Quantitative evaluation system and evaluation method for inflammation level of cells cultured in vitro

Abstract

The invention discloses a quantitative evaluation system and method for the inflammation level of cells cultured in vitro, and the system employs a modular design, and is composed of a data collection and preprocessing module, a feature gene selection module, a neural network modeling module, and an inflammation index calculation module. The evaluation method comprises the steps of data preparation, original data processing, feature gene screening, machine learning model screening, neural network model architecture, neural network weight analysis, inflammation index calculation and the like. The method has the advantages of high speed, high throughput, low cost, high accuracy, high mobility and high interpretability.

Description

Technical Field

[0001] This invention relates to an assessment of the inflammation level of in vitro cultured cells, specifically a quantitative assessment system and method for the inflammation level of in vitro cultured cells, applicable to drug efficacy verification, drug screening, etc., and belongs to the pharmaceutical field. Background Technology

[0002] Current cellular inflammation assessment technologies suffer from serious limitations, which not only affect the accuracy of inflammation detection but also restrict their widespread application in drug development. These limitations are mainly manifested in two aspects: First, traditional detection methods suffer from poor accuracy and insufficient representativeness. Traditional enzyme-linked immunosorbent assays (ELISA) and quantitative polymerase chain reactions (PCR) methods can only detect single or a few inflammatory factors, such as interleukin-6 and tumor necrosis factor-α. This single-point detection approach ignores the systemic and complex characteristics of the inflammatory response. Inflammation is a complex biological process involving the coordinated regulation of hundreds of genes, including multiple interconnected molecular networks such as the nuclear factor κB signaling pathway, the JAK-STAT signaling pathway, and the NLRP3 inflammasome activation mechanism. Single-indicator detection has a gene coverage of less than 5%, failing to capture these complex interactions, leading to significant systematic bias and incompleteness in the assessment results. Second, there are problems with high detection costs and low throughput. To comprehensively detect cellular inflammation levels, mass spectrometry techniques such as proteomics and metabolomics are required, resulting in low throughput and high costs. Furthermore, the complex sample pretreatment process, multi-step experimental operations, and specialized technical personnel often result in a comprehensive testing cost exceeding $500 per sample. This high cost severely limits the widespread application of this technology in large-scale drug screening, drug testing, and basic research, particularly in resource-constrained pharmaceutical companies and developing countries. Summary of the Invention

[0003] The purpose of this invention is to design a quantitative assessment system for the level of inflammation in in vitro cultured cells, which achieves a comprehensive breakthrough and innovation in inflammation assessment by integrating advanced machine learning algorithms, bioinformatics methods and systems biology theories.

[0004] The technical solution of this invention is as follows:

[0005] A quantitative assessment system for inflammation levels in in vitro cultured cells. The system adopts a modular design and consists of four core functional modules: a data collection and preprocessing module, a feature gene selection module, a neural network modeling module, and an inflammation index calculation module. The modules are connected through standardized data interfaces and APIs to form a complete end-to-end inflammation assessment solution.

[0006] The data preprocessing module receives raw RNA sequencing data and transforms it into a standardized gene expression matrix that can be processed by machine learning algorithms through a series of standardized bioinformatics processing procedures, including sequence quality control, genome alignment, expression quantification, and data standardization. This module employs state-of-the-art bioinformatics tools and algorithms, including STAR, TrimGalore, featureCounts, DESeq2, Seurat, fgsea, clusterProfile, GEOquery, and PCA, to ensure the accuracy and consistency of data processing.

[0007] The feature selection module is one of the core innovations of this invention. This module uses a support vector machine algorithm combined with cross-validation to systematically screen out a set of core genes closely related to inflammatory responses from tens of thousands of genes. Through rigorous statistical tests and biological verification, it ensures that the selected genes not only have good statistical significance, but also have clear biological significance and disease relevance.

[0008] The neural network modeling module adopts a multilayer perceptron architecture and builds a high-precision inflammation prediction model through deep learning. This module integrates a variety of advanced deep learning methods, including regularization, batch normalization, and adaptive learning rate adjustment, to ensure the training stability and prediction accuracy of the model.

[0009] The inflammation index calculation module converts the binary classification output of the neural network into a continuous 0-100 inflammation index. Through probability mapping and numerical transformation methods, it achieves a seamless conversion from discrete classification to continuous scoring, providing an intuitive and comparable quantitative indicator for applications such as drug screening.

[0010] A quantitative assessment method for the level of inflammation in in vitro cultured cells includes the following steps:

[0011] (1) Data preparation

[0012] Based on the construction and standardization of a large-scale multi-cell type inflammation model dataset, this dataset contains experimental data from inflammation models of six different cell types, covering the most important cell types and inflammation induction patterns in the immune system. Mouse primary microglia and human iPSC-derived microglia-like cells data are from the GSE221013 dataset; monocyte-derived macrophage data are from the GSE97672 dataset; human umbilical vein endothelial cell data are from the GSE267930 dataset; A549 lung cancer epithelial cell data are from the GSE142620 dataset; human nasal epithelial cell data are from the GSE241554 dataset; and human macrophage data are from the GSE193336 dataset.

[0013] (2) Raw data processing

[0014] Following internationally standardized RNA sequencing data analysis protocols, sequence alignment was first performed using software, employing strict quality control parameters to ensure the accuracy and reliability of the results. Next, gene expression levels were calculated using software. Quality control steps utilized single-cell and batch RNA sequencing data processing packages, employing multiple quality indicators, including gene detection rate, cell viability, and mitochondrial gene expression ratio, to filter out low-quality cells and genes, ensuring the quality of data for subsequent analyses. Data standardization employed a two-step method: first, a log2 transformation was performed with an increment of 1 to avoid zero-value issues in logarithmic operations, i.e., log2(CPM+1) transformation, where CPM represents the standardized expression level per million counts. This standardization method effectively reduces the impact of sequencing depth differences on gene expression level comparisons. Then, Z-sc... ORE standardization, by subtracting the mean and dividing by the standard deviation, transforms the expression levels of all genes into a standard normal distribution. This dual standardization strategy can minimize technical noise and batch effects, improving data comparability and analytical accuracy. The gene filtering step employs strict screening criteria to remove low-expression genes with a total count of less than 20 or detected in less than 60% of the samples. This filtering strategy can remove low-expression genes with significant technical noise while retaining biologically significant low-abundance but important regulatory genes, ultimately obtaining a high-quality expression matrix. Sample labels are clearly defined according to the experimental design: 0 represents the baseline state, i.e., the control group that did not receive inflammatory stimulation, and 1 represents the inflammatory state, i.e., the experimental group that received stimulation from various pro-inflammatory factors. This binary label design provides clear target variables for subsequent supervised learning model training.

[0015] (3) Using SVM models for feature gene screening

[0016] The Support Vector Machine (SVM) feature selection module represents a significant innovation in feature engineering in this invention. This module employs a SVM with a linear kernel function as the base classifier, using a systematic feature selection algorithm to screen the core gene set most relevant to the inflammatory response from tens of thousands of candidate genes. The choice of the SVM algorithm is based on its superior performance in high-dimensional data processing, especially in typical high-dimensional, small-sample problems such as gene expression data. SVM effectively avoids the curse of dimensionality, finding the optimal hyperplane by maximizing the classification margin. Simultaneously, its linear kernel function ensures the interpretability of the model, with the weight coefficient of each gene directly reflecting its contribution to inflammation classification. The algorithm is implemented... The current model uses a regularization parameter C of 0.01. This parameter was optimized through grid search and cross-validation, ensuring both the model's generalization ability and avoiding overfitting. A smaller C value means higher tolerance for classification errors, which is beneficial for selecting truly important feature genes rather than noisy genes. The application of the three-fold stratified cross-validation method ensures the stability and reliability of the feature selection results. Stratified sampling ensures that the proportion of inflammatory and non-inflammatory samples in each fold is consistent with the overall data, avoiding the impact of uneven sample distribution on the results. The three-fold design ensures sufficient validation while avoiding overly complex calculations. The final importance weight of each gene is calculated by averaging the results of the three-fold cross-validation, using the following formula:

[0017]

[0018] Where: K represents the total number of folds in the three-fold cross-validation;

[0019] wi,k represents the SVM model coefficients corresponding to gene i in the k-th fold of training.

[0020] The weight wi of each gene i is equal to the arithmetic mean of the absolute values ​​of the gene weights in the three-fold cross-validation. This calculation method takes into account both the magnitude and direction of the weights. The absolute value operation avoids the problem of positive and negative weights canceling each other out, and the average value calculation ensures the stability and representativeness of the results.

[0021] This invention designs and optimizes a core gene screening algorithm. Based on the results of support vector machine feature selection, it determines the final core gene set from candidate genes through a rigorous sorting and screening strategy. The algorithm first sorts genes in descending order according to their weights calculated by support vector machine; a higher weight indicates a more important contribution of the gene to inflammation classification. Then, a two-layer screening strategy is employed. The first layer of screening yields an expanded feature set of 40 genes, covering the main molecular pathways and regulatory networks in the inflammatory response. These include genes with the highest weights such as RRAD, DSE, FXYD5, GBP1, HMGN5, GBP3, SP140L, NAMPT, SYTL3, and SP140, as well as APOL6, PLAU, CD55, TRAM1, TSPAN13, C1R, DUSP5, TNFRSF9, CORO2A, NFKB2, NFKBIA, and SLC12. Genes with important biological functions, such as A8, PTP4A3, RNF24, CSF1, IER5, ATP8A1, CMPK2, NKD1, CD70, PRRG2, CFP, IL7, SOCS3, HIVEP2, SAMD9, ZNF608, FADS3, SPOCK2, and ELL2, were selected. The selection of these 40 genes was based not only on statistical weighting but also on detailed biological function verification and literature review, ensuring that each gene plays a crucial role in the inflammatory response. The second layer of screening further selected the 10 genes with the highest absolute weights from the 40 genes as the core gene set. These 10 genes are RRAD, DSE, FXYD5, GBP1, HMGN5, GBP3, SP140L, NAMPT, SYTL3, and SP140, representing the most critical molecular markers in the inflammatory response and possessing the highest application value. This dual-layer screening strategy is designed to meet the needs of different application scenarios. The 40-gene set is suitable for scientific research applications with extremely high accuracy requirements, while the 10-gene set is suitable for drug testing applications with high requirements for cost and efficiency. At the same time, it provides users with flexible options.

[0022] 40-gene feature set:

[0023] [RRAD,DSE,FXYD5,GBP1,HMGN5,GBP3,SP140L,NAMPT,SYTL3,SP140,

[0024] APOL6,PLAU,CD55,TRAM1,TSPAN13,C1R,DUSP5,TNFRSF9,CORO2A,NFKB2,

[0025] NFKBIA,SLC12A8,PTP4A3,RNF24,CSF1,IER5,ATP8A1,CMPK2,NKD1,CD70,

[0026] PRRG2,CFP,IL7,SOCS3,HIVEP2,SAMD9,ZNF608,FADS3,SPOCK2,ELL2]

[0027] 10-Gene Core Set:

[0028] [RRAD,DSE,FXYD5,GBP1,HMGN5,GBP3,SP140L,NAMPT,SYTL3,SP140]

[0029] (4) Machine learning model selection

[0030] By comparing the performance of four machine learning models—SVM, XGBoost, Random Forest, and MLP—on a classification task based on gene expression data, the AUC metric of three-fold cross-validation was used for evaluation. The results show that all models experienced a performance decline when the feature dimension was reduced from 40 genes to 10 genes, but to varying degrees. The overall trend indicates that feature selection has a significant impact on model performance, especially for models such as SVM.

[0031] (5) Neural Network Model Architecture

[0032] The multilayer perceptron neural network modeling fully considers the characteristics of gene expression data and the requirements of inflammation classification tasks. The network architecture adopts a classic feedforward neural network design. The number of nodes in the input layer is determined by the selected gene feature set; for a 40-gene feature set, the input layer contains 40 nodes, and for a 10-gene core set, the input layer contains 10 nodes. This flexible input design allows the system to adapt to different application scenarios and accuracy requirements. The first hidden layer contains 64 neurons. The choice of the number of neurons is based on extensive experimental verification and parameter optimization, ensuring both the model's expressive power and avoiding overfitting caused by over-parameterization. The activation function used is the Corrected Linear Unit (ReLU) function, which has advantages such as simple computation, good gradient propagation, and effective mitigation of the gradient vanishing problem, making it particularly suitable for training deep neural networks. The second hidden layer contains 32 neurons. The reduction in the number of neurons compared to the first hidden layer reflects the design concept of progressive feature abstraction. Higher-level abstract features are gradually extracted through hierarchical feature learning, and the ReLU activation function is also used to maintain consistency. To prevent overfitting and improve the model's generalization ability, each hidden layer... Dropout regularization layers were added after each layer, with a dropout rate of 0.5. This means that each neuron has a 50% probability of being randomly dropped during training. This randomness effectively prevents excessive dependencies between neurons, improving the robustness and generalization ability of the model. The application of the Dropout method is a classic method in deep learning to prevent overfitting, and its effectiveness has been proven in numerous practical applications. The output layer contains one neuron and uses the Sigmoid activation function. This function maps the output of the neural network to a probability range of 0 to 1, directly representing the probability that a sample belongs to an inflammatory state. This probability output provides an ideal numerical basis for subsequent calculation of the inflammation index. The model training uses the backpropagation algorithm and the Adam optimizer. The Adam optimizer combines the advantages of momentum and adaptive learning rate, and can automatically adjust the learning rate during training, improving training efficiency and stability. The loss function uses binary cross-entropy loss, which is particularly suitable for binary classification problems and can effectively guide the model to learn the correct classification boundary.

[0033] (6) Neural Network Weight Analysis

[0034] The neural network weight analysis method is a significant innovation in interpretability of this invention. By analyzing the weight parameters in the trained neural network model, this method extracts the contribution of each input gene to the final prediction result, providing interpretable biological evidence for drug screening applications and scientific research. The core of the analysis method is to extract the weight matrix from the first fully connected layer. For a model with 40 input genes, the weight matrix has a dimension of 40×64; for a model with 10 input genes, the weight matrix has a dimension of 10×64. Each element in the matrix represents the connection weight strength between the corresponding input gene and the hidden layer neuron. The absolute value of the weight reflects the importance of the connection to the model's prediction, and the positive or negative sign reflects the direction of activation or inhibition. The weight mapping step establishes a clear correspondence between each weight value in the weight matrix and its corresponding input gene. The accuracy and traceability of the weight analysis results are ensured by using gene name indexing. The comprehensive importance of each gene is obtained by calculating the absolute average weight value of the gene in all hidden neurons. The specific calculation formula is that the importance of gene i is equal to the arithmetic mean of the absolute values ​​of all hidden layer connection weights corresponding to that gene. This calculation method can comprehensively reflect the comprehensive influence of genes in the entire network and avoid the bias that may exist in single connection weights. The gene importance ranking is arranged in descending order according to the calculated importance value. The larger the importance value, the more important the gene's contribution to inflammation prediction. The ranking results provide researchers with clear information on biomarker priority, which helps to understand the model's decision logic and biological mechanisms.

[0035] (7) Inflammation index calculation

[0036] The inflammation index calculation module is the most innovative technical component of this invention. Through ingenious mathematical transformations and probability mapping techniques, this module achieves a seamless conversion from discrete binary classification output to continuous inflammation index, providing an intuitive and comparable quantitative evaluation indicator for industrial applications. The principle of this module is based on a multilayer perceptron neural network model built on the deep learning framework PyTorch. It fully utilizes the probabilistic output characteristics of the Sigmoid activation function to convert the model output, originally used for binary classification tasks, into continuous numerical values ​​with biological significance. The first step is device detection and environment configuration. The system detects whether there are CUDA-compatible GPU devices in the computing environment and prioritizes GPUs for computation to improve processing efficiency. Especially when processing large batches of samples, the parallel computing capabilities of GPUs can significantly shorten the computation time. When GPU devices are unavailable, the system automatically switches to CPU mode to ensure the versatility and deployability of the algorithm. In the data preprocessing stage, a flexible data format support strategy is adopted to identify and process the two most commonly used data formats, pandas DataFrame and numpy arrays. A type detection mechanism ensures the format compatibility of the input data. Feature selection is performed based on a predefined gene list, retaining only the expression levels of important genes used during model training as prediction input. This feature selection not only reduces the computational burden but, more importantly, ensures the accuracy and consistency of the prediction results.The data type conversion step converts the processed input data into PyTorch tensor format, specifying float32 precision to match the data type of the neural network model weight parameters, avoiding calculation errors caused by data type mismatch. Simultaneously, the tensor data is migrated to the previously determined computing device, preparing for subsequent model inference. The model inference process is executed in the torch.no_grad() context manager, which disables gradient calculation. Since backpropagation and weight updates are not required during inference, disabling gradient calculation not only saves memory but also significantly improves inference speed. Before inference, the model is set to evaluation mode (eval() mode) to ensure that regularization layers such as BatchNormalization and Dropout from training are used during inference. It exhibits correct behavior, avoiding inconsistencies between the training and inference phases. After obtaining the model's raw output, the system transfers the tensor data from the GPU to CPU memory through CPU transformation operations, and then extracts the final probability prediction values ​​through NumPy array flattening operations. These probability values ​​are located in the range of 0 to 1, directly reflecting the likelihood of the input sample being in an inflammatory state. The final calculation of the inflammation index is achieved through a simple and effective linear mapping. The specific formula is that the inflammation index equals the Sigmoid output probability value multiplied by 100. This linear transformation maps the probability values ​​in the range of 0 to 1 to the inflammation index range of 0 to 100. The higher the value, the greater the likelihood that the cell is in an inflammatory state and the more severe the inflammation. This intuitive numerical representation is easy for researchers to understand and apply, and also facilitates comparative analysis between different samples and different time points.

[0037] First, this invention achieves rapid, high-throughput, and low-cost inflammation detection capabilities. Through optimized algorithm design and technical routes, the system only requires RNA sequencing data as input to output a continuous inflammation index from 0 to 100, significantly simplifying the experimental process and technical requirements. It eliminates dependence on expensive proteomics detection equipment, reducing the cost per sample to less than 100 RMB, a reduction of over 80% compared to existing multi-omics methods. The detection time is shortened from several days using traditional methods to less than 2 hours. This significant improvement in cost and efficiency will enable inflammation assessment to be applied in a wider range of scenarios, including large-scale drug screening, routine drug testing, and basic scientific research.

[0038] Secondly, this invention significantly improves detection accuracy and cross-cell type transferability. By constructing a large-scale training dataset containing six different cell types, including representative immune-related cell types such as microglia, macrophages, endothelial cells, and epithelial cells, and employing advanced deep learning algorithms and feature engineering techniques, a universal inflammation assessment model across cell types is achieved. The average area under the receiver operating characteristic curve on the independent validation dataset reaches over 0.98, significantly surpassing the 0.85-0.90 level of existing methods. At the same time, the predictive performance stability of the model is guaranteed across different cell types, solving the key problem of lack of universality in existing technologies.

[0039] Third, this invention enhances biological interpretability and industrial application value. Through a systematic feature selection algorithm, it screens out 10 core gene feature sets and 40 extended gene feature sets. Each gene has clear biological function annotations and weight contribution analysis, providing interpretable molecular markers for drug testing and screening and scientific research. It can not only achieve accurate prediction of inflammation level, but also help researchers to deeply understand the molecular mechanism of inflammatory response, providing important theoretical guidance for the formulation of personalized treatment plans and new drug development.

[0040] The beneficial effects of this invention are as follows:

[0041] (1) Fast, high-throughput, and low-cost: Only RNA-Seq data is needed to output a continuous inflammatory index of 0-100;

[0042] (2) High accuracy and strong transferability: The average AUC across 6 cell types, including macrophages / endothelial cells / epithelial cells, is ≥0.98;

[0043] (3) High interpretability: Interpretable biomarker localization and weight calculation are achieved through a 10-gene core feature set.

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a method for quantitatively assessing the level of inflammation in in vitro cultured cells according to an embodiment of the present invention;

[0046] Figure 2 This is a representation diagram of the characteristic gene model in Example 40 of the present invention;

[0047] Figure 3 This is a representation of the core gene model in Example 10 of the present invention.

[0048] Figure 4 This is a model architecture diagram of an embodiment of the present invention. Detailed Implementation

[0049] The preferred embodiments of the present invention are described below. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] Example 1

[0051] A quantitative assessment system for inflammation levels in in vitro cultured cells is disclosed. The system adopts a modular design and consists of four core functional modules: a data collection and preprocessing module, a feature gene selection module, a neural network modeling module, and an inflammation index calculation module. The modules are connected through standardized data interfaces and APIs to form a complete end-to-end inflammation assessment solution.

[0052] The data preprocessing module receives raw RNA sequencing data and transforms it into a standardized gene expression matrix that can be processed by machine learning algorithms through a series of standardized bioinformatics processing steps, including sequence quality control, genome alignment, expression quantification, and data standardization. This module employs state-of-the-art bioinformatics tools and algorithms, including STAR (DNA sequence alignment), TrimGalore (quality control), featureCounts (reads counting), DESeq2 (differential analysis), Seurat (single-cell analysis), fgsea (enrichment analysis), clusterProfiler (GO analysis), GEOquery (data acquisition), and PCA (principal component analysis), ensuring the accuracy and consistency of data processing.

[0053] The feature selection module is one of the core innovations of this invention. This module uses a support vector machine algorithm combined with cross-validation to systematically screen a set of core genes closely related to inflammatory responses from tens of thousands of genes. Through rigorous statistical tests and biological verification, it ensures that the selected genes not only have good statistical significance, but also have clear biological significance and disease relevance.

[0054] The neural network modeling module employs a multilayer perceptron architecture and constructs a high-precision inflammation prediction model through deep learning. This module integrates various advanced deep learning methods, including regularization, batch normalization, and adaptive learning rate adjustment, to ensure the model's training stability and prediction accuracy. Additionally, it can also employ...

[0055] The inflammation index calculation module converts the binary classification output of the neural network into a continuous 0-100 inflammation index. Through probability mapping and numerical transformation methods, it achieves a seamless conversion from discrete classification to continuous scoring, providing an intuitive and comparable quantitative indicator for applications such as drug screening.

[0056] In terms of machine learning algorithm selection, the random forest algorithm is the most representative alternative. As a classic representative of ensemble learning methods, the random forest algorithm constructs multiple decision trees and obtains the final prediction result through a voting mechanism. This algorithm has a natural advantage in handling high-dimensional data and avoiding overfitting, and is particularly suitable for biomedical data such as gene expression data, which have high feature dimensions and relatively small sample sizes.

[0057] like Figure 1 As shown, a method for quantitatively assessing the level of inflammation in in vitro cultured cells includes the following steps:

[0058] (1) Data preparation

[0059] The data collection and preprocessing module is based on the construction and standardization of a large-scale multi-cell-type inflammation model dataset. The dataset was constructed following strict scientific design principles and quality control standards to ensure data representativeness, reliability, and reproducibility. This dataset contains experimental data from inflammation models of six different cell types, covering the most important cell types and inflammation induction patterns in the immune system. Mouse primary microglia and human iPSC-derived microglia-like cells data are from the GSE221013 dataset, containing 33 samples. A time-series design using lipopolysaccharide stimulation for 4 hours and 24 hours was employed, with corresponding unstimulated control groups. This design captures the dynamic changes and time-dependent characteristics of the inflammatory response. Monocyte-derived macrophage data are from the GSE97672 dataset, containing 64 samples. An infection model using three different influenza virus subtypes (H1N1, H3N2, and H5N1) was employed, with the multiplicity of infection (MNI) set to 2. Sampling was conducted at multiple time points (3 hours, 6 hours, 12 hours, and 18 hours), with a control group having a MNI of 0. This multi-virus, multi-time-point experimental approach... The design comprehensively reflects the inflammatory response patterns and temporal dynamics of macrophages to different pathogens. Human umbilical vein endothelial cell data, from the GSE267930 dataset, contains 9 samples and employs a dual inflammatory induction model of TNF-α stimulation and heat stress treatment at a temperature of 39 degrees Celsius. This design simulates the common clinical combination of cytokine-mediated and physical stress-induced inflammation. A549 lung cancer epithelial cell data, from the GSE142620 dataset, contains 36 samples and employs a long-term IL-1β stimulation model with stimulation time spans from 6 to 30 days, reflecting the development process of chronic inflammation and the molecular characteristics of long-term inflammatory states. Human nasal epithelial cell data, from the GSE241554 dataset, contains 6 samples and is treated with lipopolysaccharide stimulation, representing the inflammatory response pattern of upper respiratory tract epithelial cells. Human macrophage data, from the GSE193336 dataset, contains 8 samples and is also stimulated with lipopolysaccharide. The samples come from 4 different donors; this multi-donor design reflects individual biological differences and the individualized characteristics of inflammatory responses.

[0060]

[0061]

[0062] (2) Raw data processing

[0063] The raw data processing workflow strictly followed international standard RNA sequencing data analysis specifications. First, HISAT2 software was used for sequence alignment. This software is one of the most advanced RNA sequencing data alignment tools currently available, capable of efficiently and accurately aligning sequencing reads to the human reference genome GRCh38. Strict quality control parameters were employed during the alignment process to ensure the accuracy and reliability of the results. Next, featureCounts software was used to calculate gene expression levels. This software accurately counts the reads of each gene based on genome annotation information, providing reliable quantitative data for subsequent differential expression analysis. Quality control steps utilized professional single-cell and batch RNA sequencing data processing software packages such as Seurat and Scanpy. Multiple quality indicators, including gene detection rate, cell viability, and mitochondrial gene expression ratio, were used to filter out low-quality cells and genes, ensuring the data quality for subsequent analyses. Data standardization employed a two-step method: first, a log2 transformation was performed and 1 was added to avoid zero-value issues in logarithmic operations, i.e., log2(CPM). +1) Transformation, where CPM represents the standardized expression level per million counts. This standardization method effectively reduces the impact of sequencing depth differences on gene expression level comparisons. Then, Z-score standardization is performed, transforming the expression levels of all genes into a standard normal distribution by subtracting the mean and dividing by the standard deviation. This dual standardization strategy can eliminate technical noise and batch effects to the greatest extent, improving data comparability and analytical accuracy. The gene filtering step adopts strict screening criteria to remove low-expression genes with a total count of less than 20 or detected in less than 60% of the samples. This filtering strategy can remove low-expression genes with significant technical noise while retaining low-abundance but important regulatory genes with biological significance, ultimately obtaining a high-quality expression matrix of 68 samples and 10,287 genes. The sample labels are clearly defined according to the experimental design: 0 represents the baseline state, i.e., the control group that did not receive inflammatory stimulation, and 1 represents the inflammatory state, i.e., the experimental group that received stimulation from various pro-inflammatory factors. This binary label design provides clear target variables for subsequent supervised learning model training.

[0064] (3) Using SVM models for feature gene screening

[0065] The Support Vector Machine (SVM) feature selection module represents a significant innovation in feature engineering in this invention. This module employs a SVM with a linear kernel function as the base classifier, using a systematic feature selection algorithm to screen the core gene set most relevant to the inflammatory response from tens of thousands of candidate genes. The choice of the SVM algorithm is based on its superior performance in high-dimensional data processing, especially in typical high-dimensional, small-sample problems such as gene expression data. SVM effectively avoids the curse of dimensionality, finding the optimal hyperplane by maximizing the classification margin. Simultaneously, its linear kernel function ensures the interpretability of the model, with the weight coefficient of each gene directly reflecting its contribution to inflammation classification. The algorithm is implemented... The current model uses a regularization parameter C of 0.01. This parameter was optimized through grid search and cross-validation, ensuring both the model's generalization ability and avoiding overfitting. A smaller C value means higher tolerance for classification errors, which is beneficial for selecting truly important feature genes rather than noisy genes. The application of the three-fold stratified cross-validation method ensures the stability and reliability of the feature selection results. Stratified sampling ensures that the proportion of inflammatory and non-inflammatory samples in each fold is consistent with the overall data, avoiding the impact of uneven sample distribution on the results. The three-fold design ensures sufficient validation while avoiding overly complex calculations. The final importance weight of each gene is calculated by averaging the results of the three-fold cross-validation, using the following formula:

[0066]

[0067] Where: K represents the total number of folds in the three-fold cross-validation;

[0068] wi,k represents the SVM model coefficients corresponding to gene i in the k-th fold of training.

[0069] The weight wi of each gene i is equal to the arithmetic mean of the absolute values ​​of the gene weights in the three-fold cross-validation. This calculation method takes into account both the magnitude and direction of the weights. The absolute value operation avoids the problem of positive and negative weights canceling each other out, and the average value calculation ensures the stability and representativeness of the results.

[0070] This invention designs and optimizes a core gene screening algorithm. Based on the results of support vector machine feature selection, it determines the final core gene set from candidate genes through a rigorous sorting and screening strategy. The algorithm first sorts genes in descending order according to their weights calculated by support vector machine; a higher weight indicates a more important contribution of the gene to inflammation classification. Then, a two-layer screening strategy is employed. The first layer of screening yields an expanded feature set of 40 genes, covering the main molecular pathways and regulatory networks in the inflammatory response. These include genes with the highest weights such as RRAD, DSE, FXYD5, GBP1, HMGN5, GBP3, SP140L, NAMPT, SYTL3, and SP140, as well as APOL6, PLAU, CD55, TRAM1, TSPAN13, C1R, DUSP5, TNFRSF9, CORO2A, NFKB2, NFKBIA, and SLC12. Genes with important biological functions, such as A8, PTP4A3, RNF24, CSF1, IER5, ATP8A1, CMPK2, NKD1, CD70, PRRG2, CFP, IL7, SOCS3, HIVEP2, SAMD9, ZNF608, FADS3, SPOCK2, and ELL2, were selected. The selection of these 40 genes was based not only on statistical weighting but also on detailed biological function verification and literature review, ensuring that each gene plays a crucial role in the inflammatory response. The second layer of screening further selected the 10 genes with the highest absolute weights from the 40 genes as the core gene set. These 10 genes are RRAD, DSE, FXYD5, GBP1, HMGN5, GBP3, SP140L, NAMPT, SYTL3, and SP140, representing the most critical molecular markers in the inflammatory response and possessing the highest application value. This dual-layer screening strategy is designed to meet the needs of different application scenarios. The 40-gene set is suitable for scientific research applications with extremely high accuracy requirements, while the 10-gene set is suitable for drug testing applications with high requirements for cost and efficiency. At the same time, it provides users with flexible options.

[0071] Detailed list of genes attached:

[0072] (a) 40 gene feature sets

[0073] [RRAD,DSE,FXYD5,GBP1,HMGN5,GBP3,SP140L,NAMPT,SYTL3,SP140,

[0074] APOL6,PLAU,CD55,TRAM1,TSPAN13,C1R,DUSP5,TNFRSF9,CORO2A,NFKB2,

[0075] NFKBIA,SLC12A8,PTP4A3,RNF24,CSF1,IER5,ATP8A1,CMPK2,NKD1,CD70,

[0076] PRRG2,CFP,IL7,SOCS3,HIVEP2,SAMD9,ZNF608,FADS3,SPOCK2,ELL2]

[0077] (b) 10-gene core set

[0078] [RRAD,DSE,FXYD5,GBP1,HMGN5,GBP3,SP140L,NAMPT,SYTL3,SP140]

[0079] (4) Machine learning model selection

[0080] This study compares the performance of four machine learning models—SVM, XGBoost, Random Forest, and MLP—on a classification task based on gene expression data, using the AUC (Area Under Curve) metric based on three-fold cross-validation. The results show that all models experience performance degradation when the feature dimension is reduced from 40 genes to 10 genes, but to varying degrees. Figure 2-3 As shown, SVM's AUC decreased from 0.97 to 0.89 (a decrease of 8.2%), making it the most sensitive to changes; XGBoost decreased from 0.94 to 0.92 (a decrease of 2.1%), Random Forest decreased from 0.99 to 0.98 (a decrease of 1.0%), while MLP (Multilayer Perceptron) maintained an AUC of 0.99. This indicates that MLP has the strongest robustness when feature simplification is implemented. The overall trend shows that feature selection has a significant impact on model performance, especially for models such as SVM.

[0081] Model performance comparison table (three-fold cross-validation AUC)

[0082]

[0083] (5) Neural Network Model Architecture

[0084] Multilayer perceptron neural network modeling fully considers the characteristics of gene expression data and the needs of inflammation classification tasks. For example... Figure 4As shown, the network architecture adopts a classic feedforward neural network design. The number of nodes in the input layer is determined based on the selected gene feature set. For a 40-gene feature set, the input layer contains 40 nodes; for a 10-gene core set, the input layer contains 10 nodes. This flexible input design allows the system to adapt to different application scenarios and accuracy requirements. The first hidden layer contains 64 neurons. The choice of the number of neurons is based on extensive experimental verification and parameter optimization, ensuring both the model's expressive power and avoiding overfitting caused by excessive parameterization. The activation function used is the Corrected Linear Unit (ReLU) function, which has advantages such as simple computation, good gradient propagation effect, and effective mitigation of the gradient vanishing problem, making it particularly suitable for training deep neural networks. The second hidden layer contains 32 neurons. The reduction in the number of neurons compared to the first hidden layer reflects the design concept of feature abstraction at each level. Higher-level abstract features are gradually extracted through hierarchical feature learning. The ReLU activation function is also used to maintain consistency. To prevent overfitting and improve the model's generalization ability, each hidden layer contains 32 neurons. After each hidden layer, a Dropout regularization layer is added with a dropout rate of 0.5. This means that each neuron has a 50% probability of being randomly dropped during training. This randomness effectively prevents excessive dependencies between neurons, improving the robustness and generalization ability of the model. The application of the Dropout method is a classic method in deep learning to prevent overfitting, and its effectiveness has been proven in numerous practical applications. The output layer contains one neuron and uses the Sigmoid activation function. This function maps the output of the neural network to a probability range of 0 to 1, directly representing the probability that a sample belongs to an inflammatory state. This probability output provides an ideal numerical basis for subsequent calculation of the inflammation index. The model training uses the backpropagation algorithm and the Adam optimizer. The Adam optimizer combines the advantages of momentum and adaptive learning rate, and can automatically adjust the learning rate during training, improving training efficiency and stability. The loss function uses binary cross-entropy loss, which is particularly suitable for binary classification problems and can effectively guide the model to learn the correct classification boundary.

[0085] (6) Neural Network Weight Analysis

[0086] The neural network weight analysis method is a significant innovation in interpretability of this invention. By analyzing the weight parameters in the trained neural network model, this method extracts the contribution of each input gene to the final prediction result, providing interpretable biological evidence for drug screening applications and scientific research. The core of the analysis method is to extract the weight matrix from the first fully connected layer. For a model with 40 input genes, the weight matrix has a dimension of 40×64; for a model with 10 input genes, the weight matrix has a dimension of 10×64. Each element in the matrix represents the connection weight strength between the corresponding input gene and the hidden layer neuron. The absolute value of the weight reflects the importance of the connection to the model's prediction, and the positive or negative sign reflects the direction of activation or inhibition. The weight mapping step establishes a clear correspondence between each weight value in the weight matrix and its corresponding input gene. The accuracy and traceability of the weight analysis results are ensured by using gene name indexing. The comprehensive importance of each gene is calculated by calculating the absolute average weight value of the gene in all hidden neurons. The specific calculation formula is that the importance of gene i is equal to the arithmetic mean of the absolute values ​​of all hidden layer connection weights corresponding to that gene. This calculation method can comprehensively reflect the comprehensive influence of genes in the entire network and avoid the bias that may exist in single connection weights. The gene importance ranking is arranged in descending order according to the calculated importance value. The larger the importance value, the more important the gene's contribution to inflammation prediction. The ranking results provide researchers with clear biomarker priority information, which helps to understand the model's decision logic and biological mechanisms.

[0087] (7) Inflammation index calculation

[0088] The inflammation index calculation module is the most innovative technical component of this invention. Through ingenious mathematical transformations and probability mapping techniques, this module achieves a seamless conversion from discrete binary classification output to continuous inflammation index, providing an intuitive and comparable quantitative evaluation indicator for industrial applications. The principle of this module is based on a multilayer perceptron neural network model built on the deep learning framework PyTorch. It fully utilizes the probabilistic output characteristics of the Sigmoid activation function to convert the model output, originally used for binary classification tasks, into continuous numerical values ​​with biological significance. The first step is device detection and environment configuration. The system detects whether there are CUDA-compatible GPU devices in the computing environment and prioritizes GPUs for computation to improve processing efficiency. Especially when processing large batches of samples, the parallel computing capabilities of GPUs can significantly shorten the computation time. When GPU devices are unavailable, the system automatically switches to CPU mode to ensure the versatility and deployability of the algorithm. In the data preprocessing stage, a flexible data format support strategy is adopted to identify and process the two most commonly used data formats, pandas DataFrame and numpy arrays. A type detection mechanism ensures the format compatibility of the input data. Feature selection is performed based on a predefined gene list, retaining only the expression levels of important genes used during model training as prediction input. This feature selection not only reduces the computational burden but, more importantly, ensures the accuracy and consistency of the prediction results.The data type conversion step converts the processed input data into PyTorch tensor format, specifying float32 precision to match the data type of the neural network model weight parameters, avoiding calculation errors caused by data type mismatch. Simultaneously, the tensor data is migrated to the previously determined computing device, preparing for subsequent model inference. The model inference process is executed in the torch.no_grad() context manager, which disables gradient calculation. Since backpropagation and weight updates are not required during inference, disabling gradient calculation not only saves memory but also significantly improves inference speed. Before inference, the model is set to evaluation mode (eval() mode) to ensure that regularization layers such as BatchNormalization and Dropout from training are used during inference. It exhibits correct behavior, avoiding inconsistencies between the training and inference phases. After obtaining the model's raw output, the system transfers the tensor data from the GPU to CPU memory through CPU transformation operations, and then extracts the final probability prediction values ​​through NumPy array flattening operations. These probability values ​​are located in the range of 0 to 1, directly reflecting the likelihood of the input sample being in an inflammatory state. The final calculation of the inflammation index is achieved through a simple and effective linear mapping. The specific formula is that the inflammation index equals the Sigmoid output probability value multiplied by 100. This linear transformation maps the probability values ​​in the range of 0 to 1 to the inflammation index range of 0 to 100. The higher the value, the greater the likelihood that the cell is in an inflammatory state and the more severe the inflammation. This intuitive numerical representation is easy for researchers to understand and apply, and also facilitates comparative analysis between different samples and different time points.

Claims

1. A quantitative assessment system for the level of inflammation in in vitro cultured cells, characterized in that: The system adopts a modular design, consisting of four core functional modules: a data collection and preprocessing module, a feature gene selection module, a neural network modeling module, and an inflammation index calculation module. The modules are connected through standardized data interfaces and APIs. The data preprocessing module is responsible for receiving raw RNA sequencing data and converting it into a standardized gene expression matrix that can be processed by machine learning algorithms through a series of standardized bioinformatics processing procedures, including sequence quality control, genome alignment, expression quantification, and data standardization. The feature selection module uses a support vector machine algorithm combined with cross-validation to systematically screen a set of core genes closely related to inflammatory responses from tens of thousands of genes. Through rigorous statistical tests and biological validation, it ensures that the selected genes not only have good statistical significance, but also have clear biological significance and disease relevance. The neural network modeling module adopts a multilayer perceptron architecture and builds an inflammation prediction model through deep learning. The deep learning methods of this module include regularization, batch normalization, and adaptive learning rate adjustment to ensure the training stability and prediction accuracy of the model. The inflammation index calculation module converts the binary classification output of the neural network into a continuous 0-100 inflammation index. Through probability mapping and numerical transformation methods, it achieves a seamless conversion from discrete classification to continuous scoring, providing an intuitive and comparable quantitative indicator for applications such as drug screening.

2. A method for quantitatively assessing the level of inflammation in in vitro cultured cells, characterized by: Includes the following steps: (1) Data preparation Based on the construction and standardization of a large-scale multi-cell type inflammation model dataset, the dataset contains experimental data from inflammation models of six different cell types: mouse primary microglia and human iPSC-derived microglia-like cells from the GSE221013 dataset; macrophage data derived from monocytes from the GSE97672 dataset; human umbilical vein endothelial cells from the GSE267930 dataset; and A549 lung cancer epithelial cells from the GSE142620 dataset. Human nasal epithelial cell data came from the GSE241554 dataset; human macrophage data came from the GSE193336 dataset. (2) Raw data processing First, sequence alignment was performed, followed by gene expression level calculation. Quality control employed single-cell and batch RNA sequencing data processing software packages, using multiple quality indicators, including gene detection rate, cell viability, and mitochondrial gene expression ratio, to filter out low-quality cells and genes, ensuring data quality for subsequent analyses. Data standardization employed a two-step method: first, a log2 transformation was performed with an increment of 1 to avoid zero-value issues in logarithmic operations, i.e., log2(CPM+1) transformation, where CPM represents the standardized expression level per million counts; then, Z-score standardization was performed, converting the expression levels of all genes to a standard normal distribution by subtracting the mean and dividing by the standard deviation. Gene filtering used strict screening criteria, removing low-expression genes with a total count less than 20 or detected in less than 60% of the samples. Sample labels were clearly defined according to the experimental design: 0 represented the baseline state (control group without inflammatory stimulation), and 1 represented the inflammatory state (experimental group receiving stimulation from various pro-inflammatory factors). (3) Using SVM models for feature gene screening The support vector machine feature selection module uses a linear kernel support vector machine as the basic classifier. A systematic feature selection algorithm filters out the core gene set most relevant to the inflammatory response from tens of thousands of candidate genes. The algorithm implementation uses a regularization parameter C equal to 0.

01. A three-fold stratified cross-validation method and stratified sampling are applied. The final importance weight of each gene is calculated by averaging the three-fold cross-validation results, using the following formula: Where: K represents the total number of folds in the three-fold cross-validation; wi,k represents the SVM model coefficients corresponding to gene i in the k-th fold of training. The weight wi of each gene i is equal to the arithmetic mean of the absolute values ​​of the gene weights in the three-fold cross-validation. A core gene screening algorithm, based on the results of support vector machine feature selection, was employed to determine the final core gene set from candidate genes through a rigorous sorting and screening strategy. The algorithm first sorts genes in descending order based on their weights calculated by support vector machine; a higher weight indicates a more significant contribution of the gene to inflammation classification. Then, a two-layer screening strategy was used. The first layer yielded an expanded feature set of 40 genes, covering major molecular pathways and regulatory networks in the inflammatory response. These included genes with the highest weights: RRAD, DSE, FXYD5, GBP1, HMGN5, GBP3, SP140L, NAMPT, SYTL3, and SP140, as well as APOL6, PLAU, CD55, TRAM1, TSPAN13, C1R, etc. The genes with important biological functions are DUSP5, TNFRSF9, CORO2A, NFKB2, NFKBIA, SLC12A8, PTP4A3, RNF24, CSF1, IER5, ATP8A1, CMPK2, NKD1, CD70, PRRG2, CFP, IL7, SOCS3, HIVEP2, SAMD9, ZNF608, FADS3, SPOCK2, and ELL2. The second layer of screening further selects the 10 genes with the highest absolute weight from the 40 genes as the core gene set. These 10 genes are RRAD, DSE, FXYD5, GBP1, HMGN5, GBP3, SP140L, NAMPT, SYTL3, and SP140. (4) Machine learning model selection By comparing the performance of four machine learning models—SVM, XGBoost, Random Forest, and MLP—on a classification task based on gene expression data, the AUC metric of three-fold cross-validation was used for evaluation. The results show that all models experienced a performance decline when the feature dimension was reduced from 40 genes to 10 genes, but to varying degrees. The overall trend indicates that feature selection has a significant impact on model performance, especially for the SVM model. (5) Neural Network Model Architecture The multilayer perceptron neural network modeling adopts a classic feedforward neural network design. The number of nodes in the input layer is determined based on the selected gene feature set; for a 40-gene feature set, the input layer contains 40 nodes, and for a 10-gene core set, the input layer contains 10 nodes. The first hidden layer contains 64 neurons, and the activation function is the Modified Linear Unit (ReLU). The second hidden layer contains 32 neurons, also using the ReLU activation function to maintain consistency. A Dropout regularization layer is added after each hidden layer, with a dropout rate set to 0.

5. The output layer contains one neuron, using the Sigmoid activation function to map the neural network output to a probability range of 0 to 1, directly representing the probability that the sample belongs to an inflammatory state. Model training uses the backpropagation algorithm and the Adam optimizer; the loss function is the binary cross-entropy loss. (6) Neural Network Weight Analysis The neural network weight analysis method extracts the contribution of each input gene to the final prediction result by analyzing the weight parameters in the trained neural network model. The method extracts a weight matrix from the first fully connected layer. For a model with 40 input genes, the weight matrix has a dimension of 40×64; for a model with 10 input genes, the dimension is 10×64. Each element in the matrix represents the connection weight strength between the corresponding input gene and the hidden layer neuron. The absolute value of the weight reflects the importance of the connection to the model's prediction, and the positive or negative sign indicates the direction of activation or inhibition. The weight mapping step establishes a clear correspondence between each weight value in the weight matrix and its corresponding input gene. Gene name indexing ensures the accuracy and traceability of the weight analysis results. The overall importance of each gene is calculated by taking the absolute average weight value of that gene across all hidden neurons. The specific formula is that the importance of gene i equals the arithmetic mean of the absolute values ​​of all hidden layer connection weights corresponding to that gene. Gene importance is ranked in descending order based on the calculated importance values. (7) Inflammation index calculation The inflammation index calculation module achieves a seamless conversion from discrete binary classification outputs to continuous inflammation indices through mathematical transformations and probability mappings. A multilayer perceptron neural network model built on the PyTorch deep learning framework utilizes the probabilistic output characteristics of the sigmoid activation function to convert the model output originally used for binary classification tasks into biologically meaningful continuous values. The first step in implementation is device detection and environment configuration. The system detects the presence of CUDA-compatible GPUs in the computing environment, prioritizing GPUs for computation to improve processing efficiency. When GPUs are unavailable, the system automatically switches to CPU mode to ensure the algorithm's versatility and deployability. The data preprocessing stage employs a flexible data format support strategy, recognizing and processing pandas data. DataFrame and NumPy arrays, two of the most commonly used data formats, are used. A type detection mechanism ensures format compatibility of the input data. Feature selection is performed based on a predefined gene list, retaining only the expression levels of important genes used during model training as prediction input. The data type conversion step uniformly converts the processed input data to PyTorch tensor format, specifying float32 precision to match the data type of the neural network model weight parameters, avoiding calculation errors caused by data type mismatch. Simultaneously, the tensor data is migrated to the previously determined computing device, preparing for subsequent model inference. The model inference process is handled in torch.no_g Executed in the rad() context manager, the model is set to evaluation mode (eval() mode) before inference. After obtaining the model's raw output, the system transfers the tensor data from the GPU to CPU memory through CPU transformation operations, and then extracts the final probability prediction values ​​through NumPy array flattening operations. These probability values ​​are in the range of 0 to 1, directly reflecting the likelihood of the input sample belonging to an inflammatory state. The final calculation of the inflammation index is achieved through a simple and effective linear mapping. The specific formula is that the inflammation index is equal to the Sigmoid output probability value multiplied by 100. This linear transformation maps the probability values ​​in the range of 0 to 1 to the inflammation index range of 0 to 100.

3. The method for quantitatively assessing the level of inflammation in in vitro cultured cells according to claim 2, characterized in that, The gene feature set in step (3) 40 is: [RRAD,DSE,FXYD5,GBP1,HMGN5,GBP3,SP140L,NAMPT,SYTL3,SP140, APOL6,PLAU,CD55,TRAM1,TSPAN13,C1R,DUSP5,TNFRSF9,CORO2A,NFKB2, NFKBIA,SLC12A8,PTP4A3,RNF24,CSF1,IER5,ATP8A1,CMPK2,NKD1,CD70, PRRG2, CFP, IL7, SOCS3, HIVEP2, SAMD9, ZNF608, FADS3, SPOCK2, ELL2].

4. The method for quantitatively assessing the level of inflammation in in vitro cultured cells according to claim 2, characterized in that, The core set of 10 genes in step (3) is as follows: [RRAD,DSE,FXYD5,GBP1,HMGN5,GBP3,SP140L,NAMPT,SYTL3,SP140].

Images