprehensive performance comparison results, the best-performing molecular descriptor is selected as the optimal molecular descriptor for the thermal stability task, and candidate models that correspond to or complement it well are selected as the target property prediction models for the thermal stability task.

[0173] Subsequently, for a subset of photostable modeling data, under the same evaluation conditions, the aforementioned descriptor schemes and candidate models were repeatedly trained, validated, and their performance compared. Since photostable reflects the changes in candidate AIE fluorescent probes under illumination conditions, its correlation with chromophore stability, functional group sensitivity, and some environmental factors may differ from that of thermal stability. Therefore, the final optimal molecular descriptor and target property prediction model can differ from those for the thermal stability task. In other words, the optimal molecular descriptor for the photostable task does not need to be the same as that for the thermal stability task, nor does its target property prediction model need to be the same.

[0174] For the Stokes shift task, since this property is related to the difference between the absorption and emission peaks, it is significantly influenced by factors such as molecular conjugation structure, electron donor-acceptor distribution, and molecular conformation. Therefore, during model optimization, it is more likely to select a descriptor scheme and model type that is more suitable for characterizing spectral correlation features. Consequently, the optimal molecular descriptor and target property prediction model for the Stokes shift task can also differ from those for the aforementioned thermal stability and photostability tasks.

[0175] For the AIE enhancement factor task, since it directly reflects the difference in luminescence between candidate molecules in aggregated and dispersed states, and this property is often related to factors such as intramolecular rotational restriction, aggregated stacking mode, molecular rigidity, and conjugated structure distribution, its optimal structural characterization method and model type may differ from other property tasks. Based on the performance comparison of different descriptor schemes and candidate models for this task, the optimal molecular descriptor and target property prediction model corresponding to the AIE enhancement factor task can be further determined.

[0176] Therefore, it can be seen that the "four key properties corresponding to different optimal molecular descriptors and / or different target property prediction models" in this embodiment is not essentially a simple parallel setting of models, but rather a selection of data representation methods and classification models that are more suitable for the corresponding tasks based on the differences in prediction targets, feature sensitivity, and data distribution of the four property tasks themselves. The advantage of this approach is that it avoids the problem of a unified descriptor or model being insufficiently adapted to certain property tasks.

[0177] As a preferred embodiment, the output format of the prediction results of the four key properties in step S3 and the method of obtaining the synthetic accessibility score are further explained.

[0178] In the specific implementation process, the molecular structure information of the candidate AIE fluorescent probes to be evaluated can first be feature-transformed, and input features matching each key property task can be generated by combining experimental environment information. Subsequently, the corresponding input features are input into the thermal stability target property prediction model, the photostability target property prediction model, the Stokes shift target property prediction model, and the AIE enhancement factor target property prediction model, respectively, and the corresponding prediction results are output. The prediction results are preferably the probability of achieving the target, rather than simply outputting positive or negative class labels. That is to say:

[0179] The thermal stability target property prediction model outputs the probability that the candidate AIE fluorescent probe to be evaluated meets the thermal stability requirements, which is used as the predicted probability of thermal stability compliance. ;

[0180] The photostability target property prediction model outputs the probability that the candidate AIE fluorescent probe to be evaluated meets the photostability requirements, which is used as the photostability achievement prediction probability. ;

[0181] The Stokes shift target property prediction model outputs the probability that the candidate AIE fluorescent probe to be evaluated meets the Stokes shift requirement, which is used as the Stokes shift compliance prediction probability. ;

[0182] The probability that the AIE enhancement factor target property prediction model outputs the candidate AIE fluorescent probe to be evaluated meets the AIE enhancement factor requirements is used as the predicted probability of AIE enhancement factor compliance. .

[0183] In a specific implementation example, if a candidate AIE fluorescent probe has a positive probability of 0.82 after being predicted by the thermal stability target model, then 0.82 is taken as the predicted probability that the candidate molecule meets the thermal stability standard; if it has a positive probability of 0.67 after being predicted by the photostability target model, then 0.67 is taken as the predicted probability that its photostability standard is met; the other two properties are also obtained in the same way.

[0184] If only binary labels are used, the model output can only provide a discrete judgment of "meets the standard" or "does not meet the standard." Although this can meet the basic classification requirements, it is difficult to reflect the subtle differences between different candidates in the same property in the subsequent comprehensive scoring stage. Using the probability prediction of meeting the standard can more precisely represent the degree to which a candidate meets the requirements in each property, so that the performance differences of different candidates in the same property can be reflected in the form of continuous values.

[0185] Therefore, by using predicted probabilities to output the four key properties, we can provide richer information for subsequent comprehensive scoring and improve the distinguishability between different candidate molecules. This approach plays a positive role in solving the problem of the difficulty in unifying and refining the evaluation of multiple indicators in existing technologies.

[0186] In addition to predicting the probability of meeting the four key properties, this embodiment also calculates a synthesis accessibility score for the candidate AIE fluorescent probes to be evaluated. Preferably, the synthesis accessibility score is obtained by calculating the candidate AIE probes using cheminformatics tools and is used as one of the comprehensive scoring indicators in the subsequent weighted scoring.

[0187] Specifically, the molecular structure representation of candidate AIE fluorescent probes can be input into a cheminformatics tool. This tool, based on the complexity of the molecular backbone, functional group distribution, ring structure features, stereoscopic complexity, or other structural information related to the ease of synthesis, outputs a corresponding synthesis accessibility score. The cheminformatics tool can be a software program, algorithm module, or computing platform with synthesis difficulty assessment capabilities; its specific implementation is not limited, as long as it can provide an evaluation value characterizing the ease of synthesis of the candidate AIE fluorescent probe based on its structural information.

[0188] In a specific implementation example, the SMILES representation of a candidate AIE fluorescent probe can be input into the synthesis accessibility scoring module. The module automatically resolves the molecular structure and outputs a numerical result as the synthesis accessibility score of the candidate molecule. This score can be a continuous value, reflecting the relative ease or difficulty of synthesizing the candidate molecule in the actual process.

[0189] This invention aims to achieve a multi-index weighted scoring screening method for the optimal selection of AIE fluorescent probes. While screening based solely on four performance indicators—thermal stability, photostability, Stokes shift, and AIE enhancement factor—can reflect some of the performance characteristics of candidate molecules, it may still result in some candidates having theoretically good properties but being difficult to synthesize in practice. For AIE fluorescent probes requiring subsequent experimental verification and application development, ranking based solely on performance indicators without considering synthetic accessibility may lead to previously screened polymer targets being difficult to implement in subsequent preparation stages, thus affecting screening efficiency.

[0190] Therefore, this embodiment incorporates the synthetic accessibility score as one of the comprehensive scoring indicators into the subsequent weighted scoring process. This ensures that the comprehensive evaluation considers not only the performance of candidate molecules but also their practical preparation feasibility. In this way, the comprehensive scoring result is closer to the actual goal of "optimization," rather than simply a ranking based on performance.

[0191] Furthermore, by incorporating the synthetic accessibility score and the predicted probability of achieving the four key properties into the comprehensive scoring system, the subsequent screening results can take into account both "performance" and "implementation feasibility" factors. For this invention, this approach helps reduce screening bias caused by relying solely on performance indicators and provides a reference for the rational allocation of subsequent experimental resources, thereby improving the practicality of the screening results to a certain extent.

[0192] In this embodiment, the predicted probabilities of meeting the four key properties and the synthesis accessibility score together constitute the original input indicators for the comprehensive score. Among them, the predicted probabilities of meeting the thermal stability, photostability, Stokes shift, and AIE enhancement factor mainly reflect the degree to which the candidate AIE fluorescent probe meets the performance requirements; while the synthesis accessibility score mainly reflects the relative feasibility of the candidate AIE fluorescent probe in actual preparation.

[0193] By incorporating both types of indicators into the subsequent scoring system, "performance optimization" and "synthesis feasibility assessment" can be unified into the same screening process. This avoids the fragmented process caused by first screening based on performance and then manually judging synthesis feasibility; it also helps reduce the influence of human experience on the selection results of candidate probes, allowing subsequent ranking and screening to be based on a more complete indicator system.

[0194] After the above processing, each candidate AIE fluorescent probe can obtain five original indicators for subsequent comprehensive scoring: the predicted probability of thermal stability meeting the standard, the predicted probability of photostability meeting the standard, the predicted probability of Stokes shift meeting the standard, the predicted probability of AIE enhancement factor meeting the standard, and the synthesis accessibility score. Thus, performance and feasibility information from different sources and in different formats are organized into a set of indicators that can continue to participate in unified scoring analysis, providing a basis for subsequent weighted scoring and ranking selection.

[0195] As a preferred embodiment, the standardization process for synthesizing the accessibility score in step S4 will be further explained.

[0196] In step S4, the synthesis accessibility score is preferably standardized. Further, the standardization process includes normalization and direction alignment, so that the standardized synthesis accessibility score and the prediction results for the four key properties meet the unified evaluation direction that "a larger value indicates a better overall performance of the candidate AIE probe."

[0197] The normalization process refers to converting the original synthesis accessibility score into a standardized value within a preset range to reduce the impact of its original dimensions and numerical range on the subsequent comprehensive score. Preferably, the original synthesis accessibility score in the candidate AIE fluorescent probe sample set can be mapped to a fixed interval, such as the interval from 0 to 1.

[0198] After normalization, the synthesis accessibility score is further subjected to directional consistency processing. This directional consistency processing refers to converting the normalized synthesis accessibility score into a direction of superiority or inferiority that aligns with the prediction results of the four key properties, uniformly satisfying the principle that "the larger the value, the better the overall performance of the candidate AIE probe."

[0199] The need for directional consistency arises because the synthesis accessibility scores output by different cheminformatics tools may not be semantically consistent. In some scoring systems, smaller values ​​indicate easier molecule synthesis, while larger values ​​indicate more difficult synthesis. However, in this invention, the prediction results for the four key properties are preferably expressed using the probability of meeting the target, which generally means that larger values ​​indicate better performance or a higher probability of meeting the requirements. If the direction of the synthesis accessibility scores is not unified, subsequent comprehensive scoring may result in some indicators being considered "the larger the value, the better," while others are considered "the smaller the value, the better," leading to inconsistent scoring logic and increasing the complexity of subsequent weighted analysis.

[0200] Preferably, the standardization rule for the Synthetic Accessibility Score is as follows: when When, the standardized score is 1; when When, the standardized score is 0; when When the standardized score is calculated, it is done using the following formula:

[0201] ;

[0202] Through the above processing, the synthetic accessibility score can be transformed into an evaluation direction and scale consistent with the predicted probabilities of the four key properties, providing a foundation for subsequent multi-indicator weighted fusion.

[0203] As a preferred embodiment, the weight determination process and comprehensive score calculation process in step S5 are further explained. The comprehensive score index data includes both continuous and discrete score variables; therefore, in step S5, mixed data factor analysis is preferably used to perform a unified analysis of each score index to determine the weight corresponding to each score index.

[0204] The mixed data factor analysis is used to analyze the correlation and contribution relationships among multiple rating indicators when both continuous and discrete rating variables exist, and to extract principal components or latent factors accordingly.

[0205] In practice, the predicted probabilities of thermal stability, optical stability, Stokes shift, and AIE enhancement factor, along with the standardized synthesis accessibility score, can be used as input indicators to construct a comprehensive scoring index matrix for candidate AIE fluorescent probe samples. Subsequently, mixed data factor analysis is performed on this index matrix to extract principal components that characterize the comprehensive changes of multiple indicators, and the loadings of each scoring index on each principal component and the explanatory power of each principal component are calculated.

[0206] In this embodiment, the weight of each scoring index is preferably determined based on the loading of each scoring index on the principal component and the explanatory power of the corresponding principal component.

[0207] Specifically, after completing the mixed data factor analysis, the explanatory power of each principal component for the overall data variation, as well as the loading values ​​of each scoring indicator on each principal component, can be obtained. The loading values ​​reflect the degree of correlation between the corresponding scoring indicator and the corresponding principal component, while the explanatory power reflects the representativeness of the principal component to the overall information. Based on this, the comprehensive contribution value of each scoring indicator can be calculated by comprehensively considering both the "magnitude of the indicator's role on the principal component" and the "importance of the principal component itself," and the corresponding weights can be further determined accordingly. For example, for the thermal stability compliance prediction probability, the comprehensive contribution of the thermal stability indicator can be obtained based on the combination of its loading values ​​on each principal component and the explanatory power of the corresponding principal component; the comprehensive contribution values ​​for the light stability compliance prediction probability, Stokes displacement compliance prediction probability, AIE enhancement factor compliance prediction probability, and the standardized synthetic accessibility score are also calculated in the same way. Subsequently, the above comprehensive contribution values ​​can be normalized so that the sum of the weights corresponding to each indicator satisfies a preset constraint, such as a sum of 1, thereby forming the weight set required for subsequent comprehensive scoring.

[0208] See Figure 4 , Figure 4 This is a line graph showing the principal component explanatory power and cumulative explanatory power for different numbers of principal components when performing mixed data factor analysis on comprehensive scoring index data. It characterizes the information retention under different numbers of principal components and provides objective data support for the subsequent scientific determination of the weights of each index. Figure 4 As shown, the horizontal axis of this line graph represents the principal component number K, with a value range of 1 to 5; the vertical axis represents the information explanation rate, with a value range of 0.0 to 1.0. Figure 4 The graph contains two key curves: the solid line with diamond-shaped nodes (green curve) represents the "explanation rate" of a single principal component. It can be seen that as the principal component number K increases, the contribution of a single principal component to the explanation of the original data gradually decreases. The dashed line with circular nodes (blue curve) represents the "cumulative explanation rate" of the first K principal components. This curve increases monotonically with the increase of K, intuitively reflecting the overall completeness of information retention in the system. Figure 4It is known that when the number of principal components K=3, the cumulative explanatory power of the system has already exceeded 70%, possessing preliminary dimensionality reduction evaluation capabilities. However, the comprehensive scoring index data of this invention belongs to typical mixed-type data, which includes four continuous predictive probability variables and one standardized synthetic accessibility scoring variable. In order to minimize the loss of feature information, considering the information retention requirements near the inflection point of the cumulative explanatory power curve and the stability of subsequent weight distribution, this embodiment ultimately selects K=4 as the target number of principal components for weight analysis. With K=4, the number of principal components is accurately anchored through the cumulative explanatory power curve, effectively overcoming the technical biases of existing technologies in processing "continuous plus classification" mixed index data, such as unbalanced weight allocation, strong subjectivity, or the masking of the contribution of some variables. This design enables the final weights of each index to truly and objectively reflect their loading in the overall evaluation system, thereby greatly improving the interpretability and ranking stability of the AIE probe comprehensive scoring screening results. Under these conditions, the cumulative explanatory power of the system is close to 90%, which can fully cover the comprehensive performance characteristics of candidate AIE probes.

[0209] Figure 5 This intuitively reflects the influence of the selection of the number of principal components on the final weight allocation result of the evaluation system, and assists in determining the final weight analysis parameters. For example... Figure 5 As shown, the horizontal axis of this stacked bar chart represents the principal component number K, ranging from 1 to 5; the vertical axis represents the relative weight of each indicator, with a maximum score of 100%. Each stacked bar chart corresponds to five comprehensive scoring indicators from bottom to top: synthetic accessibility score (SA_score, orange area), thermal stability (… (green area), light stability ( (purple area), Stokes displacement ( (yellow area) and AIE enhancement factor ( (blue area).

[0210] By comparing the distribution of the histogram regions under different K values, it can be seen that as the number of principal components increases, the weight distribution of each indicator gradually tends to stabilize. This is consistent with the aforementioned cumulative explanatory power curve (see...). Figure 4 Based on the results, this embodiment preferably uses K=4 as the final weighting analysis parameter. Under the condition of K=4, the objectively calculated weight proportions of each index are as follows: thermal stability has the highest weight proportion at 26.28%; Stokes displacement is second at 24.38%; optical stability accounts for 22.00%; AIE enhancement factor accounts for 20.89%; and the synthesis accessibility score has a relatively low weight proportion at 6.44%. (Appendix) Figure 5The results strongly demonstrate the scientific validity and rationality of the FAMD weighting analysis method adopted in this invention. The objectively allocated weight distribution results are highly consistent with the actual requirements of AIE probe selection, which prioritizes key luminescent / physicochemical properties while also considering synthetic feasibility. Specifically, the four key properties occupy the dominant weights, while synthetic feasibility serves as a secondary constraint. This data-driven weight allocation method not only eliminates the subjective bias of manually setting weights but also allows for the reasonable integration of continuous variables (SA_score) and binary property variables within a unified evaluation framework. This significantly improves the interpretability, stability, and alignment with chemical intuition of the comprehensive scoring results.

[0211] In a preferred embodiment, after determining the weights of each scoring indicator, this embodiment further performs a linear weighted average of the prediction results for the four key properties and the standardized synthesis accessibility score based on the weights to obtain a comprehensive score for the candidate AIE fluorescent probe. Given that the weights of each indicator are already determined, the overall performance under the combined effect of all indicators can be directly reflected through a linear weighted summation. The comprehensive score satisfies the following relationship:

[0212] The comprehensive score of the candidate AIE probe is obtained by multiplying the predicted probability of thermal stability compliance, the predicted probability of optical stability compliance, the predicted probability of Stokes shift compliance, the predicted probability of AIE enhancement factor compliance, and the standardized synthesis accessibility score by their respective weights and then summing them.

[0213] Specifically, let the predicted probability of thermal stability meeting the standard be: The predicted probability of achieving photostability is: The probability of Stokes displacement meeting the target is... The probability of AIE enhancement factor meeting the target is: The standardized synthesis accessibility score is: ;

[0214] Correspondingly, the weights of each indicator are denoted as follows: the weight for thermal stability is... The photostability weight is Stokes displacement weights are The weight of the AIE enhancement factor is: The synthesis accessibility weight is: .

[0215] The overall score of candidate AIE probes It can be represented as:

[0216] ;

[0217] Through the above calculations, five evaluation indicators from different sources but which have undergone unified processing can be integrated into a single comprehensive score. The higher the comprehensive score, the better the overall performance of the candidate AIE fluorescent probe in terms of thermal stability, photostability, Stokes shift, AIE enhancement factor, and synthetic accessibility.

[0218] Example 2

[0219] The present invention also provides a multi-index weighted scoring system for AIE fluorescent probes, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the computer program instructions are executed by the processor, the device is triggered to perform some or all of the steps in Embodiment 1.

[0220] A processor may include one or more processing units, such as an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural network processing unit (NPU). These different processing units may be independent devices or integrated into one or more processors.

[0221] The controller can serve as the nerve center and command center of an electronic device. Based on the instruction opcode and timing signals, the controller generates operation control signals to control the fetching and execution of instructions.

[0222] The processor may also include memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or that are used repeatedly. If the processor needs to use the instruction or data again, it can retrieve it directly from the memory. This avoids repeated accesses, reduces processor waiting time, and thus improves system efficiency.

[0223] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A multi-index weighted scoring screening method for AIE fluorescent probes, characterized in that, Includes the following steps: S1. Obtain raw data including candidate AIE probe molecular structure information, experimental environment information and AIE property information, and perform data preprocessing and feature engineering on the raw data to obtain a feature dataset for modeling. S2. For the four key properties of thermal stability, optical stability, Stokes shift and AIE enhancement factor, multiple molecular descriptor characterization schemes are constructed, and candidate binary classification prediction models are established based on the characterization results of different molecular descriptors. According to the model performance evaluation results, the corresponding optimal molecular descriptor and target property prediction model are determined for each key property. S3. Input the candidate AIE probe to be evaluated into each of the target property prediction models to obtain four key property prediction results. The four key property prediction results are the prediction probabilities output by each target property prediction model. The prediction probabilities include the thermal stability compliance prediction probability, the optical stability compliance prediction probability, the Stokes displacement compliance prediction probability, and the AIE enhancement factor compliance prediction probability. Based on the molecular structure information of candidate AIE probes, their synthetic accessibility score is calculated using cheminformatics tools or rule-based algorithms. The synthetic accessibility score is used as one of the comprehensive scoring indicators in the subsequent weighted scoring to characterize the ease or difficulty of the candidate molecule in actual synthesis. S4. Construct comprehensive scoring index data based on the prediction results of the four key properties and the synthetic accessibility score, and standardize the synthetic accessibility score so that it has a consistent scoring direction with the prediction results of the four key properties. S5. Perform mixed data factor analysis on the comprehensive scoring index data to determine the weight of each scoring index, and perform comprehensive scoring on the candidate AIE probes to be evaluated according to the weight of each scoring index, so as to sort and screen the candidate AIE probes according to the comprehensive score.

2. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, The molecular structure information includes at least the SMILES representation of the molecule, the experimental environment information includes at least solvent information, and the AIE property information includes at least thermal stability, photostability, and property data related to absorption or emission spectra.

3. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, In step S1, the feature engineering process includes at least molecular structure feature construction, solvent information numerical processing, and key property label construction; The solvent information numerical processing involves mapping the solvent name to preset solvent parameters to form numerical features for machine learning modeling; the preset solvent parameters include at least one or more of Et(30), SP, SdP, SA and SB. The construction of the key property labels includes: Thermal stability is classified into two categories based on decomposition temperature; Photostable stability is classified into two categories based on whether it decomposes, changes its spectrum, or changes its molecular structure under light conditions. The Stokes shift is calculated based on the absorption wavelength and the emission wavelength, and binary classification is performed according to a preset threshold. The AIE enhancement factor is calculated based on the luminescence intensity in the aggregated and solution states, and binary classification is performed according to a preset threshold.

4. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, In step S2, the multiple molecular descriptor characterization schemes include at least two of the following: Morgan fingerprint, Daylight fingerprint, Atom-pair fingerprint, MACCS fingerprint, and the Description molecular descriptor set. The performance of different molecular descriptor characterization schemes was compared using multiple random data partitioning conditions, and the target random data partitioning conditions for modeling various key properties were determined based on the comprehensive performance index.

5. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, In step S2, the candidate binary classification prediction model includes at least two of the following: support vector machine classification model, logistic regression model, gradient boosting tree classification model, multilayer perceptron classification model, random forest classification model, and extreme gradient boosting classification model.

6. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, In step S2, the four key properties correspond to different optimal molecular descriptors and / or different target property prediction models.

7. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, In step S4, the standardization process for the synthesis accessibility score includes normalization and direction alignment.

8. The multi-index weighted scoring screening method for AIE fluorescent probes according to claim 1, characterized in that, In step S5, the comprehensive scoring index data includes both continuous and discrete scoring variables. The mixed data factor analysis is used to determine the weight of each scoring index under the mixed data conditions of continuous and discrete scoring variables. The weights of each scoring index are determined based on the loadings of each scoring index on the principal components and the explanatory power of the corresponding principal components. The prediction results on the four key properties and the standardized synthetic accessibility score are then linearly weighted based on the weights to obtain the comprehensive score of the candidate AIE probe.

9. A multi-index weighted scoring system for AIE fluorescent probes, characterized in that, It includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the multi-index weighted scoring screening method for AIE fluorescent probes as described in any one of claims 1-8.

Images