Cell survival prediction method and system based on multi-parameter conjoint analysis

Abstract

The invention relates to the technical field of drug effectiveness analysis, in particular to a cell survival prediction method and system based on multi-parameter conjoint analysis, and the method comprises the steps: putting a to-be-evaluated drug of a corresponding parameter combination to a target cell to prepare an evaluation sample; collecting cell analysis parameters for the evaluation sample; the cell analysis parameters comprise a surface plasmon resonance parameter, an elastic coefficient and a cell area; and constructing three-dimensional features according to the cell analysis parameters, and predicting the survival rate of the target cells under the parameter combination. Aiming at the problems that the existing cell survival rate statistics needs to depend on specific types of markers and the experiment cost is relatively high, three parameters capable of being used for analyzing the cell survival rate are determined through experiments. By collecting the corresponding parameters, converting the parameters into the three-dimensional features and inputting the three-dimensional features into the classifier for prediction, the survival rate of the cells under the current parameter combination can be effectively judged, the incubation time is shortened, and the experiment cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug effectiveness analysis, and in particular to a cell survival prediction method and system based on multi-parameter joint analysis. BACKGROUND

[0002] Cell viability analysis is an experimental technique that quantitatively assesses the activity of cell populations by detecting differences in specific biomarkers between live and dead cells, and is widely used in medical diagnosis, drug development and other fields. Cells continuously adapt to their microenvironment through changes in physical properties, including adhesion, morphology and mechanical elasticity. These biophysical properties reflect the intrinsic biochemical state of cells and are closely related to a variety of basic functions, such as migration, immune response, phagocytosis, apoptosis and cell-cell interaction. Accurate and real-time quantification of these parameters helps to better understand a wide range of physiological and pathological processes, including drug response, immune signal transduction and host-pathogen interaction. Traditional cell state evaluation relies heavily on fluorescence labeling techniques. For example, high-content imaging platforms extract multi-parameter morphological features through fluorescence labeling, and FRET (fluorescence resonance energy transfer) is often used to monitor intermolecular interactions and conformational changes.

[0003] In the prior art, there are corresponding devices to implement the related experimental procedures.

[0004] For example, the patent document CN202411307144.1 discloses a live cell dynamic monitoring method, which includes the following steps: obtaining a set of cell scanning pictures of live cells growing in a culture vessel under specified conditions, and pre-processing the pictures; inputting the pre-processed set of cell scanning pictures into a trained cell recognition model to output a set of cell growth state maps; inputting the set of cell growth state maps into a trained survival rate calculation model to output the survival rate of the cells in each cell growth state map; plotting the obtained cell survival rate values into a cell survival rate-time curve graph; and combining the generated cell survival rate-time curve graph with a nonlinear regression statistical method to select specified parameters to calculate the half-inhibitory concentration value of drug efficacy. The cell growth state under specified conditions is monitored, and the survival rate of the cells and the half-inhibitory concentration value of drug efficacy are calculated based on these data; the efficiency and accuracy of the fields of drug development, drug screening, treatment scheme effectiveness, etc. can be significantly improved.

[0005] ​For example, patent document CN201610168896.3 discloses a method for detecting the viability of cartilage tissue cells, including: cutting cartilage tissue into small pieces, digesting with 0.25% trypsin-EDTA2Na in a water bath; centrifuging and then continuing digestion with 0.2% type II collagenase in a water bath; waiting for the cartilage to become flocculent, adding DMEM culture medium to stop digestion, filtering with a 200-mesh sieve, centrifuging the filtrate, and discarding the supernatant; adding FDA and EB to the cell suspension and incubating in the dark; using IPP image analysis software to count the green and red cells in the image, and calculating the viability of cartilage tissue cells according to a formula. This method is conducive to fully digesting and separating cartilage cells, reducing damage to cartilage cells during the digestion and separation process, maintaining cell activity, avoiding the tediousness caused by the different excitation wavelengths of dual fluorescent dyes, and simplifying the operation; it is used to screen the pros and cons of treatment plans for articular cartilage injuries and evaluate the effect of cartilage preservation in tissue banks.

[0006] However, in actual implementation, the inventors found that traditional cell state assessment relies on fluorescence labeling technology. For example, high-content imaging platforms extract multi-parameter morphological features through fluorescence labeling, while FRET ( Resonance energy transfer (RET) is often used to monitor intermolecular interactions and conformational changes. Despite their widespread application, fluorescence methods primarily reflect changes at the biochemical level and cannot directly measure physical properties such as adhesion or stiffness. Physical changes are often inferred indirectly through signal changes. In addition, fluorescent dyes themselves may interfere with cellular function, and issues such as photobleaching and phototoxicity limit their real-time monitoring capabilities. In recent years, a variety of label-free methods have emerged to study the physical properties of cells. Electrochemical impedance spectroscopy (EIS) can infer cellular processes by measuring changes in resistance and capacitance and has been extended to single-cell measurements, but this extension often requires complex equipment and has limited spatial resolution. Real-time deformation cytometry (RT-DC) can assess the mechanical stiffness of cells by deforming them under fluid stress in microfluidic channels, but it only provides a single mechanical parameter and cannot reflect adhesion or morphological characteristics. Traction force microscopy (TFM) and micropillar array methods measure the traction forces generated by cells by tracking substrate deformation (such as the displacement of embedded beads or the offset of micropillars), providing stress information with high spatial resolution. However, these methods rely on specialized matrices and require complex calibration, limiting their throughput and scalability. More critically, most label-free techniques can only measure one physical parameter at a time, making it difficult to achieve multi-parameter, real-time phenotypic analysis at the single-cell level. Summary of the Invention

[0007] In view of the above problems existing in the prior art, a cell survival prediction method based on multi-parameter joint analysis is provided;

[0008] In another aspect, a system for implementing the above method is also provided.

[0009] The specific technical solution is as follows:

[0010] A cell survival prediction method based on multi-parameter joint analysis, comprising:

[0011] Step S1: A target cell is subjected to a corresponding parameter combination of a to-be-evaluated drug to prepare an evaluation sample;

[0012] The parameter combination comprises at least one variable parameter in concentration, dose and incubation time; Step S2: Cell analysis parameters are collected from the evaluation sample;

[0013] The cell analysis parameters consist of surface plasmon resonance parameters, elastic coefficients and cell areas;

[0014] Step S3: Three-dimensional features are constructed for the cell analysis parameters;

[0015] Step S4: The survival rate of the target cell under the parameter combination is predicted based on the three-dimensional features.

[0016] On the other hand, the step S2 comprises:

[0017] Step S21: The evaluation sample is irradiated to collect the surface plasmon resonance parameters and elastic coefficients, respectively;

[0018] Step S22: A cell image of the evaluation sample is taken, and the cell image is segmented and quantified to obtain the cell area.

[0019] On the other hand, the step S3 comprises:

[0020] Step S31: The cell analysis parameters are preprocessed to obtain preprocessed parameters, respectively;

[0021] Step S32: The preprocessed parameters are aligned in the time dimension and the concentration dimension, and then three-dimensional feature construction is performed to obtain pre-constructed features;

[0022] Step S33: The pre-constructed features are standardized and dimensionally reduced to obtain the three-dimensional features;

[0023] In the step S32, the three-dimensional construction method comprises at least one of time series modeling of parameters at multiple time points, fusion of parameters at different resolutions, and modeling of the interaction relationship between parameters;

[0024] In the step S33, the method of dimension reduction processing comprises at least one of t-SNE, PCA, UMAP and autoencoder.

[0025] In another aspect, the operation of pre-processing the surface plasmon resonance parameter in step S31 includes:

[0026] The surface plasmon resonance parameter is standardized and Gaussian noise is added, and then t-SNE parameter optimization is performed;

[0027] The elastic coefficient is standardized and Gaussian noise is added, and then t-SNE parameter optimization is performed;

[0028] The cell area is standardized and Gaussian noise is added, and then square root transformation is applied to reduce the influence of large area values, and finally t-SNE parameter optimization is performed;

[0029] The standardization process includes applying at least one of Z-score standardization, robust standardization, and minimum-maximum standardization;

[0030] The level of Gaussian noise is in the range of 0.005-0.05;

[0031] The parameters involved in the t-SNE parameter optimization include t-SNE perplexity parameter and t-SNE learning_rate parameter;

[0032] Wherein, the range of t-SNE perplexity parameter is between 50-150, and the range of t-SNE learning_rate parameter is between 1000-10000.

[0033] In another aspect, in step S33, when the method of reducing the dimension of the pre-constructed feature includes t-SNE parameter optimization, the optimized parameters include:

[0034] Dimension reduction parameters: n_components=3, perplexity=30, n_iter=30000;

[0035] Initialization method: random initialization, early_exaggeration=2.0;

[0036] Distance measurement: Use cosine distance to capture the angle relationship between features;

[0037] Optimization settings: learning_rate=200, angle=0.5 balance speed and accuracy.

[0038] In another aspect, in the step S4, a gradient boosting classifier or a random forest classifier or a support vector machine classifier or a multi-layer perceptron classifier or an XGBoost classifier or a LightGBM classifier is used as the machine learning classifier for prediction.

[0039] In the gradient boosting classifier, a learning strategy of sequential learning is integrated.

[0040] In the random forest classifier, a strategy of training multiple decision trees in parallel and making prediction by voting mechanism is integrated.

[0041] The LightGBM classifier integrates a leaf-first tree growing strategy, and performs histogram optimization and provides network parallel support.

[0042] In another aspect, before the step S1 is performed, a classifier training process is further included.

[0043] The classifier training process includes:

[0044] Step A1: a plurality of training samples are prepared by transforming the parameter combinations, and the training samples are incubated and observed to add labels corresponding to survival or not;

[0045] Step A2: training sample data corresponding to the cell analysis parameters are collected for the training samples;

[0046] The training sample data further include the labels;

[0047] Step A3: the training sample data are standardized and noise is added, and then a dataset is constructed based on the labels;

[0048] The dataset is supervised learning data, and stratified sampling is used to ensure class balance during construction of the dataset;

[0049] Step A4: the machine learning classifier is trained using the dataset, and the machine learning classifier is outputted;

[0050] In the step S4, hyperparameter optimization is performed on the machine learning classifier based on grid search or Bayesian optimization, and the machine learning classifier is evaluated based on at least one of accuracy, precision, recall, and F1-score.

[0051] In another aspect, after the step S4, the following steps are further included:

[0052] Step S5: the survival rates corresponding to a plurality of the parameter combinations are compared to generate a comparison chart output.

[0053] In another aspect, before step S1 is performed, the following steps are further included:

[0054] SHAP value calculation is performed on the cell analysis parameters to explain the contribution of each feature to the prediction result;

[0055] In addition, the feature importance of the cell analysis parameters is evaluated by feature permutation;

[0056] In addition, feature interaction analysis is performed on the cell analysis parameters to identify the synergistic effect between features.

[0057] A cell survival prediction system based on multi-parameter joint analysis for implementing the above-mentioned cell survival prediction method;

[0058] The cell survival prediction system comprises:

[0059] A sample preparation module, which uses a corresponding parameter combination of the target cell to evaluate the drug to be evaluated to prepare an evaluation sample;

[0060] The parameter combination includes at least one variable parameter in concentration, dose, and incubation time; a sampling module connected to the sample preparation module;

[0061] The sampling module collects cell analysis parameters from the evaluation sample;

[0062] The cell analysis parameters consist of surface plasmon resonance parameters, elastic coefficients, and cell areas;

[0063] A feature generation module connected to the sampling module;

[0064] The feature generation module constructs three-dimensional features for the cell analysis parameters;

[0065] A survival rate output module connected to the feature generation module;

[0066] The survival rate output module predicts the survival rate of the target cell under the parameter combination based on the three-dimensional features;

[0067] A visualization module connected to the survival rate output module;

[0068] The visualization module generates T-SNE visualization, 3D scatter plot, parameter correlation heat map, and survival curve based on the predicted survival rate;

[0069] An interactive parameter adjustment module connected to the survival rate output module and the visualization module, respectively;

[0070] The interactive parameter adjustment module adjusts the classifier parameters and the visualization parameters in real time; the batch processing module is connected to the survival rate output module;

[0071] The batch processing module processes multiple samples in parallel and outputs batch results;

[0072] The result export module is connected to the survival rate output module;

[0073] The result export module generates results and reports in corresponding formats in response to user instructions.

[0074] The above technical solution has the following advantages or beneficial effects:

[0075] Most existing label-free technologies can only measure one physical parameter at a time, making it difficult to achieve multi-parameter, real-time phenotype analysis at the single-cell level. Through experiments, three parameters that can be used to analyze cell survival rate were determined. By collecting the corresponding parameters and converting them into three-dimensional feature inputs into the classifier for prediction, the survival rate of cells under the current parameter combination can be effectively determined, and the incubation time can be shortened, reducing the experimental cost. BRIEF DESCRIPTION OF DRAWINGS

[0076] Reference is made to the accompanying drawings to more fully describe embodiments of the present application. However, the accompanying drawings are only used for illustration and explanation, and do not constitute a limitation on the scope of the present application.

[0077] Figure 1 The figure is a schematic diagram of the overall embodiment of the present application;

[0078] Figure 2 The figure is a schematic diagram of step S2 in the embodiment of the present application;

[0079] Figure 3 The figure is a schematic diagram of step S3 in the embodiment of the present application;

[0080] Figure 4 The figure is a schematic diagram of the training process in the embodiment of the present application;

[0081] Figure 5 The figure is a schematic diagram of step S5 in the embodiment of the present application;

[0082] Figure 6 The figure is a schematic diagram of the system in the embodiment of the present application;

[0083] Figure 7 The figure is a schematic diagram of the sampling module in the embodiment of the present application;

[0084] Figure 8 The figure is a schematic diagram of the feature generation module in the embodiment of the present application. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0086] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0087] The present application will be further described below with reference to the drawings and specific embodiments, but is not limited to the present application.

[0088] The present application includes:

[0089] A cell survival prediction method based on multi-parameter joint analysis, as shown in Figure 1 The method comprises the following steps:

[0090] Step S1: A target cell is subjected to a to-be-evaluated drug corresponding to a parameter combination to prepare an evaluation sample;

[0091] The parameter combination comprises at least one variable parameter in concentration, dosage and incubation time;

[0092] Step S2: A cell analysis parameter is collected for the evaluation sample;

[0093] The cell analysis parameter is composed of a surface plasmon resonance parameter, an elastic coefficient and a cell area;

[0094] Step S3: A three-dimensional feature is constructed for the cell analysis parameter;

[0095] Step S4: A survival rate of the target cell under the parameter combination is predicted based on the three-dimensional feature.

[0096] Specifically, in order to solve the problem that the cell survival rate statistics in the prior art need to rely on a specific type of marker and the experimental cost is high, in the present embodiment, first, a regression analysis method is used to statistically analyze and correlate the indexes collected by the cell without relying on a fluorescent marker, and three indexes with a high P value are screened, including a surface plasmon resonance parameter, an elastic coefficient and a cell area, which are used as indexes that can be used in the subsequent prediction process.

[0097] Among them, the surface plasmon resonance parameter (Surface Plasmon Resonance, SPR) can be used to reflect the mass and density characteristics of the cells. In the acquisition process, mainly use Kretschmann structure, microfluidic chip or other equivalent sensors based on surface plasmon wave technology to collect, by introducing the relevant light between the metal surface-sample layer to resonate, thereby collecting the optical parameters of the interaction between the cells and the metal surface to calculate.

[0098] Generally speaking, the greater the value of the surface plasmon resonance parameter, the greater the mass within 200nm distance close to the surface of the sensor Z axis, the greater the adhesion strength of the cells to the surface, and the healthier the cell state.

[0099] The scattering coefficient refers to the elastic parameters of the cells themselves collected by optical tweezers, Brillouin light scattering and other methods, which are usually related to the roughness and adhesion degree of the cell surface. The greater the value, the greater the elasticity of the cells.

[0100] The area parameter refers to the cell surface area measured by microscopic image, which directly reflects the size of the cell. Generally, healthy cells have larger area, while dead cells have smaller area.

[0101] After determining the three effective indicators according to the experiment, the corresponding survival rate analysis can be carried out based on the above indicators.

[0102] Specifically, according to the experimental requirements, such as drug concentration-cell survival rate analysis experiment, different drug parameter combinations need to be used, such as drug concentration, common values are 0, 8, 20, 30, 40, 50 μM, drug dose and incubation time, to determine the preparation method of the corresponding evaluation sample.

[0103] It should be noted that, since the traditional fluorescent marker strictly follows the different fluorescent reactions of the cells in the survival and death states, the traditional fluorescent marker usually needs a longer incubation time to make the cells change under the action of the drug.

[0104] And the parameters relied on in this scheme are mainly to observe the changes of the related indicators of the appearance of the cells under the action of the drug, so it is not strictly necessary to observe the death process of the cells.

[0105] Based on this difference, the incubation time can be effectively shortened, such as from four hours to one hour, and the shorter incubation time can improve the experimental efficiency and reduce the experimental cost.

[0106] After adding the drug according to the corresponding concentration and dosage, the target cells are incubated to obtain an evaluation sample. Then, the evaluation sample is subjected to image shooting, parameter measurement and the like by using a microfluidic system or a microscope and the like, so that the three cell analysis parameters are obtained.

[0107] Based on the three cell analysis parameters, the three-dimensional features can be assembled, and the three-dimensional features are predicted based on a classifier, so that the survival rate of the cells is predicted and output.

[0108] In one embodiment, as shown in Figure 2 Step S2 includes:

[0109] Step S21: irradiating the evaluation sample to respectively collect the surface plasmon resonance parameter and the elastic coefficient;

[0110] Step S22: shooting a cell image of the evaluation sample, and segmenting and quantifying the cell image to obtain a cell area.

[0111] Specifically, to achieve better parameter collection efficiency, the embodiment provides a cell analysis parameter collection method.

[0112] Specifically, for the surface plasmon resonance parameter and the elastic coefficient, a specific wavelength laser can be used for irradiation, excitation and measurement of the resonance signal change and the scattering light intensity, respectively.

[0113] The surface plasmon resonance parameter is obtained according to the resonance signal change, and the elastic coefficient is obtained according to the scattering light intensity.

[0114] Subsequently, a cell image of the evaluation sample is shot, which is an optical microscopic image. After gray scale conversion, filtering, binarization and the like, a segmentation model such as U-net can be used to separate the cell region and the background part, and the cell area is calculated by counting the pixel number of the segmentation result. For the subsequent analysis process, the cell areas of multiple cells in the field of view are usually averaged, but multiple cell areas can also be extracted to assemble a vector for more accurate prediction.

[0115] It should be noted that the above processing method is usually a processing method for a single time point or a single data frame. Considering that some classifiers can improve the prediction accuracy of the survival rate after a certain period of time by inputting sequence data, during the collection of the cell analysis parameters of the evaluation sample, data of a certain period of time can be collected, and then the above processing method is performed on each data frame. The three-dimensional vector is assembled and predicted based on the three groups of parameter sequences.

[0116] In one embodiment, as shown in Figure 3As shown, step S3 comprises:

[0117] Step S31: preprocessing respectively for cell analysis parameters to obtain preprocessed parameters;

[0118] Step S32: aligning the preprocessed parameters in time dimension and concentration dimension, and then constructing three-dimensional features to obtain pre-constructed features;

[0119] Step S33: standardizing and dimensionality reducing the pre-constructed features to obtain three-dimensional features;

[0120] In step S32, the three-dimensional construction method comprises at least one of the following: time series modeling of parameters at multiple time points, fusion of parameters at different resolutions, and modeling of interaction between parameters.

[0121] In step S33, the method of dimensionality reduction comprises at least one of the following: t-SNE, PCA, UMAP, and autoencoder.

[0122] Specifically, to realize the combination of multiple cell analysis parameters and facilitate subsequent prediction, in this embodiment, first, for a single cell analysis parameter, preprocessing is performed respectively to form preprocessed parameters. The preprocessing process usually includes standardization processing, adding Gaussian noise to increase data robustness, dimensionality reduction processing, etc.

[0123] Specifically, in step S31, the preprocessing operation for the surface plasmon resonance parameter comprises:

[0124] Standardizing the surface plasmon resonance parameter and adding Gaussian noise, and then performing t-SNE parameter optimization;

[0125] Standardizing the elastic coefficient and adding Gaussian noise, and then performing t-SNE parameter optimization;

[0126] Standardizing the cell area and adding Gaussian noise, and then applying square root transformation to reduce the influence of large area value deviation, and finally performing t-SNE parameter optimization;

[0127] Generally, the optional processing operation comprises:

[0128] The standardization processing comprises at least one of the following: Z-score standardization, robust standardization, and minimum-maximum standardization;

[0129] The level of Gaussian noise is in the range of 0.005-0.05;

[0130] The parameters involved in t-SNE parameter optimization include T-SNE perplexity parameter and t-SNE learning_rate parameter;

[0131] wherein the t-SNE perplexity parameter ranges between 50-150, and the T-SNE learning_rate parameter ranges between 1000-10000.

[0132] The combination of the above parameters achieves better processing results.

[0133] In a preferred embodiment, the combination of processing parameters includes:

[0134] The surface plasmon resonance parameters are subjected to Z-score standardization, and Gaussian noise with a level of 0.01 is added, followed by t-SNE parameter optimization, and the optimization parameters include perplexity=80, learning_rate=5000, and cosine distance is used;

[0135] The elastic coefficient is subjected to RobustScaler to reduce the influence of outliers, and Gaussian noise with a level of 0.015 is added, followed by t-SNE parameter optimization, and the optimization parameters include perplexity=100, learning_rate=5000;

[0136] The cell area is subjected to StandardScaler for standardization, and Gaussian noise with a level of 0.02 is added, followed by square root transformation to reduce the deviation of large area values, and finally t-SNE parameter optimization, and the optimization parameters include perplexity=120, learning_rate=5000.

[0137] Meanwhile, for the three groups of pre-processing parameters, alignment processing is required to ensure the consistency of the three groups of parameters at the same concentration and time point, and then the pre-constructed features are combined, and then standardized and subjected to T-SNE parameter optimization for dimension reduction, and finally three-dimensional features are obtained.

[0138] In one embodiment, in step S33, when the method for dimension reduction processing of the pre-constructed features includes t-SNE parameter optimization, the optimization parameters include:

[0139] Dimension reduction parameters: n_components=3, perplexity=30, n_iter=30000;

[0140] Initialization method: random initialization, early_exaggeration=2.0;

[0141] Distance measurement: cosine distance is used to capture the angle relationship between features;

[0142] Optimization setting: learning_rate = 200, angle = 0.5 balance speed and accuracy.

[0143] In one embodiment, in step S4, a gradient boosting classifier or a random forest classifier or a vector machine classifier or a multi-layer perception classifier or an XGBoost classifier or a LightGBM classifier is used for prediction.

[0144] When a gradient boosting classifier is used, the parameters of the classifier include:

[0145] The base learner is built based on decision stumps, and the ensemble strategy includes sequential learning, each new model corrects the errors of the previous model, and the log loss function is applied to optimize the binary classification problem. Regularization includes controlling overfitting through learning rate, and supports subsampling.

[0146] When a gradient boosting classifier is used, the parameters of the three features and the combined three-dimensional features described above include:

[0147] SPR parameters:

[0148] n_estimators = 150, learning_rate = 0.05, max_depth = 4, subsample = 0.8; SC parameters:

[0149] n_estimators = 100, learning_rate = 0.08, max_depth = 3, subsample = 0.85; AREA parameters:

[0150] n_estimators = 180, learning_rate = 0.03, max_depth = 5, subsample = 0.75; three-dimensional features:

[0151] n_estimators = 200, learning_rate = 0.04, max_depth = 6, subsample = 0.8.

[0152] When a random forest classifier is used, the parameters involved include:

[0153] Decision trees are used as base learners, Bagging-based ensemble learning method, multiple decision trees are used for voting, and the ensemble strategy is to train multiple decision trees in parallel, and make the final prediction through the voting mechanism. When selecting features, a subset of features is randomly selected at each split.

[0154] When the random forest classifier is selected, the parameters of the three features and the combined three-dimensional features described above include:

[0155] SPR parameters:

[0156] n_estimators=200, max_depth=8, min_samples_split=5;

[0157] SC parameters:

[0158] n_estimators=150, max_depth=6, min_samples_leaf=3;

[0159] AREA parameters:

[0160] n_estimators=250, max_depth=10, min_samples_split=4; 3D combination:

[0161] n_estimators=250, max_depth=8, max_features='sqrt.

[0162] When the support vector machine classifier is selected, the parameters involved include:

[0163] The radial basis function kernel is selected to handle nonlinear classification problems, and the optimization goal is to maximize the classification interval while minimizing classification errors. By selecting the random forest classifier, high-dimensional data classification can be achieved, and good performance can be achieved for small sample data.

[0164] Under the random forest classifier, the parameters of the three features and the combined three-dimensional features described above include:

[0165] SPR parameters:

[0166] C=1.0, kernel='rbf', gamma='scale';

[0167] SC parameters:

[0168] C=1.5, kernel='rbf', gamma='auto';

[0169] AREA parameters:

[0170] C=2.0, kernel='rbf', gamma='scale';

[0171] Three-dimensional features:

[0172] C=1.5, kernel='rbf', gamma='scale'.

[0173] When the multi-layer perception classifier is selected, the non-linear expression capability is enhanced by the ReLU activation function, and the Adam optimizer is used for adaptive learning rate optimization, and the regularization can use L2 regularization to avoid overfitting.

[0174] Under the multi-layer perception classifier, the parameters of the three features and the combined three-dimensional features described above include:

[0175] SPR parameters:

[0176] hidden_layer_sizes=(100, 50), alpha=0.0001, max_iter=1000; SC parameters:

[0177] hidden_layer_sizes=(80, 40), alpha=0.001, max_iter=1000;

[0178] AREA parameters:

[0179] hidden_layer_sizes=(120, 60), alpha=0.0005, max_iter=1000;

[0180] Three-dimensional features:

[0181] hidden_layer_sizes=(100, 50, 25), alpha=0.0001, max_iter=2000.

[0182] When the XGBoost classifier is selected, the second-order gradient information can be configured as row sampling and column sampling, and the parameters of the three features and the combined three-dimensional features described above include: SPR parameters:

[0183] n_estimators=100, max_depth=5, learning_rate=0.05, subsample=0.8; SC parameters:

[0184] n_estimators=120, max_depth=4, learning_rate=0.03, subsample=0.85; AREA parameters:

[0185] n_estimators=150, max_depth=6, learning_rate=0.04, subsample=0.75; three-dimensional features:

[0186] n_estimators = 180, max_depth = 6, learning_rate = 0.03, subsample = 0.8.

[0187] When the LightGBM classifier is selected, the leaf-preferred tree growth strategy is adopted, and histogram optimization is performed. The parameters of the three features and the combined three-dimensional features described above include: SPR parameters:

[0188] n_estimators = 100, max_depth = 5, learning_rate = 0.05, subsample = 0.8; SC parameters:

[0189] n_estimators = 120, max_depth = 4, learning_rate = 0.03, subsample = 0.85; AREA parameters:

[0190] n_estimators = 150, max_depth = 6, learning_rate = 0.04, subsample = 0.75; three-dimensional features:

[0191] n_estimators = 200, max_depth = 7, learning_rate = 0.025, subsample = 0.85.

[0192] In order to achieve better prediction effect, in the scheme, after testing the recognition performance of a plurality of classifiers, the test results shown in Table 1 are obtained:

[0193]

[0194] Table 1

[0195] From the test results in Table 1, it can be seen that the gradient boosting classifier has the highest prediction accuracy and the strongest explanation, can provide detailed feature importance and decision path, and the training time is moderate and the memory occupation is reasonable. Therefore, it is selected as the classifier used in the actual scheme, and the random forest can be used as a suboptimal classifier.

[0196] However, according to other selection criteria, other models can also be selected, such as:

[0197] Accuracy first:

[0198] GradientBoosting > XGBoost > LightGBM > RandomForest > SVM > MLP;

[0199] Speed priority:

[0200] LightGBM > XGBoost > RandomForest > GradientBoosting > SVM > MLP;

[0201] Memory efficiency:

[0202] LightGBM > SVM > GradientBoosting > MLP > XGBoost > RandomForest;

[0203] Interpretability priority:

[0204] GradientBoosting > RandomForest > XGBoost > LightGBM > SVM > MLP;

[0205] Stability priority:

[0206] RandomForest > GradientBoosting > LightGBM > XGBoost > SVM > MLP.

[0207] On this basis, the above parameter configuration can achieve better training effect of the model.

[0208] In an embodiment, before step S1, the classifier training process is further included;

[0209] As shown in Figure 4 , the classifier training process includes:

[0210] Step A1: a plurality of training samples are prepared by transforming parameter combinations, and the training samples are incubated and observed to add labels corresponding to survival or not;

[0211] Step A2: training sample data corresponding to cell analysis parameters are collected from the training samples;

[0212] Labels are also added to the training sample data;

[0213] Step A3: the training sample data are standardized and noise is added, and then a dataset is constructed based on the labels;

[0214] The dataset is supervised learning data, and stratified sampling is used to ensure class balance during the construction of the dataset;

[0215] Step A4: a machine learning classifier is trained using the dataset, and the machine learning classifier is outputted;

[0216] In step S4, the machine learning classifier is subjected to hyperparameter optimization based on grid search or Bayesian optimization, and the machine learning classifier is evaluated based on at least one of accuracy, precision, recall, and F1-score.

[0217] Specifically, to implement the training process of the above model, in the embodiment, before starting the identification, a plurality of training samples are prepared by constructing different parameter combinations, and then incubation is performed on the training samples, so that after the drug fully acts on the cells, it is observed whether the cells survive to perform manual labeling.

[0218] Subsequently, the training samples are subjected to cell analysis parameter acquisition to form training sample data according to the same parameter acquisition method, and then the data are processed based on the same preprocessing method to construct a data set, which is divided into a training set, a test set, and a validation set in proportion.

[0219] The data set is used to train the gradient boosting classifier, a loss function is used to quantify the LOSS of each round and perform parameter optimization, and the classifier is output after the iteration condition is reached.

[0220] Further, to illustrate the effectiveness of the above three cell analysis parameters, after the above three indicators are obtained through regression analysis screening, whether a single indicator can be used to judge the cell survival rate is further verified.

[0221] The experimental process specifically includes: a combination of drugs of a specific concentration is added to sample cells, then incubation is performed, then a single parameter is processed in the same way, the parameter is predicted by the classifier, and the evaluation results as shown in Table 2 are obtained:

[0222] Parameter Accuracy Precision Recall F1 AUC-ROC SPR 0.847 0.832 0.863 0.847 0.912 SC 0.791 0.776 0.809 0.792 0.871 AREA 0.823 0.811 0.837 0.824 0.896

[0223] Table 2

[0224] Among them, SPR is the surface plasmon resonance parameter, SC is the elastic coefficient, and AREA is the area. According to Table 2, the above three parameters can predict the survival rate to a certain extent as single parameters. According to the importance ranking, the importance of SPR is 0.42, which is the most important feature, the importance of SC is 0.35, which is a secondary feature, and the importance of AREA is 0.23, which is an auxiliary prediction feature.

[0225] After combining the above three parameters, the overall parameter accuracy measurement process is performed, as shown in Table 3:

[0226] Model combination Training accuracy Validation accuracy Test accuracy Degree of overfitting SPR+SC+AREA 0.924 0.887 0.891 Low (3.7%) SPR+AREA 0.908 0.879 0.883 Low (2.9%) SPR+SC 0.896 0.851 0.856 Moderate (5.3%)

[0227] Table 3

[0228] It can be seen that the accuracy increases significantly with the increase of parameters.

[0229] Meanwhile, the prediction accuracy under different concentration combinations is also tested, as shown in Table 4:

[0230] Concentration Sample number Predicted accuracy Standard deviation 95% confidence interval 0 156 0.923 0.034 (0.889,0.957) 8 142 0.901 0.041 (0.860,0.942) 20 138 0.884 0.038 (0.846,0.922) 30 129 0.876 0.045 (0.831,0.921) 40 121 0.859 0.049 (0.810,0.908) 50 108 0.833 0.052 (0.781,0.885)

[0231] Table 4

[0232] It can be seen that the combination of the above features has good prediction accuracy under any concentration, and the data standard deviation is small. After cross-validation of the data combination, the average accuracy of 5-fold cross-validation is in the range of 0.887±0.021, the standard deviation between each fold is less than 2.5%, the model stability is good, the test set performance is basically consistent with the validation set, and the generalization ability is strong.

[0233] In one embodiment, as shown in Figure 5 , after step S4, it further includes:

[0234] Step S5: comparing the survival rates corresponding to the plurality of parameter combinations to generate a comparison chart output.

[0235] A cell survival prediction system based on multi-parameter joint analysis is used to implement the above-mentioned cell survival prediction method;

[0236] As shown in Figure 6 , the cell survival prediction system includes:

[0237] A sample preparation module 1, which puts the corresponding parameter combination of the target cell into the drug to be evaluated to prepare an evaluation sample;

[0238] The parameter combination includes at least one variable parameter in concentration, dose, and incubation time;

[0239] A sampling module 2 connected to the sample preparation module 1;

[0240] The sampling module collects cell analysis parameters for the evaluation sample;

[0241] The cell analysis parameters consist of surface plasmon resonance parameters, elastic coefficients, and cell area;

[0242] A feature generation module 3 connected to the sampling module 2;

[0243] The feature generation module 3 constructs a three-dimensional feature for the cell analysis parameters;

[0244] A survival rate output module 4 connected to the feature generation module 3;

[0245] The survival rate output module 4 predicts the survival rate of the target cell under the parameter combination based on the three-dimensional feature.

[0246] Visualization module 5, visualization module 5 is connected to survival rate output module 4;

[0247] The visualization module 5 generates a T-SNE visualization graph, a 3D scatter plot, a parameter correlation heat map, and a survival curve graph based on the predicted survival rate;

[0248] Interactive parameter adjustment module 6, the interactive parameter adjustment module 6 is connected to the survival rate output module 4 and the visualization module 5 respectively;

[0249] The interactive parameter adjustment module 6 adjusts the classifier parameters and visualization parameters in real time;

[0250] Batch processing module 7, batch processing module 7 is connected to survival rate output module 4;

[0251] The batch processing module 7 processes multiple samples in parallel and outputs batch results;

[0252] A result export module 8, the result export module 8 is connected to the survival rate output module 4;

[0253] The result export module 8 responds to user instructions to generate results and reports in corresponding formats.

[0254] In one embodiment, Figure 7 As shown, the sampling module 2 includes:

[0255] a first sampling module 21 , which irradiates the evaluation sample to collect surface plasmon resonance parameters and elastic coefficients;

[0256] The second sampling module 22 takes cell images of the evaluation sample, and segments and quantifies the cell images to obtain cell areas.

[0257] like Figure 8 As shown, the feature generation module 3 includes:

[0258] The single parameter preprocessing module 31 preprocesses the cell analysis parameters to obtain preprocessing parameters;

[0259] A three-dimensional vector generation module 32, the three-dimensional vector generation module 32 is connected to the single parameter preprocessing module 31;

[0260] The three-dimensional vector generation module 32 aligns the preprocessing parameters in the time dimension and then performs vectorization processing to obtain three-dimensional vector features;

[0261] The three-dimensional vector processing module 33 is connected with the three-dimensional vector generation module 32, and performs standardization and dimension reduction processing on the three-dimensional vector features to obtain three-dimensional features.

[0262] The above merely describes the preferred embodiments of the present application, but does not limit the embodiments and protection scope of the present application. It should be understood by those skilled in the art that any equivalent replacement and obvious changes made according to the content of the present application should be included in the protection scope of the present application.

Claims

1. A cell survival prediction method based on multi-parameter joint analysis, characterized in that: include: Step S1: administering the drug to be evaluated with a corresponding parameter combination to the target cells to prepare an evaluation sample; The parameter combination includes at least one variable parameter among concentration, dosage and incubation time; Step S2: collecting cell analysis parameters for the evaluation sample; The cell analysis parameters consist of surface plasmon resonance parameters, elastic coefficient and cell area; Step S3: constructing three-dimensional features based on the cell analysis parameters; Step S4: predicting the survival rate of the target cells under the parameter combination based on the three-dimensional features.

2. The cell survival prediction method according to claim 1, wherein The step S2 comprises: Step S21: irradiating the evaluation sample to collect the surface plasmon resonance parameters and elastic coefficients respectively; Step S22: photographing a cell image of the evaluation sample, and segmenting and quantifying the cell image to obtain the cell area.

3. The cell survival prediction method according to claim 1, wherein The step S3 comprises: Step S31: performing preprocessing on the cell analysis parameters to obtain preprocessing parameters; Step S32: aligning the preprocessing parameters in the time dimension and the concentration dimension, and then performing three-dimensional feature construction to obtain preconstructed features; Step S33: performing standardization and dimensionality reduction processing on the pre-constructed features to obtain the three-dimensional features; In step S32, the three-dimensional construction method includes at least one of performing time series modeling on parameters at multiple time points, fusing parameters at different resolutions, and modeling the interaction relationship between parameters; In step S33, the dimensionality reduction processing method includes at least one of t-SNE, PCA, UMAP, and autoencoder.

4. The cell survival prediction method according to claim 2, characterized in that: In step S31, the operation of pre-processing the surface plasmon resonance parameters includes: The surface plasmon resonance parameters are normalized and Gaussian noise is added, followed by t-SNE parameter optimization; The elastic coefficients were normalized and Gaussian noise was added, followed by t-SNE parameter optimization; The cell areas were normalized and Gaussian noise was added, followed by a square root transformation to reduce the bias of large area values, and finally t-SNE parameter optimization was performed; The standardization process includes applying at least one of Z-score standardization, robust standardization, and minimum-maximum standardization; The level of the Gaussian noise is in the range of 0.005-0.05; The parameters involved in the t-SNE parameter optimization include the T-SNE perplexity parameter and the t-SNElearning_rate parameter; Among them, the range of t-SNE perplexity parameter is between 50-150, t-SNE The range of the learning_rate parameter is between 1000-10000.

5. The cell survival prediction method according to claim 2, characterized in that: In step S33, when the method of performing dimensionality reduction processing on the pre-constructed features includes performing t-SNE parameter optimization, the optimization parameters include: Dimensionality reduction parameters: n_components = 3, perplexity = 30, n_iter = 30000; Initialization method: random initialization, early_exaggeration=2.0; Distance metric: Use cosine distance to capture the angular relationship between features; Optimization settings: learning_rate = 200, angle = 0.5 to balance speed and accuracy.

6. The cell survival prediction method according to claim 1, wherein In step S4, a gradient boosting classifier, a random forest classifier, a support vector machine classifier, a multilayer perceptron classifier, an XGBoost classifier, or a LightGBM classifier is used as a machine learning classifier for prediction; The gradient boosting classifier is integrated with a sequential learning strategy; The random forest classifier integrates a strategy of training multiple decision trees in parallel and making predictions through a voting mechanism; The LightGBM classifier integrates a leaf-first tree growing strategy, performs histogram optimization, and provides network parallel support.

7. The cell survival prediction method according to claim 6, characterized in that: Before executing step S1, a classifier training process is also included; The classifier training process includes: Step A1: preparing multiple sets of training samples by changing the parameter combination, incubating and observing the training samples to add labels corresponding to whether they are alive or not; Step A2: collecting training sample data corresponding to the cell analysis parameters for the training samples; The annotation is also added to the training sample data; Step A3: normalizing and adding noise to the training sample data, and then constructing a data set based on the annotation; The dataset is supervised learning data, and stratified sampling is used in the process of constructing the dataset to ensure class balance; Step A4: training the machine learning classifier using the data set and outputting the machine learning classifier; In step S4, hyperparameter optimization is performed on the machine learning classifier based on grid search or Bayesian optimization, and the machine learning classifier is evaluated based on at least one of accuracy, precision, recall and F1-score.

8. The cell survival prediction method according to claim 1, wherein After step S4, the method further includes: Step S5: Comparing the survival rates corresponding to a plurality of the parameter combinations to generate a comparison chart output.

9. The cell survival prediction method according to claim 1, wherein: Before executing step S1, the method further includes: SHAP value calculation is performed on the cell analysis parameters to explain the contribution of each feature to the prediction results; and, evaluating feature importance of the cell analysis parameters by feature scrambling; and performing feature interaction analysis on the cell analysis parameters to identify synergistic effects between features.

10. A cell survival prediction system based on multi-parameter joint analysis, characterized in that: Used to implement the cell survival prediction method according to any one of claims 1 to 9; The cell survival prediction system comprises: A sample preparation module, which delivers a drug to be evaluated with a corresponding parameter combination to target cells to prepare an evaluation sample; The parameter combination includes at least one variable parameter among concentration, dosage and incubation time; a sampling module, the sampling module being connected to the sample preparation module; The sampling module collects cell analysis parameters from the evaluation sample; The cell analysis parameters consist of surface plasmon resonance parameters, elastic coefficient and cell area; A feature generation module, the feature generation module is connected to the sampling module; The feature generation module constructs three-dimensional features based on the cell analysis parameters; A survival rate output module, the survival rate output module is connected to the feature generation module; The survival rate output module predicts the survival rate of the target cells under the parameter combination based on the three-dimensional features; A visualization module, the visualization module is connected to the survival rate output module; The visualization module generates a t-SNE visualization graph, a 3D scatter plot, a parameter correlation heat map, and a survival curve graph according to the predicted survival rate; An interactive parameter adjustment module, wherein the interactive parameter adjustment module is connected to the survival rate output module and the visualization module respectively; The interactive parameter adjustment module adjusts the classifier parameters and visualization parameters in real time; A batch processing module, the batch processing module is connected to the survival rate output module; The batch processing module processes multiple samples in parallel and outputs batch results; A result export module, the result export module is connected to the survival rate output module; The result export module responds to user instructions to generate results and reports in corresponding formats.

Images